Abstract:In view of the need to simulate multiple signals in communication testing, a large-bandwidth, multi-channel signal source hardware architecture is proposed. FPGA is used as the hardware control core and baseband signal generator. The high-speed serial transceiver GTY in FPGA is used to implement the JESD204C interface protocol and complete the single-lane 16.5 Gbps data rate transmission. The CX8242K RF transceiver is used as the RF transmitter and configured through Microblaze to achieve high-quality generation of modulated signals from baseband to high-IF. Frequency planning is carried out for the target frequency band, and clock schemes and multi-channel synchronization schemes are designed. Finally, a domestically produced multi-channel vector signal source with wide bandwidth, high precision and low spurious is realized. Tests have proved that the vector signal source can normally generate various vector modulation signals, and realize carrier frequency adjustment in the range of 10 M~6 GHz, supporting a single-channel transmission signal with a maximum bandwidth of up to 800 MHz, which has high application value in actual engineering tests.

Abstract:Solving the magnetic target parameters (position, magnetic moment) using a three-axis magnetic sensor array is a typical nonlinear least squares optimization problem, and the LevenbergMarquardt (LM) optimization algorithm is often used. However, when solving this problem, it is affected by observation errors such as sensor noise, which leads to less reliable results. In order to reduce the influence of observation errors, measurement noise and other effects on the solution results, in this paper, the principle of geometric accuracy factor is used to optimize the layout of the magnetic sensor array, in order to reduce the ability of noise and other observation errors to interfere with the magnetic parameter solving, and the accuracy of the magnetic target parameter solving is improved by decreasing the mean and variance of the geometric accuracy factor of the observation plane. After simulation experiments, we compare the mean and variance of the geometric accuracy factor of the observation plane with different arrangements of the three sensors, and find a reliable optimal layout scheme for the sensor array: the radius of the reference circle r is 20%~30% of the observation plane; the sensor array is located in the arc of the reference circle and arranged in a positive triangle; and the distance between the plane in which the reference circle is located and the observation plane d is 0. It is verified that when the sensor arrays are not arranged in an optimized layout, the solution deviation ［mx,my,mz］ of the three-axis magnetic moment vectors using the LM algorithm is up to ［0.034 4,0.027 9,0.028 8］A·m2, and that of the positional solution deviation ［x,y,z］ is up to ［3.37,3.14,3.31］cm; the optimized array layout reduces the maximum values of the solution deviation ［mx,my,mz］ and the maximum value of ［x,y,z］ are reduced by ［75.37%,78.66%,76.74%］ and ［72.67%,92.83%,85.76%］ respectively.The optimization of the sensor array layout is instructive for improving the accuracy of magnetic target parameter solving.

Abstract:Aiming at the influence of various parameter changes on the transmission efficiency of the magnetic coupling resonant wireless power transmission system and the problem of low transmission efficiency. With the goal of improving the transmission efficiency of the system, the influence of frequency, resistance, coil mutual inductance and other parameters on the transmission efficiency was analyzed through Matlab simulation software, and the genetic simulated annealing algorithm was introduced to optimize the system operating frequency, coil turns, coil radius, coil distance and load resistance. Finally, the genetic simulated annealing algorithm was simulated in Python software, and a set of optimal solutions corresponding to the system parameters were obtained, so that the maximum transmission efficiency of the system can reach 94.27%. The simulation results show the feasibility and effectiveness of the optimization algorithm, which can effectively optimize the system parameters and thus improve the transmission efficiency of the wireless power transmission system, and has certain practical application value.

Abstract:Aiming at the high cost of the traditional liquid lens control system and the inability to meet the problems of personalised customisation, a control system for automatic focusing of liquid lenses is developed. The system selects ATmega32U4 microcontroller as the main control chip, LTC2662 chip and H-bridge circuit as the current output module for controlling the liquid lens to achieve zoom, and provides two operation modes, manual and automatic, for users to choose. The current output module of this system can output a stable working current of DC -300~300 mA, with an average error of 0.64% and a current stability of 0.327 7%; and the average time consumed by the auto-focusing of the focus control software is about 1.3 s, with an image resolution of up to 45.3 lp/mm. Compared with the existing liquid lens control system, this system realizes the automatic adjustment of the focus of the liquid lens, which is simple in operation, stable in performance, and low in cost. Compared with the existing liquid lens control system, this system can automatically adjust the focal length of the liquid lens with simple operation, stable performance and lower cost, providing more flexible and personalised customisation options for liquid lenses.

