Speed Sensorless State Estimation for Induction Motors: A ...web.mit.edu/leizhou/~lei/publications/2016_ACC_MHE_final.pdf · A dual-stage adaptive MHE, which performs parameter and

Speed Sensorless State Estimation for InductionMotors: A Moving Horizon Approach

Lei Zhou1 and Yebin Wang2

Abstract—This paper investigates the speed sensorless stateestimation problem for induction motors. Aiming at developingnew state estimation means to improve the estimation bandwidth,this paper proposes various moving horizon estimation (MHE)-based state estimators. Applying the MHE for induction motorsis not straightforward due to the fast convergence requirement,external torque disturbances, parametric model errors, etc. Toimprove speed estimation transient performance, we propose anMHE based on the full induction motor model and an assumedload torque dynamics. We further formulate an adaptive MHEto jointly estimate parameters and states and thus improverobustness of the MHE with respect to parametric uncertainties.A dual-stage adaptive MHE, which performs parameter and stateestimation in two steps, is proposed to reduce computationalcomplexity. Under certain circumstances, the dual-stage adaptiveMHE is equivalent to the case with a recursive least squarealgorithm for parameter estimation and a conventional MHEfor state estimation. Implementation issues and tuning of theestimators are discussed. Numerical simulations demonstrate thatthe proposed estimators can effectively estimate the inductionmotor states at a fast convergence rate, and the dual-stageadaptive MHE can provide converging state and parameterestimation despite the initial model parametric errors.

I. INTRODUCTION

In the speed sensorless control of induction motors, wherethe motor speed and position are not measured, the con-vergence rate of the state estimation is the key limitationto the motor’s tracking bandwidth. This fact motivates thedevelopment of new state estimation solutions for inductionmotor systems.

Speed sensorless state estimation for induction motors is achallenging problem since the motor dynamics is multivariableand nonlinear, and the motor parameters are often not exactlyknown. Through the years, numerous estimation schemes havebeen studied for induction motors. The classic model referenceadaptive system (MRAS) approach treats the motor speed asa time-varying parameter to avoid nonlinearity [1]–[4], butit often suffers from slow converging due to the adaptiveestimation. The sliding mode observer (SMO) treats the non-linear terms as bounded uncertainties and has achieved robustperformance [5], [6], but this often leads to an unnecessaryworst case design and degraded estimation accuracy. Theextended/unscented Kalman filter (EKF/UKF) schemes have

1Lei Zhou is with the Department of Mechanical Engineering, Mas-sachuestts Institute of Technology, Cambridge, MA 02139, USA. Email:[email protected]. This work was done while Lei Zhou was an internwith Mitsubishi Electric Research Laboratories.

2Yebin Wang is with with Mitsubishi Electric ResearchLaboratories, 201 Broadway, Cambridge, MA 02139, USA. Email:[email protected]

also been studied [7], [8], where the rotor mechanical equationis often not included. This formulation allows state estimationwithout knowing the motor mechanical parameters, but oftenresults in slow transient. Reference [9] performed EKF for in-duction motor with the mechanical dynamics included, whichhelps improving the transient performance and the estimationaccuracy at low speed.

In this paper we study the moving horizon estimation(MHE) for speed sensorless state estimation of inductionmotors, targeting at improving the convergence rate of thespeed estimation. The MHE has been initially introducedin [10] inspired by its widely used dual problem recedinghorizon control (RHC), and is receiving growing interest inthe past decade due to the advances in numerical optimizationsand computational capability of computers. References [11]and [12] have provided comprehensive studies of the MHEfor general linear and nonlinear systems, respectively.

The MHE for induction motor state estimation has beenexplored in [13] to achieve better estimation accuracy andbandwidth against MRAS and EKF estimators. However in[13] the motor speed is assumed to be constant over theestimation horizon, which may limit the speed estimationconvergence rate. Also [13] assumed exact knowledge ofmodel parameters, which is not always available in practice.

