Scholarlink

In the context of Industry 4.0, machine maintenance and product quality control are crucial for manufacturing efficiency and reliability. This paper introduces a novel approach based on heatmap transformation and deep neural networks for product defect classification and machine failure prediction from tabular data, including static numerical and time series data. Unlike existing approaches that analyze numerical values using either a single record or a sequence of records as input, our method converts these inputs into heatmaps. This allows for visualizing multivariate process parameters and detecting signs of defects or failures through color variations using image-based classification models. The method also incorporates an explainability approach that leverages existing image-based explainability techniques to identify specific parameters and values associated with defects or failures. This provides operators with valuable insights to help identify the root causes of problems. The approach has shown promising results when applied to two public datasets from real industrial use cases.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Toward an Explainable Heatmap-Based Deep Neural Network for Product Defect Classification and Machine Failure Prediction in Industry 4.0

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tojo Valisoa Andrianandrianina Johanesa,
  • Lucas Equeter,
  • Sidi Ahmed Mahmoudi,
  • Pierre Dehombreux

摘要

In the context of Industry 4.0, machine maintenance and product quality control are crucial for manufacturing efficiency and reliability. This paper introduces a novel approach based on heatmap transformation and deep neural networks for product defect classification and machine failure prediction from tabular data, including static numerical and time series data. Unlike existing approaches that analyze numerical values using either a single record or a sequence of records as input, our method converts these inputs into heatmaps. This allows for visualizing multivariate process parameters and detecting signs of defects or failures through color variations using image-based classification models. The method also incorporates an explainability approach that leverages existing image-based explainability techniques to identify specific parameters and values associated with defects or failures. This provides operators with valuable insights to help identify the root causes of problems. The approach has shown promising results when applied to two public datasets from real industrial use cases.