Abstract:Condensation particle counter (CPC) is a kind of instrument for detecting submicron and nanometer aerosol particles, which has important applications in environmental monitoring and meteorological research. Some international institutions have realized the mass production of CPC. However, the relevant research in China lags behind. Based on the research results of domestic and foreign scholars on CPC, this paper discusses the development of CPC in the past 40 years. Different optimization schemes for CPC structure are introduced, and the influence of growth temperature, particle concentration and working fluid properties on the performance of CPC detection is discussed. It is found that the addition of sheath flow can significantly reduce the cutting particle size of CPC. The properties of working fluid directly determine the counting efficiency of CPC. Reasonable control of temperature difference between condenser and saturator and particle concentration can also improve the counting efficiency of CPC. After that, several mainstream CPC calibration devices and calibration technologies are introduced, and their advantages and disadvantages are compared. Finally, the test performance and main applications of different types of CPC are discussed.

Abstract:Design of angular displacement measuring system for precision reducer

Abstract:In the manufacturing process of nanometer time-grating sensor, the manufacturing errors caused by the processing technology are mainly manifested as the electrode geometric errors. By using the method of sectional area integral to build the mathematical model, the influence of electrode geometric errors on the measurement accuracy is analyzed in detail. It is revealed that there is an averaging effect when the signal is picked up by multiple induction electrodes, which can effectively homogenize the measurement errors caused by the random variation of electrode geometric errors. Two prototype sensors with a range of 200 mm are fabricated by micro-nano-machining technology with a manufacturing accuracy of 1 μm and printed circuit board (PCB) technology with a manufacturing accuracy of 10 μm, respectively. Accuracy comparison experiment is carried out and experimental results show that, due to the averaging effect, the prototype made by PCB technology has achieved a measurement accuracy of ±250 nm in the full range after simple linear compensation, which is close to the measurement accuracy of the prototype made by micro-nano-machining technology. Experimental results verify the effectiveness of the averaging effect of multiple induction electrodes.

Abstract:In order to solve the problem that it is difficult to effectively compensate the error caused by the angle in ultrasonic ranging, based on the theory and method of function approximation, a compensation method of the error caused by the angle in ultrasonic ranging is proposed. Firstly, the propagation and incidence process of ultrasonic pulse are simulated. The simulation results show that the propagation speed of ultrasonic pulse is different under different ranging angles, which becomes the medium of error introduction. Then, the correlation relationship between the ranging angle and the error under the action of this medium is analyzed through experiments, the basis function model combination method is used to construct the ultrasonic ranging angle error model. Finally, aiming at the problem that the model independent variables (measured distance and ranging angle) must be known values, and error compensation cannot be achieved in practice, the measured value of ranging is taken as the variable of the iterative calculation, the model is taken as the relational expression of the iterative calculation, and a compensation algorithm of the error caused by the ultrasonic ranging angle is designed. The actual measurement verifies that the algorithm can make the average value of the ranging error less than 1. 1 mm when the ranging angle changes, which effectively compensates the error caused by the ranging angle and improves the accuracy of ultrasonic positioning.

Abstract:When the electronic equipment works, the electromagnetic property is greatly affected by temperature. How to reconstruct the three-dimensional temperature field of electronic equipment with sparse sensor measurement data in real time is the key problem to realize health monitoring and electromagnetic property control. Aiming at this problem, taking active phased array antenna (APAA) as the research object, a temperature field temporal spatial reconstruction method is proposed based on proper orthogonal decompositionregularized least square-Kalman filter (POD-RLS-KF), which fuses simulation data and sparse sensor measurement data. Firstly, POD is used to separate the simulated transient temperature field data into temporal mode coefficients ( TMC) and spatial basis functions ( SBF). Secondly, the data collected by sparse sensors are used to carry out the optimal estimation of the TMC in real time with RLS and KF methods. Finally, the temperature field is reconstructed in real time combining the SBF and updated TMC. The simulation and experiment results show that when the simulation state is different from the actual state, the method can calculate the three-dimensional temperature field of the antenna in real time according to a small amount of data provided by sparse sensors. Compared with existing reconstruction methods, the proposed method shows better reconstruction accuracy and noise suppression ability, and the average of the root mean square error of the reconstruction temperature field is 7. 18% of that of Kalman filter-linear stochastic estimation (KF-LSE) method and 1. 53% of that of the Gappy method.

