Abstract:The number of escalators in China has been increasing in recent years. However, regular inspection, supervision, spot check, and other common methods are hardly to inspect the potential failure inside the escalator. An improved fault diagnosis method based on CNN-LSTM neural network is proposed in this paper. Besides, a software for escalator monitoring and fault diagnosis is designed with LabVIEW. Furthermore, an improved method to fuse the shallow and deep data features to improve the accuracy of fault diagnosis based on the CNN-LSTM neural network algorithm is proposed in this paper. The fault diagnosis method proposed is tested with the data of Case Western Reserve University. The results show that the method is efficient and effective, and the fault diagnosis accuracy is 99.4%. Moreover, the software for escalator monitoring and fault diagnosis is designed. there vibration sensors are set on the key parts of the escalator, and monitoring software is used for data acquisition, data display and data storage. After collecting a large amount of operating condition data, the of typical faults can be diagnosed.

Abstract:An extreme gradient boosting (XGBoost) method for photovoltaic array fault diagnosis is proposed to solve the problems of low precision and poor model performance due to the nonlinear output of photovoltaic array and the influence of maximum power point tracking. Firstly, based on the single diode model of photovoltaic cell, a simulation model of photovoltaic array was established, and the output characteristics and fault causes of photovoltaic array were systematically simulated and analyzed by using PVsyst software, and the fault characteristic parameters were obtained and the validity of the selected fault feature parameters is verified by the feature importance ranking. Secondly, the fault characteristics of photovoltaic array under different fault states are extracted, and the fault diagnosis model based on XGBoost is constructed. Finally, grid search and cross validation were used to optimize the hyperparameters of the diagnostic model, and the performance of the diagnostic model was evaluated by confounding matrix calculation. Compared with decision tree, random forest and gradient lifting tree, the results show that the proposed method not only can accurately detect all kinds of faults, but also has better generalization ability and higher diagnosis accuracy.

Abstract:A new serial-fed omnidirectional slot array antenna is proposed. Eight annular slots are etched back to back on both sides of the stripline, and a Y-shaped slot is loaded inside the slot to extend the length of the Y-shaped slot. The top is etched with horizontal grooves to form the final Y-T slot structure to improve the unit radiation performance. Genetic algorithm (GA) is used to further optimize the array so that its working bandwidth is further widened and the gain is improved. According to the design results, the antenna samples are fabricated. The actual measurement shows that the impedance bandwidth |S11|<-10 dB of the antenna is 7.93% (5.56~6.02 GHz), and the peak gain in the bandwidth is higher than 8.99 dBi. Operating at 5.8 GHz, the E-plane sidelobe level is lower than -10.11 dB, the peak gain is 9.43 dBi, and it has stable omnidirectional radiation performance.

Abstract:In order to verify the hardware safety integrity level of a functional safety gas detector used in the explosive atmosphere, a method of FMEDA is adopted to analyze the failure modes and effects and calculate the random hardware failure rates of the gas detector elements, the functional safety related parameters are quantified based on the adopted fault online diagnosis measures. The analysis results show that the safe failure fraction of the gas detector which is 1oo1 architecture, B type system and low demand mode of operation reaches 94%, the average probability of a dangerous failure on demand of the safety function is equal to 6.8×10-4, therefore the hardware design of this functional safety gas detector satisfies the requirements of the safety integrity level SIL 2 specified in the standard IEC 61508.

Abstract:ZPW-2000A track circuit plays an important role in the process of ensuring the safety of train operation, once the failure will cause unpredictable consequences. Therefore, fault prediction of track circuit is of great significance. In this paper, an improved grey GM(1,1) prediction model is proposed to predict and analyze the red band phenomenon of track circuit, which solves the problems of low prediction accuracy and certain error of the traditional grey GM(1,1) prediction model. By introducing the weakening factor to reduce the prediction error caused by the original data fluctuation, and using the rectangle method to optimize the background weight of the traditional model, the optimal background parameters under the constraints were obtained based on the genetic algorithm, and the improved GM(1,1) prediction model was obtained. The performance of the improved prediction model is verified by combining the rail outlet voltage data collected from the signal workshop of railway bureau. The results show that compared with the traditional grey GM(1,1) model, the average relative error of the improved model is reduced by 28.3%, and the improved model has higher prediction accuracy and practical value.

