Virtual vehicle sensors based on neural networks trained using data generated by simulation models

• US 6,236,908 B1
• Filed: 05/07/1997
• Issued: 05/22/2001
• Est. Priority Date: 05/07/1997
• Status: Expired due to Fees

• Current Assignee: Ford Global Technologies, LLC (Ford Motor Company)
• Original Assignee: Ford Global Technologies, LLC (Ford Motor Company)
• Inventors: Cheng, Jie, Tascillo, Anya Lynn, LaCrosse, Stephanie Mary, Newman, Charles Edward Jr., Davis, George Carver
• Primary Examiner(s): LOUIS JACQUES, JACQUES H
• Application Number:
  US08/852,829
• Time in Patent Office: 1,476 Days
• Field of Search: 701/1, 701/2, 701/36, 701/29, 701/31, 701/33, 701/34, 701/23, 701/27, 701/40, 701/101, 701/102, 701/103, 701/106, 706/20, 706/21, 706/22, 706/23, 706/24, 706/25, 706/10, 706/14, 706/16, 706/31, 706/41, 706/905, 364/150, 364/151, 364/528.01, 364/528.15, 382/157, 382/190, 382/280
• US Class Current: 701/1
• CPC Class Codes: F02D 2041/1433 using a model or simulation of the system
• CPC Class Codes: F02D 2041/1437 Simulation
• CPC Class Codes: F02D 41/0062 Estimating, calculating or determining the internal EGR rate, amount or flow
• CPC Class Codes: F02D 41/0072 Estimating, calculating or determining the EGR rate, amount or flow sensors in EGR systems F02M26/45
• CPC Class Codes: F02D 41/1405 Neural network control
• CPC Class Codes: G05B 13/027 using neural networks only
• CPC Class Codes: Y02T 10/40 Engine management systems

Abstract

A virtual vehicle sensor includes a neural network which produces a sensor output based on a linear combination of non-linear physical signals generated by conventional physical sensors. Instead of determining an output directly, the neural network determines the polynomial coefficients as functions of the physical signals indicative of other engine operating parameters. The sensor is manufactured using relatively limited data collection to calibrate a simulation model. The output of the simulation model is used for model-based mapping to generate more comprehensive maps used for training the neural network. The trained neural network is embedded in a controller and acts as the virtual sensor to monitor engine parameters which are difficult to measure or for which conventional physical sensors do not currently exist. The virtual sensor may be used to sense parameters such as in-cylinder residual mass fraction, emission levels, in-cylinder pressure rise during combustion, and exhaust gas temperature.

Claims

1. A method of manufacturing a sensor for use with a vehicle component having a controller in communication with a plurality of physical sensors each generating a signal indicative of first operating parameters, the sensor determining values for a second operating parameter based on values for the plurality of first operating parameters, the method comprising:

generating test data during operation of the vehicle component representative of values for the plurality of first operating parameters for a first set of operating conditions;

calibrating a simulator for simulating operation of the vehicle component using the test data;

generating at least one map which characterizes performance of the vehicle component as a function of predetermined parameters, the map being based on output of the simulator for a second set of operating conditions;

adjusting weights corresponding to nodes of a neural network based on the at least one map so as to develop a trained neural network; and

embedding the trained neural network into the controller by storing a representation of the trained neural network in computer readable media, the representation including a plurality of instructions executable by a microprocessor and data representing the weights corresponding to the nodes of the neural network, such that the trained neural network determines values for the second operating parameter based on values for the plurality of first operating parameters.

2. The method of claim 1 wherein the vehicle component comprises an engine and wherein the step of generating test data comprises operating the engine on a dynamometer.

3. The method of claim 1 wherein the step of adjusting weights comprises adjusting weights corresponding to monotonically increasing piecewise differentiable function nodes of the neural network.

4. The method of claim 1 wherein the step of generating test data comprises generating test data at a plurality of representative anchor points.

5. The method of claim 4 wherein the step of generating at least one map comprises interpolating data based on the plurality of representative anchor points.

6. A method for controlling a vehicle component using a plurality of physical sensors for sensing first operating parameters and a controller in communication with the plurality of physical sensors, the method comprising:

monitoring signals generated by the plurality of physical sensors to determine values for the first operating parameters;

processing the values for the first operating parameters using a neural network embedded in the controller to determine a value for residual mass fraction, the value for residual mass fraction being based on a linear combination of the plurality of values for the first operating parameters such that the neural network functions as a sensor for the second operating parameter; and

controlling the vehicle component based on the value of the second operating parameter.

7. The method of claim 6 wherein the step of processing comprises processing values representative of engine speed, manifold pressure, EGR, and air/fuel ratio.

8. The method of claim 6 wherein the step of processing comprises determining polynomial coefficients for residual mass fraction, the coefficients being functions of the first plurality of operating parameters.