Abstract:Most methods for predicting the Remaining useful life of deep learning of rotating units usually assume that the data distribution of training data and test data is the same, resulting in low prediction accuracy of the model under different working conditions. For the above problems, this paper proposed a model transfer method based on adversarial training, where the transfer object is a rotating multi-unit RUL prediction model. Aiming at the transfer scenario where the source domain and target domain have different working conditions and the target domain lacks label samples, a domain classifier was introduced to extract the common features of the source domain and target domain data. The feature extraction network in the RUL prediction model was retrained by combining the labeled data in the source domain with the unlabeled data in the target domain. In the training process, auto association and correspondence constraints were added to improve the ability to extract common features, thus realizing the migration application of the model in different scenarios. The test results of the transfer model using the XJTU-SY public dataset revealed that the prediction accuracy of the method described in this paper is higher than that of the original prediction model under the new working conditions. Compared with other transfer methods, the prediction error of this method is smaller, and it has a better effect on predicting the remaining useful life of rotating units under variable working conditions.

Abstract:The abnormal sound in the trains running can be used as information source to reflect the status of the vehicle equipment. For the reason that, a recognition model based on 1D-CNN was proposed to identify the abnormal sound of trains, and a set of recognition system for abnormal sound of trains was designed. Firstly, the experimental sample library of audio data was constructed. Then MFCC was used to extract the characteristic information of abnormal sound data samples. Aiming at the complex mapping relationship between train noise features and vehicle state types, a 1D-MCNN based on MFCC input was constructed to identify and classify the fault information contained in abnormal sound. Finally, the model parameters such as MFCC order, learning rate and batch size are determined by experiments and optimization. The t-SNE algorithm and confusion matrix were used to analyze the model feature extraction ability. The results show that the method is effective for the identification and diagnosis of abnormal sound of trains and its accuracy rate reaches 98.38%.

Abstract:A constant transconductance input and output rail-to-rail fully differential amplifier with continuous time common-mode feedback is proposed. The input stage complementary difference pair uses cross conduction to keep the total input transconductance constant in the whole common-mode input range. The intermediate stage adopts folded common source and common gate structure to achieve high gain and full swing. At the same time, A continuous-time common-mode feedback circuit with A class A output structure is designed to enable FDA to work in large swing and high impedance systems. Based on SMIC 0.18 μm process, the designed FDA was simulated and verified and the layout was drawn. When the supply voltage was 3.3 V and the load capacitance was 5 pF, the DC gain of the circuit was 92.2 dB, the unit gain bandwidth was 5.55 MHz, the input-output rail-to-rail range was close to 100%, and the input-stage transconductance change rate was only 4.53%. The setup time is 218 and 195.7 ns, respectively, and the corresponding swing rate is 15 and 16.7 V/μs, respectively.

Abstract:Petrochemical equipment always coated with a coating composed of thermal insulation layer and metal protective layer for thermal insulation. The application of pulsed eddy current testing technology can realize the wall thickness detection without removing the coating. However, in actual testing, the material and thickness of the metal protective layer are inconsistent, which will affect the testing results and testing errors. In this paper, a pulsed eddy current experiment platform was built to analyze the impact of different types and thicknesses of metal protective layers on the pulsed eddy current testing signal under different insulation layer thicknesses. The results show that when the metal protective layers are aluminum and stainless steel plates, the increase in their thickness or insulation layer thickness will increase the dispersion of characteristic values, but will not affect the trend of characteristic curves. The thickness of the tested part can be calculated through the characteristic curves; when the metal protective layer is galvanized steel plate, the shielding effect and induced eddy current generated by its high magnetic permeability characteristics will affect the detection signal and characteristic curve. With the increase of the thickness of the galvanized steel plate, the difference of the late attenuation of the detection signal of the specimens with different thicknesses is smaller. When the thickness of the insulation layer increases to 50 mm, the detection signals are basically coincident, and the thickness of the tested piece cannot be calculated through the characteristic curve.

