Application of Deep Belief Networks for Precision Mechanism Quality Inspection

Jianwen Sun, A. Steinecker, Philipp GlockerLecture Notes in Computer Science 2014 · DOI · OpenAlex

Abstract

This paper presents Tilear, an automated fault detection system for precision mechanisms based on Deep Belief Network (DBN) auto-encoders. Traditional inspection methods are often subjective and require large datasets of defective samples. Tilear addresses these challenges by employing an anomaly detection approach, training exclusively on "good" samples to identify defects through reconstruction error. Experimental results on industrial datasets demonstrate superior performance compared to traditional machine learning models.

Introduction

In industrial manufacturing, quality inspection for precision mechanisms—such as micro-motors or greasing units—is a critical yet challenging task. Traditionally, this process relies on human experts who assess product quality based on acoustic or vibration signals. This manual approach is subjective, inconsistent, and prone to human error due to fatigue.

While machine learning offers a potential solution, standard discriminative models like Support Vector Machines (SVMs) typically require balanced datasets containing both healthy and defective parts. In high-quality manufacturing environments, defective parts are rare, making it difficult to collect sufficient data for training. Furthermore, the variety of possible mechanical defects makes it nearly impossible to capture every failure mode. Tilear reframes quality inspection as an anomaly detection problem, learning only from healthy samples.

Methodology

The core of the Tilear system is a Deep Belief Network (DBN) auto-encoder. The architecture consists of stacked Restricted Boltzmann Machines (RBMs) that are unfolded symmetrically to form an auto-encoder structure.

The architecture of the Tilear DBN auto-encoder, showing the encoder, bottleneck layer, and decoder stages.

The process involves two primary phases:

  1. The Teacher Phase (Training): The model is trained exclusively on signals from "good" samples. It learns to compress input data into a low-dimensional "bottleneck" representation and then reconstruct it.
  2. The Tester Phase (Decision-Making): During testing, the system attempts to reconstruct the signal of a new part. Healthy parts yield a low Root Mean Squared Error (RMSE), while defective parts—which the model cannot accurately represent—produce a high reconstruction error, triggering a fault alarm.

Experimental Setup

Tilear was evaluated on two distinct industrial datasets:

  • Precision Electromotors: 36 samples (21 healthy, 15 defective). Input data were cepstrograms derived from vibration signals.
  • Greasing Control Units: 47 samples (24 healthy, 23 defective). Input data were spectrograms derived from acoustic signals.

Results and Discussion

Tilear demonstrated significant improvements in both accuracy and reliability. By learning from "success" rather than "failure," the model establishes a robust mathematical definition of a working machine without needing prior knowledge of specific fault types.

Comparison of Area Under the Curve (AUC) performance between Tilear and a standard SVM on the electromotor dataset.

On the electromotor dataset, Tilear achieved an AUC of 0.960, outperforming a state-of-the-art SVM (0.941). For greasing control units, the system reached an AUC of 0.995, misclassifying only one defective sample. Additionally, Tilear benefited from GPU acceleration, reducing training time to approximately 40 minutes compared to several hours for the CPU-based LIBSVM implementation.

Conclusion

Tilear provides a reliable, unsupervised approach to industrial quality inspection by leveraging DBN auto-encoders for anomaly detection. While the current study utilized small datasets, the results are promising for large-scale deployment. Future work will focus on expanding the model to perform fault classification and conducting more rigorous hardware-controlled performance comparisons.

Summary updated

BibTeX
@inproceedings{sun2014application,
  author = {Jianwen Sun and A. Steinecker and Philipp Glocker},
  title = {{Application of Deep Belief Networks for Precision Mechanism Quality Inspection}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2014},
  doi = {10.1007/978-3-662-45586-9\_12},
}
Figure