New Techniques for Testing and Operational Support

of AESA Radars

Stephane Kemkemian, Cyrille Enderli Alain Larroque

Thales Airborne Systems Thales Airborne Systems
Elancourt, France Pessac, France
kemkemian@ieee.org, cyrille.enderli@fr.thalesgroup.com Alain.larroque@fr.thalegroup.com

Abstract— The Active Antennas (AESA) technology has

dramatically increased the operational capability of modern II. TEST STRATEGY AT PRODUCTION LEVEL
radars. Nevertheless, minimize production costs and cost of
ownership of these systems is also a major industrial objective. A. Testing Method used up to now
The paper is organized in two parts. The first one deals with This testing method was being carried out on all radars
improvements achieved so far for reducing the costs and the manufactured by TAS for over forty years. It has mainly three
complexity of industrial testing. In the second part, a data test steps:
mining method, based on Bayesian Networks, is presented. It
aims at processing all data issued from the Built-In-Test (B.I.T.) • Functional tests at subsets level including printed
in order to accurately detect some defects impossible to catch boards and other electronic assemblies.
with current methods, such as transient failures or initiation of
youth’s defects. Originally planned for testing in production, • Performances verifications at subset level and
this new performing method could replace, in the future, the elimination of youth’s defects on subsets.
current B.I.T. processing, which is used in operational support.
• Verification of Sensor system performances.
The targeted evolution consists in physically integrating
I. INTRODUCTION
the entire radar in one testing step.
Unlike the centralized systems where the transmit signal is
generated by a central transmitter and applied to the antenna; B. The Evolution of the Test Strategy
the AESA generates itself the transmit power and receive
The evolution of the industrialization strategy and
capability in each individual module. This arrangement
integration tests of the AESA goes along with functional
significantly reduces RF losses compared to systems with
evolutions of the radar and relies on the following principles:
central transmitter and receiver. Thanks to the instant beam
steering agility, multiple-modes can operate at the same time, • A calibration of the AESA subset via the radar.
something not possible with conventional mechanically
scanned systems. • A test mode to characterize the sensor.
The reliability of the centralized systems was mainly • Functional tests based on the Built-In-Test of the
limited by the reliability of power devices (especially the radar.
transmitter). This is no longer the case with AESA radars: • The use of data mining to detect more accurately
because the antenna remains operational even if some modules youth’s defects in particular of the AESA.
have failed (graceful degradation), the antenna lifespan can
run into thousands of hours. 1) The calibration Means :
Despite these improvements, industrial production tests are Classically, they are:
always critical, even more complex than with conventional • The Near Field Chamber (NFC): A probe scans the
systems (hundreds or thousands of active modules to be tested radiating face of the antenna at short distance. The
in operation). corresponding measurements are used to calibrate all
Moreover, the reliability of AESA makes even more the antenna’s channels, both on transmit and receive.
critical the early detection of youth’s faults which may affect
• The Compact Far Field Chamber (CFFC) to
other radar subsets. A method developed by Thales Airborne
characterize the antenna pattern after the previous
Systems (TAS) is presented below.
step.

2013 IEEE Radar Conference (RadarCon13)

978-1-4673-5794-4/13/$31.00 ©2013 IEEE
 Reconstruction methods from near field measurements C. New Smart Detection of Defects
only have been developed over the last years and can now 1) Limitation due to the current B.I.T.
carry out all the qualifications of an antenna without the NFC
step (calibration, antenna diagram, measurement of the EIRP The test method, which has been previously described,
at transmission and G/T at reception) [1], [2], [3]. suffers from limitations that are inherent to the nature of the
B.I.T. itself. Indeed, the B.I.T. is optimized for the operational
2) The Test Mode use and not for testing in production phase. Thus, to avoid
The test mode is an embedded software application in detecting too many intermittent faults, which would be
which the radar makes use of its own resources (signal, data unmanageable in operational support, several “software
processing and waveform generation). It thus performs a self- filters” are used for masking too many transient faults.
characterization of the radar sensor without external means by To overcome this problem, one solution is to use directly
generating and processing synthetic echoes. the raw data used by the B.I.T. from the different subsets right
The test mode allows: after their aggregation. These raw data have clearly much
richer content; nut no one (“human” agent) can neither
• In production phase: reduction of the test means and effectively exploit such content nor find the inference relations
reduction of test time. existing within the data. Indeed, the flow of raw data to B.I.T.
• In operational use: improvement of operational is of the order of 150 words each millisecond.
availability and reduction of maintenance resources. 2) Data Mining Experimentation
The measurements resulting of the test mode are derived To validate this principle, works have been performed
either from the software processing of test signals (e.g. using COTS Artificial Intelligence software called
spectral purity, inspection of the waveform, receiver response, BAYESIALAB™ (version 4.6.5) developed by the BAYESIA
etc.), or from integrated measurement probes in the subsets Company. The process has mainly two steps:
(e.g. currents and temperatures within AESA, etc.). They are
1st step - Supervised learning:
then aggregated and processed by the Built-In-Test.
By running a Radar deemed “good” during hours in
For the production of new AESA, the industrial objective
various environmental conditions, a learning file is created.
is to go on simplifying the means of calibration and
The learning file has therefore a size of few GB.
characterization of antennas. The roadmap for the means of
test and calibration (cf. Figure 1. ) consists of two phases: At first, the software builds a “raw” Bayesian Network
(BN). Each node of the BN is one of the 150 variables. That is
• The development of an external calibration caisson to
to say it automatically searches for probabilistic inference's
calibrate the AESA installed on the radar.
relations among the 150 variables issued from the B.I.T.
• Development of a self-calibration method embedded during hours. Note that the details of the algorithms used for
in the radar. inferences searching are property of BAYESIA Company.
To give an order of idea, the learning data base which was
used corresponds to about one hour of recording and it takes
about 15 minutes to build the network with a standard
desktop. The learning is performed on the binary words of the
B.I.T. without any knowledge of the meaning of each bit from
the software. In addition, the causality of phenomena has not
been taken into account in this learning: indeed, the time
aspect has not been taken into account.

