System and method for forecasting battery state with imperfect data

Patent metadata

Field	Value
Publication number	US-12154389-B2
Application number	US-202117194802-A
Country	US
Kind code	B2
Filing date	Mar 8, 2021
Priority date	Mar 8, 2021
Publication date	Nov 26, 2024
Grant date	Nov 26, 2024

Abstract

An approach to forecasting battery health as a dynamic time-series problem as opposed to a static prediction problem is presented. Systems and methods disclosed herein forecast a trajectory to failure by predicting a path to failure as opposed to only predicting when the battery may fail. A machine-learning model is implemented that extracts unique features taken from time-series data, such as time snippets of charging data. The raw time-series data may include current voltage and temperature with complex transformations and without capturing a full cycle, which permits wider applicability to instances of varying depth of discharge (DoD).

First claim

The invention claimed is:

1. A method of forecasting a state-of-health for a vehicle battery, the method comprising:
obtaining a first battery dataset assembled from one or more vehicle battery cells;
selecting an arbitrary time window from a subset of the first battery dataset, including historical data;
selecting at least one feature from the arbitrary time window;
structuring the subset for feature extraction by a convolutional filter;
capturing an evolution of the at least one feature using the convolutional filter;
generating a state-of-health model; training the state-of-health (SOH) model by inputting the at least one feature, the evolution of the at least one feature, and a ground truth future state-of-health as variables, in which the state-of-health model comprises a gaussian-process convolutional neural network (GP-CNN) having convolutional layers, tuned to capture short-term and long-term aging effects within the arbitrary time window of measurement;
training a gaussian process regressor using the GP-CNN to generate SOH estimates to capture a relation between identified points through covariance functions, and generate forecasts with error estimates;
embedding a gaussian process regressor in a battery management system of a vehicle; and
predicting, using electrochemical measurements collected during cycling/operation of battery cells of a vehicle battery of the vehicle, a trajectory to failure for the vehicle battery based on the gaussian process regressor.

2. The method of claim 1 wherein the at least one feature includes a time-stamp relative to a reference age of a vehicle battery cell.

3. The method of claim 1 wherein the at least one feature is a raw time-series.

4. The method of claim 1 wherein the at least one feature comprises current data.

5. The method of claim 1 wherein the at least one feature comprises voltage data.

6. The method of claim 1 wherein the at least one feature comprises temperature data.

7. The method of claim 1 wherein the at least one feature comprises a derived metric.

8. The method of claim 1 wherein the at least one feature comprises a change in current during a constant-voltage hold.

9. The method of claim 1 wherein the evolution is a temporal evolution.

10. The method of claim 1 further comprising generating a recommendation from the state-of-health model to maximize battery performance.

11. The method of claim 1 wherein the one or more vehicle battery cells are cycled to end-of-life.

12. The method of claim 1 further comprising tuning the convolutional neural network using a second dataset.

13. A system for forecasting a state-of-health for a vehicle battery, comprising: one or more processors; a memory communicably coupled to the one or more processors and storing:
a behavioral forecast system including instructions that when executed by the one or more processors cause the one or more processors to generate trajectory to failure of the vehicle battery by:
obtaining a first battery dataset assembled from one or more vehicle battery cells;
selecting an arbitrary time window from a subset of the first battery dataset, including historical data;
selecting at least one feature from the arbitrary time window; structuring the subset for feature extraction by a convolutional filter;
capturing an evolution of the at least one feature using the convolutional filter;
generating a state-of-health model;
training the state-of-health (SOH) model by inputting the at least one feature, the evolution of the at least one feature, and a ground truth future state-of-health as variables, in which the state-of-health model comprises a gaussian-process convolutional neural network (GP-CNN) having convolutional layers tuned to capture short-term and long-term aging effects within the arbitrary time window of measurement;
training a gaussian process regressor using the GP-CNN to generate SOH estimates to capture a relation between identified points through covariance functions, and
generate forecasts with error estimates embedding a gaussian process regressor in a battery management system of a vehicle; and
predicting, using electrochemical measurements collected during cycling/operation of battery cells of a vehicle battery of the vehicle, a trajectory to failure for the vehicle battery based on the gaussian process regressor.

14. A battery monitoring system of a vehicle comprising:
one or more processors;
a memory communicably coupled to the one or more processors and storing:
a behavioral forecast system including instructions that when executed by the one or more processors cause the one or more processors to generate a trajectory to failure of the battery by:
obtaining a first battery dataset assembled from one or more vehicle battery cells;
selecting an arbitrary time window from a subset of the first battery dataset, including historical data;
selecting at least one feature from the arbitrary time window, the at least one feature comprising at least one of current data, voltage data, capacity data, temperature data or an engineered metric;
structuring the subset for feature extraction by a convolutional filter;
capturing an evolution of the at least one feature using the convolutional filter;

generating a state-of-health model;
training the state-of-health (SOH) model by inputting the at least one feature, the evolution of the at least one feature, and a ground truth future state-of-health as variables, in which the state-of-health model comprises a gaussian-process convolutional neural network (GP-CNN) having convolutional layers-tuned to capture short-term and long-term aging effects within the arbitrary time window of measurement;
training a gaussian process regressor using the GP-CNN to generate SOH estimates to capture a relation between identified points through covariance functions, and generate forecasts with error estimates;
embedding a gaussian process regressor in a battery management system of the vehicle; and
predicting, using electrochemical measurements collected during cycling/operation of battery cells of a vehicle battery of the vehicle, a trajectory to failure for the vehicle battery to enable a second life battery application based on the gaussian process regressor.

Assignees

• Toyota Res Inst Inc

Inventors

• Teo Linnette
• Balaji Gopal Chirranjeevi

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• B60L58/12

  responding to state of charge [SoC] · CPC title

• B60L2240/549

  Current · CPC title

• B60L2240/547

  Voltage · CPC title