Abstract:Aiming at the problems of insufficient topic feature representation ability and high feature dimension caused by high dimensionality of data in topic-based text classification task, this paper improves the input feature representation and Convolutional Neural Networks (CNN) structure. In the feature representation, LDA model is proposed to calculate the inverse topic space frequency to get the topic vector matrix of text, which reduces the feature representation of noisy topic and enhances the weight of key topic. The topic vector matrix and word vector matrix of the text are respectively used as the input of CNN model. A two-layer CNN network structure is proposed, and a multi-channel pooling layer is added after the pooling layer of each layer of CNN to integrate the pooling results of each layer of CNN, and obtain more local salient features while reducing the feature dimension. Finally, the Attention mechanism is used to weight the fused features and input them to the fully connected layer for classification. According to the experimental results, the accuracy and recall of the improved model in text classification tasks are above 98%, and the F1 value is increased by 6% compared with the benchmark experiment.

Abstract:In order to solve the problems of limited duration and location of communication experiments of the Electronic Information Specialty in Universities and help students better understand the principle of spectrum analysis, a small program-based signal generator and spectrum analyzer were developed, and the advantages of the Internet were used to realize the effect of conducting experiments at any time. The signal generator can generate 6 kinds of commonly used signals including sine wave, square wave, triangle wave, random noise, Lorentz pulse and Simpson pulse in a certain frequency range, and output them through the mobile phone speaker or headphone jack. The speech received by the microphone is subjected to windowed spectral analysis, and the spectral results are displayed on the screen. Using a fixed type of mobile phone to test, the results show that the relative error of the frequency of the signal generator is 0% when generating a sine wave of 100 Hz to 10 kHz, and the relative error of the frequency of other waveforms is less than 1%. During spectrum analysis, the peak frequency measurement error is within 5 Hz, which has good experimental teaching guiding significance and field application applicability.

Abstract:In order to improve the estimation accuracy of lithium battery state of health, a lithium battery SOH estimation method based on IGWO-SVR is proposed. Firstly, aiming at the problem of kernel parameter selection of support vector regression (SVR), the improved gray wolf (IGWO) algorithm is used to optimize the kernel parameters of support vector regression (SVR). The SVR estimation model realizes the estimation of the SOH of lithium batteries. Based on the NASA battery dataset, the model is trained and validated and compared with the SVR and GWO-SVR methods. The results show that the IGWO-SVR method can effectively improve the accuracy and stability of SOH estimation, and the maximum estimation error does not exceed 2%.

Abstract:To solve the problems of nonlinearity, strong coupling and large lag in the microbial degradation process of fermentation tank, a decoupling control method of fermentation tank based on fuzzy gray wolf PID algorithm was proposed. First, the control system model of the fermentation bin is identified, and the coupling effect of temperature and humidity generated in the control process is eliminated through dynamic decoupling compensation. Then, combining the advantages of PID and fuzzy control, the fuzzy PID control system model is established. Finally, the fuzzy PID initialization parameters are optimized by using the iterative optimization ability of the gray wolf optimization algorithm. The simulation results show that, compared with the traditional control method, the proposed control strategy reduces the temperature loop regulation time from 1 623 s to 596 s, the relative humidity regulation time is reduced by 910 s, and has a better decoupling effect, which can effectively realize the accurate control of degradation environment.

Abstract:To accurately assess pollution status of the surface of luminescent port for navigation-aid lamps, this paper considers the impact of different pollutants on the difficulty of cleaning and proposes a pollution grade assessment method based on fusion weight method by fuzzy evaluation. The homogeneity was used as the discrimination standard to distinguish different pollutants in the blocked image. Fusion of subjective and objective weights through fuzzy evaluation, according to which the calculation formula of pollution degree was determined. Cleaning experiment was conducted on polluted samples according to the pollution grade method of this paper, and the cleaning rate is 93%. And the area method and entropy method will cause at least 32% and 29% waste of resources respectively. The area method does not consider the different cleaning difficulties of different pollutants. The pollution degree evaluation method proposed in this paper can distinguish pollutants and consider the subjective and objective influence of pollution assessment, which means the evaluation results are more scientific and accurate.

