Monitoring the physical layer is key to detect bit error rate (BER) degradation caused by failures and to identify the cause of the failure and localize the failed elements. Once the failure has been detected, actions can be taken to reduce as much as possible its impact on the network. Commercially available optical equipment are able to correct degraded optical signals by means of Forward Error Correction (FEC) algorithms. A value of pre-FEC BER over a pre-defined threshold would imply a non-error-free post-FEC transmission and, as a result, communication would be disrupted. Therefore, a prompt detection of lightpaths with excessive pre-FEC BER can help to greatly reduce such SLA violations, in particular when supporting vlinks. As a result of the above, it would be desirable to anticipate such degradations and apply re-optimization to re-route those affected demands according to their SLAs in order to reduce the affected traffic after the degradation is detected. Designing algorithms capable of promptly detect distinct BER anomaly patterns would be desirable. The objective would be to anticipate intolerable BER values as much as possible aiming at leaving enough time to plan a re-routing procedure during off-peak hours. In this paper, we propose an effective machine learning-based algorithm to localize and identify the most probable cause of failure impacting a given service, as well as a re-optimization algorithm to re-route affected demands, targeting at reducing SLA violation. Results show that the proposed detection and re-routing algorithms noticeably reduce bandwidth and number of demands affected.