Abstract:In the field of precision measurement and aerospace, the magnetic-grating-like displacement sensor has high requirements on the uniformity of magnetic field generated by permanent magnet. While the magnetic field uniformity of common rectangular permanent magnet is poor, which affects the high precision and reliable measurement of hydraulic cylinder displacement. Therefore, this paper designed a curved permanent magnet with a unique structure. First based on the molecular current hypothesis and the Biot-Savart Law, to derive the expression of the magnetic induction intensity of a curved permanent magnet at any point in space; then, the structure size of the permanent magnet was optimized using the Maxwell electromagnetic field analysis software, to determine the basic size of the curved permanent magnet; finally, the experimental platform is built to measure the magnetic induction intensity at the corresponding position of the permanent magnet. By comparing the measured data and Maxwell simulation data, the results show that the simulation is basically consistent with the experimental results. The curved permanent magnet can produce a uniform magnetic field, and the relative error of the uniform field does not exceed 3%. In the practical application of the Magnetic-grating-like displacement sensor, the sensitive element tests the uniform magnetic field, ensuring the signal quality for the subsequent subdivision of the sensor, thus improving the test accuracy of the sensor, which further proves the practical significance of the arc permanent magnet.

Abstract:CAT-1 is the first domestic magnetic confinement fusion device to use a magnetic levitation dipole field magnet design. Based on the overall goals and parameter design requirements of the CAT-1 device, this paper uses a simplified line current model, and employs vector magnetic field, mechanical equilibrium, and dynamics methods to complete the equilibrium stability analysis of the superconducting magnetic levitation system, and provides parameter design results. A simplified permanent magnet levitation experimental device has been preliminarily designed to verify the stability of the superconducting magnetic levitation system and the reliability of the parameters. The results show that for the CAT-1 device′s levitation magnet weighing 1 200 kg, current of 5 MA, and levitation distance of 2.0 m design goals, the optimal value of the lifting coil radius is 1.7 m, and the corresponding current is 3.49 kA. In order to effectively reduce and control the displacement of the levitated magnet, the working area near the equilibrium point should be restricted to vertical displacement| Δz| < 0.1 m, horizontal displacement |Δer| <0.05 m, and tilt angle displacement |Δα|<π/24. The influence of TSR coil on the suspended coil is analyzed. The calculation shows that TSR and the suspended coil drift in the opposite direction, the drift amplitude is related to the radial position, and the originally closed magnetic field line on the side of the TSR coil is destroyed, resulting in the loss of the transported particles.A conceptual design of a levitation experiment device using 1.5 kg permanent magnets was completed, and the analysis showed that the distance between the permanent magnet and the copper levitation coil of 0.1 m, levitation coil current of 895 A satisfying the requirements for control response speed.

Abstract:Static contact force is an important index to evaluate the performance of pantograph, and also an important factor to measure the contact quality of pantograph, which directly affects the safety of train running. Traditional measurement methods have problems such as low efficiency and poor accuracy. Based on the idea of cloud, tube and terminal of the internet of things, this thesis develops a vehicle-mounted pantograph performance detection device, and proposes an adaptive segmental data fitting algorithm to optimize detection accuracy. The algorithm performs data segmentation and curve model optimization at the same time, and realizes adaptive segmentation at the best point by gradually expanding interval length and evaluating five fitting models: polynomial, exponential, Gaussian, Fourier and power function. The experimental results show that the average error rate of the measuring device is reduced from 1.91% to 0.21%, which is better than other matching methods in the thesis and shows a good precision compensation effect.