Abstract:To address the problems of inefficiency and lack of robustness to the environment in the current traditional image processing algorithms for tread surface defect detection, this paper proposes an improved tread surface defect detection method based on the Faster RCNN. The improved network first uses Resnet50 as the feature extraction network, and adds a self-attention mechanism to the feature fusion output part of the Feature Pyramid Network to enhance the detection ability of the detection network for small defects, and finally uses the K-means++ clustering algorithm to cluster the anchor frames of the tread defect dataset, and uses the clustering results to customize anchor frames that are more suitable for wheel-to-tread defects.The experimental results show that the improved Faster RCNN network has an average detection speed of 68 ms, an average accuracy (mAP) of 97.3% and an accuracy of 39.3% for the detection of small target defects.

Abstract:Aiming at the problems of low output precision and poor tracking performance of an array system composed of multiple MEMS gyroscopes under dynamic conditions, a new dynamic filtering model and filtering method were proposed. By analyzing the error characteristics of the MEMS gyroscope and modeling the angular velocity dynamically, a random error dynamic filtering model of the array gyroscope based on the angular velocity estimation was constructed. Due to the uncertainty of the model in the dynamic situation, the tracking performance of the traditional method is degraded. A multiple fading factor adaptive Kalman filter based on variational Bayesian method algorithm was designed. The variational Bayesian method and strong tracking theory were used to improve the estimation accuracy, convergence speed and robustness. Finally, the static and dynamic experiments were carried out on the high-precision turntable. The experimental results show that under static conditions, the variance is reduced to 4% of a single gyro, and the zero bias instability is reduced to 47.2%; Under the dynamic condition, it can effectively track the change of angular velocity, and the angular velocity residual variance is reduced to 6.2% of that of a single gyro. This algorithm can effectively improve the output accuracy of MEMS Gyro array system.

Abstract:Virtual synchronous generator provides inertia and damping support for the system by simulating the external characteristics of the VSG, so that the inverter system can run safely and stably. Due to the existence of nonlinear load in the new energy inverter system, the output harmonics of the system are polluted, which affects the output power quality. In order to optimize the system output, this paper combined with VSG control algorithm, introduced a random excitation in the noisy system PWM modulation wave, and then compared with the carrier to generate the driving signal. The harmonic distribution is analyzed by voltage power spectral density. Simulation results show that the random excitation method is adopted to VSG inverter output voltage waveform is improved to a certain extent, compared with before introducing random excitation, the output of A, B, C threephase voltage total harmonic factor（THD）was reduced by 11.32%, 11.60% and 11.42%, respectively, show the rationality and validity of this method and suitable for engineering applications.

Abstract:This paper proposes a variable step-size dual-mode blind equalization algorithm that runs the SMMA algorithm and the DD algorithm in parallel. On the basis of maintaining the high performance of the original two algorithms, this algorithm improves the traditional DD algorithm, runs the two algorithms in a parallel structure, and at the same time adds an error control function to control the proportion of the SMMA algorithm, reduce the steady-state error. Theoretical analysis and simulation experiments show that compared with the traditional method, ISI is reduced to -29.8 dB, and the convergence speed is improved, and the algorithm completes the convergence at about 630 symbols. After the completion of the simulation experiment, demodulation equalization of the measured signals was conducted, and the EVM was reduced to 1.69%.

Abstract:Aiming at the limitation of a single distance metric in the traditional indoor WiFi fingerprinting algorithm and the relationship between dBm representation and power is not considered, an indoor WiFi fingerprinting algorithm based on voting mechanism is proposed. After collecting the received signal strength (RSS) data, first, preprocess the RSS data. Then, based on the voting mechanism, the nearest neighbors selected by each distance metric are intersected to form common neighbors, and count each the frequency of common neighbor points. Finally, the final positioning result is obtained by probability weighting. Experimental results show that the proposed method achieves a localization accuracy of 1.63 m, and the average localization accuracy is improved by 10%, 33%, and 58%, respectively, compared with the localization accuracy of KNN, Spearman, and KTCC methods. Furthermore, the localization accuracy is improved by 12% compared to the optimal localization accuracy of 1.86 m in the MAN2 dataset.