Abstract:The path planning of UAV area coverage in unknown environment is studied based on deep reinforcement learning method. By building a grid environment model, randomly deploying UAV and no-fly zone in the environment, and using a double deep Q-network(DDQN) to train the coverage strategy of UAV, a set of UAV coverage path planning framework base on DDQN is obtained. The simulation experiment shows that the designed UAV unknown area coverage path planning framework can achieve full coverage in the environment without no fly zone, and can also better complete the area coverage task in the environment with an unknown number of no fly zones. Compared with DQN method, its average coverage rate can be 2% higher under the same training conditions and training rounds, higher than Q Learning method and Sarsa method in the environment without no fly zone.

Abstract:Aiming at the memory bandwidth bottleneck of FPGA hardware accelerator of convolutional neural network algorithm, this paper proposes a Secondary Cache-Row Recombination (SC-RR) based on secondary cache. By analyzing the performance of SDRAM memory, FPGA hardware acceleration principle and memory bandwidth bottleneck, a secondary cache mechanism is established. This mechanism can serve the stacked access requests during the acceleration process, reducing the additional overhead of Active and Precharge operations by merging access requests from the same Bank/Row. The experimental test results show that under the SC-RR scheduling strategy, the memory access time is reduced by 32.87%, the power consumption is reduced by 31.71%, and the effective bandwidth utilization is increased to 91.3%. In the case of similar performance, hardware resource consumption is reduced by 83.8%, which meets the design requirements.

Abstract:The detection of concrete interface position is the key to control the overfilling height of the pile. The traditional method of determining the concrete interface position by manual experience has a large error and cannot effectively control the filling height. It is difficult to discriminate the concrete interface during pouring, which leads to the problem of underfilling or overfilling of concrete. The first method of overfilling detection based on magnetic detection technology is proposed. Firstly, the magnetic target is designed by using the density difference of different media in the borehole, and then by measuring and processing the three-component data of magnetic field, the concrete interface position is detected to control the concrete filling height. This paper designs a set of overfill detection system based on magnetic detection technology, completes the selection of measurement points through indoor tests, and designs the magnetic target specific gravity as 2.1 g/cm3, and the alarm threshold is calibrated as 333 mG. Through construction site tests, it is verified that the system achieves rapid detection of concrete interface to the pile elevation and overfill ceiling position, and accurately completes two levels of warning to achieve automatic detection, which solves the problem of It solves the problem of concrete underfilling or overfilling at the construction site, and effectively reduces the waste of concrete.

Abstract:For the problem of poor initial filtering effect of normalized least mean square (NLMS) algorithm and poor robustness of non-local means (NLM) filtering, this paper proposed an improved model based on variational modal decomposition (VMD)-Hausdorff distance non-local means (HDNLM) filtering. For the power line interferenceand white Gaussian noise in lower extremity the EMG signal, VMD was used to decompose the noisy signal, HDNLM was used to filter the decomposed signal, and the filtered output signal was superimposed, finally, the performance of the algorithm was evaluated by signal-to-noise ratio (SNR) and improved root mean square error (IRMSE).The experimental results show that the NLM and its improved NLM (INLM) are better filtered on average compared to VMD-HDNLM and NLMS when the noise amplitude is 0.1~0.2 M in 16 muscle EMG signals,but when the EMG noise amplitude was 0.3~0.5 M, the IRMSE values of VMD-HDNLM increased by 0.64%, 1.84%, 3.11% and 13.95%, 12.77, 11.07% and 1.05%, 1.74%, 2.85% on average relative to NLM, NLMS and INLM.At the same time, the VMD-HDNLM algorithm has a wider range of parameters than the NLM and INLM algorithms to obtain a smaller value of IRMSE, its robustness is better, and the probability of obtaining a better value in actual situations is greater.

Abstract:In the non-electrolytic capacitors drive system, the film capacitors with small volume and long life are often used to store the input energy at the grid side. However, this topology easily leads to the problem that the grid-side input power and the inverter side output power are prone to coupling. In order to upgrade current quality of the grid, this paper proposes an inverter output power control strategy with phase compensation based on second-order generalized integral phase-locked loop and proportional integral resonant controller to depress the grid-net set harmonic content and increase the power factor. First of all, by analyzing the topology of the drive system, the high-power factor condition without electrolytic capacitor is clarified. Secondly, Secondly, the second order generalized integral phase-locked loop is used to obtain the phase and amplitude information of the grid side voltage, and the Kirchhoff current law is used to calculate the phase compensation angle of the inverter output power; And then, the inverter output power control loop based on the proportional integral resonance controller is established to adjust the output power of the inverter close to the ideal value. At last, the effectiveness of the control strategy is verified by the comparative experiments.

