Student Seminar 202612

[Student Seminar 202612] A Probe Employing Elliptical Excitation and High-Density TMR Array Sensors for the Inspection of Small-Diameter Tubes

Speaker:  Jingyi Wang

Time:     15:00-17:00, Nov. 29th

Location:  SIST 1A-200

Host:      Prof. Ye

Abstract:

Neutron flux tubes are situated within nuclear reactors and are susceptible to deterioration and potential rupture due to exposure to high-pressure water currents and nuclear radiation. Therefore, the periodic assessment of neutron flux tubes is crucial for ensuring the safety of nuclear power plants. Nevertheless, the quantitative assessment of neutron flux tubes is still a challenging problem, due to the difficulties in constructing array probes given the tube's small inner diameter of 5 mm. In this study, we developed a probe using high-density integrated tunneling magnetoresistive (TMR) sensors and an elliptical excitation for the evaluation of tiny tubes. The probe consists of 16 sensors, which are bare die chips cut from a wafer and positioned along the circumference of an elliptical printed circuit board (PCB). The dimensions of the TMR sensors are 0.5 mm in length and 0.5 mm in width, allowing them to fit into a small diameter tube while achieving a high-resolution measurement of the magnetic field. The probe's novel elliptical design improves its sensitivity to defects with uncertain orientations and offers increased space for sensor insertion. In contrast to a circular configuration, which only responses to the defect boundary, the elliptical configuration responses to the whole defect region and yields a more prominent defect indication.

Defect characteristics are derived from the obtained magnetic field images, including the signal's morphology in axial and circumferential directions, peak values, and average values, which are used for defect identification and quantification. The depth is quantified using a fitting formula that correlates the signal with the width and depth. The Probe is experimentally assessed using tube samples with defects that vary in shape, length, depth, and width. The results indicate an average depth quantification error of 3.41\% and a mean relative length measurement error of 2.96\%. The low quantification error facilitates accurate wear assessment, hence improving the operational safety of nuclear power plants and extending the lifespan of neutron flux tubes.

Bio:

Jingyi Wang received the B.S. degree in Electronic Information Engineering from Shanghaitech University, Shanghai, China, in 2023. She is currently pursuing the MS. degree with ShanghaiTech University, Shanghai, China. Her current research interests include the development of a small-diameter array probe and array signal processing.

[Student Seminar 202612] Pediatric Epilepsy Automated Classification model via Latent-space Enhanced Learning

Speaker:  Chenghao Xue

Time:     15:00-17:00, Nov. 29th

Location:  SIST 1A-200

Host:      Prof. Ye

Abstract:

Precise identification of epilepsy subtypes is fundamental for effective clinical intervention, yet distinguishing between specific syndromes remains a significant challenge. Traditional diagnostic approaches largely depend on the visual interpretation of long-term electroencephalography (EEG) by clinicians, which is labor-intensive and susceptible to subjective variability. While deep learning has emerged as a promising tool for automated diagnosis, existing models often struggle to capture the subtle electropathological boundaries between related syndromes. To address these limitations, we propose a robust deep learning framework for the automated classification of epilepsy subtypes, including West syndrome (WS), Benign Epilepsy with Centrotemporal Spikes (BECTS), and healthy controls. Our framework leverages the BIOT (Biosignal Transformer) foundation model, incorporating critical architectural refinements to enhance diagnostic precision. The proposed architecture consists of three primary modules: 1) a pre-trained BIOT encoder for high-capacity spatio-temporal feature extraction from raw EEG; 2) a contrastive learning module designed to maximize the discriminative distance between different seizure phenotypes in the latent space; and 3) an Attentive Feature Aggregator that replaces traditional global pooling with a learnable attention mechanism to dynamically distill key diagnostic biomarkers. Experimental results demonstrate that our approach achieves state-of-the-art performance, significantly outperforming conventional supervised methods. This work underscores the potential of integrating foundation models with attentive contrastive paradigms to provide objective, high-precision diagnostic support for epilepsy subtyping.

Bio:

Chenghao Xue was born in Shanghai, China, in 2001. He received the bachelor’s degree from University of Shanghai for Science and Technology, Shanghai, China, in 2023. At present, he is pursuing a master's degree at ShanghaiTech University.

[Student Seminar 202612] Multi-Frequency Stacked Array Eddy Current Measurement and Machine-Learning-Based Oxide Film Thickness Estimation Under Multi-Parameter Coupling

Speaker:  Liuxin Ge

Time:     15:00-17:00, Nov. 29th

Location:  SIST 1A-200

Host:      Prof. Ye

Abstract:

Accurate, non-destructive quantification of oxide film thickness on zirconium alloy cladding is essential for ensuring the integrity of nuclear fuel assemblies, yet remains challenging due to multiparameter coupling among oxide film thickness, crud deposition, and the material characteristics. This paper presents a novel method that combines a micro/nano-fabricated stacked array eddy current testing probe with a machine-learning-based quantification algorithm to achieve high-precision oxide

film thickness prediction under complex conditions. The method is based on multi-frequency data matrices that are obtained using the stacked array probe. Then, a CatBoost regression model is trained to calculate the oxide film thickness. The performance of the model is validated using a hybrid test set of experimental and simulated data, achieving a maximum absolute error of 2.42 µm and a mean absolute error (MAE) of 1.33 µm for samples without crud layer. For the case involving a crud layer, the model achieved a maximum absolute error of 3.88 µm and MAE of 1.46 µm for simulation data. The results demonstrate that the proposed approach effectively captures nonlinear, multi-parameter interactions, enabling accurate oxide film thickness estimation. This methodology provides a highsensitivity sensing tool and a robust data-driven quantification strategy for in-service nuclear fuel cladding inspection.