In this paper, the MHE considering the full dynamics ofthe induction motor is being studied, where the rotor speedis estimated as a state using the rotor’s equation of motion.Comparing with the constant speed assumption, the inclusionof the mechanical equation can improve the speed estimationconvergence rate and can improve the estimation accuracy atlow speed. This formulation, however, increases the estima-tor’s sensitivity with respect to the mechanical uncertaintiessuch as load variations and friction torque disturbances. Toaddress this, in our work the load torque is being estimated asa state variable with an assumed dynamics.

Another contribution of this paper is the inclusion of on-lineparameter estimation. It is well known that the performanceof MHE is significantly influenced by the model accuracy. Inorder to increase the estimator’s robustness in terms of para-metric uncertainties, the adaptive MHE is being studied, wherethe parameters are being estimated together with the states.Different formulations of the adaptive MHE for inductionmotors are introduced and discussed, and a dual-stage adaptiveMHE that decomposes state and parameter estimations isproposed. Our simulation shows that the dual-stage estimatordesign can effectively lower the implementation difficulty ofMHE for induction motors and can achieve accurate estimation

despite initial parameter errors.This paper is organized as follows. The induction motor

model and the general MHE formulation are briefly introducedin Section II. The MHE for induction motor state estimationincluding the mechanical dynamics is presented in SectionIII. Several adaptive MHE formulations for induction motorare presented in Section IV. Section V discusses the designand tuning of the estimators, and Section VI verifies theperformances of MHE and dual-stage adaptive MHE throughnumerical simulations. Conclusion is drawn in Section VII.

II. INDUCTION MOTOR MODEL AND GENERAL MHEA. Induction Motor Model

The induction motor model in the stationary two-phasereference frame can be written as

ids = −γids + αβψdr + βψqrω + uds/σ

iqs = −γiqs − βψdrω + αβψqr + uqs/σ

ψdr = αLmids − αψdr − ψqrωψqr = αLmiqs + ψdrω − αψqr

ω =µ

J(−idsψqr + ψdriqs)−

TLJ

y =[ids, iqs

]T,

(1)

where (ψdr, ψqr) are the rotor fluxes, (ids, iqs) are the statorcurrents, (uds, uqs) are the stator voltages, all defined in thestationary d-q frame. ω is the rotor speed, J is the rotor inertia,TL is the load torque, and y is the measurement. The restvariables in (1) denote model parameters, where σ = Ls(1−L2

m

LsLr), α = Rr

Lr, β = Lm

σLr, γ = Rs

σ + αβLm, µ = 32Lm

Lr;

(Rs, Ls) and (Rr, Lr) are the resistance and inductance of thestator and the rotor, respectively; Lm is the mutual inductance.

Speed sensorless estimation problem for an induction motoris roughly formulated as: design an estimator to reconstruct thefull state of the induction motor system (1) from measuringonly the stator currents (ids, iqs) and voltages (uds, uqs).

B. General MHE formulationThis section briefly introduces the general MHE formula-

tion to make this paper self-contained. Consider a nonlinearstochastic discrete-time system

xk+1 = fk(xk, uk) + wk

yk = hk(xk) + vk,(2)

where k is the time step, xk is the state, uk is the controlinput, yk is the output, wk is the process noise, and vk is themeasurement noise. The MHE at time T can be formulated asthe following constrained optimization problem

minz,wkT−1

k=T−N

ZT−N (z) +

T−1∑k=T−N

Lk(wk, vk)

subject toxk+1 = fk(xk, uk) + wk, k = T −N, ..., T − 1

vk = yk − hk(xk) ∈ Vk, k = T −N, ..., T − 1

xk(k; z, wj, uk) ∈ Xk, k = T −N, ..., Twk ∈Wk, k = T −N, ..., T − 1,

(3)

where N is the length of the estimation horizon definedbetween T − N and T − 1, and z = xT−N is the state atthe beginning of the estimation horizon. The sets Xk, Wk

and Vk denote the constraints on state, process noise, andmeasurement noise, respectively.