Abstract:Atomic force microscope ( AFM) uses the interaction force between the probe and the object to be measured to achieve imaging, and obtains the relevant geometric parameters through obtaining the high-resolution imaging of the rectangular nano grating measurement standard instrument and performs calibration, so as to realize the quantity transfer from standard measurement instrument to working measurement instrument. In the AFM scanning process, due to the influence of the needle tip, the scanned image is the result of the interaction of the probe and the sample, rather than a true description of the sample morphology. Aiming at this phenomenon, this paper proposes an AFM image restoration method based on long short-term memory (LSTM) network. This method trains the scan lines of the simulation image obtained with the expansion method, and then obtains the AFM image restoration model suitable for rectangular nano grating. Experiment results show that for a rectangular nano-grating with a line width of 20 nm and a height of 40 nm, the relative error of the grating line width after restoration with the proposed method is 7. 40% , the proposed method further improves the measurement accuracy compared with the traditional restoration method.

Abstract:In order to detect isopropanol gas harmful to human health, a high sensitive isopropanol gas sensor based on In2O3 and Fe2O3 loaded In2O3(Fe2O3@ In2O3 ) hollow sphere was successfully prepared. The crystal phase, morphology and chemical composition of the prepared In2O3 and Fe2O3@ In2O3 sample were analyzed with characterization methods. The experiment test results show that the hollow sphere gas sensor of Fe2O3 @ In2O3 has better gas-sensing characteristics to isopropanol than pure In2O3 under the optimal operating temperature of 200℃ . The response value of the hollow sphere gas sensor of Fe2O3@ In2O3 can reach 28. 2 (which is 1. 75 times that of pure phase In2O3 ), and the response time and recovery time are 1 s and 2 s, respectively, the gas sensor also has good repeatability and stability. An n-n heterojunction is formed between the two oxides of Fe2O3 and In2O3 , which increases the initial resistance of the sensor, thereby optimizes the gas-sensing characteristics of In2O3 . The prepared hollow sphere gas sensor of Fe2O3@ In2O3 has a wide application prospect in detecting isopropanol gas.

Abstract:In modern robotics and flexible electronic applications, there is an urgent need for flexible tactile sensors with high sensitivity, good flexibility and three-dimensional force measurement. A tactile sensor based on polyvinylidene fluoride (PVDF) piezoelectric film is proposed to achieve dynamic three-dimensional force measurement. The sensor consists of three PVDF piezoelectric sensitive units with an angle of 120 degrees to form three pressure sensitive areas, and the three piezoelectric sensitive units are arranged on the bottom surface of the polydimethylsiloxane (PDMS) hemispherical structure. The triaxial contact force transmitted from the top of the PDMS hemisphere causes the change of piezoelectric charge in the three piezoelectric sensitive units, thereby calculating the direction and magnitude of the contact force. This paper derives a three-dimensional force algorithm of the tactile sensor with three sensitive units, and gives a calibration for the prepared sensor to realize the three-dimensional dynamic force test. Results show that the average error of angle θ is 7. 75% , the average error of angle φ is 12. 17% and the average error of force F is 7. 48% . The three-dimensional force tactile sensor has good application potential in wearable electronics, health care, and human-computer interaction and other fields.

Abstract:Due to the advantages of low power consumption, fast response and high measurement accuracy, Micro flowmeter have a wide application in the fields of automotive industry, aerospace engineering, biological research, clinical diagnosis and so on. Calorimetric micro flowmeter determines the flow velocity value by measuring the temperature difference between the upper and lower reaches of heater. However, when the flow velocity exceeds a certain threshold, the temperature difference will not increase with the flow velocity, which limits the measuring range of the micro flowmeter. In this paper, numerical simulations for a wide range of flowrates (0. 01-160 SCCM) are firstly carried out to investigate the response of temperature difference to flowrate. It is found that the response of temperature difference to flow velocity is almost linear when diffusion is dominant. With the increase of flow velocity, the response of temperature difference decreases gradually and is nonlinear. As the flow rate increases further, the response of temperature difference becomes saturated and the calorimetric operation mode fails. Secondly, on the basis of numerical simulation, a dual-mode micro-heat flow sensor is proposed, which adopts hot wire operation mode at high flow rate and calorimetric mode at medium and small flow rate. Finally, a micro flowmeter is manufactured by 0. 18 μm CMOS-MEMS technology, and the threshold for mode-switch is defined by the sensitivity in experiments. Compared with standard flowmeter, the error is less than 2% , which is in line with the practical application requirements, but the range is doubled.

