Error Cause Analysis of Laboratory Results with the Help of AI - Master Thesis in Computer Engineering

Aug 2, 2026

Upload by Patrick

Electronics laboratories must test many different devices, generating large characterization datasets for each unit. Traditional root cause analysis relies on expert manual inspection, making it costly and slow. Rule-based automatic evaluations using hard-coded signal-condition logic still require expert effort to tailor formulas per product and become difficult to scale due to data complexity. This thesis explores early steps toward a machine learning root cause analysis workflow, establishing a foundation for transitioning from expert-driven logic to data-driven ML methods.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages86

Reading Time3 h 37 min

AccessFree to download

Introduction

Motivation

Research Questions

Machine Learning Theory

Supervised and Unsupervised Learning

Anomaly Detection and Performance Measures

PCA

DBSCAN

Isolation Forest

K Nearest Neighbors

Artificial Neural Networks

Convolutional Neural Networks

Autoencoders

Model Analysis and Experimental Results

Unsupervised Approach

Semi-Supervised Approach

Conclusions and Future Works

Master Thesis in Computer Engineering Error Cause Analysis of Laboratory Results with the Help of AI February 27, 2023 Master Candidate Supervisor Andrea Matteazzi Prof. Gian Antonio Susto Student ID 2010655 University of Padova External Supervisors Sandra Mack Inﬁneon Dr. Anja Zernig KAI Academic Year To my family and friends Abstract In the electronics laboratory, a large amount of diﬀerent devices need to be tested, and the characterization of each of them generates a large amount of data. Classical root cause analysis is inherently ineﬃcient because it requires manual inspection by experts, turning out to be costly and time consuming. Furthermore, automatic evaluations through sequences of conditions for the signals, ending up in hard-coded logical formulas, are still inappropriate. This is due to the further necessity for experts, in order to design such formulas speciﬁcally for each diﬀerent product, and to the complexity of the data itself, which may lead to the infeasibility of such approach. For these reasons, in this thesis, ﬁrst steps towards a machine learning (ML) approach are investigated, laying the foundation into the transition to a ML root cause analysis approach. Contents List of Figures xi List of Tables xiii List of Acronyms xix 1 Introduction 1 1. 1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1. 1 Industrial Problem . . . . . . . . . . . . . . . . . . . . . . . 1 1.1.2 Understanding the Data .................... 3 1.2 Research Questions ........................... 5 2 Machine Learning Theory 7 2.1 Machine Learning ............................ 7 2.1.1 Supervised and Unsupervised Learning ........... 7 2.1.2 Learning Algorithm and Cost Function ........... 8 2.1.3 Parameters and Hyperparameters .............. 9 2.1.4 Data Split and Generalization Error ............. 10 2.1.5 Overﬁtting and Underﬁtting ................. 11 2.2 Anomaly detection ........................... 13 2.3 Performance Measures . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3. 1 Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3.2 F1 Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.3.3 Matthews Correlation Coeﬃcient . . . . . . . . . . . . . . . 21 2.4 PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 2.5 DBSCAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 2.6 Isolation Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.7 K Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . 30 CONTENTS 2.8 Artiﬁcial Neural Network . . . . . . . . . . . . . . . . . . . . . . . 31 2.8.1 Deep Feed-Forward Neural Network . . . . . . . . . . . . . 32 2.8.2 Activation Function . . . . . . . . . . . . . . . . . . . . . . . 35 2.8.3 Regularization . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.9 Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . 38 2.9. 1 Convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 2.9.2 Convolutional and Locally Connected Layers . . . . . . . . 42 2. 10 Autoencoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.10.1 Autoencoders in Anomaly Detection . . . . . . . . . . . . . 43 2.10.2 Dense Autoencoder . . . . . . . . . . . . . . . . . . . . . . . 46 2.10.3 Convolutional and Locally Connected Autoencoders . . . 46 2.10.4 Variational Autoencoder . . . . . . . . . . . . . . . . . . . . 47 3 Model Analysis and Experimental Results 51 3.1 Unsupervised Approach ........................ 53 3.2 Semi-Supervised Approach ...................... 53 4 Conclusions and Future Works 67 References 69 List of Figures 1.1 Error cause analysis of diﬀerent devices from a speciﬁc product . 1 1.2 Error cause analysis of diﬀerent products .............. 2 1.3 Case study device and testing procedure ....

FAQs

Why are classical root cause analysis methods inefficient in electronics laboratories?

They require manual inspection by experts, which is costly and time consuming. Additionally, rule-based evaluations need experts to design product-specific hard-coded formulas.

What is the main goal of using machine learning in this thesis?

To investigate initial steps for a machine learning-based root cause analysis approach, enabling a transition from expert-defined logic to data-driven methods.

Which machine learning concepts and models are covered in the theory section?

The thesis includes core ML concepts, anomaly detection, performance measures, PCA, DBSCAN, Isolation Forest, k-nearest neighbors, and neural network models such as ANN, CNN, and autoencoders.

Aug 2, 2026

Upload by Patrick