A Two-Step Computationally Efficient Procedure for IMU ... · Procedure for IMU Classification and...

Post on 31-Jul-2020

2 views 0 download

Transcript of A Two-Step Computationally Efficient Procedure for IMU ... · Procedure for IMU Classification and...

A Two-Step Computationally EfficientProcedure for IMU Classification and Calibration

Gaetan Bakalli1

joint work withAhmed Radi2, Prof. Stephane Guerrier3,Yuming Zhang3, Dr. Roberto Molinari3

& Prof. Sameh Nassar2

1 University of Geneva2 University of Calgary

3 Pennsylvania State University

IEEE/ION PLANS , Monterey 2018

April 25, 2018G.Bakalli MGMWM April 25, 2018 1 / 29

Introduction Outline

Outline

1 Introduction1 Current Framework2 The GMWM

3 Near-Stationarity1 Motivation2 Definition3 Test and calibration4 Model Selection

4 Implementation1 mgmwm package and shiny

web-app2 Case Study

5 Conclusion

G.Bakalli MGMWM April 25, 2018 2 / 29

Introduction Current Framework

Wavelet Variance (Allan Variance) log-log plot

Scale τ

Wav

elet

Var

ianc

e ν2

Haar Wavelet Variance Representation

21 23 25 27 29 211

10-4

10-3

10-2

10-1

100

Empirical WV ν2

CI(ν2, 0.95)

G.Bakalli MGMWM April 25, 2018 3 / 29

Introduction GMWM

The GMWM estimator and its Model Selection Criterion

Definition: GMWM EstimatorThe GMWM estimator is the solution of the following optimizationproblem

θ = argminθ∈Θ

‖ν − ν(θ)‖2Ω,

in which Ω, a positive definite weighting matrix, is chosen in a suitablemanner such that the above quadratic form is convex.

The Wavelet Variance Information Criterion

WVIC = E[E0[‖ν0 − ν(θ)‖2

Ω

]],

where E[·] and E0[·] represent specific probabilistic expectations whichmeasure how well the WV implied by the estimated model fits the WVobserved on a future replicate.

G.Bakalli MGMWM April 25, 2018 4 / 29

Near-Stationarity Motivation

What we observe in reality

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

Run 1Run 2Run 3Run 4Run 5Run 6

Figure: Data coming from an ADIS16405 IMU.

Remarks:Several sequences ofthe error signalsissued from an IMU.

G.Bakalli MGMWM April 25, 2018 5 / 29

Near-Stationarity Motivation

Current Assumption on IMU’s Error Signal

2

G.Bakalli MGMWM April 25, 2018 6 / 29

Near-Stationarity Motivation

What we observe in reality

2

G.Bakalli MGMWM April 25, 2018 7 / 29

Near-Stationarity Definition

Near stationarity

ContextLower grade IMUs.Several recording of IMU error signal.Static conditions.

Near stationarityWe define a nearly stationary time series, as one which exhibit thefollowing properties:

1 Same model, but with different parameter values for eachsequences.

2 The vector of parameter θ has a probability distribution G(θ0),where E[θ] = θ0.

3 The distribution G(θ0) can be interpreted as the internal sensormodel, which may account for unobserved factors (e.g. temperature).

G.Bakalli MGMWM April 25, 2018 8 / 29

Near-Stationarity Test and calibration

Near-Stationarity test

2 2

H0 : k = 0 for all k<latexit sha1_base64="TbXna530UtRp3PxBdKeqGvCNo2g=">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</latexit><latexit sha1_base64="K7UhVGdOMS+/azk4C99O+vA1tt8=">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</latexit><latexit sha1_base64="K7UhVGdOMS+/azk4C99O+vA1tt8=">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</latexit><latexit sha1_base64="lEpoaz6G3eDo981VHdS3Kh/ebQA=">AAACKHicbVDJSgNBFOxxjXGLevTSGARPYcaLCwhBLzlGMEbIhOFN541p0rPQ/UYMw/yOF3/Fi4JKrn6Jk0Vwq1N11Xu8rvITJQ3Z9siam19YXFourZRX19Y3Nitb29cmTrXAlohVrG98MKhkhC2SpPAm0Qihr7DtDy7GfvsOtZFxdEXDBLsh3EYykAKokLxK3Q2B+gJU1sg9+5S7faAsc6mPBHnuDfgZ519Pz+Yu4T1lPIg1B6V4zgdepWrX7An4X+LMSJXN0PQqL24vFmmIEQkFxnQcO6FuBpqkUJiX3dRgAmIAt9gpaAQhmm42SZrz/ULpTc4HcUR8on7fyCA0Zhj6xeQ4l/ntjcX/vE5KwXE3k1GSEkZieihIFaeYj2vjPalRkBoWBISWxV+56IMGQUW55aIE53fkv6R1WDupOZdOtX4+a6PEdtkeO2AOO2J11mBN1mKCPbAn9srerEfr2Xq3RtPROWu2s8N+wPr4BIVxplo=</latexit>

