Control Loop Abnormality Detection for Valve Stiction Using CNN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Control valves in industrial facilities suffer from non-linearities such as stiction, which lead to oscillations, product quality variations, and equipment instability, making it difficult for existing methods to accurately detect and address valve stiction in a timely and efficient manner.
Innovation Solution
A convolutional neural network (CNN) based method is employed to detect valve stiction and other control loop abnormalities by analyzing time series data of Process Value, Set Point, and Controller Output, allowing for automated, non-invasive, and rapid identification without the need for preprocessing, and can be trained and improved using feedback from control engineers in a cloud-based environment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual analysis of PV/SP/OP plots by control engineers is used to detect valve stiction, then detection accuracy can be maintained at human expert level, but the detection process is time-consuming and cannot handle large volumes of loops efficiently
Solution Approach 1:
The patent replaces the mechanical human visual inspection process with an automated computational system using machine learning algorithms. The system processes PV/SP/OP time series data through trained models that automatically identify stiction patterns, eliminating the need for manual plot analysis while maintaining or improving detection accuracy and enabling high-throughput processing of thousands of loops.
Solution Approach 2:
The patent creates a virtual copy of the expert control engineer's detection capability through machine learning models trained on labeled data. The trained model replicates and codifies the pattern recognition skills of human experts, allowing automated detection that matches or exceeds human performance while dramatically increasing processing speed and scalability.
2Productivity
If automated detection methods are implemented to increase processing speed, then productivity improves, but detection accuracy and reliability may deteriorate compared to expert manual analysis
Solution Approach 1:
The patent performs preliminary training of machine learning models using extensively labeled training data that captures various stiction patterns and normal operations. This pre-training phase allows the automated system to learn from diverse examples before deployment, ensuring high detection accuracy is achieved beforehand, enabling the system to maintain expert-level performance while providing automated high-speed detection.
Solution Approach 2:
The patent implements feedback mechanisms where the automated detection system's performance is continuously evaluated and refined. The model can be retrained with additional labeled data, and its predictions can be validated against expert reviews, creating a feedback loop that improves accuracy over time while maintaining high productivity through automated processing.
3Reliability
If complex preprocessing and feature extraction are applied to improve detection robustness, then measurement precision may improve, but device complexity and processing time increase
Solution Approach 1:
The patent extracts and removes the complex preprocessing and feature extraction steps from the detection pipeline. By using raw or minimally processed PV/SP/OP time series data directly as input to the machine learning models, the system achieves robust detection without the complexity of traditional signal processing stages, simplifying the overall architecture while maintaining reliability.
4Device complexity
If traditional detection methods are used to maintain simple system architecture, then device complexity remains low, but the ability to handle noise and distinguish different oscillation causes deteriorates
Solution Approach 1:
The patent replaces traditional simple detection algorithms with sophisticated machine learning models that inherently possess noise filtering and pattern recognition capabilities. These models can distinguish between stiction-induced oscillations and other types of loop behavior by learning from training data, providing robust noise handling while maintaining a relatively simple deployed system architecture that requires only data input and model inference.
Data Source
AI summary
A trained machine learning algorithm processes time series production data. The time series production data are representative of a control process within a facility control loop. The machine learning training algorithm is trained using positive training data that are representative of a normal operation of components within the facility control loop and negative training data that are representative of an abnormal operation of components within the facility control loop. Output of the trained machine learning algorithm identifies abnormalities in the facility control loop.