In a paper published in the journal Manufacturing Letters, researchers addressed limitations in machine learning (ML) for manufacturing by introducing a quantum-based strategy. They applied quantum support vector machines to identify machine states in fused filament fabrication using acoustic data and used quantum convolutional neural networks (CNN) to detect spatters in laser powder bed fusion through optical images. Their findings showed that quantum methods matched classical accuracy but required exponentially fewer parameters.
Study: Quantum machine learning for additive manufacturing process monitoring. Image Credit: Pixel B/Shutterstock.com
Related Work
Past work in data-driven manufacturing has shown that while ML effectively monitors complex processes, it faces limitations such as high data requirements, imbalanced datasets, and the curse of dimensionality. ML in manufacturing is limited by the need for large training datasets, imbalanced data due to rare anomalies, and the curse of dimensionality, which increases data demands as model complexity grows. Additionally, models struggle with real-time adaptability to new machine states.
Quantum ML for Advanced Manufacturing
In this study, a quantum support vector machine (QSVM) was applied for anomaly detection in the fused filament fabrication (FFF) process. Using acoustic emission (AE) sensor data, QSVM identified various machine states, such as 'normal,' 'run-out-of-material,' 'semi-blocked,' and 'blocked.' These states were determined by analyzing AE data based on three main attributes: acoustic burst signal (ABS)-energy, root mean square (RMS), and AE counts.
A time window of 500 ms was used to compute the mean and standard deviation of these attributes, resulting in six features. The QSVM was trained on 75% of the data, and the remaining 25% served as test data. Results showed distinct patterns, particularly in the 'blocked' and 'run-out-of-material' states.
The QSVM implementation utilized a quantum feature map based on the Havlíček et al. approach, encoded with the ZZ feature map using Qiskit. For encoding, four qubits were used for the mean and standard deviations of RMS and ABS-Energy in the first case, and six qubits were employed in the second case to include AE counts.
QSVM was trained with a quantum kernel and compared to two classical SVMs with linear and radial basis function (RBF) kernels. This comparison revealed that QSVM maintained competitive classification accuracy while leveraging quantum-specific feature mapping.
The study also introduced a QCNN for spatter detection in the laser powder bed fusion (LPBF). Images of the LPBF melt pool were captured at 1,000 frames per second, with each frame manually labeled as 'Spatter' or 'No Spatter.' Images were cropped to focus on the melt pool, resized, and encoded using 8 qubits. QCNN was trained on a balanced dataset, ensuring robust classification of spatter events.
QCNN’s architecture included convolution and pooling layers that optimized information retention while reducing the qubits used. Convolution layers with rotation and controlled-NOT (CNOT) gates enhanced feature extraction, and a quantum pooling circuit reduced computational cost by encoding two qubits into one.
The QCNN was trained over 100 epochs using the constrained optimization by linear approximations (COBYLA) optimizer and was compared to a classical neural network with a similar structure. Both networks performed well, but QCNN demonstrated efficient performance with reduced computational parameters, highlighting the potential for quantum ML in advanced manufacturing applications.
Quantum Anomaly Detection
The study applied quantum ML methods, specifically QSVM and QCNN, to detect anomalies in the FFF and LPBF processes. The QSVM demonstrated competitive performance in anomaly detection, achieving 85.7% accuracy with four features and 89.5% accuracy with six features, outperforming classical SVMs with linear and RBF kernels.
The results indicated that increasing the number of features improved classification accuracy, highlighting the advantage of mapping data into high-dimensional Hilbert space for more accurate classification. Additionally, QCNN exhibited effective performance in spatter detection, achieving a training accuracy of 75.0% and a test accuracy of 64.6% with 63 parameters, showing that it could match the accuracy of classical neural networks while requiring exponentially fewer parameters.
The architecture of the QCNN was structured as an n-level binary tree, significantly enhancing its efficiency compared to traditional neural networks, which typically exhibit higher space complexity. The findings suggested that both QSVM and QCNN could mitigate the overfitting challenges commonly encountered in ML, especially in manufacturing applications where training data may be limited.
The study emphasized the unique sampling complexity of quantum algorithms, which must be considered alongside time and space complexities typical in classical approaches. By leveraging the advantages of quantum-specific feature mapping and fewer parameters, these quantum ML methods showed promise for advanced manufacturing applications, paving the way for more robust anomaly detection and classification.
Conclusion
To sum up, this study demonstrated the potential of quantum ML in manufacturing applications, highlighting the effectiveness of QSVM for identifying FFF machine states and QCNN for spatter detection in LPBF processes. The results showed that QSVM's quantum kernel offered superior mapping to high-dimensional feature spaces, while QCNN achieved comparable accuracy to classical neural networks with exponentially fewer parameters, mitigating overfitting issues. Despite current limitations in quantum computing, the findings suggested that tailored quantum ML models could address scalability challenges and enhance predictions in dynamic manufacturing systems.
Journal Reference
Choi, E., et al. (2024). Quantum machine learning for additive manufacturing process monitoring. Manufacturing Letters, 41, 1415-1422. DOI: 10.1016/j.mfglet.2024.09.168, https://www.sciencedirect.com/science/article/pii/S221384632400258X
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.