Abstract:Aiming at the problems of low efficiency and inconsistent standards in manual visual inspection of aviation lock-wire twisting direction in typical maintenance scenarios, an automatic detection model AFE-YOLOv7 is constructed. Using YOLOv7 as the basic model, the convolutional block attention mechanism CBAM is integrated into the SPPCSPC spatial pooling pyramid block to enhance the network′s ability to pay attention to information between different channels and spatial information; embedding a CA coordinate attention block between the neck network and the head prediction network to enhance the network′s perception of the direction and position information of aviation fastener lock-wire; optimizing the bounding box loss function to Focal-EIoU Loss to improve the robustness of the model. Using a self-built aviation lock-wire twisting directional dataset, the comparative and ablation experiments are conducted on the AFE-YOLOv7 model. The results show that AFE-YOLOv7 achieves the highest accuracy of 83.33%, and compared to YOLOv7, the proposed model has improved accuracy, recall, and mAP values by 7.67%, 8.68%, and 10.25%, respectively; compared with widely used object detection methods such as YOLOv5s, it can better adapt to lock-wire twisting direction detection in multiple scenarios, with a running speed of 30.1 fps, meeting real-time detection requirements, promoting the construction of smart civil aviation.

Abstract:The internal circulating current of modular multilevel converters (MMC) increases system losses and exacerbates distortion of bridge arm currents, thereby increasing system costs. Addressing issues such as oscillations and limited control accuracy in the sliding mode control (SMC) circulating current controller, this paper proposes a fuzzy sliding mode control (FSMC) circulating current controller based on traditional sliding mode variable structure control. Firstly, the special working characteristics of MMC are introduced, and the mechanism of MMC circulating current is analyzed. Then, the control principle of the fuzzy sliding mode circulating current controller is elucidated through modeling analysis and decoupling control. Subsequently, under the same conditions, simulation studies are conducted on MMC systems with traditional PI, SMC, and FSMC circulating current suppression strategies, respectively, under two different operating conditions: AC-side output disturbance and load transient. Finally, through simulation and experimental results, it is demonstrated that the FSMC circulating current controller outperforms the other two circulating current suppressors in terms of both circulating current suppression effectiveness and disturbance rejection capability, thereby enhancing system robustness, stability, and enabling rapid response operation.

Abstract:Aiming at the problem that adversarial samples generated by generative adversarial sample generation methods have low authenticity and poor attack effect, an adversarial sample generation method ACGAN based on AdvGAN and CGAN is proposed. First, attacking a specific target, ACGAN generates targeted adversarial samples in the frequency domain by introducing additional target labels in the training and attack stages. Secondly, the gated convolutional neural network is introduced in the generator and discriminator to help the ACGAN model capture more accurate data features, thereby improving the success rate of the attack. Finally, the perceptual loss function is introduced to minimize the difference in speech feature representation between the model output and the target output, thereby improving the auditory quality of the generated samples. Experimental results show that compared with the existing methods in targeted attacks, the ASR is improved by 1.5%, and the SNR and PESQ are improved by 10.5% and 11.1% respectively, which proves the effectiveness and potential of ACGAN in the field of adversarial sample generation.

Abstract:In order to study the performance of inverters under unbalanced voltage in photovoltaic systems, a new reaching law sliding mode control phase-locked loop is introduced on the basis of virtual synchronous generator control. Firstly, the new sliding mode phase-locked loop is used in the virtual synchronous generator to regulate the positive sequence components of the voltage separated from the decoupled dual synchronous reference coordinate system; then, the positive and negative sequence components of the voltage and the current are separated using the decoupled dual synchronous reference coordinate system and the symmetric component method respectively; and then they are individually controlled and generated as voltage modulated signals. Simulation results show that the phase-locking error of the new sliding mode phase-locked loop are only 0.019 7 and 2.401 5×10-10, which are much smaller than 1.485 3 and 1.640 5 of the decoupled double synchronous phase-locked loop, when the grid voltage drops and the phase jumps occur.The introduction of a new sliding mode phase-locked loop can enhance the stability and robustness of the system and improve the operational performance of PV inverters under voltage imbalance conditions, which provides a certain reference for the design and optimization of PV system inverters.

