Industrial Process State Analysis for Heterogeneous Time Series
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The heterogeneity of time series data from industrial processes complicates the analysis and training of machine learning models, as different operating states require distinct models and assumptions, leading to poor performance when data from one state is used to train models intended for another.
Innovation Solution
A computer-implemented method for identifying and analyzing operating states of industrial processes using time series data, allowing for the determination of specific operating states and training of models tailored to these states, which improves prediction accuracy and anomaly detection by segmenting data into relevant operating states and using probabilistic transitions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning models are trained on heterogeneous time series data from different operating states, then the model can handle more diverse process conditions, but the prediction accuracy deteriorates because data from one state is used to train models intended for another state
Solution Approach 1:
The patent segments the heterogeneous time series data into distinct subsets based on identified operating states. Each segment contains data from a specific operating state, allowing separate machine learning models to be trained on each segment. This segmentation resolves the contradiction by enabling the model to maintain high prediction accuracy for each specific state while collectively handling diverse process conditions through multiple specialized models.
2Device complexity
If a single machine learning model is used to analyze all operating states, then the device complexity is reduced, but the analysis accuracy deteriorates due to the heterogeneous nature of time series data from different states
Solution Approach 1:
The patent implements a dynamic model selection mechanism that automatically identifies the current operating state and selects the appropriate specialized model for analysis. This dynamic approach allows the system to adapt the model structure to the specific operating conditions, maintaining high analysis accuracy while managing complexity through intelligent routing rather than requiring a single overly complex universal model.
3Quantity of substance
If time series data from all operating states is used for training, then the training data quantity is maximized, but the model fitness to specific use cases deteriorates
Solution Approach 1:
The patent segments the training data into state-specific subsets based on operating state identification. Each specialized model is trained exclusively on data from its corresponding operating state, ensuring high fitness to the intended use case. The system maximizes the utility of available data by creating multiple focused training sets rather than mixing all data together, thereby maintaining both adequate data quantity for each model and high reliability for specific applications.
Data Source
Figure 1
Figure 2
Figure 3a~3b
AI summary
A computer-implemented method (100) for identifying and/or analyzing operating states (2) of an industrial process (1) that is being executed on an industrial plant, comprising the steps of: • obtaining (110) at least one time series (3a) of measurement values of at least one process variable (3) of the industrial process (1); • determining (120), from the at least one time series (3a), using a given classifying logic (4), for at least one point in time in the time series (3a), at least one operating state (2) of the industrial process (1); and • from one or more of: ∘ at least one statistical quantity (2a) computed over multiple operating states (2); ∘ at least one sequence (2b) of operating states (2); ∘ the duration (2c) for which at least one operating state (2) lasts; and ∘ a combination of at least one operating state (2) and measurement values from the time series (3a) that relate to this operating state (2), determining (130) one or more of: ∘ at least one quantity of interest (5) that further characterizes the operating state (2) of the industrial process (1); ∘ a fitness (6a) of at least one given model (6) for explaining the time series (3a) of measurement values; and ∘ a fitness (3b) of at least one measurement value of the time series (3a) as a training example for the training of at least one machine learning model (1), and/or ∘ training (140) a model (6) that describes the behavior of the industrial process (1) in the operating state (2) determined by the classifying logic (4).