Decentralized Fleet Analytics Through Plant-Level Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Centralized fleet analytics approaches incur high network traffic and storage costs, sensitive data sharing issues, and unreliable connections, particularly in fleets with distinct entity ownership.
Innovation Solution
A decentralized approach where each plant builds a machine learning model from local data streams, transmitting the model to a cloud server for ensemble learning, which updates and distributes back to plants, reducing data transmission and preserving privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If all data from individual plants is sent to a centralized cloud server for analysis, then comprehensive fleet-wide analytics can be achieved, but network traffic and storage costs increase significantly
Solution Approach 1:
The patent divides the centralized analytics system into distributed edge computing nodes at each plant. Each plant builds and trains its own machine learning model locally using its data, then only the model parameters (not the raw data) are transmitted to the cloud. This segmentation reduces network traffic from voluminous raw data to compact model parameters, while still achieving fleet-wide analytics through aggregated model updates.
Solution Approach 2:
The patent extracts only the essential information (machine learning model parameters) from the raw data at each plant, leaving the voluminous raw data local to each plant. This extraction approach allows the cloud to receive and aggregate model parameters from multiple plants without transmitting the actual operational data, significantly reducing storage requirements and network bandwidth consumption.
2Productivity
If voluminous data is transmitted to the cloud server, then comprehensive analytics can be performed, but storage costs and network overhead increase
Solution Approach 1:
The patent performs preliminary data processing and model training at the edge devices before cloud transmission. Each plant pre-trains its machine learning model using local historical data, then only the trained model parameters are sent to the cloud. This preliminary action at the edge eliminates the need to transmit raw operational data, reducing data transmission volume while maintaining analytics capability.
Solution Approach 2:
The patent uses machine learning model parameters as a compact representation (copy) of the information contained in voluminous raw data. Instead of transmitting actual operational data, the system transmits compressed model parameters that capture the essential patterns and insights, achieving the same analytics goal with minimal data transmission.
3Adaptability or versatility
If plants share data with centralized cloud servers, then fleet-wide machine learning models can be trained, but sensitive data privacy concerns arise
Solution Approach 1:
The patent introduces machine learning model parameters as an intermediary between raw plant data and centralized fleet analytics. Instead of directly sharing sensitive operational data with the cloud, each plant transforms its data into model parameters locally, which serve as an intermediate representation that captures fleet-wide patterns without exposing proprietary operational details. This intermediary approach enables fleet-wide training while protecting data privacy.
Solution Approach 2:
The patent extracts only the necessary analytical insights in the form of machine learning model parameters from sensitive raw data, leaving the actual sensitive operational data local to each plant. This extraction separates the valuable analytical information from the sensitive context, allowing fleet-wide model training without direct exposure of proprietary data.
4Ease of operation
If unreliable network connections are used to transmit data to the cloud, then connectivity flexibility is maintained, but data transmission reliability decreases
Solution Approach 1:
The patent treats machine learning model parameters as lightweight, easily retransmittable information units compared to voluminous raw data. These compact model parameters can be quickly retransmitted if transmission fails, making the system more resilient to unreliable network connections. The small size and low cost of retransmitting model parameters compensate for potential transmission failures.
Solution Approach 2:
The patent performs preliminary model training and validation at the edge before transmission, ensuring that only validated model parameters are sent to the cloud. This preliminary validation reduces the need for retransmission and improves overall transmission reliability, while the compact nature of model parameters makes retry transmissions more feasible over unreliable connections.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Technologies for decentralized fleet analytics are disclosed. In at least one embodiment, a local cloud service at a plant site builds a first machine learning model of one or more first streams of data associated with the plant site. The local cloud service sends the first machine learning model to a cloud service connected to the plant site and other plant sites. The local cloud service receives a second machine learning model from the cloud service. The second machine learning model is trained as a function of the first machine learning model and one or more machine learning models built by the other plant sites. The local cloud service updates the first machine learning model based on the second machine learning model.