“Quantum machine learning for recognition of defects in ultrasonic imaging” by Prof. Prabhu Rajgopal

About the Author

Trained at IIT Madras (IITM – BTech & DD MTech) and Imperial College London (PhD, Postdoc), Dr. Prabhu Rajagopal has expertise in non-invasive diagnostics. Full Professor since 2019, currently he is the Advisor for Innovation & Entrepreneurship, heading Automation & Data Engineering at the Center for NDE​ at IITM and the spin-out company Plenome. With 30+ funded projects, 225+ technical articles and 33 granted IPs, he is widely recognised for his work on digital transformation in the industrial (energy and mobility) and social (water, sanitation and health) contexts. He is recipient of India’s most prestigious award for mid-career scientists, the Shanti Swarup Bhatnagar prize for Technology and Innovation (2024). 

This paper- ‘Quantum machine learning for recognition of defects on ultrasonic imagin’, written by Prof. Prabhu Rajgopal, is the first Quantum Machine Learning Paper from CNDE Lab, IIT Madras. 

Summary of the paper

The paper presents the potential use of a quantum machine learning (QML) algorithm, in particular a Variational Quantum Classifier (VQC), for automated weld defect recognition using ultrasonic phased arrays. The application performs classification with high accuracy and fast calculation times by harnessing the inherent parallelism of qubits capable of processing multiple states simultaneously.

Weld quality is important across many industries and its evaluation is often performed by human inspectors which can be time consuming and error prone. Automated defect recognition (ADR) based on artificial intelligence (AI) has recently been proposed as an alternative to improve decision making speed and reduce human error rates. Convolutional neural networks (CNNs) have commonly been used for the analysis of weld images obtained from ultrasonic phased array inspections. However, CNNs are classical algorithms that struggle with increasing data set sizes and may not consistently achieve accurate solutions required for safety critical applications.

The VQC is demonstrated as a potent substitute for traditional classifiers such as MLPs, autoencoders, and RCNNs. Because the training of MLP models, autoencoders or RCNN are inherently based on information propagation through feedforward layers (for MLP), learning compressed representations of inputs (for autoencoders), and locating defect regions in images (for RCNN), these models become increasingly time-consuming to train with the increasing size of datasets.

In this study, a simulation-assisted dataset using FE modeling and deep learning via DCGAN was generated.

Two data augmentation methodologies for TFM imaging of weld porosity and slag

are applied:

• Full Matrix Capture (FMC): A shear wave wedge is used to launch shear waves in multiple azimuthal directions. The received A-scan signals are used as input data for imaging and sizing the defects using TFM.
• Synthetic Image Generation: The FE-generated TFM images act as ground truth images for training AI algorithms to generate synthetic images.

A feature selection technique is applied to extract the most discriminative features from the generated TFM images, which are then provided as an input to an encoder circuit and processing in quantum space using a Quantum Convolution Neural Network. The VQC has been trained by varying the sizes of dataset in order to enhance the

classification performance.

The results show that VQC performs significantly better compared to classical machine learning algorithms. It recorded an overall accuracy of 97% on non-noisy weld defect images and 85% on noisy datasets, indicating its strong performance for the given task even with noise in the dataset. The influence of noise caused a drop in classification accuracy for all algorithms2 , hence calling for further investigation especially in a noisy industrial environment.

VQC was also benchmarked against MLPs11 , autoencoders12 and RCNNs13 . Classical algorithms have been reported to give good results in various applications but their biggest drawback is the tremendous increase of training time as data increases along with scalability. On the other hand, VQC possesses inherent efficiency characteristics due to its quantum nature. 

The results demonstrate the prospects of QML in non-destructive testing and indicate that due to specific advantages of quantum computing like superposition and entanglement, VQC can perform more effective analysis with complex datasets when compared to classical approaches. This paper also presents insights enabling deployment of VQC in industry along with concerns associated with feasibility. 

Three research directions are discussed:

1. Noise Analysis: Study on the effect of different types of noise on the classification accuracy.
2. Scalability: Theoretically it is still not clear how scalable VQC is for larger datasets and more complex defect types.
3. Integration with Existing Systems: Develops approaches to ameably integrate developed QML techniques within the current inspection technologies can be a likely next task.

This study presents the potential use of quantum machine learning (QML) for automated weld defect recognition as initial research. The results show that QML can not only achieve similar performances compared with classical machine learning methods but also could increase the accuracy when applied to small-sample-size-based data. As quantum computer’s software technology and hardware technology penetrating into applications further, the high-performance NDT driven by QML is becoming a part of reality.

In conclusion, this paper does not fully elaborate on how QML works. Nonetheless, we strategy highlight its potential performance for one industrial issue providing evidence worth studying deeper. It is demonstrated that performing classification via a quantum approach would provide superior performance based on precision and recall rates than via a classical approach at least under 750 training samples per class.