A Wide-Deep-Sequence Model-Based Quality Prediction Method in Industrial Process Analysis.

Area of Science:

• Industrial Intelligence
• Manufacturing Process Analysis
• Data-Driven Quality Prediction

Background:

• Accurate product quality prediction is crucial for industrial process adjustment and optimization.
• Data-driven predictive models are gaining traction for analyzing complex industrial data.
• Extracting relevant quality features from diverse industrial data (supply chain, machining) is challenging.

Purpose of the Study:

• To propose a novel data-driven method for reliable industrial product quality prediction.
• To effectively handle high-redundancy industrial data and extract quality-relevant features.
• To improve the reduction of defective products through enhanced prediction accuracy.

Main Methods:

• A Wide-Deep-Sequence (WDS) model integrating Wide-Deep (WD) and Long Short-Term Memory (LSTM) components.
• Data reduction techniques applied to high-redundancy industrial variables.
• WD model for time-invariant features; LSTM for time-domain features.
• Joint training strategy with a penalty mechanism for unreliable predictions.

Main Results:

• Demonstrated effectiveness of the WDS model on a real-world manufacturing dataset.
• Successful extraction of quality features from both time-invariant and time-domain industrial data.
• Improved prediction accuracy, particularly in identifying and reducing defective products.

Conclusions:

• The proposed WDS model offers a reliable approach for industrial product quality prediction.
• The method effectively integrates diverse data types and addresses data redundancy.
• This approach contributes to enhanced industrial intelligence and manufacturing efficiency.