Abstract:In order to improve the working efficiency of special equipment testing workers to measure the flexible value under the main beam of the crane, ensure the measurement accuracy and realize the intelligent processing of measurement data, a measurement method based on horizontal laser and photoelectric sensor is proposed, and the principle prototype of the measuring system is designed. The measuring system transforms the arch value of the crane into the height difference of horizontal laser irradiation on the photoelectric sensor array and improves the measurement resolution by the arrangement of the horizontal beam; ensures the horizontal illumination by the focus instead of the horizontal laser in the result with the function of LoRa and 5G wireless communication technology. In the field of a crane girder detection, the system prototype designed in this paper and the traditional theodolite method contrast measurement experiment, the results show that the measurement method and this paper designed the system prototype compared with the traditional theodolite method, the measurement error of each point are less than 0.6 mm, in the experiment, and the total measurement time reduced by about 70%, ensure the measurement accuracy and reliability, significantly improve the measurement efficiency, implements the measurement data real-time upload to the cloud, improve the intelligent degree of the detection work.

Abstract:The C-field current stability of rubidium atomic fountain clock can affect the second-order Zeeman frequency shift of the clock. Traditional methods to optimize the physical system of the C-field are complicated and difficult to meet the miniaturization requirements. Starting from the circuit system of rubidium atomic fountain clock, this paper puts forward the method of optimizing C-field circuit by using chip current source. Firstly, the influence of the chip current source output fluctuation on the second order Zeeman frequency shift of the rubidium atomic fountain clock is analyzed and the relationship between the second order Zeeman frequency shift of the rubidium atomic fountain clock and the output current of the C-field chip current source is obtained； secondly, the measurement experiment of VC12MA current source is carried out. The experiment shows that when the C field is generated by VC12MA current source, the Allan variance of the output current value is 2.24×10-9, and the relative disturbance to the second-order Zeeman frequency shift of the rubidium atomic fountain clock is 1.78×10-17. The frequency stability of the second-order Zeeman shift of the rubidium atomic fountain clock is improved from the original 10-16 order to the optimized 10-17 order. The method presented in this paper has great application value in the performance improvement and miniaturization of the rubidium atomic fountain clock.

Abstract:In response to the current industrial field, a SocFPGA architecture solution with high-speed interconnect bus is proposed to address the problem of slow processing speed in surface defect detection in current workpiece inspection. Firstly, to remove defects in the image and highlight the defective pixels, the image data is processed by grayscale, median filtering, adaptive threshold segmentation and morphological filtering. Secondly, the pixel threshold algorithm and voting algorithm were used to label the locations of surface defects in the image. Finally, a surface defect detection system for workpieces was built using copper plate samples and OV5640 cameras. Experiments show that when the defect diameter is not less than 0.5 mm, defect detection rate is about 90.24%, and the processing time per frame is about 0.62 μs. It realizes the real-time online detection of workpiece surface defect image, which provides a certain reference for the research in this field.

Abstract:A node expansion method for preventing scratching and collision with obstacles is provided, considering the influence of the robot′s own size and volume during its walking process. The traditional A* search method from the start point to the end point is changed to a simultaneous bidirectional search method from the start point to the end point and from the end point to the start point. The "end point" in the search process is set as the current point of the synchronous search in both forward and reverse directions, and the distance between the current point and the current end point is introduced into the evaluation function, significantly reducing the number of reciprocal searches and the number of search nodes, and improving the search efficiency. Based on the path obtained by the improved bidirectional A* search algorithm, path optimization is carried out, including the removal of redundant points on the path and the arc transition of the inflection point path successively. The simulation results show that the obstacle avoidance path obtained based on the above method has fewer search nodes, high search efficiency, smooth and stable path, and is easy for the robot to complete the obstacle avoidance path walking.

