Abstract:Surface electromyography (sEMG) signals of spasticity patients can be problematic due to involuntary spikes and the poor signal quality. Also, the length of sEMG signals can be very short. In order to solve these problems, a stretch reflex onset (SRO) detection method based on modified sample entropy is proposed: Firstly, sEMG signals are framed by a fixedlength sliding window and the sample entropy of each frame is calculated. Afterwards, adaptive threshold is set to determine the SRO. The results show that the modified sample entropy achieves improved performance in SRO detection compared with the standard sample entropy, and shows better robustness in processing shorter time data series and against spurious background spikes. The recognition accuracy rate reaches 8906% using modified sample entropy but only reaches 4818% when using standard sample entropy. The findings from this study show that the proposed method can provide insight as to the mechanisms underlying the passive resistance.

Abstract:In order to overcome the limitations of small parameter detection conditions in classical stochastic resonance, this paper mainly introduces the single sideband modulation technology based on Hilbert transform combined with a secondorder stochastic resonance system, and the application of modulation stochastic resonance to realize large signal detection in engineering is proposed. At the same time, the paper discusses the cases of small sampling frequency and large sampling frequency, and makes relevant research on the selection of fundamental frequency signals. It demonstrates that the Hilbert single sideband frequency modulation has a good detection effect under the condition of small sampling frequency combined with the secondorder system, but under the condition of large sampling frequencies, it can be promoted combined with scale transformation. Numerical simulation analysis shows that the baseband signal has a higher and more stable output signaltonoise ratio when it takes about 10~60 times the frequency axis resolution. The effect of this technique combined with secondorder systems is significantly better than firstorder systems. Finally, the scale transformation with Hilbert single sideband frequency modulation technology is applied to the bearing inner and outer ring fault signal detection, and it is found that the large frequency fault signal can be detected obviously and accurately.

Abstract:The vibration signal of planetary gearboxes is complex with timevarying transmission paths, which makes fault feature extraction of planetary gearboxes difficult. For this issue, a fault feature extraction method based on windowed vibration separation and variational mode decomposition (VMD) is proposed. Firstly, the angle synchronous average is used to eliminate asynchronous components. Then, the windowed vibration separation is implemented. The parameter K is determined by observing the center frequency, and the VMD is applied to separated vibration signals to select the component containing fault feature. Finally, the selected feature component is demodulated by Hilbert demodulation analysis to achieve the fault feature extraction. By analyzing the measured signal of planet gear root crack fault, it is proved that this method can effectively extract the fault feature of the planet gear.

Abstract:In order to make the locomotive work in the optimal adhesion state, a sliding mode control algorithm based on extended Kalman filter (EKF) is proposed. Firstly, based on the wheelrail dynamics and the locomotive adhesion model, the EKF is used to approximate the wheelset speed and the vehicle body speed of the locomotive. Then, in order to overcome the external disturbance on the system, a sliding mode control algorithm based on exponential convergence disturbance observer is proposed. Considering the problem that the optimal wheelrail creep speed is unknown, a variable steplength search algorithm is designed according to the locomotive sticking model to track the optimal creep speed of the current track surface. Simulation results show that the EKFbased sliding mode control algorithm can effectively improve the stability of locomotive operation and the utilization of adhesion.

Abstract:Bearings are critical components of rotary machinery equipment. Numerous studies have been conducted on bearing fault diagnosis.Some of these methods can only be used for diagnosis of a certain type of bearing failure and cannot detect other failures. The diagnostic accuracy rate for most methods can be further improved. A new method is proposed for bearing fault diagnos is based on wavelet packet energy entropy and deep belief network (DBN).The bearing vibration signal is processed using wavelet packet transform to get the energy entropy feature vector. The feature vector represents the vibration energy in different frequency bands, which can be used to distinguish the fault type. The deep model based on DBN is adopted to recognize fault types.The proposed method achieves 100% and 995% fault recognition accuracy on two bearing datasets, respectively.The experimental results show that the proposed method has good versatility and can achieve high diagnostic accuracy.

