Battery State Estimation Using Impedance Spectroscopy and Neural Networks

The patent applies parameter changes by training the neural network with impedance spectra measured at different temperatures to create temperature-dependent reference data. This allows the system to account for temperature variations in impedance characteristics, thereby maintaining estimation accuracy without requiring additional temperature sensors. The neural network learns to distinguish between impedance changes caused by temperature versus those caused by cell variance or state of charge.

The patent replaces the mechanical/physical temperature sensing system with an electrical measurement system based on impedance spectroscopy and neural network processing. Instead of using physical temperature sensors that directly measure thermal state, the system uses electrical impedance measurements combined with computational algorithms to infer temperature, thereby reducing hardware complexity while maintaining measurement capability.

The patent makes the impedance measurement system multi-functional by enabling it to perform multiple tasks: determining state of charge, estimating temperature, and monitoring cell health. The same impedance measurement chip and neural network infrastructure used for SoC estimation are extended to also provide temperature estimation, eliminating the need for separate temperature sensing hardware and reducing overall system complexity.

The system uses the battery cells' own electrical impedance characteristics as the measurement basis for temperature estimation. Instead of requiring external temperature sensors to measure the cells, the cells themselves provide the measurement signal through their impedance response to small test currents. The neural network processes these self-generated signals to extract temperature information.

The patent applies preliminary action by performing offline training of the neural network using impedance spectra from multiple battery cells measured at known temperatures. This pre-processing step creates temperature-dependent reference patterns that capture cell-to-cell variations. During operation, the pre-trained network can quickly estimate temperature without complex real-time calculations, maintaining both simplicity and reliability.

The patent creates virtual copies of temperature sensor functionality through the neural network model. The network learns to replicate the temperature measurement function by processing impedance data, effectively copying the information that would otherwise require physical temperature sensors. This virtual copying approach maintains measurement reliability while avoiding the complexity of physical sensor installation in each cell.