Abstract:In order to solve the problem of single feature tracking failure under the condition of background illumination change and partial occlusion in the road, a vehicle tracking algorithm based on multi-feature fusion and Kalman prediction is proposed. Multi-feature fusion includes tracking the color, edge and texture of the vehicle, using color histogram to describe the color distribution, using Local Binary Patterns (LBP) with rotation invariance to describe the texture distribution, using the improved Canny operator to calculate the edge distribution information, establishing the feature fusion function, and using the average peak correlation energy to construct the best feature description of this tracking. When the feature matching between two adjacent frames is greater than the set threshold in the vehicle tracking process, occlusion and tracking interruption are determined. Kalman filter is used to predict the current position. Finally, different experiments show the effectiveness of the algorithm.

Abstract:In this paper, a new capacitive MEMS ring-shaped wave gyroscope with high sensitivity is studied, Based on the outstanding advantages of the solid wave gyroscope such as high sensitivity, good reliability and wide dynamic range. First, the sensitive part and the whole structure of this MEMS ring-shaped wave gyroscope are designed, and the working characteristics and sensitive principle are analyzed. Then, the numerical model of the ring-shaped sensitive part is established, and the relationship between the sensitivity and the main design parameters is studied. Finally, it is concluded that the resonant frequency is 6.036 0 kHz, the mechanical sensitivity is 0.003 6 μm/° with a driving displacement of 2.906 8 μm, and the maximum stress is 19.79 MPa in the ultimate displacement of the gyroscope, by studying the motion mode, driving method, mechanical sensitivity and ultimate displacement of the ring sensitive structure. The results show that the sensitive structure of the gyroscope is reasonable and has a high sensitivity, it can be used as the basis for the further development of high sensitivity MEMS gyroscope.

Abstract:Aiming at the problem of low coverage of wireless sensor networks deployed randomly, a Multi Strategy gray wolf (MSGWO) algorithm for wireless sensor network coverage optimization is proposed. Firstly, in order to balance the global and local search, a nonlinear convergence factor of hyperbolic tangent is proposed; Secondly, the bounding step size is reconstructed by differential mutation to reduce the probability of the algorithm falling into local optimization; Then, in order to speed up the convergence speed and accuracy of the algorithm, the gray wolf position is updated by using the transient search optimization equation; Then, Levy flight strategy is integrated to increase the diversity of space search; Finally, the boundary offside strategy is introduced to avoid the relocation of gray wolf individuals. The simulation results show that compared with SSA, LGWO, PSO and PSOGWO, the average coverage increment of MSGWO algorithm is 12.52%, 6.054%, 7.53% and 3.45% respectively. This algorithm has higher average coverage and better node distribution.

Abstract:Convolution neural network (CNN) is sensitive to spatial features, while Inception has the advantage of multi-scale feature extraction compared with CNN, long-short-term memory network (LSTM) is sensitive to temporal features, and deep short-term memory network (DLSTM) has deeper feature extraction advantages than LSTM. In order to fully extract the spatial and temporal characteristics of rolling bearing vibration signals in multi-scale, a dual-channel rolling bearing fault diagnosis model Inception-DLSTM based on the combination of Inception channel and DLSTM channel is proposed. For the Inception channel, the time-frequency diagram generated by the wavelet transform of the bearing vibration signal is used as the input, and the multi-scale Inception network is used to extract the spatial feature information of the time-frequency diagram; for the DLSTM channel, the bearing vibration signal is directly taken as the input, and the DLSTM network is used to fully extract the time feature information of the signal.Then the feature information output from the two channels is connected into a spatio-temporal feature vector, and finally the classifier is used to diagnose and identify the bearing fault. Comparing the bearing fault data can be obtained, and the fault identification accuracy of the Inception-DLSTM dual channel can reach 100%, and has good fault diagnosis and feature extraction capabilities.

Abstract:Aiming at the problem of fault diagnosis caused by noise pollution and the vague of fault characteristic frequency. A new method for fault diagnosis of rolling bearings is proposed. Firstly, the Gini Index (GI) is used to evaluate the health status of rolling bearings, and the vibration signal with abnormal state is used for noise reduction preprocessing using the optimal parameter Maximum Correlated Kurtosis Deconvolution (MCKD) to highlight impact component.Then,calculate the hierarchical entropy (HE)of the preprocessed signal to form a feature matrix.Finally, the cuckoo search algorithm is used to optimize the relevant parameters of the support vector machine, and the intelligent diagnosis of the fault state of the rolling bearing is completed.The feasibility of the proposed method is verified by experimental analysis, and it has a high accuracy.