Ha : H0 is false<latexit sha1_base64="VuD0dw1/P9GQrwGWEAe7WeBF82k=">AAACFnicbVC7SkNBEJ3r2/iKWmqxGASrcK+NjypoY6lgVEhCmLuZq4t7H+zOFcMlf2Hjn4iNhYqt2Nn5KW4SC18HFs6eM8PMnDDTyrLvv3sjo2PjE5NT06WZ2bn5hfLi0olNcyOpLlOdmrMQLWmVUJ0VazrLDGEcajoNL/f7/ukVGavS5Ji7GbViPE9UpCSyk9rlajNGvpCoi4NeG3fF968vmkzXXAhlRYTakui1yxW/6g8g/pLgi1Rqq3dHHwBw2C6/NTupzGNKWGq0thH4GbcKNKykpl6pmVvKUF7iOTUcTTAm2yoGd/XEulM6IkqNewmLgfq9o8DY2m4cusr+2va31xf/8xo5R9utQiVZzpTI4aAo14JT0Q9JdJQhybrrCEqj3K5CXqBByS7Kkgsh+H3yX1LfrO5Ug6OgUtuDIaZgBdZgAwLYghocwCHUQcIN3MMjPHm33oP37L0MS0e8r55l+AHv9RMGbaGi</latexit><latexit sha1_base64="S7E+Rs8RU9w0hfGeFOVGpYg72JQ=">AAACFnicbVDLSgNBEJz1bXxFPSoyGARPYdeLj1PQS44RjApJCL2T3mTI7IOZXjEsOfoHXvwT8eJBxat48xv8CSeJB40WDNRUddPd5SdKGnLdD2dicmp6ZnZuPrewuLS8kl9dOzdxqgVWRaxifemDQSUjrJIkhZeJRgh9hRd+92TgX1yhNjKOzqiXYCOEdiQDKYCs1MwX6yFQR4DKyv0mHPGfX5fXCa8p49LwAJRB3m/mC27RHYL/Jd43KZQ2708/b7buK838e70VizTEiIQCY2qem1AjA01SKOzn6qnBBEQX2lizNIIQTSMb3tXnO1Zp8SDW9kXEh+rPjgxCY3qhbysHa5txbyD+59VSCg4amYySlDASo0FBqjjFfBASb0mNglTPEhBa2l256IAGQTbKnA3BGz/5L6nuFQ+L3qlXKB2zEebYBttmu8xj+6zEyqzCqkywW/bAntizc+c8Oi/O66h0wvnuWWe/4Lx9AeM9owg=</latexit><latexit sha1_base64="S7E+Rs8RU9w0hfGeFOVGpYg72JQ=">AAACFnicbVDLSgNBEJz1bXxFPSoyGARPYdeLj1PQS44RjApJCL2T3mTI7IOZXjEsOfoHXvwT8eJBxat48xv8CSeJB40WDNRUddPd5SdKGnLdD2dicmp6ZnZuPrewuLS8kl9dOzdxqgVWRaxifemDQSUjrJIkhZeJRgh9hRd+92TgX1yhNjKOzqiXYCOEdiQDKYCs1MwX6yFQR4DKyv0mHPGfX5fXCa8p49LwAJRB3m/mC27RHYL/Jd43KZQ2708/b7buK838e70VizTEiIQCY2qem1AjA01SKOzn6qnBBEQX2lizNIIQTSMb3tXnO1Zp8SDW9kXEh+rPjgxCY3qhbysHa5txbyD+59VSCg4amYySlDASo0FBqjjFfBASb0mNglTPEhBa2l256IAGQTbKnA3BGz/5L6nuFQ+L3qlXKB2zEebYBttmu8xj+6zEyqzCqkywW/bAntizc+c8Oi/O66h0wvnuWWe/4Lx9AeM9owg=</latexit><latexit sha1_base64="S08iFJm3LbAsifM4mu3OaVySKdc=">AAACFnicbVC5TsNAEF1zEy4DJc2KCIkqsmk4qgialCARiJRY0XgzTlasD+2OEZGVv6DhV2goANEiOv6GTXARjiet9Pa9Gc3MCzMlDXnepzMzOze/sLi0XFlZXVvfcDe3rkyaa4FNkapUt0IwqGSCTZKksJVphDhUeB3enI3961vURqbJJQ0zDGLoJzKSAshKXbfWiYEGAlTRGHXhhE9/Pd4hvKOCS8MjUAb5qOtWvZo3Af9L/JJUWYnzrvvR6aUijzEhocCYtu9lFBSgSQqFo0onN5iBuIE+ti1NIEYTFJO7RnzPKj0epdq+hPhEne4oIDZmGIe2cry2+e2Nxf+8dk7RUVDIJMsJE/E9KMoVp5SPQ+I9qVGQGloCQku7KxcD0CDIRlmxIfi/T/5Lmge145p/4Vfrp2UaS2yH7bJ95rNDVmcNds6aTLB79sie2Yvz4Dw5r87bd+mMU/Zssx9w3r8A+KmfXQ==</latexit>