Figure 1. Characterization and calibration AESA roadmap.

3) The functional Tests

Functional tests have been developed in production for
detecting youth’s defects of radars and for lowering the
removal rate as original equipment.
The principle is to run the radar, as it would be in
operational configurations, by stimulating its inputs via the Figure 2. Refined Bayesian Network modeling the radar behavioir.
previous test mode according to scenarios of a few hours.
During testing, the behavior of the radar is monitored via the The learning is supervised in the sense that a “human”
Built-In-Test (B.I.T.) under different environmental conditions expert can interacts and add a priori knowledge in the learning
(ambient, cold, hot, vibrations, wet atmosphere, etc.) phase: once this “raw” BN constructed, the intervention of an
expert has led to retain only 29 “main” variables with
significant inference relationships, and then a second refined Probability Log-Likelihood
BN is computed. This refined Bayesian Network constitutes
the “reference model” (cf. Figure 2. ).
CASE #3
2nd step – Detection and localization of failures:
By running now the Radar to be tested, each line of Quite frequent
variables issued from the B.I.T. is tested through the “model” event with very
BN and the likelihood of this line of variables, with regard to low likelihood.
the reference model, is calculated. A high likelihood value
indicates a good concordance, while a low value indicates a
fault. The analysis of the statistical deviations between the
reference model and the B.I.T. data-flow obtained from the
radar under test allows determining when and where failures
occur.
The experimental study was conducted by generating
twenty test cases. Two of them included simulated transient
failures and a third one was recorded while the antenna was
perturbed by an incoming spurious signal during calibration.
Figure 4. Likelihood Graph of case #3.
The data files were analyzed blindly by BAYESIALAB™.
The software detected three cases of deviation from the
reference. The first two cases of deviation (simulated transient III. CONCLUSION AND PERSPECTIVES
failure) showed concordances with the contingency table of
This paper presents the developments already achieved
respectively 82% and 86% when all cases expected “good”
and planned for testing of radars in production phase,
had a concordance value greater than 97%. The third case
especially those fitted with an active antenna:
(perturbation by spurious signal) of deviation had a very low
concordance (< 50%) and a graph of likelihood very different • The antennas can now be fully tested only using
from the reference (cf. Figure 3. and Figure 4. ). “Near Field Chamber” and radar's own resources.
Probability Log-Likelihood
• The other subsets can also be tested using the radar's
own resources. However, the testing efficiency is
currently limited by issues of efficient processing of
the huge dataflow from the B.I.T.
REFERENCE
An experiment was conducted in order to process these
Almost data by COTS Bayesian Network software that provides a
permanent Likelihood scale much finer analysis of the behavior of the radar to be tested.
behavior with (negative LOG scale)
high likelihood. The short term objective is to perform this processing in
delayed time within the test-bed. The next objective would be
to implement such software in the radar and replace the
existing B.I.T. The best accuracy of analysis would allow
detecting incipient failures and thus significantly improve the
operational availability of the radar.

ACKNOWLEDGEMENTS
The authors wish to thank the BAYESIA Company for
their collaboration in the work on data mining.
Figure 3. Likelihood Graph of reference.
REFERENCES
These tests have clearly demonstrated the interest of using
raw data instead of filtered binary states (e.g. quantized power [1] Renard, Silvy, “Testing an airborne phased array military Satcom
antenna with ARAMIS near-field range”, 12th AMTA (Antenna
value of a signal rather than an averaged state with respect to a Measurement Techniques Association) Symposium, Philadelphia, PA,
threshold). This constraint should be fulfilled in the radar USA, Oct.8-10, 1990.
architecture (both hardware and software) from scratch. [2] Patent 08-02985 (FR), 30 Mai 2008, “Procédé et dispositif de mesure
en champ proche du facteur de mérite d’une antenne”.
In this first experiment, we did not attempt to determine
[3] Chabod, Renard, “SOSTAR-X Active Antenna: Results and Lessons
“when” and “where” the failure occurred. The next goal will Learned” , EuMW / EuRAD 2008 Conference, Amsterdam,
be to localize doubtful lines in the data file and the targeting of Netherlands, Oct.27-31, 2008
suspicious variables enabling the localization of defects.