Signals integration is used in digital communication systems with data fusion to enhance the performance and reduce the multipath effect. The two main approaches for signals integration in digital communication systems with data fusion are full and semi-full signals integration. In full signals integration systems, there are multiple receivers producing very large number of bits and the entire signals integration system is closely resembled analog multiple receiver implementations. This approach achieves the optimum performance at the expense of high cost and complexity. In semi-full signals integration systems, only few numbers of bits are used after preliminary processing of signals at each individual receiver. This method could reduce system complexity and cost at the expense of overall performance degradation. This paper provides performance analysis of full and semi-full signals integration approaches in digital communication systems in case of non-coherent frequency shift keying (NCFSK) receivers with Gaussian noise and Rician fading stochastic model. The performance loss due to semi-full signals integration is analyzed for different number of information bits.
Accurate and robust pose estimation of non-cooperative spacecraft is critical for autonomous rendezvous and on-orbit servicing. While monocular vision-based methods have attracted growing interest owing to their low cost and structural simplicity, achieving high-precision pose estimation under large scale variations in target distance and complex illumination conditions remains a formidable challenge. In this paper, we propose a novel dual-path prediction network reinforced with a geometric consistency constraint to address these issues. Our framework features two distinct yet complementary pathways. The first path employs a feature pyramid network to extract multi-resolution representations, from which stable keypoints are detected and subsequently integrated with a PnP solver, thereby enabling accurate pose estimation across targets with large scale variations. The second path employs an adaptive-weighted feature pyramid network augmented with a spatial self-attention module to effectively fuse multi-scale information and strengthen global contextual reasoning. Its output is processed by two direct regression heads for rotation and translation, hence improving accuracy and robustness under occlusion and degraded geometric conditions. To ensure coherence between the two pathways, we further introduce a geometric consistency loss that enforces alignment of their outputs during training, thereby improving stability and generalization. Experimental results on SPEED and SwissCube datasets demonstrate that our framework achieves substantial improvements over existing methods, particularly under extreme conditions.
Fatigue crack growth under high-cycle fatigue is one of the most severe problems in the design, maintenance, and safe operation of aircraft structures. During operation, these structures experience millions of loading cycles, which cause the gradual growth of nucleated cracks leading to ultimate failure. Thus, accurate modeling of fatigue crack growth is necessary for ensuring structural integrity, maximizing inspection intervals, and extending the service life of aerospace structures. Conventional methods of modeling crack growth under high-cycle fatigue use linear elastic fracture mechanics to arrive at the cyclic stress intensity factor, which is then used in the Paris Law describing steady state crack growth. Paris law is highly non-linear, consisting of two constants, C and m, under fully reversible loading. These parameters are evaluated using Euler integration of Paris law and linear regression of scattered crack growth measurements from standard tests. However, a lack of constraints during the calibration can render the parameter estimates to be inaccurate particularly when the data is significantly scattered. To address this limitation, Physics-Informed Machine Learning (PIML) architectures are employed to calibrate the parameters of Paris law. However, before utilizing this calibration approach, the accuracy of Physics-Informed Neural Networks (PINNs) to integrate Paris law was tested. To this end, the predictions from Physics-Infused Long-Short Term Memory (PI-LSTM) and Implicit Euler Transfer Learning (IETL) architecture were also compared to Euler integration, and a reasonable agreement was obtained. Following this, these methods were applied to obtain the parameters from numerically generated data using some assumed C and m values. It was observed from the study that the method was not only able to calibrate the parameters, but also that the network could be used to predict crack growth when the amplitudes were modified. Finally, scattered data was artificially generated by choosing distributions of C and m. Subsequently, IETL was applied to the scattered data to calibrate the parameters and showed a satisfactory comparison. In summary, this study exemplifies the merits and demerits of different PIML methods when applied to predict crack growth from the Paris law. Furthermore, the approach allows both crack growth evolution and Paris constants to be predicted from limited experimental data, thereby reducing the need for repeated costly tests across different loading cases. Finally, the reliability of the PIML framework to predict crack growth for various amplitudes and block loading is demonstrated.
