Search in the TU Delft Repository Collections
Recently Added Records
The AltiCube+ mission involves the on-orbit assembly of multiple CubeSats to build two radar interferometers. To accomplish this, vision-based relative navigation is required both for target localization during rendezvous and for post-docking relative displacement estimation of structural deformation. This paper presents the design, simulation, and experimental validation of a compact monocular vision-based system capable of fulfilling both functions using a single camera architecture. The proposed system consists of a camera mounted on the chaser spacecraft and an active LED fiducial marker configuration on the target spacecraft. A non-planar marker geometry combined with colour-based segmentation enables robust detection over distances ranging from 10 m to close contact. Sub-pixel marker localization is performed using weighted centroiding and Gaussian fitting, followed by pose estimation through an Efficient Perspective-n-Point (EPnP) algorithm with iterative refinement. The results show that the proposed architecture can support both rendezvous localization and micrometre-scale structural deformation monitoring, providing a compact sensing solution for CubeSat assembly missions.
...
The AltiCube+ mission involves the on-orbit assembly of multiple CubeSats to build two radar interferometers. To accomplish this, vision-based relative navigation is required both for target localization during rendezvous and for post-docking relative displacement estimation of structural deformation. This paper presents the design, simulation, and experimental validation of a compact monocular vision-based system capable of fulfilling both functions using a single camera architecture. The proposed system consists of a camera mounted on the chaser spacecraft and an active LED fiducial marker configuration on the target spacecraft. A non-planar marker geometry combined with colour-based segmentation enables robust detection over distances ranging from 10 m to close contact. Sub-pixel marker localization is performed using weighted centroiding and Gaussian fitting, followed by pose estimation through an Efficient Perspective-n-Point (EPnP) algorithm with iterative refinement. The results show that the proposed architecture can support both rendezvous localization and micrometre-scale structural deformation monitoring, providing a compact sensing solution for CubeSat assembly missions.
Probabilistic Nonlinear Spatial Filtering for Multi-Talker Speech Enhancement
Multi-Source Speech Presence Probability Estimation and DNN-Guided MVDR Beamforming in Multichannel Reverberant Environments
A linear MVDR beamformer followed by a single-channel post-filter is MMSE-optimal only when the interference is Gaussian. That assumption fails when competing talkers are present, and the optimal estimator is then nonlinear. This thesis derives that estimator from a composite hypothesis structure which splits the per-bin signal-activity state into two questions: whether the target is present, and which interferer dominates. In closed form the estimator is a set of target-directed MVDR–Wiener filters, one per interference hypothesis, each weighted by an interferer posterior, with the weighted sum multiplied by a target posterior. These two posteriors are the multi-source speech presence probabilities, and the chain rule supplies the product structure exactly, without assuming that they factorise.
The estimator is shown to be structurally equivalent to the nonlinear spatial filter of Tesch and Gerkmann. Two sub-networks with no shared parameters estimate only these posteriors; all filtering is performed by the model-based filters. On a simulated five-talker reverberant corpus captured with a three-microphone array, the system improves SI-SDR by 11.26 dB and PESQ by 0.53, exceeding an end-to-end complex-mask baseline of comparable cost and more than doubling the gain of an oracle MVDR. Because the network outputs posteriors rather than a mask, the estimated quantities can be inspected directly and are shown to recover the intended signal-activity decomposition. ...
The estimator is shown to be structurally equivalent to the nonlinear spatial filter of Tesch and Gerkmann. Two sub-networks with no shared parameters estimate only these posteriors; all filtering is performed by the model-based filters. On a simulated five-talker reverberant corpus captured with a three-microphone array, the system improves SI-SDR by 11.26 dB and PESQ by 0.53, exceeding an end-to-end complex-mask baseline of comparable cost and more than doubling the gain of an oracle MVDR. Because the network outputs posteriors rather than a mask, the estimated quantities can be inspected directly and are shown to recover the intended signal-activity decomposition. ...
