Computational Intelligence Approaches to Defect Detection in 3D Printing

• Published in: 2025 IEEE Symposium on Computational Intelligence on Engineering/Cyber Physical Systems (CIES) pp. 1 - 7
• Main Authors: Awonuga, Oluwaseun, Madden, Kyle, Coleman, Sonya, Kerr, Dermot, Quinn, Justin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.03.2025
• Subjects: 3D printing; Computational intelligence; Computational modeling; correlation; defect; Defect detection; Intelligent sensors; sensor; Solid modeling; Subtraction techniques; Temperature distribution; Temperature sensors; Three-dimensional printing; Time series analysis; unsupervised methods
• DOI: 10.1109/CIES64955.2025.11007634

Digitalization in smart manufacturing is driving the use of Internet of Things (IoT) in many 3D printing environments. These sensors facilitate collection of data in the form of time series that can reflect a normal condition or faulty state. The ability to identify the normal conditions or faulty states by analysing sensor data is vital to minimise defects in additive manufacturing processes. However, detecting a defect based on correlated behaviour of multiple sensors is complex and an active area of ongoing research utilising multivariate time series. Currently, no comparative studies exist between machine learning and deep learning approaches that consider the potential correlation between multiple sensor data while constructing a fault detection model. In this work, we propose a unique computational intelligence approach to defect detection in a multi-sensor fused deposition modelling 3D printer. We decompose temperature and humidity sensor data into residual components using a seasonal-trend procedure with locally estimated scatterplot smoothing. A subtraction technique is then utilised to reduce two time series into one, by focusing directly on a "deviation from correlated behavior" of both sensor data. Five unsupervised models were used to detect defective state using the joint feature of temperature and humidity as training data. The test results demonstrated that Long Short-Term Memory-AutoEncoder outperformed other models with a recall rate of 94% in identifying all possible defects from the correlated behaviour of the sensors during print activity.