G.Bakalli MGMWM April 25, 2018 9 / 29

Near-Stationarity Test and calibration

Multisignal GMWM (MGMWM)

MGMWMConsidering k = 1, . . . ,K , the number replicates recorded from the same IMU in staticcondition, we define the Multisignal GMWM estimator as the solution of the followingminimization problem:

θ = argminθ∈Θ

‖ν − ν(θ)‖2Ω,

where ν represent the stack of WV computed and each sequences, and Ω a blockdiagonal weighting matrix.

MGMWM considering independent signals

θ = argminθ∈Θ

1K

K∑k=1

‖νk − ν(θ)‖2Ωk .

Let νjk , respectively νj (θ) be the j th elements of the vectors νk , the empirical WVcomputed on the k th replicates of size j = 1, . . . , Jk .

G.Bakalli MGMWM April 25, 2018 10 / 29

Near-Stationarity Model Selection

Back to the WVIC

The Wavelet Variance Information Criterion

WVIC = E[E0[‖ν0 − ν(θ)‖2

Ω

]],

where E[·] and E0[·] represent specific probabilistic expectations whichmeasure how well the WV implied by the estimated model fits the WVobserved on a future replicate.

G.Bakalli MGMWM April 25, 2018 11 / 29

Near-Stationarity Model Selection

WVIC with Multiple Signal Replicates

G.Bakalli MGMWM April 25, 2018 12 / 29

Implementation Package and Shiny web-app

mgmwm.smac-group.com

G.Bakalli MGMWM April 25, 2018 13 / 29

Implementation Case Study

Empirical WV

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

Run 1Run 2Run 3Run 4Run 5Run 6

Figure: Empirical WV of 6 replicates coming from an ADIS 16405 IMU gyroscopes Y-axisrecorded a 100hrz

G.Bakalli MGMWM April 25, 2018 14 / 29

Implementation Case Study

First Model estimated

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

Implied WVAR1AR1AR1RWWN

Figure: Empirical and Implied WV (in orange) of 6 replicates coming from an ADIS 16405 IMUgyroscopes Y-axis recorded a 100hrz. Model fitted is composed of 3 AR1 + RW + WN. Plainlines represent the contribution of each individual process.

G.Bakalli MGMWM April 25, 2018 15 / 29

Implementation Case Study

Selected model trough WVIC

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

Implied WVAR1AR1AR1

Model selected

Figure: Empirical and Implied WV (in orange) of 6 replicates coming from an ADIS 16405 IMUgyroscopes Y-axis recorded a 100hrz. Model selected is composed of 3 AR1. Plain linesrepresent the contribution of each individual process.

G.Bakalli MGMWM April 25, 2018 16 / 29

Implementation Case Study

WVIC comparison for all nested models.

CV-WVIC (Model) - CV-WVIC (3*AR1)

CI for CV-WVIC

1 100 10000

2*AR1+RW

WN+3*AR1

WN+2*AR1+RW

2*AR1

3*AR1+RW

WN+3*AR1+RW

WN+2*AR1

WN+AR1+RW

AR1+RW

WN+AR1

AR1

WN+RW

WN

RW

|

|

|

|

|

|

|

| |

| |

| |

| |

| |

| |

||

Figure: Comparison of the WVIC criteria for every model nested in 3 AR1 + RW + WN withthe selected model (log scale). The dots represent the value of the WVIC, the lines alongside itsrespective confidence interval. Model in green represent ”equivalent” models with less orequivalent model complexity.

G.Bakalli MGMWM April 25, 2018 17 / 29

Implementation Case Study

Implied WV for equivalent models

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

2*AR12*AR13*AR12*AR13*AR12*AR1+RW

2*AR13*AR12*AR1+RWWN+2*AR1+RW

2*AR13*AR12*AR1+RWWN+2*AR1+RWWN+2*AR1

Figure: Fit comparison for ”equivalent” models.

G.Bakalli MGMWM April 25, 2018 18 / 29

Implementation Case Study

WVIC comparison for equivalent nested models.