The cost function in (3) consists of two parts: the arrivalcost ZT−N (z) and the sum of the stage costs Lk(wk, vk)over the horizon. The stage cost Lk(wk, vk) penalizes onthe estimation errors wk and vk at each time step inside theestimation horizon, and the arrival cost ZT−N (z) summarizesthe past data that are not explicitly accounted for in theobjective function. A true arrival cost is defined as

ZT−N (z) = minx0,wkT−N−1

k=0

T−N−1∑k=0

Lk(wk, vk) + Γ(x0) (4)

and subject to constraints in (3) from 0 to T −N . Here Γ(x0)is the initial cost, penalizing on the deviation of the initialstate estimate from its true value. The MHE with the truearrival cost ensures that it has the same solution with the full-information estimation.

Remark 2.1: When MHE is used for nonlinear or con-strained systems, the exact expression for the true arrivalcost cannot be established [12]. An approximation of thearrival cost, denoted by ZT−N (z), is usually used. The arrivalcost approximation can significantly influence the estimationaccuracy and the stability of the estimator.

III. MHE FOR INDUCTION MOTORS

Work [13] considers the MHE for the induction motor withan assumed speed dynamics ω = 0. This treatment helpsto ameliorate numerical stability of the optimization problemderived from the MHE, however compromises the estimationperformance. In this section, we formulate the MHE using theinduction motor model with the mechanical equation included.

Assuming the load torque is slowly time-varying comparedto the motor states, we have TL = 0. By combining (1) andTL = 0, we can obtain a 6th-order induction motor model withthe state variables given by x = [ids, iqs, ψdr, ψqr, ω, TL]T .By discretizing this model and including the process andmeasurement noises, we can get a discrete-time stochasticmodel of the induction motor as

xk+1 = f(xk) +Buk + wk

yk = Cxk + vk.(5)

Note that in (5), B and C are constant matrices, while f(·) isa smooth vector field.

Remark 3.1: The main goal of including the load torque as astate variable is to improve the estimator’s robustness towardsmechanical uncertainties. When the motor is running, the loadtorque may be time-varying, and the Coulomb friction is alsoknown to deteriorate the estimator’s performance especiallyduring low speed operation. In order to maintain accurateestimation despite these uncertainties, the load torque is treatedas a state variables. We selected an assumed dynamics TL = 0since the motor load torque variation during operation isusually slow compared with the required speed bandwidth.

The MHE for the induction motor with the rotor speeddynamics can be formulated as

minz,wkT−1

k=T−N

ΦT = ZT−N (z) +

T−1∑k=T−N

Lk(wk, vk) (6)

subject to the system dynamics (5). A quadratic stage cost isselected as Lk(wk, vk) = wTkQ

−1wk + vTk R−1vk, where Q

and R are positive definite matrices and can be regarded asdesign parameters of the estimator. Specifically, when wk andvk are zero mean, independent Gaussian variables, the matricesQ and R can be selected as their covariance matrices.

The induction motor model is nonlinear. As is mentioned inRemark 2.1, there does not exist a closed-form expression forthe exact arrival cost. Here we use the filtering form of arrivalcost approximation introduced in [12]. Define the cost for theinitial estimation error as Γ(x0) = (x0 − x0)TΠ−10 (x0 − x0).The approximate arrival cost can be calculated by

ZT−N (z) = (z − xT−N )TΠ−1T−N (z − xT−N ) + Φ∗T−N ,

where Φ∗T−N is the computed optimal cost of the problem (6)at time T −N . The matrix ΠT−N is updated according to thefollowing matrix Ricatti equation

Πk+1 =Q+AkΠkATk

−AkΠkCT (R+ CTΠkC

T )−1CΠkATk ,

in which Ak = ∂f(xk)/∂xk.Remark 3.2: The MHE (6) does not include inequality

constraints for two reasons. First, this does not significantlyimprove the estimation performance, because the inequalityconstraints on induction motor states are loose and almostalways satisfied. Second, removing the inequality constraintscan simplify the optimization problem and greatly reduce thecomputational load.

IV. ADAPTIVE MHE FOR INDUCTION MOTORS

In induction motor systems, the model parameters areoften not exactly known as well as time-varying during theoperation. For example the electric heating incurs significantvariations of both the stator and rotor resistance values. On theother hand, it is well known that the MHE is a model-basedestimation scheme, and its performance highly relies on themodel accuracy. In order to improve the estimator’s robustnesswith respect to parametric model errors, we present an adaptiveMHE for the speed sensorless estimation, where the systemparameters are estimated together with states.

