Industrial Fault Detection With Runtime-Loaded Device ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning (ML) model deployment solutions for industrial devices face inefficiencies and high costs when scaling up, particularly due to the need for individually training and managing thousands of ML models, leading to excessive storage and computational overhead.
Innovation Solution
A flexible multi-model machine learning deployment approach where trained ML models are stored outside docker images and executed at runtime, with a placeholder script in the image, allowing for efficient management and re-training of individual models without re-creating the docker image, and utilizing a centralized platform like ABB Ability ™< to manage and orchestrate these models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If individually trained ML models are stored inside docker images for each device, then model specificity and portability are maintained, but storage costs and computational overhead increase significantly when scaling to thousands of devices
Solution Approach 1:
The trained ML model weights are extracted from individual docker images and stored centrally in a database. Each docker image contains only a placeholder script that references the central storage location. This separates the model artifacts from the execution environment, eliminating redundant storage while maintaining model specificity through centralized management.
Solution Approach 2:
A single docker image template serves multiple devices by containing a placeholder script that dynamically loads different models from central storage. This universal image structure can be deployed to any device, with the specific model for each device loaded at runtime from the centralized model repository, eliminating the need for device-specific docker image creation.
2Adaptability or versatility
If thousands of individual docker images are created and managed for each device's ML model, then model deployment flexibility is maintained, but device complexity and management overhead increase
Solution Approach 1:
A single universal docker image template with a placeholder script is deployed to all devices. The placeholder script dynamically loads the appropriate model from central storage based on device identity. This eliminates the need to create and manage thousands of individual docker images while maintaining the ability to deploy models to any device.
Solution Approach 2:
A centralized model management platform acts as an intermediary between the universal docker images and individual device models. This platform stores models centrally, manages model versions, and provides runtime model loading services, simplifying the deployment architecture and reducing management complexity.
3Reliability
If ML models are re-trained and re-deployed by re-creating docker images, then model updates are achieved, but loss of time and productivity decrease due to repeated image creation processes
Solution Approach 1:
The model re-training process is decoupled from the docker image creation process. Only the model weights are re-trained and updated in central storage, while the universal docker image template remains unchanged. This eliminates the time-consuming step of re-creating docker images for every model update.
Solution Approach 2:
The universal docker image template is prepared in advance with a placeholder script that is designed to load models from central storage. This preliminary setup enables rapid model updates without requiring repeated image creation, as the infrastructure is already in place to support dynamic model loading.
Data Source
Figure 1
Figure 2
AI summary
The invention relates to an industrial data monitoring apparatus. The apparatus comprises an input unit; a processing unit; and an output unit. The input unit is configured to receive industrial monitor data from a plurality of devices. The processing unit is configured to access a machine learning environment. The processing unit is configured to access a plurality of trained machine learning models, wherein the plurality of trained machine learning models each correspond to a different device of the plurality of devices, and wherein the plurality of trained machine learning models are separate to the machine learning environment. Upon receipt of industrial monitor data from a particular device of the plurality of devices, the apparatus is configured to implement the machine learning environment and run the trained machine learning model for the particular device to analyse the industrial monitor data received from the particular device of the plurality of devices. The output unit is configured to output an analysis result associated with the analysis of industrial monitor data received from the particular device of the plurality of devices.