CV-WVIC (Model) - CV-WVIC (3*AR1)

CI for CV-WVIC

-0.2 0.0 0.2 0.4 0.6

2*AR1+RW

WN+2*AR1+RW

2*AR1

WN+2*AR1

| |

| |

| |

| |

Figure: Comparison of the WVIC criteria for ”equivalent” model nested in 3 AR1 + RW + WNwith the selected model.

G.Bakalli MGMWM April 25, 2018 19 / 29

Implementation Uncertainty Evaluation

Uncertainty Evaluation: Near-Stationary Case

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

Run 1Run 2Run 3Run 4Run 5Run 6

G.Bakalli MGMWM April 25, 2018 20 / 29

Implementation Uncertainty Evaluation

Uncertainty Evaluation: Near-Stationary Case

Scale(s)

Wav

elet

Var

iance

2

ADIS 16405 IMU 100 Hz Gyroscope Y-Axis

10-1 100 101 102 103

10-5

10-4

10-3

10-2

10-1

Run 1Run 2Run 3Run 4Run 5Run 6

G.Bakalli MGMWM April 25, 2018 21 / 29

Implementation Uncertainty Evaluation

Confidence Intervals

0.9990 0.9992 0.9994 0.9996 0.9998

Sm

oot

hed

Den

sity

1

G.Bakalli MGMWM April 25, 2018 22 / 29

Implementation Uncertainty Evaluation

Confidence Intervals

0.17 0.18 0.19 0.20 0.21 0.22

Sm

oot

hed

Den

sity

3

G.Bakalli MGMWM April 25, 2018 23 / 29

Conclusions

Conclusion and Further Development

ConclusionsConcept of near stationarity.Framework for Multisignal calibration and model selection.Implementation in mgmwm package and web-app.

DevelopmentsSpeed-up the computation in the R package.Create a data-loader for various imu’s.Extend this set-up for GNSS data.

G.Bakalli MGMWM April 25, 2018 24 / 29

Conclusions

Conclusion and Further Development

ConclusionsConcept of near stationarity.Framework for Multisignal calibration and model selection.Implementation in mgmwm package and web-app.

DevelopmentsSpeed-up the computation in the R package.Create a data-loader for various imu’s.Extend this set-up for GNSS data.

G.Bakalli MGMWM April 25, 2018 25 / 29

Conclusions

Related Presentations

Session A6An Optimal Virtual Inertial Sensor Framework using WaveletCross Covariance: Yuming Zhang, Pennsylvania State University,USA; Haotian Xu, University of Geneva, Switzerland; Ahmed Radi,University of Calgary, Canada; Roberto Molinari, Stephane Guerrier,Pennsylvania State University, USA; Naser El-Sheimy, University ofCalgary, Canada.Construction of Dynamically-Dependent Stochastic ErrorModels: Philipp Clausen, Swiss Federal Institute of Technology inLausanne, Switzerland; Samuel Orso, University of Geneva,Switzerland; Jan Skaloud, Swiss Federal Institute of Technology inLausanne, Switzerland; StÃľphane Guerrier, Pennsylvania StateUniversity, USA.

G.Bakalli MGMWM April 25, 2018 26 / 29

Conclusions Thanks!

Thank you very much for your attention!

A special thanks toProf. Naser El-Sheimy (U. of Calgary)Justin Lee (Penn State University)

Any questions?

More info...

SMAC-group.comgithub.com/SMAC-Groupgaetan.bakalli@unige.ch@SMAC_Group

G.Bakalli MGMWM April 25, 2018 27 / 29

Conclusions Thanks!

Two Estimators: Average GMWM vs MGMWM

Average GMWM vs MGMWM

We define θ as the Average GMWM (AGMWM) defined as:

θ = 1K

K∑k=1

θk ,

with θk = argminθk∈Θ ‖νk − ν(θk)‖2Ωk

. Remeber that we define theMultisignal GMWM (MGMWM) as

θ? = argminθ∈Θ

1K

K∑k=1‖νk − ν(θ)‖2

Ωk .

G.Bakalli MGMWM April 25, 2018 28 / 29

Conclusions Thanks!

Two Estimators: Average GMWM vs MGMWM

PropertiesIt turns out that the MGMWM appears far more appropriate than the AGMWM for twomain reasons:

The MGMWM is more efficient than the AGMWM, i.e.

tr(

minΩi var[θ])

tr(

minΩi var[θ?]) P7−−→ c > 1.

From Jensen inequality implies that

If ν(θ) is linear, i.e for stochastic processes WN, DR and QN, or ifG(θ0) is a Dirac function,

θ? − θP7−−→ 0.

If ν(θ) is not linear i.e for stochastic processes RW and AR1, than

θ? − θP7−−→ δ 6= 0.

G.Bakalli MGMWM April 25, 2018 29 / 29