W1G.1.pdf OFC 2018 © OSA 2018 Learning from the Optical Spectrum: Soft-Failure Identification and Localization [Invited] Luis Velasco * , Behnam Shariati, Alba P. Vela, Jaume Comellas, and Marc Ruiz Optical Communications Group (GCO), Universitat Politècnica de Catalunya (UPC), Barcelona, Spain e-mail: lvelasco@ac.upc.edu Abstract: The availability of coarse-resolution cost-effective Optical Spectrum Analyzers (OSA) allows its widespread deployment in operators’ networks. In this paper, several machine learning approaches for failure identification and localization that take advantage of OSAs are presented. © 2018 Optical Society of America 1 OCIS codes: (060.0060) Fiber optics and optical communications, (060.1155) All optical networks 1. Introduction Failure identification and localization can reduce failure repair times greatly. Failure localization techniques have been proposed mainly for hard failures, while significant work is still required for soft failure detection, identification, and localization. Note that some soft failures could affect signal QoT and eventually evolve to hard failures. In a recent work [1], the authors proposed monitoring the performance of lightpaths at the transponders side to verify their proper operation, as well as to detect BER degradations thus, anticipating connection disruptions. The authors analyzed several soft failure causes affecting signal QoT, such as laser drift, filter shift, and tight filtering, and propose algorithms to detect and identify the most probable failure. However, monitoring the signal at the egress node does not allow localizing failures and therefore, monitoring techniques to analyze and evaluate QoT in-line are required. In this regard, the availability of a new generation of cost- effective OSAs with sub-GHz resolution, integratable in the optical nodes [2], allows real-time monitoring of the optical spectrum of the lightpaths and computing their corresponding OSNR. Note that, when a signal is properly configured, its central frequency should be around the center of the assigned spectrum slot to avoid filtering effects, and it should be symmetrical with respect to its central frequency. Therefore, optical spectrum features can be exploited by machine learning-based algorithms to detect degradations and identify failures. In this paper, we analyze DP-QPSK and 16QAM modulated signals and, from [4], study approaches to detect filtering related failures, i.e., Filter Shift and Tight Filtering; the optical spectrum would be asymmetrical in the case of filter shift, and its edges get noticeably rounded in the case of tight filtering. These changes allow distinguishing optical spectra suffering from such failures from properly configured ones. However, some of these effects, in particular for the case of tight filtering, can be observed when a properly configured signal passes through several filters (filter cascading). Therefore, it is of paramount importance to devise solutions to cope with this issue preventing the misclassification of a properly configured signal as a failed one. We study several alternatives solving this issue, which can be used individually or combined. Ultimately, the optical spectrum analysis can be used by sophisticated algorithms able to identify and localize failures. These algorithms can be deployed in the network controller, as well as in nodes’ agents, close to the observation points, to reduce the amount of monitoring data to be conveyed to the control/management plane [3]. 2. Failure detection and identification with OSAs Let us firstly overview our proposed soft failure detection and identification process that utilizes the optical spectrums captured by OSAs deployed in the intermediate nodes [4]. The process involves both, modules running in node agents and modules running in the controller; this follows the architecture proposed in [3]. Fig. 1a shows an example of optical spectrum acquired by an OSA with 625 MHz resolution. In general, QPSK and 16QAM optical signals present, once filtered, a flat spectral region around the central frequency, sharp edges, and a round region between the edges and the central one. Running in node agents, a module named as Feature Extraction (FeX) receives the C-band optical spectrum acquired by the local OSA and extracts the data for the portion of the spectrum allocated to the lightpath under study; data consists of an ordered list of frequency-power (<f, p>) pairs. After equalizing power, so the maximum power to be 0 dBm, the derivative of the power with respect to the frequency is computed; Fig. 1b illustrates the derivative of the example optical signal, where sharp convexity can be observed close to the edges. Next, the FeX module characterizes the mean (μ) and the standard deviation (σ) of the power around the central frequency (fc±Δf), as well as a set of primary features computed as cut-off points of the signal with the following power levels: i) equalized noise level, denoted as sig (e.g., -60dB + equalization level); ii) edges of the signal computed using the derivative, denoted as ∂; iii) a family of power levels computed w.r.t. μ-kσ, denoted as kσ; and iv) a family of power levels computed with respect to μ-mdB, denoted as -mdB. Each of these power levels generates a couple of cut-off 1The research leading to these results has received funding from the EC through the METRO-HAUL project (G.A. nº 761727), from the Spanish MINECO TWINS project (TEC2017-90097-R), and from the Catalan Institution for Research and Advanced Studies (ICREA).