TLDR: A new study demonstrates how Explainable AI (XAI) can significantly improve industrial injection molding processes. By applying XAI techniques like SHAP, Grad-CAM, and LIME to an LSTM model, researchers successfully reduced the number of input features from 19 to 9 and 6. This reduction not only maintained high quality classification performance, with the 9-feature model even outperforming the full model, but also led to faster prediction times. The findings suggest that XAI can make AI-driven quality control more interpretable, efficient, and feasible for industrial settings, especially those with limited sensor data, and simplify the creation of synthetic training data.
Artificial intelligence (AI) is rapidly transforming industrial processes, especially in quality control. However, the sophisticated nature of many machine learning models often makes them difficult to understand, limiting their practical use in real-world industrial settings. This lack of transparency, coupled with incomplete data from older industrial machines, presents a significant hurdle for widespread AI adoption. A recent study tackles this challenge by using Explainable Artificial Intelligence (XAI) to make AI-driven quality control more transparent and efficient in industrial injection molding processes.
Unlocking AI’s Potential with Explainable AI
The research, titled “Improving Industrial Injection Molding Processes with Explainable AI for Quality Classification,” explores how XAI techniques can provide clear insights into how AI models make decisions. This is crucial for building trust in AI systems and for identifying the most important factors influencing product quality. The authors, Georg Rottenwalter, Marcel Tilly, and Victor Owolabi from Rosenheim Technical University of Applied Sciences, focused on the complex process of injection molding, where precise quality classification is vital.
The Problem: Black-Box Models and Limited Data
Traditional machine learning models, while powerful, often act as ‘black boxes,’ meaning it’s hard to understand why they arrive at a particular conclusion. In an industrial context, this makes it difficult for human operators to trust predictions, troubleshoot errors, or improve processes. Furthermore, many older industrial machines lack the extensive sensor technology needed for comprehensive data collection, making it challenging to train robust AI models. The goal of this study was to overcome these limitations by identifying the most critical data points for quality prediction.
XAI in Action: SHAP, Grad-CAM, and LIME
The researchers applied three popular XAI methods—SHAP (SHapley Additive exPlanations), Grad-CAM, and LIME (Local Interpretable Model-agnostic Explanations)—to an Long Short-Term Memory (LSTM) model. This model was trained on real production data from injection molding cycles, initially using 19 different input features like temperature, pressure, and injection speed. The XAI techniques helped to analyze which of these features were most influential in the model’s quality classification decisions.
Feature Reduction: A Path to Efficiency
A key aspect of the study involved reducing the number of input features. By analyzing the importance scores from SHAP, Grad-CAM, and LIME, the researchers created two smaller datasets: one with the 9 most important features and another with the 6 most important features. The aim was to see if a simpler model, using fewer data points, could maintain or even improve performance while also speeding up the prediction process.
Surprising Results: Better Performance with Less Data
The findings were significant. While the original model with 19 features achieved an average validation accuracy of 86% and an F1 score of 89%, the model trained with only 9 features actually performed better. It achieved an average validation accuracy of 91% and an F1 score of 92%, indicating improved generalization. The model with 6 features, though slightly less stable, still maintained strong performance with an 84% validation accuracy and 86% F1 score.
Beyond accuracy, feature reduction also led to efficiency gains. The 9-feature model showed a 7% improvement in inference time (how quickly it makes predictions), and the 6-feature model was 13% faster compared to the full 19-feature model. Memory consumption, however, saw only minimal reductions, as it’s primarily determined by the model’s architecture rather than the number of input features.
Also Read:
- Enhancing Industrial Injection Molding with Synthetic Data for Smarter Production
- Unlocking AI’s Decisions: A Transparent Approach to Early Sepsis Detection
Implications for Industry
This research demonstrates that XAI can successfully identify the most critical parameters in complex industrial processes like injection molding. By focusing on these essential features, manufacturers can simplify data acquisition, making AI-based quality control more accessible for existing industrial systems, including those with limited sensor capabilities. Furthermore, reducing the number of input parameters simplifies the generation of realistic synthetic data, which is crucial for training robust AI models when real-world data is scarce or difficult to obtain. This approach enhances the feasibility and quality of AI-driven quality prediction in manufacturing.
For more details, you can read the full research paper here.
