Data processing: artificial intelligence – Neural network – Learning task
Reexamination Certificate
2000-02-11
2003-10-21
Davis, George B. (Department: 2121)
Data processing: artificial intelligence
Neural network
Learning task
C706S021000, C706S016000
Reexamination Certificate
active
06636841
ABSTRACT:
BACKGROUND OF THE INVENTION
Field of the Invention
The present invention relates generally to telecommunications system fault location, and more particularly relates to a system and method for telecommunications system fault diagnostics employing a neural network.
Telecommunication systems are generally complex electrical systems which are subject to failure from a variety of fault modes. The rapid and accurate classification and isolation of a fault within a telecommunications system is desired to minimize dispatch and repair costs associated with such faults. Therefore, it is a long standing objective within the telecommunications industry to provide a system which can use measured data to automatically diagnose one of several failure modes.
The accurate diagnosis of faults within a telecommunications system is hampered by the limited accessibility of test points within the system as well as the complex relationships between faults and measurable system parameters. An automated line test system (LTS) that is currently used to perform this function is illustrated in FIG.
1
. In the LTS of
FIG. 1
, a remote test unit (RTU)
2
is employed at each local exchange (EX)
4
within the telecommunication system. The RTU
2
is a hardware device which generates test signals. These test signals are introduced into the system through the EX
4
. The test signals propagate through a main distribution frame (MDF)
6
and into the telecommunications lines
8
. The signals typically pass through a cross connect switch
10
, to one or more distributing points (DP)
12
. Ultimately the signals reach various customer apparatus (CA)
14
such as a modem, facsimile machine, telephone handset and the like.
The telecommunications system, when operating normally, exhibits characteristic parameters in response to the RTU
2
test signal. These parameters include voltage values, current values, resistance values, capacitance values and the like. The RTU
2
samples and evaluates these parameters through the use of software. During a fault condition, these parameters change in response to a given fault.
The diagnostic software
16
implements a simple heuristic algorithm. The algorithm includes decision rules which compare one or more measurements with predetermined (by an engineer) threshold values to determine whether a fault exists. As an example, the algorithm may compare measured resistance values between a pair of lines against a set of expected threshold values which are stored in the program to decide whether a fault exists in either an exchange
4
or customer apparatus
14
. The algorithm uses linear decision rules to perform these functions.
The LTS is also capable of recording the measured parameters in a database
18
for future reference. Additionally, the LTS has the capability of accepting manually entered data regarding each fault from an operator via a keyboard. This information may include customer fault reports and service personnel “clear off” codes indicating the actual location of a fault. In this way, a large amount of data is assembled regarding fault history and parameter values associated with various fault locations. However, the LTS is unable to use this data to improve its own operation. If desired, the data stored in database
18
may be evaluated by an engineer periodically and the decision thresholds employed by the algorithm may be manually updated. This is an extremely labor intensive, and therefore expensive, operation. Therefore, it is a long standing objective in the field of telecommunication system diagnostics to develop a system which can overcome this limitation.
In diagnostic and fault location systems unrelated to telecommunications systems, neural networks have been employed to improve system performance. A neural network is a data processing system largely organized in parallel. The neural network includes a collection of processing elements, or neurons, which are mutually interconnected to one another. The various connections are known as neuronal interconnects. The network is typically formed with an input layer of neurons, an output layer of neurons and one or more hidden layers of neurons.
An important characteristic of neural networks is that they can “learn” by means of neural network training. During training, previously acquired measurement data is applied to the neural network input layer. An error signal is generated at the output layer and is back propagated through the hidden layers of the network. During this operation, the various weights associated with each neuronal interconnect are adjusted to minimize the error signal. If sufficient data is applied to the neural network, the neural network is able to classify unknown objects according to parameters established during training.
In U.S. Pat. No. 5,440,566 to Spence et al., a neural network is employed to perform fault detection and diagnosis for printed circuit boards. The neural network disclosed in the '566 patent is used to process thermal image data from an energized printed circuit board. The neural network is trained by applying data from a printed circuit board with known faults to the network. Once trained, the neural network is then able to analyze new data and classify the new data into one of a plurality of printed circuit board faults.
In U.S. Pat. No. 5.537,327 to Snow et al., a neural network is used in connection with a method and apparatus for detecting high impedance faults in electrical power transmission systems. The system disclosed in the '327 patent employs a trained neural network to evaluate fast Fourier transforms (FFT) of continuously acquired current measurements. The neural network continuously monitors the FFT data and activates a fault trigger output in the event a high impedance fault is detected.
Neural Network Technology
In general, neural networks can be viewed as a powerful approach to representing complex nonlinear discriminant functions in the form y
k
(x; W
k
) where x is an input parameter and W
k
is an optimizing parameter within the neural network. One form of neural network is referred to as a multilayer perceptron (MLP) network. The topology of an MLP neural network is illustrated in FIG.
2
. The MLP network includes an input layer
24
, an output layer
26
and at least one hidden layer
28
. These layers are formed from a plurality of neurons
22
. The input layer
24
receives input parameters and distributes these parameters to each neuron
22
in the first hidden layer
28
. The hidden layers
28
process this data and establish probability estimates for each of a plurality of output neurons which make up the output layer
26
.
Within the MLP network, each single neuron
22
is a discrete processing unit which performs the discriminant function by first performing a linear transformation and then a nonlinear transformation on the input variable x as follows:
u
k
=
ϕ
(
W
k
T
x
+
W
kO
)
=
ϕ
(
∑
j
=
1
d
w
kj
x
i
+
w
kO
)
Eq
.
1
Where &phgr; is a nonlinear function having the form:
ϕ
(
ν
)
=
1
1
+
exp
(
-
ν
)
Eq
.
2
The general network function for the MLP neural network of
FIG. 2
is as follows:
y
k
=
ϕ
{
∑
s
w
ks
(
2
)
ϕ
[
∑
q
w
sq
(
1
)
ϕ
(
∑
j
w
qj
(
0
)
x
j
+
w
qo
(
0
)
)
+
w
so
(
1
)
]
+
w
ko
(
2
)
}
Eq
.
3
Once a network topology is established, it is necessary to “train” the network by applying previously collected training data to the input layer
24
and output layer
26
of the neural network. Optimal network parameters, or interneural weights, are estimated from this known training data. Preferably this is accomplished using a back propagation method. In this process, known data is applied to the input of the neural network and is propagated forward by applying the network equation as previously stated in equation 3. The input data results in output vectors for each layer of the MLP network. The output vectors are evaluated for all output
Austin James
Zhou Ping
Cybula Ltd.
Hoffman & Baron LLP
LandOfFree
System and method for telecommunications system fault... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with System and method for telecommunications system fault..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and System and method for telecommunications system fault... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3124764