Prof. Shestakevych presented a framework for using AI-assisted scoping to support the selection and integration of multimodal sensing technologies in digital psychotherapy, with a particular focus on virtual reality (VR) art therapy and the assessment of anxiety-related conditions. Her presentation addressed the challenge of translating complex physiological and behavioural measurements into interpretable information that can support psychologists in monitoring therapeutic sessions and making better-informed clinical decisions.

The proposed approach combines conceptual modelling, structural analysis, impact mapping and multi-criteria evaluation to examine the relationships between sensor characteristics, psychophysiological indicators, clinical requirements and operational constraints. The presentation outlined a conceptual architecture comprising a sensor data acquisition layer, a pipeline for synchronising and preprocessing heterogeneous data streams, an analytical core using signal-processing and machine-learning methods, and a presentation layer providing interpretable indicators through a clinical decision-support interface. Potential data sources include VR headsets, wearable devices, physiological sensors and eye-tracking technologies. The framework also considers practical factors such as measurement accuracy, sampling frequency, patient comfort, device cost, API availability and the reliability of the relationships between physiological signals and psychological states.

An important aspect of the proposed framework is its emphasis on interdisciplinary collaboration. Psychologists would define clinical requirements and provide expert annotations, data scientists would develop and validate analytical methods, and hardware developers would use the identified needs to improve existing devi or design new sensing technologies. The presentation also explored possible implementation scenarios, ranging from an initial knowledge base and decision-support tool for psychologists to a more comprehensive technical infrastructure for multimodal data acquisition and analysis. The work was presented as a conceptual framework and development roadmap, with the potential to support more personalised, evidence-based and minimally invasive approaches to digital psychotherapy.

Sensor_Fusion_VR_ART.pdf

Prof. Muliarevych presented approaches to real-time decision-making in distributed systems, focusing on the transition from sensor data fusion to reliable actions under conditions of latency, uncertainty and system failure. His presentation addressed a central challenge in distributed intelligent systems: obtaining a sufficiently accurate and up-to-date view of the environment and making decisions within strict time constraints, even when data sources, communication channels and computational resources are distributed across multiple devices and services.

He introduced an end-to-end decision pipeline connecting sensor acquisition, state estimation, decision policy, distributed delivery and actuation. The presentation highlighted that successful sensor fusion is not sufficient on its own: the resulting state must also be timely, reliable and accompanied by information about uncertainty. Prof. Muliarevych discussed the importance of data freshness, event ordering and stateful processing, as well as the consequences of delayed, conflicting or incomplete observations. He also examined how decision-making can be distributed across edge devices, fog infrastructure and cloud services, emphasising the need to place decision authority at the lowest level that has sufficient information and authority to act safely. The selection of an appropriate architecture involves balancing latency, computational cost, communication overhead, privacy and reliability.

The presentation further explored mechanisms for improving the correctness and observability of distributed decisions. These included deadline-aware processing, freshness-aware sensor fusion, explicit representation of uncertainty, handling conflicting events, and monitoring operational metrics through tools such as Grafana. An anonymised production case study illustrated how these principles could be applied to event-stream processing and distributed decision pipelines. Connecting these concepts with his previous research on motion detection and tracking, neural-network-based target recognition, resource allocation, communication formats and protection against prompt-injection attacks, Prof. Muliarevych presented a common research challenge: ensuring that decisions remain reliable as data, resources, latency and trust conditions change in real time. His contribution provided a systems-level perspective relevant to the development of robust intelligent applications, including those that rely on heterogeneous sensor data and autonomous decision-making.

Real-time_Decision-Making.pdf

Dr Basystiuk presented approaches to multimodal data fusion for intelligent industrial systems, focusing on fault-tolerant and explainable deep learning for autonomous condition monitoring. His presentation addressed a key challenge in Industry 4.0: individual sensors provide only partial information about the state of industrial equipment. For example, temperature measurements may reveal a problem relatively late, vibration signals can respond rapidly to mechanical changes, and acoustic emissions may provide even earlier indications of emerging faults. Combining these complementary sources can improve the reliability of monitoring and reduce false alarms.

He discussed different fusion strategies, including early fusion of sensor features, late fusion of individual model outputs, and a proposed hybrid architecture based on a shared latent representation and adaptive weighting. The architecture combines specialised processing pipelines for different modalities, including convolutional neural networks for acoustic and thermal data, wavelet analysis and recurrent networks for motor-current signals, and transformer-based processing of operator logs. Cross-modal attention is used to integrate heterogeneous representations, while adaptive gating assigns greater weight to more reliable data streams and suppresses the influence of noisy or failing sensors. The resulting representation supports several monitoring tasks within a unified framework, including fault detection, quality control and predictive maintenance.

An important part of the presentation concerned the practical deployment and interpretability of multimodal systems. Dr Basystiuk discussed industrial data infrastructure and real-time processing, including MQTT-based communication, streaming technologies and time-series storage, as well as feature extraction in the time, frequency and time–frequency domains. He also addressed explainability through SHAP and LIME, which can help engineers understand the contribution of individual features to a model’s predictions. Additional mechanisms, including uncertainty estimation and analysis of adaptive fusion weights, were presented as ways to identify distribution shifts and potential over-reliance on particular sensor modalities. Industrial examples included 3D printing, laser metal deposition and digital twins for equipment diagnostics. The presentation identified synchronisation of data streams with substantially different sampling rates and the deployment of inference directly on machine controllers as important remaining challenges. Overall, the proposed approach aims to combine predictive performance, resilience to sensor degradation and interpretable decisions in industrial condition-monitoring systems.

Multimodal_Data_Fusion_Presentation.pdf

Prof. Olena Pavliuk presented research on human-centric sensor fusion and federated learning for the safe, adaptive and energy-efficient control of autonomous guided vehicles (AGVs) in industrial environments. She holds a professorial position at Lviv Polytechnic National University (LPNU) and is also an assistant professor (adiunkt) at the Silesian University of Technology (SUT). Within TUAI, she coordinates cooperation between the two universities, helping to connect their research activities and strengthen LPNU’s involvement in the network.

Her presentation addressed the challenges of Industry 5.0, where autonomous mobile robots must operate safely alongside people while responding to changing environmental conditions and resource constraints. The proposed approach integrates heterogeneous data from AGVs, including battery measurements, inertial sensors, wireless connectivity indicators, odometry and industrial communication systems, with human-related information such as movement patterns and physiological signals. Combining these sources can improve situation awareness, support anomaly detection and enable control decisions that account for both machine performance and human activity.

A central element of the work is a distributed data-processing and federated-learning architecture spanning edge, fog and cloud levels. The presentation covered data preprocessing, signal filtering, anomaly detection, missing-data recovery and adaptive consensus-based correction, alongside federated models for human activity recognition, battery-state estimation and prediction of wireless communication conditions. These outputs can inform multi-agent path planning, helping AGVs select routes while considering constraints such as battery state, connectivity, safety and energy consumption. The presentation also discussed reliability analysis as a means of assessing the robustness and availability of the integrated system.

The proposed framework brings together sensor fusion, privacy-preserving distributed learning and predictive-adaptive control to support more reliable and human-aware industrial robotics. Future research directions include edge AI for real-time operation on resource-constrained AGVs, improved resilience to sensor failures, explainable AI for transparent decision-making, and the use of digital twins for simulation and optimisation. The work illustrated how collaboration between LPNU and SUT within TUAI can connect complementary expertise in sensor fusion, federated learning and intelligent industrial systems.

Human-Centric_Sensor_Fusion.pdf