Abstract:Neural networks possess strong function approximation capabilities and can be applied to Flush Air Data Sensing systems for air data estimation. Addressing the issues of random initial weights and thresholds, local optima during training, and high training data requirements inherent in traditional BP neural networks, an improved particle swarm optimization algorithm-based neural network is proposed to enhance the prediction accuracy of the FADS system. The performance of this algorithm is validated through Computational Fluid Dynamics simulations, using pressure data from aircraft in both conventional flight and high angle of attack flight states. The results indicate that, under conditions of limited training data, the PSO-optimized neural network significantly improves air data prediction accuracy in both flight states. In conventional flight, the prediction errors for static pressure, Mach number, angle of attack, and sideslip angle are reduced by 54.88%、 60.46%、 53.76% and 62.12%, respectively; while in the high angle of attack flight state, the prediction errors are reduced by 71.96%、 47.52%、 66.96% and 53.41%, respectively. Furthermore, with the same data samples, the PSO-optimized neural network exhibits smaller error fluctuations across multiple training runs, demonstrating higher stability and reliability.

Abstract:Steady-state visual evoked potential (SSVEP) is an essential signal type in brain-computer interface (BCI) systems, widely utilized in BCI research due to its high stability and ease of operation. While previous studies have achieved significant progress in SSVEP signal classification, challenges such as low signal-to-noise ratio, non-stationarity, and individual variability still persist. To further enhance the accuracy and practicality of SSVEP classification, this paper proposes a novel neural network architecture—Gam-EEGNet—that combines a global attention mechanism with EEGNet. EEGNet, known for its compact, efficient, and adaptive structure, plays a critical role in SSVEP signal processing. By incorporating a global attention mechanism into EEGNet, Gam-EEGNet can more accurately extract and represent SSVEP signal features, effectively reducing individual variability and noise interference. Experiments were conducted using SSVEP EEG data encompassing 12 different frequencies, and the performance of Gam-EEGNet was compared with that of other mainstream deep learning methods, including CCNN, FB-tCNN, and SSVEPNet. The results demonstrate that GamEEGNet outperforms these methods in terms of classification accuracy and information transfer rate (ITR) across different time windows, particularly achieving a classification accuracy of 86.58% within a short 0.7 s time window. In a 1 s time window, the average recognition accuracy across multiple subjects exceeded 95%, with an ITR above 189 bits/min. Moreover, Gam-EEGNet showed better convergence and stability during training, with faster convergence and lower training errors. These results indicate that Gam-EEGNet offers significant performance improvements in SSVEP signal classification, making it especially suitable for real-time BCI systems requiring rapid response, with broad application potential.

Abstract:To address issues such as the imbalanced allocation of exploration and exploitation, as well as insufficient data utilization in traditional double deep Q-Network algorithms for path planning, an improved DDQN path planning algorithm is proposed. Firstly, the concept of exploration success rate is introduced into the adaptive exploration strategy, dividing the training process into exploration and exploitation phases to allocate exploration and exploitation effectively. Secondly, the double experience pool mixed sampling mechanism partitions and samples experience data based on reward size to maximize the utilization of beneficial data. Finally, a reward function based on artificial potential field is designed to enable the robot to receive more single-step rewards, effectively addressing the issue of sparse rewards. Experimental results show that the proposed algorithm achieves higher reward values, greater success rates, and shorter planning times and steps compared to the traditional DDQN algorithm and the DDQN algorithm based on experience classification and multi-steps, demonstrating superior overall performance.

