Jiale Liu

Intelligent fault diagnosis functions as a necessary tool to prevent substantial damage to industrial products and enhance system reliability. While Artificial Neural Networks (ANNs) have been extensively studied in this context, they still face substantial challenges in resource consumption, robustness, and generalizability. To overcome these limitations, researchers have developed the third-generation neural network based on brain’s structure, namely the Spiking Neural Network (SNN), which leverages the concept of time steps and spiking signals for enhanced spatiotemporal feature processing and energy efficiency. This paper proposes the Wide Spiking Residual Grouping Attention Framework (WSRGA-FW), which incorporates the advantages of both ANN and SNN. The WSRGA-FW employs Extended Gramian Representation for signal encoding to reduce noise impact, followed by a tailored ANN with wide convolutional kernels, optimized residual structures, and Grouped Perception Generation (GPG) Layers. These augmentations increase the network’s representation and robustness, particularly in noisy environments. The backbone ANN is transformed into an SNN model, allowing deployment in portable and miniaturized devices with improved application prospects. Performance evaluations across various noisy scenarios using bearing fault datasets demonstrate that WSRGA-FW surpasses existing networks. Visualization of firing rates and energy consumption calculation contribute to the interpretability and intrinsic energy efficiency advantages of SNNs.

Gravitational Wave (GW) detection is a pivotal field in astrophysics, gaining prominence for its role in deciphering cosmic phenomena. Traditional machine learning algorithms in GW detection, while effective, grapple with challenges in parallelism, scalability, efficiency in handling high-dimensional data and noise interference. To mitigate these limitations, this paper proposes the Quantum-Gravitational Transformer (QGFormer), a novel quantum-classical hybrid transformer for gravitational wave detection. The QGFormer consists of a vision transformer encoder and a quantum classifier following the autoencoder architecture, which first segments the original input signal into small patches and maps them to higher dimensions, and then inputs the higher-dimensional feature map into the quantum classifier for prediction. The excellent parallelism inherent in the multi-head self-attention mechanism and quantum computing makes QGFormer expected to achieve high computational efficiency and scalability on quantum computing devices. Additionally, the super-density coding and quantum state compression techniques that could be used in these devices help mitigate the effects of noise. In experiments on the Gravitational Spy dataset, which utilizes data collected by the LIGO telescope, our results show that QGFormer outperforms the comparison models in accuracy and F1-score, demonstrating the promising future of quantum machine learning approaches in this field.