A linear MVDR beamformer followed by a single-channel post-filter is MMSE-optimal only when the interference is Gaussian. That assumption fails when competing talkers are present, and the optimal estimator is then nonlinear. This thesis derives that estimator from a composite hypothesis structure which splits the per-bin signal-activity state into two questions: whether the target is present, and which interferer dominates. In closed form the estimator is a set of target-directed MVDR–Wiener filters, one per interference hypothesis, each weighted by an interferer posterior, with the weighted sum multiplied by a target posterior. These two posteriors are the multi-source speech presence probabilities, and the chain rule supplies the product structure exactly, without assuming that they factorise.
The estimator is shown to be structurally equivalent to the nonlinear spatial filter of Tesch and Gerkmann. Two sub-networks with no shared parameters estimate only these posteriors; all filtering is performed by the model-based filters. On a simulated five-talker reverberant corpus captured with a three-microphone array, the system improves SI-SDR by 11.26 dB and PESQ by 0.53, exceeding an end-to-end complex-mask baseline of comparable cost and more than doubling the gain of an oracle MVDR. Because the network outputs posteriors rather than a mask, the estimated quantities can be inspected directly and are shown to recover the intended signal-activity decomposition.
The estimator is shown to be structurally equivalent to the nonlinear spatial filter of Tesch and Gerkmann. Two sub-networks with no shared parameters estimate only these posteriors; all filtering is performed by the model-based filters. On a simulated five-talker reverberant corpus captured with a three-microphone array, the system improves SI-SDR by 11.26 dB and PESQ by 0.53, exceeding an end-to-end complex-mask baseline of comparable cost and more than doubling the gain of an oracle MVDR. Because the network outputs posteriors rather than a mask, the estimated quantities can be inspected directly and are shown to recover the intended signal-activity decomposition.
Industrial quality control needs inspection systems that are accurate, explainable, and adaptable to new product categories with few labelled data. Traditional deep-learning detectors, such as memory-bank, normalising-flow, and reconstruction-based methods, reach near-perfect AUROC on the established MVTec AD benchmark, yet they emit only a binary defect verdict and a heatmap. They give no human-readable account of why a sample was flagged or what type of defect it carries, and that lack of interpretability is a primary obstacle to deploying automated inspection in safety- and audit-critical manufacturing.
This thesis asks whether an open vision-language model (VLM), fine-tuned on structured reasoning traces and refined with reinforcement learning, can deliver explainable anomaly detection that is competitive with the state of the art, on products it was not trained on and without per-product retraining or reference images. We build an end-to-end framework that fine-tunes the open-weight Qwen2.5-VL family on structured chain-of-thought reasoning traces. Each trace follows a six-phase template (Framing, Scan, Focus, Evaluate, Alternatives, Decide) and emits an XML output with a free-text rationale, a defect-type tag, a spatial location, and a binary verdict. Using Gemini 2.5-Flash we generate 14,472 such traces from the Real-IAD multi-view dataset, released here as the AnomalyThink dataset, and partition them into product- and class-stratified splits of 6,000 for supervised fine-tuning, 4,236 for GRPO, and 4,236 held out.
We run a controlled four-factor comparison of fine-tuning configurations across model size (3B versus 7B), vision-encoder freezing, supervised-data scale (6K versus the full 14.5K union), and training epoch (up to four), with the unfrozen encoder evaluated only at the larger scale. This is followed by Group Relative Policy Optimisation (GRPO) with an unweighted two-function reward scoring four bounded signals: format, verdict, semantic type, and location. Evaluation uses the DS-MVTec (1,670 samples) and VisA (2,141 samples) subsets of MMAD, together with a held-out Real-IAD split.