Abstract:Due to the popularity and difficulty of research in the field of computer vision, pedestrian detection has been widely used in auxiliary driving and traffic monitoring. Traditional feature extraction method for pedestrians in complex environment is difficult to effectively capture the distinct characteristics information. And convolutional neural network which is popular at present has some influence on generalization performance because BP algorithm is easy to fall into local minimum value,and with the increase of the network layer, some significant feature information is decreasing layer by layer. In view of the above problems, this paper proposes a pedestrian detection algorithm combining deep sensing features with kernel extreme learning machines. Firstly, on the basis of CNN structure, the front layer features and the deep layer features are fused in two stages, and then sent to the followup layer for learning, the directed acyclic graph network(DAGnet) network is constructed. Then, the depth feature information is classified by the kernel extreme learning machine with high realtime performance and strong generalization ability, and the parameters are optimized by Kfold cross validation;In the detection phase, graphbased visual saliency(GBVS) saliency detection algorithm is used on the feature map learned by the DAGnet network to quickly mark the pedestrian area in the test image, and then sliding window is used to identify the precise position of the pedestrian in the salient area.The experimental results show that the positive detection rate on the INRIA data set and the Caltech data set is higher than 90%,and the detection speed is improved significantly if the accuracy is guaranteed.

Abstract:Aiming at the high reliability and high maintenance cost of the main bearing in the wind turbine, a maintenance optimization model based on the lowest maintenance cost per unit time is proposed. Firstly, by analyzing the vibration signal of the main bearing of the wind turbine, the characteristic value which can reflect the degradation process of the bearing is extracted, and then establish the Weibull proportional hazard model. Secondly, the failure rate curve of the main bearing during the degradation period is analyzed to determine the failure replacement threshold of the main bearing; then improve the traditional failure rate by modifying the traditional agereturn factor, and analyze the model operation results to avoid the phenomenon of undermaintenance during regular maintenance; Finally, determine the maintenance cycle with the best maintenance cost per unit time. Analysis of the maintenance optimization model and the calculation results show that the maintenance cost per unit time can be reduced by 14.4% through model optimization.

Abstract:Aiming at the two key points (feature extraction and fault recognition) of bearing fault diagnosis, a new car motor bearing fault diagnosis method was proposed. At the feature extraction link: a feature extraction method based on LCD decomposition and symbol entropy was proposed. At the fault identification link: in order to improve search ability of fruit fly optimization algorithm (FOA) to relevance vector machine (RVM), study of “history” strategy was introduced to FOA, then, FOA with history study ability (HSAFOA) was proposed and effectively improved the classification performance of RVM. Different fault types and different fault degrees of rolling bearing fault diagnosis experiment results show that the LCD symbol entropy can represent fault effectively and HSAFOA improved the identification accuracy of RVM, it has a certain superiority when compared with some other methods.

Abstract:Gale weather can easily cause highspeed train accidents such as derailment and rollover. Therefore, the ultra shortterm prediction of wind speed is of great significance for the safe operation of highspeed rail. A prediction model based on long shortterm memory (LSTM) networks is proposed in this paper. The maximum wind speed data per minute collected by WindLog wind speed sensor is preprocessed. The proposed model was trained using wind speed data of Haidian District, and the wind speeds 1, 5 and 10 min ahead were predicted. The mean absolute error (MAE) of 1min ahead prediction was 0467 m/s with the accuracy rate of 100%. The MAE of 5 min ahead prediction is 0543 m/s with the accuracy rate of 996%, the MAE of 10 min ahead prediction is 0627 m/s, and the accuracy rate was 988%. The experimental results show that the prediction model has better adaptability and higher prediction accuracy.

Abstract:For the instability of the extreme learning machine in predicting the remaining useful life of lithiumion batteries, this paper proposes a hybrid particle swarm optimization algorithm to optimize the prediction model of extreme learning machines.The optimized particle swarm optimization algorithm is used to optimize the input of the extreme learning machine, which not only can significantly improve the prediction accuracy of the model, but also greatly increase the credibility of the single prediction result of the remaining useful life lithiumion battery.In this paper, the lithiumion battery data published by NASA PcoE is used to carry out simulation experiments and evaluate the prediction performance of the model, and compare it with the prediction results of standard extreme learning machine prediction model.The statistical results show that the prediction error is controlled by about 2%.

