IoT

The accurate prediction of the power of High-Pressure Grinding Rolls (HPGR) is considered essential for evaluating equipment operating conditions and optimizing energy consumption. Existing power prediction methods, however, are considered to have notable limitations: mechanistic models exhibit limited adaptability under dynamic conditions, and purely data-driven models lack physical logic constraints, often resulting in physically inconsistent predictions during extreme conditions such as shutdowns. To address these problems, proposed in this paper is a power prediction method that combines Long Short-Term Memory (LSTM) networks with Physics-Informed Neural Networks (PINN) under a sliding-window rolling prediction framework, termed LSTM-PINN. This method takes HPGR as the research object and integrates the LSTM-PINN framework with the physical mechanism of high-pressure roller mills for power prediction. Specifically, based on the physical operating mechanism of HPGR, four constraint relationships-monotonicity, linear prior, thermodynamic, and energy conservation-are introduced and embedded into the loss function of LSTM-PINN. While LSTM-PINN has been applied in other domains, its combination with the HPGR physical mechanism for power prediction has not been previously reported. The LSTM network extracts dynamic temporal features of power evolution, while the four mechanism-driven physical constraints guide the model toward physically consistent predictions. Experiments using actual operational data from a mining company demonstrate that: (1) the pure LSTM model outputs non-zero power during shutdown, whereas LSTM-PINN brings predictions to zero; (2) compared with pure LSTM, LSTM-PINN reduces RMSE by 23.13%, 13.02%, and 11.86%, and reduces MAE by 21.67%, 13.70%, and 10.85% under the single-feature, multi-feature, and all-feature input combinations, respectively. Among all configurations, the multi-feature combination achieves the lowest absolute errors with the smallest variance, representing the optimal trade-off between accuracy and stability. Overall, the proposed method substantially improves prediction accuracy, interpretability, and reliability, providing a trustworthy basis for energy optimization and intelligent control of HPGRs.

Internet of Things (IoT) malware and intrusions generate network-observable flow patterns, but high benchmark accuracy does not by itself establish transfer to new environments. This study presents Protocol-conditioned Analysis with Behavioral Threat Identification (PA-BTI) as an evidence-bounded framework combining dataset-specific Random Forest/XGBoost detection, protocol-conditioned direction-invariant CorrAUC (dCorrAUC) diagnostics, a River streaming diagnostic, and context-aware alert prioritization. Across the primary evaluations, selected operating points reached 99.77% accuracy with 0.48% false-positive rate (FPR) on ToN-IoT and 99.44% accuracy with 0.08% FPR on a balanced Bot-IoT holdout, whereas the official NSL-KDD split reached 79.55% accuracy and 66.20% attack recall. To address testbed-artifact sensitivity, a new official UNSW-NB15 train/test ablation obtained 87.67% accuracy, 98.59% attack recall, and 25.70% FPR with all features; removing TTL/state variables changed accuracy to 87.22%, and additionally removing the ct_* windowed features changed it to 86.31%, showing no collapse but a persistently high official-split FPR. A new integrated common-flow replay propagated real UNSW detector outputs and training-only protocol evidence through the contextual layer using explicitly synthetic CTI/asset context. On a disjoint 41,166-record evaluation half, calibrated fusion did not improve context-graded nDCG@100 over confidence alone (0.610 versus 0.627), demonstrating component interaction while bounding any claim of prioritization benefit. The River diagnostic also exposed a minority-class failure mode (91.56% accuracy but attack-class F1 = 0.003). PA-BTI is therefore presented as a reproducible, bounded evaluation framework rather than evidence of cross-dataset generalization or robust portability.

Deepfake manipulation has become increasingly difficult to detect as generative models continue to produce highly realistic facial content. This work presents the Frequency-Aware Lightweight Hybrid Network (FALH-Net Lite), a compact deepfake classification framework that combines spatial- and frequency-domain representations for resource-constrained edge deployment. The proposed framework integrates grayscale intensity with three high-frequency Discrete Wavelet Transform (DWT) sub-bands (LH, HL, and HH) to form a four-channel input, which is processed using an adapted YOLO11n-cls backbone. A Frequency-Aware Squeeze-and-Excitation (FreqSE) module is incorporated to recalibrate channel responses, while Focal Loss is employed during training to address class imbalance. The Celeb-DF-v2 dataset is partitioned at the video level before frame extraction to reduce the risk of near-duplicate leakage between training and validation data. Under the corrected video-level evaluation, FALH-Net Lite achieves an accuracy of 87.7%, with a balanced accuracy of 62.1%, an ROC-AUC of 73.6%, an F1-score of 93.2%, and an MCC of 36.7%. The model contains approximately 1.26 million parameters and requires 0.4885 GFLOPs, providing a compact computational profile for edge inference. Deployment experiments on the NVIDIA Jetson Orin Nano and Raspberry Pi 5 further evaluate the feasibility of executing the proposed classification pipeline on resource-constrained hardware. Cross-dataset validation and broader multi-run statistical evaluation remain directions for future work.