Abstract:The key of improving the transmission stability of the light wave phase difference in the optical path of the optical voltage sensor is the suppression of the parasitic interference phase difference of the system optical path. To this end, an optical voltage sensor based on the Pockels effect of crystal material and combined with the sagnac optical path structure is studied. Based on the analysis of the reciprocity mechanism of the optical path structure, it is proposed to change the system optical path to use panda-type polarization maintaining fiber and low the combined form of the birefringent rotating fiber can eliminate the mode phase difference generated by the light-sensitive circuit section, thereby improving the optical reciprocity of the sensor. A mathematical model of the temperature field and thermal stress of the optical path of the system is established, and the influence of the external environment temperature on its phase difference propagation is studied by finite element simulation. It is obtained that for every 1℃ increase in the external temperature, the optical path birefringence decreases by 2. 275 21×10 -7 on average. Based on this, an experimental study on the reciprocity optical path has been carried out, and experiment shows that the maximum influence of temperature on the measurement result is 1. 733 9% / ℃ . When the ambient temperature rises by 10℃ , compared to the optical path, all panda-type polarization-maintaining fibers are used for sensor measurement. The accuracy is increased from 6. 008% to 1. 53% ; when the temperature rises by 20℃ , the accuracy is increased from 52. 016% to 8. 13% . The feasibility of the reciprocal optical path can be preliminarily proved.

Abstract:Multi solution is a difficult problem in minimal solution method for extrinsic calibration of 2D laser radar and camera. This paper proposes three improvement measures for this problem and applies them to the minimum solution calibration process. Firstly, the observable spatial constraints of the sensor are used to eliminate the false solutions in which checkerboard cannot be observed by these sensors. Then, the error evaluation function of true solution selection is improved by using the constraint between the boundary laser points and boundary line segments. Finally, the least square solution of the translation parameters is solved by using chessboard boundary constraints. Simulation data and real data are tested in the experiment, and results show that the method can improve the hit rate of the true solution and calibration accuracy compared with Francisco method. When the noise variance of laser point is less than 20 mm, the hit rate of the true solution is about 98% , which meets the requirements of practical application. Compared with the traditional minimum solution calibration method, this paper uses checkerboard boundary constraints and observable spatial constraints to choose multiple solution sets, and optimizes the translation parameters. As a result, it can effectively improve the performance of the minimum solution calibration method.

Abstract:The satellite signals of multi-UAV systems are susceptible to interference in complex environments. To address this issue, a cooperative navigation method based on loop sum-product algorithm is proposed. According to sensor′s characteristics, the absolute navigation based on the GNSS / INS tight combination and UWB-assisted satellite double-difference tight combined relative navigation are designed. Each platform′s absolute navigation information and relative navigation information between platforms are used to construct a collaborative navigation filter. Message transmission rules are designed to reduce the flow between platforms based on loop sum-product algorithm. Calculation tasks are distributed and run on each platform, which reduces calculation pressure. Finally, simulation and UAV experiments are carried out for different complex scenes. Results show that this scheme′s cooperative navigation accuracy is significantly better than the result without cooperative navigation. With the increase of observable information, this method′s navigation accuracy is getting higher and higher.

Abstract:According to recent research, plantar pressure can reflect various characteristics of the human body, which is promising for biometric identification. In our study, the feasibility and methodology of biometric identification by plantar pressure is discussed. Insoles with eight pressure sensors are used to collect over 14 000 steps of 14 participant as database. After that, the scientificity of classification is discussed by unsupervised learning, and the influence of ground conditions on pressure data is discussed. Convolutional neural network (CNN) is used as the classifier to evaluate the classification performance, and the effects of gait segmentation and multiple gait cycles on improving the accuracy are studied. Experimental results show that the accuracy of data classification after segmentation is 98. 8% , while that without segmentation is 93. 6% . When using 3 and 5 gait cycles for classification, the accuracy rise up to 99. 4% and 99. 8% . The results suggest that CNN with segmented data and selecting multiple gait cycles for classification has practical value in biometric identification utilizing plantar pressure.

