SOM-Based Process Operation Prediction for Nonlinear Batch Installations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting the operation of process engineering installations are inefficient due to the need for expensive modeling of complex non-linear dynamic processes, with limited reliability and applicability, especially in dynamic and batch processes, where simulation-based predictions are rarely realized and data-based identification of non-linear models is challenging.
Innovation Solution
A data-based method using self-organizing maps (SOMs) to predict the operation of technical installations by learning from historical data, eliminating the need for rigorous modeling, and allowing for the representation and prediction of both continuous and batch processes, including those with non-linear behaviors, through unsupervised learning and visualization of high-dimensional data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If simulation-based prediction methods are used, then prediction reliability is improved, but device complexity and modeling effort increase significantly
Solution Approach 1:
The patent replaces complex mechanical modeling and simulation systems with a data-driven neural network approach. Instead of building detailed process models, automation models, and control models, the system learns directly from historical process data, substituting the traditional modeling mechanism with a learning-based mechanism that achieves similar predictive reliability without the complexity overhead
Solution Approach 2:
The patent creates a virtual copy of the process behavior through neural network training on historical data. The network learns to replicate the process dynamics, automation responses, and control actions without explicitly modeling each component, providing a simplified copy that captures essential behavior for prediction purposes
2Ease of manufacture
If data-based identification of non-linear dynamic models is attempted, then modeling effort is reduced, but prediction reliability deteriorates due to inability to capture complex non-linear behaviors
Solution Approach 1:
The patent transforms the approach from parameter-based modeling to pattern-based learning. Instead of identifying model parameters for non-linear differential equations, the system learns directly from input-output patterns in historical data, allowing it to capture complex non-linear behaviors through data-driven relationships rather than explicit mathematical models
Solution Approach 2:
The patent implements a dynamic learning system where the neural network adapts to capture time-varying process behaviors. The system learns dynamic relationships between process variables, automation responses, and control actions through sequential data processing, enabling reliable prediction of non-linear dynamic processes without requiring explicit dynamic models
3Ease of operation
If linear dynamic models with individual non-linearities are used, then ease of operation is improved, but adaptability to complex non-linear processes deteriorates
Solution Approach 1:
The patent creates a universal prediction system using neural networks that can handle multiple types of processes (continuous, batch, complex non-linear) through a single unified approach. The system is not limited to specific process types or linear assumptions, providing versatile applicability across different manufacturing scenarios while maintaining ease of operation through automated learning
Data Source
AI summary
Method and system for predicting the operation of a technical installation in which a process-engineering process having a process step runs, wherein datasets are acquired and stored in a memory and, during a learning phase, a self-organizing map is learned, symptom threshold values and associated tolerances are ascertained and stored for each neuron per SOM and all permissible temporal profiles of victor neurons are ascertained and stored per process step or batch for all timestamps, where a time starting from which operation of the technical installation should be predicted is determined during an evaluation phase, a victor neuron profile for the overall process step or batch is ascertained using current datasets of a process step via learned victor neuron profiles, and values, stored in neurons of this victor neuron profile, of the process variables are displayed after the previously determined time as a predicted profile.