Millimeter-wave radar, with advantages such as non-contact penetration detection and privacy protection, has become a promising solution for unobtrusive monitoring in the field of smart elderly care. To solve the problem of whether there are human body in the elderly care scene, this study proposed a method for judging the presence of a human body based on adaptive dual thresholds to reduce invalid vital sign detection in an empty environment. This method used a low-frequency energy ratio as the core judgment basis. It combined adaptive thresholds to accurately judge the presence of human targets, effectively reducing false detections caused by background interference. In addition, given the defect that variational mode extraction (VME) needs to rely on manual parameter adjustment based on empirical values, the crown porcupine optimization (CPO) algorithm is introduced to optimize the VME parameters adaptively, and the optimized VME is used to reconstruct the heartbeat signal to improve the signal purity. Then, the multiple signal classification (MUSIC) algorithm was used for spectrum analysis to improve the accuracy of heart rate estimation. The results show that in the experimental judgment of personnel, the miss rate in the case of personnel presence is 2.2%, and the false alarm rate in the case of no personnel is only 3.2%; the root mean square error and mean absolute error of the proposed heart rate (HR) estimation method are reduced by 4.4 beat per minute and 3.05 beat per minute respectively compared with the traditional VME, verifying its excellence.

Motion intention recognition is considered the key technology for enhancing the training effectiveness of upper limb rehabilitation robots for stroke patients, but traditional recognition systems are difficult to simultaneously balance real-time performance and reliability. To achieve real-time and accurate upper limb motion intention recognition, a multi-modal fusion method based on surface electromyography (sEMG) signals and arrayed flexible thin-film pressure (AFTFP) sensors was proposed. Through experimental tests on 10 healthy subjects (5 males and 5 females, age 23±2 years), sEMG signals and human-machine interaction force (HMIF) signals were collected during elbow flexion, extension, and shoulder internal and external rotation. The AFTFP signals based on dynamic calibration compensation and the sEMG signals were processed for feature extraction and fusion, and the recognition performance of single signals and fused signals was compared using a support vector machine (SVM). The experimental results showed that the sEMG signals consistently appeared 175±25 ms earlier than the HMIF signals (p<0.01, paired t-test). In offline conditions, the recognition accuracy of the fused signals exceeded 99.77% across different time windows. Under a 0.1 s time window, the real-time recognition accuracy of the fused signals was 14.1% higher than that of the single sEMG signal, and the system’s end-to-end delay was reduced to less than 100 ms. The AFTFP sensor is applied to motion intention recognition for the first time. And its low-cost, high-density array design provided an innovative solution for rehabilitation robots. The findings demonstrate that the AFTFP sensor adopted in this study effectively enhances intention recognition performance. The fusion of its output HMIF signals with sEMG signals combines the advantages of both modalities, enabling real-time and accurate motion intention recognition. This provides efficient command output for human-machine interaction in scenarios such as stroke rehabilitation.

Microsphere assisted microscopy (MAM) has been rapidly developed to meet the measurement needs of microstructures. MAM can be integrated with optical interference microscopy (OIM) to achieve high lateral resolution surface profile measurement. However, the microspheres introduce intricate phase changes, resulting in optical path asymmetry which is very challenging to compensate for. This limitation constrains the application of MAM in OIM. In this paper, simulation analysis reveals that the phase transmission of the microsphere is influenced by parameters such as microsphere diameter and its relative position to the sample. It is concluded that a unique compensation process must be adopted for each individual microsphere. Addressing this issue, we proposed a phase compensation algorithm based on the three-dimensional position control of the microsphere and integrated it into our combined system of MAM and white light interferometry (WLI), reducing the phase errors introduced by the microspheres while enhancing the lateral resolution of optical system. This approach improved the profile measurement accuracy, offering a perspective for optically measuring the surface profile of intricate microstructures.

To address the stochasticity and nonlinearity of solar collector power systems, a soft sensor prediction model with a hybrid convolutional neural network (CNN) and long short-term memory network (LSTM) was constructed, and the hyperparameter optimization of the hybrid neural network (CNN-LSTM) was carried out by using the sparrow search algorithm (SSA). The model utilized the powerful feature extraction and non-linear mapping capabilities of deep learning to effectively handle the complex relationship between input and target variables. The batch normalization technique was used to speed up the training and improve the stability of the soft-sensing model, and the random discard technique was used to prevent the soft-sensing model from overfitting. Finally, the mean absolute error (MAE) was used to assess the accuracy of the soft sensor model predictions. This study compared the proposed model with soft sensor prediction models like Bp, Elman, CNN, LSTM, and CNN-LSTM, using dynamic thermal performance data from the solar collector field of the molten salt linear Fresnel photovoltaic demonstration power plant. The deep learning-based soft sensor model outperformed the other models according to the experimental data. Its coefficients of determination (namely R²) are higher by 6.35%, 8.42%, 5.69%, 6.90%, and 3.67%, respectively. The accuracy and robustness have been significantly improved.

