Abstract:Atrial fibrillation is the most common cardiac arrhythmia in clinical practice, while some realtime automatic detection algorithms have been developed rapidly for improving clinical diagnosis. However, it is difficult to judge the advantages and disadvantages of the existing atrial fibrillation detection algorithms for their lack of comparisons between models or databases. Three different machine learning algorithms, including support vector machine, random forest and logistic regression, were selected to build three separate models to detect atrial fibrillation. The three models were trained on the MITBIH atrial fibrillation database, and were tested and compared on three independent databases respectively. Furthermore, the influence of feature selection on model performance was analyzed. Experimental results showed that when applied 12 features (3 domain features and 9 nonlinear features), the sensitivity, specificity, accuracy and F1score of the three models reached more than 95% on the China physiological signal challenge 2018 public database and the wearable dynamic ECG database. In addition, the random forest algorithm has stronger stability and generalization ability compared with the other two algorithms.

Abstract:The brachialankle pulse wave velocity (baPWV) is usually used as an evaluation index of the elasticity of the blood vessel wall, the commonly used measurement method of baPWV is the method of extremity brachial and ankle polyconduction signals. Aiming at the problem of expensive detection instrument and complex operation, a new method is proposed to detect the degree of arteriosclerosis by collecting the photoelectric volume pulse wave at the fingertip. Clustering analysis is carried out by using independent pulse pressure index P1 and P2 combined with the characteristic parameter K value, which has correlation with the elasticity of the vessel wall. By improving K_means algorithm to select representative waveforms in several waveform cycles, the characteristic parameters are extracted for regression modeling by support vector machine. The results show that the average absolute relative deviation of the detection value is 421%. This results indicating that this method is a simple and effective method for the detection of vessel wall elasticity, which fully meets the actual clinical application requirements.

Abstract:This paper proposes an integrated enzymefree electrochemical system for the detection and analysis of glucose based on the technologies of functional modification, MEMS and wireless communication. The nickel hydroxide (Ni (OH)2) modified screen printed electrodes were used as the sensing part, and an integrated circuit was designed for capturing, conversing and amplifying the weak signals. This circuit is composed of a microcontroller, a signal generator circuit, a constant potential circuit, an I/V conversion circuit, an amplifier circuit and an AD conversion circuit. All of these were integrated into a module with the size of a common U disk. The data was transmitted wirelessly via bluetooth and the analysis results were displayed on the upper device (PC or smart phone). The experimental results of glucose detection show that the sensitivity for glucose determination is about 1112 μA/mmol/L, and the detection limit is 459 μmol/L. Through systematic experimental evaluation, it is proved that the proposed system has a good repeatability and antiinterference ability, and the electrode activity can last for more than 48 hours. The system is expected to play a greater role in home prevention and clinical diagnosis of diabetes due to its small size, low cost and portability.

Abstract:A quantitative assessment method of spasticity based on the fusion of acceleration and angle signals is proposed and corresponding device is developed to solve the difficult problem of accurate quantitative assessment of spasticity in stroke patients. By analyzing the limb acceleration and joint angle signal data, the joint angle (θp) and joint mobility (θrom) corresponding to the acceleration mutation point are extracted, and the θp/θrom value is calculated for the evaluation of spasticity status. The modified Ashworth scale (MAS) and the device were used in clinical experiments on 27 subjects, and different spasm detection methods were used for different muscle groups. The correlation analysis between MAS level and θp/θrom value shows that θp/θrom value can be used to assess spasticity state, and the device has good validity, the correlation meets r=-0869, P<005. Based on MAS and θp/θrom values, the MAS grade indicating spasticity state was modified again. The results showed that there was significant difference between the modified MAS grades (P<0001), which could realize quantitative evaluation of spasticity state.