Abstract:Harris Hawk algorithm has problems such as easy precocious puberty, falling into local optimal traps, and poor stability. In order to improve the performance of the algorithm, this paper proposes an improved Harris Hawk algorithm using deep deterministic policy gradient (DDPG).DDPGHHO combines deep reinforcement learning with heuristic algorithm, trains neural network by using deep deterministic policy gradient, dynamically generates key parameters of HHO through neural network, balances global search and local search, and endows the algorithm with the ability to jump out of local optimal traps in the later period. Through the comparative experiments of function optimization and path planning, the results show that the DDPGHHO has certain generalization and excellent stability, and can search the better path in different environments.

Abstract:Aiming at the problems of redundant input characteristic data, poor anti-interference ability and limited predictive accuracy in short-term photovoltaic power prediction, short term forecasting of photovoltaic output based on feature factor selection and IVMD-GLSSVM is proposed. Firstly, GRA-KCC is used to analyze the characteristic factors that affect the photovoltaic power, and extract the extremely relevant characteristic factors that affect the photovoltaic power. Then IVMD is used to decompose the photovoltaic power data to reduce the impact of data nonlinearity and volatility on prediction accuracy. Then each modal component is input into the GLSSVM prediction model for prediction, and the superposition of the prediction results of each subsequence is the final prediction result. Finally, the prediction model and other models are verified and analyzed in MATLAB. The results show that the proposed prediction model has strong anti-interference ability and high prediction accuracy.

Abstract:In order to improve the stable tracking control of airborne Opto-electronic platform, a model predictive control algorithm based on linear quadratic enhanced Kalman filter is proposed. A dynamic model of airborne Opto-electronic platform is established. Base on the state Kalman filter, a linear quadratic regulator gain is introduced to reduce the phase delay of the estimated state, which makes the state estimation more accurate. The maximum error between the simulation results of tracking target and the state estimation results of Kalman filter is reduced by 58.14%, and the maximum error between the extended Kalman filter and the state estimation is reduced by 52.62%. The simulation results show that the algorithm can effectively improve the tracking and control performance of the airborne optoelectronic platform, and realize the stable tracking and control of the airborne optoelectronic platform.

Abstract:Aiming at the problems that the basic particle swarm optimization (PSO) Algorithm is fast in convergence and easy to premature maturity, and is prone to fall into local misunderstandings, this paper proposes a particle swarm-artificial bee swarm hybrid (PSO-ABC) algorithm, which is applied to path planning in the three-dimensional of UAV. Based on the improved PSO, the algorithm integrates the ABC algorithm to plan globally the three-dimensional path of UAV. First, the nonlinear inertia weight and shrinkage factor are introduced to improve the particle velocity formula, and then the search operator of the ABC is used to search for the optimal solution again, which solves the problem that the PSO algorithm falls into a local misunderstanding due to its poor local search ability. In this paper, two groups of experiments are set up in a three-dimensional environment to compare the path optimization performance of PSO-ABC algorithm, PSO and ABC algorithm. The experimental results show that the path optimization ability of the algorithm proposed in this paper has been improved, which is 6.1% higher than that of the PSO, and 6.9% higher than that of the ABC.

Abstract:The internal defects of solar cells are the main reason for reducing the current conduction efficiency. The intensity of image defects after imaging by Electrolumine-scence （EL) or Photoluminescence （PL) varies greatly, and direct threshold segmentation will cause missed detection. This paper proposes a full-attention FSA-UNet network for hybrid defect segmentation in solar cells. Aiming at the characteristics of defect stratification, a feature enhancement module is designed to improve the ability to distinguish weak defects, and at the same time improve the backbone feature extraction network to speed up the detection efficiency of strong defects. This algorithm can accurately segment a variety of defects in single crystal silicon wafers. In order to verify the effectiveness of the algorithm in this paper, comparing the algorithm in this paper with U-net and DeepLabV3+, the best MIOU reaches 77.9%, which highlights the advantages of this algorithm.

