Machine learning applications in aerospace domain

Machine learning applications in aerospace domain

• 1.
  Machine Learning Applications inAerospace Domain 2018. 6. 김 홍 배 Artificial Intelligence Laboratory
• 2.
  Expert Knowledge inData + Labels Model (mostly) determined by D + L Machine Learning Technology Spectrum Expert Knowledge in Model Details Data refines model parameters Real-world systems often combine several techniques Machine Learning Model-Based Understanding Model-DrivenData-Driven Open Exploration Unsupervised Exploratory Parameterization Naïve Stats Supervised ML Physics-Based Modeling Reinforcement Learning UQ Assimilative Models
• 3.
  Object detection, SemanticSegmentation, Change Detection, Super-resolution, etc Applications DL Based Satellite Image Analysis
• 4.
  • Detect faultsand failures in complex aerospace systems • Key challenges  Data is extremely large, noisy, and unlabelled  Most of applications exhibit temporal behaviour  Detected anomalous events typically require immediate intervention Applications Anomaly Detection of Aircraft & Spacecraft
• 5.
  Applications Properties of TelemetryData  Multimodality : A satellite system (or each of its subsystems) has a number of different operational modes and changes from one mode to another over time.  Heterogenuity : variables in a satellite’s housekeeping data are divided into two types: Continuous variables that take real values and discrete variables that take categorical values.  High Dimensionality : Numbers of Continuous and Status Telemetry Variables usually over 1,000 !!  Temporal Dependence  Trivial Outliers : data occasionally contain exceptionally large abnormal values caused by errors in data conversion or transmission.
• 6.
  Applications Machine Learning inPlanetary Exploration Rovers • Make intelligent decisions about what data to gather and transmit. • Develop onboard rover traverse science data analysis system for data prioritization and opportunistic science Constraint on Rovers : limited downlink bandwidth and communication time delay between Earth and the rovers.(The average distance between the two planets is 225 million km ~ 750 sec. time delay) Autonomous Exploration for Gathering Increased Science(AEGIS)
• 7.
  1. Researchers areinterested in identifying the existence of certain pre- specified signals of scientific interest. 2. The second criterion is the identification of unexpected, or anomalous, features, as these can lead to new scientific discoveries. 3. Finally researchers want to capture a description of the typical characteristics of a region. Three classes of data evaluation criteria Applications Machine Learning in Planetary Exploration Rovers
• 8.
  Architecture • Mapping ft’n,which maps physical space to latent space • Clustering with Gaussian mixture density X Z 𝜇 𝑘, 𝐶 𝑘 𝜇1, 𝐶1 𝜇 𝐾, 𝐶 𝐾 Parametric modeling of Clusters,M(θ)Mapping ft’n + 𝑤 𝐾 𝑤 𝑘 𝑤1 𝑓(𝑥) 𝜇 𝑘, 𝐶 𝑘, 𝑤 𝑘 Query(x) 𝑧1 𝑧2 Physical space Latent space Architecture for Anomaly detection & Auto Exploration
• 9.
  Σ Input layer Hidden layer (RBFs) Outputlayer W1 W2 WM x1 x2 xn No weight f(x) Each of n components of the input vector x feeds forward to m basis functions whose outputs are linearly combined with weights w (i.e. dot product x∙w) into the network output f(x). The output layer performs a simple weighted sum (i.e. w ∙x). If the RBFN is used for regression then this output is fine. However, if pattern classification is required, then a hard- limiter or sigmoid function could be placed on the output neurons to give 0/1 output values Input data set ∶ 𝑋 = { 𝑥1 𝑥2 … 𝑥 𝑁} Architecture Architecture for Anomaly detection & Auto Exploration
• 10.
  Σ Architecture Architecture for Anomalydetection & Auto Exploration query 0.2 query 0.9
• 11.
  Σ Σ Category 1Category 2 Category 1 Category 2 Architecture Architecture for Anomaly detection & Auto Exploration
• 12.
   predict thepower or fuel consumption of the spacecraft  Three years of spacecraft telemetry are released … can you predict the fourth year ?  The ultimate goal is to automate operations and extend satellite life time, which in turn increases the scientific return. Applications Machine Learning in Spacecraft Engineering “Get the data, make a model and predict the budgets of Subsystem”
• 13.
  Mars Express Challenge Themost promising approaches are ensemble selections where different models are merged to produce the prediction. The ensemble of random forest, LSTM and another deep neural network model provided a better result than each one of them separately. Comparison of the two best models Applications
• 14.
  Exploring Generative 3DShapes Using Auto-encoder Networks • Purpose : find modes of 3D objects • Key idea : Parametric modeling of 3D objects with a fixed dimension regardless of shape Applications
• 15.
  Machine learning frameworkwhich predicts aerodynamic forces and velocity and pressure fields given a three dimensional object shape and Reynolds number input. Applications Learning Three-Dimensional Flow for Interactive Aerodynamic Design
• 16.
  Input layer Gaussian Processing Output layer: Y x1 x2 xn d(x) Input data set ∶ 𝑋 = { 𝑥1 𝑥2 … 𝑥 𝑁} N v(x) N p(x) N . . . . . . Gaussian Process (GP) regression for inferring the CFD simulation data Applications Learning Three-Dimensional Flow for Interactive Aerodynamic Design Three regressors : for drag coefficient, non-dimensionalized velocity, and pressure. Input : Parametric modeling vector of car + Reynolds No. y Output data set ∶ 𝑋 = {𝑦1 𝑦2…𝑦 𝑁}
• 17.
  • Assist butrespect models : Machine learning should be used to correct/improve existing models, not to replace them. • Cost effective & exact solution : Turbulent flow & Solid Mechanics modeling  Optimal design of Aircraft & Rocket engine Applications Physics- Informed Machine Learning
• 18.
  Phase I : Trainingwith Machine learning Phase II : Prediction with ML assisted RANS Simulation Data : features q responses δ𝑅(ϒ,ΔΛ,Q) q δ𝑅 Neural Nets Query q’ Corrected Reynold stress δ𝑅’