Accurate, non invasive estimation of core body temperature ( T c ) is vital for real time physiological monitoring and heat illness prevention, yet gold standard measurements are invasive and impractical for wearables. We propose a physics guided hybrid framework that embeds a two node thermoregulation model within a neural residual learning pipeline. The physical layer uses heart rate (HR) to estimate metabolic heat production and directly incorporates wearable head skin temperature, while the residual learner (1D-CNN/LSTM/GRU) corrects systematic model discrepancies. Evaluated on a controlled dataset with strict participant-level splits, the Hybrid-CNN (full input) achieved the best overall accuracy (RMSE = 0.265 ∘ C, MAE = 0.202 ∘ C, R 2 = 0.638 ), outperforming CNN/LSTM/GRU-only networks, an Extended Kalman Filter (RMSE = 0.365 ∘ C), and a physics-only baseline (RMSE = 0.604 ∘ C, negative R 2 ). A simplified Hybrid-CNN that omits ambient sensors performed similarly (RMSE = 0.268 ∘ C, R 2 = 0.632 ), indicating deployability with just HR and skin temperature. Noise robustness tests, conducted by injecting Gaussian noise into HR ( σ = 5 -50 bpm), showed Hybrid-CNN leading for σ ≤ 20 bpm, whereas Hybrid-LSTM was most resilient under extreme noise ( σ = 50 bpm; RMSE = 0.353 ∘ C). The residual learners are lightweight (120-140 KB; 250k-700k FLOPs/step), supporting real-time, on-device inference. Overall, coupling interpretable thermophysiology with targeted residual learning yields accurate, robust, and computationally efficient T c monitoring for wearable health systems.

This work presents an always-on CNN processor featuring compute-in-memory (CIM) and layer-fusion (LF) techniques. It demonstrates an end-to-end neural network (NN) inference while eliminating memory accesses for both the weight and inter-layer activation, thus significantly saving associated energy. The analog LF-CIM processor combines charge- and time-domain multiply-accumulate (MAC) arrays in a fusional manner, avoiding the need for the interfacial analog-to-digital and digital-to-analog converters; it also supports a highly sparsity-adaptive network and exhibits low-voltage-supplied tolerance to circuit noise and variations, which facilitates an energy-efficient and robust CNN processor, simultaneously. Furthermore, the full-precision activation originating from the LF reduces the NN parameters and operations, while the inference accuracy is improved by the presented ensemble NN. The prototype processor is fabricated in a 28-nm CMOS process, demonstrating a fully on-chip MNIST inference with 97.9% accuracy. It operates at 3,508 frames per second while consuming$11.6~\mu $W at a 0.5-V supply; the achieved efficiency of 3.32 nJ/frame is over 50-fold than the state-of-the-art end-to-end MNIST accelerators.

This paper presents a weak PUF-assisted strong PUF that combines the metrics between weak and strong PUFs. Unlike the conventional strong PUFs that rely on the nonlinear combination of a large number of entropy cells for high modeling attacks resilience, the presented strong PUF utilizes unique key streams to bitwise encrypt the raw responses from the conventional strong PUF to facilitate an inherent immunity to modeling attacks thus fulfilling high-level security. A device-specific pseudo-random number generator (P-RNG) configured by a dedicated weak PUF array generates a unique key stream (K). Since the weak PUFs are inherently immune to machine learning or deep learning-based modeling attacks, the final encrypted responses of the proposed strong PUF are also inherently immune to modeling attacks. Moreover, we propose a two-to-one (2-to-1) selection scheme and the digitally-controlled-delay-line (DCDL)-based stability checker to suppress the bit-error-rate (BER) of the weak PUF array and improve the efficiency of the spatial majority voting (SMV)-based error correction scheme, thus achieving high stability for our proposed strong PUF. Fabricated in 65nm CMOS GP technology, the proposed weak PUF-assisted strong PUF shows a high energy efficiency of 3.05 pJ/bit at a 2M bit rate. Meanwhile, it demonstrates an ultra-low average worst-case BER of$8.9 \times 10^{-11}$for the temperature range of −20°C to 120°C and a supply voltage variation of ±10% with the proposed stabilization schemes. The proposed strong PUF occupies a core area of 0.075mm2.

As one of the most important auxiliary heating facilities of tokamak, Neutral Beam Injector (NBI) has the highest heating efficiency and clearest operating mechanism. Almost all important tokamak devices in the world are equipped with NBI, including Experimental Advanced Superconducting Tokamak (EAST). The mismatch of the operating parameters of the NBI ion source will lead to the instability of the plasma, and maybe cause breakdown in the ion source, which will limit the operation of the NBI long pulse and high power, and it will even be difficult to heat the plasma in the tokamak. In order to stabilize the plasma in the NBI ion source, data from multiple rounds of experiments along with a priori information obtained from a predictive plasma model are used. This paper proposes a method based on Self-Organizing Map (SOM) and Back Propagation(BP) type neural network to estimate the pulse width during the beam extraction process of the NBI ion source under given parameters by training historical data, and adjust the operating parameters accordingly to reduce the ratio of ion source breakdown. The SOM approach mainly relies on current and voltage sensors's data instead of a priori information, and tries to estimate the beam extraction pulse width with less calculations. A BP neural network is also designed to reduce the uncertainty of the SOM algorithm. The algorithms have been tested on off-line data, obtained from experimental shots at EAST-NBI, Hefei, China.