A. Augmented state MHE

One way to implement the adaptive MHE is throughaugmented state MHE, which is defined on the basis ofan augmented system dynamics. Define the vector of modelparameters as p = [α, β, γ, σ]T . We first expand the statex by including model parameters as an augmented statex′T = [x, p]T . Also define the augmented process noisesw′ = [w,wp]

T , where wp represent the mismatch betweenthe true model parameters and its estimate p. Consequently,assuming that the parameters are slowly time varying, we havethe augmented system dynamics given by (5) and p = 0. The

augmented state MHE is therefore formulated as the followingoptimization problem:

minz′,w′k

T−1k=T−N

Φ′T = Z ′T−N (z′) +

T−1∑k=T−N

L′k(w′k, v′k), (7)

and subject to the augmented system dynamics.In the augmented state MHE, the stage and arrival costs

are calculated using the same formulas as the non-adaptiveMHE (6), except that the augmented state x′ and processnoises w′ are used instead of x and w. The covariancematrix of the augmented process noises is defined as Q′ =diag(Q,Qp), where each diagonal component of the matrixQp = diag(Qα, Qβ , Qγ , Qσ) represents the weight on theestimation error of the individual model parameter.

B. Dual-stage adaptive MHE

In the augmented state MHE, the inclusion of parametersin the states results in a higher order and highly non-convexoptimization problem. This fact, however, adds significant dif-ficulties to the optimization problem solving. In order to makethe problem tractable, a dual-stage adaptive moving horizonestimator is proposed, where the parameter estimation andthe state estimation are decomposed into two sequential steps.Comparing with the original augmented state MHE, the dual-stage MHE can effectively reduce the size and complexity ofthe optimization problems, and therefore makes them relativelyeasy to solve with established nonlinear programming (NLP)solvers.

In the dual-stage adaptive MHE, two optimization problemsare solved sequentially at every time step for parameter andstate estimation. The parameter estimation can be achieved bysolving the optimization problem

minp

ΦpT = Φp∗T−Np+

T−1∑k=T−Np

vTk R−1vk, (8)

where Np is the length of the parameter estimation horizon.The arrival cost in (8) is selected as Φp∗T−Np

, which impliesthe estimator is totally forgetting the initial guesses. Thisselection is made because there is no dynamics involved inthe parameters to propagate the covariance of the parameterestimation error, i.e., p = 0. The stage cost in (8) is selectedas a quadratic form of the output error vk, and the penaltyon wk is not included. This is because wkT−1k=T−Nx

are thedecision variables of the state estimation and thus are fixedin the parameter estimation, so the quadratic term wTkQ

−1wkdoes not directly penalize on the model parameters. In theformulation (8), the parameter vector p is constant in theparameter estimation horizon, therefore the size of the cor-responding optimization problem is fixed and is independentto the horizon length Np.

The state estimation problem is given by

minz,wkT−1

k=T−Nx

ΦxT = ZT−Nx(z) +

T−1∑k=T−Nx

Lxk(wk, vk)

and subject to the induction motor model (5). Fig. 1 showsa block diagram of the dual-stage MHE. In this design, thestate estimation is based on the estimated parameter p from the

StateMHE(𝑁#)

ParameterMHE(𝑁%)

𝑥(*+,-*𝑢*+,-,

*+0 𝑦*+,-*

𝑢*+,2,*+0 𝑦*+,2

*

𝑧+0

Fig. 1. Block diagram of the dual-stage MHE.

previous iteration. By separating out the parameter estimation,the state estimation in the dual-stage MHE is reduced to theconventional MHE for state estimation.

Remark 4.1: The parameter estimation can be simplifiedby redefining the parameter vector as θ = [γ, αβ, β, 1/σ]T

and using only the first two state equations in (5), i.e., thestator current dynamics. The reason is two fold. First, allfour parameters are appearing in the stator current dynamics,therefore these two equations are sufficient to estimate allparameters. Second, the elements of θ are linearly involved inthe induction model equations, therefore the parameter estima-tion can be an unconstrained linear estimation problem. Withthis modification, the complexity of solving the correspondingoptimization problem can be significantly reduced.