Aiming at the problem of inaccurate crowd counting and location in dense scenes, a dynamic region-sensing crowd counting and location method based on high-resolution fusion was proposed. Firstly, U-HRNet was used as the main backbone to extract high-resolution features of the population and enhance the ability of feature extraction with different resolutions. Then, the dynamic regional awareness attention module was designed to make full use of the global and local feature information, refine the differentiated learning of target feature and background feature, reduce the interference of background feature, and improve the positioning performance of the model. Finally, the predicted threshold map and confidence map were input into the binarization module to output the prediction and counting results of the crowd independent individual target. Experimental results showed that the proposed method achieved good performance of counting and positioning in different scenarios.

Accurate measurement of bean particle size is essential for automated grading and quality control in agricultural processing. However, existing image segmentation methods often suffer from low efficiency, over-segmentation, and high computational cost. We proposed a distance-gradient dual constrained watershed algorithm for precise segmentation and measurement of bean particles. The method integrated distance transform-based seed extraction with gradient-constrained flooding, effectively suppressing noise-induced region fragmentation and improving the separation of adherent particles. An experimental platform was constructed using an industrial camera and an image-processing pipeline to evaluate performance. Compared with the conventional watershed algorithm, the proposed method improves segmentation accuracy by 7.2% and reduces the mean particle size error by 27.8% (0.13 mm, representing a relative error of 2.4%). Validation on three soybean varieties confirmed the robustness and generalizability of the approach. The results indicated that the proposed algorithm provided an efficient and accurate technique for agricultural particle size analysis, offering potential for integration into practical low-cost inspection systems.

For image compression sensing reconstruction, most algorithms use the method of reconstructing image blocks one by one and stacking many convolutional layers, which usually have defects of obvious block effects, high computational complexity, and long reconstruction time. An image compressed sensing reconstruction network based on self-attention mechanism (SAMNet) was proposed. For the compressed sampling, self-attention convolution was designed, which was conducive to capturing richer features, so that the compressed sensing measurement value retained more image structure information. For the reconstruction, a self-attention mechanism was introduced in the convolutional neural network. A reconstruction network including residual blocks, bottleneck transformer (BoTNet), and dense blocks was proposed, which strengthened the transfer of image features and reduced the amount of parameters dramatically. Under the Set5 dataset, when the measurement rates are 0.01, 0.04, 0.10, and 0.25, the average peak signal-to-noise ratio (PSNR) of SAMNet is improved by 1.27, 1.23, 0.50, and 0.15 dB, respectively, compared to the CSNet+. The running time of reconstructing a 256×256 image is reduced by 0.147 3, 0.178 9, 0.231 0, and 0.252 4 s compared to ReconNet. Experimental results showed that SAMNet improved the quality of reconstructed images and reduced the reconstruction time.

Aiming at the problem of insufficient feature extraction in single scale neural network model and the problem that convolutional neural network cannot process sequential tasks in the classification of EEG signals in depression, a hybrid model (BFTCNet) of dual-branch convolutional neural network (Bi_CNN) and temporal convolutional network (TCN) based on feature recalibration (FR) was proposed to classify EEG signals of depressed patients and healthy controls. Firstly, Bi_CNN module was used to extract the mixed EEG features between different frequency bands and different channels. Secondly, FR module was used to enhance the features extracted by Bi_CNN. Finally, TCN with dilated causal convolution was used for the sequence learning to capture the temporal dependency between features. In this study, 128 EEG channels of resting-state (closed-eye) EEG data from the public dataset MODMA were used as experimental data, including 29 healthy controls and 24 depression patients. The performance of the model was evaluated by the 10-fold cross validation method. The proposed BFTCNet achieves a classification accuracy of 95.98%, F1 score value of 95.47%, sensitivity and specificity of 94.21% and 97.50%, respectively. Compared with the single-scale network model EEGNet-8,2, the classification accuracy and F1 value are improved by 1.5% and 1.48%, respectively. Meanwhile, the ablation experiment proved that each sub-module had its contribution to the improvement of the model’s classification ability .

