Suitability of Artificial Neural Network for MEMS
Transcription
Suitability of Artificial Neural Network for MEMS
sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Suitability of Artificial Neural Network for MEMS-based Flow Control J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON FEMTO-ST Institute University of Franche-Comté, France dMEMS 2012. April 2nd-3rd, Besançon, France J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 1 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Outline 1 Introduction 2 Problem Specification and CFD Model 3 Predicting the Flow using an Artificial Neural Network 4 Optimizing the Flow 5 Conclusion J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 2 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Active control of turbulent flows Objective Improve the aerodynamic performance of moving objects Various potential benefits Better maneuverability Increasing the range or payload capability Environmental compliance A solution for performing active fluid control: sets of distributed MEMS MEMS actuators are adjusted to control the flow MEMS sensors may provide information on the real evolution of the flow J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 3 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Application of MEMS for active flow control - Vehicle 2005 Citroen C-SportLounge Concept Car Virtual airfoil on the tail using 40 MEMS actuators producing pulsed microjets Improve the vehicle’s drag and stability at high speeds e free g ima N on fre N ei on m ag fre e ei m ag e Non J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 4 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Application of MEMS for active flow control - Aircraft J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 5 / 27 Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion sciences & technologies MEMS-based flow control is a scientific challenge A closed-loop flow control system requires A model that captures the actuation effects A controller Fluid models are described by Navier-Stokes equations Highly nonlinear CFD tools cannot solve them in real-time Prohibitive computation time Changing inflow velocity A nonlinear model that can predict the forced flow in real-time is needed Our objectives 1 Study the suitability of a neural network to model the process 2 Provide a relevant strategy to control the MEMS J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 6 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Outline 1 Introduction 2 Problem Specification and CFD Model 3 Predicting the Flow using an Artificial Neural Network 4 Optimizing the Flow 5 Conclusion J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 7 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Simplified vehicle model: a backward facing step Simplified ground vehicle geometry Backward facing step (FEMTO-ST simulation) Ahmed body k − ω model Fluent v = 40 m.s−1 v = 40 m.s−1 height h = 288 mm Re = 747000 J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON k − RNG model height h = 288 mm Neural Network for MEMS-based Flow Control 8 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Computational fluid dynamics model - Geometry H h y 0 x l Backward facing step characteristics t L Flow characteristics height h = 0.288 m upstream section H = 0.712 m length l = 0.5 m downstream section L = 2.5 m J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 9 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Computational fluid dynamics model - Setup 5 microjet actuators located on the step near the corner, 10, 30, 50, and 70 mm away Microjet velocity discrete value in {0, 33, 66, 99} m.s−1 Uniform velocity profile at inlet inflow velocity in {30, 35, 40} m.s−1 Outlet outflow boundary conditions 19.5 mm 9.5 mm Microjet actuators Wall functions at the bottom wall Computation with FLUENT CFD code 150,000 cells 5 minutes on a regular Linux Workstation J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 10 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Objectives Focus on pressure data on the back face 244 measure points are aggregated into a singular force f Objective 1 - Model Predict the force f (C, v ) for any inflow velocity v in [30, 40] any configuration C of actuators in {0, 33, 66, 99}5 Find a nonlinear mapping between inputs defining the current inflow and an actuation configuration an output describing the flow evolution Objective 2 - Control Control the actuators to maximize the force f (C, v ) when the inflow velocity varies J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 11 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Outline 1 Introduction 2 Problem Specification and CFD Model 3 Predicting the Flow using an Artificial Neural Network 4 Optimizing the Flow 5 Conclusion J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 12 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Overview Why choosing a neural network? Successful modeling of complex relationships between data inputs and outputs Can model linear and nonlinear systems, static and dynamic ones Universal approximator for nonlinear systems Analogy with biological neural networks A set of interconnected neurons processes the inputs A training process iteratively adapts the neurons parameters to the data Multilayer feedforward networks Can approximate any nonlinear function to any arbitrary precision MultiLayer Perceptron network J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 13 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Multilayer perceptron Hidden layer x1 Input vector x1 Output layer Output vector −1 x2 w 1j w 2j bias n ϕj ... bias x 0 −1 w 0j i=0 w ij x i y j −1 xn x2 w nj y1 −1 Supervised training using FLUENT results −1 x3 −1 y2 Mean square criterion ⇒ reflects the prediction mismatch Training