Abstract:The remote field eddy current (RFEC) technique is widely applied for defect inspection in the wall of metallic pipes. However, it is difficult to identify the defect’s specific location because this technique has the same sensitivity to inner diameter (ID) and outer diameter (OD) defects. To meet the demand of inservice inspection for pressure piping, this paper presents an outside inspection method based on pulsed remote field eddy current (PRFEC) for pipe defects and makes a contribution on depth quantification and location identification of ID and OD defects. First, based on the characteristic of magnetic field propagation, the principles of inside and outside RFEC inspection were compared and analyzed. Afterwards, the repetitive frequency and duty cycle of the excitation pulse were optimized by using finite element simulation. Then, the magnetic perturbations caused by ID and OD defects and their response signals were studied, and meanwhile the correlations of signal peak and time of zerocrossing (TZC) with the defect depth and location were revealed. Finally, a PRFEC system was set up and experiments were carried out on a carbon steel pipe with prefabricated ID and OD defects. The results show that: i) the peak value of defect signal increases monotonically as the defect depth increases, which can be used to quantify defect depth; ii) the signals’ TZCs of ID defects are always greater than those of OD defects with the same depths, and therefore, by using this feature, the defect location can be identified.

Abstract:Aiming at the problem of traditional gait recognition algorithm due to the change of clothing and the change of covariate factors such as perspective, this paper proposes a gait recognition algorithm based on improved deep convolutional neural network. The algorithm uses the layered processing mechanism to extract the gait features from the gait data, which can reduce the impact of common changes and occlusion on the recognition accuracy. At the same time, the algorithm determines the optimal number of features of each layer in the network according to experiments. The optimal size of the graph and the type of input features to be used for gait recognition can handle relatively small data sets without any enhancement or finetuning. CASIAB gait database simulation experiments show that CNN proposed in this paper covering the gait recognition problem of cross view gait recognition and no subject, it can overcome the covariate factor problem related to gait recognition, and has better gait recognition accuracy.

Abstract:In order to solve the problem of automatic detection for pipeline stress, a pipeline stress detection equipment based on magnetic memory method (MMM) is developed. Firstly, the principle of MMM to detect the stress concentration was introduced. And then an embedded system based on S3C2440 MCU was designed. The weak magnetic field probe is anisotropic magneto resistance (AMR) sensor HMC 5883L. The visualization terminal based Linux system was developed. According to GB/T 35090, the criterion of stress concentration is carried out for the quantitative evaluation of pipeline. There is no significantly difference from that of TSC series detector from the validation experiment. Finally, the residual stress XRD method was used to verify the stress concentration of the samples. The experiment shows that the equipment can reflect the stress concentration in the material and is suitable for the detection of the stress concentration in the conveying pipeline.

Abstract:Current lane detection algorithm is easily affected by lane line wear, occlusion, shadow and so on, which results in low accuracy and robustness of the detection algorithm, a lane detection scheme for parallel Snake coupled with Kalman filter was proposed. Firstly, in order to obtain the parallel properties of left and right boundaries of road, expectation maximization operator was introduced, and the vanishing point was estimated by minimizing the objective function to estimate its homography matrix. In the homogeneous coordinate space, the homography transformation was used to complete the change of the lane line perspective to the bird's eye view. Then, a lane model was established by parameter prediction operator. A parallel Snake lane detection method was constructed by adding parallelism constraints to active contour model, eventually converge on the left and right edges of the road. Finally, taking into account the continuity of the parameters between the front and back frames, the Kalman filter was used to track and optimize, and the noise was suppressed to improve the recognition accuracy of the lane line. Experimental results show that: compared with the commonly used lane detection algorithms, the proposed scheme has improved accuracy and robustness, and achieves good performance in lane datasets such as shadow, illumination change and boundary damage.

Abstract:Sevenlevel inverter has a complex structure and several fault attributes cross each other, thus reducing the discrimination among similar fault classes. Given this, a supervised kernel shared nearest neighbor (SKSNN) algorithm was proposed and applied to locality preserving projection (LPP), which formed a new feature extraction algorithm for The IGBT opencircuit fault feature extraction of cascade sevenlevel inverter. Firstly, the threephase current of the AC side was collected as the original signal corresponding to each fault status. On that basis, the lowdimensional sensitive features embedded in raw data would be extracted by the SKSNNLPP algorithm. Then, the extracted fault features were taken as the input of support vector machine （SVM） to establish fault diagnosis model. Finally, through the comparative analysis of the diagnostic effects, it can be shown that the proposed method is superior to traditional signal processing and statistical analysis methods, which can effectively reduce the misdiagnosis rate of similar fault categories and can achieve 964% diagnostic accuracy.