Abstract:To address the issues of uneven dispersion and low coverage rates that arise from the random deployment of nodes in wireless sensor networks, an improved snake optimization algorithm for WSN coverage optimization is proposed. First, Circle mapping is utilized for population initialization to bolster the diversity of the population. Moreover, an adaptive spiral search algorithm is employed during the exploration phase of the snake to extend the search range. Then, the introduction of pheromones from the black widow algorithm mitigates the tendency of the snake’s exploitation phase to fall into local optima. Note that the use of a differential evolution strategy enhances the capability for optimization. Finally, by applying the improved snake optimization algorithm to the deployment of sensor nodes and using the coverage model of sensors, the maximum coverage rate is determined. Experiments indicate that the improved algorithm can effectively enhance the node coverage and expand the WSN coverage area to reduce the node energy consumption and extend network lifetime.

Abstract:To achieve high fault tolerance of wireless power transfer system during coupler offset and ensure output stability, a wireless power transfer system structure with high anti-offset performance based on dipole coils has been proposed in this paper. Firstly, a new type of dipole coil has been proposed, which has been studied using finite element method and its magnetic pole direction and structural size are optimized and analyzed. Secondly, based on the new dipole coil, the characteristics of input current, output power and efficiency of LCC-S and S-S compensation topologies were compared when the input voltage is constant. Finally, experimental researches were conducted on the new dipole coil under the same conditions of leeds-line quantity and size, compared with the square coil, the designed dipole coil has a reduction of 8.51% in coupling coefficient change rate, 11.12% in power change rate when the offset range of the X and Y axes is -60 to +60 mm, and the efficiency always greater than 80% during the offset process.

Abstract:Aiming at the problem of sensorless control in the full speed domain of permanent magnet synchronous motor, it is difficult to estimate the position of the observer when the motor starts. A sensorless control strategy based on current-frequency ratio combined with Romberg observer is proposed. First, in the low-speed domain, I/F control is used to start; in the medium-high speed domain, Romberg observer and phase lock loop are introduced for position estimation. Secondly, to solve the oscillation problem caused by direct transition between two control methods, a new smooth transition strategy is proposed. For the problem of poor coordination and anti-interference ability of PI regulator, ADRC is introduced and combined with a new smooth transition strategy to form a compound transition method. Finally, through Simulink and motor experiment platform verification, simulation and experiment results show that the transition process of the composite control method is stable and the speed overdrive and error are significantly reduced.

Abstract:Due to the special application environment of the power distribution room inspection robot, the traditional heuristic algorithm used for multi-objective point path planning may deteriorate the solution results, thus failing to obtain the globally optimal solution in practical applications. In response to the above issues, this paper proposes a multi-objective point path planning algorithm based on the serial fusion of improved grey wolf optimization and A*. Firstly, the pre-A* algorithm is used in conjunction with the grid distance formula to calculate the grid distance between any two target points. Then, an improved grey wolf algorithm, which has modified the input variable encoding method and convergence factor formula, is adopted to plan the optimal cruise sequence vector for multiple target points. Finally, the path between adjacent target points in the optimal cruise sequence vector is planned in sequence, and the globally closedloop planning path for multiple target points is finally obtained. The simulation results show that the optimal path length obtained by the improved grey wolf algorithm is reduced by a maximum of 18.1% compared with the traditional grey wolf algorithm. Compared with the traditional single algorithm optimization, the fusion algorithm reduces the traversed grids by 808 and the path length by 76.4%.

Abstract:To address the problems that the control performance of conventional sliding mode speed control deteriorates drastically with the changes of the operating parameters of permanent magnet synchronous motor, resulting in large speed jitter, current harmonics and torque pulsations, a novel adaptive fractional-order reaching law sliding mode controller is designed. Firstly, a novel fuzzy adaptive fractional-order reaching law is designed by combining the fractional-order theory and the general power convergence law function, and the level of the adaptive parameters is analysed by orthogonal experiment. Secondly, the parameters are adjusted online using a fuzzy controller to improve the robustness of the controller. Since the introduction of the fuzzy control introduces more current harmonics, the current predictive control with optimal duty cycle is used instead of the traditional current inner loop PI controller. Finally, the simulation model is built and compared with the traditional power reaching law sliding mode controller, and the simulation results show that the three-phase current harmonics of the proposed controller in this paper are reduced by 7.93%, and the speed landing after sudden load addition is reduced by 31.2%, which makes the robustness stronger.