C. RLS-based dual-stage adaptive MHE

The simplification to the parameter estimation in Re-mark 4.1 formulates the parameter estimation of the inductionmotor as an unconstrained linear system identification prob-lem, in which case, the recursive least square (RLS) estimationmethod can be readily applied for the parameter estimation.

For a linear equation y = φT θ, where φ is the input vectorand y is the vector of measurements, the RLS estimation givesthe estimated parameter θ that minimizes the accumulatedmean squared error

minθ

ΦRLST =1

T

T∑k=1

(yk|θ − yk

)2. (9)

To perform RLS-based parameter estimation for the induc-tion motor, the discretized stator current equations can bewritten as the following linear regression form[

ik+1ds −i

kds

dtik+1qs −i

kqs

dt

]︸ ︷︷ ︸

y

=

[−ikds ψkds ψkqsω

k ukds−ikqs ψkqs −ψkdsωk ukqs

]︸ ︷︷ ︸

φT

γαββ

1/σ

,︸ ︷︷ ︸

θ

(10)

where dt is the sampling interval. The RLS estimation algo-rithm can then be applied to (10) and identify the parameters.

Remark 4.2: By comparing the cost functions of the MHEparameter estimation formulation (8) and the RLS parameterestimation given in (9), we can conclude that with the sim-plification in Remark 4.1, the RLS parameter estimation isequivalent to the MHE parameter estimation with an infiniteparameter estimation horizon length and with the matrix R inthe cost function being an identity matrix.

V. DISCUSSION

A. Arrival cost

The arrival cost in MHE plays a crucial role in determiningthe behavior of the overall estimation process. Since a closed-

form expression for the true arrival cost does not exist fornonlinear or constrained systems, an approximation to thearrival cost need to be used. According to the stability analysisof the MHE in [12], asymptotic convergence of the estimationerror can be established if the approximated arrival cost isbounded by the true arrival cost.

Although this condition allows systematic stability analysis,a practical arrival cost synthetic method that meets this condi-tion is hard to find. In our development of MHE for inductionmotors, the filtering arrival cost approximation in [12] and [14]is being used, as was discussed in Section III.

Another commonly used approximation of the arrival costis ZT−N = Φ∗T−N . This arrival cost is independent of zand is totally ignoring the initial guesses. This arrival costapproximation satisfies the inequality conditions and thereforeasymptotic convergence can be guaranteed. However, thisselection does not necessarily give satisfactory performance[15]. With this approximation, the horizon length need to besufficiently large to achieve faster convergence.

As an alternative to the filtering approximation, a smoothingarrival cost approximation was proposed to eliminate theperiodic behavior of the estimator [11], [14]. In this design,more data are used in the update of z, where the arrivalcost covariance uses ΠT−N |T−1. In our implementation thisapproximation was not selected, mainly because the periodicalbehavior of the MHE does not significantly deteriorate theconvergence rate of the estimation.

B. Horizon length

Another important design parameter of the MHE is thehorizon length. Similar to its dual problem RHC, a largehorizon length is preferable for the MHE. Nevertheless, along estimation horizon will lead to a large scale optimizationproblem and overload the computational resources. Usuallythe horizon length is determined by balancing the trade-offbetween estimation performance and the computational time.

A longer estimation horizon allows the estimator to usemore data, and therefore the estimation is less dependent onthe approximation of the arrival cost. For the unconstrainedMHE, selecting a horizon length of N = 1 will reduce theMHE estimator to the EKF, where only the measurementsat the current time step are used in the estimation process.Intuitively, one can deduce that comparing with unconstrainedMHE having a horizon length larger than 1, the EKF is moresensitive to the initial error in states and covariances.

As a design parameter of the estimator, the horizon lengthmay be dynamically changed in the estimation process. Thisdegree-of-freedom is particularly interesting for the dual-stageadaptive MHE, where the model accuracy changes alongwith the parameter estimation process. In our implementation,different horizon lengths are selected for the state estimation inthe initial parameter estimation transient and in steady state.During the parameter estimation transient, a horizon lengthof 2 is selected for less trust to the model accuracy. Afterthe parameter estimation converges, a longer horizon length isused to provide more precise estimation.

