Training of neural network for multi-source data fusion

• US 5,276,770 A
• Filed: 11/19/1992
• Issued: 01/04/1994
• Est. Priority Date: 08/11/1989
• Status: Expired due to Term

• Current Assignee: Hughes Electronics Corporation (Rtx Corporation)
• Original Assignee: Hughes Aircraft Company (Rtx Corporation)
• Inventors: Castelaz, Patrick F.
• Primary Examiner(s): MacDonald, Allen R.
• Application Number:
  US07/979,394
• Time in Patent Office: 411 Days
• Field of Search: 395/23, 395/24, 395/27, 395/22
• US Class Current: 706/25
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology

Abstract

A method of training a multilayer perceptron type neural network to provide a processor for fusion of target angle data detected by a plurality of sensors. The neural network includes a layer of input neurons at least equal in number to the number of sensors plus the maximum number of targets, at least one layer of inner neurons, and a plurality of output neurons forming an output layer. Each neuron is connected to every neuron in adjacent layers by adjustable weighted synaptic connections. The method of training comprises the steps of (a) for each sensor, designing a plurality of the input neurons for receiving any target angle data, the number of designated input neurons for each sensor being at least as large as the maximum number of targets to be detected by the sensor; (b) for a known set of targets having a known target angle for each sensor, applying a signal related to each known target angle to the designated input neurons for each of the sensors, wherein the output neurons will produce an initial output; (c) for a selected one of the sensors, designating a plurality of the output neurons to correspond to the input neurons designated for the selected sensor and applying the signal related to the known target angles for the selected sensor to the designated output neurons to provide a designated output signal wherein the difference between the initial output and the designated output signal is used to adapt the weights throughout the neural network to provide an adjusted output signal; and (d) repeating steps (a)-(c) until the adjusted output signal corresponds to a desired output signal.

Claims

1. A method of training a neural network of the multilayer perceptron type to provide neural network processor for fusion of target angle data from targets, said targets being detected by a plurality of sensors which provide target angle data for each detected target, said neural network including a plurality of input neurons forming a first layer, the number of input neurons being at least equal to the number of sensors plus the maximum number of targets to be detected by the sensors, at least one layer of inner neurons, and a plurality of output neurons forming an output layer, each neuron being connected to every neuron in the adjacent layer of neurons by weighted synaptic connections which are capable of increasing or decreasing the connection strength between individual neurons, said method of training comprising the steps of;

(a) for each sensor, designating a plurality of the input neurons for receiving any target angle data from said sensor, the number of said designated input neurons for each sensor being at least as large as the maximum number of said targets to be detected by said sensor;

(b) for a known set of targets, each target having a known target angle for each sensor, applying a signal related to each said known target angle to the designated input neurons for each of said sensors, wherein the output neurons will produce an initial output;

(c) for a selected one of said sensors, designating a plurality of said output neurons to correspond to the input neurons designated for said selected sensor and applying said signal related to said known target angles for the selected sensor to the designated output neurons to provide a designated output signal wherein the difference between the initial output and the designated output signal is used to adapt the weights throughout the neural network to provide an adjusted output signal; and

(d) repeating steps (a)-(c) until the adjusted output signal corresponds to a desired output signal.

2. A method of training a neural network as recited in claim 1 further comprising the steps of:

(a) repeating steps (a)-(c) for at least one additional set of known targets 1

3. A method of training a neural network as recited in claim 1 wherein step (b) comprises:

for a known set of targets, each target having known target angle for each sensor, applying a normalized signal related to each said known target angle to the designated input neurons for the appropriate one of said sensors, wherein the output neurons will produce an initial output.