Millimeter wave (mm-Wave) radar has been widely employed for non-contact vital sign detection. This research proposes a novel processing scheme to improve the accuracy and stability of non-contact heart rate detection. Firstly, we propose an adaptive range bin selection method based on variance which can select accurate range bins containing phase information of vital signs. Secondly, the smooth spline is utilized to fit the original phase signal to obtain a relatively pure template signal. Subsequently, the pure template signal is used as the input signal for the matched filter to convolve with the phase signal. The matched filtering can eliminate interference and noise such as breathing and random body movements (RBM). Next, the heartbeat signal is extracted by the Variable Mode Extraction (VME) algorithm. Finally, a novel frequency measurement technique named as Double-Chirp Z-Transform (Double-CZT) is proposed. The frequency of the heartbeat signal is measured by the Double-CZT. Compared with the traditional frequency measurement techniques, the Double-CZT improves the accuracy of frequency measurement and allows for more accurate heart rate measurements. Many subjects were invited to participate in the experiment, and the mean absolute error (MAE) of heart rate was less than 1 beat per minute (bpm) on average. The experimental results demonstrate that the proposed scheme can improve the signal-to-noise ratio (SNR) of heartbeat signal and provide accurate heart rate estimation.

Energy consumption has become a crucial issue due to the large-scale deployment of small base stations (SBSs) in dense small cell networks (DSCNs). In this paper, a joint optimization problem involving sleep mode in subframes and power allocation to minimize the DSCN energy consumption while guaranteeing users' rate requirements is formulated as a mixed integer nonlinear programming. To address this problem, we propose a cooperative sleep and power allocation approach by decomposing it into two subproblems. First, we derive the optimum number of active subframes for each SBS and present a centralized heuristic coalition formation algorithm to manage SBSs to form coalitions. In such a case, SBSs can transmit data in active subframes and sleep in others. We then obtain the SBSs' optimum transmit power in the active subframes relying on a distributed price-based power allocation algorithm. System-level simulation results show that our proposed cooperative scheme can yield significant performance gains in terms of energy saving compared with the maximum power allocation and the non-cooperative power allocation (NCPA) approaches. In addition, the effects of target rates on coalition size and energy consumption are also analyzed.

For large-scale high-dimensional positioning scenes, the massive number of grid points brings challenges to the multi-target positioning algorithms based on compressed sensing. To cope with the challenges, a fast multi-target localization method based on direction of arrival is proposed. A compressed sensing model is constructed for multi-target localization based on the DOA sequence measured by positioning nodes. Then, a two-stage fast matching pursuit algorithm is presented for sparse reconstruction, which consists of preliminary estimation and supports rectification. A process similar to orthogonal matching pursuit algorithm is adopted to get preliminary estimate result, but no nonlinear operations is employed for complexity reduction. Then another iterative process is carried out to rectify the chosen supports in preliminary result sequentially. Simulation results verify the effectiveness and accuracy of the proposed method for multi-target localization.

To address the triple bottlenecks of data scarcity, oversized models, and slow inference that hinder Cantonese automatic speech recognition (ASR) in low-resource and edge-deployment settings, this study proposes a cost-effective Cantonese ASR system based on LoRA fine-tuning and INT8 quantization. First, Whisper-tiny is parameter-efficiently fine-tuned on the Common Voice zh-HK training set using LoRA with rank = 8. Only 1.6% of the original weights are updated, reducing the character error rate (CER) from 49.5% to 11.1%, a performance close to full fine-tuning (10.3%), while cutting the training memory footprint and computational cost by approximately one order of magnitude. Next, the fine-tuned model is compressed into a 60 MB INT8 checkpoint via dynamic quantization in ONNX Runtime. On a MacBook Pro M1 Max CPU, the quantized model achieves an RTF = 0.20 (offline inference 5 × real-time) and 43% lower latency than the FP16 baseline; on an NVIDIA A10 GPU, it reaches RTF = 0.06, meeting the requirements of high-concurrency cloud services. Ablation studies confirm that the LoRA-INT8 configuration offers the best trade-off among accuracy, speed, and model size. Limitations include the absence of spontaneous-speech noise data, extreme-hardware validation, and adaptive LoRA structure optimization. Future work will incorporate large-scale self-supervised pre-training, tone-aware loss functions, AdaLoRA architecture search, and INT4/NPU quantization, and will establish an mJ/char energy–accuracy curve. The ultimate goal is to achieve CER ≤ 8%, RTF < 0.1, and mJ/char < 1 for low-power real-time Cantonese ASR in practical IoT scenarios.

In dense small cell networks (DSCNs), small cells are heterogeneous due to the unplanned deployment and various traffic loads, resulting in different energy efficiency (EE) preferences. In this paper, taking into account the heterogeneity of small cells, energy saving (ES) and EE are jointly optimized through subchannel allocation, subframe configuration and power allocation. In order to quantify the effects of small cells’ heterogeneous information on EE, an EE preference function is first defined. Then, the joint ES and EE optimization problem is formulated as a multi-objective optimization problem. Due to the coupling of ES and EE, obtaining the solution is non-trivial. Therefore, we propose a heterogeneity-aware ES and EE (HESEE) optimization algorithm, where subchannel allocation is optimized to ensure the fairness of active subframes required by the users in the same small cell base stations (SBSs). Then subframe configuration is conducted via group formation sleep mechanism. Particularly, address the non-concave sum-of-ratios optimization for system EE, the concave-convex procedure (CCCP) method is adopted. Simulation results show that the proposed HESEE algorithm can optimize the SBSs’ EEs according to their EE preferences. In addition, the HESEE algorithm can achieve good performance in reducing energy consumption as well as improving the system EE.

With the rapid development of Automatic Speech Recognition (ASR) technology, end-to-end speech recognition systems have gained significant attention due to their ability to directly convert raw speech signals into text. However, such systems heavily rely on large amounts of labeled speech data, which severely limits model training performance and generalization, especially in low-resource language environments. To address this issue, this paper proposes an end-to-end speech recognition approach based on deep fusion, which tightly integrates an external language model (LM) with the end-to-end model during the training phase, effectively compensating for the lack of linguistic prior knowledge. Unlike traditional shallow fusion methods, deep fusion enables the model and the external LM to share representations and jointly optimize during training, thereby enhancing recognition performance under low-resource conditions. Experiments conducted on the Common Voice dataset show that, in a 10 h extremely low-resource scenario, the deep fusion method reduces the character error rate (CER) from 51.1% to 17.65%. In a 100 h scenario, it achieves a relative reduction of approximately 2.8%. Furthermore, ablation studies on model layers demonstrate that even with a reduced number of encoder and decoder layers to decrease model complexity, deep fusion continues to effectively leverage external linguistic priors, significantly improving performance in low-resource speech recognition tasks.