TABLE IPARAMETERS OF INDUCTION MOTOR MODEL.

Parameter ValueStator resistance Rs 11.05ΩRotor resistance Rr 2.133ΩStator self-inductance Ls 0.23 HRotor self-inductance Lr 0.23 HMutual Inductance Lm 0.22 HRotor inertia J 0.0012 kgm2

Number of pole pairs p 2Motor power 250 W

𝜔"#$

-

-+

ωSpeedPI +

-TorquePI

FluxPI

Inductionmotor

InverseParkTransformation

ParkTransformation

MovingHorizon

Estimation

3 phase->

2phase

𝜔 %

𝜔 %

𝑖'"#$ 𝑢'

𝑢)

𝑢𝑑𝑠, 𝑖'-𝑖'

𝑖)

𝑢.𝑢/𝑢0

𝑢., 𝑖.

𝜃2

𝑢𝑞𝑠, 𝑖)-𝑢/, 𝑖/𝑢0, 𝑖0

𝑢'-

𝑢)-

+

𝑖)"#$2 phase

->3 phase

Fig. 2. Block diagram of induction motor vector control.

VI. NUMERICAL VALIDATION

A. Setup description

Numerical simulations are used to test the proposed MHEschemes. Table I shows the system parameters of the induc-tion motor used in the simulations. The simulation runs ata sampling rate of 10 kHz, and the Matlab OptimizationToolboxTM is used for solving the optimization problems. Theprocess and measurement noises are assumed to be zero-meanGaussian random processes, with the covariance matricesbeing Q = diag

(1× 10−4A2, 1× 10−4A2, 1× 10−4(V · s)2,

1 × 10−4(V · s)2, 1 × 10−4(rad/s)2, QTL(Nm)2

)and R =

diag(1 × 10−6A6, 1 × 10−4A2

), where QTL

can be selectedaccording to the motor’s operation conditions.

Fig. 2 shows a block diagram of the speed sensorless induc-tion motor system that is used in the numerical evaluations.The controllers form a standard indirect field oriented control,and thus the details are omitted. The simulation is conductedwith the control loops closed using the measured currents andthe estimated speed, and the proportional-integral (PI) trackingcontroller gains are kept the same under different test cases.

B. MHE state estimation

The proposed MHE formulation for induction motor stateestimation is compared with the transient performance ofEKF. Note that both estimators have the mechanical equationincluded in the model. In this simulation, the induction motorparameters are assumed to be exactly known. The initial statesare selected as ids = iqs = 1 A, ψdr = ψqr = 0 V · s,ω = 5 rad/s, TL = 0 Nm. The initial values of the estimatedstates were selected to be all zero values. The covariancematrix for initial state estimation error is Π0 = I6×6 × 10−3,and QTL

is selected as 1× 10−4 for not including the torqueestimation. An estimation horizon length of 20 time steps is

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Mo

tor

Sp

ee

d (

rad

/s)

0

50

100

150

Reference SpeedPlant SpeedEstimated Speed

Time (s)0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Sp

ee

d E

stim

atio

n E

rro

r (r

ad

/s)

-2

0

2

4

6MHE Speed Estimation ErrorEKF Speed Estimation Error

Fig. 3. State estimation for induction motor with MHE.

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Moto

r S

peed (

rad/s

)

0

50

100

150

Reference SpeedPlant Speed Q

TL = 1

Plant Speed QTL

= 10

Plant Speed QTL

= 100

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4S

peed E

stim

atio

nE

rror

(rad/s

)

-20

-10

0

10

20Speed Error Q

TL = 1

Speed Error QTL

= 10

Speed Error QTL

= 100

Time (s)0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Load torq

ue (

Nm

)

-0.5

0

0.5

1

1.5

True load torqueEstimated, Q

TL = 1

Estimated, QTL

= 10

Estimated, QTL

= 100

Fig. 4. Induction motor with MHE under step torque disturbance.

selected for the MHE. The simulated operating condition isspeed step responses, where the reference speed is 100 rad/sduring the time interval [0, 0.2s], and a reference speed stepof 20 rad/s is added at t = 0.2 s.

