What question did this study set out to answer?

The aim is to develop a machine learning framework to detect layer shifting in FFF 3D printing using sensor data.

April 22, 2026

Machine learning-based detection of layer-shifting in FFF 3D printing using MEMS vibration and acoustic sensors

Key Points

The aim is to develop a machine learning framework to detect layer shifting in FFF 3D printing using sensor data.
Implemented a Cartesian FFF 3D printer with MEMS sensors to capture vibration and acoustic signals.
Printed 54 specimens with variations using a Taguchi L9 design and extracted five time-domain features.
Trained and evaluated eight ML classifiers on sensor data using various performance metrics.
Random Forest classifier achieved 97.43% accuracy, indicating excellent performance on accelerometer data.
Precision and F1 score for Random Forest were 97.71% and 97.31%, respectively.
ROC-AUC score of 0.9951 reflects high reliability in fault detection.

Abstract

Purpose Layer shifting is one of the most critical machine-induced defects in fused filament fabrication (FFF) three-dimensional (3D) printing, primarily caused by belt drive slippage along the x-axis. This defect leads to dimensional inaccuracies and structural misalignments in printed parts. The present study aims to develop a machine learning (ML)-based framework supported by sensor data to detect layer shifting at an early stage of the printing process. Design/methodology/approach A Cartesian FFF 3D printer was equipped with two low-cost MEMS sensors, an ADXL335 accelerometer and a MAX4466 microphone, mounted near the x-axis stepper motor to capture vibration and acoustic signals under normal and faulty conditions. Using a Taguchi L9 design, 54 specimens were printed with controlled variations. The acquired signals were sampled at 1,600 Hz, and five time-domain features (Root Mean Square, Mean, Standard Deviation, Peak-to-Peak and Crest Factor) were extracted. Eight ML classifiers (SVM, RF, GB, LR, DT, KNN, NB and XGBoost) were trained and evaluated for both sensor data sets using accuracy, precision, recall, F1 score and ROC-AUC. Findings Among the models tested, Random Forest demonstrated superior performance on accelerometer data with 97.43% accuracy, 97.71% precision, 97.31% F1 score and a ROC-AUC of 0.9951, indicating high fault detection reliability. Originality/value This study presents a novel approach that combines vibration and acoustic sensing with ML-based classification to identify machine-related defects in FFF printing. It further proposes a conceptual real-time feedback framework for future integration into intelligent, fault-tolerant 3D printing systems.

Demander à l'IA

Bookmark