Abstract:To address the issues of low detection accuracy and poor real-time performance in current fall detection systems, a fall monitoring system based on multisensor information fusion has been designed. The system is centered around the ESP32 microprocessor and utilizes sensors embedded in smartphones, pressure film sensors, and MPU6050 sensors for data collection. Health data is displayed in real-time through a mini-program interface, providing monitoring and alert functions. A collaborative cloud-edge fall detection method has been proposed, combining a local multi-threshold algorithm with an improved SSA-LSTM-Transformer algorithm and data fusion weights in the cloud. This algorithm has been validated on a public dataset, achieving an accuracy rate of 99.13%. Finally, system validation was performed through experiments, and the results showed that the system′s fall detection accuracy is 97.67%. It effectively detects falls and provides real-time positioning and alerts.

Abstract:In order to solve the problems of slow convergence speed of the basic black-winged kite algorithm (BKA) and easy to fall into local optimum, an enhanced black-winged kite algorithm (EBKA) with multi-strategy improvement was proposed. Firstly, the tracking prey location update strategy is introduced to improve the global search ability of the algorithm and accelerate the convergence speed. Secondly, an adaptive t-helix strategy is proposed in the attack stage to prevent the algorithm from falling into local optimum. Finally, in the migration stage, when the leader of the black-winged kite loses its leadership role, the Levy tangent flight strategy is proposed to avoid the premature convergence of the algorithm. In order to verify the improvement effect of the algorithm, 8 test functions were selected for testing and compared with 5 swarm intelligence algorithms. Experimental results show that compared with other swarm intelligence algorithms, EBKA can quickly find the theoretical optimal value of 0 on the single-peak function, converge to the optimal value in about 30 times in the multimodal function F5、F6 and F8， and converge to the theoretical optimal value of 0 in the F6 and F7. It is proved that EBKA has good convergence performance, stability and global optimization ability.

Abstract:In this paper, a method for driving scene modeling and behavior decision-making based on ontology is proposed to solve the problem that autonomous vehicles have difficulty in effective navigation and decisionmaking planning on unstructured roads. First, an ontology model of each element in the unstructured road is established, in which the eight-direction model is used to describe the positional relationship between the unmanned vehicle and obstacles in the road scene. Then, the Cartesian coordinate system of the grid map in the autonomous vehicle is converted into the Frenet coordinate system, and the risk function is defined with the combined spring model as the framework to evaluate the risk value of the vehicle driving in the current scene. Then, the photoelectric information data and prior driving knowledge are integrated to form an ontology knowledge base. Finally, the Prolog inference engine is used to infer the final behavior decision result, which must meet the safety and rationality evaluation. Experimental results show that in unstructured roads, this method can give a decision result that is more in line with the driver′s behavior at the decision level and also performs well in assisting path planning.

Abstract:In modern power systems, instability modes have become increasingly diversified following disturbances, necessitating the accurate identification of various instability modes to implement appropriate control measures and prevent significant losses. Therefore, a transient stability assessment method for power systems based on an improved Swin Transformer is proposed in this paper. Firstly, time-domain simulations are conducted to collect voltage magnitude and phase angle characteristics following disturbances, which are used to construct a feature matrix. Then, building upon the Swin Transformer, a spatial cross-scale convolutional attention module is introduced to replace the original multi-head self-attention module. This new module utilizes a series of convolutional layers with different kernel sizes to effectively extract features across multiple dimensions, leading to more accurate prediction results. Finally, simulation experiments on the modified New England 10-machine 39-bus system and IEEE 50 -machine 145-bus system show prediction accuracies of 99.05% and 99.00%, respectively, with multi-swing instability misjudgment rates of 0.35% and 0.27%. These results demonstrate that the proposed method not only accurately predicts different instability modes but also exhibits superior robustness in the presence of noise and missing PMU features.

Abstract:While image and audio data often dominate fault diagnosis research, the exploration on fault diagnosis of tabular data remains of paramount significance. In the field of tabular fault diagnosis, prior endeavors primarily focused on traditional supervised learning methods, and the evaluation of cross-condition fault diagnosis tasks was insufficient. In this paper, we introduce a self-supervised learning method customized for cross-condition fault diagnosis in tabular data, which combines contrastive learning strategy and tabular masking modeling strategy with a Transformer-based autoencoder architecture. The results of diagnostic instance on the Case Western Reserve University datasets demonstrate that after proper fine-tuning, our method can generally outperform the diagnostic accuracy of the supervised learning baselines in the target tasks. Compared with the self-supervised learning baselines, the introduction of contrastive learning strategy and tabular masking modeling strategy increases the average diagnostic accuracy of the autoencoder by 0.74% and 3.35% respectively in the three target tasks. Furthermore, our comprehensive analysis and discussion on the fidelity and utility of the proposed method serve to demonstrate its rationality.

