Earthquakes are sudden releases of energy in the Earth's crust that generate seismic waves recorded as time‐series signals called seismograms. Earthquake analysis depends on interpreting the temporal structure of seismic phases, yet many machine-learning approaches require large manually labeled datasets that are difficult to obtain at scale. To address this, we propose a self-supervised framework that learns waveform continuity and ordering cues without human annotations. Our approach consists of two steps: preprocessing and model training. In preprocessing, each one-dimensional waveform is converted into a standardized waveform image and partitioned into a fixed number of consecutive segments with small gaps to simulate missing observations and create non-contiguous inputs. In model training, these segments are randomly shuffled and used to train a Siamese convolutional neural network to predict the shuffled order through a permutation-classification objective, where pseudo-labels are generated automatically from the applied shuffle. We evaluate the framework across multiple segment settings and summarize its performance trends as task difficulty increases. Overall, this study demonstrates that permutation-based self-supervision can enable learning useful waveform structure from unlabeled data, and it suggests a practical direction for developing phase-aware representations that can support future seismic analysis when labeled data are limited.

To address the nonlinear and non-stationary characteristics of vibration signals during rolling bearing operation and the issue of insufficient degradation information representation, this paper designs a prediction framework that combines Adaptive Variational Mode Decomposition (AVMD) with a SETCN-BiGRU multi-head temporal attention mechanism for Remaining Useful Life (RUL) prediction. First, AVMD is used to decompose the bearing horizontal vibration signal into five Intrinsic Mode Functions (IMFs). Time-domain and frequency-domain statistics are extracted from each IMF and concatenated into a 115-dimensional degradation feature sequence. Subsequently, the model processes in parallel: a TCN-SENet branch extracts local temporal features and adaptively adjusts channel weights, while a BiGRU with multi-head temporal attention sub-network captures global bidirectional dependencies and critical degradation periods within the degradation sequence. Finally, the two types of features are fused, and the RUL prediction result is output. Experimental results demonstrate that the proposed model achieves an RMSE, MAE, and R² of 0.0582, 0.0477, and 0.9483 respectively on the IEEE PHM 2012 dataset, and average values of 0.0780, 0.0559, and 0.9133 on a self-built laboratory bearing dataset, indicating good prediction accuracy, robustness, and generalization ability.

This study investigates how accelerated sea-level rise interacts with tidal dynamics along the East Coast of the United States, which are regions influenced by the Gulf Stream. Using 2004–2019 satellite altimetry, NOAA tide-gauge data, and the Atlantic Meridional Overturning Circulation (AMOC) index, the analysis reveals that both sea-surface height (SSH) and tidal range (TR) exhibit consistent increases. Although their year-to-year variations are weakly linked, SSH and TR share a strong long-term slope change. We speculate that increases in SSH may deepen the water and are associated with increased TR, although the underlying mechanisms remain uncertain. Integrating AMOC reanalysis data indicates that the weakening of its circulation likely drives the SSH rise and tidal changes. A potential regime transition around 2020 introduces new uncertainty, suggesting the AMOC system may be entering a different dynamical state. These findings highlight the need for improved multi-decadal modeling to support coastal resilience and future ocean renewable energy station planning.