Operational Anomaly Detection Using Process-State Device Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing motor current signature analysis and operational deviation detection methods in industrial automation systems require extensive training times, often weeks or months, leading to process disruptions and increased costs, and rely on control systems for process state identification, which can be unreliable in continuous processing environments.
Innovation Solution
Implement a container-based machine learning system that uses acquired data to identify process states without control system input, enabling faster training and more accurate deviation detection by deploying containers closer to data sources, allowing for continuous and discrete/batch processing without disrupting operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning training methods are used to train device models on actual operating conditions, then the accuracy of operational deviation detection is improved, but the training time increases to weeks or months and causes process disruptions
Solution Approach 1:
The system performs preliminary training actions by collecting and storing operational data during normal operations without interrupting the process. Device models are trained in advance using historical data accumulated from multiple sources, allowing the training to be completed before deployment without causing process disruptions.
Solution Approach 2:
The system creates copies of operational data from multiple sources (control systems, sensors, maintenance records) to train device models. Instead of training on live operational data which would require process interruption, the system uses replicated historical datasets that can be processed offline, maintaining training accuracy while avoiding process disruption.
2Measurement precision
If traditional machine learning training methods are used with process interruptions, then the device models can be trained on actual operating conditions, but the deployment time increases due to waiting for suitable disruption windows
Solution Approach 1:
The system maintains continuous operation of industrial processes while performing data collection and model training in parallel. Multiple data sources continue to operate uninterrupted, feeding data to the training system which processes information continuously without requiring process stoppages or waiting for deployment windows.
Solution Approach 2:
The system introduces an intermediary data collection and processing layer that sits between operational sources and training requirements. This intermediary continuously accumulates and preprocesses data from control systems and sensors, allowing training to proceed on prepared datasets without direct interruption to operational processes.
3Ease of operation
If control systems are used to identify process states for device model training, then the process state information is available, but the reliability decreases in continuous processing environments where control systems may not accurately reflect actual asset states
Solution Approach 1:
The system uses multiple data sources serving multiple functions: control systems provide operational commands, sensors provide actual physical measurements, and maintenance records provide contextual information. By cross-referencing these diverse sources, the system achieves reliable process state identification that works across both discrete and continuous processing environments.
Solution Approach 2:
The system implements feedback mechanisms where sensor data from actual asset operation continuously validates and corrects process state identification. When sensor measurements indicate a state different from control system reports, the system uses this feedback to adjust its state identification, improving reliability in continuous processing where control systems may not reflect actual conditions.
Data Source
AI summary
Systems and methods described herein may involve monitoring an asset based on multiple device models representing the asset as operated in different process states. The systems and methods may involve receiving acquired data corresponding to a current operation of the asset and identifying a device model of the multiple device models based on the acquired data. The device model may correspond to a process state of the different process states, an operational parameter that the asset is operated in, and a training status indication.