Abstract:With the application of deep learning method in the field of image processing, the image related intelligent interaction technology has also been rapidly developed. Visual question answering (VQA) collects the image information by asking questions related to the image and ultimately achieves the purpose for enriching the image understanding. Through comprehensive analysis and comparison of related methods of VQA in recent years, the method can be constructively divided into four types according to the model structure: basic model, attention mechanism related model, modular model and external knowledge base model. At the same time, it also points out directions for visual and semantic information processing and future research on visual reasoning in VQA from three aspects.

Abstract:Aiming at the low accuracy of Beidou pseudo ranges that is caused by some noises with unknown reason or inaccurate statistical characteristics, and in order to reduce the influence of state estimation on filtering process because of the inaccuracy of statistical characteristics of noises, a new method of positioning calculation that unscented Kalman filter (UKF) combines with noise statistics estimator is proposed. The method introduces an improved noise estimation algorithm SageHusa and combines it with UKF, then the algorithm adaptively estimates the noise of the system and the observation in real time, and resists the error caused by inaccurate noise in positioning calculation. Finally, a convergent factor is added to UKF when the algorithm updates the filtering state in order to ensure the convergence of the algorithm. The experiment indicates that the adaptive method increases the precision of pseudo ranges positioning around 40% comparing with the traditional unscented Kalman filter, the method can deal with the influence of inaccurate noise, and the algorithm convergence speed of the algorithm is promoted at the same time, it can be also used in positing system with timevarying noise and unknown statistical noise.

Abstract:A method for detecting defects of monomer thermal battery inside the thermal battery is proposed in this paper, which aimed at the problem of the low accuracy of internal assembly fault detection at home and abroad. This detection includes three defects, the overall flipchip of the monomer thermal battery, the assembly sequence of the monomer thermal battery, and the leakage of the monomer thermal battery part are analyzed. Using the improved gray level cooccurrence matrix, HU invariant moment, template matching to analyze the defects of monomer thermal batteries. Finally, proposing a detection method based on weight distribution parameters, which is using CART (Classification and Regression Tree) decision tree for detection. The experimental results show that the accuracy of this method reaches 975% and meets the testing requirements, which provides an effective way for thermal battery defect detection.

Abstract:In order to solve the target tracking problem in complex scenes such as target changes and nonuniform illumination of the scene, this paper proposes a target tracking algorithm based on dynamic adaptive correlation filtering in complex scenes. The algorithm first constructs a set of geometrically distorted target reference images, and then combines the constructed optimal templates for each graph based on the idea of the combined filter. At the same time, in order to avoid the need to prespecify the target expected position in the next frame, the algorithm uses a time seriesbased prediction mechanism to improve the tracking accuracy by considering the kinematics of the target. Finally, the algorithm is designed with a reinitialization mechanism to restart the system in the event of a system failure. The simulation results show that the proposed algorithm has better tracking accuracy and tracking efficiency than the existing ones, which verifies the effectiveness and feasibility of the proposed algorithm.

Abstract:It is difficult to establish an accurate nonlinear model for switched reluctance motor with double salient structure and high saturated magnetic field, and traditional design methods are difficult to obtain optimal scheme. In order to optimize the torque characteristics of switched reluctance motor (SRM) and shorten the starting time of the engine, firstly, a SRM is designed with traditional methods and the design parameters are selected. Then, the kernel extreme learning machine (KELM) is used to establish a nonparametric model of the SRM. Finally, simulated annealing particle swarm optimization (SAPSO) is used to optimize the structural parameters of the motor for higher average torque and lower torque ripple. Simulation results show that the model has the advantage of a higher precision and a faster speed of regression. The average torque is increased by 395 N·m, and the torque ripple is decreased by 023. The conclusion is that the combination of the KELM and the SAPSO is suitable for the design and optimization of the motor, and it has fewer regulation parameters, strong generalization, avoidance of local optimal solution and less timeconsuming.

Abstract:The interested objects or events is often related to moving objects in the scene, so moving object detection become a key step and challenging problem. This paper studies the detection of moving objects with a moving camera, and the homography transform is used as the background motion model. Only the background motion is involved in the estimation of the model without the interference of the foreground motion during the estimation of the model parameters. A dualmode singleGaussian model is adopted to prevent the background model from being contaminated by foreground pixels and the background model is transferred in continuous frames using the motionfusion based compensation method. A foreground probability map is also erected according to the temporal and spatial attributes of the foreground object and is intended for an implementation of the adaptive decision threshold constructed to the selected pixels. Compared with other algorithms on a unified video sequence, the experimental results show that the proposed algorithm has a good performance on detection of which is proved to be a realtime approach, and the precision, the recall and the Fmeasure value increased by 052%, 3337% and 2014% respectively.

