Federated Learning Cross-Factory Model Training 2026
Federated Learning Cross-Factory Model Training
Federated learning enables geographically separated manufacturing facilities to train shared ML models without exposing raw operational data. Active filings from IBM, Intel, Google, Nokia, and Ericsson span 2019–2026 across 7 jurisdictions.
Cross-Factory FL: From Academic Prototype to Industrial Deployment
Cross-factory federated learning deploys FL architectures across multiple independent manufacturing or warehouse facilities—each operating as a distinct data silo—to train shared predictive models without centralizing raw production data. The field spans horizontal FL, vertical FL (VFL), and federated transfer learning, each addressing different inter-plant data distribution scenarios.
Supporting paradigms include hierarchical aggregation architectures using client–edge–cloud stacks, heterogeneity-aware client selection algorithms, communication compression schemes, and privacy-preserving aggregation via trusted execution environments (TEEs) or differential privacy. These components collectively address the latency, trust, and data heterogeneity challenges inherent in multi-plant deployments.
Publication dates in this dataset range from 2019 to 2026, indicating an active and still-accelerating filing cycle. The technology has moved from foundational hierarchical frameworks such as HierFAVG (2020) and FedCluster through industrialization-focused heterogeneity handling (2021–2022) to platform and governance layers (2023–2024) and cross-domain, multi-task specialization (2025–2026).
In this dataset, 13 unique assignees are identified across 7 jurisdictions. Core FL infrastructure IP is concentrated among large technology companies including IBM, Google, Intel, and Huawei in retrieved records, while application-layer filings from Dell, MOKSA.AI, Nokia, and Ericsson reflect a more distributed pattern of domain-specialist innovation.
Technology Cluster Distribution and Filing Activity Over Time
Patent filings in this dataset cluster across four primary technical areas: vertical federated learning, hierarchical edge aggregation, concurrent multi-site training, and privacy-preserving aggregation. Filing activity accelerated from 2021 onward, with 2025–2026 showing a pivot toward cross-domain and multi-task specialization.
Patent Count by Technology Cluster — Dataset Snapshot
In this dataset, privacy-preserving aggregation and VFL each account for three patent records, while hierarchical edge aggregation and concurrent multi-site training each contribute two to three records, reflecting distributed technical investment across core FL infrastructure clusters.
↗ Click bars to explorePatent Filings by Period — Cross-Factory FL Dataset Snapshot
In this dataset, filing activity was sparse in 2019–2020 with foundational concepts, accelerated substantially in 2021–2022 with industrialization-focused patents, and reached the highest concentration in 2023–2026 with platform, governance, and cross-domain specialization filings.
↗ Click bars to exploreKey Deployment Domains for Cross-Factory Federated Learning
Within this dataset, cross-factory FL applications span industrial manufacturing and production line monitoring, video-based visual inspection, automotive edge systems, and telecommunications network automation. Each domain reflects distinct technical requirements for data privacy, model heterogeneity, and edge deployment architecture.
Industrial Production Line Monitoring
Literature from 2021 demonstrates FedSVM and FedRF algorithms for failure prediction in production lines, showing FL performance is statistically comparable to centralized learning on Bosch production data. Cohort-based FL (2021) organizes factories by data distribution similarity, delivering Federated Learning as a Service via an Industrial FL Services API deployed to edge devices. Dell Products’ 2023 and 2024 US patents explicitly cover cross-organization edge FL for warehouse and factory event detection.
Industrial ManufacturingFederated Video Visual Inspection
MOKSA.AI’s 2026 filings across US, EP, and IN jurisdictions introduce a parent-child FL architecture where a parent model is trained on stratified aggregated video data from multiple clusters and child models capture cluster-specific invariant and variant features. This architecture is directly applicable to federated defect detection across factories with different visual inspection setups. The multi-jurisdiction filing strategy (US, EP, IN) signals commercial deployment intent.
