The project aims at defining, designing and validating a machine-learning-based method for the detection of radio-frequency anomalies, and for the identification of the associated root causes. The main purpose is to accelerate the diagnostic activity of domain experts in their analysis throughout the whole development and test phases of an antenna system.

Objectives

As of today, performances of antenna systems are evaluated analysing both measured and modelled data. If an anomaly occurs, an iterative process is made to determine the root cause. The goal of this project is to develop and validate AIDA (Artificial-Intelligence-Assisted Performance and Anomaly Detection and Diagnostic), a machine-learning-based software for the detection of RF anomalies and the identification of the associated root causes.

AIDA capabilities

AIDA intends to contribute to the antenna experts analysis and to reduce the diagnosis time by implementing the following software capabilities

  • Early identification of antenna system anomalies, using an AI approach to classify patterns data, implementing generalization strategies in order to foster re-use of a trained model for different antennas under test (AUTs).
  • Accurate anomaly quantification, thanks to a wide labelled database which is used for this purpose.
  • Verification of the anomaly classification and quantification output, comparing the measured patterns with the EM model data updated with the AIDA diagnostic output.

Benefits

Classification performances

AIDA shows very good classification performances when testing antennas of the same operative band as the training dataset

Generalization approach

AIDA can receive as input data from antennas operating at different frequency bands or having different characteristic dimensions, thanks to the generalization approach implemented

Reliability verification

AIDA gives the user the possibility to check the reliability of an anomaly detection directly in the software front-end, computing a comparison between the measured pattern and the updated EM model pattern

Easy to implement algorithms

The implemented algorithms do not need experts’ knowledge on Artificial Intelligence to be configured, but only limited information and knowledge to define the relevant antennas anomalies to be considered

Anomaly quantification

AIDA provides anomaly quantification with good accuracy for specific anomaly classes

Features

AI learning method

AIDA determines the anomaly which characterizes a test input antenna using an AI model which has been trained with a fully supervised learning method

Generalisation to different platforms

From the input antenna raw patterns, derived quantities are computed in order to generalize the use of the software to similar antennas mounted on different platforms

Extensive database

Once the anomaly classification is concluded, a search for the most similar antenna at database to the antenna under test is made in order to compute the anomaly quantification

Generalisation to different antennas

Additional methods are implemented in order to manage data from antennas at different frequency bands and with different characteristic dimensions

Dynamic EM model

The anomaly classification and quantification output is usable by the user to update the EM model of the antenna

Diagnosis verification

Once the EM model is updated, AIDA gives the possibility to verify the diagnosis by comparing the updated patterns with the measured data

System Architecture

AIDA is characterised by a typical three tier architecture composed by:

  • The AIDA database, which collects all the antennas patterns uploaded into the system (training and test data);
  • The AIDA Diagnostic SW, which executes the anomaly classification and quantification, and it contains the function usable for the computation of the reconstruction error; the AIDA Diagnostic SW functionalities can be accessed both from the command terminal and from the developed AIDA Front-end SW.
  • The AIDA Front-end SW, which is a web-based application from which all the features of the AIDA Diagnostic SW can be reached. Moreover, this software allows the user to investigate data from the database, plotting patterns and investigating diagnostic results.

The AIDA Training SW is an external module, which is responsible for the training of an AI algorithm, either imposing the training hyper-parameters, or using an hyper-parameter optimization; in the current version implemented, the AIDA Training SW functionalities can be accessed from the command terminal.

Partners

University of Venice (UNIVE), Italy

T.A.S.I. S.p.A., Italy