Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Boosting Defect Detection in Manufacturing using Tensor Convolutional Neural Networks

Pablo Martin-Ramiro, Unai Sainz de la Maza, Sukhbinder Singh, Roman Orus, Samuel Mugel

TL;DR

The paper tackles automated defect detection in mass-manufacturing where CNNs are often too parameter-heavy for edge deployment. It introduces Tensor Convolutional Neural Networks (T-CNNs) that replace standard convolution weight tensors with Tucker-decomposed cores and factor matrices, enabling training in a compressed parameter space. On a real Bosch ultrasonic sensor component dataset, T-CNNs achieve CNN-like defect-detection performance while reducing parameters by up to about fivefold and speeding up training, with slip-through rates dropping from 10% toward 4.6%. The work demonstrates practical advantages for real-time quality control, including deployment on edge devices and faster inference, while outlining future directions for hyperparameter optimization and broader applicability.

Abstract

Defect detection is one of the most important yet challenging tasks in the quality control stage in the manufacturing sector. In this work, we introduce a Tensor Convolutional Neural Network (T-CNN) and examine its performance on a real defect detection application in one of the components of the ultrasonic sensors produced at Robert Bosch's manufacturing plants. Our quantum-inspired T-CNN operates on a reduced model parameter space to substantially improve the training speed and performance of an equivalent CNN model without sacrificing accuracy. More specifically, we demonstrate how T-CNNs are able to reach the same performance as classical CNNs as measured by quality metrics, with up to fifteen times fewer parameters and 4% to 19% faster training times. Our results demonstrate that the T-CNN greatly outperforms the results of traditional human visual inspection, providing value in a current real application in manufacturing.

Boosting Defect Detection in Manufacturing using Tensor Convolutional Neural Networks

TL;DR

The paper tackles automated defect detection in mass-manufacturing where CNNs are often too parameter-heavy for edge deployment. It introduces Tensor Convolutional Neural Networks (T-CNNs) that replace standard convolution weight tensors with Tucker-decomposed cores and factor matrices, enabling training in a compressed parameter space. On a real Bosch ultrasonic sensor component dataset, T-CNNs achieve CNN-like defect-detection performance while reducing parameters by up to about fivefold and speeding up training, with slip-through rates dropping from 10% toward 4.6%. The work demonstrates practical advantages for real-time quality control, including deployment on edge devices and faster inference, while outlining future directions for hyperparameter optimization and broader applicability.

Abstract

Defect detection is one of the most important yet challenging tasks in the quality control stage in the manufacturing sector. In this work, we introduce a Tensor Convolutional Neural Network (T-CNN) and examine its performance on a real defect detection application in one of the components of the ultrasonic sensors produced at Robert Bosch's manufacturing plants. Our quantum-inspired T-CNN operates on a reduced model parameter space to substantially improve the training speed and performance of an equivalent CNN model without sacrificing accuracy. More specifically, we demonstrate how T-CNNs are able to reach the same performance as classical CNNs as measured by quality metrics, with up to fifteen times fewer parameters and 4% to 19% faster training times. Our results demonstrate that the T-CNN greatly outperforms the results of traditional human visual inspection, providing value in a current real application in manufacturing.
Paper Structure (17 sections, 10 equations, 4 figures, 2 tables)

This paper contains 17 sections, 10 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Problem overview
  3. Dataset description
  4. Problem Formulation
  5. Tensor convolutional neural networks
  6. Convolutional Neural Networks
  7. Tensor network basics
  8. Tensorizing a CNN
  9. Parameter Counting
  10. Experimental setup
  11. Model architecture
  12. Training setup
  13. Performance metrics
  14. Results
  15. Performance of the T-CNN
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Tensor network diagrams. Each rank-$n$ tensor is denoted by an object with $n$ connected links each of which represents an individual dimension of the tensor. A scalar, a vector, a matrix and a rank-$n$ tensor have respectively, $0$, $1$, $2$, $n$ connected legs.
  • Figure 2: Contraction of tensors which is equivalent to tensor trace over their shared indices is represented graphically by connecting the shared links. Here, $R$ and $S$ tensors are connected along the shared leg $\beta$. This contraction operation is equivalent to matrix multiplication.
  • Figure 3: The SVD decomposition of a matrix to two rank-$3$ tensors which are called matrix product operator (MPO).
  • Figure 4: The HOSVD (Tucker) factorization of the convolution weight tensor to a core tensor and four factor matrices. The $\chi_i$s are the factorization (truncation) ranks of the factor matrices.