Abstract:Aiming at the problems of poor response speed, insufficient anti-disturbance performance and poor observation accuracy of traditional sliding mode observer in maximum power tracking of offshore wind turbines, a control strategy of improved dual-loop anti-disturbance control and adaptive full-order sliding mode observer is proposed. The fal function is replaced by a smooth continuous function to weaken the jitter; the adaptive rate is designed to obtain a smoother back electromotive force, and a phase-locked loop algorithm is introduced to estimate the rotor speed and position angle. Comparative experiments show that the strategy proposed in this paper improves the speed tracking response speed by about 24.3%, the maximum speed error is reduced from 0.70 to 0.34 rpm, and the maximum rotor position error is reduced from 0.045 to 0.012 rad. The observation accuracy is also significantly improved in steady state. Under this strategy, the system anti-disturbance performance is enhanced, the response speed is improved, and the observation accuracy is significantly improved, which improves the maximum power tracking effect.

Abstract:Pedestrian streets, as emerging urban structures, have made life more convenient but also pose significant fire risks. To address the problem of evacuating people during sudden fire incidents in commercial pedestrian streets, a path planning algorithm based on an improved ant colony algorithm in a smoky environment is proposed. Firstly, the Gaussian plume model is used to calculate the concentration of smoke on the roads. Based on this, the equivalent distance is used instead of the Euclidean distance to quantify the harm of gases to humans. At the same time, a modified heuristic function is proposed by considering the impact of crowd density on speed. Given that traditional ant colony algorithms exhibit slow convergence, a tendency to get trapped in local optima, and an excess of redundant nodes in path planning, the A* algorithm is further integrated to adjust the initial pheromone concentration of the ant colony algorithm. The path selection and pheromone update rules are also improved, and a deadlock prevention mechanism is introduced, enhancing the global search capability and increasing search efficiency. Finally, the obtained path is smoothed to reduce the extra path length caused by redundant nodes. Simulation experiments have validated that the algorithm not only significantly improves performance but also effectively plans escape routes according to the fire environment.

Abstract:In univariate time series classification tasks, effectively utilizing the multi-scale and time-dependent features of time series is crucial for enhancing classification accuracy. Addressing the limitations in existing models regarding the comprehensive use of multi-scale and time-dependent features, this paper introduces a new hybrid model MHAGRU-MCCE that combines the multi-scale conditional convolution and enhancement (MCCE) module with a multi-head attention mechanism based GRU (MHAGRU). MCCE captures rich temporal features across different scales, while MHAGRU focuses on extracting the dependency relationships within the time series data. On 85 public datasets from UCR, comparative validation with six mainstream deep learning-based time series classification models, including MACNN, AFFNet, OS-CNN, LITETime, MLP, and LSTM-FCN, demonstrates that MHAGRU-MCCE achieves respective improvements in mean accuracy (MA) of 0.66%、2.04%、3.45%、2.70%、12% and 2.89%. It also achieved the highest arithmetic mean rank (AMR)=2.45 and geometric mean rank (GMR)=1.98, fully demonstrating the effectiveness and superiority of MHAGRU-MCCE in handling univariate time series classification problems.

Abstract:Aiming at the problem of target misdetection and missing detection caused by large size changes and mutual occlusion in UAV aerial images, a lightweight target detection algorithm based on YOLOv8s is proposed by integrating multi-scale features. In the backbone network, lightweight multi-scale convolutional EMSC is used to replace Bottleneck in C2f module, which enhances the expression ability of multi-scale features. The lightweight upsampling operator Dysample is introduced into the neck network to capture the fine features of the image. Task Aligned Assigner hyperparameters are optimized to solve the problem of sample imbalance during training. Finally, the system of visual interface is designed, and the object of aerial photography is detected by visual interface. The simulation on the data set VisDrone2019 shows that the accuracy and recall rate of the algorithm are improved by 2.4% and 3.3% respectively compared with the benchmark algorithm, and mAP0.5 is improved by 3.5%, effectively improving the effect of aerial photography target detection. The model generalization experiment is carried out on UAVDT data set, and the effect is better than other classical algorithms.

