Quantum neural networks (QNNs) achieved 98% accuracy in detecting air pressure leaks, according to a new collaboration between the Washington Institute for STEM Entrepreneurship Research (WISER) and the Fraunhofer Institute for Industrial Mathematics ITWM. The partnership focused on applying new quantum computing techniques to anomaly detection in manufacturing, aiming to improve quality control and reduce downtime in complex production systems. The researchers evaluated the performance of QNNs using the NASA bearing fault dataset and demonstrated the potential of these models for real-world engineering challenges. “Quantum neural networks have the potential to integrate quantum principles into machine learning,” said WISER’s Vardaan Sahgal. “However, critical gaps exist in understanding the practical limitations of QNNs in terms of trainability and approximation capabilities.” This study identified binary and exponential encoding as effective strategies to balance model performance and feasibility.
Quantum neural networks for industrial anomaly detection
Researchers can analyze sensor data from industrial equipment to identify anomalies early and potentially improve predictive maintenance strategies. The team’s systematic evaluation of QNN demonstrated competitive performance against traditional machine learning approaches, achieving 98% accuracy in detecting air pressure leaks. This success extended beyond a single application. QNN also showed strong ROC-AUC performance when tested against the NASA bearing failure dataset, indicating potential broader applicability in aerospace engineering and beyond.
A key aspect of this work included optimizing the QNN design, where the researchers identified binary versus exponential encoding as an effective tradeoff between the model’s ability to represent complex data and the feasibility of training the model on current quantum hardware. “This research demonstrates how quantum machine learning can be applied to real-world industrial problems today, while also highlighting the potential to improve the quality of decision support in complex production environments as quantum hardware continues to evolve,” explained Dr. Pascal Halffmann from Fraunhofer ITWM. This collaboration leverages a Fisher Technique factory model and intentionally introduces artificial leaks to provide realistic sensor data for evaluating QNNs in a controlled environment, solidifying the practical focus of the research.
Data encoding strategies and trainability of QNNs
The pursuit of practical quantum machine learning algorithms is increasingly focused on optimizing performance within the constraints of short-term quantum hardware. Researchers are concentrating on how to build QNNs that can effectively learn from real industrial data. This is evidenced by the team applying QNN to challenges such as detecting air pressure leaks and identifying faults in rotating machinery, and leveraging sensor data to evaluate performance in real-world settings. In this study, we achieved 98% accuracy in identifying air pressure leaks using QNNs and demonstrated the potential for high-performance anomaly detection in manufacturing processes. This level of accuracy is compelling considering the complexity of industrial systems and the need for reliable fault diagnosis. Dr.
Quantum neural networks (QNNs) hold promise for integrating quantum principles into machine learning. However, significant gaps exist in understanding the practical limitations of QNNs in terms of trainability and approximation capabilities. Our work provides a roadmap for choosing analyzes that balance expressive power on both synthetic and real-world datasets while using a limited number of qubits to address noise issues.
Vardhan Sagar, Sage