Abstract:A non-isolated bidirectional DC-DC converter is proposed in this paper, which realizes zero ripple of low voltage side current by optimizing the coupling inductor. The proposed converter also has the advantages of high voltage gain (G=(1+D)/(1-D) in boost mode), reduced switch (only three), lower voltage stress (about half of the high voltage side voltage), and zero-voltage-switching for all switches. In addition, it requires only one magnetic core, and the input and output sides share the common ground. The working principle, steady-state characteristics, conditions of low voltage side zero current ripple, soft switching conditions and parameter design methods of the proposed converter are depth analyzed, and then verified by a 250 W/100 kHz simulation prototype.

Abstract:Traditional explosion vibration velocity prediction models are mostly applied in mines and rock mass blasting. There is little research on the surface vibration velocity model of near earth blasting, and the prediction accuracy is low. In order to study the surface vibration intensity model of near earth blasting, a simulation model of near earth blasting surface vibration velocity was established based on LS-DYNA analysis software. Through dimensional analysis, the influence factor of vibration wave propagation velocity was introduced, and an improved model of surface vibration peak velocity was established. The velocity model was subjected to multiple nonlinear regression analysis through simulation data. Finally, the accuracy of the improved model in this paper was verified by experimental application. The results showed that the average relative error predicted by the improved model in this paper was 11.33% through simulation data fitting, while the average relative errors predicted by the two classical models were 34.05% and 31.67%, respectively. The average relative error of the improved prediction model in predicting the vibration velocity of the experimental measurement point was 8.28%, and the prediction errors of both classical models were higher than 44%. Therefore, the improved model in this article has a significant improvement in prediction accuracy compared to existing classical models, and can better characterize the intensity of surface vibration and reflect the attenuation law of peak velocity of surface vibration. It can provide some theoretical basis for predicting the peak velocity model of surface vibration in near earth explosion tests.

Abstract:Aiming at the problems of slow model training speed, large amount of computation and untimely response of quadrotor UAV vision obstacle avoidance system based on deep reinforcement learning, a lightweight and fast model training system is designed. The system first takes the depth image and the UAV′s own state information as input, and then uses a GRU structure-based A3C algorithm (GRU-A3C) to output continuous action space and combine the curriculum learning method for training acceleration. Finally, A3C was used as the baseline for ablation experiments. The experimental results are as follows: after 1 000 rounds of training, the success rate of GRU-A3C algorithm trained using curriculum learning method is 0.28, and the success rate of A3C algorithm is 0.2. After 5 000 rounds of training, the success rate of GRU-A3C algorithm trained using curriculum learning method was 0.72, and the success rate of A3C algorithm was 0.62. The data show that this system can effectively accelerate the model convergence speed, shorten the training time and improve the training effect.

Abstract:Due to the limited terrain conditions, an increasing number of railway lines have to be designed with long inclines, rendering traditional wheel-rail force measurement technology, which only addresses lateral and vertical wheel-rail forces, inadequate for addressing the subsequent operational and maintenance issues. Further research into longitudinal wheel-rail force measurement technology is required to effectively address these challenges. Therefore, this paper introduces a longitudinal wheel-rail force measurement method based on the LM algorithm. Initially, the measurability of longitudinal force is validated and its governing principles are investigated through finite element simulation calculations. Subsequently, a longitudinal force measurement bridge is designed as a foundation for transforming the problem of solving longitudinal force into that of solving an overdetermined multivariate nonlinear equation system. Then, the degree decomposition is used for fitting calibration coefficient under arbitrary angle and the rotation angle and wheel-rail force is solved through the LM algorithm. By employing SIMPACK to simulate train uphill operation under diverse conditions, this study validated the proposed method and quantified the influence of contact point on longitudinal force measurement error. The results demonstrate that the proposed method exhibits high precision, with a relative error of less than 6%.

