IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 6, DECEMBER 2001 1773 A Wavelet Networks-Based Method for the Digital Telecommunication System Monitoring Pasquale Daponte, Senior Member, IEEE, Gianpaolo Mercurio, and Sergio Rapuano Abstract—This paper deals with an automatic monitoring method for digital telecommunication systems. The method is based on the use of wavelet networks (WNs) for signal classifi- cation. After some theoretical considerations on the WN theory, the procedure for the implementation of the method is described. The method, designed for quadrature amplitude modulation signals, can be easily extended to further modulation schemes. The metrological characterization of the method is carried out by means of numerical and experimental tests. The results highlight the good performance of the method. Index Terms—Digital communications, quality-of-service, time–frequency representations, wavelet networks. I. INTRODUCTION T HE GOAL of any digital telecommunication system man- ager is to win the market competition offering a high- quality service. Therefore, anomalous events, called stresses, which reduce the quality of the transmitted signal and even com- pletely degrade it, are to be detected and removed as soon as possible through a reliable in-service monitoring (ISM) system [1]. A second reason for establishing ISM over the telecommu- nication system is that the in-service stress conditions cannot be reproduced in out-of-service tests [2]. The transmission state can be monitored by means of tradi- tional multimonitor systems [3], [4]. At the present time, they present some limits: i) high cost, ii) complexity, and iii) need of an expert human operator able to interpret the huge data quan- tity produced by the monitoring system. The paper proposes an innovative method which overcomes the above-mentioned limits and, in particular, the need of an expert human operator, allowing the automatic monitoring of the stresses overlapped to the quadrature amplitude modulation (QAM) signals under test. The method is able i) to detect the presence of stresses, ii) to perform the automatic classification of stresses overlapped to the QAM signal under test, by means of a wavelet network (WN) based algorithm, and iii) to measure the parameters character- izing the stresses [5]–[8]. The paper is organized as follows: Section II reports some theoretical observations about the WNs. Section III analyzes the proposed method and the implemented algorithms. Section IV describes the obtained results on simulated and actual stressed QAM signals. Manuscript received December 1, 2000; revised September 25, 2001. The authors are with Facoltà di Ingegneria, Università del Sannio, Benevento, Italy (e-mail: daponte@unisannio.it). Publisher Item Identifier S 0018-9456(01)10960-5. Fig. 1. Wavelet network scheme. II. THEORETICAL BACKGROUND ON WNS An artificial neural network (ANN) structure classifies the input waveform following shape-based criteria in the time-do- main. On the other hand, a time–frequency analysis best fits the signals addressed in this work. So, the ANN structure can be improved by means of a preprocessing input filter block based on a time–frequency representation (TFR). That block furnishes the time–frequency information on the signal necessary to the ANN for a correct classification. In order to avoid a manual arrangement of the time–frequency parameters of the preprocessing filters, a WN can be used [9]. In a WN structure (Fig. 1) the sigmoidal activation function of each neuron of the input layer is replaced with a modified ver- sion of an opportune base function, called the mother wavelet. The input layer can be seen as a filter bank, and the parameters characterizing the input nodes are: , representing a transla- tion parameter, and , representing a scale parameter inversely proportional to the characteristic frequency of the node (central band frequency of the filter). The input node number is established by the number of the most important time–frequency characteristics. In particular, a wavelet node for each time–frequency characteristic to be rec- ognized is necessary. During the training phase, the WN develops appropriate de- cision-making capability through the weighting coefficients ad- 0018–9456/01$10.00 © 2001 IEEE