algorithm ⇒ minimizes the criterion x4 −1 How many hidden neurons and which training algorithm ? J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 14 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Network topology and implementation Network features 3rd order HPU network ⇒ better approximation of functions with sharp variations Single hidden layer Number of neurons found by trial and error process Sigmoidal activation Output layer Number of neurons specified by the problem Linear activation Training process L-BFGS quasi-Newton algorithm with Wolfe linear search Holdout validation to avoid overfitting ⇒ dataset split in training, validation and test subsets J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 15 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Training setup Global dataset of 3072 samples obtained with FLUENT simulations 3 inflow and 4 microjet velocities → 3 × 45 = 3072 samples Dataset randomly partitionned in training subset ≈ 65% of the samples validation subset ≈ 10% of the samples test subset ≈ 25% of the samples Network topology 6 initial inputs = (C, v ) → 83 ones with HPU 1 output = f (C, v ) Hidden layer of 15, 20, 25, and 30 neurons Control of the training process 500, 750, and 1000 training epochs J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 16 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Prediction results Performances of the different networks - Test subset of 769 data samples Average on 20 trainings Random subsets and initial network parameters Topology Hidden neurons 15 20 25 30 6 initial/83 HPU inputs and 1 output Epochs % CVRMSE R2 500 10.26 0.9475 750 10.21 0.9480 1000 10.23 0.9477 500 10.20 0.9481 750 10.14 0.9487 1000 10.15 0.9486 500 10.31 0.9469 750 10.28 0.9473 1000 10.26 0.9475 500 10.25 0.9476 750 10.20 0.9480 1000 10.20 0.9481 J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 17 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Prediction results Observation Prediction 20 Resulting force f 0 -20 -40 -60 -80 0 100 200 300 400 500 Test data samples indexes J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON 600 700 800 Neural Network for MEMS-based Flow Control 18 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Prediction results Observation Prediction 20 Resulting force f 0 -20 -40 -60 -80 500 550 600 Test data samples indexes J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON 650 700 Neural Network for MEMS-based Flow Control 19 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Outline 1 Introduction 2 Problem Specification and CFD Model 3 Predicting the Flow using an Artificial Neural Network 4 Optimizing the Flow 5 Conclusion J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 20 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Controller - Objective Hypotheses Continuously varying inflow velocity v Changing configuration C is time consuming Objective Real-time maximization of the back face step force f (C, v ) by modification of the actuators Definition (Optimal configuration) The optimal configuration C ∗ (v ) maximizes the force f (_, v ): C ∗ (v ) = f ((c0 , . . . , c4 ), v ) argmax (c0 ,...,c4 )∈{0,33,66,99}5 J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 21 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Controller - Local optimal configuration Global optimization is not suited Ideally, instantaneously change the current configuration to C ∗ (v ). Takes time δt practically Inflow velocity at time t + δt: v 0 6= v → C ∗ (v 0 ) 6= C ∗ (v ) Optimization using a neighborhood structure Search the “best” configuration in the neighborhood of C Neighborhood set N(C) = {C 0 = (c00 , . . . , c40 ) | |cj0 − cj | ≤ 33 and 0 ≤ j ≤ 4} For each pair (C, v ) the local optimal configuration is C̃(v ) = argmax f (C 0 , v ) C 0 ∈N(C) Greedy algorithm to find C̃(v ) J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 22 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Validating the approach - Protocol Scenario Random initial configuration C0 Random sequence of inflow velocities vi [30, 40] s.t. −kd < vi+1 − vi < ka Comparison cumulative optimized force σ= i=n−1 X f (C̃(vi−1 ), vi ) i=1 cumulative force obtained with the static configuration C0 σ̄ = i=n−1 X f (C0 , vi ) i=1 J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 23 / 27 Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion sciences & technologies Validating the approach - Experiments Setup ka = 2.5m.s−1 and kd = 5m.s−1 20 random scenarios with each time a sequence of 10 velocities Experimental observations Optimized forces ⇒ always positive σ = 118 N in average Without optimization ⇒ always negative σ̄ = −475 N in average Average improvement of almost 125% J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 24 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Outline 1 Introduction 2 Problem Specification and CFD Model 3 Predicting the Flow using an Artificial Neural Network 4 Optimizing the Flow 5 Conclusion J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 25 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Conclusion Summary A neural network can provide suitable predictions in a simplified case Only 5 actuators Relevance of active flow control The proposed control strategy improves the aerodynamic performances Future works Neural network design Incremental training Online training using information from sensors Control strategy Scalability of the method Distributed approach J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 26 / 27 sciences & technologies Introduction Problem Specification and CFD Model Predicting the Flow using an Artificial Neural Network Optimizing the Flow Conclusion Thank you for your attention Questions? J.-F. C OUCHOT, K. D ESCHINKEL, and M. S ALOMON Neural Network for MEMS-based Flow Control 27 / 27