Quantum AI Boosts Weld Inspection, Spotting Defects with Greater Accuracy

Akshaya Srinivasan and colleagues at the RPTU Kaiserslautern-Landauhave combined quantum and classical computing to improve industrial quality control. They demonstrate two hybrid quantum-classical machine learning approaches for identifying defects in aluminium TIG welding images, comparing their effectiveness to a conventional deep learning model. The research uses convolutional neural networks to simplify complex image data before processing it with quantum algorithms, specifically a Variational Quantum Linear Solver and a variational quantum circuit with angle encoding. Although a classical convolutional neural network performed strongly, the results indicate hybrid models can achieve competitive performance, suggesting a viable path towards near-term quantum solutions for real-world industrial defect detection and quality assurance.

Quantum machine learning matches classical performance identifying welding defects

A competitive performance between hybrid quantum-classical models and a conventional convolutional neural network has been achieved, matching the accuracy of Model-1, a classical system which previously outperformed 13 other architectures. Achieving comparable results opens avenues for utilising quantum computing in practical applications, as previous quantum approaches struggled to surpass established classical techniques in complex image classification tasks. Both quantum models, employing a Variational Quantum Linear Solver and angle encoding respectively, were benchmarked on classifying defects in aluminium TIG welding images, a key process in industries demanding high structural integrity. The significance of this work lies in demonstrating that quantum computation can, at the very least, equal the performance of state-of-the-art classical machine learning in a specific, industrially relevant application, a crucial step towards realising the potential of quantum technologies.

Model-1, a convolutional neural network specifically designed for weld image classification, validated these competitive results. The CNN acted as a feature extractor, reducing the dimensionality of weld images before quantum processing and generating feature vectors enabling comparison with the quantum models. This dimensionality reduction is critical, as it addresses the challenge of feeding high-resolution image data directly into quantum algorithms, which are often limited by qubit availability and coherence times. The Variational Quantum Linear Solver (VQLS) approach incorporated a quantum kernel method, mapping data into a higher-dimensional ‘Hilbert space’ to enhance the support vector machine (SVM) optimisation process; analysis of the quantum kernel condition number revealed its impact on classification stability. A well-conditioned kernel is essential for stable and accurate classification, preventing numerical issues during the SVM training process. A variational quantum circuit (VQC) directly encoded classical features as angles within quantum gates, trained with a classical optimiser for model refinement, and tested across both binary and multiclass defect identification scenarios. The choice of angle encoding allows for a natural mapping of classical data onto quantum states, but requires careful calibration to ensure optimal performance.

Quantum kernel performance mirrors optimised classical networks for weld defect detection

Automated visual inspection is vital for maintaining quality in manufacturing, particularly for critical welds where hidden flaws can have catastrophic consequences. In industries such as aerospace, automotive, and construction, weld integrity is paramount, and even minor defects can compromise structural safety. Investigators are now investigating whether quantum computing can offer an edge in detecting these defects, moving beyond matching the performance of existing systems. Like many in the field, this work demonstrates parity rather than superiority to a carefully optimised classical convolutional neural network, a benchmark system that already outperformed thirteen other image analysis architectures. The challenge isn’t necessarily to immediately replace classical methods, but to identify scenarios where quantum algorithms can provide a demonstrable advantage, either in terms of accuracy, speed, or resource efficiency.

Detailed examination of these approaches, specifically exploring how quantum kernels impact classification accuracy and refining angle encoding techniques, are valuable steps towards using quantum power for complex image analysis tasks. The quantum kernel method, by mapping data into a higher-dimensional Hilbert space, potentially allows the model to capture non-linear relationships that might be missed by classical SVMs. However, the effectiveness of this approach depends heavily on the choice of kernel and the characteristics of the data. Optimising the angle encoding scheme within the VQC is also crucial, as it directly affects the expressibility and trainability of the quantum circuit. The ability to effectively encode classical information into quantum states is a fundamental requirement for hybrid quantum-classical machine learning.

Hybrid quantum-classical approaches can achieve performance comparable to conventional deep learning when classifying defects in aluminium TIG welding images, and further development will begin to unlock the full potential of quantum computing in industrial applications. Investigators explored two distinct quantum models by initially simplifying image data with a convolutional neural network; one utilised a Variational Quantum Linear Solver and the other employed angle encoding within a quantum circuit. Competitive results with established techniques signifies a shift towards viable near-term quantum solutions for industrial quality control, paving the way for further investigation into quantum algorithms for image analysis. The use of aluminium TIG welding images provides a realistic and challenging test case, as weld defects can be subtle and varied, requiring sophisticated image analysis techniques for accurate detection. Future research could focus on exploring different quantum algorithms, optimising the hybrid architecture, and scaling the approach to larger and more complex datasets. The ultimate goal is to develop quantum-enhanced quality control systems that can improve manufacturing efficiency, reduce costs, and enhance product safety.

The research highlights the potential for quantum machine learning to contribute to the ongoing advancements in automated industrial inspection. While current quantum hardware limitations prevent a clear demonstration of quantum supremacy in this task, the achieved parity with a strong classical baseline is a significant milestone. The exploration of both VQLS and VQC approaches provides valuable insights into the strengths and weaknesses of different hybrid quantum-classical strategies. Continued investigation into these techniques, alongside advancements in quantum hardware, could ultimately lead to the development of robust and scalable quantum solutions for industrial quality control, offering benefits beyond matching existing classical performance.

The study demonstrated that hybrid quantum-classical models can perform competitively with conventional convolutional neural networks when classifying defects in aluminium TIG welding images. This is important because it suggests quantum computing may offer a viable route towards automated quality control in industrial settings. Researchers achieved this by combining classical image processing with two distinct quantum approaches, including one utilising a Variational Quantum Linear Solver. Future work will focus on exploring different quantum algorithms and optimising the hybrid architecture to further refine these techniques.