Automotive engine misfire detection system including a bit-serial based recurrent neuroprocessor

• US 6,434,541 B1
• Filed: 04/21/1999
• Issued: 08/13/2002
• Est. Priority Date: 10/23/1996
• Status: Expired due to Fees

• Current Assignee: California Institute of Technology, Ford Global Technologies, LLC (Ford Motor Company)
• Original Assignee: Ford Global Technologies, LLC (Ford Motor Company)
• Inventors: Puskorius, Gintaras V., Tawel, Raoul, Aranki, Nazeeh, Feldkamp, Lee A., James, John V., Marko, Kenneth A.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US09/284,843
• Time in Patent Office: 1,210 Days
• Field of Search: 706/30, 706/43
• US Class Current: 706/30
• CPC Class Codes: G01M 15/11 by detecting misfire

Abstract

An engine diagnostic system includes a bit-serial based recurrent neuroprocessor for processing data from an internal combustion engine in order to diagnose misfires in real-time and reduces the number of neurons required to perform the task by time multiplexing groups of neurons from a candidate pool of neurons to achieve the successive hidden layers of the recurrent network topology.

Claims

1. A system comprising:

a neuroprocessor for implementing a reconfigurable network topology including a plurality of hidden layers containing neurons interconnected in a recurrent configuration;

the neuroprocessor being responsive to an input pattern corresponding to processed real time engine operating conditions for determining whether a fault condition has occurred;

the neuroprocessor comprising a neural module including a plurality of bit-serial neurons;

a global controller for time multiplexing groups of neurons from the neural module to form first and second hidden layers of the network topology, the controller controlling application of the input pattern to the first hidden layer and controlling storage of the output of said first hidden layer for subsequent application as input to said second hidden layer.

2. The system defined in claim 1 configured to detect engine misfire, said input pattern includes a crankshaft acceleration input, an engine load input, an engine speed input and a cylinder identification input, the neuroprocessor producing an output indicating whether a misfire has occurred.

3. The system defined in claim 2 wherein each neuron comprises a bit-serial multiplier for multiplying first and second inputs, the controller sequentially applying to said first input of said multiplier one input of said input pattern or the activation output of a neuron on a previous layer, the controller applying a synaptic weight appropriate for said one input of said input pattern to said second input of said multiplier, each neuron further comprises a bit-serial accumulator for accumulating the output of said multiplier.

4. The system defined in claim 3 wherein the input bits are provided to the multiplier in parallel and the input synaptic weight bits are provided to the multiplier serially, the multiplier computing the product of the two inputs and making the results available to the accumulator on a bit-serial basis.

5. The system defined in claim 4 wherein the accumulator comprises a cyclical shift register with an adder at the input stage allowing the pipelining of outputs from the multiplier to be accumulated as the data is available from the multiplier.

6. The system defined in claim 5 wherein upon completion of all multiply and accumulate operations, a tri-state data latch stores the relevant accumulated sum until required by the controller.

7. The system defined in claim 6 wherein inputs are obtained from an engine controller.

8. An engine misfire detection system comprising:

a neuroprocessor for implementing a reconfigurable network topology including a plurality of hidden layers containing neurons interconnected in a recurrent configuration;

the neuroprocessor being responsive to an input pattern for determining whether a misfire has occurred, said input pattern corresponding to processed real time engine operating conditions;

the neuroprocessor comprising a neural module including a plurality of bit-serial neurons;

a global controller for time multiplexing groups of neurons from the neural module to form first and second hidden layers of the network topology, the controller controlling application of the input pattern to the first hidden layer and controlling storage of the output of said first hidden layer for subsequent application as input to said second hidden layer;

said global controller comprising a configuration controller, and a run-time controller, the configuration controller including a plurality of configuration registers containing data that explicitly defines the topology of each layer of the recurrent neural network architecture.

9. The system defined in claim 8 wherein the configuration controller includes a data bus, an address bus, and receives a configuration control signal, a clock signal, and a reset signal, the address on the address bus internally selecting a configuration register as the destination of data on the data bus.

10. The system defined in claim 9 wherein the neuroprocessor further includes a neuron state RAM module for storing the contents of the output layer of the neural network.

11. The system defined in claim 10 wherein the neuron state RAM module is a single port RAM module.

12. The system defined in claim 9 wherein the run-time controller comprises a current layer register selector, a finite state machine for sequencing high-level inter-layer operations, an intra-layer propagation controller for controlling execution of neuronal multiply and accumulates, and an intra-layer specific neuron output data storage controller for controlling calculation of non-linear activations for the neurons whose linear state is given in accordance with the following equation:

$x_j=\underset{i}{\sum <br><br>}<br><br>\text{\hspace{1em}<br><br>}<br><br>y_i<br><br>w_{\mathrm{ij}}$

and for subsequently storing the resulting quantities in the neuron state RAM.

13. The system defined in claim 12 wherein the neuroprocessor further includes a sigmoid activation look-up-table for performing the non-linear activation function.

14. The system defined in claim 13 wherein said input pattern includes a crankshaft acceleration input, an engine load input, an engine speed input, a cylinder identification input, and a bias input.

15. The system defined in claim 14 wherein each neuron comprises a bit-serial multiplier for multiplying first and second inputs, the controller sequentially applying one input of said input pattern or the activation output of a neuron on a previous layer to said first input of said multiplier, the controller applying a synaptic weight appropriate for the input to said second input of said multiplier, each neuron further comprises a bit-serial accumulator for accumulating the output of said multiplier.

16. The system defined in claim 15 wherein the input bits are provided to the multiplier in parallel and the input synaptic weight bits are provided to the multiplier serially, the multiplier computing the product of the two inputs and making the results available to the accumulator on a bit-serial basis.

17. The system defined in claim 16 wherein the accumulator comprises a cyclical shift register with an adder at the input stage allowing the pipelining of outputs from the multiplier to be accumulated as the data is available from the multiplier.

18. The system defined in claim 17 wherein upon completion of all multiply accumulates, a tri-state data latch stores the relevant accumulated sum until required by the controller.

19. The system defined in claim 18 wherein certain ones of the inputs in said pattern are sensor data processed by an engine controller prior to input to said neuroprocessor.

20. An engine misfire detection system comprising:

a neuroprocessor for implementing a reconfigurable network topology including a plurality of hidden layers containing neurons interconnected in a recurrent configuration;

said neuroprocessor being responsive to an input pattern for determining whether a misfire has occurred, said input pattern corresponding to processed real time engine operating conditions;

said neuroprocessor comprising a neural module including a plurality of bit-serial neurons;

a global controller for time multiplexing groups of neurons from the neural module to form first and second hidden layers of the network topology, the controller controlling application of the input pattern to the first hidden layer and controlling storage of the output of said first hidden layer for subsequent application as input to said second hidden layer;

said global controller comprising a configuration controller, and a run-time controller, the configuration controller including a plurality of configuration registers containing data that explicitly defines the topology of each layer of the recurrent neural network architecture;

said global controller including a run-time controller for initiating a neurocomputation using the data in said configuration registers, and a neuron state RAM module for storing the contents of the output layer of the neural network;

said run-time controller comprising a current layer register selector, a finite state machine for sequencing high-level inter-layer operations, an intra-layer propagation controller for controlling execution of neuronal multiply and accumulates, a sigmoid activation look-up-table for performing a non-linear activation function and an intra-layer specific neuron output data storage controller for controlling calculation of non-linear activations for the neurons and for subsequently storing the resulting quantities in said neuron state RAM, said input pattern including a crankshaft acceleration input, an engine load input, an engine speed input, a cylinder identification input, and a bias input.