Video AnalyticsAutomotive Edge FL Deployment
Real-time end-to-end FL for steering wheel angle prediction (2021 literature) demonstrates asynchronous FL in automotive production and validation contexts. Samsung Electronics’ 2023 WO and IN patents cover optimal split FL for DNN partitioning between client and edge devices, addressing resource-constrained automotive edge deployments. The split FL approach determines the optimal cut layer to balance computation and communication overhead across distributed vehicle or factory nodes.
Automotive SystemsTelecom Network Automation FL
Nokia Technologies filed FL multi-task learning patents in WO (2025), IN (2026), and GB (2025), covering systems that train multiple personalized global models simultaneously for different tasks using different FLDN subsets. Ericsson’s 2025 and 2026 WO filings address cross-domain VFL coordination between network function instances, with a coordinator NF receiving intermediate results and transmitting training parameters back to participant NFs. These telecom deployments serve as proxy environments for distributed, privacy-sensitive multi-site model training with direct parallels to cross-factory settings.
Network AutomationKey Patent Assignees in Cross-Factory Federated Learning (Retrieved Records)
In this dataset, International Business Machines Corporation holds 4 patent records — the largest single-assignee count in retrieved records — spanning cross-party XGBoost training, hierarchical framework management, and TEE-based decentralized aggregation. Intel Corporation and Nokia Technologies Oy each hold 3 records in this dataset, covering edge FL client selection and federated multi-task learning respectively.
Top Assignees by Filing Count — Cross-Factory FL (Dataset Snapshot)
↗ Click bars to exploreInternational Business Machines Corporation
IBM holds 4 patent records in this dataset spanning US (×3) and DE (×1) jurisdictions, covering filings from 2022 to 2024. Key patents include federated XGBoost training using global histogram fusion (US 2022, US 2024), a hierarchical framework builder with multi-channel parameter propagation (US 2024), and a multi-TEE decentralized aggregation system with runtime memory encryption and dynamic gradient fragment shuffling (DE 2024). IBM’s portfolio represents the broadest enterprise-grade cross-party FL IP in this dataset.
United StatesIntel Corporation
Intel holds 3 patent records in this dataset across WO (×1, 2021) and US (×2, 2023 and 2026) jurisdictions, all related to systems and methods for distributed learning for wireless edge dynamics. The patents cover edge computing nodes that process client reports, select candidate clients per training epoch, and train on coded client data — addressing straggler and communication bottleneck challenges in multi-factory deployments. The 2026 US continuation confirms Intel’s sustained long-term investment in edge FL client selection infrastructure.
United StatesFive Emerging Directions in Cross-Factory Federated Learning
The 2025–2026 filing cohort in this dataset signals a pivot from core FL infrastructure toward cross-domain adaptability, VFL interoperability standards, concurrent pipeline architectures, multi-task personalization, and video-based federated inspection. Each direction addresses a gap in current cross-factory deployment capability.
Cross-Domain Federated Incremental Learning for Dynamic Factory Networks
Northwestern Polytechnical University’s 2026 CN filing introduces a federated domain incremental learning framework for dynamic devices moving between edge environments, using K-medoid coreset replay, elastic weight consolidation, progressive knowledge distillation, and maximum mean discrepancy-based cross-domain feature alignment. This directly addresses factories onboarding to or leaving a FL federation over time — one of few retrieved records in this dataset targeting dynamic factory network membership. The framework is particularly relevant for supply chains with changing factory configurations or product mix transitions.
VFL Interoperability and Negotiation Protocol Standardization
Ericsson’s 2026 WO filing introduces explicit interoperability signaling between VFL participant nodes — exchanging model type and supported training method metadata before training commences. This standardization layer is identified as a near-term bottleneck for large-scale cross-factory VFL where factories may use heterogeneous ML stacks. Ericsson’s earlier 2025 WO filing established the coordinator NF architecture for cross-domain VFL, and the 2026 filing extends this with training negotiation protocols — signaling that telecom infrastructure vendors are positioning to own this standardization layer.
