Di-CNN: Domain-Knowledge-Informed Convolutional Neural Network for Manufacturing Quality Prediction

Shenghan Guo, Dali Wang, Zhili Feng, Jian Chen, Weihong Guo2023 · DOI

TL;DR

This paper introduces Di-CNN, a novel convolutional neural network that integrates domain knowledge with real-time sensor data for improved manufacturing quality prediction. By adaptively weighting different information sources, Di-CNN achieves higher accuracy and provides interpretable insights into which factors are most critical for quality.

What Problem Does This Paper Address?

In manufacturing, Convolutional Neural Networks (CNNs) are powerful tools for analyzing sensor data (especially images) to predict product quality. However, traditional CNNs are purely data-driven. This means they often require vast amounts of training data, can struggle with noisy or limited data, and their predictions can be difficult to interpret. Crucially, they don't inherently incorporate valuable "domain knowledge" – information about the manufacturing process from design stages, physical constraints, or expert understanding. This lack of domain knowledge integration can limit prediction accuracy and make it hard to understand why a certain prediction was made, hindering practical application and process improvement.

Methods

The core innovation is the Di-CNN (Domain-Knowledge-Informed CNN). It builds upon a standard CNN architecture but enhances it by:

  • Multi-Source Input Fusion: Di-CNN accepts not only real-time image data (processed by a base CNN to extract features) but also "domain knowledge" inputs. These can include design features, build parameters, or operational modes.
  • Adaptive Weighting Scheme: This is the key differentiator. Instead of treating all input sources equally, Di-CNN learns to assign weights to each input source during training. This adaptive weighting scheme optimizes these weights to minimize the prediction error (Mean Squared Error - MSE) on the training data. Essentially, it learns which information sources are most relevant to predicting quality.
  • Fusion Layer: A dedicated fusion layer combines the weighted features from all input sources before the final prediction layers. This allows the model to learn complex interactions between different types of data.

The adaptive weighting is formulated as an optimization problem solved iteratively during training. This allows the model to dynamically adjust the importance of each input source as it learns.

Datasets

The study uses a real-world dataset from resistance spot welding (RSW), a common automotive manufacturing process. The data includes:

  • Real-time thermal videos: Captured during the welding process, serving as the primary image input.
  • Domain knowledge inputs:
    • Current intensity (categorical, converted to dummy code).
  • Coating condition (categorical, converted to dummy code).
  • Number of sheets being welded (numerical).

The quality metrics predicted were weld thickness and diameter, measured through destructive testing. The dataset was preprocessed to create spatial-temporal instances from the thermal videos. The study employed sixfold cross-validation for robust evaluation.

Key Insights and Ideas

  • Domain Knowledge is Not Just "Extra Data": The paper argues that domain knowledge isn't just another input; it's a crucial component that can fundamentally improve model performance, especially when real-time data is limited or noisy.
  • Adaptive Weighting Unlocks Interpretability: The learned weights provide direct insights into the relative importance of different data sources for quality prediction. This moves beyond a "black box" model to one that can inform decision-making.
  • Not All Data Sources Are Equal: The adaptive weighting scheme effectively identifies and prioritizes the most informative inputs, allowing the model to focus its learning capacity. This is particularly useful when some data sources are less reliable or relevant than others.
  • Bridging the Physics-vs-Data Gap: While not a full physics-based model, Di-CNN offers a practical way to integrate "physics-like" domain knowledge into data-driven models, making them more robust and interpretable for complex manufacturing processes where full physical modeling is challenging.

Key Results

The Di-CNN with adaptive weights significantly outperformed both a conventional CNN and a Di-CNN with fixed, equal weights:

  • Mean Squared Error (MSE):
    • Di-CNN (adaptive weights): 6.8866 (mean), 6.1916 (median)
    • Di-CNN (fixed weights): 13.6171 (mean), 13.1343 (median)
    • Base CNN: 27.2935 (mean), 25.6117 (median)
  • Interpretability: The learned weights in the case study indicated that the image data (Input 1) and the number of sheets (Input 4) were the most relevant factors for predicting weld quality. Coating condition and current intensity received lower weights, suggesting they were less influential in this specific scenario.

Limitations and Future Work

  • Limitations:
    • Numerical Metrics Only: The current Di-CNN is designed for predicting numerical quality metrics. Extending it to handle categorical quality predictions would broaden its applicability.
  • Computational Burden: Like most deep learning models, Di-CNN can be computationally intensive to train.
  • Unregulated Weight Updating: The weight updating is solely based on training loss, which could be made more sophisticated.
  • Future Work:
    • Categorical Quality Prediction: Adapt the model for non-numerical outputs.
  • Sparsity Integration: Reduce computational burden by incorporating sparsity into the model structure.
  • Regulated Weight Updating: Develop more advanced weighting schemes that consider computational efficiency and regulated objectives.
  • Guidance for Data Acquisition: Use the learned weights to inform future sensor placement and data collection strategies.
  • Broader Domain Knowledge Integration: Explore integrating more diverse forms of external information.

Summary updated

BibTeX
@misc{guo2023di,
  author = {Shenghan Guo and Dali Wang and Zhili Feng and Jian Chen and Weihong Guo},
  title = {{Di-CNN: Domain-Knowledge-Informed Convolutional Neural Network for Manufacturing Quality Prediction}},
  year = {2023},
  doi = {10.3390/s23115313},
}

Loading related content...

Figure