Our main result comes from a three-stage pipeline in which each stage feeds the next: supervised fine-tuning, GRPO, and a second supervised stage trained on data the GRPO model produced. Supervised fine-tuning on the 6,000-trace dataset reaches 80.16 % balanced accuracy on DS-MVTec and 64.78 % on VisA. GRPO with our reward functions lifts this to 82.73 % and 70.39 %. We then roll that GRPO model out at k = 8 over all 10,236 training images, which is the 6,000-image SFT split plus the 4,236-image GRPO split. We keep the traces it gets right and have a teacher correct or rewrite the rest. The new corpus is rebuilt on the 6,000 SFT images. Training the base Qwen2.5-VL-7B on it reaches 82.80 % and 72.07 %, the best result on that backbone, and beats the released IAD-R1 model on both subsets (81.92 % and 71.34 %) under the same evaluation harness. We repeat the same recipe end to end on a second backbone, LLaVA-OneVision-7B-SI. There the curated corpus reaches 87.32 % and 72.65 % at the selected epoch against 87.66 % and 72.38 % for SFT plus GRPO, and 74.29 % on VisA by epoch 4. On both backbones the supervised stage matches the reinforcement-learned one and exceeds it on VisA. The finding is therefore not specific to one model family. The last stage is a supervised one. What GRPO gave us was not the final model but the data we used to train it.
GRPO on its own helps when the supervised baseline is weak. Applied to the earlier supervised baseline it raises balanced accuracy to 82.73 % on DS-MVTec, 70.39 % on VisA, and 80.87 % on the held-out Real-IAD split, exceeding IAD-R1 on DS-MVTec. Applied on top of the stronger curated dataset, GRPO did not improve binary accuracy in our experiments. Reward kept rising while balanced accuracy stayed 1.6 to 4.0 points below the supervised initialisation. This held for all three advantage estimators we tried, within a single 120-step probe on one initialisation. Two further results follow. Within our data, the 6,000-trace split beats the full 14.5K union by 8.5 balanced-accuracy points on DS-MVTec at matched epochs, a data-count effect rather than a trace-length one. And freezing the vision encoder consistently beats unfreezing it, most clearly on the harder VisA benchmark.
The framework is reproducible on commodity hardware, two RTX A6000 GPUs with ZeRO-3 CPU offloading, and yields a single 7B model that detects, localises, and explains industrial defects in natural language, with no per-product retraining and no pixel-level masks at inference. ...
This thesis asks whether an open vision-language model (VLM), fine-tuned on structured reasoning traces and refined with reinforcement learning, can deliver explainable anomaly detection that is competitive with the state of the art, on products it was not trained on and without per-product retraining or reference images. We build an end-to-end framework that fine-tunes the open-weight Qwen2.5-VL family on structured chain-of-thought reasoning traces. Each trace follows a six-phase template (Framing, Scan, Focus, Evaluate, Alternatives, Decide) and emits an XML output with a free-text rationale, a defect-type tag, a spatial location, and a binary verdict. Using Gemini 2.5-Flash we generate 14,472 such traces from the Real-IAD multi-view dataset, released here as the AnomalyThink dataset, and partition them into product- and class-stratified splits of 6,000 for supervised fine-tuning, 4,236 for GRPO, and 4,236 held out.
We run a controlled four-factor comparison of fine-tuning configurations across model size (3B versus 7B), vision-encoder freezing, supervised-data scale (6K versus the full 14.5K union), and training epoch (up to four), with the unfrozen encoder evaluated only at the larger scale. This is followed by Group Relative Policy Optimisation (GRPO) with an unweighted two-function reward scoring four bounded signals: format, verdict, semantic type, and location. Evaluation uses the DS-MVTec (1,670 samples) and VisA (2,141 samples) subsets of MMAD, together with a held-out Real-IAD split.
Our main result comes from a three-stage pipeline in which each stage feeds the next: supervised fine-tuning, GRPO, and a second supervised stage trained on data the GRPO model produced. Supervised fine-tuning on the 6,000-trace dataset reaches 80.16 % balanced accuracy on DS-MVTec and 64.78 % on VisA. GRPO with our reward functions lifts this to 82.73 % and 70.39 %. We then roll that GRPO model out at k = 8 over all 10,236 training images, which is the 6,000-image SFT split plus the 4,236-image GRPO split. We keep the traces it gets right and have a teacher correct or rewrite the rest. The new corpus is rebuilt on the 6,000 SFT images. Training the base Qwen2.5-VL-7B on it reaches 82.80 % and 72.07 %, the best result on that backbone, and beats the released IAD-R1 model on both subsets (81.92 % and 71.34 %) under the same evaluation harness. We repeat the same recipe end to end on a second backbone, LLaVA-OneVision-7B-SI. There the curated corpus reaches 87.32 % and 72.65 % at the selected epoch against 87.66 % and 72.38 % for SFT plus GRPO, and 74.29 % on VisA by epoch 4. On both backbones the supervised stage matches the reinforcement-learned one and exceeds it on VisA. The finding is therefore not specific to one model family. The last stage is a supervised one. What GRPO gave us was not the final model but the data we used to train it.