In camera calibration, accurate estimation of homography matrix between the world coordinates of the calibration board and its image coordinates is a key step in high-precision calibration of intrinsic camera parameters. The existing homography matrix estimation methods have problems such as dependence on thresholds, low computational efficiency, and initial model or sorting quality affecting results. In this paper, a homography matrix estimation method based on adaptive genetic algorithm was proposed. Firstly, a new circular grid calibration board was designed and the strategy of first sampling of data sets was optimized. Secondly, a mathematical model for the estimated homography matrix was established according to the adaptive genetic algorithm. Thereby the optimal homography matrix between the calibration board and its image was obtained. Finally, the intrinsic camera parameters were calculated based on Zhang’s calibration method. The experimental results show that compared with the results of three traditional estimation methods RANSAC, PROSAC, and LMEDS, the reprojection error of the images by our estimation method is reduced by about 4.11%—7.85%, 11.94%—16.91%, and 10.19%—17.82%, respectively; and the average running time of the algorithm decreases by about 25.85%—37.47%, 11.99%—22.71%, and 46.50%—53.35%, respectively. In addition, the homography matrix estimation method in this paper was applied to camera calibration. The results show that compared with the traditional estimation method, the average accuracy of the camera during the calibration process increases by about 5.48%, 15.06%, and 11.47%, respectively; and the average calibration efficiency of the camera is improved by about 10.13%, 5.71%, and 14.26%, respectively. The homography matrix estimation method proposed in this paper not only obtained reliable results, but also had certain value and significance in improving the estimation accuracy and calculation efficiency in camera calibration.

To improve the accuracy of indoor localization methods with channel state information (CSI) images, a localization method that used CSI images from selected multiple access points (APs) was proposed. The method had an off-line phase and an on-line phase. In the off-line phase, three APs were selected from the four APs in the localization area based on the received signal strength indication (RSSI). Next, CSI data was collected from the three selected APs using a commercial Intel 5 300 network interface card. A single-channel sub-image was constructed for each selected AP by combining the amplitude information from different antennas and the phase difference information between neighboring antennas. These sub-images were then merged to form a three-channel RGB image, which was subsequently fed into the convolutional neural network (CNN) for training. The CNN model was saved upon completion of training. In the on-line phase, the CSI data from the target device was collected, converted into images using the same process as in the off-line phase, and fed into the well-trained CNN model. Finally, the real position of the target device was estimated using a weighted centroid algorithm based on the model’s output probabilities. The proposed method was validated in indoor environments using two datasets, achieving good localization accuracy.

In response to the problems of low sampling efficiency, strong randomness of sampling points, and the tortuous shape of the planned path in the traditional rapidly-exploring random tree (RRT) algorithm and bidirectional RRT algorithm used for unmanned aerial vehicle (UAV) path planning in complex environments, an improved bidirectional RRT algorithm was proposed. The algorithm firstly adopted a goal-oriented strategy to guide the sampling points towards the target point, and then the artificial potential field acted on the random tree nodes to avoid collision with obstacles and reduced the length of the search path, and the random tree node growth also combined the UAV’s own flight constraints, and by combining the triangulation method to remove the redundant node strategy and the third-order B-spline curve for the smoothing of the trajectory, the planned path was better. The planned paths were more optimized. Finally, the simulation experiments in complex and dynamic environments showed that the algorithm effectively improved the speed of trajectory planning and shortened the length of the trajectory, and could generate a safe, smooth and fast trajectory in complex environments, which could be applied to online trajectory planning.

Fluxgate current sensors (FGCSs) are increasingly employed in power systems due to their high-precision characteristics, yet their measurement flexibility remains constrained by conventional closed-core designs. To address this limitation, we proposed a split-core sensor structure comprising four magnetic core strips, which achieved non-intrusive current measurement while maintaining detection accuracy. An analytical model of the induced electromotive force was established based on the probe’s geometric configuration, followed by finite element simulations to optimize key parameters including core radius, core width, excitation coil turns, and sensing coil configuration. A complete prototype integrating the measurement probe, excitation circuit, and signal processing circuitry was developed and experimentally validated. The experimental results show a sensitivity of 0.109 9 V/A, a hysteresis error of 0.559%, and a repeatability error of 1.574% over a measurement range of ±10 A. After polynomial fitting-based error compensation, the nonlinearity error was reduced to 0.208%, achieving performance comparable to closed-core sensors. This work provided a practical solution for applications demanding both high measurement accuracy and installation flexibility.

