Virtual Sensors for Diesel Engine and After-treatment Management
Federico Covassin Magneti Marelli S.p.A. - Powertrain
Reggio Emilia June 27th 2016
Powertrain Index
■ Company Overview ■ Magneti Marelli Powertrain (Diesel) System ■ Standard Emission trend and technologies ■ O2 intake Virtual Sensor for Diesel engine management ■ NOx Virtual Sensor for after-treatment management ■ NOx Virtual Sensor adaptivity concepts ■ Conclusions
Jun 27th, 2016 2 Company Overview
Magneti Marelli is an international company committed to the design and production of hi-tech systems and components for the automotive sector.
AUTOMOTIVE POWERTRAIN ELECTRONIC SUSPENSION LIGHTING (Gasoline and Diesel SYSTEMS SYSTEMS engine control, Electric Motor and Transmission) (Headlamp, Rearlamp and Body (Instrument clusters, Infotainment & (Suspension Systems, Shock Computer) Telematics) Absorbers, Dynamic Systems)
EXHAUST PLASTIC AFTERMARKET MOTORSPORT PARTS & SERVICES (Injectors, ECU, KERS, DST, SYSTEMS COMPONENTS Hydraulic Gearbox Control (Mechanical, Body Work, Electrics and Electronic (Manifolds, Catalytic converter, AND MODULES Unit…) DPF and Mufflers) and consumables) (Bumper, Dashboard, Central Consol, and Fuel system)
Jun 27th, 2016 3 Magneti Marelli Worldwide Presence
Sales 2015 7.3 bn €
R&D Centers 12
R&D (of sales) 4.6%
Production units 89
Application Centers 30
Investments (of sales) 5.7%
Employees 40,418
Jun 27th, 2016 4 Magneti Marelli Powertrain Worldwide Presence
BOLOGNA (I)
CREVALCORE(I) SHANGHAI (CN) VENARIA (I) CHANGCHUN(CN) CORBETTA (I) SMMP (CN) BARI (I) GUANGZHOU(CN) TRAPPES (F) WUHU(CN) ARGENTAN (F) MANESAR (IN) WOLFSBURG (D) PUNE(IN) KECHNEC
AUBURN HILLS (USA)
SANFORD (USA)
SALTILLO (MEX)
HORTOLANDIA (BR)
Production plant R&D Center Application Center
Sales 2015 919 mio € Production units 15 R&D Centers 4 Application Centers 9 R&D* 5.7% Investments* 13.7% Employees worldwide 5.293
* % of “make” sales Jun 27th, 2016 5 Customer Brand Portfolio
Jun 27th, 2016 6 Magneti Marelli Powertrain – Diesel Systems
CONTROL SYSTEMS SENSORS Hybrid Propulsion
ECU Smart Sensors Power Motor-Generator Inverter
Controls Transmission & SW
AMT DCT
Engine Management Calibration
Components for Combustion Models SCR/SCRoF System & Spray targeting ACTUATORS Intake man. w var. Throttle Body VGT Actuator Swirl Actuator Jun 27th, 2016 7 Emission Standard trend
EPA European Other main legislation Union Japan
HC India CO Brazil
NOX China NH3 PM PM Other main requirements Number limit Pollutant SO2 On Board Diagnostic Requirement Pb (OBD II , EOBD)
Fuel consumption, years CO2 Reduction
Euro1 Euro2 Euro3 Euro4 Euro5 Euro6b Euro6d PC Euro I Euro II Euro III/IV Euro V Euro VI LDT LCV Stage I Stage II Stage III Stage IV Stage V
Tier 0 Tier 1 Tier 2 Tier 3 Tier 4 NRMM HD LEV I LEV II LEV III Jun 27th, 2016 8 Technologies for reaching the target
Two main ways for NOx reduction on Diesel Engine ‹ NOx Engine Out reduction ‰ Combustion improvement (e.g. EGR HP, EGR LP) ‹ After-treatment technologies (e.g. LNT, SCR, SCRonF) could be combine for reach the target
NOx VS LNT DPF
Tsensor Tsensor
UEGO UEGO
∆PTrap
NOx VS NOx VS SCR DOC DPF (or LNT) NOx Sensor Tsensor DEF NOx Sensor Tsensor (opt) Injector
Mixer UEGO ∆PTrap
NOx VS DOC SCRoF ufSCR (or LNT) DEF NOx Sensor NOx Sensor NO Sensor Injector T x sensor (opt) O2 Virtual Sensor
Mixer UEGO
∆PTrap NOx Virtual Sensor Jun 27th, 2016 9 Intake O2 and NOx Virtual Sensors
NOx VS DOC SCRoF ufSCR
Adaptive control
Combustion Management – EGR Control Management
Partnership with University of Salerno
Jun 27th, 2016 10 Intake O2 Virtual Sensor
Purpose of intake O2 measure: ‹ Estimation of effective EGR ratio (e.g. EGR Valve Control Management) ‹ Enhancement of conventional and advanced combustion control (PCCI, HCCI,LTC) ‹ Improvement of NOx prediction during engine transients, suitable for both dynamic adjustments of EMS strategies and management of after-treatment devices.
