Adaptive machine learning method to predict and control engine combustion

• US 10,030,602 B2
• Filed: 07/22/2015
• Issued: 07/24/2018
• Est. Priority Date: 07/22/2014
• Status: Active Grant

• Current Assignee: Board of Regents of the University of Michigan (University of Michigan)
• Original Assignee: Board of Regents of the University of Michigan (University of Michigan)
• Inventors: Vaughan, Adam
• Primary Examiner(s): Vilakazi, Sizo
• Assistant Examiner(s): Steckbauer, Kevin R
• Application Number:
  US14/806,037
• Publication Number:
  US 20160025028A1
• Time in Patent Office: 1,098 Days
• CPC Class Codes: F02D 2041/1412 using a predictive controller
• CPC Class Codes: F02D 35/021 using an ionic current sensor
• CPC Class Codes: F02D 35/023 by determining the cylinder pressure
• CPC Class Codes: F02D 35/028 by determining the combustion timing or phasing
• CPC Class Codes: F02D 41/008 Controlling each cylinder individually
• CPC Class Codes: F02D 41/1405 Neural network control
• CPC Class Codes: F02D 41/263 the program execution being modifiable by physical parameters
• CPC Class Codes: F02D 41/401 Controlling injection timing F02D41/402 takes precedence
• CPC Class Codes: Y02T 10/40 Engine management systems

Abstract

A method is provided for controlling an internal combustion engine on a cycle-by-cycle basis. The method includes: maintaining training data used to identify a mapping function for the engine; populating a buffer with adaptive data for a given cylinder, where the adaptive data are measures of the inputs and the output of the mapping function that were captured during a number of recent operating cycles of the given cylinder; combining training data for the mapping function with the adaptive data; identifying the mapping function from the combined data set using a weighted least squares method; predicting a combustion feature of the given cylinder in the next cycle using the mapping function and measures of the inputs during the present cycle; and controlling the engine based in part on the predicted combustion feature of the given cylinder in the next cycle using one or more actuators.

Claims

1. A method for controlling an internal combustion engine on a cycle-by-cycle basis, comprising:

(a) providing, by a controller, a mapping function which defines a relationship between inputs and an output associated with the internal combustion engine, where the output is a given combustion feature of a given cylinder in the engine at a next cycle and the inputs include the given combustion feature of the given cylinder in a present combustion cycle and parameters indicative of mixture state and composition or engine actuator settings from one cycle to the next cycle;

(b) maintaining, by the controller, training data used to identify the mapping function in a data store associated with the controller, where the training data is measures of the inputs and the outputs of the internal combustion engine over a range of operating conditions;

(c) populating, by the controller, a buffer with adaptive data for the given cylinder in the present combustion cycle, where the adaptive data are measures of the inputs and the output of the mapping function that were captured during m most recent combustion cycles of the given cylinder;

(d) combining, by the controller, the training data with the adaptive data to form a combined data set, where the training data excludes the adaptive data;

(e) identifying, by the controller, the mapping function from the combined data set using a weighted least squares method;

(f) predicting, by the controller, the given combustion feature of the given cylinder in the next cycle using the mapping function and measures of the inputs during the present cycle;

(g) controlling, by the controller, the engine by actuating fuel injectors or another actuator in the engine based in part on the predicted combustion feature of the given cylinder in the next cycle; and

(h) repeating steps c) - g) for subsequent combustion cycles of the internal combustion engine.

2. The method of claim 1 wherein the given combustion feature is selected from a group consisting of combustion timing, combustion duration, total heat release quantity, pressure rise rate, location of peak pressure, ion current, knock intensity, and flame vs. auto ignition fraction.

3. The method of claim 1 wherein the engine has one or more pistons driven by a crankshaft, the output for the mapping function is defined as combustion timing of the given cylinder in the next cycle, and the inputs for the mapping function being selected from a group consisting of combustion timing of the given cylinder in present cycle, an indicator of fuel injected into the given cylinder in the present cycle, crank angle(s) at which fuel injection occurred in the present cycle and one or more pressure measures made inside the given cylinder.

4. The method of claim 1 wherein providing a mapping function further comprises

defining the mapping function, wherein the inputs of the mapping function are selected from a group consisting of combustion timing of the given cylinder in present cycle, an indicator of fuel injected into the given cylinder in the present cycle, crank angle at which fuel injection occurred in the present cycle and one or more pressure measures made inside the given cylinder;

capturing measurements of the inputs during a present cycle and corresponding combustion timing of the given cylinder at the next cycle while varying operating conditions of the engine;

storing the captured measurements of the inputs and the output as the training data;

determining a matrix that scales one or more transformations of the inputs of the mapping function to the output of the mapping function using a least squares method, such that the matrix serves as the mapping function for the engine; and

downloading the training data into the data store associated with the controller.

5. The method of claim 1 wherein populating a buffer further comprises capturing measures of pressure inside the given cylinder during the recent operating cycles.