GRPO on its own helps when the supervised baseline is weak. Applied to the earlier supervised baseline it raises balanced accuracy to 82.73 % on DS-MVTec, 70.39 % on VisA, and 80.87 % on the held-out Real-IAD split, exceeding IAD-R1 on DS-MVTec. Applied on top of the stronger curated dataset, GRPO did not improve binary accuracy in our experiments. Reward kept rising while balanced accuracy stayed 1.6 to 4.0 points below the supervised initialisation. This held for all three advantage estimators we tried, within a single 120-step probe on one initialisation. Two further results follow. Within our data, the 6,000-trace split beats the full 14.5K union by 8.5 balanced-accuracy points on DS-MVTec at matched epochs, a data-count effect rather than a trace-length one. And freezing the vision encoder consistently beats unfreezing it, most clearly on the harder VisA benchmark.
The framework is reproducible on commodity hardware, two RTX A6000 GPUs with ZeRO-3 CPU offloading, and yields a single 7B model that detects, localises, and explains industrial defects in natural language, with no per-product retraining and no pixel-level masks at inference. ...
Industrial quality control needs inspection systems that are accurate, explainable, and adaptable to new product categories with few labelled data. Traditional deep-learning detectors, such as memory-bank, normalising-flow, and reconstruction-based methods, reach near-perfect AUROC on the established MVTec AD benchmark, yet they emit only a binary defect verdict and a heatmap. They give no human-readable account of why a sample was flagged or what type of defect it carries, and that lack of interpretability is a primary obstacle to deploying automated inspection in safety- and audit-critical manufacturing.
This thesis asks whether an open vision-language model (VLM), fine-tuned on structured reasoning traces and refined with reinforcement learning, can deliver explainable anomaly detection that is competitive with the state of the art, on products it was not trained on and without per-product retraining or reference images. We build an end-to-end framework that fine-tunes the open-weight Qwen2.5-VL family on structured chain-of-thought reasoning traces. Each trace follows a six-phase template (Framing, Scan, Focus, Evaluate, Alternatives, Decide) and emits an XML output with a free-text rationale, a defect-type tag, a spatial location, and a binary verdict. Using Gemini 2.5-Flash we generate 14,472 such traces from the Real-IAD multi-view dataset, released here as the AnomalyThink dataset, and partition them into product- and class-stratified splits of 6,000 for supervised fine-tuning, 4,236 for GRPO, and 4,236 held out.
We run a controlled four-factor comparison of fine-tuning configurations across model size (3B versus 7B), vision-encoder freezing, supervised-data scale (6K versus the full 14.5K union), and training epoch (up to four), with the unfrozen encoder evaluated only at the larger scale. This is followed by Group Relative Policy Optimisation (GRPO) with an unweighted two-function reward scoring four bounded signals: format, verdict, semantic type, and location. Evaluation uses the DS-MVTec (1,670 samples) and VisA (2,141 samples) subsets of MMAD, together with a held-out Real-IAD split.