Abstract:To solve the problems of difficulty in segmentation of complex areas, loss of edge details, and insufficient generalization ability in polyp segmentation by existing models.This paper proposed a polyp segmentation model based on fusion of local and global features.Convolutional neural network and Transformer are used as parallel encoders to make the model take into account both the local detail features and global semantic features of multiple scales.The attention enhancement block and the multi-scale residual block are constructed at the jump junction. The former enhances the model′s focus on important information, while the latter efficiently explores the target regions and accurately predicts theirs boundaries, while promoting the interaction between different levels of features.The residual-based stepwise upsampling feature fusion method is used in the decoding stage to gather the features of each stage, which further enhanced the perception ability of the model and enriched the polyp features.Finally, the efficient prediction head is used to promote the fusion of shallow features and output the segmentation results.The model performs best in the comparative experiments. Compared with the sub-optimal model, on the Kvasir and CVC-ClinicDB datasets, it achieved an average mDice improvement of 1.21% and an average mIoU improvement of 1.82%; on the CVC-ColonDB and ETIS datasets, it achieved an average mDice improvement of 2.67% and an average mIoU improvement of 2.83%. The experimental results show that the proposed model has better segmentation accuracy and generalization performance than the existing mainstream models.

Abstract:Aiming at the problem of low accuracy and poor classification effect of small size defect detection of offshore wind turbine blades, an improved early defect detection model of offshore wind turbine blade surface based on EfficientNet is proposed. Firstly, the asymmetric convolution is introduced into the EfficientNet feature extraction network to replace the ordinary 3 × 3 convolution, which enhances the convolution kernel skeleton information and improves the ability of the network to extract defect information. Secondly, a hybrid spatial channel attention module is proposed to focus on space and channel information, and the BiFPN feature fusion module is used to fuse the semantic information of different depths to improve the multi-scale feature fusion ability of the algorithm. Finally, Focal-EIOU and Focal Loss functions are introduced to calculate the position loss and classification loss, so as to improve the positioning accuracy and solve the problem of imbalance between positive and negative image samples in the model training process. The experimental results show that the average accuracy of the proposed algorithm model is 97.6%, and the detection performance of early defects on the surface of wind turbine blades is significantly improved.

Abstract:FOD poses a great threat to airports. Detecting and removing FOD accurately and timely is the key point of airports’ safety work. A FOD detection algorithm that based on YOLOv7 is proposed to meet the requirements of accuracy and real-time. Firstly, the CBAM module is introduced into the backbone network to focus on the extraction of small target feature information from two aspects: spatial attention and channel attention. Secondly, the idea of AFPN is integrated into the feature extraction network and SA-PANet structure is proposed in combination with it. SA-PANet can asymptotically fuse adjacent effective feature layers and alleviate the semantic gap between them. Thirdly, the BiFormer module is introduced into the down-sampling branch of PANet, which can focus on further fusion extraction of small target feature information in the feature extraction network. Lastly, MPDIoU Loss is introduced into the boundary frame positioning loss calculation, which can not only accelerate the convergence of the model but also improve the detection accuracy and location accuracy. Experiments on FOD datasets show that the mAP50 of the improved YOLOv7 algorithm is 98.76%, which is 9.09 percentage points higher than the original YOLOv7. Comparing with other algorithms for FOD detection, the improved YOLOv7 algorithm has higher detection accuracy and the increase of Params and GELOPs is controlled within the acceptable range, which meet the accurate and fast requirements of FOD detection tasks.