Abstract:Aiming at the abdominal electrode-sourced ( AES) signals, a fetal electrocardiogram ( FECG) signal extraction method utilizing extended Kalman filter (EKF) combined with FastICA is proposed. Firstly, the raw maternal abdominal mixed signals are preprocessed to suppress baseline drift, power frequency interference and pulse artifacts. Then, FastICA is used to separate the maternal electrocardiogram (MECG) signal estimation and FECG signal estimation containing residual MECG component and other noise from maternal abdominal mixed signals. The EKF is used to filter the FECG signal estimation to obtain the residual MECG component estimation, which is suppressed to obtain the noisy FECG signal estimation. At last, the clear FECG signal is extracted using EKF again. The clinical data were adopted to verify the proposed FECG extraction method. The sensitivity, positive predictive value and F1 score of the proposed FECG extraction method are 99. 27% , 94. 35% , 96. 71% , respectively, and the signal-to-noise ratios based on cross correlation and singular value decomposition are 6. 145 4 dB and 6. 509 6 dB, respectively. The experiment results show that the proposed method is better than traditional FECG signal extraction methods both on subjective visual effect and objective assessment index

Abstract:Current non-contact sleeping electrocardiogram (ECG) monitoring technology has a limitation when used to effectively acquire the multi-positional human ECG signals. This study developed a type of ECG monitoring system that can measure sleeping ECG signals under multi-positional conditions through clothes and bed sheets. A flexible conductive fabric is used as the electrode material. The shape of ECG electrode is similar as an arched door. The variation of electrical field is sensed by capacitive coupling. The ECG signals are coupled from human body to the fabric electrodes. Combined with a large resistance of 10 GΩ as a bias resistor, an instrumentation amplifier with high input impedance is used for differential signal processing. A digital filtering technology is used to remove the noise in the raw ECG signals, while the finite state machine (FSM) logic algorithm is applied to extract the features of ECG signals. The ECG signals detected from human subjects experimentally by the proposed system are compared with those detected by the strip-type noncontact ECG electrodes as well as traditional contact ECG acquisition system. Experiment results show that this system can obtain decent sleeping ECG signals under non-contact conditions, and the detection rate of R wave of ECG signal is over 95% .

Abstract:This paper aims to establish a milling temperature model of the spinal lamina, which mainly considers the influence of bone density and milling parameters ( bone cutting depth and milling feed speed) on the milling temperature. Firstly, based on the analysis of the lamina′s layer-by-layer bone-cutting process, a series of bone-cutting experiments with different parameters are designed by the full factor experiment method, and the emissivity of the cancellous bone materials is calibrated. Then, the temperature data needed to establish and validate the model are collected by cutting layers of artificial cancellous bone materials with different densities at different depths and feed rates using a robot and an orthopedic ball-end milling cutter. A thermal imaging camera measures two kinds of milling temperatures during the process of the layer-by-layer cutting by the robot, which are the temperature of the milling cutter and the temperature of the bone surface. Finally, the influence of the bone density and milling motion parameters on these two types of milling temperatures is analyzed, and a prediction model of lamina′s cancellous bone milling temperature is established using experimental data and a neural network. Experimental results show that the goodness of fit between the temperature value estimated by the model and the measured value of the bone cutting experiment is 0. 97. The proposed model can help surgeons or robots to select appropriate milling motion parameters when milling cancellous bone with different densities layer by layer to improve the safety of laminectomy.

Abstract:Knee osteoarthritis (OA) is one of the main causes of activity limitation and physical disability in the elderly. Early diagnosis and intervention of knee osteoarthritis can help patients slow down the deterioration of OA. At present, the early diagnosis of knee osteoarthritis is detected by X-rays and scored according to the Kellgren-Lawrence (KL) grade. However, doctors′ scores are relatively subjective and vary from doctor to doctor. Grade classification of knee osteoarthritis is a matter of orderly classification. The ordinal penalty loss function assigns higher penalty weights to the classes that are further away from the ground truth, which is more suitable for knee osteoarthritis classification. In existing works, the penalty weights no longer change during training procedure, so the training model often fails to reach the expected results. In this paper, an adaptive ordinal penalty adjustment strategy is proposed to address the shortcomings of the ordinal penalty loss, in which the penalty weights are automatically tuned in reverse according to the confusion matrix obtained at each stage (epoch). Furthermore, the performance of the proposed method is validated on several classical CNN models such as ResNet, VGG, DenseNet and Inception by X-ray image data from Osteoarthritis Initiative (OAI). Experimental results show that the adaptive ordinal penalty adjustment strategy proposed in this paper can effectively improve the classification accuracy (AC) and mean absolute error (MAE) of the model when the initial weight score difference is small.