IBM vs. Intel: Enterprise FL Infrastructure Approaches Compared
Click any row to explore further.
| Dimension | IBM (International Business Machines) | Intel Corporation |
|---|---|---|
| Filing Count in Dataset | 4 patents (US ×3, DE ×1) | 3 patents (WO ×1, US ×2) |
| Filing Date Range | 2022–2024 | 2021–2026 |
| Core Technical Focus | Cross-party XGBoost training, hierarchical framework management, TEE-based decentralized aggregation | Edge computing FL client selection, coded training data, straggler mitigation |
| Privacy Mechanism | Multi-TEE aggregation with runtime memory encryption; dynamic gradient fragment shuffling across aggregators per iteration | Coded client data for communication bottleneck reduction; client selection per epoch |
| Aggregation Architecture | Hierarchical ML model management; cloning of primary model weights into secondary models; multi-channel parameter propagation | Edge computing node aggregates client reports; selects candidate clients per training epoch |
| Cross-Factory Applicability | Multi-factory tabular data via global histogram fusion; multi-plant model hierarchies; TEE trust for competing organizations | Multi-factory edge deployment addressing straggler and communication challenges in wireless edge environments |
| Key Patent Example | Trusted and Decentralized Aggregation for Federated Learning (DE, 2024) | Systems and Methods for Distributed Learning for Wireless Edge Dynamics (US, 2026) |
| Jurisdiction Strategy | US-primary with DE filing for TEE aggregation patent | WO international route (2021) with US continuation filings (2023, 2026) |
Frequently Asked Questions: Cross-Factory Federated Learning Patents
Cross-factory federated learning deploys FL architectures across multiple independent manufacturing or warehouse facilities — each operating as a distinct data silo — to train shared predictive or analytical models without centralizing raw production data. The approach spans horizontal FL (same feature space across factories), vertical FL (different feature sets, overlapping samples), and federated transfer learning for factories with dissimilar data distributions.
In this dataset of 16 retrieved patent records, IBM holds 4 records (the most of any single assignee), followed by Intel, Nokia Technologies, and MOKSA.AI with 3 records each, and Ericsson, Google, Samsung, and Dell with 2 records each. These counts reflect this dataset snapshot only and should not be interpreted as a comprehensive industry ranking.
VFL is architecturally suited to cross-factory scenarios where different plants hold different feature sets about overlapping production batches. A coordinator node receives intermediate results from participant nodes, updates a coordinator-side model, and transmits gradients back — keeping raw data siloed at each factory. Ericsson’s 2025 and 2026 WO patents cover VFL coordination between network function instances, directly applicable to multi-factory deployments where Factory A holds process parameters and Factory B holds quality inspection records.
IBM’s 2024 DE patent covers multi-TEE aggregation with runtime memory encryption. Parties partition model updates at parameter granularity and dynamically shuffle fragments across aggregators per training iteration, preventing any single aggregator from reconstructing full gradients. This mechanism addresses the trust requirement when competing manufacturers must share a federated infrastructure without exposing proprietary production data.
Nokia’s 2025–2026 filings in WO, IN, and GB cover FL systems that train multiple personalized global models simultaneously for different tasks using different FLDN subsets. For factories with varying production tasks — such as Plant A running welding inspection and Plant B running assembly defect detection — this enables a single federated infrastructure to serve differentiated model objectives without retraining from scratch.
According to this dataset, incremental and domain-adaptive FL for factories joining or leaving a federation over time is identified as a critical unsolved problem. Northwestern Polytechnical University’s 2026 CN filing is one of few retrieved records addressing this, using K-medoid coreset replay, elastic weight consolidation, progressive knowledge distillation, and maximum mean discrepancy-based cross-domain feature alignment. This gap represents a high-value R&D investment area for organizations with changing factory configurations.
Data and insights on this page are based on a limited patent and literature dataset and are for reference only. Figures may not represent the complete technology landscape.