Conference Agenda

Real-time fault detection in electromagnetic compatibility is critical for dependable electronic system operation. This paper evaluates anomaly detection algorithms applied to time-domain electromagnetic compatibility measurements, meanwhile also addressing the lack of standardized datasets in this area. A novel publicly available dataset was generated using a custom test bench, injecting controlled disturbances to simulate real-world anomalies of varying severity and frequency. In terms of representation, the raw three-phase current signals were compared to a lower-dimensional representation obtained by leveraging Clarke-Park transformations. This study systematically compares a variety of classical statistical methods and modern machine learning techniques. The performance of these methods was assessed using five key anomaly detection performance metrics, providing insights in the models' effectiveness. The results demonstrate the importance of appropriate preprocessing and relation to key characteristics of the anomaly detection methods.

The detection of anomalies in automotive sensor signals distorted by intentional electromagnetic interference (IEMI) is investigated through the support of autoencoders. These are designed to extract the most important features of data in a dimensionally reduced latent representation. Furthermore, several classification methods for electromagnetic signal disturbances are analyzed and compared in this compressed latent space. The performance of the methods is compared to a baseline model and evaluated on different metrics. The aim of the investigated methods is to achieve a high recall (sensitivity) and minimal false negative misclassifications, thus detecting all possible anomalies, for which a neural network classifier is shown to be the best-performing model.

This paper presents a novel approach for optimizing the design of power distribution networks (PDNs) in heterogeneous multi-chiplet systems by leveraging a graph neural network (GNN). The proposed method predicts self-impedances at observation ports and optimizes decoupling capacitor placement to minimize PDN impedance across multiple voltage domains. Addressing the inherent complexity of PDN design in multi-chiplet architectures, a GNN-based surrogate model is employed to efficiently explore the high-dimensional design space, streamlining capacitor selection and placement. The optimization framework integrates PDN design objectives and constraints into a feedback-driven deep reinforcement learning process, enabling impedance reduction while minimizing the total number of capacitors. This approach ensures adherence to key design rules while achieving optimal PDN performance within a targeted bandwidth. By combining GNN-based modeling with reinforcement learning, this work represents a significant advancement in PDN design methodology, offering a faster and more cost-effective solution for the heterogeneous integration of multi-chiplet systems.

Reliable acquisition and processing of sensor data are crucial for numerous technical systems, particularly in safety-critical applications. However, intentional electromagnetic interference (IEMI) can distort sensor signals, leading to inaccurate measurements or communication failures. This paper investigates anomaly detection in sensor signals affected by IEMI, using data generated from LTspice circuit simulations. The simulations model both undisturbed signals and disturbed signals by injecting double-exponential electromagnetic pulses into a PSI5 sensor communication system. We compare traditional statistical methods with deep learning-based approaches. Two LSTM-based models were implemented: (1) a regression-based approach detecting anomalies through prediction errors and (2) a classification-based approach directly identifying anomalous windows. Both approaches were evaluated against a z-score-based baseline. The results show that the classification-based model achieves the highest anomaly detection performance, with an area under the receiver operating characteristic curve of 0.92 and an average precision of 0.79, significantly outperforming the baseline. Future work will explore hybrid models integrating autoencoders and LSTMs in the latent space to enhance robustness.