Our main result comes from a three-stage pipeline in which each stage feeds the next: supervised fine-tuning, GRPO, and a second supervised stage trained on data the GRPO model produced. Supervised fine-tuning on the 6,000-trace dataset reaches 80.16 % balanced accuracy on DS-MVTec and 64.78 % on VisA. GRPO with our reward functions lifts this to 82.73 % and 70.39 %. We then roll that GRPO model out at k = 8 over all 10,236 training images, which is the 6,000-image SFT split plus the 4,236-image GRPO split. We keep the traces it gets right and have a teacher correct or rewrite the rest. The new corpus is rebuilt on the 6,000 SFT images. Training the base Qwen2.5-VL-7B on it reaches 82.80 % and 72.07 %, the best result on that backbone, and beats the released IAD-R1 model on both subsets (81.92 % and 71.34 %) under the same evaluation harness. We repeat the same recipe end to end on a second backbone, LLaVA-OneVision-7B-SI. There the curated corpus reaches 87.32 % and 72.65 % at the selected epoch against 87.66 % and 72.38 % for SFT plus GRPO, and 74.29 % on VisA by epoch 4. On both backbones the supervised stage matches the reinforcement-learned one and exceeds it on VisA. The finding is therefore not specific to one model family. The last stage is a supervised one. What GRPO gave us was not the final model but the data we used to train it.
GRPO on its own helps when the supervised baseline is weak. Applied to the earlier supervised baseline it raises balanced accuracy to 82.73 % on DS-MVTec, 70.39 % on VisA, and 80.87 % on the held-out Real-IAD split, exceeding IAD-R1 on DS-MVTec. Applied on top of the stronger curated dataset, GRPO did not improve binary accuracy in our experiments. Reward kept rising while balanced accuracy stayed 1.6 to 4.0 points below the supervised initialisation. This held for all three advantage estimators we tried, within a single 120-step probe on one initialisation. Two further results follow. Within our data, the 6,000-trace split beats the full 14.5K union by 8.5 balanced-accuracy points on DS-MVTec at matched epochs, a data-count effect rather than a trace-length one. And freezing the vision encoder consistently beats unfreezing it, most clearly on the harder VisA benchmark.
The framework is reproducible on commodity hardware, two RTX A6000 GPUs with ZeRO-3 CPU offloading, and yields a single 7B model that detects, localises, and explains industrial defects in natural language, with no per-product retraining and no pixel-level masks at inference.
This thesis asks whether an open vision-language model (VLM), fine-tuned on structured reasoning traces and refined with reinforcement learning, can deliver explainable anomaly detection that is competitive with the state of the art, on products it was not trained on and without per-product retraining or reference images. We build an end-to-end framework that fine-tunes the open-weight Qwen2.5-VL family on structured chain-of-thought reasoning traces. Each trace follows a six-phase template (Framing, Scan, Focus, Evaluate, Alternatives, Decide) and emits an XML output with a free-text rationale, a defect-type tag, a spatial location, and a binary verdict. Using Gemini 2.5-Flash we generate 14,472 such traces from the Real-IAD multi-view dataset, released here as the AnomalyThink dataset, and partition them into product- and class-stratified splits of 6,000 for supervised fine-tuning, 4,236 for GRPO, and 4,236 held out.
We run a controlled four-factor comparison of fine-tuning configurations across model size (3B versus 7B), vision-encoder freezing, supervised-data scale (6K versus the full 14.5K union), and training epoch (up to four), with the unfrozen encoder evaluated only at the larger scale. This is followed by Group Relative Policy Optimisation (GRPO) with an unweighted two-function reward scoring four bounded signals: format, verdict, semantic type, and location. Evaluation uses the DS-MVTec (1,670 samples) and VisA (2,141 samples) subsets of MMAD, together with a held-out Real-IAD split.
Our main result comes from a three-stage pipeline in which each stage feeds the next: supervised fine-tuning, GRPO, and a second supervised stage trained on data the GRPO model produced. Supervised fine-tuning on the 6,000-trace dataset reaches 80.16 % balanced accuracy on DS-MVTec and 64.78 % on VisA. GRPO with our reward functions lifts this to 82.73 % and 70.39 %. We then roll that GRPO model out at k = 8 over all 10,236 training images, which is the 6,000-image SFT split plus the 4,236-image GRPO split. We keep the traces it gets right and have a teacher correct or rewrite the rest. The new corpus is rebuilt on the 6,000 SFT images. Training the base Qwen2.5-VL-7B on it reaches 82.80 % and 72.07 %, the best result on that backbone, and beats the released IAD-R1 model on both subsets (81.92 % and 71.34 %) under the same evaluation harness. We repeat the same recipe end to end on a second backbone, LLaVA-OneVision-7B-SI. There the curated corpus reaches 87.32 % and 72.65 % at the selected epoch against 87.66 % and 72.38 % for SFT plus GRPO, and 74.29 % on VisA by epoch 4. On both backbones the supervised stage matches the reinforcement-learned one and exceeds it on VisA. The finding is therefore not specific to one model family. The last stage is a supervised one. What GRPO gave us was not the final model but the data we used to train it.