Fig. 3 shows the simulation results of the MHE for inductionmotor. In Fig. 3, the top plot shows the reference, plant andestimated speeds, and the bottom plot shows the estimationerror of the MHE and that of the EKF with the same initialconditions. It can be seen that the MHE with the mechanicalequation included can correctly estimate the speed of the in-duction motor and demonstrated a faster convergence transientcomparing with EKF. However when the reference speed stepis happening, the estimation error of the MHE experiences asmall transient (peak 0.25 rad/s), while the estimation error ofthe EKF barely deviates from zero.

The proposed MHE was also compared with the baselineMHE formulation in [13]. However with a step-type speedreference, our simulation shows that the baseline MHE hasa relatively slow estimation transient, and consequently thetracking controllers have to be tuned slower than the proposedMHE to ensure system stability. This observation coincideswith the fact that the baseline MHE will suffer slow transientdue to inherent adaptation-based speed estimation.

C. Load torque estimation

We also simulate the proposed MHE with the mechanicalequation to verify its ability to sustain step-type load torque

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4Mo

tor

Sp

ee

d (

rad

/s)

0

50

100

150

ReferencePlant speedEstimated

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Sp

ee

d E

stim

atio

nE

rro

r (r

ad

/s)

-20

0

20

40

Time0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Pa

ram

ete

r E

stim

atio

n E

rro

r (%

)

-50

0

50

γαβσ

Fig. 5. Simulation results of the dual-stage adaptive MHE.

disturbances, and the results are shown in Fig. 4. In this test,an estimation horizon of 10 time steps is selected. The initialconditions for the state and its estimate are taken as x =[1, 1, 0, 0, 5, 0]T and x = [0, 0, 0, 0, 0, 0]T , and the initial guesson error covariance is Π0 = I6×6 × 10−3.

In Fig. 4, the top plot shows the reference and plantspeed of the motor with estimators using different QTL

, themiddle plot shows the corresponding speed estimation errors,and the bottom plot presents the true load torque and theirestimates. The data show that the MHE formulation withload torque included in the state variables can successfullyreject disturbances in the load torque, and the load torquemodel error covariance QTL

determines the convergence rateof the torque and speed estimation. This observation matcheswith the performance of the 6th order EKF with load torqueestimation included [16], where a larger error covariance termQTL

gives a faster estimation transient.

D. Dual-stage adaptive MHE

The RLS-based dual-stage adaptive MHE is simulated withthe induction motor system. In this test case, the initial valuesof the parameter estimates are σ0 = 0.8σ, γ0 = 0.8γ,α0 = 0.9α, β0 = 0.9β. The horizon length of the MHE stateestimator is selected as Nx = 2 when 0 s ≤ t ≤ 0.1 s, andNx = 10 when 0.1 s < t ≤ 0.4 s.

The simulation results of the RLS-based dual-stage adaptiveMHE are shown in Fig. 5, where the top plot shows theplant and estimated speeds, the middle plot shows the corre-sponding speed estimation error, and the parameter estimationpercentage error is shown in the bottom plot. This simulationdemonstrates that the dual-stage adaptive MHE can success-fully incorporate the parameter estimation and give convergingestimation despite the existence of initial model parametricerrors, while the under these conditions non-adaptive MHE(6) and EKF fail to provide convergent state estimation.

VII. CONCLUSION AND FUTURE WORK

In this work, a moving horizon estimation (MHE) schemefor induction motor state estimation with the rotor mechanical

dynamics included was introduced, and a dual-stage adaptiveMHE formulation that offers parameter on-line estimation wasproposed. Simulation results show that the proposed MHEcan provide a relatively fast converging estimation transientand can reject torque disturbances, which will allow the usageof high bandwidth tracking controllers and therefore improvethe speed control bandwidth of the motor. The test resultsof the dual-stage adaptive MHE for induction motor showthat the proposed estimation scheme can successfully achieveconverging estimation when the model parameters are notexactly known initially. Future work should consider analysisand better tuning of the dual-stage adaptive MHE, whichwill significantly resolve the difficulties of the experimentalimplementation of MHE for induction motors.

