Neural filter architecture for overcoming noise interference in a non-linear, adaptive manner

• US 5,442,543 A
• Filed: 04/11/1994
• Issued: 08/15/1995
• Est. Priority Date: 08/11/1992
• Status: Expired due to Fees

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Tresp, Volker
• Primary Examiner(s): Gordon, Paul P.
• Application Number:
  US08/211,656
• Time in Patent Office: 491 Days
• Field of Search: 364/148, 364/149, 364/150, 364/151, 364/572, 395/21, 395/23, 395/24
• US Class Current: 700/29
• CPC Class Codes: H03H 21/0016 Non linear filters

Abstract

The non-linear filter architecture according to the invention provides a neural network for modelling a non-linear transfer function, there being supplied to the neural network, on the input side, the filter input signals f(n), . . . , f(n-i), . . . , f(n-M), a time index signal i and the values p(n), . . . , p(n-i), . . . , p(n-M) of the parameter vector p. The neural network uses these values to calculate, at each time i, output values which are summed for the M+1 times i=0, . . . , M, as a result of which the filter output function g(n) is formed. The invention can be used for implementing a method for overcoming noise signals in digital signal processing, by using a circuit arrangement or a software system. Specifically, the invention can be used in a method for suppressing cardio-interference in magneto-encephalography. The invention can furthermore be used for overcoming motor noise.

Claims

1. A non-linear filter architecture, comprising:

a) a memory means for storing values for input to a neural network;

b) supply means connected to said memory means for successively supplying said values from said memory means to said neural network;

c) a neural network having inputs connected said supply means to receive an input signal f(n-i) associated with time n-i, a time signal for time i and a parameter vector p(n-i) associated with time n-i are linked, at time i;

d) an accumulator means connected to said neural network for summing signals from said neural network;

e) the neural network being constructed such that an output signal g(n) from said accumulator means at time n results from summation of M+1 output signals, which are associated with times n, . . . , n-M, in accordance with a formula ##EQU13## where N[f(n-i),i,p(n-i)] designates an output function of the neural network.

2. An adaptation method for a non-linear filter architecture, comprising the steps of:

a) inputting an input signal f(n-i) to a memory of the non-linear filter, which input signal is associated with time n-i, a time signal for a time i and a parameter vector p(n-i) associated with time n-i are linked, at time i, via said memory to input nodes of a neural network;

b) summing in an accumulator connected to an output of the neural network results M+1 output signals, which are associated with the times n, . . . , n-M, of the neural network, in accordance with a formula ##EQU14## where N[f(n-i),i,p(n-i)] designates an output function of the neural network to obtain an output signal g(n) of the non-linear filter at time n;

varying each weighting w of the neural network in such a manner that a predetermined error E, which represents a value of deviation, determined over a predetermined time interval, of the output signals g(n) of the filter from required output signals gm(n), is minimized by varying the weightings of the neural network.

3. A circuit arrangement for implementing a non-linear filter architecture, having:

a) a neural network having inputs connected to receive an input signal f(n-i) associated with time n-i, a time signal for time i and a parameter vector p(n-i) associated with time n-i are linked, at time i;

b) the neural network being constructed such that an output signal g(n) at time n results from summation of M+1 output signals, which are associated with times n, . . . , n-M, in accordance with a formula ##EQU15## where N[f(n-i),i,p(n-i)] designates an output function of the neural network, comprising;

a) a memory having an input for storing an instantaneous filter input signal f(n), instantaneous values p(n) of the parameter vector, last M filter input signals f(n-1), . . . , f(n-M) and parameter vectors p(n-1), . . . , p(n-M);

b) supply means connected to said memory for, over M+1 successive times n-M, . . . , n, supplying memory contents associated with these times to input nodes to the neural network;

c) summing means connected to an output of the neural network are provided for summation of output values of the neural network over the times n-M, . . . , n.

4. A method for overcoming noise signals in digital signal processing, comprising the steps of:

a) inputting an input signal f(n-i) to a memory for sequential application via a supply means to the non-linear filter, which input signal is associated with time n-i, a time signal for a time i and a parameter vector p(n-i) associated with time n-i are linked, at time i, to input nodes of a neural network;

b) summing in an accumulator means connected to an output of the non-linear filter results M+1 output signals, which are associated with the times n, . . . , n-M, of the neural network, in accordance with a formula ##EQU16## where N[f(n-i),i,p(n-i)] designates an output function of the neural network to obtain an output signal g(n) of the non-linear filter at time n;

varying each weighting w of the neural network in such a manner that a predetermined error E, which represents a value of deviation, determined over a predetermined time interval, of the output signals g(n) of the filter from required output signals gm(n), is minimized by varying the weightings of the neural network.

5. Use of a method as claimed in claim 4, for suppressing cardio-interference in magneto-encephalography.

6. Use of a method as claimed in claim 4, for overcoming motor noise.