Bio:

Liuxin Ge was born in Ganzhou, China, in 2001. She received the bachelor's degree from School of Information Science and Technology, ShanghaiTech University in 2023.  At present, she is pursuing a master's degree at ShanghaiTech University.

[Student Seminar 202612] Multi-view Contrastive Learning for Cell Line-specific Synthetic Lethality Recommendation

Speaker:  Yingfan Rui

Time:     15:00-17:00, Nov. 29th

Location:  SIST 1A-200

Host:      Prof. Ye

Abstract:

Synthetic lethality (SL) is a promising strategy for targeted cancer therapy, as inhibiting SL partners of genes harboring cancer-specific mutations can selectively eliminate cancer cells while sparing normal cells. Although supervised learning-based SL prediction methods have demonstrated strong performance, they always focus on prediction of gene pairs. However, the core demand of researchers is not identifying which gene pairs are Synthetic Lethality (SL), but rather among a large number of genes, which SL partner genes should be prioritized for validation. Hence, we propose a method aiming to upgrade SL prediction (a binary classification task) to a precise Top-N recommendation task by developing a multimodal representation learning framework. Our method leverages multimodal information, integrates modality-specific and cross-modal collaborative perspectives to characterize gene features from multiple dimensions, and realizes the recommendation of synthetic lethality partner genes based on multi-modal contrastive learning.

Bio:

Yingfan Rui was born in Shanghai, China, in 2001. She received her bachelor’s degree from ShanghaiTech University, Shanghai, China, in 2023 and is currently pursuing a master’s degree at the same institution. Her research focuses on bioinformatics, data science, and machine learning, with a particular interest in synthetic lethality prediction for precision cancer therapy. 

[Student Seminar 202612] A Online Noise2Self-Based Framework for Real-Time MCG and MMG Signal Filtering

Speaker:  Xingshen Hou

Time:     15:00-17:00, Nov. 29th

Location:  SIST 1A-200

Host:      Prof. Ye

Abstract:

Bio-magnetic signals such as magnetocardiography (MCG) and magnetomyography (MMG) are extremely weak and are often severely contaminated by environmental interference, power-line noise, baseline drift, and non-stationary physiological artifacts. While conventional filtering techniques and supervised deep-learning-based denoising methods have shown effectiveness in specific scenarios, they either rely on strong prior assumptions, require clean reference signals, or fail to satisfy the stringent latency constraints of real-time bio-magnetic applications.

In this work, we propose a real-time, self-supervised bio-magnetic signal denoising framework based on Noise2Self, termed dl_n2s_rt, specifically designed for continuous online processing. A lightweight one-dimensional blind-spot convolutional network (TinyDenoiser) is developed to satisfy strict real-time constraints, and a dual blind-spot strategy combining input masking and center-zero convolution is employed to rigorously enforce the self-supervised learning assumption. To enable stable online adaptation, the framework integrates block-wise training with overlap-add (OLA) reconstruction, time-budget-aware online updates, event-preserving constraints, and frequency-domain regularization to maintain waveform fidelity and physiological interpretability.

Extensive simulations are conducted on synthetic MCG and MMG signals under a wide range of noise conditions from -50 dB to 50 dB. The proposed method is systematically evaluated in terms of computational latency, convergence behavior, power-line interference suppression, spectral structure preservation, signal-to-noise ratio improvement (both proxy and true SNR), in-band energy consistency, and time delay. Experimental results demonstrate that dl_n2s_rt consistently meets real-time requirements, achieving stable convergence within a limited training window while maintaining near-zero temporal delay. Compared with traditional FIR/IIR filters and supervised deep learning baselines, the proposed method exhibits superior or comparable denoising performance across most noise conditions, particularly in low- and medium-SNR regimes. Moreover, the analysis reveals distinct performance characteristics between MCG and MMG signals, highlighting the impact of signal non-stationarity on self-supervised denoising behavior and delineating the applicability boundaries of Noise2Self-based methods.

Overall, this work presents the first practical realization of Noise2Self for real-time bio-magnetic signal filtering, offering a robust, label-free, and computationally efficient solution for online MCG/MMG processing. The proposed framework provides a viable foundation for real-time bio-magnetic monitoring, event detection, and future deployment in wearable and multi-channel OPM-based systems.

Bio:

Xingshen Hou was born in Shanghai, China, in 2000. He received the bachelor’s degree from ShanghaiTech University, Shanghai, China, in 2022. At present, he is pursuing a master’s degree at ShanghaiTech University. His main research interests include wireless sensor networks, signal processing, and bio-magnetic signal measurement.