Abstract:The extraction of the centerline from multi-line structured light is a critical technique in three-dimensional measurement technologies. Reflectivity and other environmental factors on the surface of the object being measured commonly result in low accuracy and instability in extracting the centerline. This thesis proposes an enhanced laser centerline extraction method. It begins by harnessing global features from line laser images, extracted through the Transformer backbone branch within the encoding layer of the SegFormer network. Additionally, the method integrates the Vgg16 backbone branch to capture shallow contour details from the line laser images. The incorporation of the MSASPP module significantly refines the model′s ability to segment linear targets, thus elevating the segmentation accuracy within the laser stripe area. This refined SegFormer network model supplies a superior image source for subsequent centerline extractions, utilizing the Steger method to achieve precise detections. Experimental evidence indicates a 42% enhancement in computational speed over the Steger algorithm, with a notable increase in extraction accuracy by approximately 0.3 pixel. This method proves effective in diverse and complex environments, satisfying industrial demands for precision and stability in inspections.

Abstract:To tackle the performance decline of models detecting rail surface defects due to the similarity between the characteristics of defect areas and background areas, this paper explores the high real-time, lightweight object detection network YOLOv8n and proposes a multi-modal rail surface defect detection algorithm, named RailBiModal-YOLO. Improvements to the YOLOv8n model involve the construction of a dual-stream backbone network structure that allows for the parallel extraction of multi-scale depth and RGB information; a plug-and-play dual-modal feature interaction and revision fusion module is designed to minimize the interference of low-quality image features and to fully leverage the complementary information from both modalities; the EVCBlock is introduced during the multi-scale feature construction phase to enhance the intra-layer information interaction within the RGB-D feature layers, thereby improving the detection of small defects. The Northeastern University NEU-RSDDS-AUG dataset is utilized for experiments, which has been custom-divided into four typical defect types, with mean average precision (mAP), frames per second (FPS), and the number of parameters serving as the primary evaluation metrics. It is demonstrated by the experimental results that the proposed model, in comparison to the original model, not only maintains high detection speed but also achieves enhancements in mAP@50 and mAP@50:95 by 1.8% and 3.2%, respectively, along with exhibiting increased robustness.

Abstract:Although the dung beetle optimization algorithm (DBO) has unique advantages, there are also some problems, such as low convergence accuracy and easy to fall into local optimum. In order to solve these problems, an improved dung beetle optimization algorithm named MSIDBO is proposed to enhance the optimization effect and maintain the balance between global and local search. An adaptive fitness distance balance strategy is proposed, which effectively avoids the dilemma of the algorithm falling into the local optimal solution by optimizing the foraging and stealing behavior of dung beetles. At the same time, the guided learning strategy and the local optimal perturbation scheme are introduced to accelerate the convergence speed of the algorithm and balance the relationship between the local development and global exploration ability of the algorithm. In order to evaluate the performance of MSIDBO algorithm, CEC2017 test function is used for simulation experiments. In three practical engineering design problems, MSIDBO algorithm is used at the same time, and compared with other five optimization algorithms. The results show that MSIDBO algorithm has significant advantages in convergence speed, solution accuracy and stability, which fully verifies its efficiency and reliability in practical application.

Abstract:To address the issue of the currently popular residual network having low accuracy in identifying gearbox faults in complex noise environments, and the slow convergence speed and poor global search capability of the traditional whale optimization algorithm(WOA), this paper proposes an intelligent fault diagnosis method based on the gramian angular difference field （GADF） and a hybrid whale-particle swarm optimization algorithm combined with a CBAM attention mechanism residual network. First, the collected one-dimensional vibration signals are overlap-sampled to obtain sufficient signal samples. Then, the gramian angular difference field encoding technique is used to convert the one-dimensional data into two-dimensional image data, constructing a two-dimensional image dataset under different faults. Artificial noise is added to expand the sample size and verify the impact of noise on the diagnostic method. Next, a CBAM attention mechanism module is added to the traditional ResNet network to enhance useful features and suppress irrelevant features, thus improving the model′s representation capability. The image dataset is then input into the HWP algorithm-optimized CBAM-ResNet model for training. Finally, the trained CBAM-ResNet model is used to classify the spiral bevel gearbox fault dataset, outputting diagnostic results. Experimental results show that this method can accurately identify spiral bevel gearbox faults without manual denoising, achieving an accuracy rate of 100%, and maintaining 95.38% accuracy in complex noise environments. Compared to other methods, it has higher accuracy, faster network convergence speed, and better robustness.

