Upload
others
View
10
Download
0
Embed Size (px)
Citation preview
WIR BEGEISTERN UNSERE KUNDEN MIT EMOTIONALEN FAHRERLEBNISSEN.
PRÄZISE - CHARAKTERSTARK - INNOVATIV
MACHINE LEARNING AS A EXAMPLE FOR OVERRIDEDETECTIONSUPERVISED MACHINE LEARNING WITH MATLAB.
Machine Learning – override detection Seite 2
MOTIVATION.
Evaluation of the "Statistics and Machine Learning" Toolbox from MATLAB
Large number of recorded vehicle measurements (unlabeled) available
Machine Learning - override detection Seite 3
MACHINE LEARNING.
Machine Learning
Supervised Learning
Classification
Support Vector Machines
Discriminant Analysis
NeuralNetworks
Nearest Neighbor
Decision Trees
Regression
Unsupervised Learning
Clustering
Algorithms used.
Machine Learning - override detection Seite 4
WORKFLOW.
Record themeasurements
Loading thedata
Feature
extraction
Training themodel
Application of the model to test data
set
Validation ofprediction
Machine Learning - override detection Seite 5
GENERATION OF THE MEASURED DATA. TRACK
Train a model
Training data set:
- Handling course Miramas
- 259.000 data points
≙ 43 minutes
Test the trained model
Test data set:
- Handling course Aschheim
- 150.000 data points
≙ 25 minutes
Machine Learning - override detection Seite 6
GENERATION OF THE MEASURED DATA.
Insert a trigger signal.
0 0 1 1 1 1 1 0 0 0
Signal 1
Signal 2
Signal 3
Signal 4
Override
t
Machine Learning - override detection Seite 7
GENERATION OF THE MEASURED DATA.
Insert a trigger signal.
0 0 1 1 1 1 1 0 0 0
Signal 1
Signal 2
Signal 3
Signal 4
Override
t
Machine Learning - override detection Seite 8
FEATURE EXTRACTION.
Filter.
Suppress signal noise
Training and test data sets have to filterin the same way
Machine Learning - override detection Seite 9
FEATURE EXTRACTION.
Peak Analysis.
- Use the FindPeaks function
- Minimum distance between the peaks
Machine Learning - override detection Seite 10
FEATURE EXTRACTION.
Principal Component Analysis (PCA).
Transformation in the directions of Principal Components
Machine Learning - override detection Seite 11
MODEL SELECTION.
K-Fold Crossvalidation.
Blue= Training data setRed= Test data set
Results: Average error
Source: Machine Learning for Evolution Strategies, Kramer, 2016, S.39
Machine Learning - override detection Seite 12
MODEL SELECTION.
confusion matrix.
Goal:100 % on the green diagonal
Receiver-Operating-Characteristic-Curve.
Goal :AUC = 1
Machine Learning - override detection Seite 13
VIDEO: PROCEDURE.
Machine Learning - override detection Seite 14
RESULTS.confusion matrix: K-Nearest Neighbor & PCA Feature Extraktion
Machine Learning - override detection Seite 15
RESULTS.confusion matrix: Support Vector Machine
Machine Learning - override detection Seite 16
RESULTS.confusion matrix: Quadratic Discriminant analysis model
Machine Learning - override detection Seite 17
RESULTS.confusion matrix: Complex Decision Trees
Machine Learning - override detection Seite 18
SUMMARY.
Learning must always be carried out from the beginning of the measurements, no adaptive learning
Generate C code from the learned algorithm possible
Fast results with little previous knowledge