|
NSC/DOC(94)23
NEURAL NETWORK BENCHMARK FOR SMORN-VII
CONDITION MONITORING IN NUCLEAR REACTORS
USING NEURAL NETWORK
Organisation for Economic Co-operation and Development
Nuclear Energy Agency
Nuclear Science Committee
I. INTRODUCTION
The neural networks are getting widespread application almost in all
engineering fields as the researches are reported in the literature. Perhaps
the most important feature they exibit is that they can efficiently process
data where parallel processing is desirably required due to size and complexity
of the data. Among their outstanding appealing properties, often their
ability to process multivariate data with nonlinear as well as linear
dependence among the variables and their fault-tolerance are referred.
Although the reported works with neural networks indicate the recognition
of this new technology due to the successful outcomes, the degree of success
and the limitations at the same time involved are neither clearly reported
nor well understood. This is due to complex information processing mechanism
of neural networks in the multidimensional space so that the detailed
modelling of the mechanism has not been identified yet, in spite of focused
attention on this issue. As result of this, often the outcomes of the
applications reflected the effectiveness of the neural network methodology
rather than the engineering predictions before the application at hand,
as the latter is the case in most other engineering applications as it
should be.
On the above described perspective, the present benchmark is designed
to identify the effective utilization of neural network technology in
nuclear engineering field to investigate the possibilities to improve
the existing methodologies in use. In this respect plant monitoring is
selected as the application area and participants are requested to report
their way of neural network utilization using the data distributed and
their reasonings for their approach yielding their neural network structure
in use.
Thus, in brief the goal of the benchmark can be stated as follows: The
benchmark is intended to identify the applicability of the neural networks
in the nuclear industry for monitoring.
II. BENCHMARK DATA
Data are obtained from a fuel cycle operation and correspond to different
operations. This includes start-up (I), normal operation (II), shutdown
(III), start-up (IV), normal operation (V), and again shutdown (VI) relevant
to the end of the fuel cycle. Identification of the signals are given
in Table 1 and locations of the sensors are given
in Fig. 1.
Data file comprise 44 variety of signals, 32 of which are given in each
of 6 particular operations. Data is formed by combining these six operations
sequentially so that the whole data consist of 1169 time steps where at
each time step 32 signals are given.
For neural network utilization at each time step a suitable number of
signals are selected which are subject to analysis. Therefore each time
step is used to form a pattern, the total number of patterns being maximum
1169. Operations and related information are given in Table
2, where actual 32 channel identification is given for each operation.
In the first operation time, step is selected as 10 minutes and in the
following operations time steps are 1, 2, 1, 1 and 1 minute respectively.
Data are given in ASCII format on a floppy suitable for PC-reading.
A sample printout of the starting part of the file is given in
Fig. 2 for 7 time steps.
The full data is available here
III. BENCHMARK TASKS
Benchmark tasks are divided into two parts. The first part contains
the basic tasks which are requested to be carried out as a minimum to
participate in the Benchmark.
In the whole benchmark analysis 14 signals will be used ( see Fig.3),
which are shown as bold in Table I and Table
II (channels: 1, 7, 8, 14, 15, 16, 17, 18, 19, 24, 25, 26, 27, 29).
It is important to note that the temperature information at the input
and the output of the core will be used in differential form; that is
the neural network will receive the information as difference between
channel 15 and channel 14 as well as difference between channel 17 and
channel 16 (see on figures 4 and 5).
Second part of the benchmark contains optional tasks which are required
to be performed in a similar way as before with some additional analyses,
so as to reflect the outcomes of your further studies.
As the data is obtained from a real process the relative measurement
errors in signals is approximately the same for all signals and it is
less than 1 % for nominal operational conditions.
A. Primary (basic) Tasks
| Task a |
Use first 600 patterns from the beginning for training your neural
network and perform recall (i.e. neural
network estimation) for both 942 patterns and total 1169 patterns.
The neural networks structures to be used in their analysis are given
in Fig 4 (11 inputs, 1 output) and Fig.
5 (autoassociative network with 12 input and 12 outputs). |
| Task b |
Carry out the same analysis in Tasks a using 1000
patterns in learning. In this case the recall is
requested only for total 1169 patterns for both network structures. |
| Task c |
Instead of taking 8 nodes in the hidden layer use another number
instead of eight if eight for any reason is concluded to be not appropriate
and/or optimal. |
Reporting the results and complementary questions
- Report in files (ASCII on a floppy) the estimated values by neural
network for 942 and 1169 patterns seperately together with the respective
errors (difference in physical units) for the same number of patterns.
- Report in a file the estimated values of Task b for
1169 patterns together with respective errors as defined above for the
same number of patterns.
- Report in a file the weights of each neural structure used after the
respective training processes. This implies, for instance, in Fig. 5
reporting the followings:
| WI(L,M) |
= |
input weights; L is the number of input (L=12), M is the number
of nodes in the hiddenlayer (M=8) |
| WTI(M) |
= |
bias in the sigmoid function at hidden layer (M=8) |
| WTO(N) |
= |
bias in the sigmoid function at output (N=12). |
- Give sum of the squared errors averaged over the number of patterns
used during the training.
Also, give enough information (e.g.) normalization (if it is used
at all) of the input and the output data etc, about the introduction of
the data to the neural network so as to be able to reproduce the results
at hand during later evalution of the reported results.
B. Secondary (optional) Tasks
| Task d |
Carry out sensitivity analysis for cases where neural network is
trained in autoassociative mode. Sensitivity is defined as the variation
of each output with respect to the inputs. It is not ment for the
variation of the synaptic weights. |
| Task e |
The data delivered contained redundant signals so that one might
consider it is worth to repeat the benchmark experiments in Task
a and Task b once more including these redundant
signals. |
Reporting the results and complementary questions
- Interpret the results of Task d and explain the importance
and robustness of this analysis.
- Report the results of Task e in the form as described
in Task c.
Final notes:
- Give your name and adress and your E-mail, phone number in a README.TXT
file.
- Give a brief introduction about your work.
- In addition to the results reported in ASCII files it is desirable
to present relevant plots in hard copy form for later convenience during
benchmark evaluation.
- The reported results should appropriately be in different files for
easy identification. Information to identify the files should be clearly
indicated in an INFO.TXT. Each file containing the reported results
should desirably include an informative introduction or a header.
- Give information about the neural network program used indicating
if it is commercial or your own. If it is commercial, its name; version
number; type of computer used; training algorithm; the necessary time
for training and further noteworthy information.
Organizers: E. Türkcan (ECN) and Ö. Ciftcioglu
(ITU)
Information:
- E. Türkcan
Netherlands Energy Research Foundation ECN
Nuclear Energy, Dynamic Signal Analysis
P.O.Box 1, 1755 ZG Petten
The Netherlands
E-MAIL : turkcan@ecn.nl
Telephone: (+31) 2246-4385 / 4262.
Telefax: (+31) 2246-3490
|