Abstract:With the significant increase in the complexity and diversity of modulation types in modern wireless communication environments, higher requirements are placed on the performance of automatic modulation recognition technology. This paper proposes a hybrid neural network model consisting of a convolutional neural network, a squeeze and excitation module, a long short-term memory network, a gated recurrent unit, and a fully connected layer network to improve the efficiency and accuracy of AMR technology. First, to address the problem of limited modulation signal recognition accuracy in low signal-to-noise ratio environments, a singular value decomposition algorithm is introduced to denoise the received I/Q signal, thereby improving the recognition accuracy of modulation signals under low signal-to-noise ratios while improving signal quality. Then, a convolutional neural network is used to extract multi-channel spatial features from the denoised signal. Then， a squeeze and excitation module is added to improve the pertinence of feature extraction. The gated recurrent unit and the long short-term memory network are combined to capture the time series characteristics of the signal. Finally, the extracted features are mapped to the classification space of the modulation mode through a fully connected layer network for classification and recognition. Experimental results show that the proposed network model significantly improves the modulation recognition accuracy in a low signal-to-noise ratio environment. The average recognition accuracy on the RadioML2016.10b dataset reaches 64.63%. At the same time, it enhances and improves the distinction and recognition accuracy of QAM16 and QAM64.

Abstract:Doppler speed measurement radar has the advantages of wide range of speed measurement，high speed measurement accuracy and strong reliability, and is widely used in the field of wheel-rail and maglev transportation in China. It is necessary to consider the interference of clutter to Doppler signal in the process of speed measurement radar, so it is important to study the signal processing method under the background of track ground clutter to improve the accuracy of speed measurement and ensure the safety of driving. In this paper, three typical probabilistic and statistical models, namely Kernel distribution, Weibull distribution and Gamma distribution, are theoretically analyzed, clutter measurement experiments are carried out using 77 GHz+24 GHz dual-band vehicle-mounted velocity radar. The results show that the statistical characteristics of the track ground clutter data of the vehicle-mounted speed measurement radar developed in this paper follow the Kernel distribution. In the background of clutter, firstly, the least mean square adaptive filtering method is used to de-noise the measured signal, and the improved Burg algorithm is used to estimate the spectrum to achieve high-precision velocity measurement. Experiments have verified that the proposed algorithm can effectively suppress clutter and improve the signal-to-noise ratio, and the final velocity measurement error is less than 0.5 km/h at low speed and less than 0.5% when the speed is greater than 50 km/h.

Abstract:Traditional power grid anomaly detection methods rely on converting expert knowledge into fixed rules and thresholds, which cannot meet the demands of rapidly evolving power grid systems. The current anomaly detection research mainly focuses on electricity theft and equipment failures as the main analysis objects, but the analysis of overcurrent anomalies is insufficient. This paper analyzes the characteristics of overcurrent anomalies, and discusses the problems and deficiencies of traditional experience-based rules. Through feature engineering, we determines the feature variables, and proposes an XGBoost-based power grid overcurrent anomaly detection model. Through experimental data testing and evaluation, the indicators of the model proposed in this paper outperform the detection methods based on traditional experience-based rules. In the 5-fold cross-validation, the minimum F1 score of the proposed model showed a 19.2% improvement compared to traditional rules, while the average value demonstrated a 15.1% improvement. The experimental results did not show significant performance differences, confirming the effectiveness of the model in anomaly detection. Compared to other commonly used machine methods for anomaly detection, the proposed model in this paper achieved an improvement of 6.4% to 8.7% in F1 score, demonstrating advantages in terms of stability and accuracy. The extreme case testing with training data significantly less than the testing data, along with the conducted interpretability analysis of the model, demonstrated that the proposed model exhibits high transparency and reliability. Moreover, it shows good generalization performance, making it suitable for effective deployment in real-world environments for overcurrent anomaly detection.

