Coherent optical receivers continuously run adaptive digital signal processing (DSP) to recover transmitted data. As a by-product, the internal states of this DSP, such as adaptive equalizer coefficients and polarization estimates, form a rich multivariate time series. This side-channel information reflects physical disturbances acting on the fibre, such as vibrations and mechanical stress, yet it is normally discarded. This project uses machine learning to extract this side-channel information at the receiver and turn existing communication links into environmental sensors, without any dedicated sensing hardware. We train neural networks to detect, classify and estimate disturbance events from receiver-side DSP states. The model families include temporal convolutional networks, recurrent networks and Transformer-based sequence models, together with model-based architectures that embed known signal-processing structure for interpretability. The models are trained and validated on labelled datasets recorded in the laboratory under controlled disturbances. The computational work consists of training these models, systematic hyperparameter and ablation studies, and evaluation across dataset splits. This requires many independent GPU training runs. The expected outcome is accurate and interpretable learning-based methods that enable joint communication and sensing. This project is part of the PhD studies of Jun Cao at Chalmers University of Technology, and is an extension of his Master's thesis work at Chalmers. Main supervisor: Christian Häger, Department of Electrical Engineering, Chalmers University of Technology.