Neural Networks for EMC Modeling of Airplanes Vlastimil Koudelka Department of Radio Electronics FEKT BUT Metz,
2 Outline Introduction to artificial neural networks (ANNs) Neural networks abilities Evaluation of EM immunity of layered str. Application of neural classifier in EMC Regression neural network based optimization Related problems Conclusion
3 Introduction to ANNs (1) Highly parallel structures Basic element: Neuron Organized to the layers Adaptive nonlinear mapping (learning) Nonlinear separable classification Optimization features W 1,2 (2) W 2,2 (2) W 3,2 (2) b 2 (2) f(n) n
4 Introduction to ANNs (2) RegressionClassification Multi layered perceptron (MLP) Radial basis network (RBF) General regression network (GRN) MLP Probabilistic NN (PNN) Self organizing map (SOM)
5 Introduction to ANNs (3) Optimization Self organizing map Self adopted GRNN Hopfield neural network
6 NN applications: objectives and motivations Behavioral modeling: continuous models, computational efficiency Neural models: composite materials, equipment input impedances, aircraft fuel gauge, field levels Pre-processing, post-processing and optimization tools Offline training / Online responses Suitable for direct parallel implementation Noise suppression (Bayesian regularization) Multidimensionality, adaptability, generalization, robustness
7 Evaluation of EM immunity of layered str. (1) At three virtual probes the electromagnetic fields values are estimated by ANN Harmonic and pulsed wave excitations (Gaussian pulse) MLP, RBF, PNN
8 Evaluation of EM immunity of layered str. (2) Harmonic wavesPulsed waves Dependencies of EM field values on the electrical parameters Pulsed waves: the electric field intensity is expressed in its effective values to respect the mean stress of a virtual device. MLP, RBF, PNN
9 Application of neural classifier in EMC Probability density function of exceeding prescribed limit Classification of structures with various electrical parameters Probability of EM structure successfulness is estimated by ANN Probabilistic neural network
10 Optimization example 1) Initial set of trial solutions consisting of n+1 samples 2) Founded minima of the regression surface is taken as a new training pattern 3) After several iterations the GRNN is well adapted to unknown function f (x)
11 Related problems Black box modeling (interpretation of NN weights and biases) Validation techniques: cross validation NN error estimation (perturbation analyses) Training set compilation: preprocessing and initialization A number of neurons: clustering problem, regularization Efficiency and performance of training algorithms: Benchmarks Stability and robustness of dynamical NNs
12 Conclusion Regression, classification, optimization Computational efficiency, good generalization abilities Pre-processing, post-processing and optimization tools Wide area of applications (neural network adaptation) Shielding efficiency, EM structure classification, material modeling Black box modeling, validation, benchmarking
13 Contact Department of Radio Electronics, Brno University of Technology Purkynova 118, Brno, Czech Republic Tel: Fax: Vlastimil Koudelka
page 14 This work was supported by the project CZ.1.07/2.3.00/ Communication Systems for Emerging Frequency Bands