Abstract:Aiming at the problems of poor reconstruction completeness and insufficient generalization ability of multi-view stereo reconstruction in complex scenes such as uneven illumination, weak texture, and non-Lambertian surfaces, this paper proposes a multiview stereo reconstruction algorithm based on the attention mechanism. In the feature extraction stage, the algorithm adopts a multi-scale feature extraction module based on depth-separable convolution and self-attention mechanism, which enhances the spatial feature relationships among multiple views while expanding the sensory field, thus improving the network′s ability to characterize features in complex scenes to achieve more accurate feature matching. In the cost volume regularization stage, this paper introduces the channel attention mechanism to adaptively adjust the weights of different channels, so as to reduce the interference of irrelevant information on the model and filter the background noise to improve the generalization ability of the model. On the DTU dataset, the completeness and overall metrics of this paper′s algorithm are 0.286 and 0.334, respectively, which are improved by 25.71% and 5.92% compared to the benchmark algorithm CasMVSNet. The structure of the reconstructed point cloud is also more complete in complex scenes compared to other state-of-the-art (SOTA) algorithms. On the Tanks and Temples intermediate dataset, the reconstructed point cloud composite index F-score is 61.49, indicating that the algorithm in this paper has better robustness and generalization ability.

Abstract:Tire defect detection is of great significance for the identification of tire safety performance, and researching high-performance tire anomaly detection methods is extremely important for the safety performance of automobiles. This article proposes a network model, UDGANomaly, based on U-Net discriminators, which is based on generative adversarial networks. Firstly, encoding and decoding are introduced into the discriminator. The encoder module performs image-by-image classification, and the decoder module outputs pixel-by-pixel classification decisions, providing spatially coherent feedback to the generator. Secondly, a self-attention mechanism is introduced in the encoder and decoder of the generator to further focus on the representative information contained in multi-scale features. Finally, an improved generator loss function based on structural similarity was designed to address visual inconsistency and enhance the robustness of irregular texture detection. After comparative research, it was found that the network structure proposed in this paper has significantly better anomaly detection performance than other traditional network models on the same tire dataset, and the average testing accuracy is as high as 95.6%.

Abstract:In response to the challenges of off-road and unknown outdoor environments, which lack rich prior information and exhibit highly variable road types without clear boundaries, the paper explores a method for detecting drivable areas based on road topological constraints. Firstly, in order to meet the requirements of global task planning, road skeletons are semi-automatically extracted from satellite images to construct a global road network and plan the global trajectory. Secondly, the current driving direction, which is derived from the road topology network and real-time positioning, serves as a strong constraint for the adaptive growth of the drivable area, ensuring that the extension of the drivable area aligns with the road patterns. Finally, a bayesian-based method is proposed for the fusion of LiDAR and visual data to enhance the robustness of drivable area detection. The experiments show that the proposed method significantly improves road pattern consistency and detection rate. The detection loss is 0.2‰. The average proportion factor and average divergence factor under straight road is 92.14% and 12.38%.The average proportion factor and average divergence factor under winding road is 85.46% and 20.75%.

Abstract:Aiming at the problem that it is difficult to fully extract deeper features using single-scale convolution and traditional ReLU activation function when using deep learning models to identify EEG signals of adolescent schizophrenia. Put forward a kind of multi-scale convolutional neural network model with adaptive ReLU(MSAPNet) for adolescent patients with schizophrenia and healthy adolescent brain electrical signal classification. Firstly, a multi-scale cascade module is used to extract the input 3D feature matrix containing the original EEG spatial information. Secondly, the features at different levels were fused through the designed feature fusion module. Multi-scale down sampling module is then used to decrease the dimension of feature maps. Finally, using the classification module to complete identification and detection of disease. The experimental results show that the MSAPNet of disease identification accuracy, sensitivity, specificity and accurate rates and F1 score can be achieved respectively 97.21%, 97.51%, 96.86%, 97.29% and 97.40%, compared with the related research has better detection performance, proved the effectiveness of the proposed method.

