ABB Group

Deep models deployed on edge devices frequently encounter resource variability, which arises from fluctuating energy levels, timing constraints, or prioritization of other critical tasks within the system. State-of-the-art machine learning pipelines generate resource-agnostic models, not capable to adapt at runtime. In this work we introduce Resource-Efficient Deep Subnetworks (REDS) to tackle model adaptation to variable resources. In contrast to the state-of-the-art, REDS use structured sparsity constructively by exploiting permutation invariance of neurons, which allows for hardware-specific optimizations. Specifically, REDS achieve computational efficiency by (1) skipping sequential computational blocks identified by a novel iterative knapsack optimizer, and (2) leveraging simple math to re-arrange the order of operations in REDS computational graph to take advantage of the data cache. REDS support conventional deep networks frequently deployed on the edge and provide computational benefits even for small and simple networks. We evaluate REDS on seven benchmark architectures trained on the Visual Wake Words, Google Speech Commands, Fashion-MNIST and CIFAR10 datasets, and test on four off-the-shelf mobile and embedded hardware platforms. We provide a theoretical result and empirical evidence for REDS outstanding performance in terms of submodels' test set accuracy, and demonstrate an adaptation time in response to dynamic resource constraints of under 40$\mu$s, utilizing a 2-layer fully-connected network on Arduino Nano 33 BLE.

This paper presents a physics-informed neural network approach for dynamic modeling of saturable synchronous machines, including cases with spatial harmonics. We introduce an architecture that incorporates gradient networks directly into the fundamental machine equations, enabling accurate modeling of the nonlinear and coupled electromagnetic constitutive relationship. By learning the gradient of the magnetic field energy, the model inherently satisfies energy balance (reciprocity conditions). The proposed architecture can universally approximate any physically feasible magnetic behavior and offers several advantages over lookup tables and standard machine learning models: it requires less training data, ensures monotonicity and reliable extrapolation, and produces smooth outputs. These properties further enable robust model inversion and optimal trajectory generation, often needed in control applications. We validate the proposed approach using measured and finite-element method (FEM) datasets from a 5.6-kW permanent-magnet (PM) synchronous reluctance machine. Results demonstrate accurate and physically consistent models, even with limited training data.

Effective thermal management is one of the primary design bottlenecks for high-power transformers. While rectangular windings offer good fill factor and current capacity, a characterization of their thermal performance is limited. This paper presents a comparative analysis of edgewise and flatwise rectangular windings for transformers through FEM. Six 1-kW transformer designs with different conductor thickness, which results in "high-profile" and "low-profile" transformers, were analyzed in terms of thermal performance. The analysis revealed that the flatwise structure provides a parallel, simultaneous thermal path that is consistently superior, achieving lower peak temperatures and more uniform heat distribution across all geometries in both natural and forced-air cooling. The edgewise structure, conversely, presents a series thermal path that impedes heat dissipation from the inner turns, leading to localized hotspots that worsen as the profile becomes higher. A representative edgewise and flatwise transformer with a square window area was implemented in a 100 V:100 V, 1-kW CLLC bidirectional resonant converter at 28.5 kHz switching frequency for experimental comparison under natural and forced-air cooling conditions. The experimental results in both validated the thermal superiority of the flatwise structure, demonstrating lower peak winding temperatures and reduced thermal gradients between the winding and the outer surfaces of the transformer.

Grid design and battery integration for industrial microgrids pose significant challenges due to fluctuating loads and the intermittency of renewable energy sources. Preliminary frontend engineering and design (preFEED) enables cost-effective assessments based on high-level system modelling. This study introduces a techno-economic preFEED framework tailored to industrial microgrids, with a focus on battery energy storage systems (BESS) that balance energy shifting and grid stability under uncertainty. The framework emphasizes cost-optimal operation while incorporating grid-stability considerations, drawing insights from a comprehensive meta-study on microgrid design. We apply our framework to a mining use case, in which electrification by means of renewables and battery energy systems shall reduce the carbon footprint and contribute to more sustainable mining. Our results show that complex and costly FEED design can be streamlined through an interactive and user-friendly preFEED approach, enabling early-stage techno-economic evaluations. This methodology serves as a practical entry point for site managers and facilitates informed decision-making for a transition to low-carbon industrial site operation.

Modern software systems increasingly depend on open-source and commercial third-party components, which can expose them to cybersecurity risks present within these external dependencies. The vulnerability assessment of such components remains predominantly manual, creating bottlenecks that slow remediation efforts. This work introduces VISOR, an automated, LLM-driven framework that collects vulnerability information from various online sources and orchestrates the selection and invocation of multiple custom-designed vulnerability-assessment tools to generate structured and actionable reports for human experts. The framework supports two workflows: a linear agentic workflow that generates and then refines a well-parameterized tool plan, and a non-linear agentic workflow that adaptively revises the analysis through iterative refinement, dynamic tool selection, and evidence-guided adjustments. At its core, VISOR collects and correlates evidence to determine whether a reported vulnerability is applicable to a target software repository and whether the vulnerable code is reachable within the codebase. By design, the framework does not automatically apply mitigations to fix vulnerabilities; instead, it assists cybersecurity experts by providing evidence-based assessments and proposed mitigation actions in a structured report, enabling informed remediation decisions. Experiments show that the framework achieves 76% agreement with cybersecurity expert assessments while substantially reducing manual effort during initial vulnerability triage.