KPI Time-Series Pattern Monitoring for Real-Time Plant Event Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional plant models are inadequate for monitoring time series Key Performance Indicators (KPIs) to detect and diagnose undesirable operating events in industrial processes due to their complexity, cost, and limited capability in handling extreme operational conditions, leading to inefficient maintenance and false alerts.
Innovation Solution
A modeling approach that builds event pattern models with configurable event signatures for KPIs, allowing users to define patterns through known events, sister units, or pattern libraries, and applies unsupervised or supervised pattern discovery methods to identify and deploy models online, using the Aspen Tech Distance (ATD) for similarity scoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional first-principle models are used for detailed dynamic predictions, then model accuracy is improved, but development complexity and cost increase significantly
Solution Approach 1:
The patent creates simplified copies of complex first-principle models by using historical data to train empirical models that replicate the behavior of detailed dynamic models without requiring the same level of complexity. These empirical models serve as lightweight substitutes that maintain prediction accuracy for monitoring purposes while dramatically reducing development cost and complexity.
Solution Approach 2:
The patent transforms the modeling approach by changing from fixed first-principle parameters to adaptive empirical parameters that are calibrated using historical process data. This allows the model to capture complex dynamic behavior through data-driven parameter relationships rather than requiring detailed physical understanding of all processes, thereby reducing development complexity while maintaining prediction capability.
2Reliability
If traditional empirical models are trained with normal operational data, then model calibration is improved, but detection of extreme operational events deteriorates
Solution Approach 1:
The patent implements dynamic model calibration that adapts to different operational conditions. The system automatically adjusts model parameters and selection based on the current operating regime, allowing the same model framework to be calibrated for normal operations while simultaneously maintaining capability to detect extreme events through appropriate parameter selection and ensemble methods.
Solution Approach 2:
The patent creates a universal monitoring framework that can handle both normal operational calibration and extreme event detection within a single system. By using multiple empirical models trained on different data subsets and combining their predictions, the system achieves both reliable calibration for normal operations and adaptability for detecting rare extreme events that fall outside the normal training data distribution.
3Reliability
If statistical models are used for anomaly detection, then detection capability is improved, but fault identification capability deteriorates
Solution Approach 1:
The patent segments the monitoring task into two distinct functions: anomaly detection and fault identification. The empirical model first detects whether an anomaly exists by comparing predicted versus actual values, and when an anomaly is detected, the system then segments the analysis to identify specific fault conditions by examining which model components deviated most from predictions, thereby preserving fault identification information that would otherwise be lost.
Solution Approach 2:
The patent implements feedback mechanisms where the results of anomaly detection trigger further diagnostic analysis. When the empirical model detects an anomaly through residual analysis, the system provides feedback to initiate detailed fault identification procedures, including examining individual model predictions, analyzing pattern matches against known fault signatures, and prioritizing potential root causes based on the nature and magnitude of deviations.
4Reliability
If traditional multivariate statistical approaches are used, then monitoring capability is improved, but periodic recalibration requirement increases
Solution Approach 1:
The patent performs preliminary actions by pre-training multiple empirical models on comprehensive historical data covering normal and extreme conditions before deployment. During online monitoring, the system selects and applies the most appropriate pre-trained model based on current operating conditions, eliminating the need for periodic recalibration while maintaining monitoring capability across varying operational regimes.
Solution Approach 2:
The patent discards the requirement for periodic recalibration by using a library of pre-trained empirical models that cover different operational scenarios. When operating conditions change, the system recovers the appropriate pre-trained model from the library rather than recalibrating an existing model, thereby maintaining monitoring capability without the time loss associated with recalibration cycles.
5Measurement precision
If Fuzzy-reasoning approaches are used for fault detection, then detection accuracy is improved, but system complexity increases
Solution Approach 1:
The patent replaces complex mechanical-like fuzzy-reasoning systems with data-driven empirical models that directly compute fault detection results. Instead of using multiple fuzzy rules, event signature disassembly, and similarity matrix calculations, the system uses trained empirical models that automatically learn the complex relationships between process variables and fault conditions from historical data, achieving the same detection accuracy with significantly reduced system complexity.
Data Source
AI summary
Embodiments are directed to computer methods and systems that build and deploy a pattern model to detect an operating event in an online plant process. To build the pattern model, the methods and systems define a signature of the operating event, such that the defined signature contains a time series pattern for a KPI associated with the operating event. The methods and systems deploy the pattern model to automatically monitor, during online execution of the plant process, trends in movement of the KPI as a time series. The methods and systems determine, in real-time, a distance score between a range of the monitored time series and the time series pattern contained in the defined signature. The methods and systems automatically detect the operating event in the online industrial process based on the determined distance score, and alter parameters of the process (e.g., valves, actuators, etc.) to prevent the operating event.