Abstract:Proposes a SA-PSO-XGBoost prediction model for forecasting the temperature in Nanjing, based on ECMWF meteorological data from January 1, 2016, to December 31, 2017. The meteorological data is divided into training and testing sets. The PCA dimension reduction method is applied to compress and reduce the features of the meteorological data. The SA-PSO-XGBoost model optimizes the hyperparameters using a hybrid algorithm combining simulated annealing and particle swarm optimization. The testing set is then used to compare the performance of the SA-PSO-XGBoost model with XGBoost, GRU, and LSTM neural networks. Experimental results demonstrate that the SA-PSO-XGBoost model outperforms others in terms of accuracy and robustness in predicting the temperature 6 hours ahead.

Abstract:The current multi-agent reinforcement learning algorithm cannot fully consider the cooperative relationship between multi-agents in the value decomposition algorithm, and the stochastic strategy used in the exploration process is prone to cross the optimal point and fall into the local optimal solution. Aiming at the above problems, this paper proposes a deep communication multi-agent reinforcement learning algorithm. This paper designs a communication mechanism in value decomposition network by using convolution and fully connected structure to enhance the cooperation between multi-agents. Then, a new adaptive exploration strategy is proposed in this paper. In order to balance the contradiction between data exploration and utilization, a periodic decay strategy is added. Finally, simulation results verify that the proposed method achieves 25.8% performance improvement in some scenarios, and improves the cooperation capability of multi-agent.

Abstract:In order to improve the accuracy of local stereo matching, a stereo matching algorithm based on improved Census cost and optimized guided filtering is proposed. Aiming at the problem that the traditional Census cost calculation is not accurate in the cost calculation of disparity discontinuous region, the neighborhood pixel validity label is carried out in the Census transformation process, and the influence of invalid pixels on the overall cost is reduced by calculating different generation values for different validity points. In the stage of guided filtering cost aggregation, two sizes of windows are used to calculate the linear coefficients, and then different linear coefficients are selected according to the results of image region division, which solves the problem of inadaptability of local region cost aggregation caused by fixed window size. Finally, the final disparity map is obtained by disparity calculation and disparity optimization. Experiments were carried out on the Middlebury v3 stereo matching evaluation platform. The results show that the average mismatch rates of the proposed algorithm in the non-occluded area and in all areas are 18.17% and 23.81%, respectively, which are better than many existing algorithms.

Abstract:In the application of path planning for mobile robots, jump point search (JPS) is widely used because of its simple, rapid-and easy-to-implement characteristics. However, the standard JPS algorithm with a low efficiency of heuristic function is prone to hold redundant nodes in the planned path. Besides, the JPS often neglects the path security. To address these issues, this paper proposes an improved JPS algorithm. The proposed algorithm designs a novel heuristic function combining diagonal distance and direction information to improve the path-finding efficiency, and further smooths the planned path for guaranteeing safety and reliability. The path-planning experiment for mobile robot in the obstacle environments demonstrates that our algorithm obtains higher path-search efficiency with reducing the planning time and improving the safety effectively, compared with other advanced path planning algorithms.

Abstract:High-density electrode channels are commonly used in brain-computer interface (BCI) systems to obtain high spatial resolution EEG signals, but at the same time, too many noise channels are introduced, which affect the decoding performance of electroencephalogram (EEG). In order to eliminate irrelevant noise channels, a channel selection method based on Tikhonov regularized co-spatial pattern (TRCSP) and L2 norm for motor imagery EEG is proposed this paper. Firstly, the optimal spatial filter is obtained based on TRCSP and classifier, and then the weight values of each channel obtained by the spatial filter are sorted based on L2 norm. The data of the first K channels are used for CSP feature extraction, and the optimal K value is determined according to the classification accuracy of the classifier, so as to obtain the optimal number of channels and channel combination. In the experiment, six classifiers were used on the BCI competition III (2005) dataset IVa and self-collected dataset from our laboratory to verify the effectiveness of the proposed channel selection method. The average classification accuracy of the proposed method on the two datasets reached 87.57% and 74.32%, respectively, better than other existing methods.