Abstract:Aiming at the problem of driver facial fatigue detection, a driver fatigue detection algorithm based on genetic algorithm optimized probabilistic neural network (PNN) was proposed. The face detector based on HOG feature is used to detect the face, and ERT algorithm is used to locate the key points. The four fatigue characteristic parameters including PERCLOS value, blink frequency, the proportion of yawning time per unit time and the frequency of nodding were calculated and input into PNN for fatigue discrimination, and the genetic algorithm was used to optimize the smoothing factor of PNN. Improve the accuracy of fatigue classification. NHTU-DDD dataset and YawDD dataset were used to train the network, and self-collected samples were used to verify the generalization performance of the model. Compared with SVM, BP neural network and unoptimized PNN model, the accuracy rates of SVM, BP neural network and unoptimized PNN were 95.67%, 97.67% and 95.33%, respectively. The accuracy of the proposed optimized PNN model is 98.67%, which verifies the effectiveness of the proposed algorithm.

Abstract:In view of the current problems that the video extensometer is difficult to mark in practical applications and the mark is easy to fall off in the test, a set of video extensometer system that can realize high-precision displacement tracking measurement based on the texture characteristics of the rebar itself without special manual marks was designed for the performance test of rebar. The hardware composition, selection and basic working principle of the video extensometer measurement system were introduced. The image matching algorithm used was mainly studied, and the selection of the position and size of the matching template, the sub-pixel positioning algorithm based on surface fitting, and the update of the matching template in dynamic test were analyzed and discussed. Finally, static test, rigid displacement test and dynamic test all verify that the designed video extensometer system can meet the real-time and accuracy requirements of the measurements.

Abstract:To tackle the difficulty in deploying on the side devices of the ARM platform triggered by the huge computation amount and memory occupation of deep learningbased object detection algorithms, this paper presents a lightweight method based on ARM platform object detection. Innovatively, this research adds constraints to the scaling factor of the batched normalized layer and the convolution kernel parameter of the convolution layer in the network, performs sparse training, and uses the scaling factor and the convolution kernel parameter as two criteria for judging the importance of channels and thus pruning the unimportant channels. Furthermore, CBAM attention is adopted to achieve lightweight structure replacement of layers with poor pruning effect. On this basis, the model processed with structure replacement is re-trained to eventually build the final model. Lastly, the optimized YOLOv5n and YOLOv5s are tested respectively in single-object detection and multi-object detection scenarios. The test results show that the method proposed in this research is superior to the conventional lightweight method on ARM devices. In the character detection scenario, the size of the optimized YOLOv5n model is just 0.68 MB, and the detection speed can reach 45 fps when the single-core CPU is deployed on ARM devices, which can well meet the realtime requirements and also greatly reduce the difficulty and hardware cost of the deployment on side devices.

Abstract:Aiming at the problem of fast and accurate positioning of car headlights at car inspection stations and preventing car from replacing inspection, a car front detection dataset Car-Data is established. To solve the problems of car detection in the vehicle inspection station scene, a lightweight car front detection algorithm based on YOLOv5m is proposed. First, the convolution block of the original network is replaced by an improved cross stage depth separable convolution block to reduce the parameters and computation of the network as a whole. Then, the spatial pyramid pooling module in the backbone extraction network of YOLOv5m is replaced with the spatial pyramid dilated convolution module of the enhanced receptive field, thereby improving the object detection accuracy of the network. Finally, the upsampling method is modified in the neck feature enhancement network, and an upper and lower layer feature fusion module is proposed to reduce the loss of feature information. The experimental results on the Car-Data show that compared with the original YOLOv5m, the size of the improved algorithm is reduced by 48%, the number of detection frames per second is increased by about 10 frames, and the detection accuracy is still improved by 2.02 percentage points. Therefore, the improved algorithm can meet the needs of car front detection in the car detection scene of the car inspection station.