GRPO on its own helps when the supervised baseline is weak. Applied to the earlier supervised baseline it raises balanced accuracy to 82.73 % on DS-MVTec, 70.39 % on VisA, and 80.87 % on the held-out Real-IAD split, exceeding IAD-R1 on DS-MVTec. Applied on top of the stronger curated dataset, GRPO did not improve binary accuracy in our experiments. Reward kept rising while balanced accuracy stayed 1.6 to 4.0 points below the supervised initialisation. This held for all three advantage estimators we tried, within a single 120-step probe on one initialisation. Two further results follow. Within our data, the 6,000-trace split beats the full 14.5K union by 8.5 balanced-accuracy points on DS-MVTec at matched epochs, a data-count effect rather than a trace-length one. And freezing the vision encoder consistently beats unfreezing it, most clearly on the harder VisA benchmark.
The framework is reproducible on commodity hardware, two RTX A6000 GPUs with ZeRO-3 CPU offloading, and yields a single 7B model that detects, localises, and explains industrial defects in natural language, with no per-product retraining and no pixel-level masks at inference.
Conventional Life Cycle Assessment (LCA) is limited by its reliance on static and retrospective background datasets and by its predominantly inside-out modelling approach, where models are built strictly from internal factory operations and primary data, often missing sudden external disruptions. These limitations reduce their capability in context of the Critical Raw Material (CRM) value chains affected by geopolitical conflicts and climate-related disruptions. While external macro-events may alter Life Cycle Inventory (LCI) parameters, they are usually reported in fragmented and unstructured sources like daily news. Additionally, public news rarely reports direct LCA parameters or exact numerical changes, making it difficult to capture external events that genuinely impact LCA models. More recently, Large Language Models (LLMs) have emerged as a promising solution. However, without historical references and the capacity for complex reasoning or real-world interaction, LLMs are highly prone to fabricating or hallucinating missing numbers when asked to generate scenario values. This makes these events difficult to translate into structured, quantitative, and LCA-ready scenario inputs.
This study conducts a methodological exploration and develops an Event Knowledge Graph (EKG)-driven Agentic AI framework as a human-governed outside-in sensing layer for LCA scenario exploration. Because relying solely on LLMs to quantify qualitative news inevitably leads to numerical hallucinations, the objective is to examine how qualitative real-world macro-events can be reliably translated into structured, quantitative, and LCA-ready parameter-change scenarios without unconstrained AI guessing.
The main contribution of this research is a novel, validated, and reusable blueprint for the structured, low-hallucination risk translation between macro-disruptions in CRM value chains reported in real-time news and quantitative changes in LCI parameters. Crucially, the system explicitly positions Agentic AI as a supportive co-pilot, proving that human expertise remains indispensable in defining parameter scopes, setting scenario boundaries, and performing the final LCA interpretation. ...
This study conducts a methodological exploration and develops an Event Knowledge Graph (EKG)-driven Agentic AI framework as a human-governed outside-in sensing layer for LCA scenario exploration. Because relying solely on LLMs to quantify qualitative news inevitably leads to numerical hallucinations, the objective is to examine how qualitative real-world macro-events can be reliably translated into structured, quantitative, and LCA-ready parameter-change scenarios without unconstrained AI guessing.
The main contribution of this research is a novel, validated, and reusable blueprint for the structured, low-hallucination risk translation between macro-disruptions in CRM value chains reported in real-time news and quantitative changes in LCI parameters. Crucially, the system explicitly positions Agentic AI as a supportive co-pilot, proving that human expertise remains indispensable in defining parameter scopes, setting scenario boundaries, and performing the final LCA interpretation. ...
