Federated Learning for Predictive Maintenance and Anomaly Detection Using Time Series Data Distribution Shifts in...

Area of Science:

• Industrial Engineering
• Data Science
• Machine Learning

Background:

• Equipment failure significantly impacts manufacturing productivity, necessitating robust predictive maintenance strategies.
• Distributed industrial environments present challenges due to data heterogeneity across diverse equipment, complicating predictive maintenance efforts.

Purpose of the Study:

• To develop and evaluate an effective predictive maintenance framework for distributed manufacturing environments with heterogeneous equipment.
• To address data distributional shifts in time series data for improved anomaly detection and maintenance prediction.

Main Methods:

• Proposed a novel 1DCNN-Bilstm model for time series anomaly detection, combining 1D convolutional neural networks and bidirectional LSTMs for feature extraction.
• Integrated a federated learning framework with the 1DCNN-Bilstm model to handle data heterogeneity and distributional shifts in industrial settings.
• Evaluated the combined framework using a pump dataset to assess performance in anomaly detection and predictive maintenance.

Main Results:

• The proposed federated learning framework integrated with the 1DCNN-Bilstm model achieved a high test accuracy of 97.2%.
• Demonstrated the model's effectiveness in extracting features from time series data and accurately detecting anomalies.
• Validated the potential of the approach for real-world predictive maintenance applications.

Conclusions:

• The combined federated learning and 1DCNN-Bilstm approach offers a promising solution for predictive maintenance in complex industrial environments.
• The framework effectively handles data heterogeneity and distributional shifts, leading to high accuracy in anomaly detection.
• This research highlights the potential for significant improvements in manufacturing productivity and equipment reliability through advanced AI techniques.