Abstract:To solve the problem of realtime arrhythmia diagnosis, this paper designs and works out a realtime arrhythmia diagnosis system and proposes a realtime arrhythmia diagnosis algorithm based on feed forward neural networks(FFNN) and onedimensional convolution neural network (1DCNN). The system uses a wearable ECG (electrocardiogram) acquisition device to collect ECG signals, and then wirelessly transmits the ECG to the client software in real time for arrhythmia diagnosis, and finally automatically uploads the diagnosis results to the server. The algorithm takes raw ECG signals as input and segments with a 200ms time window, then uses a classifier based on the FFNN model to detect the position of the R peak in real time, the algorithm extracts the ECG sequence between three R peak and resamples it to a length of 360 as the ECG_RRR feature, finally uses a classifier based on the 1DCNN model for realtime arrhythmia detection. This paper uses the MITBIH database to train the algorithm model and test the system. The results show that the realtime arrhythmia diagnosis system and algorithm proposed in this paper have the characteristics of high accuracy, strong realtime performance and easy deployment. The recall rate, precision rate, and the overall accuracy of the system’s interpatient R peak position predictions are 980%, 995%, and 976% respectively. The overall accuracy for 5class interpatient arrhythmia classification is 915%.

Abstract:The electrocardiogram signal (ECG) intuitively reflects the physiologically electrical activities of heart, and has important reference value in diagnosing heart diseases. In this paper, we proposed a kind of twoclass classification method for ECG signals using convolutional neural networks. The network convolution layer used different convolution kernels to maximize the use of local features for classification and detection of abnormal heart beats. The method has utilized the MITBIH Arrhythmia Database proposed by Massachusetts Institute of Technology. Calculating performance metrics through confusion matrix and applying crossvalidation against three traditional machine learning methods, experiments show that the model accuracy rate can even reach 9686%, which increases 339%, compared with the support vector machine dichotomy method with the highest accuracy performance. This method simplified the feature extraction process and fully improved the accuracy of abnormal heartbeat detection.

Abstract:In order to achieve rapid gait state judgment and analysis to better perform highprecision gait recognition and control of the exoskeleton of the lower limbs, the algorithm research based on wearable inertial measurement device to detect human body posture change is carried out. Through the measurement experiment of nonperiodical gait changes such as falling, turning, squatting and standing up of the lower limbs of the human body, data of the subjects’ body angle, lower limb joint angular velocity and acceleration changes during the experiment were obtained, and then random forests were applied. Four classic classification algorithms for machine learning have performed a comparative analysis of activity recognition on subjects. The results show that compared with other algorithms, the decision tree supervised learning algorithm can quickly and accurately detect and judge a variety of nonperiodic changes in the human body in the active state, the previous recognition accuracy can reach more than 99%. Research can provide a theoretical basis for the development and application of wearable smart equipment.

Abstract:In this paper, a direct frequency domain approximation method of wavelet filter was proposed based on the beetle antennae search algorithm (BAS), aiming at the problems of less methods to construct wavelet filter in frequency domain, no guarantee of filter stability and low approximation accuracy. Firstly, the approximation rational formula was designed according to the basic requirements of the wavelet filter circuit. Then, based on the rational formula, a frequency domain optimization model of wavelet filter approximation was established considering the constraint conditions of filter circuit stability, positive reality and zero initial value. Finally, Gaussian and Marr wavelet were taken as examples, in which the BAS algorithm was used to solve the optimization mathematical model to obtain rational parameters. The approximation results are compared with other frequencydomain methods in time and frequency domain. The simulation and calculation results show that the proposed approach has a good approximation effect under the same conditions. The mean square errors (MSE) of the fifthorder Gaussian wavelet and seventhorder Marr wavelet are only 3446 8×10-4 and 7346 2×10-4 respectively.

Abstract:Aiming to solve the problem that the increase of the size of the radiator is limited by the internal space of the directcurrent (DC) power verification instrument, the global response surface method (GRSM) is used to optimize multiobjectives of the thickness of the substrate and the fins, the number of fins and the airflow. Firstly, a finite element model (FEM) of 1 kW switching power supply is proposed and the orthogonal experiment is designed, with variable factors of substrate thickness H2, fin thickness H3, number of fins N, and fan airflow V, and the objective of the maximum temperature of the power chip and switching power supply. Secondly, based on the experimental results, the multiobjective optimization of the variable factors is conducted by using the GRSM. The results show that the maximal temperature can be decreased to 9161 ℃ and 6273 ℃, respectively, under the conditions of H2=30 mm,H3=10 mm,N=10 and V=2400 CFM. Finally, the experiment is carried out with the optimized parameters, and the experimental results agree well with the optimization ones with marginal errors of -253 ℃ and 370 ℃, respectively. Our scheme presents a significant achievement since the maximum temperature is reduced by nearly 30%, and the obtained optimized parameters are useful.