Abstract:Aiming at the deficiency of using empirical threshold to detect the starting and ending points of active segment of surface electromyography (sEMG) signal in traditional methods, an adaptive detection method of sEMG starting and ending points based on time-frequency point density is proposed. Time-frequency point density is innovatively proposed as the characteristic parameter of surface EMG signal in this method. Firstly, butterworth bandpass filtering and wavelet threshold denoising are used to preprocess the sEMG signals in Ninapro DB8 dataset. Short-time Fourier transform is used for time-frequency analysis of signals. Secondly, the sEMG signal is divided into several continuous unit time-frequency windows, the number of frequency points in the windows is counted, and the time-frequency point density (TFPD) characteristic parameters are extracted. Finally, the TFPD results are adaptively normalized in the interval ［-1,1］, and the start and end of EMG signals are detected by using the binary judgment method based on sliding window. The experimental results show that this method can detect the start and end of sEMG signal activity segment in quasi-real time within 0.5 s, and the accuracy is nearly 100%. Compared with other common algorithms, the proposed method has better accuracy. The influence of individual differences can be eliminated through normalized positive and non-positive values, and the adaptability is strong. In addition, the proposed method has strong practicability in gesture recognition system.

Abstract:To address the issues of insufficient initial population diversity, weak global optimization capability in the early stages, low convergence accuracy in the later stages, and susceptibility to local optima in the snake optimizer (SO) for solving robotic path planning problems, a multi-strategy enhanced snake optimizer (MSESO) is proposed. The MSESO utilizes a good point set method to initialize the snake population, thereby increasing the diversity of the initial population and ensuring more comprehensive coverage of the search space. It introduces two oscillation factors to balance the process of global search and local exploitation, dynamically updating the search range. Additionally, the integration of an adaptive elite opposition-based learning strategy effectively leverages valuable information from the population to improve population quality, enhancing the likelihood of approaching the optimal solution, accelerating the algorithm′s convergence speed, and improving convergence accuracy. The MSESO is applied to robotic path planning, beginning with ablation experiment to verify the effectiveness of the proposed strategies. Subsequently, comparative experiments on maps of varying complexity are conducted to assess the pathfinding performance of MSESO against other algorithms, demonstrating the superiority of the improved algorithm. Ablation experiment results show that the proposed strategies in MSESO significantly enhance path planning performance. Comparative experiment results indicate that MSESO outperforms the control algorithms in terms of average path length, path length variance, and average number of iterations, validating the robustness and superiority of MSESO in path planning.

Abstract:To monitor the health status of the boiler in real-time and accurately obtain the temperature and stress situation of the boiler pipeline, a fiber optic grating boiler status monitoring technology based on deep learning is proposed. A sensor structure with dual fiber Bragg grating cascaded packaging and its fixing method have been designed to improve the measurement performance of the sensor. A feature fusion parallel transformer regression prediction model was constructed to process the temperature and strain signals of sensors, achieving accurate recognition of the temperature and strain of sensing units. The experimental results show that the sensitivity of the two gratings in the sensor to temperature is 12.31 pm/℃ and 11.63 pm/℃, and the sensitivity to strain is 1.2 pm/με and 0, eliminating the influence of temperature on strain measurement, with temperature compensation effect. By introducing deep learning algorithms, the difficult problem of high-order mixing terms in the sensitivity of fiber Bragg gratings to temperature and strain mixing in high-temperature environments has been solved. The model′s coefficient of determination is greater than 0.9, and the average absolute error and mean square error are 0.23 and 0.31, respectively, effectively improving the sensor′s recognition accuracy for temperature and stress. In summary, this technology has achieved accurate measurement of temperature and strain in high-temperature environments, making up for the shortcomings of traditional measurement methods such as high-temperature failure and single point measurement. It provides an effective solution for real-time monitoring of boiler working health status.