[Paper] ContinuumConductor : Decentralized Process Mining on the Edge-Cloud Continuum

Source: arXiv - 2512.07280v1

Overview

Process mining has long relied on gathering all event logs in a single, powerful data center before extracting insights. ContinuumConductor flips that script: it shows how to run the same mining steps directly on the edge‑cloud continuum that powers modern IIoT deployments, preserving privacy and cutting latency while still delivering high‑quality process models.

Key Contributions

• ContinuumConductor framework – a layered decision matrix that tells a system when and where (edge, fog, or cloud) to execute each mining sub‑task (pre‑processing, correlation, discovery).
• Trade‑off analysis for every step of the classic process‑mining pipeline, quantifying impacts on bandwidth, latency, compute load, and data privacy.
• Decision criteria (e.g., event‑rate thresholds, resource‑availability signals, privacy tags) that can be evaluated automatically at runtime.
• Prototype implementation spanning Raspberry‑Pi‑class edge nodes, a fog layer on Kubernetes, and a central cloud analytics service.
• Real‑world validation in an inland‑port logistics scenario, demonstrating measurable gains in response time and network usage without sacrificing model accuracy.

Methodology

1. Layered Architecture – The authors define three logical layers:

   • Edge (resource‑constrained devices that generate raw sensor events).
   • Fog (intermediate nodes with moderate compute, often on‑premise).
   • Cloud (elastic, high‑performance analytics cluster).

2. Decision Framework – For each mining phase, the framework evaluates a set of metrics (event volume, CPU/memory pressure, privacy level, required latency) against pre‑configured thresholds. The outcome is a placement decision (edge → fog → cloud).

3. Decentralized Mining Primitives – The team re‑implemented core mining algorithms (e.g., event‑log filtering, case correlation, process discovery via the Inductive Miner) as lightweight micro‑services that can run on any layer.

4. Dynamic Orchestration – A lightweight controller monitors the metrics in real time and migrates tasks across layers when conditions change (e.g., a sudden spike in sensor data pushes preprocessing to the fog).

5. Evaluation Setup – A synthetic workload mimicking port operations (container handling, truck arrivals, customs checks) was streamed from edge gateways to a fog node and then to a cloud cluster. The same workload was also processed centrally for baseline comparison.

Results & Findings

Key takeaways

• Shifting preprocessing and case correlation to the edge/fog slashes bandwidth and latency while keeping the discovered process models just as accurate.
• The decision framework reacts within seconds to load spikes, automatically re‑balancing work without manual intervention.

Practical Implications

• Real‑time process monitoring – Operators can get near‑instant alerts (e.g., bottleneck detection) without waiting for batch uploads to a central server.
• Data‑privacy by design – Sensitive manufacturing logs never leave the premises; only anonymized, high‑level traces are sent to the cloud, easing GDPR/CCPA compliance.
• Cost savings – Reduced upstream traffic translates into lower WAN bills, and edge‑level compute can be performed on existing IoT gateways, avoiding expensive cloud compute spikes.
• Scalable IIoT pipelines – The framework can be embedded into existing edge‑orchestrators (K3s, Azure IoT Edge, AWS Greengrass) as a plug‑in, letting developers adopt decentralized mining with minimal code changes.
• Plug‑and‑play for developers – The micro‑service APIs follow the OpenAPI spec; developers can swap in their own discovery algorithms or custom privacy filters without touching the orchestration logic.

Limitations & Future Work

• Resource heterogeneity – The current prototype assumes relatively uniform edge hardware; extreme low‑power nodes (e.g., BLE beacons) may still need to offload everything.
• Static thresholds – Decision criteria are configured manually; learning‑based threshold adaptation is left for future research.
• Security of orchestration – While data privacy is addressed, the control channel that moves tasks between layers could be a target for attacks; hardened authentication mechanisms are planned.
• Broader domain validation – The study focuses on a single logistics use‑case; upcoming work will test ContinuumConductor in manufacturing, smart‑grid, and healthcare pipelines.

ContinuumConductor opens the door for developers to embed process‑mining intelligence directly into the fabric of edge‑cloud systems, turning raw IoT streams into actionable process insights without the latency and privacy penalties of a monolithic cloud‑only approach.

Authors

• Hendrik Reiter
• Janick Edinger
• Martin Kabierski
• Agnes Koschmider
• Olaf Landsiedel
• Arvid Lepsien
• Xixi Lu
• Andrea Marrella
• Estefania Serral
• Stefan Schulte
• Florian Tschorsch
• Matthias Weidlich
• Wilhelm Hasselbring

Paper Information

• arXiv ID: 2512.07280v1
• Categories: cs.DC
• Published: December 8, 2025
• PDF: Download PDF