Abstract:In order to improve the operation efficiency and power efficiency of artificial intelligence accelerator, proposes a new convolutional neural network (CNN) accelerator, and realizes a computing-in-memory method. Firstly, a neural network architecture is designed, which has the characteristics of highly parallel computing and efficient operation of MAC unit. Secondly, in order to reduce power consumption and die size, a symmetric SRAM array and an adjustable data flow structure are adopted to realize the efficient computation of multi-layer network in SRAM, which reduces the times of external memory access and the power consumption of SoC system. Operation efficiency is improved as well. Through the 40 nm process of SMIC, the SOC design, tape and test are completed. Results show that the computational power can reach 288 GOPS at 500 MHz, the power consumption at full speed is 89. 4 MW, the area is 1. 514 mm 2 , the computational power consumption ratio is 3. 22TOPS / W and the 40nm computational power area ratio is 95. 1 GOPS / mm 2 . Compared with results in other literatures, the power consumption and area of computing power increase by at least 4. 54% and 134% , respectively, which is more suitable for embedded ends.

Abstract:Aiming at the problem of large number of micro ring resonator (MRR) required by current all-optical full adder, an optical full adder with three Cascaded MRR structures is proposed for the first time. In view of the fact that MRR is sensitive to temperature fluctuation and process variation, the MRR fault model is established. The mean error distance of reliability metric ofoptical full adder (OFA) is designed, and the influence of MRR single fault model on OFA is analyzed. Simulation results of insertion loss show that the proposed OFA architecture is superior to existing OFA architecture in general. Compared with the current scheme, the hardware overhead of the proposed OFA architecture is reduced between 50% and 70% . Experimental results show that the average error distance of scheme 1 and scheme 2 is large, while the average error distance of the proposed scheme is moderate. For the multi-bit binary full adder, the absolute value of the average error distance increases with the number of bits of the multi-bit binary full adder in the single fault model with the highest bit. For the single fault model with the lowest bit, the absolute value of the average error distance remains unchanged with the number of bits of the multi bit binary full adder. Physical verification and experiments based on Modelsim platform verify the correctness of the effect of MRR fault on the performance of full adder.

Abstract:For solid tantalum capacitors with 3,4-ethylenedioxythiophene (PEDOT) conductive polymer, temperature and humidity are used as accelerated test stresses in this paper. Variable interval measurement method is used to build a constant stress accelerated degradation test platform for tantalum capacitors under four stress levels, which are 85℃ / 85% RH、95℃ / 70% RH、 95℃ / 85% RH、 110℃ / 85% RH. Degradation data of the performance degradation parameters capacitance and loss factor are obtained. Aiming at the non-single change trend of the degradation parameters of tantalum capacitor under accelerated stress of temperature and humidity, ordered clustering algorithm is used to divide the degradation interval. Optimal classification number is determined by the change rate of error function slope, and the stable degradation interval of tantalum capacitor is obtained. Based on data reconstruction and Wiener process, the capacitance and the dissipation factor are fitted and the fitting accuracy reaches 97% and 95% , respectively. As a result, the effectiveness of Wiener process model is verified. Combined with Copula function, a binary accelerated degradation model of tantalum capacitor based on random effect Wiener process is established. Reliability evaluation method is researched and the product life under normal stress level which meets the product life requirement of 12 ~ 15 years is deduced. Results show that the binary accelerated degradation model can finish degradation performance analysis and life prediction of tantalum capacitor.