This paper presents a new type of ultra-material microwave pressure sensor designed for extreme environments, and conducts a systematic study on its structural design, manufacturing process, working mechanism, and experimental performance. The sensor is based on the cross-slot ultra-material resonant structure. Platinum-based conductive patterns are precisely fabricated on a high-purity alumina ceramic substrate through screen printing, and a strong bond between metal and ceramic is achieved through high-temperature sintering. Thanks to the high-temperature stability of the ceramic material and the high precision of the process, this sensor maintains excellent structural integrity and performance consistency in harsh environments. The working mechanism of the sensor is based on the microstructural deformation induced by pressure. When external pressure is applied to the ceramic cavity, the deformation of the cavity will change the equivalent electromagnetic boundary conditions inside, thereby causing perturbations in the resonant modes of the metamaterial, resulting in a continuous measurable shift in the resonant frequency. Based on this mechanism, the change in pressure can be precisely mapped to the frequency change, enabling wireless and passive pressure measurement. By utilizing the intrinsic resonant radiation of the metamaterial to achieve coupled readings, the complexity of sensor integration is significantly reduced and its working reliability in high-pressure, high-temperature, and strong electromagnetic interference environments is improved. During the design stage, the influence laws of the geometric parameters of the metamaterial and other factors on the resonant performance and pressure sensitivity were analyzed through finite element coupling simulation. Experimental verification shows that the sensor exhibits excellent linear pressure response within the range of 0-500 kPa, and maintains good repeatability and frequency stability in the high-pressure zone. The maximum sensitivity reaches 135 kHz/kPa, and the frequency drift is minimal during multiple loading-unloading cycles, fully demonstrating that the structural strength and reliability of the design meet the engineering requirements. The sensor proposed in this study could achieve long-term stable operation in aerospace engine compartments, high-temperature metallurgical furnaces, deep mine pressure monitoring, petrochemical high-corrosion pipelines, and extreme environment equipment. This research not only demonstrated the potential of integrating metamaterials with advanced ceramic processes to construct wireless passive sensors, but also provided new design ideas and process routes for the engineering application of microwave sensing technology in harsh environments.

Data collected in fields such as cybersecurity and biomedicine often encounter high dimensionality and class imbalance. To address the problem of low classification accuracy for minority class samples arising from numerous irrelevant and redundant features in high-dimensional imbalanced data, we proposed a novel feature selection method named AMF-SGSK based on adaptive multi-filter and subspace-based gaining sharing knowledge. Firstly, the balanced dataset was obtained by random under-sampling. Secondly, combining the feature importance score with the AUC score for each filter method, we proposed a concept called feature hardness to judge the importance of feature, which could adaptively select the essential features. Finally, the optimal feature subset was obtained by gaining sharing knowledge in multiple subspaces. This approach effectively achieved dimensionality reduction for high-dimensional imbalanced data. The experiment results on 30 benchmark imbalanced datasets showed that AMF-SGSK performed better than other eight commonly used algorithms including BGWO and IG-SSO in terms of F1-score, AUC, and G-mean. The mean values of F1-score, AUC, and G-mean for AMF-SGSK are 0.950, 0.967, and 0.965, respectively, achieving the highest among all algorithms. And the mean value of G-mean is higher than those of IG-PSO, ReliefF-GWO, and BGOA by 3.72%, 11.12%, and 20.06%, respectively. Furthermore, the selected feature ratio is below 0.01 across the selected ten datasets, further demonstrating the proposed method’s overall superiority over competing approaches. AMF-SGSK could adaptively remove irrelevant and redundant features and effectively improve the classification accuracy of high-dimensional imbalanced data, providing scientific and technological references for practical applications.

Accurate and real-time fire detection is crucial for industrial production and daily life. However, the variable form of fire and the significant differences in visual characteristics across its different stages pose great challenges to precise fire prevention and control. To address this issue, a multi-scale fire target detection algorithm using YOLO-fire was proposed by improving the YOLOv8 model. This model introduced new layer structures and attention mechanism, replaced new feature fusion modules and loss functions. By introducing a small-target detection P2 layer, the model’s ability to detect early-stage fires is improved. The coordinate attention mechanism is integrated into the layer structures of multi-scale target detection, enhancing the capture of target location information and channel relationships, thereby focusing more on the target regions. The Neck network structure was optimized by adopting a BiFPN_F strategy for different feature layers, which strengthened the cross-scale representation of fire features and controlled the parameter count of the designed model. The WIoU loss function was employed to optimize the regression process, improving fire source localization accuracy in complex scenarios, enhancing model robustness, and increasing detection precision. Experimental results on fire datasets demonstrated that YOLO-fire could effectively detect multi-scale fire targets in various scenarios. Compared to the baseline model (YOLOv8n), YOLO-fire achieves improvements of 1.37% in accuracy, 1.25% in mAP50-95, and 0.35% in F1-score, while reducing parameters by 3.79%. Furthermore, compared to current mainstream target detection algorithms, YOLO-Fire achieved optimal detection performance while reducing network parameters and computational complexity. This research provided effective technical support for fire safety prevention and control in related fields.