Sleep spindles are key reference signals in electroencephalography (EEG), typically occurring during non-rapid eye movement sleep. This study evaluates an automated detection method based on image-level analysis using an advanced vision transformer. A novel deep network, SpindleSwin-Net, is proposed for end-to-end and flexible sleep spindle detection by integrating the advantages of wavelet decomposition and the Swin Transformer architecture. The Montreal Archive of Sleep Studies (MASS) and DREAMS datasets are utilised. During image transformation, 1.5 s time-frequency maps containing annotated spindles are generated using a designed continuous wavelet transform filter, while randomly sampled EEG segments without annotations are labelled as non-spindles. In the detection stage, all time-frequency maps are input into a dual-branch Swin Transformer network to classify spindle versus non-spindle signals and to predict coordinates. Five-fold cross-validation is applied during training. SpindleSwin-Net is evaluated through both single-dataset and cross-dataset analyses. The method achieves recall, precision, F1-score, and IoU of (E1: 0.71, 0.62, 0.65, 0.66; E2: 0.99, 0.92, 0.95, 0.67) on the MASS dataset and (E3: 0.84, 0.52, 0.63, 0.60; E4: 0.88, 0.69, 0.77, 0.77) on the DREAMS dataset. Crossdataset experiments demonstrate F1-scores of 0.96 and 0.63 on MASS and DREAMS, respectively. This vision transformer-based approach bypasses the cumbersome steps of conventional signal processing and reduces annotation time, offering potential as a practical tool to alleviate the workload of sleep clinicians.
Groundwater-induced internal erosion of finer soil fractions is critical for initiating granular slope collapse, but developing accurate computational models to capture coupled unsaturated seepage, particle migration, and hydro-mechanical property evolution of soils remains challenging. This study proposes a novel three-phase five-component mathematical model based on mixture theory to describe multiphysics phenomena in seepage-induced erosion in gap-graded granular soils. To enhance model accuracy, an improved micromechanics-based multiscale model is integrated to account for the specific moisture content and grading-dependent mechanical behavior of the porous medium experiencing seepage erosion. The computational framework innovatively employs an enhanced stabilized finite element method to eliminate numerical oscillations, validated via comparison with analytical solutions. The spatio-temporal evolution of the eroded zone and hydro-mechanical response of a reconstituted flume slope were successfully identified. Numerical results show volumetric settlements and shear sliding are major consequences of internal erosion, with soil property degradation near the slope toe accelerating failure. Neglecting internal erosion delays predicted failure onset and underestimates collapse severity. The proposed framework proves efficient and reliable for predicting seepage-induced slope failure initiation.
This paper addresses the boundary stabilization problem for a coupled linear Korteweg-de Vries (KdV) system. The model consists of two coupled dispersive equations, with the input applied at the right Neumann boundary and the output given by the left Neumann boundary. First, a full-state boundary feedback controller is designed using a backstepping transformation, ensuring exponential stability of the closed-loop system. Second, an observer is constructed using the boundary measurements, leading to an output feedback control law. Finally, the stability of the combined plant-observer system is established through a Lyapunov analysis.
The increasing need for adaptive and nonlinear security mechanisms has stimulated interest in chaos-based cryptographic primitive generation. This paper proposes a fractional-order hyperchaotic neural framework for dynamic keystream generation, affine S-box parameterization, and bidirectional diffusion. The proposed method combines a four-dimensional fractional-order hyperchaotic oscillator, discretized using the Grünwald–Letnikov formulation, with a bounded feedforward neural transformation to generate chaotic-neural sequences. These sequences are used to derive round-dependent keystreams, permutation vectors, and invertible affine transformations over GF(2)8. Instead of constructing a fundamentally new S-box family through chaotic sorting, the proposed approach applies key-dependent and round-dependent invertible affine transformations around a strong public base S-box. The resulting S-boxes are affine-equivalent to the base S-box and therefore inherit its principal substitution properties while introducing temporal variability within the multi-round substitution-diffusion process. The dynamical behavior of the fractional-order system is investigated through Lyapunov exponent analysis, dissipativity, Kaplan–Yorke dimension, and bifurcation analysis, confirming the presence of complex hyperchaotic dynamics. The cryptographic behavior is evaluated through entropy, pixel correlation, key and plaintext sensitivity, S-box quality metrics, NIST randomness tests, and ablation analysis. The results indicate that the proposed framework produces noise-like ciphertexts, strong image-domain diffusion behavior, robust sensitivity to small parameter variations, and stable affine-equivalent S-box characteristics. These results are interpreted as empirical evidence of randomness and diffusion, not as formal cryptographic security proofs. The method is not intended to replace standardized ciphers such as AES; rather, it is presented as a nonlinear adaptive primitive-generation framework that links fractional-order hyperchaotic dynamics, neural nonlinear transformation, and dynamic substitution–diffusion design.