Abstract:The wear of bearings and gears in aeroengine is the main cause of engine failures. The debris generated by wear contain important information about the wear status of rotating parts, online sensing of wear debris in lubricant is an effective method to diagnose potential failures of rotating components. Firstly, the mechanism and characterization of wear debris are analyzed, the relationship between wear characteristics and wear particle characteristics and the typical characteristics of debris are described. Secondly, several online sensing technologies of wear debris in lubricant are introduced, including optical method, electromagnetic method, acoustic method and energy method. The monitoring mechanism, technical characteristics, research progress and limitations of the online sensing methods are investigated. Finally, the development trends and challenges of online sensing for lubricating wear debris are systematically discussed.

Abstract:Aiming at the problems of few labeled fault samples, unbalanced dataset of pipeline operation state data set and high cost of sample labeling in the process of urban buried drainage pipeline blockage fault detection, an classification and recognition method of drainage pipeline blockage fault based on active learning is proposed. This method adopts the improved committee sample query strategy, and established an active learning model based on consensus entropy to realize the learning of unbalanced data set. After fully considering the uncertainty of the samples and mining the most informative of unlabeled samples for labeling, the committee composed of several random forest classifiers was used to classify and identify the unlabeled samples. The vote entropy, uniform entropy and randomly selected sample query strategy are compared and verified on the pipeline operation data set collected by the laboratory. The experimental results show that the committee query strategy based on consensus entropy has faster convergence speed and better stability under the initial training set of class distribution equilibrium, and also has good recognition effect under the initial training set with unbalanced distribution of categories.

Abstract:Detection of the loose particles is urgently required in the production of sealed electronic equipment. Particle impact noise detection is a national aerospace electronic component loose particles detection method. Aiming at the problem of the large volume of the sealed electronic equipment and the difficulty in determining the position of the loose particle, this paper uses the parameter optimized kNN algorithm to locate the loose particle. After building a positioning experiment system and designing a specimen model, a multichannel loose particle signal is obtained, and the time domain and frequency domain features with excellent performance are extracted as the data set for kNN algorithm learning. The grid search method is used to find the optimal k value selection, distance measurement and weight setting of the kNN algorithm, then the kNN algorithm of parameter optimization is used to establish the plane and space positioning models respectively. The experimental results show that using the kNN algorithm of parameter optimization for loose particle positioning, the accuracy of plane and space positioning reaches 8118% and 7934% respectively, which effectively improves the positioning accuracy under traditional conditions.

Abstract:A 625 MS/s, 12 bit twochannel time interleaved ADC is designed in 40nm CMOS process. The single channel is pipeline ADC with no sampleandholdamplifier (SHA) frontend for lowpower consumption. A wideband and highlinearity foreground input buffer and a high speed and high precision bootstrapped switch are used for ensuring the effective input bandwidth of the interleaved system. A background calibration algorithm based on reference channel is applied for sampling time mismatch calibration between channels. This background calibration method is appropriate for completely random input signals. The core area of the system is 069 mm2. The postsimulation results show that the 625 MS/s, 12 bit time interleaved ADC achieves 67 dB of SFDR and 585 dB of SNDR with the Nyquist sampling at full sampling speed, while its power consumption is 295 mW, which meets the design targets and confirms the effectiveness of the design.

Abstract:As an important security component in the chip, true random number generator (TRNG) plays an increasingly important role in modern encryption systems. For the design of TRNG, the key is to require an entropy extractor to stably generate entropy under severe environmental changes (such as process fluctuations, voltage and temperature (PVT)). Based on the Xilinx FPGA platform, a lowcost, highefficiency true random number generator based on ring oscillator was proposed in this paper. The proposed TRNG improves the efficiency of entropy extraction through fast carry logic on the one hand, and optimizes the circuit structure and delay on the other hand to achieve considerable throughput and randomness with relatively low resource overhead. The TRNG proposed was verified on multiple Xilinx Virtex6 FPGAs and Xilinx Spartan6 FPGAs. Experimental data test results show that the proposed TRNG can exhibit good robustness in a wide range of PVT and generate random bit streams. Random bits only passed the NIST SP80022 statistical test suite with a fairly high P value, but also passed the latest NIST SP80090B test.