REFERENCES

[1] D. J. Atkinson, P. P. Acarnley, and J. W. Finch, “Observers for inductionmotor state and parameter estimation,” IEEE Trans. Ind. Appl., vol. 27,no. 6, pp. 1119–1127, 1991.

[2] C. Schauder, “Adaptive speed identification for vector control of in-duction motors without rotational transducers,” IEEE Trans. Ind. Appl.,vol. 28, no. 5, pp. 1054–1061, 1992.

[3] H. Kubota, K. Matsuse, and T. Nakano, “Dsp-based speed adaptive fluxobserver of induction motor,” IEEE Trans. Ind. Appl., vol. 29, no. 2, pp.344–348, 1993.

[4] H. Kubota and K. Matsuse, “Speed sensorless field-oriented controlof induction motor with rotor resistance adaptation,” IEEE Trans. Ind.Appl., vol. 30, no. 5, pp. 1219–1224, 1994.

[5] A. Benchaib, A. Rachid, E. Audrezet, and M. Tadjine, “Real-timesliding-mode observer and control of an induction motor,” IEEE Trans.Ind. Electron., vol. 46, no. 1, pp. 128–138, 1999.

[6] M. Tursini, R. Petrella, and F. Parasiliti, “Adaptive sliding-mode observerfor speed-sensorless control of induction motors,” IEEE Trans. Ind.Appl., vol. 36, no. 5, pp. 1380–1387, 2000.

[7] Y.-R. Kim, S.-K. Sul, and M.-H. Park, “Speed sensorless vector controlof induction motor using extended kalman filter,” IEEE Trans. Ind. Appl.,vol. 30, no. 5, pp. 1225–1233, 1994.

[8] K. Shi, T. Chan, Y. Wong, and S. Ho, “Speed estimation of an inductionmotor drive using an optimized extended kalman filter,” IEEE Trans. Ind.Electron., vol. 49, no. 1, pp. 124–133, 2002.

[9] M. Barut, S. Bogosyan, and M. Gokasan, “Speed-sensorless estimationfor induction motors using extended kalman filters,” IEEE Trans. Ind.Electron., vol. 54, no. 1, pp. 272–280, 2007.

[10] K. R. Muske, J. B. Rawlings, and J. H. Lee, “Receding horizon recursivestate estimation,” in Proc. 1993 ACC, pp. 900–904.

[11] C. V. Rao, J. B. Rawlings, and J. H. Lee, “Constrained linear stateestimation: a moving horizon approach,” Automatica J. IFAC, vol. 37,no. 10, pp. 1619–1628, 2001.

[12] C. V. Rao, J. B. Rawlings, and D. Q. Mayne, “Constrained stateestimation for nonlinear discrete-time systems: Stability and movinghorizon approximations,” IEEE Trans. Automat. Control, vol. 48, no. 2,pp. 246–258, 2003.

[13] D. Frick, A. Domahidi, M. Vukov, S. Mariethoz, M. Diehl, andM. Morari, “Moving horizon estimation for induction motors,” in Proc.3rd IEEE Int. Sym. on Sensorless Control for Electrical Drives, 2012.

[14] M. J. Tenny and J. B. Rawlings, “Efficient moving horizon estimationand nonlinear model predictive control,” in Proc. 2002 ACC, pp. 4475–4480.

[15] E. L. Haseltine and J. B. Rawlings, “Critical evaluation of extendedkalman filtering and moving-horizon estimation,” Ind. & Eng. Chem.Research, vol. 44, no. 8, pp. 2451–2460, 2005.

[16] Y. Zhang, Z. Zhao, T. Lu, L. Yuan, W. Xu, and J. Zhu, “A comparativestudy of luenberger observer, sliding mode observer and extendedkalman filter for sensorless vector control of induction motor drives,”in Proc. Energy Conversion Congress and Exposition, 2009, pp. 2466–

2473.