recursive least squares algorithm dedicated to early
play

RECURSIVE LEAST SQUARES ALGORITHM DEDICATED TO EARLY RECOGNITION OF - PowerPoint PPT Presentation

RECURSIVE LEAST SQUARES ALGORITHM DEDICATED TO EARLY RECOGNITION OF EXPLOSIVE COMPOUNDS THANKS TO MULTI- TECHNOLOGY SENSORS Aurlien MAYOUE, Aurlie MARTIN, Guillaume LEBRUN and Anthony LARUE ICASSP 2013, VANCOUVER OUTLINE OVERVIEW


  1. RECURSIVE LEAST SQUARES ALGORITHM DEDICATED TO EARLY RECOGNITION OF EXPLOSIVE COMPOUNDS THANKS TO MULTI- TECHNOLOGY SENSORS Aurélien MAYOUE, Aurélie MARTIN, Guillaume LEBRUN and Anthony LARUE ICASSP 2013, VANCOUVER

  2. OUTLINE • OVERVIEW • Context • Prototype • RECURSIVE LEAST SQUARES ALGORITHM • Theorical basis & Principle • One dimensional signal case • Regularization • Multisensor adaptation • EXPERIMENTS • Description • Results ICASSP 2013 | MAYOUE Aurélien | 2

  3. CONTEXT compounds detection EGDN DNT e-nose identification NM TNT gas sensor + algorithm quantification interferents signal GAS SENSOR ALGORITHM detection Recursive Least Squares interferents EGDN NM DNT TNT - analyte decision - sensitive material - technology ICASSP 2013 | MAYOUE Aurélien | 3

  4. PROTOTYPE Prototype based on a gas sensor array: technology active layers Fluorescence 1 (OPTO) Quartz Crystal 2 Microbalance (QCM) Surface Acoustic 2 Wave (SAW) QCM OPTO SAW inhaler ICASSP 2013 | MAYOUE Aurélien | 4

  5. THEORICAL BASIS first order response linear drift   ( ) t − α β = δ − + α + β f t Q e t  τ  , , , Q. . 1 . Langmuir model: τ δ   , = + • parameters depending on the absorption affinity between the unknown gas and the sensor: used to build • τ time constant models • δ sensitivity • parameters depending on experimental conditions: • Q concentration of the compound estimated by • α slope of the sensor linear drift RLS algorithm • β sensor offset ICASSP 2013 | MAYOUE Aurélien | 5

  6. PRINCIPLE interferents interferents NM EGDN NM EGDN τ c δ c ( ) ( ) ( , ) c α c β c Q ( ) ( ) ( ) ( , , ) are set using are estimated in training such a way examples to each model build models best fits the real data ACQUISITION DNT TNT TNT DNT MODELS RECURSIVE LEAST SQARES e DNT P DNT DECISION P TNT e TNT e EGDN P EGDN P NM e NM e interferents P interferents ICASSP 2013 | MAYOUE Aurélien | 6

  7. RLS: ONE DIMENSIONAL SIGNAL CASE Least Squares : = θ + Ε H Z Μ   • Z acquisition vector     Q   • H model matrix t   - = δ α + Ε τ   • θ vector of parameters (1 - e ) t 1     • E error Μ β       − θ = T T ˆ H H H Z 1 Pseudo-inverse solution: ( ) Recursive Least Squares : Z H  Ε       k   k   k  = θ +       ε z h       k + k + k + 1 1 1 ⇔ = θ + Ε Z H + + + k k k 1 1 1   0   T θ ˆ = P h h P   0 0 = − + +   θ = θ + − θ P P k k k k ˆ ˆ P h T z h ˆ 1 1 with   Solution: ( ) 0 + k k k + k k + k + k + k + k + T 1 h P h 1 1 1 1 1 1 k + k k + 1 1 P = Id 0 ICASSP 2013 | MAYOUE Aurélien | 7

  8. RLS: REGULARIZATION Q, α and β can freely evolve: the sensor drift and the exponential part cannot be discriminated correctly Real Data Hyp. TNT Hyp. EGDN Hyp. EtOH Hyp. DCM Hyp. MEK Estimated Data ( )  Γ  0 0   Q Reguralization : θ = θ − 2 + Γ θ 2 ˆ H Z with Γ = Γ   arg min 0 0 α   θ Γ 0 0   β Г Q , Г α and Г β are used to set each parameter inertial. Real Data Hyp. TNT Hyp. EGDN Hyp. EtOH Hyp. DCM Hyp. MEK Estimated Data   0   T θ = P h h P ˆ   0 0 = −   P P k k + k + k θ = θ + T − θ ˆ ˆ P h z h ˆ 1 1 with  0  Solution: ( ) + k k + + + + + + T k k k k k k k 1 h P h 1 1 1 1 1 1 + + k k k 1 1 = Γ T Γ − P 1 ( ) 0 ICASSP 2013 | MAYOUE Aurélien | 8

  9. RLS: MULTISENSOR CASE Z ICASSP 2013 | MAYOUE Aurélien | 9

  10. RLS: MULTISENSOR CASE Z ICASSP 2013 | MAYOUE Aurélien | 10

  11. RLS: MULTISENSOR CASE Z - work in real time - process samples from sensors with different sampling frequencies - discriminate compounds with different kinetics and/or amplitude ratio from the multi-sensor ICASSP 2013 | MAYOUE Aurélien | 11

  12. EXPERIMENTS: DESCRIPTION Compounds : EtOH DCM MEK TNT EGDN Protocol : - lab condition Training set : only h100 acquisitions - vapour generation cell Test set : h100, h50 and h10 acquisitions - different concentrations ICASSP 2013 | MAYOUE Aurélien | 12

  13. EXPERIMENTS: RESULTS 1) Identification rate : Quantification : Explosives TNT EGDN Theoritical values Concentration h100 h50 h10 h100 h50 h10 Estimated values Identification rate 3/3 3/3 3/3 3/3 3/3 3/3 Identification time (s) 47 43 47 31 32 32 Interferents EtOH DCM MEK Concentration h100 h50 h10 h100 h100 Identification rate 3/3 3/3 2/3 2/3 3/3 - identification rate: 94% Identification time (s) 35 32 31 31 34 - identification time < 60s 2) Identification rate : - robustness to variations of Explosives TNT EGDN concentration Concentration h100 h50 h10 h100 h50 h10 QCM+SAW 0/3 0/3 0/3 3/3 3/3 3/3 - performances are deteriorated OPTO+SAW 3/3 3/3 3/3 3/3 2/3 0/3 when a technology is missing OPTO+QCM 3/3 3/3 3/3 3/3 2/3 1/3 Interferents EtOH DCM MEK Concentration h100 h50 h10 h100 h100 QCM+SAW 3/3 3/3 2/3 2/3 3/3 OPTO+SAW 2/3 3/3 2/3 2/3 1/3 OPTO+QCM 1/3 1/3 2/3 2/3 0/3 ICASSP 2013 | MAYOUE Aurélien | 13

  14. Video TNT, EGDN vs. EtOH, DCM, MEK Aurelien.mayoue @cea.fr Recursive Least Squares Algorithm CEA LIST Dedicated to Early Recognition of 91191 Gif-sur-Yvette Cedex, France Explosive Compounds thanks to Multi-technology Sensors ICASSP 2013

Recommend


More recommend