Abstract:In road traffic motorcycle accidents, failure to wear a helmet is the leading cause of fatal injuries to riders. Aiming at the problems of false detection and missed detection in the current helmet detection due to the similarity in color and shape of black hair, hat and helmet, a motorcycle helmet detection algorithm with triplet attention mechanism and bidirectional cross-scale feature fusion is proposed. First, a triplet attention mechanism is introduced into the backbone network of YOLOV5s, which extracts semantic dependencies between different dimensions, eliminates the indirect correspondence between channels and weights, and improves detection accuracy by paying attention to the differences between similar samples. Second, the EIOU bounding loss function is used to optimize the detection effect of occluded and overlapping objects. Finally, the weighted bidirectional feature pyramid network structure is adopted in the feature pyramid to achieve efficient bidirectional cross-scale connection and weighted feature fusion, which enhances the network feature extraction capability. The experimental results show that the improved algorithm achieves 98.7% mAP@0.5 and 94.0% mAP@0.5:0.95. Compared with the original algorithm, the improved algorithm′s mAP@0.5 increases by 3.9% and mAP@0.5:0.95 increases by 7.6%, with higher accuracy and stronger generalization ability.

Abstract:Aiming at the poor estimation of the existing stereo matching algorithms on the details of objects and the fact that supervised algorithms relying on a large number of groundtruth disparity maps, this paper proposes a self-supervised stereo matching algorithm combining deep and shallow features. The algorithm embeds Efficient Channel Attention in the feature extraction network to extract shallow and more expressive deep features of the picture. The cost volume predicting initial disparities are constructed based on the deep features, and the shallow features are used to guide the optimization of the initial disparities. In addition, in the loss function section, on the basis of the left and right disparity consistency loss, this paper proposes the left and right feature consistency loss, which strengthens the constraint effect of shallow feature information on disparity maps and improves the robustness of the algorithm. This article trains and evaluates on the KITTI 2015 dataset and applies it to the actual scenes taken by us. Experimental results show that the proposed method can achieve better results than other algorithms, especially in the details where the disparity changes suddenly.

Abstract:Aiming at the problems of high cost, complex installation, and poor quality of 3D contour reconstruction at present, a vehicle contour dimension measurement method based on binocular vision is proposed. The method firstly corrects the vehicle image pair collected by the binocular camera, and calculates and generates the vehicle disparity map through the improved stereo matching algorithm. Based on the principle of 3D measurement of binocular vision, the disparity information of the vehicle contour is calculated for 3D reconstruction, and the vehicle point cloud is generated. Aiming at the problem of missing vehicle contour data in camera blind area, a point cloud symmetry repair method based on license plate recognition is designed to generate a complete three-dimensional vehicle contour. The experimental results show that the measurement and indication errors of the three vehicle models are all less than 1%, and the vehicle model reconstruction integrity is high.

Abstract:Defect detection on ceramic tile surface is an indispensable process during its production. To meet the needs of automatic detection, this paper proposes an algorithm which is based on adaptive wavelet transform, to perform defect bulge detection on ceramic tile surface with complex stereoscopic texture interference. Firstly, the red channel image of the ceramic tile is extracted and preprocessed by Gaussian filtering to suppress noise. Secondly, adaptive wavelet transform and linear median filtering are used to enhance the contrast between the bulge and the background area. Finally, the methods of binarization and morphology are applied to obtain the information of the bulge area. It is noted that the algorithm can detect the bulging defects of ceramic tiles of complex three-dimensional texture, with an accuracy of up to 98.5% and recall rate over 95%.

