Parameter Efficient Multi-Class Intelligent Scheduling for Multimodal Online Distributed Industrial Anomaly Detection
The MODIAD framework addresses the gap in industrial anomaly detection for distributed, continuously generated multimodal data. It proposes a complete workflow involving a Multi-class Intelligent Scheduling (MIS) algorithm to coordinate model updates and a Resource Efficient Class-Wise Low Rank Adaptation (REC-LoRA) strategy to minimize computational overhead. This enables efficient, collaborative detection on edge devices, demonstrating superior performance on standard datasets under realistic
Analysis
Background
Traditional industrial anomaly detection methods are designed for centralized, offline settings. This is misaligned with real-world environments where data is generated continuously by distributed, heterogeneous sensors (e.g., 2D images and 3D point clouds). The rise of edge intelligence creates an opportunity for distributed, collaborative learning at the data source, but requires new algorithms to manage this process efficiently.
Key Points
The MODIAD framework tackles two core challenges: resource-constrained coordination and training efficiency.
Multi-class Intelligent Scheduling (MIS): This is formulated as an optimization problem to decide which anomaly class models to update on which edge devices at each time step. The goal is to balance two competing factors:
- Data Sufficiency: Prioritizing classes with enough new data to ensure a meaningful model update.
- Update Frequency: Preventing class models from becoming stale by ensuring they are updated regularly.
The proposed Sequential Marginal Gain Greedy (SMG) algorithm efficiently solves this MIS problem, enabling practical multi-class training under resource limits.
Resource Efficient Class-Wise Low Rank Adaptation (REC-LoRA): To reduce the cost of training, REC-LoRA applies parameter-efficient fine-tuning. Instead of updating all model parameters, it introduces and trains a small set of low-rank matrices for each anomaly class. This drastically cuts the computational and communication overhead during distributed updates while maintaining detection performance.
Validation: The system's effectiveness is proven on two multimodal datasets, MVTec 3D-AD (industrial parts) and Eyecandies (packaged goods), showing superior performance and efficiency in the MODIAD scenario.
Significance
This work is significant because it bridges a critical gap between theoretical anomaly detection and practical industrial deployment. It formally defines and solves the distributed online learning problem for industrial quality control, moving beyond static benchmarks. The SMG algorithm and REC-LoRA strategy provide a concrete, efficient blueprint for implementing collaborative edge intelligence. This enables real-time, scalable monitoring of modern, data-intensive industrial systems, directly supporting advancements in automation and predictive maintenance.
Disclaimer: The above content is generated by AI and is for reference only.