Abstract:In order to extract the purer rain sound signal from the rain sound signal mixed with various noises, this paper proposes a denoising method of rain sound signal based on the combination of improved fully adaptive noise set empirical mode decomposition (CEEMDAN) and wavelet threshold. Methods Crosscorrelation function was introduced to find the optimal decomposition level F value of CEEMDAN, and the signal was decomposed into multiple intrinsic mode components (IMF) with high frequency to low frequency by CEEMDAN algorithm. Using wavelet threshold, the noise component in the high frequency IMF component is filtered out, and finally, the denoised high frequency IMF component and the denoised low frequency IMF component are reconstructed to extract a relatively pure rain sound signal. The experiment shows that the denoising effect of this method is superior to the traditional methods such as empirical mode decomposition (EMD) denoising algorithm and wavelet threshold denoising algorithm, and the denoised rain signal can accurately reflect the characteristics of environmental rain, thus improving the accuracy of rain analysis.

Abstract:This paper focuses on the problem of large positioning error of the DV-Hop algorithm. To overcome the defects of DV-Hop algorithm when solving the average jump distance and the unknown node position, this paper proposes an improved DV-Hop positioning algorithm based on ranging correction and bat optimization. Firstly, the minimum mean squared error criterion is used to solve the average jump distance between anchor nodes while a correction factor is added to reduce the ranging error. Secondly, the chaotic mapping strategy is applied to initialize the population and set the threshold M to control the number of mappings, and besides, the speed-weighted strategy is used to control the step length of the search to enhance the ability of the bat algorithm to jump out of the local optimum. Finally, the Improved Bat Algorithm is used to determine the location of the unknown node. The simulation results show that the proposed positioning algorithm has higher positioning accuracy, which is improved by 32.35%, 18.80% and 8.16% compared with DV-Hop algorithm, BADV-Hop algorithm and PSODV-Hop algorithm, respectively.

Abstract:In order to solve the problem of insufficient transmission performance in satellite communication and deep space communication in the system of cascaded binary low-density parity-check (LDPC) codes and high-order continuous phase modulation (CPM), a multivariate LDPC-CPM optimization method is proposed. Firstly, cascading to form a multivariate LDPC-CPM system achieves better transmission error performance. Secondly, considering that the problem of the traditional decoding algorithm of nonbinary LDPC codes is too complex and hard to implement in hardware, an improved Mixed-Log-FFT-BP algorithm is designed. The solution of the ratio reduces the decoding complexity. Finally, in view of the problem of too many parameters in multivariate LDPC-CPM system, a parameter optimization method with gradually convergent performance is proposed. The simulation results show that, compared with the system before optimization, the optimized system has a 1~1.2 dB improvement in the bit error performance under the medium and high signal-to-noise ratio and the system complexity is lower.

Abstract:The object of remote sensing image has the characteristics of complex background and changeable direction. The process of object detection in remote sensing image using traditional methods is complex and time-consuming, with low accuracy and high rate of missed detection. To solve the above problems, we propose an improved YOLOv5AC algorithm. This algorithm bases on the YOLOv5s model. First, an asymmetric convolution structure is built in the original Backbone to enhance the robustness of the model to flipped and rotated targets; Secondly, coordinate attention mechanism is introduced into C3 module of backbone network to improve feature extraction capability, and Acon (Activate Or Not) adaptive activation function is used for activation; Finally, we use CIOU as the location loss function to improve the positioning accuracy of the model. The improved YOLOv5-AC model was tested on NWPU VHR-10 and RSOD datasets, and the average accuracy reached 94.0% and 94.5%, respectively, 1.8% and 2.3% higher than the original YOLOv5s, which effectively improved the object detection accuracy of remote sensing images.

Abstract:To address the problem of low detection accuracy in realtime detection of defects on the surface of flat ceramic films, this paper proposes a YOLOv5 ceramic film defect detection method that incorporates coordinate attention and adaptive features. By adding a coordinate attention mechanism to the backbone network of the original YOLOv5 model, the relationship between location information and channels is established to obtain the region of interest more accurately. The adaptive feature fusion mechanism is incorporated into the prediction network of the original network to improve the detection capability of the model for multi-scale defects. Replace the spatial pyramid pooling module in the original network with the spatial pyramid pooling module of the null space convolution pooling module to improve the convolutional kernel field of view to obtain more useful information. The experimental results show that the average accuracy of this model is 97.8%, the number of detection frames is 32 FPS, and the average accuracy is improved by 5.5% compared with the original YOLOv5 model. The model proposed in this paper improves the detection accuracy of the model under the condition of satisfying the real-time detection of flat ceramic film defects, which has certain reference value for promoting the development of flat ceramic film defect detection.