6. The method of claim 1 wherein the buffer is further defined as a ring buffer.

7. The method of claim 1 wherein identifying the mapping function further comprises machine learning the mapping function with a single-hidden layer feedforward neural network with randomly chosen weights for each of the hidden nodes in the neural network.

8. The method of claim 1 further comprises repeating steps (c) - (g) for subsequent combustion cycles in the engine.

9. The method of claim 1 wherein controlling the engine further comprises determining crank angle at which to inject fuel during the next cycle in accordance with

10. The method of claim 1 wherein controlling the engine further comprises determining crank angle at which to inject fuel during the next cycle using gradient descent with multiple actuators.

15. The method of claim 1 wherein the inputs of the mapping function are identified by unsupervised learning, deep learning, or autoencoders.

11. A method for predicting combustion timing of an internal combustion engine, the engine having two or more pistons driven by a crankshaft, comprising:

defining, by a controller, a mapping function for the internal combustion engine, where the mapping function defines combustion timing of a given cylinder in the engine at a next cycle in terms of inputs selected from a group consisting of combustion timing of the given cylinder in present cycle, an indicator of fuel injected into the given cylinder in the present cycle, crank angle at which fuel injection occurred in the present cycle and one or more pressure measures made inside the given cylinder;

capturing, by the controller, measurements of the inputs and corresponding combustion timing of the given cylinder at the next cycle during varied operating conditions of the engine;

determining, by the controller, a vector that scales the inputs of the mapping function to the output of the mapping function using a least squares method, wherein the vector identifies the mapping function;

storing the captured measurements of the inputs and the output as the training data;

downloading the training data and the vector into a data store in a vehicle having an internal combustion engine;

populating a buffer with adaptive data for the given cylinder in the present cycle, where the adaptive data are measures of the inputs and the output of the mapping function that were captured during m most recent combustion cycles of the given cylinder;

combining the training data with the adaptive data to form a combined data set, where the training data excludes the adaptive data;

updating the vector using a combined data set;

predicting combustion timing of the given cylinder in the next cycle using the vector and measures of the inputs during the present cycle; and

controlling combustion timing in the internal combustion engine of the vehicle based in part on the predicted combustion timing.

12. The method of claim 11 wherein the combustion timing of a given cylinder at the next cycle is further defined as crank angle at the time fifty percent of net heat release has occurred.

13. The method of claim 12 wherein the one or more pressure measures include a pressure inside the given cylinder during carryover from one combustion cycle to another combustion cycle, a pressure inside the cylinder before gas is fully expanded, and a pressure inside the given cylinder when gas is being compressed.

14. The method of claim 11 wherein determining a vector further comprises machine learning the mapping function with a single-hidden layer feedforward neural network with randomly chosen weights for each of the hidden nodes in the neural network.

16. A controller for an internal combustion engine residing in a vehicle, comprising:

a data store configured to store training data used to identify a mapping function and associated with the controller, where the training data is measures of the inputs and the outputs of the internal combustion engine over a range of operating conditions, the mapping function defines a relationship between inputs and an output associated with the engine, the output is combustion timing of a given cylinder of the engine in a next cycle, and the inputs include combustion timing of the given cylinder in the given cylinder in a present cycle and one or more parameters indicative of mixture state and composition from the present cycle to the next cycle;

a buffer configured to store adaptive data for the given cylinder, where the adaptive data are measures of the inputs and corresponding output of the mapping function captured during m most recent cycles of the given cylinder and excludes the training data;

an adaptor module configured to receive adaptive data for the given cylinder and populate the buffer with the received adaptive data;

a predictor module having access to the buffer and the data store, the predictor module configured to combine the training data with the adaptive data residing in the buffer, determine a matrix that scales one or more transformations of the inputs of the mapping function to the output of the mapping function using a weighted least squares method, and predict combustion timing for the given cylinder in the next cycle using the matrix and measures of the inputs during the present cycle;

a control module interfaced with one or more fuel injectors and configured to control the fuel injectors based in part on the predicted combustion timing for the given cylinder.

17. The engine controller of claim 16 further comprises a pressure sensor or an ion current sensor residing in the given cylinder and interfaced with the adaptor module.

18. The engine controller of claim 16 further comprises a position sensor interfaced with the adaptor module, wherein the internal combustion engine has one or more pistons driven by a crankshaft and the position sensor is configured to determine crank angle of the crankshaft.

19. The engine controller of claim 18 wherein the inputs for the mapping function are selected from a group consisting of combustion timing of the given cylinder in present cycle, an indicator of fuel injected into the given cylinder in the present cycle, crank angle(s) at which fuel injection occurred in the present cycle and one or more pressure measures made inside the given cylinder.

20. The engine controller of claim 19 wherein the mapping function is implemented by a single-hidden layer feedforward neural network with randomly chosen input weights for each of the hidden nodes in the neural network.

21. The engine controller of claim 20 wherein the control module controls the fuel injector by determining crank angle of the crankshaft at which to inject fuel during the next cycle in accordance with