Electric Machine Temperature Estimation With Stable Hybrid Neural Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining the temperature of electric machine components, such as the rotor and stator, face challenges in stability and accuracy, particularly in hybrid vehicles and electric vehicles, where physical sensors are costly and unreliable.
Innovation Solution
The development of hybrid neural network models that incorporate physical laws, specifically using neural ordinary differential equations (NODE) to define the right-hand side of ordinary differential equations, allowing for the integration of neural networks and mechanistic equations to create stable and accurate temperature estimation models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If physical temperature sensors are used to measure rotor temperature, then measurement accuracy is improved, but manufacturing cost and installation complexity increase
Solution Approach 1:
The patent creates a virtual copy of the temperature sensor by training a neural network model to replicate the behavior of physical temperature sensors. The model learns from training data containing operational parameters and corresponding sensor measurements, then predicts temperatures without requiring physical sensors in the rotor, thereby eliminating installation complexity while maintaining measurement accuracy
Solution Approach 2:
The patent replaces the mechanical/physical temperature sensing system with a computational model. Instead of using physical sensors that require installation in difficult-to-reach rotor locations, the system uses a neural network that processes easily accessible operational data (currents, voltages, speeds) to infer temperatures, substituting a mechanical measurement system with an information-processing system
2Device complexity
If blackbox neural network models are used for temperature estimation, then device complexity is reduced, but model stability and reliability deteriorate
Solution Approach 1:
The patent implements feedback by continuously monitoring operational parameters and feeding them into the neural network model in real-time. The model receives ongoing input data (currents, voltages, speeds) and produces continuous temperature estimates, allowing the system to adapt to changing operating conditions while maintaining stable and reliable predictions throughout the operational lifecycle
Solution Approach 2:
The patent performs preliminary action by training the neural network model offline before deployment. During the training phase, the model learns from extensive datasets containing operational parameters and corresponding temperature measurements, preparing it to make accurate and stable predictions during actual operation without requiring complex runtime adjustments or interventions
3Reliability
If complex physical models are used to ensure stability, then model reliability is improved, but device complexity and computational requirements increase
Solution Approach 1:
The patent applies parameter changes by transforming the complex physical model into a neural network representation with learnable parameters. Instead of using fixed complex physical equations that are difficult to solve in real-time, the system uses a neural network with parameters optimized during training to capture the essential thermal dynamics, achieving both stability and computational efficiency
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
The proposed solution ensures input-to-state stability (ISS), preventing physically implausible behaviors, and reduces model complexity, making it suitable for embedded systems, while providing reliable and accurate temperature determination for electric machine components.
Implementation Method 1
The first principle of thermodynamics is applied, a general model can be defined as follows: dTstator/dt = Ploss + gc(Tc - Ts) + gr(Tr - Ts) wherein Ploss are power losses in the stator, and gc and gr are the thermal conductivities from the stator to the coolant (gc) or to the rotor (gr)
Implementation Method 2
It is thereby assumed that the stator interacts with the rotor and the coolant only through convection. Convection with other bodies can be considered by adding additional convection terms.
Data Source
AI summary
A method for determining a temperature of a component of an electric machine includes providing a plurality of trained neural networks, providing a current measurement of a plurality of operational parameters of the electric machine, and processing the operational parameters using the neural networks. The method further includes integrating output parameters issued by the neural networks over a time period, and issuing the integrated parameters as a temperature of the component.