Abstract:This paper aims at the problems of small target proportion, complex background and low detection accuracy in aerial images. Proposes a small aerial target detection algorithm based on receptive field enhancement and parallel coordinate attention. A receptive field enhancement module is designed to expand the receptive field range by using cavity convolution of different sizes and integrate effective channel attention mechanism to improve the feature extraction ability of the network. The feature fusion structure is improved to improve the detection ability of small targets. A parallel coordinate attention module is designed to improve the ability of aerial photography dense small target detection and anti-background interference. Experiments are conducted on VisDrone dataset with different input resolutions. The experimental results show that mAP0.5 and mAP0.5:0.95 of the proposed algorithm are improved by 5.4% and 4.2% compared with YOLOv5 algorithm, and mAP0.5 can reach 54.9% at input resolution 1 536×1 536. It can achieve better effect of small target detection.

Abstract:To address the problem of fast stitching of tile images with different textures, an image stitching algorithm based on frequency domain is studied. Firstly, the relationship between speed and accuracy in image stitching is analyzed, and the Gaussian pyramid scale space is constructed to reduce the running time of the program. Subsequently, ceramic tile images are registered by phase correlation algorithm. In this process, the subpixel refinement algorithm is combined to compensate for the registration error caused by the decrease in resolution, and a high-frequency feature highlighting strategy based on the Laplacian operator is proposed to enhance the robustness of the phase correlation algorithm. Experiments on tile image stitching with multiple colors and textures show that the proposed algorithm improves the stitching success rate by 60% and 65% in comparison with the algorithms based on SIFT and ORB. Compared with the SURF, H-SURF and FFT algorithms, it improves 9.04%, 7.58% and 4.02% in the SSIM. In terms of running speed, the proposed algorithm is 4 times, 3 times and 2 times higher than SIFT, SURF and H-SURF algorithm, and is only 52% of the traditional Fourier transform-based phase correlation method. The algorithm achieves fast and high-quality image stitching, which meets the imaging requirements of high-resolution tiles in industrial applications.

Abstract:In order to solve the problem of low detection accuracy of traffic signs due to small targets in real traffic scenes, a traffic sign detection algorithm based on improved SSD was proposed. First, a deeper ResNest network was used to replace VGG16, the backbone network of the original SSD algorithm, to enhance the strong characterization of weak target features. Then, the RFB module was used in the additional layer of SSD to increase the receptive field of small targets. Secondly, bi-FPN weighted bidirectional feature pyramid network is used to effectively combine deep and shallow feature information to improve the detection performance of small targets. Finally, K-means++ clustering algorithm was used to adjust the size of the default window, which effectively avoided the problem that the original default window was too large but the traffic signs were small and could not be matched, so as to improve the detection efficiency. Experimental results show that the proposed model achieves 95.33% mAP on the China Traffic Signs Data Set (CCTSDB). Compared with the original SSD model, the proposed model can better adapt to traffic signs detection under natural background.

Abstract:The corrugated compensator is a key component for compensation in modern pipelines. The corrugated compensator is a compensation device that absorbs the dimensional changes in pipelines due to thermal expansion and contraction, excessive pressure, etc. through the effective expansion and contraction of its own elastic element, and detects the fault conditions caused by excessive axial dimensional changes of corrugated compensators in industrial pipelines in real time. In this paper, an improved edge detection operator based on machine vision technology is designed to detect the axial dimension of the corrugated compensator, using hybrid filtering instead of Gaussian filtering to filter the noise, refining the image by morphological processing, and finally using the OTSU algorithm to achieve the separation of the object to be measured and the background. The experiment proves that the machine vision technology can detect the axial dimension of the corrugated compensator, and the method can solve the problem that the axial dimension of the corrugated compensator in industrial pipelines is difficult to detect in real time, while the detection equipment is easy to install, safe to use, and the accuracy can reach the expected index.

