Quantum Kernel Methods for Industrial Anomaly Detection

A. Senese, Elena Esposito, Saverio De Vito +2 more

Kuan-Cheng Chen, Samuel Yen-Chi Chen, Chen-Yu Liu +1 more

Takao Tomono, Kazuya Tsujimura

Machine learning is widely used for industrial anomaly detection; however, when anomaly distributions overlap, high false positive rates (FPR) often occur, hindering practical implementation. While quantum kernel methods have been proposed as a solution, the specific conditions under which they outperform conventional methods remain unclear. This study investigated the conditions under which quantum kernels outperform the classical RBF kernel in industrial acoustic anomaly detection. Using autoregressive (AR) features and an unsupervised One-Class SVM, we compared a conventional radial basis function (RBF) kernel with two types of quantum kernels: QK1 (employing linear entanglement) and QK2 (employing all-to-all entanglement). Two datasets with distinct characteristics were used for the experiments: “Open Belt Drive (OBD),” where the anomaly class is separable, and “Mini 4WD (M4W),” where anomaly distributions overlap. While all methods achieved perfect classification for the OBD dataset, the quantum kernels reached optimal performance using fewer features. In contrast, for the M4W dataset, QK2 significantly outperformed the RBF kernel, achieving an F1 score of 0.893 and an FPR of 0.033 (compared to 0.428 and 0.600 for RBF, respectively). t-SNE analysis confirmed that the M4W anomaly data were difficult to separate in the conventional feature space. These results suggest that quantum kernels offer the greatest advantage in problems involving overlapping distributions and that analyzing separability via t-SNE helps identify anomaly detection tasks suitable for the application of quantum technology.

Albert Nieto-Morales, Rodrigo Gil-Merino

Takao Tomono, Kazuya Tsujimura

We developed a novel workflow that leverages Quantum Kernel feature space expressiveness combined with classical dimensionality reduction techniques. This workflow enables detection and visual identification of multiple simultaneous anomalies using acoustic data from a single non-contact sensor. Our newly developed method provides an intuitive interface for operators to identify specific anomalies. By combining conventional Mel-frequency cepstral coefficients (MFCC) with principal component analysis (PCA), the newly constructed Quantum Kernel achieves near-perfect classification (F1 = 1.0 under specific conditions: j ≥ 9 features, file-level data splits) of complex multi-source anomalies, significantly outperforming the classical RBF kernel (F1 ≈ 0.75-0.76) and 1D-CNN autoencoder baseline (F1 ≈ 0.60-0.85). The Quantum Kernel's superior representational power enables accurate anomaly detection with minimal training data. Our results demonstrate simulation-based evidence of potential quantum advantage under ideal conditions, where the performance of our method dramatically exceeds that of classical approaches as the number of features increases, pending validation on noisy intermediate-scale quantum (NISQ) hardware. This hybrid quantum-classical machine learning approach demonstrates significant potential for industrial applications, particularly for complex time-series data in data-scarce regimes where classical methods exhibit limited performance.

Mingchao Guo, Shi‐Jie Pan, Wenmin Li +5 more