Session: 03-11-02: Future of Smart Manufacturing

Paper Number: 147406

147406 - Privacy-Preserving In-Situ Monitoring in Additive Manufacturing Through Hyperdimensional Computing 

Additive Manufacturing (AM) enables the creation of complex parts unattainable through traditional methods but struggles with issues like slower production, higher costs, and quality control, leading to a need for in-situ sensing to detect and correct defects in real-time, thereby enhancing efficiency and reducing costs. However, while in-situ sensing presents a promising solution for real-time monitoring and quality assurance in AM, it raises significant privacy concerns due to the potential for sensitive data exposure and the vulnerability of sensor data to unauthorized access, necessitating robust privacy mechanisms to protect proprietary information and ensure security against industrial espionage. To combat privacy issues in additive manufacturing's in-situ sensing, various mechanisms like cryptographic techniques, data anonymization, and differential privacy (DP) have been explored. While cryptographic methods ensure data confidentiality, they can burden system performance with heavy computational loads. Anonymization techniques, on the other hand, reduce the risk of data breaches by stripping sensitive information from datasets but may lead to irreversible data loss and incomplete privacy guarantees. DP emerges as a superior alternative, introducing noise to data processing to make individual records indistinguishable, enhancing privacy without significantly compromising data utility. While differential privacy offers enhanced security by embedding noise into data, thereby ensuring user privacy, its application in AM can inadvertently affect the precision of detecting defects due to the introduced noise. This paper proposes a groundbreaking method that amalgamates differential privacy with Hyperdimensional Computing (HDC), a model inspired by the human brain's functionality, to optimize real-time monitoring within AM in-situ sensing while safeguarding privacy. Our studies demonstrate this combination's effectiveness in monitoring specific anomalies like overhangs in the AM process leveraging rapid melt pool data analysis. The results reveal a remarkable balance achieved between maintaining operational efficiency, ensuring high prediction accuracy, and upholding stringent data privacy standards. For example, when applying the same level of privacy protection (with a privacy budget set at 1), our HDC approach achieved an F-score of 94.30%, markedly surpassing the performance of traditional models such as ResNet50 (42.59%), DenseNet201 (65.41%), EfficientNet-b2 (28.11%), and AlexNet (48.09%). This underscores the unparalleled efficiency and privacy preservation capabilities of the HDC model in the realm of AM in-situ sensing. Interestingly, the initial F-scores without the application of differential privacy for HDC, ResNet50, DenseNet201, EfficientNet-b2, and AlexNet were 97.25%, 97.70%, 98.88%, 98.07%, and 96.45%, respectively, showing comparable performance levels. This indicates that while the initial performance metrics of these models were similar, the integration of HDC with differential privacy uniquely maintains high performance, demonstrating its exceptional ability to balance effective in-situ sensing with stringent privacy requirements in additive manufacturing processes.

Presenting Author: Fardin Jalil Piran University of Connecticut

Presenting Author Biography: Fardin Jalil Piran is a Ph.D. student in the School of Mechanical, Aerospace, and Manufacturing Engineering at the University of Connecticut.

Authors:

Fardin Jalil Piran University of Connecticut
Prathyush Poduval University of California Irvine
Hamza Errahmouni Barkam University of California Irvine
Mohsen Imani University of California Irvine
Farhad Imani University of Connecticut