Abstract:The segmentation and extraction of PV panel region information from infrared images of PV panels can greatly improve the accuracy of PV panel fault detection. However, the traditional semantic segmentation algorithm is not effective in processing the boundary information of PV panels, and there are cases that the boundary of PV panels is wave-like, sticking to each other, and the background is mis-segmented. To solve this situation, this paper proposes a semantic segmentation algorithm model for PV panels based on improved DeepLabV3+, which changes the backbone network to MobileNetV2, introduces the Canny edge detection algorithm to output new shallow feature semantic information, and designs the SE-ASPP module to re-calibrate the feature channels to enhance the network expression capability, and increase the number of channels of shallow feature semantic information to strengthen the attention to shallow feature semantic information. Experimental results show that the precision, mIoU, recall and F1 score of the improved DeepLabV3+ algorithm model reach 99.50%、99.21%、99.61% and 99.55%, respectively, which are 2.24%、1.58%、1.57% and 1.72% higher than the original DeepLabV3+ model, respectively. Improved DeepLabV3+ model performs well in real segmentation tasks and has higher detection accuracy and reliability.

Abstract:In order to solve the problems of low potential recognition rate and poor robustness caused by similar skin tones and lighting changes in complex environments, a dynamic gesture segmentation and recognition method based on elliptical skin color model was proposed. Firstly, the Cr component in the YCrCb color space combined with the OTSU threshold segmentation algorithm was used to segment the hand region. Secondly, in view of the problem that the direct application of morphological processing may lead to the loss of details and affect the accuracy of recognition in the case of different gesture complexity and finger thickness, the traditional Canny algorithm was improved, and the morphological processing was combined with the filling of hand edges. Then, by combining Kalman and the improved CamShift algorithm to track the gestures, the dynamic gesture segmentation is completed. Finally, the segmented dynamic gestures are recognized by the BP neural network, and the implementation of the optimization algorithm on the GPU is used to accelerate the computation-intensive tasks such as image processing, feature extraction and forward propagation of the neural network by using the parallel processing capability of the GPU. This optimization measure significantly improves the real-time performance of the dynamic gesture recognition method, making it better suitable for various application scenarios with high real-time requirements. Experimental results show that the proposed method has strong robustness and anti-interference ability in response to complex background and lighting environment changes, and the average recognition rate can reach 94.67%.

Abstract:Unsafe behaviors of individuals are the most contributing cause of accidents in the laboratories of universities, with improper use of personal protective equipment being the most prominent manifestation. Existing methods are primarily used in construction and industrial settings, focusing solely on whether personnel are wearing protective equipment, without effectively identifying the completeness and effectiveness of the personal protective equipment wearing. This study divides the wearing state of PPE into fine labels and proposes a two-stage personal unsafe behavior detection methods based on object detection and attribute recognition algorithms. In the first stage, the improved YOLOv5-DETR-LPE algorithm is used to achieve the precise detecting of personnel under complex background and low-quality image conditions in the laboratory. In the second stage, the attribute recognition algorithm based on EfficientNet-B3 is used to recognize the unsafe behaviors of detected personnel. The results shows that YOLOv5-DETR-LPE achieves a 1.15% improvement in accuracy and a 5.11% increase in mAP50:95 in the self-built dataset compared to YOLOv5n, with only a slight increase in model parameters and computational load. The EfficientNet-B3 algorithm maintains high accuracy in the recognition of all 11 labels of three attributes. Finally, the recognition and early warning system is designed and implemented in an actual environment, verifying the effectiveness and feasibility of the system in practical scenarios.