Abstract:Unmanned reclaimers in bulk materials ports has the problems of low reciprocating reclaiming efficiency. Meanwhile, the existing machine learning classification models are not effective because of high noise, frequent fluctuations, and unbalanced data of millimeter-wave radar sensing datasets. In this paper, a stack boundary sensing method based on improved fuzzy twin support vector machine combined with 1-Nearest Neighbor algorithm is proposed. Firstly, the millimeter-wave radar is used to obtain the stack boundary scan data and preprocess it. According to the spatial distribution and operation characteristics, the 10-dimensional features of the point cloud are extracted to form the stack point cloud sample dataset; secondly, the improved fuzzy membership function is introduced. The fuzzy twin support vector machine divides the pile point cloud sample dataset into overlapping and non-overlapping regions. Then, the fuzzy twin support vector machine decision boundary and 1-nearest neighbor algorithm are used to classify the non-overlapping and overlapping region samples respectively to improve the classification ability of unbalanced datasets. Finally, the classification results obtained are added to the perception link to achieve the purpose of sensing the boundary of the pile. Experiments on the dataset collected by manual operation radar show that the proposed perception method effectively improves the ability to recognize minority categories. Field experiments show that the improved perception method is closer to the operator′s judgment, the idle time of the bucket wheel is reduced by 15. 1% , which improves the operating efficiency of the unmanned reclaimer and has reference significance for the construction of unmanned bulk materials ports.

Abstract:Positioning technology based on geomagnetic signal has the advantages of no need to set up signal transmitting equipment, low cost and wide coverage of indoor positioning. Aiming at the present situation that indoor geomagnetic positioning method is complicated and requires repeated measurement to accurately map two-dimensional plane coordinate points, a method based on mobile terminal image visualization mapping and automatic insertion of acquisition data was proposed to quickly collect indoor geomagnetic fingerprints and dynamically establish fingerprint database matching indoor two-dimensional maps. Based on this acquisition method, the improved dynamic time warping (DTW) algorithm is used for geomagnetic sequence matching in the positioning phase, which can reduce the amount of calculation and improve the accuracy of positioning results. Subsequently, particle filter is used to fuse the positioning results based on geomagnetic sequence matching and pedestrian dead reckoning (PDR) results to achieve fast and accurate positioning on the intelligent mobile terminal carrier. Simulation and experimental results show that the proposed method has a pedestrian tracking and positioning efficiency of 65 ms per step and an average positioning accuracy of 1. 4 m in large indoor areas.

Abstract:To give a simple and accurate explicit model for the output characteristic curve of cadmium telluride ( CdTe) photovoltaic modules, the idea of using two second-order Bezier function trajectories to fit the left and right side curves of the maximum power point of the module output characteristics is proposed. The influence rule of the control point on the Bezier function is analyzed, and the existence of the optimal control point is defined. Then, the optimal output characteristics of seven CdTe modules from different manufacturers and types are given. According to the Bezier modeling results under the control points, the fitting law between the optimal control point position of Bezier function and the module filling factor is found out by using the similar triangle theory, and an explicit model describing the output characteristics of CdTe photovoltaic modules is established. Finally, four new CdTe modules are used to evaluate the proposed rules, and the iterative results are compared with the existing models. Experimental results show that the average relative error of CdTe module modeling method based on Bezier function is between 0. 49% and 1. 5% , while the average relative error of existing models is between 2. 45% and 9. 19% , which demonstrates the simplicity and correctness of the proposed model.

Abstract:A wearable and flexible NFC tag antenna is fabricated by using gallium-based liquid metal as flexible conductive material. Through theoretical, simulation and experimental analysis, the NFC tag antenna designed has good flexibility and high bandwidth. Its inductance change and resonance offset degree are very small under large stretching and bending deformation, and the working stability is well. Different from traditional NFC antenna, the antenna designed has good flexibility and stretchability, which can be gently attached to the surface of human skin. The NFC tag prepared by this antenna can communicate wirelessly at the operating frequency of 13. 56 MHz. The maximum communication distance is about 3. 36 cm, and the maximum communication distance after contacting with human wrist skin is about 3. 12 cm. Due to its unique advantages, the flexible NFC tag antenna fabricated has great potential applications in information exchange, personal health data monitoring and wearable biosensors.

Abstract:In order to effectively solve the relocalization problem of mobile robot, a robot relocalization method fusing deep learning and particle filtering is proposed. Firstly, a 3-DOF mobile robot relocalization framework is proposed, which mainly includes two progressive stages: relocalization model construction and robot online relocalization. Secondly, a 3-DOF mobile robot relocalization network model, G_PoseNet, is proposed and constructed based on PoseNet. The pose result predicted by G_PoseNet is used as the initialization state of particle filter localization algorithm to support the subsequent relocalization process. Then, a data model based kidnapping state judging method is proposed to determine whether to start the relocalization process. Finally, a large number of experiments on public datasets and real environment were performed to verify the proposed method. The result shows that the G_PoseNet model can guarantee a certain degree of location prediction accuracy and improve the pose angle prediction accuracy, the success rate of robot relocalization achieves 87% .