Conventional Life Cycle Assessment (LCA) is limited by its reliance on static and retrospective background datasets and by its predominantly inside-out modelling approach, where models are built strictly from internal factory operations and primary data, often missing sudden external disruptions. These limitations reduce their capability in context of the Critical Raw Material (CRM) value chains affected by geopolitical conflicts and climate-related disruptions. While external macro-events may alter Life Cycle Inventory (LCI) parameters, they are usually reported in fragmented and unstructured sources like daily news. Additionally, public news rarely reports direct LCA parameters or exact numerical changes, making it difficult to capture external events that genuinely impact LCA models. More recently, Large Language Models (LLMs) have emerged as a promising solution. However, without historical references and the capacity for complex reasoning or real-world interaction, LLMs are highly prone to fabricating or hallucinating missing numbers when asked to generate scenario values. This makes these events difficult to translate into structured, quantitative, and LCA-ready scenario inputs.
This study conducts a methodological exploration and develops an Event Knowledge Graph (EKG)-driven Agentic AI framework as a human-governed outside-in sensing layer for LCA scenario exploration. Because relying solely on LLMs to quantify qualitative news inevitably leads to numerical hallucinations, the objective is to examine how qualitative real-world macro-events can be reliably translated into structured, quantitative, and LCA-ready parameter-change scenarios without unconstrained AI guessing.
The main contribution of this research is a novel, validated, and reusable blueprint for the structured, low-hallucination risk translation between macro-disruptions in CRM value chains reported in real-time news and quantitative changes in LCI parameters. Crucially, the system explicitly positions Agentic AI as a supportive co-pilot, proving that human expertise remains indispensable in defining parameter scopes, setting scenario boundaries, and performing the final LCA interpretation.
This study conducts a methodological exploration and develops an Event Knowledge Graph (EKG)-driven Agentic AI framework as a human-governed outside-in sensing layer for LCA scenario exploration. Because relying solely on LLMs to quantify qualitative news inevitably leads to numerical hallucinations, the objective is to examine how qualitative real-world macro-events can be reliably translated into structured, quantitative, and LCA-ready parameter-change scenarios without unconstrained AI guessing.
The main contribution of this research is a novel, validated, and reusable blueprint for the structured, low-hallucination risk translation between macro-disruptions in CRM value chains reported in real-time news and quantitative changes in LCI parameters. Crucially, the system explicitly positions Agentic AI as a supportive co-pilot, proving that human expertise remains indispensable in defining parameter scopes, setting scenario boundaries, and performing the final LCA interpretation.
A threshold graph is generated from a single node by repeatedly adding either a node i connected to all existing nodes with a common link weight wi > 0 or a node i connected to none. Let Gw be a weighted threshold graph encoded by the weight vector w = (w1, w2, …, wN ) with wi ≥ 0. A closed-form expression for the pseudoinverse of its Laplacian matrix Qw is derived via spectral decomposition, which yields an explicit formula for the effective resistance matrix Ωw. We present a detailed structural characterization of the matrix Ωw and determine a subset of the spectrum of the matrix Ωw in terms of the weights wi . As an application, we show that when the missing links of a threshold graph are sequentially added in nondecreasing order of effective resistance, the threshold property of the graph is preserved at each step until the complete graph of the same size is obtained.
...
A threshold graph is generated from a single node by repeatedly adding either a node i connected to all existing nodes with a common link weight wi > 0 or a node i connected to none. Let Gw be a weighted threshold graph encoded by the weight vector w = (w1, w2, …, wN ) with wi ≥ 0. A closed-form expression for the pseudoinverse of its Laplacian matrix Qw is derived via spectral decomposition, which yields an explicit formula for the effective resistance matrix Ωw. We present a detailed structural characterization of the matrix Ωw and determine a subset of the spectrum of the matrix Ωw in terms of the weights wi . As an application, we show that when the missing links of a threshold graph are sequentially added in nondecreasing order of effective resistance, the threshold property of the graph is preserved at each step until the complete graph of the same size is obtained.