m Pamb Tamb & air
O2exh
n m& f
uEGR
Combustion Mng - EGR Control Mng
Pman Jun 27th, 2016 11 Intake O2 Virtual Sensor
MEAN VALUE MODEL ‹ Filling and Emptying for ‹ Low computational intake/Exhaust manifolds demand and identification issues ‹ Mass and energy conservation in any control volume ‹ Homogeneous pressure, temperature ‹ Suitable for on- and chemical composition in the intake board manifold implementation ‹ Instantaneous and perfect mixing of incoming flows ‹ Model accuracy fully satisfactory ‹ No heat transfer through manifold walls
Jun 27th, 2016 12 Mean Value model Main equation
Oxygen mass fraction in the intake manifold
RairTman O&2man = [()O ,2 exh − O ,2 man m& egr +()O ,2 amb − O ,2 man m& air ] PmanVman Intake manifold temperature
RairTamb T&man = [(kairTegr −Tman )m& egr + ()kairTic,out −Tman m& egr − ()kair −1 Tmanm& cyl,in ] PmanVman
Prediction of Exhaust O2 concentration in place of UEGO measurement
RexhTexh O& ,2 exh = [(O ,2 cyl,out − O ,2 exh )m& cyl,out ] PexhVexh
m& cyl,inO ,2 man −α st m& f O ,2 man O ,2 cyl,out = m& cyl,in + m& f
Jun 27th, 2016 13 O2 Experimental results
Common-Rail Diesel 1.3 – EGR/HP - VGT
60 c Engine Key characteristics
/str] 40
3 60 Rated Power and 95Hp and Max. Torque 210Nm@1500rpm a
[mm 20 Cylinders 4 in line f 3
m /str] 40 Displacement 1248 cm 0 3 Valves per cylinder 4 (DOHC) 600 200 400 600 800 1000 1200 1400 Combustion Diesel Direct Injection System
[m m 20 f Compression Ratio 16.8:1 40 m Synchronisation • Crankshaft position sensor 0 system • Camshaft position sensor 20 60 Fuel Injection Common Rail Solenoid Injectors System (160 MPa) EGR EGR POS [%] 0 Intake Air System • Electrical Throttle Body 40 actuator 0.21 • Intake manifold pressure and temperature sensor b 0.2 Turbo charging VGT turbocharger, vacuum 20 System controlled with vacuum electro- 0.19 modulator with VGT position sensor EGR POS [%] EGR System • DC-Motor EGR valve + 0.18 0 position feedback
man [%] 0.22 • Air Flow Meter, before Turbo- 2
O 0.17 Compressor c 0.2 Exhaust Gas • 1 linear oxygen sensor 0.16 measured System (UHEGO) in the exhaust pipe, predicted downstream of turbine just 0.15 0.18 before catalyst
0 200 400 600 800 1000 1200 1400 m an [% ] After-treatment • Oxidant Catalyst + Diesel 2 measured Time [s] System Particulate Filter (coated soot O 0.16 filter, without additive),predicted close- coupled 560 570 580 • 590DPF Differential600 pressure610 620 Time [s]sensor *O2 intake measured with O2 sensor • DPF inlet temperature sensor
Jun 27th, 2016 14 O2 Experimental results
Common-Rail Diesel 2.3 – EGR/HP - VGT 50 Engine Key characteristics Rated Power and 150Hp and Max. Torque 320Nm 40
Cylinders 4 in line /str] Displacement 2286 cm3 3 Valves per cylinder 4 (DOHC) 30 Combustion Diesel Direct Injection System Qref [mm Qref Compression Ratio 19:1 20 Synchronisation • Crankshaft position sensor system • Camshaft position sensor 10 Fuel Injection Common Rail Solenoid Injectors 0 5 10 15 20 25 30 System (160 MPa) Intake Air System • Electrical Throttle Body actuator • Intake manifold pressure and temperature sensor Turbo charging VGT turbocharger, vacuum System controlled with vacuum electro- modulator with VGT position sensor EGR System • DC-Motor EGR valve + position feedback • Air Flow Meter, before Turbo- Compressor Exhaust Gas • 1 linear oxygen sensor System (UHEGO) in the exhaust pipe, downstream of turbine just 0 before catalyst After-treatment • Oxidant Catalyst + Diesel 0 System Particulate Filter (coated soot man [%] man
filter, without additive), close- 2
coupled O 0 • DPF Differential pressure sensor 0 • DPF inlet temperature sensor 0 200 400 600 800 1000 1200 1400 0 Time [s]
Jun 27th, 2016 15 NOx Virtual Sensor
Purpose: ‹ After-Treatment Device Management (LNT, SCR, SCRoF, etc.) ‹ Diagnosis After-treatment system: ‹ NOx Sensor diagnosis: plausibility check & functional diagnosis
NOx VS LNT DPF Tsensor Tsensor
UEGO UEGO
∆PTrap
NOx VS DOC SCRoF ufSCR DEF NOx Sensor NOx Sensor Injector T NOx Sensor sensor (opt)
Mixer UEGO
∆PTrap
Jun 27th, 2016 16 Recurrent Neural Network in one slide
Neural Network: • Black Box Model • Basic elements (neurons) are combine together with connections and are placed in different layers depending on the architecture • Right inputs are fundamental for good result • The training and validation phases are important to make more or less strong connections X • A lot of NN parameters impact the final result: 4 number of neuros, number of layers, NN architecture, training algorithm, epoch of training, initial conditions, etc.