Abstract:At present, the health state recognition of industrial robot harmonic reducer is mainly based on vibration signals, which requires additional test system, increases the difficulty and cost of data acquisition, and its accuracy and effectiveness are affected by the installation location of sensors. Based on this, the health state recognition method of harmonic reducer based on depth feature learning of voltage signal is proposed. The industrial robot motor voltage signal is used to characterize the health state of harmonic reducer, and the continuous wavelet transform is used to transform the voltage signal into time-frequency diagram to obtain the time-frequency information of voltage signal under different health state of harmonic reducer, and the data sample set is constructed. The convolutional neural network is used to self-learn the time-frequency information of the voltage signal, and the network parameters are supervised to adjust. In this way, the health state of harmonic reducer can be recognized while the depth characteristics of voltage signal under different health state of harmonic reducer are obtained. Experiment results show that the recognition accuracy of the proposed method reaches 90% above, which proves that the proposed method can effectively recognize the health state of harmonic reducer, and has good generalization ability and robustness.

Abstract:Aiming at the problems of low efficiency and poor robustness of multi-object detection and segmentation for traffic scenes in intelligent driving, a fast Multi-object detection and segmentation for traffic scene based on improved Mask R-CNN is proposed. Firstly, in order to effectively reduce network parameters and compress model volume, the lightweight MobileNet is used as the backbone network to improve the ability of transplant algorithms on the subsequent embedded side. Secondly, by optimizing the convolution structure of FPN and backbone network to ensure the complete transfer of feature information between high-level layers, the improved network model for multi-object detection and segmentation in traffic scenes is obtained by adjusting hyperparameters. Comparative experiments are conducted under different traffic scenarios, the improved network can accurately realize the detection and segmentation of multiple objects and the average detection accuracy can reach 85. 2% . Migration experiments are carried out on the ApolloScape and NuScence dataset to improve the network, which show good generalization capabilities. The improved backbone structure and network structure optimization proposed in this paper can adapt to a variety of complex traffic scenarios and complete the fast detection and segmentation of multiple object in traffic scenarios. It provides theoretical basis and technical solutions for intelligent driving.

Abstract:To solve the problems of multiple water gauge stitching, incomplete water gauge shooting and poor environmental adaptability in water level recognition of water gauge images, a photogrammetric method is proposed to measure the water level of indefinite water gauge image under complex lighting conditions. Firstly, the perspective projection relationship of the water gauge between the pixel coordinate system and the world coordinate system is constructed through the manually selected identification points. Then, the water gauge and the background are segmented and binarized by the color characteristics of the water gauge. According to the binarized image, the average and variance threshold method is used to detect the position of the water level. Finally, the expression of the middle line on the water gauge in the water level image is calculated through the projection relationship. The length of the middle line on the water gauge on the water surface is calculated by combining the position of the water level line to obtain the water level value. In this paper, the measurement experiment is carried out on the real water gauge image of the hydrological station, and the detection results are compared with the remotely measured water level value. Results show that this method could improve the reliability of the water level detection of the indefinite water gauge image. The measurement accuracy is up to 1 cm and the detection resolution is 1 pixel, which basically meet the water level monitoring requirements of hydrological stations.

Abstract:The visual / inertial positioning method represented by the visual / inertial odometry has been widely used recently. The traditional visual / inertial odometry obtains the fixed distortion parameters of camera through offline calibration methods. When the calibration of camera distortion parameters is inaccurate or the camera distortion parameters are changed, the positioning accuracy will decrease. Aiming at this problem, a robust visual / inertial odometry method facing to online self-calibration of camera distortion parameters is proposed. Firstly, the camera distortion parameters are added to the variables to be optimized in the visual / inertial odometry. The Jacobian matrix of the visual reprojection error with respect to the variables to be optimized is derived. Then through the factor graph optimization method, online self-calibration of camera distortion parameters and real-time optimization solution of carrier navigation information are realized. Finally, the effectiveness of the proposed method was verified through the EuRoC dataset and realworld experiments. The real-word experiment result demonstrates that compared with the traditional visual / inertial odometry method, the accuracy of the proposed method is improved by 65. 40% in an outdoor wide open scene.