Page 37 - IJAMD-1-2

International Journal of AI for
Materials and Design
AI-driven quality assurance in AM

4.2. Methodology 4.3. Results

4.2.1. Data collection 4.3.1. Accuracy
In-situ monitoring systems equipped with high- The trained CNN model achieved an accuracy of 95% on
resolution cameras and infrared sensors were installed the test dataset, effectively identifying common defects
on the SLM machine. These systems captured layer-by- such as porosity, surface roughness, and incomplete
layer images and thermal data during the manufacturing fusion. The confusion matrix showed high precision
process. A comprehensive dataset comprising 10,000 (average 94%) and recall rates (average 96%) for all defect
images, annotated with defect types and locations, was classes.
compiled for training and testing the CNN model.
4.3.2. Real-time implementation
• Image data: Captured using high-resolution cameras
to provide detailed visual information of each layer. The model was integrated into the SLM machine’s
• Thermal data: Collected using infrared sensors to control system for real-time defect detection. During
monitor temperature variations, which are indicative production, the in-situ monitoring system fed images into
of potential defects. the CNN, which processed and identified defects within
milliseconds. When defects were detected, the system
4.2.2. CNN architecture alerted the operator and suggested corrective actions, such
A deep CNN architecture tailored for image recognition as adjusting laser power or scan speed.
tasks was employed. The architecture included. • Defect alert system: Implemented to notify operators
• Input layer: Handling images of size 256×256 pixels. immediately upon defect detection, allowing for
• Convolutional layers: Four convolutional layers on-the-fly adjustments.
with rectified linear unit activation functions, each • Corrective measures: Suggested adjustments were
followed by max-pooling layers to extract hierarchical based on defect type, e.g., reducing laser power
features. for overheating issues or adjusting scan speed for
• First layer: 32 filters, kernel size 3×3, stride 1, padding incomplete fusion.
“same.”
• Second layer: 64 filters, kernel size 3×3, stride 1, 4.3.3. Improvements in quality assurance
padding “same.” • Defect reduction: The real-time detection and
• Third layer: 128 filters, kernel size 3×3, stride 1, corrective measures reduced the incidence of porosity
padding “same.” and surface roughness by 40%.
• Fourth layer: 256 filters, kernel size 3×3, stride 1, • Production efficiency: Early detection of defects
padding “same.” minimized the need for post-processing and rework,
• Fully connected layers: Two fully connected layers improving overall production efficiency by 30%.
with 512 and 256 neurons, respectively, to interpret • Operator assistance: The system provided valuable
the features and classify defects. feedback to operators, enhancing their ability to
• Dropout layers: Dropout layers with a rate of 0.5 after maintain consistent quality across batches.
each fully connected layer to prevent overfitting.
• Output layer: A softmax layer to predict the probability 4.4. Challenges and future work
of different defect types.
4.4.1. Challenges
4.2.3. Training • Data quality: The quality of the training data
The CNN model was trained using the annotated dataset. significantly impacted the model’s performance.
Data augmentation techniques, such as rotation, scaling, Ensuring high-quality, accurately annotated data were
and flipping, were applied to increase the diversity of the crucial.
training data. The model was trained for 50 epochs with a • Model interpretability: While CNNs are powerful,
batch size of 32, using the Adam optimizer and categorical their black-box nature makes it challenging to
cross-entropy loss function. interpret the decision-making process.
• Learning rate: Set at 0.001, with learning rate decay to • Computational resources: Real-time processing
fine-tune the training process. requires significant computational power,
• Validation split: 20% of the dataset was used for necessitating the use of specialized hardware like
validation during training. graphics processing units (GPUs).

Volume 1 Issue 2 (2024) 31 doi: 10.36922/ijamd.3455