Abstract
Enabled by the increasing availability of sensor data monitored from production machinery, condition monitoring and predictive maintenance methods are key pillars for an efficient and robust manufacturing production cycle in the Industrial Internet of Things (IIoT). The employment of machine learning models to detect and predict deteriorating behavior by analyzing a variety of data collected across several industrial environments shows promising results in recent works, yet also often requires transferring the sensor data to centralized servers located in the cloud. Moreover, although collaborating and sharing knowledge between industry sites yields large benefits, especially in the area of condition monitoring, it is often prohibited due to data privacy issues. To tackle this situation, we propose an Autoencoder-based Federated Learning method utilizing vibration sensor data from rotating machines, that allows for a distributed training on edge devices, located on-premise and close to the monitored machines. Preserving data privacy and at the same time exonerating possibly unreliable network connections of remote sites, our approach enables knowledge transfer across organizational boundaries, without sharing the monitored data. We conducted an evaluation utilizing two real-world datasets as well as multiple testbeds and the results indicate that our method enables a competitive performance compared to previous results, while significantly reducing the resource and network utilization.
Introduction
In the era of Industry 4.0, Condition Monitoring (CM) has emerged as a critical component for maintaining the health of industrial machinery. By identifying early signs of equipment failure, companies can transition from reactive repairs to proactive maintenance, significantly reducing costly downtime. However, the traditional approach to CM—streaming high-frequency vibration data to centralized cloud servers—is increasingly unsustainable.
This centralized paradigm faces three primary obstacles:
- Bandwidth and Latency: Industrial sites often operate in remote locations with limited connectivity (e.g., 4G or satellite). Transferring gigabytes of raw sensor data is often cost-prohibitive or technically impossible.
- Data Privacy: Raw vibration data can inadvertently leak proprietary information about production cycles and machine utilization, making organizations reluctant to share data with third-party cloud providers.
- Resource Constraints: While edge computing hardware like Raspberry Pis are becoming more common, they typically lack the memory and processing power required for standard deep learning training pipelines designed for high-end server environments.
The Proposed Framework: Federated LSTM-Autoencoders
To address these challenges, this research proposes a Federated Learning (FL) architecture utilizing Long Short-Term Memory (LSTM) Autoencoders.
The Model
The core of the detection system is an LSTM-Autoencoder (AE). Unlike supervised models that require labeled failure data, the AE is trained exclusively on "normal" operational data. It learns to compress and then reconstruct the vibration signal. When a machine begins to fail, the incoming signal deviates from the learned "normal" patterns, causing the reconstruction error to spike. This spike serves as the trigger for an anomaly detection alert.
Federated Learning Process
Federated Learning allows for model training without the need for raw data transfer. In this framework, "Training Nodes" (edge devices) perform local training on their own sensor data. Instead of sending the data to the cloud, they transmit only the model weights—the mathematical parameters of the neural network—to a central "Aggregation Node."
The cloud server aggregates these weights using the FedAvg algorithm to create a refined "global model," which is then redistributed to all participating devices. This cycle enables different industrial sites to benefit from a collective intelligence that has encountered various failure modes without any participant ever exposing their raw data.
Edge Optimization
To ensure the model can run on low-power hardware, the authors introduced a windowed single-batch training approach. Standard training often requires processing multiple preceding batches of data simultaneously, which is memory-intensive. By splitting single batches into smaller temporal windows, the model can capture necessary dependencies without exhausting the limited RAM of an edge gateway.
Experimental Evaluation
The framework was evaluated using two distinct datasets to ensure robustness:
- NASA IMS Bearing Dataset: A public benchmark containing high-frequency vibration data from bearings run until failure.
- Rotating Machine (RM) Dataset: A real-world dataset from an industry partner featuring vibration metrics from five different types of rotating machinery, including pumps and motors.
Key Results
The experimental results demonstrate that the Federated Learning approach is not only viable but superior in resource-constrained environments:
- Bandwidth Efficiency: The FL approach reduced network utilization by up to 99.22% compared to centralized data streaming. This drastic reduction makes sophisticated condition monitoring feasible for remote sites with limited connectivity.
- High Accuracy: Despite the distributed training, the FL model achieved an average F1-score of 99.4% on the industry dataset, matching the performance of centralized state-of-the-art baselines.
- Reduced Hardware Requirements: The optimized training method required only 2-3GB of RAM, a significant improvement over the 9-11GB required by centralized baselines. This allows the system to operate on standard edge devices like a Raspberry Pi.
Conclusion and Future Work
This research proves that deep learning for industrial maintenance is no longer restricted to the cloud. By combining LSTM-Autoencoders with Federated Learning, it is possible to achieve high-accuracy anomaly detection while preserving data privacy and minimizing bandwidth usage.
Future research will focus on addressing Non-IID (Independent and Identically Distributed) data, where different machines may have varying "normal" baselines due to age or environmental factors. Additionally, the team plans to explore more sophisticated aggregation algorithms beyond FedAvg and test the framework across a broader range of IIoT hardware architectures to ensure universal deployability.
Summary updated