Abstract:In recent years, research on remote sensing object detection has mainly focused on improving the representation methods for bounding boxes, while overlooking the unique prior knowledge present in remote sensing scenes. To further enhance the detection accuracy of two-stage models while maintaining inference complexity, this paper presents improvements in feature representation and training strategies based on the feature extractor LSKNet constructed with large kernel convolutions. First, the RFA module is introduced to extract scale-invariant contextual information, alleviating the background noise introduced by LSK and enhancing the model′s robustness to noise. Then, the CS loss is proposed to implement a consistent supervision training strategy that reduces the semantic gap between features of different scales, enabling the model to possess multi-scale capabilities while focusing more on small objects. The proposed method achieves a single-scale result of 79.03% mAP50 on the large remote sensing image dataset DOTA, demonstrating the effectiveness of the proposed approach with almost no increase in inference complexity.

Abstract:Aiming at the existing problems of landslide semantic segmentation network of remote sensing image, such as large number of model parameters, slow training speed, fuzzy recognition of landslide boundary region, and differentiation of multi-scale semantic information classification of remote sensing image, this paper proposes an improved lightweight semantic segmentation model of RTformer. The cavity convolution attention ASPP module and channel attention SE module were embedded among the modules at different levels of the model to capture semantic information at different scales and to enhance the feature representation ability and improve the feature extraction ability of the model, making it more suitable for landslide remote sensing image recognition. Cityscapes data set was used to conduct comparative experiments on the expansion rate setting of the cavity convolution in the model and different batch sizes to obtain the optimal solution. A self-supervised training task was designed using the Bijie landslide disaster data set as the pre-training data set, and the model was fine-tuned and the segmentation performance of the model against the landslide disaster remote sensing images was tested. The resulting model achieved the best performance on both Cityscapes dataset and Bijie landslide disaster dataset. Compared with the original RTformer model, the mean crossover ratio (mIOU) of the two datasets increased by 2.26% and 4.34%, respectively. Compared with the classical semantic segmentation models such as FCN, U-Net, DeeplabV3 and SegFormer, the improved model realizes the recognition task with the fewest parameters and the fastest reasoning speed, and achieves the optimal segmentation effect.

Abstract:Aiming at the problem that most classic visual SLAMs are not robust enough in indoor dynamic environments, a visual SLAM that can distinguish between high and low dynamic environments is proposed based on the ORB-SLAM3 algorithm framework. First, an algorithm is proposed to distinguish whether the prior dynamic objects in indoor environments are in high or low dynamics based on the reprojection error of the pose transformation between multiple consecutive frames. Then, according to the high and low dynamics of the environment, it is decided whether to combine the YOLOv8-Seg instance segmentation network to remove the dynamic features in the dynamic environment to ensure the tracking accuracy of the SLAM system. Finally, in order to deal with the repeated map points in the map caused by dynamic features, a repeated map point elimination algorithm is added to the local map tracking to delete the repeated map points in the dynamic environment, further ensuring the stable tracking of the system. Experimental results on the public dataset TUM RGB-D show that the improved algorithm has improved the positioning accuracy compared with the ORB-SLAM3 algorithm, with a maximum improvement of 60.41% in low dynamic environments and a maximum improvement of 94.65% in high dynamic environments. Compared with other dynamic feature removal algorithms, higher positioning accuracy is achieved in most sequences, and it is also more advantageous in real-time performance. The proposed algorithm effectively solves the problem of SLAM coping with indoor dynamic environments and improves the positioning accuracy of SLAM.