Abstract:Aiming at the problem of low detection accuracy of conveyor belt defect detection of belt conveyor due to the lack of public data sets, the diversification of defect shapes and the different lengths of tearing, this paper will use linear array camera and use linear laser as an auxiliary tool in the shooting process to reduce the influence of harsh environment on the image, and put forward an improved YOLOv5 conveyor belt defect detection algorithm to ensure the production safety of coal mine. Firstly, on the basis of the existing data, the method of combining multiple data enhancement methods is extended. Then, in the feature extraction stage, the C3 module in Backbone is replaced with a C3_A similar to the attention mechanism to improve the overall performance. Then, in the feature fusion stage, the shortcircuit method is used to combine the PAN structure of Backbone and Neck to reduce the loss of feature information. Finally, the fine-tuned intersection-union ratio is integrated into the loss function and two parameters are set. The original intersection-union ratio is scaled and cropped, which effectively constrains the position relationship between the model prediction box and the real box, and further improves the accuracy of the model′s boundary box regression. The experimental results show that the average accuracy of conveyor belt defect detection is 88.1%, the accuracy rate is 88%, and the recall rate is 86.5%, which meets the detection requirements of conveyor belt defects.

Abstract:Aiming at the problems of low efficiency and accuracy of manual inspection of medical transparent square cups and wear on products, visual inspection technology is introduced, combining image processing and deep learning, a medical transparent square cup posture and defect detection system is designed and developed. For square cup posture detection, firstly, the edge area of the square cup image is accurately segmented based on the improved dynamic threshold. Then, the high and low threshold design of the Canny algorithm is improved to detect more edges with subtle amplitude differences, and the threshold condition for merging collinear edges is set, and then the two edge straight lines are fitted, and the virtual central axis is calculated based on the two edges. Finally, the center is roughly located by the minimum circumscribed rectangle, and the foot of the rough positioning point to the axis is used as the square cup precise positioning coordinate, and the square cup image is corrected to the reference posture according to the affine transformation. In terms of defect detection, the local defects of the image are grayscale spliced to construct an image data set that is balanced between the three defect classes. After training three neural networks, SqueezeNet, Inception-V3 and ResNet-50, through comprehensive evaluation, it was found that the SqueezeNet model had the best performance, with an average accuracy of 98.6% in the test set, and recognition accuracy and recall rates of 99.8% and 98.8% respectively. The experimental verification results show that the detection speeds for the pose and defects of a single image are 770.5 ms and 553.1 ms respectively. After the improvement, the pose detection accuracy is higher, and the defect detection accuracy rate reaches 94%, which has good real-time performance and stability and can meet the detection requirements. This research can provide technical support for the pose and defect detection of medical transparent square cups.

Abstract:Aiming at the problems of many types of defects on the surface of steel plate, large defect differences, high leakage detection rate, etc., a defect detection algorithm to improve YOLOv9 is proposed. Firstly, the algorithm improves the RepNCSPELAN4 module in the feature extraction network through the FasterBlock in FasterNet, and the RepNCSPELAN4-FB module is designed to realize the multi-scale feature fusion, so as to reduce the number of parameters of the model, and secondly, using the inverse residual structure of iRMB and a kind of highly efficient multi-scale attention module, EMAttention, to combine to form a new iEMA module that improve the accuracy of the network, and finally, using the Inner-WIOU loss function to improve the bounding box regression loss, which improves the model′s detection performance for inhomogeneous distributions and target defects at different scales. Through experiments on the GC10-DET dataset, the improved algorithm improves the precision, recall and map@0.5 by 3.5%、 3% and 2.1% compared with the original algorithm.The model shows good performance in steel surface defect detection.

Abstract:Helicopters are widely used in the field of general aviation due to their excellent low altitude performance. As a key capability for helicopter flight performance, near ground maneuvering flight is an important subject in helicopter flight testing. In order to ensure the safety and controllability of helicopters during ground maneuvering flight tests, real-time monitoring of some key parameters during helicopter flight is necessary. In response to the testing requirements for close range measurement of rotor tail rotor and real-time monitoring and alarm of aircraft near ground height during helicopter flight tests, key technologies such as matrix laser ranging technology and platform control algorithms were studied to determine the design plan for key testing technologies for helicopter near ground flight. Based on this, a helicopter near ground flight safety monitoring and alarm system was developed to achieve high-precision measurement within a hundred meters of ground height during helicopter maneuvering test flights, with accuracy and time delay meeting the requirements of flight test. This has important reference significance for the design of future helicopter flight test systems.