Abstract:To solve the problem of low accuracy of tradition transformer fault diagnosis methods, a transformer fault diagnosis method based on the improved sparrow search algorithm was proposed. First, the oppositionbased learning (OBL) is introduced to optimize the selection of the population to improve the global optimization ability of the sparrow search algorithm.Then use the ISSA to dynamically optimize the kernel function parameters and penalty coefficients of the support vector machine, and obtain the fault diagnosis model of the support vector machine optimized by the ISSA based on DGA. The original data is processed through very sparse random projection to remove redundant features. At last input the processed data into ISSASVM for fault diagnosis, and compare it with GWOSVM, PSOSVM and SSASVM. The results show that the fault diagnosis rate of the ISSASVM is 92%, which is 1067%, 8% and 533% higher than that of GWOSVM, PSOSVM and SSASVM. So it can predict the operating status of the transformer more accurately.

Abstract:The optimization design of diagnosis strategy is an important part in the process of testability design. Unreliable test factors seriously affect the optimization design process. This paper summarizes previous research results. Aiming at the problem that heuristic search algorithm is difficult to solve the problem of diagnosis strategy optimization under unreliable testing conditions, this paper proposes a diagnosis strategy optimization algorithm based on the essence ant system. This paper establishes a mathematical model for the optimization of the diagnosis strategy under unreliable conditions, and then constructs the optimized target with the cost of testing and the cost of error. Then, it uses the improved ant system algorithm to solve the problem. Finally, the algorithm is applied to an equipment for instance analysis. Compared with greedy algorithm and common ant colony algorithm, it shows the advantages of the algorithm in precision and convergence speed, and verifies the feasibility and effectiveness of the algorithm.

Abstract:Aiming at the problem that part contour recognition in the recognition of thinwalled parts in industrial production lines is greatly affected by light, the color constancy technology is applied to the contour recognition of industrial production lines. Based on the basic principles of Retinex, HSI and edge detection algorithms, a contour feature recognition algorithm for thinwalled parts under complex lighting conditions is proposed for image restoration and contour recognition of thinwalled parts. First, the method uses HSI color space to extract the brightness of the image. Then the improved Retinex algorithm is used to perform adaptive image enhancement on the acquired part image and the light change information in the image is filtered out. Then the image is grayed out on this basis. Finally, the Canny algorithm is used to identify the edges of thinwalled parts, and the valid outline features of thinwalled parts are further extracted. Experimental results show that the algorithm can quickly and accurately identify the outline information of thinwalled parts under complex lighting conditions, and meet the needs of industrial assembly line inspection.

Abstract:In recent years, more and more attention has been paid to the safety of motorcyclists. Wearing helmets is very important for their own safety. In order to improve the accuracy and robustness of the detection network, in this paper, the mainstream onestep detection network SSD net is introduced with similar visual mechanism module, and the weight of network feature map is reselected in channel and space. The RFB module is also added to the network, which is similar to the human visual eccentricity mechanism. We also use Mosaic method for data enhancement and cosine attenuation learning rate to optimize the network. The experimental results show that the MAP value of the improved network is about 4% higher than that of the original SSD net. And it has better application effect.

Abstract:Imagebased action recognition is still a very challenging task because it is disturbed by the differences in the background information of the images in the class and the similarity of the behavior between the classes. Some action categories are very similar in terms of human poses and facial expressions, so extracting salient features from various parts of the image that are rich in semantic information is essential to improve the accuracy of action recognition. Drawing on the advantages of the bilinear pooling model in finegrained image classification, and to avoid this model which containing a lot of background noise to affect the recognition accuracy, an improved bilinear pooling model is proposed for action recognition in the paper. The model uses channel and spatialwise attention mechanism to focus on the important targets in the image, and generates RoI by integrating multilayer attention mask, which can effectively suppress the background noise information in the image and improve the accuracy of action recognition. Our method achieves the accuracy of 8524% on the Stanford-40 dataset, and the accuracy of 8457% on the custom 60 kind of action dataset.