Abstract:In the face of the challenges and security risks posed by the complex environment at the border, the deployment of unmanned monitoring systems is crucial for our country′s border defense. The existing unmanned border defense systems encounter challenges such as high rates of false positives and missed detections, as well as insufficient real-time capabilities, primarily due to significant variations in image scale caused by differences in distance between the cameras and the intrusion targets, along with the use of occlusion strategies by the monitored targets. A FDB-YOLOv5 occlusion small target detection algorithm with higher average accuracy, fewer parameters, and stronger universality is proposed to address this issue. Firstly, a dataset is constructed by collect a large number of personnel samples with different occlusion areas; secondly, a new structure called Faster_C3 has been introduced to reduce the delay and parameter count of the occlusion small target detection network, thereby improving the detection speed and universality of the model; in addition, a Dysample upsampler based on point sampling is introduced into the neck network to obtain more local details and semantic information, enhance the detection capability of the model for small targets, while reducing the computational overhead. Finally, a spatial pyramid pooling method based on multi-scale feature extraction BSPPF is used to effectively solve the problems of scale invariance and loss of feature information of the occluded targets, so as to better capture key information and im-prove the stability and robustness of the model for detecting occluded small targets. The experimental results indicate that compared to the baseline YOLOv5, FDB-YOLOv5 mAP@0.5% reaching 91.5%; experimental outcomes demonstrate that compared to the baseline YOLOv5, FDB-YOLOv5 exhibited superior performance with an mAP@0.5 score reaching 91.5%. There was also a reduction in the number of parameters and computations by 19.07% and 18.41%, respectively, and an increase in model detection speed by 8.83%. When compared to Faster R-CNN、SSD、YOLOv5s and YOLOv8, FDB-YOLOv5 showcases outstanding capabilities, offering valuable insights for unmanned border target detection technologies.

Abstract:Forgings are prone to various surface defects such as cracks during the manufacturing process, which affects product quality. Aiming at the problem that small target cracks are easily missed in complex visible light environments, and considering the requirement of efficient deployment in production line, the LSC-PoolFormer algorithm is proposed. First, the magnetic particle inspection images from the automobile steering knuckle production line of Hubei Sanhuan Forging Co., Ltd. were collected, annotated and made into a FDMPI data set; then, an encoder based on the PoolFormer backbone network was used to achieve lightweight and efficient feature extraction; secondly, the asymptotic feature pyramid is introduced as the neck network to reduce the semantic gap between features of different scales; finally, DS-Seghead is proposed as the decoding head based on dynamic snake convolution to enhance the model′s perception of stripe cracks, and a DDS training strategy is proposed, reducing the probability of missing small target cracks. The experimental results of LSC-PoolFormer on FDMPI show that compared with the baseline model, the parameter amount and calculation amount of this algorithm decreased by 9.2% and 48.78% respectively, and the F1 score and IoU increased by 1.1% and 1.69% respectively. At the same time, the performance on the public data set NEU-Seg also proves the generalization ability of the algorithm. Compared with the baseline model, the mF1 score and mIoU increased by 0.66% and 1.04% respectively while greatly reducing the number of parameters and calculations. Experiment shows that the algorithm in this paper significantly reduces the complexity of the algorithm while maintaining detection accuracy, which is beneficial to the actual deployment.

Abstract:To address the issues of existing conveyor belt foreign object detection models in coal mines, which perform poorly in low-light environments, miss elongated and small foreign objects, and have a large model size that hinders deployment on edge devices, this paper proposes a low-light coal mine conveyor belt foreign object detection algorithm based on an improved YOLOv8. First, image enhancement techniques are applied to preprocess low-light images to enhance the effective feature information of foreign objects on the coal mine conveyor belt. Next, dynamic snake convolution is introduced into the model′s backbone network to dynamically adjust the convolution kernel shape, improving the model′s focus on elongated foreign objects. Additionally, a slim-neck design paradigm is used to modify the neck network, significantly reducing the model′s parameters while maintaining learning capability. Finally, the Inner-CIoU loss function is employed to replace the CIoU loss function, accelerating the model′s convergence and improving its detection performance for elongated and small foreign objects. Experimental results show that, compared to the baseline model, the improved algorithm increases the average detection accuracy by 1.6%, reduces the model size by 29.7%, and improves the detection speed (FPS) by 59%, validating its effectiveness. In comparison with other advanced models, it is proved that the proposed algorithm still has strong recognition ability in complex environment.