Recurrent Neural Network: • The current output also depends on the previous outputs y ( t − ,1 θ ),..., y ( t − i , θ ), x ( t ),.., x ( t − j + 1 ), y ( t , θ ) = F 1 1 x ( t ),..., x ()t − j + 1 , 2 2 x 3 ( t ),..., x 3 ( t − j + )1
Jun 27th, 2016 17 Recurrent Neural Network training
Recurrent Neural Network NOx m • Delay compensation from measured data & deb
• RNN identification/ training: mf, rpm, ° Parametric analysis and selection of SOI the most suitable neural network structure; ° Deterministic analysis to set RNN weights initial conditions; O2 intake ° RNN Pruning by means of the Optimal Brain Surgeon (OBS) algorithm.
RNN main parameters: • N. of Layers: 3 • N. of neurons in hidden layer: 18 • N. of epoch: 80 • Lag input space: 2 • Lag output space: 2
Jun 27th, 2016 18 NOx Experimental Result
Common-Rail Diesel 1.3 – EGR/HP - VGT
300 40 measured TEST NEDC predicted 35 250 30 200 25
[g/h] 150 20 sp NOx
mf [mm3/str] 15 100 10 50 5
0 0 500 600 700 800 900 1000 1100 1200 0 20 40 60 80 100 120 140 Time [s] 350 measured 300 predicted
250
200
MSE R2 ERR.INT [g/h] sp 150
NEDC 16,14 0,9746 0,0909 NOx 100 TEST 90,21 0,9756 0,1707 50
0 0 20 40 60 80 100 120 140 Time [s] *NOx Engine Out measured with NOx production sensor
Jun 27th, 2016 19 Recursive Least Square based Adaptation
Off/On-line parameters update (adaptivity) • Implementation of RLS methods. • Development of methods for neural network.
• Adaptivity based on gain/offset parameter already implemented. ‰ Low CPU load
Adaptive Control • RNN re-training procedure ‰ more powerful but high CPU load
Jun 27th, 2016 20 Recursive Least Square based Adaptation
NOx measured, plus imposed drift of -30%.
20 k Update Validation 0 k 1 15
10 500 500 Neural network k0=0.701 Neural network k =1 measured measured 400 parameters 0 Least Square 400 k =3.282 5 1 Least Square k1=0 300 300 0 200 200 NOx [ppm] NOx [ppm] 1500 0 NEDC200 profile400 600 Train800 profile1000 1200 1400 Neural1600 network 100 time [s] 100 measured Least Square 0 0 400 450 500 550 600 1550 NEDC1600 profile 1650 1700 1000 time [s] time [s] NEDC RMSE [ppm] IE [%]
Least Square 26, 06 12,7958 NOx [ppm] 500 Neural Network 66,007 55,2771
0 0 200 400 600 800 1000 1200 1400 1600 time [s] Jun 27th, 2016 21 Conclusion
• Emissions standard trend requests strict pollutant limit • Engine Management System and after-treatment will be more complex and more expensive and Virtual Sensors represent a good opportunity • O2 intake estimation with model based approach gives good result and could be implemented on ECU for real-time estimation • NOx engine out estimation with Recurrent Neural Network also present good results with benefits on calibration effort • Parameters update (adaptivity) is implemented with good result recovering the system dispersion
Jun 27th, 2016 22 Thank You Backup