Abstract:The traditional calibration method of transformer on-load tap-changer tester has difficulties in synchronous calibration of transition resistance, transition time and synchronization. In this paper, three variable resistor modules based on FPGA control DAC are designed to simulate different resistors. After frequency compensation, the variable resistor module can realize the rapid switching of resistors, a6nd realize the synchronous verification of analog transition resistance, transition time and synchronization. At the same time, the protection circuits of overvoltage, overcurrent and antireverse connection are designed to improve the reliability of the whole device. After testing, the maximum allowable error of transition resistance simulated by the calibration device is better than ±(0.2% RD+5 mΩ), and the absolute error of transition time and synchronization is less than 10 μs, which fully meets the needs of laboratory calibration and solves the calibration problem of transformer on-load tap-changer tester.

Abstract:As the key component of the support structure of the onshore wind turbine, the service quality of the tower directly affects the operation safety of the wind turbine. Monitoring tower vibration, evaluating service quality and early warning of damage are important means to improve the operation safety of wind turbine. Firstly, the type of tower, the load and dynamics of tower are briefly introduced. Secondly, the methods and characteristics of tower contact and noncontact vibration monitoring are summarized, different monitoring methods were compared, and the limitations of existing tower vibration monitoring schemes are put forward. Finally, the research status and existing difficulties of tower vibration monitoring data in state identification, fault diagnosis and fault warning are analyzed, and the future development trend of tower vibration monitoring is forecasted.

Abstract:Aiming at the problem that the information missing gestures recognition based on deep learning needs a large amount of labeled. The deeper the network needs more parameters, we first collect a data set called IMG_NUIST which consists of information missing gestures and full gestures. Then we propose a new gesture recognition model CLGR, the inter-class and intra-class similarities constraints enhance the feature learning performance of the model. Extensive experiments are conducted on two classic datasets (ASL Alphabet and NUS I) and the proposed IMG_NUIST dataset. The experimrnt results are shown as follows: 1) in the ablation study, contrastive learning can effectively improve recognition accuracy up to over 98.60% and the model convergence speed are significantly accelerated. 2) In the comparative experiments with two recent works and two contrastive learning models, the computational complexity of CLGR is 41.4% simpler than that of the two comparison models on average. CLCR can recognize the gestures with missing information and works well for those gestures with cluttered backgrounds. The gesture recognition accuracy of CLCR on the NUS I and IMG_NUSIT data sets outperforms the four comparison methods and is only 0.43% lower than the best result on ASL. Especially on the NUS I dataset, CLCR increases the recognition accuracy of gestures by 17.35% on average. The experimental results show that the proposed model is significantly effective for gesture recognition tasks with missing information and cluttered background with fast convergence speed and low computational complexity, and it is practical.

Abstract:According to the real-time and accurate measurement of oxygen concentration in narrow space during flight test，the existing problems of oxygen concentration measurement system used in flight test are analyzed firstly, TDLAS is proposed to measure the oxygen concentration of fuel tank. Secondly, the internal structure of the system is optimized to achieve gas-liquid separation and miniaturization. The final volume is more than ten times smaller than the existing equipment, and the wavelength modulation technology and temperature/pressure compensation algorithm are used to improve the oxygen concentration detection capability of the system in the airborne environment. Finally, the reliability of the system was verified through ground and flight tests. The maximum measurement error is 0.28%, reaching the international advanced level of airborne oxygen concentration measurement, laying the foundation for the subsequent identification of various types of inerting systems.

Abstract:Indoor positioning systems have a wide range of applications in drones, robotics, medicine, and VR. Lighthouse spatial positioning technology was originally developed as a tracking and positioning system for VR devices. It has great advantages in accuracy and delay. However, the implementation of this algorithm often relies on centralized official tracking software and deployment of multiple base stations. The process is complicated, and the tracker of the VR device is large and expensive. For this, this paper proposes an attitude estimation algorithm for a single base station, and designs a low-cost, lightweight, and scalable positioning tracker based on FPGA. The positioning tracker captures the optical signal from the base station, performs filtering, decoding, data synchronization and attitude estimation on it, and outputs highprecision positioning data in real time through the serial port, which improves the integration and computing efficiency of the positioning system. By building an experimental scene, the positioning tracker and attitude estimation algorithm designed in this paper are verified. The experimental results show that the accuracy of the positioning system reaches millimeter level, and the position jitter range is less than 4%.