Abstract:The classification and segmentation of 3D laser point cloud have a positive role in promoting the development of 3D reconstruction and automatic driving technology. The 3D laser point cloud data has the characteristics of disorder, irregularity, and sparsity, so the research of 3D laser point cloud classification and segmentation faces many challenges. The point cloud transformer (PCT) classification network uses the scalar attention mechanism to extract the local features of 3D laser point clouds. It has a good 3D laser point cloud feature learning ability and shows advanced classification accuracy in 3D laser point cloud classification and segmentation tasks. However, when PCT downsamples the 3D laser point cloud data, it ignores the influence of its sparsity on the geometric structure, so it cannot fully extract the local features, resulting in the degradation of the classification and segmentation accuracy of the 3D laser point cloud. To solve this problem, this paper proposes a three-dimensional laser point cloud classification and segmentation network GAM-PCT based on the attention mechanism. Specifically, the GAM-PCT network uses the vector attention mechanism to adjust the weight of the single channel features and uses the subtraction relationship and neighborhood location coding to obtain the attention features of the three-dimensional laser point cloud neighborhood, At the same time, a plug and play geometric affine (GAM) module is inserted to solve the sparsity problem of the local area of the three-dimensional laser point cloud when downsampling the whole point cloud, thereby improving the classification accuracy of the network. The experimental results show that, compared with the PCT three-dimensional laser point cloud classification and segmentation network, the classification accuracy of the proposed GAM-PCT network on the data set modelnet40 is increased by 0.3%, while the classification accuracy on the ScanObjectNN data set is increased by 1.9%, and the average intersection ratio of segmentation on the shipment data set is increased by 0.2%. At the same time, the network parameters and the flops index are reduced by 0.31 g and 0.69 m respectively. The experimental results show that the complexity of the improved network is simplified, which fully verifies the effectiveness of the improved method.

Abstract:For distributed optical fiber temperature measurement system, the back Raman scattering signal containing temperature information generated by it is extremely weak, so it is easy to be masked by white noise. In this paper, a joint noise reduction algorithm based on VMD-SVD is designed, and the important parameters of the algorithm are selected with better noise reduction performance indicators by comparing the noise reduction of the constructed simulation signal. Experiments show that when the joint algorithm and EMD, VMD, EMD-SVD algorithm are used to reduce the noise of multiple groups of test signals respectively, the VMD-SVD noise reduction algorithm has an improvement of more than 15 dB compared with the EMD algorithm, and more than 9 dB compared with the VMD algorithm. Finally, compared with the EMD-SVD algorithm, the VMD-SVD algorithm has an improvement of more than 1 dB. Finally, when the above algorithm is applied to reduce the noise of multiple groups of backward Raman scattering signals measured by the distributed optical fiber temperature measurement system, VMD-SVD noise reduction algorithm can also effectively eliminate the white noise of the signal, laying the foundation for subsequent high-quality temperature measurement.

Abstract:Poststroke depression (PSD) is one of the common complications after stroke, which seriously affects the rehabilitation of stroke patients. At present, the diagnosis of PSD is mainly based on the clinical manifestations of patients with various scales, but this method has certain subjectivity. Electroencephalography (EEG) combined with deep learning techniques has the potential to provide objective criteria for the diagnosis of PSD. In this study, we collected EEG signals from 28 subjects without poststroke depression (PSND) and 38 subjects with poststroke mild depression (PSMD), and proposed an end-to-end PSD diagnostic framework, which combines long short-term memory (LSTM) based on attention mechanism with convolutional neural network (EEGNet). LSTM model is used to learn the time-series dependencies of the EEG signal. attention mechanism assigns weights to the time domain information to improve the utilization of useful information. Finally, EEGNet module is used to extract more representative deep features in the EEG signal. The results showed that the accuracy, precision, recall, F1-Score and Kappa coefficient obtained by 10-fold cross-validation were 95.90%, 95.75%, 96%, 95.82% and 91.60%. Compared with the basic deep learning model for EEG-based PSD classification, our method maintains stable model performance and has high accuracy for the diagnosis of PSD, which provides a certain reference for the screening and diagnosis of PSD.

Abstract:Magnetic flux leakage detection technology is a common detection method in the field of pipeline internal detection. The analysis base on magnetic flux leakage signal is of great significance for pipeline safety evaluation. Establishing a two-dimensional simulation model of magnetic flux leakage detection, and studying the influence of excitation intensity on magnetic flux leakage signal in magnetic flux leakage detection; Based on the physical model of the influence of magnetization state on MFL signal, the excitation intensity is divided into three stages, namely, initial growth stage, nonlinear growth stage and saturated linear growth stage. The results show that the dividing points of the three stages is only affected by the depth of defects, and the simulation results are in good agreement with the physical model. The divided three stages effectively measure the influence of excitation intensity in pipeline magnetic flux leakage detection, which has guiding significance for magnetic flux leakage signal acquisition and analysis.