Abstract:Aiming at the problem of false alarms caused by the low probability of abnormal events in existing anomaly detection methods, a novel video anomaly detection approach is proposed based on the Gaussian process regression framework. By integrating the structural properties of deep learning with the flexibility of kernel methods, a new deep learning technology called deep Gaussian process regression that fully encapsulates CNN structure is introduced to extract features and detect anomaly in one model. The results on the popular Avenue dataset and on a recently introduced realevent video surveillance dataset show that the detection model based on deep Gaussian process regression has achieved 839% framelevel AUC and 344% framelevel AUC on the two dataset, respectively, and has reached the state of the art in performance.

Abstract:Aiming at the deficiency of the measurement accuracy of the wave velocity in the current traveling wave ranging method and the defect of the detection algorithm of the wave head, by analyzing the changing law of traveling wave speed with frequency, the frequency range when the wave speed is stable is derived, so as to eliminate the influence of the wave velocity on the ranging result. On the basis of studying the relationship between the arrival time of the traveling wave of the highvoltage direct current (HVDC) transmission line fault and the specific instantaneous amplitude of the traveling wave, the mechanism of the instantaneous amplitude affecting the traveling wave location is analyzed, and a fault location method of the transmission line based on the instantaneous amplitude is proposed. The iterative method forms a timeamplitude diagram of the highfrequency current fault signal. According to the timeamplitude diagram, the wave head time and the specific instantaneous amplitude at that time are determined to form a HVDC transmission line fault location algorithm. MATLAB/Simulink simulation results show that the relative error of this method is within 03%, and compared with the positioning method of Hilbert Huang transform, this method has higher positioning accuracy and is basically not affected by fault distance and transition resistance.

Abstract:Aiming at the issues of data privacy and security, storage and transmission efficiency in the health smart home system, this article proposed a key parameter encryption method for smart home. The key sequence is obtained by extracting the main peak characteristics of the pulse waveform data and the STFT frequency domain characteristics and splicing them together. The randomness test shows that the 24 s original data can generate a 128bit key sequence with good randomness, and then the three parameters of the original data are divided into blocks Compress data first and then use the key sequence to combine AES symmetric encryption and ECC asymmetric encryption. The experiment compares the three encryption methods CECC, AECC and CAECC of the three parameters. The experimental results show that the CECC time and space overhead are the largest; when the parameter data amount is not greater than 64 KB, the AECC time overhead is about 08 times CAECC, and the space overhead is about 3 times of CAECC, AECC time overhead is minimum, CAECC space overhead is minimum; when the data volume is greater than 64 KB, CAECC time and space overhead is minimum.

Abstract:In view of the fact that it is difficult for the existing frequency corrector to correct the two signals whose frequencies have a small frequency difference and changes dynamically over time, a highprecision dynamic frequency correction technology is proposed, to achieve the measurement and control of the phase difference between two different frequency signals that change over time and achieve a small frequency correction of the order of 10-12. The design mainly uses the ADC as a digital phase detector function to sample a large amount of effective data in the linear region using the edge effect. The FPGA and the microcontroller preprocess and store the data. The VCOCXO corrects the frequency according to the voltagecontrolled voltage provided by the singlechip microcomputer after the data is stored and preprocessed. The final experimental results show that the new frequency corrector has the characteristics of low noise, high resolution, high conversion efficiency, high stability, and high adjustability.

Abstract:In order to study the polarization characteristics of silver, platinum, glassy carbon, gold and copper electrodes and determine the best application scenario of each material electrode, a detailed theoretical analysis, model simulation and flume measurement of surface microelectrode types are carried out, and a polarization impedance evaluation method of surface microelectrode materials is obtained. The results show that the measurement accuracy of platinum and glassy carbon electrode is the best, the performance of silver electrode is slightly inferior to that of platinum and glassy carbon electrode, and the gold electrode can only be used above 50 kHz. In terms of accuracy, sensitivity and reliability, the comprehensive performance of glassy carbon electrodes is the best, followed by silver, platinum, gold and copper electrodes. Platinum and glassy carbon electrodes are more suitable for impedance spectroscopy applications, while silver, platinum and glassy carbon electrodes are very suitable for electrical impedance scanning or tomography. Silver, platinum and glassy carbon electrodes can be selected for lower frequency impedance measurement, while gold, platinum and glassy carbon electrodes can be selected for higher frequency.

