Additive Manufacturing Fault Detection Using Multiple Process Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional direct metal laser melting (DMLM) process quality control is inadequate for real-time monitoring, as it struggles to detect process deviations and faults due to measurement variability from part geometry, scan path history, sensor viewing angle, and optic-induced distortions, leading to potential defects and increased costs.
Innovation Solution
The implementation of time-dependent and spatially-dependent models for in-process monitoring, using sensors like photodiodes and pyrometers, along with a multiple model hypothesis testing algorithm to differentiate between normal and faulty conditions, and generate probabilistic fault indications and quality metrics by comparing sensor data to trained models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional between-build diagnostics are used for quality control, then device complexity is reduced, but measurement precision and reliability deteriorate due to inability to detect process deviations in real-time
Solution Approach 1:
The patent segments the monitoring task by dividing the build process into discrete layers and further into laser-on transients within each layer. This segmentation allows the system to focus computational resources on specific, critical moments in the process where deviations are most detectable, thereby improving measurement precision without proportionally increasing overall system complexity.
Solution Approach 2:
The system performs preliminary action by pre-training multiple process models (nominal and fault conditions) before actual monitoring begins. These models are prepared in advance with expected sensor responses for various conditions, enabling real-time comparison and detection without requiring complex real-time computation during the actual build process.
2Measurement precision
If steady state averaged measurement is used for monitoring, then device complexity is reduced, but measurement precision deteriorates due to inability to resolve individual parameter changes
Solution Approach 1:
The patent employs periodic action by analyzing laser-on transients that occur periodically during the layer-by-layer build process. Each laser-on transient represents a periodic event where the laser is activated to create a melt pool, and these periodic events are captured and analyzed to detect parameter shifts. This approach provides more granular measurement precision compared to steady-state averaging while maintaining manageable system complexity through focused analysis of these periodic events.
3Reliability
If multiple diagnostic methods are added to discriminate parameter shifts, then measurement precision improves, but device complexity and computation time increase
Solution Approach 1:
The system performs preliminary action by pre-training multiple process models (nominal and fault conditions) before actual monitoring begins. These models are prepared in advance with expected sensor responses for various conditions, enabling real-time comparison and detection without requiring complex real-time computation during the actual build process.
Solution Approach 2:
The patent creates copies of the process under different conditions through multiple process models. Each model represents a hypothetical process state (nominal or various fault conditions), and these model copies are compared against actual sensor data to identify deviations. This copying approach allows comprehensive fault detection without requiring complex real-time analysis of every possible condition.
4Productivity
If in-situ process monitoring is implemented, then productivity is improved through real-time feedback, but measurement precision deteriorates due to variability from part geometry, scan path history, sensor viewing angle, and optic-induced distortions
Solution Approach 1:
The patent applies local quality by training separate process models for different locations on the build plate. Each model is trained on data specific to its local region, accounting for variations in part geometry, scan path history, and sensor viewing angle that are characteristic of that particular location. This localized modeling approach maintains measurement precision while enabling comprehensive real-time monitoring across the entire build plate.
Solution Approach 2:
The system addresses measurement variability by incorporating it as explicit parameters in the model training process. Rather than treating geometry, scan path, viewing angle, and optical distortions as sources of error to be eliminated, the patent changes the approach to include these parameters in the model development, allowing the models to learn and compensate for these systematic variations.
Data Source
Figure 1~2
Figure 3
Figure 4A
AI summary
Generating fault indications for an additive manufacturing machine (110) based on a comparison of the outputs of multiple process models (150) to measured sensor data.