Abstract:At present, many remote sensing image fusion algorithms mainly use the pixel features of independent pixels to complete the fusion of image subbands, ignoring the regional correlation of image subbands, resulting in the fusion of image discontinuity and fuzzy effects. Therefore, a remote sensing image fusion algorithm based on IHS transform coupled with adaptive region features is proposed in this paper. The IHS transform is introduced to decompose the multispectral image to obtain the intensity component, and then fuse it with the panchromatic image. The nondownsampling contourlet transform is used to decompose the PAN image and the intensity components to obtain the high and low frequency subband information. The adjustment factor of the lowfrequency subband fusion model is adaptively adjusted by using the regional energy and spatial characteristics of the image, so that the fusion subband can contain more spatial information. A highfrequency subband fusion model is constructed based on the region variance feature of the image, which makes the fusion highfrequency subband contain more texture information. Experimental results show that the proposed algorithm has better spectral and spatial characteristics than the current remote sensing image fusion algorithm.

Abstract:As the key component of wireless communication on wireless networkonchip (WiNoC), wireless router will seriously affect the network performance when congestion occurs. Thus, a highperformance wireless router design scheme deploying cooperative congestion control capacity (C3WR) is proposed. To relieve buffer congestion happened at the receiving end of the wireless interface, a scheme that the wireless input port is connected to other local input buffer is designed to provide multipath transmission for the congested data. Then a WCC algorithm is proposed to mitigate wireless receiver congestion by shunting packets. For the buffer congestion in the sender side of the wireless interface, a dynamic routing policy is adopted to reduce the traffic injection rate of the congested wireless router. Experimental results show that compared with the comparison objects, the proposed scheme can improve the throughput by 3380% and reduce the average latency by 3317% at most, while the area overhead and power consumption are acceptable. This scheme balances the traffic distribution of the network with coordinated congestion control, it ensures the system performance while making full use of the available resources.

Abstract:A study strategy of hybrid state estimation for power system based on twostage linear iteration and basic principles of Kalman filter and linear iteration was proposed, it is designed to solve the problems of lower accuracy of power system mixed state estimation, poorer filtering effect and weaker convergence ability, etc. In the first stage, the measurement data of phasor measurement unit (PMU) were used for linear estimation. In the second stage, where the content of the first stage was combined with the traditional measurement value to estimate the state, and the high frequency characteristic of PMU was used to sample the twostage measurement data iteratively. Which was tested in IEEE 14 and IEEE 57 bus test system, and the results were compared with other hybrid models. The results showed that the estimation accuracy, data convergence and measurement parameter error of this strategy were better than those of other hybrid models.

Abstract:The hybrid actuating system with HA/EHA dissimilar redundancy configuration is the development trend of the future multielectric aircraft actuating system. In order to solve the problem of high energy consumption and serious overheat in the hybrid actuating system, first, the efficiency analysis of HA and EHA that constitute hybrid actuation system should be made, by comparing the efficiency characteristics, an optimization method is proposed to automatically switch to a high efficiency channel according to different working conditions; Then, the operating characteristics of air flight load of four quadrant are analyzed, to load F and actuator velocity conditions constitute the V network as a starting point, proposes a hybrid energy recovery and adaptive conditions of actuator system work mode. Compared with the traditional working mode, the energy consumption is reduced by more than 50%. The effectiveness of the method can be verified by simulation.

Abstract:Aiming at the distributed hierarchical data grid model as tree data grid, an optimal replica placement algorithm is proposed in this paper, in which the number of replicas is k, which is specified by the user. The proposed algorithm consists of two phases. In phase 1,all nodes of binary tree are visited in reverse breadthfirst order, and based on whether a replica of object i is placed on a node or not, the total replication cost including both read and storage cost are calculated in a bottom up manner. In phase 2, based on a recursive process, the read cost and the storage cost calculated in phase 1 are used as input, and replicas are placed using top down procedure in order to minimize the total replication cost. Theoretical analysis and simulation results show that the optimal replica placement algorithm proposed in this paper not only has the lower time complexity, but also outperforms several typical replica placement algorithms at present in terms of the normalized placement cost, effective network usage and percentage of local access.