Abstract:Aiming at the problem of delay estimation for singlechannel mixed signals, an algorithm based on recursive least squares (RLS) is proposed, which divides the delay estimation into coarse estimation based on cyclic statistics and fine estimation based on RLS, using the timeaveraged criterion of power of two to iteratively update the rough estimate of time delay to complete the entire estimation process. The algorithm has a simple structure, fast convergence speed, and obvious improvement in estimation accuracy, but it is more dependent on the delay difference of the two signals. Simulation experiment results prove the effectiveness of the algorithm.

Abstract:Automatic Xray security inspection is an important method to maintain public safety. Current research of prohibited item inspection on Xray images only works on predefined classes in the dataset and cannot be generalized to unseen categories. The imbalance problem in the dataset will also affect the performance of models. In order to solve above defects, the paper proposes a segmentation model for prohibited item inspection in Xray Images based on fewshot learning. The model first embeds the test image and annotated support images to a common space, then measures the spatial pixelwise similarity and regional similarity, finally segments out suspected areas in the test image. To deal with uncertain numbers of support images, a fusion method based on the ConvGRU is proposed to integrate the similarity information for the test image and different support images. Experiments show that the proposed model improves 20% and 22% meanIoU compared to the stateoftheart methods under 1shot task and 5shot task, which demonstrates the ability to recognize unseen categories.

Abstract:High precision network traffic prediction is the basis of modern network intelligent management. Targeting at the problem of parameter optimization of SVM in the process of network traffic prediction modeling to improve the network traffic prediction results, this paper proposes the network traffic prediction model of SVM optimized by Improved Gray Wolf algorithm. Firstly, collect the historical data of network traffic, and preprocess the data with phase space reconstruction and normalization, then introduce the improved gray wolf algorithm to quickly search the relevant parameters of the global optimal support vector machine, and learn the historical data of network traffic after preprocessing according to the optimal parameters, and establish a prediction model that can mine the history data of network traffic including the law of change after that, the network traffic prediction model of SVM optimized by other algorithms is compared and analyzed. The results show that the prediction accuracy of the improved gray wolf algorithm optimized support vector machine is more than 90%, much higher than the compared model, and the training time of the prediction modeling process is less than the compared model, which can meet the requirements of high accuracy and realtime network traffic management.

Abstract:In order to realize that multiple seismic source vehicles follow in the form of a fleet during the field construction, this paper adopts a pilot driving model to record the state information of the pilot vehicle in real time and send it to subsequent vehicles for automatic followup driving. The autofollowing vehicle is controlled according to the acquired speed and heading angle. This paper firstly proposes a control algorithm that combines PID control and model predictive control (MPC), and designs a controller to determine the execution rules of the control algorithm to ensure that the vehicle Safe distance. Through Carsim/Matlab simulation verification, the control method proposed in this paper is compared with linear quadratic control (LQR). The results show that the control method proposed in this paper can meet the following driving conditions in acceleration, uniform speed, and deceleration in the longitudinal control, and can follow the corners such as 180 degrees and 90 degrees in the lateral control, and maintains good performance with the pilot vehicle. safe distance.

Abstract:A narrowband imaging method for space target with compound micromotion is proposed. Due to the advantages of narrowband radar in target detection and tracking, narrowband radar is widely used in space target detection. For space target with micromotion, there is timevarying Doppler modulation induced by micromotion, which contains the important structural information of the target. By applying the inverse Radon transform (IRT) method on the timefrequency image, the position of scattering centers of the target can be obtained and the narrowband imaging can be achieved. Narrowband imaging reduces the requirement of radar bandwidth and has advantages in space target detection. However, in the real detection scene, the target motion is composed of micromotion and translation, which makes the narrowband imaging method invalid. In this paper, based on the radar echo model of space target with compound micromotion, the timevarying Doppler modulation characteristics of the target are analyzed, and a narrowband imaging method for compound micromotion is proposed. Firstly, the micromotion period is estimated based on the timefrequency correlation coefficient. Then the translational influence is removed by the Doppler cancellation method, the target translational parameters are estimated, and then the translational compensation is achieved. Finally, the narrowband imaging is achieved based on the IRT. The proposed method is not affected by the target translation, and can effectively achieve the narrowband imaging for space target with compound micromotion.