Hybrid Neural Process Modeling for Transient Industrial Dynamics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data-driven modeling techniques, such as LSTM and ResNet, struggle to efficiently capture the dynamics of industrial processes due to structural limitations, and conventional static neural networks fail to represent both transient and steady-state behaviors.
Innovation Solution
A hybrid neural network model incorporating a first-principle model with ordinary differential equations is used to predict industrial process outputs, combining a memoryless nonlinear block and a dynamic model to account for controller inputs and disturbances, enabling accurate prediction of process outputs over time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional static neural networks are used to model industrial processes, then the model structure is simple, but the model cannot represent transient (time-dependent) responses and only describes steady-state behavior
Solution Approach 1:
The patent transforms the static neural network into a dynamic model by incorporating time-dependent ordinary differential equations that describe the rate of change of process variables. This allows the model to capture transient responses while maintaining the neural network's ability to learn from data, effectively making the system dynamic rather than static.
Solution Approach 2:
The patent creates a hybrid model that combines the strengths of data-driven neural networks with physics-based ordinary differential equations. This composite approach integrates the adaptive learning capability of neural networks with the temporal dynamics of ODEs, resulting in a model that can represent both steady-state and transient behaviors accurately.
2Reliability
If LSTM and ResNet are used to add dynamics to neural networks, then the model can capture transient behavior, but the memory usage is high and dynamics are not captured efficiently due to structural limitations
Solution Approach 1:
The patent extracts the essential dynamic behavior from complex LSTM/ResNet structures by using ordinary differential equations that directly model the rate of change of process variables. This extraction eliminates the need for bulky memory mechanisms while retaining the ability to capture transient responses efficiently.
Solution Approach 2:
The patent replaces the mechanical memory structures of LSTM (gates, cell states) and ResNet (residual connections) with a more efficient mathematical representation using ordinary differential equations. This substitution achieves the same dynamic modeling goal with significantly reduced computational overhead and memory usage.
3Adaptability or versatility
If data-driven modeling techniques are used, then the model can be updated with new data, but the model may not accurately represent the underlying physical processes and controller interactions
Solution Approach 1:
The patent merges data-driven neural network components with physics-based ordinary differential equations in a hybrid model. This combination allows the model to be updated with new data through neural network training while simultaneously maintaining adherence to physical process constraints and dynamics described by the ODEs, thus achieving both adaptability and physical fidelity.
Solution Approach 2:
The patent incorporates feedback mechanisms where the hybrid model's predictions are continuously refined by comparing with actual process data. The neural network components learn from measurement errors and update their parameters, while the ODE components ensure that updates remain consistent with physical process behavior, creating a feedback loop that improves both adaptability and reliability.
Data Source
AI summary
A method and system for modelling industrial processes, including closed loop feedback processes. The system includes a sensor for measuring an input for the industrial process; and a processor configured to receive a measurement of the input from the sensor at an input time. The processor is also configured to implement a hybrid neural network model to output a derivative of the output at the input time, wherein the neural network model incorporates at least one neural network block and a first-principle block incorporating dynamic model having an ordinary differential equation defining the rate of change over time of the output as a function of the or each associated input; to input the derivative to an ordinary differential equation solver to predict the output at a subsequent time; and to output the prediction of the output at the subsequent time using the measured input.