Abstract:The traditional transmission tower assembly monitoring system has the idle state of data transmission with invalid energy consumption, which makes its life time difficult to meet the engineering needs. Therefore, based on the idea of turning the sensors in the idle time period into low-power sleep state, this paper proposes an M/M/1/N/∞queuing model suitable for data transmission of transmission tower assembly monitoring system, deduces the distribution equation of data transmission in the monitoring system, and determines the corresponding rotation times of data transmission by solving the ratio of mon-itoring data waiting length to effective input rate, The idle time of data transmission of each sensor is determined, and the low-power working mode of turning the sensor in the idle time period into sleep is realized. According to this method, the developed transmission tower assembly monitoring system consumes 24% of power in the 6-day actual monitoring process of Zhouning Ningde 500 kV Line Project, and the endurance time is increased by 1.9 times compared with the traditional method, meeting the engineering demand that the endurance time of the monitoring system is greater than the construction cycle of single base tower assembly.

Abstract:Aiming at the problems of low accuracy and low detection efficiency of manual quality detection of sandpaper surface defects in the current industrial production process, an automatic detection method of sandpaper surface defects based on the YOLOv5 network model and CA attention mechanism is proposed. Firstly, the surface images of sandpaper in the process of sandpaper production are sampled, and the collected surface defect images are divided into four defect types, namely, sand removal, sand piling, scratch, and fold, to make the surface defect datasets of sandpaper. Secondly, the C3 module in the YOLOv5 backbone network is improved to the CAC3 module by combining it with the CA attention mechanism. Finally, the network models before and after the improvement are trained and verified on the self-built sandpaper surface defect datasets. The experimental results show that the values of P, R, mAP@0.5, mAP@05:0.95, and S of the improved YOLOv5+CAC3 network model are 96.2%, 92.9%, 95.8%, 65.0%, and 16.8 ms, which are 1.1%, 2.2%, 0.6%, 1.7% and 4.5 ms higher than the YOLOv5 network model before improvement. This method has high precision, fast speed, and stable detection in the detection of surface defects of sandpaper, which meets the requirements of the detection of surface defects of sandpaper in the production process.

Abstract:Aiming at the problem of feature extraction and fault diagnosis of rolling bearing vibration signals, a fault diagnosis method of rolling bearing based on wavelet packet information entropy and support vector machine (SVM) optimized by snake optimization algorithm (SO) is proposed. The collected vibration signals are processed by using the wavelet packet, the energy spectrum entropy and the coefficient entropy of the wavelet packet are constructed, and the constructed feature vectors are input into the SO-SVM for identification and classification; Finally, the multi-fault pattern recognition is realized and the diagnosis results are output. The simulation results show that the diagnostic accuracy of this method for five different groups of samples reaches 99.17%~100%, and compared with FOA-SVM and PSO-SVM, it has a higher effect of fault recognition and classification.

Abstract:Due to its special double salient structure and serious nonlinear problems, Switched reluctance motor has a large torque ripple, which severely limits its application range. The traditional torque sharing function is affected by torque characteristics and voltage limitation, and there are still large torque ripples during the commutation period. In different speed segments, affected by current change rate and commutation period, the torque ripple will be more obvious when the commutation overlap angle is fixed. Based on this phenomenon, this paper proposes a control strategy in which the commutation overlap angle changes with the speed in real-time. By detecting the angle corresponding to the actual torque crossing zero during the operation of the motor, the corresponding commutation overlap angle is calculated, and a look-up table between torque load-motor speed-commutation overlap Angle is established. According to this reference table, the most suitable commutation overlap angle can be look-up with the change of motor speed under different loads to minimize the torque ripple of switched reluctance motor. Finally, in order to verify the effectiveness and feasibility of the strategy, a 3 kW, 3-phase 12/8-pole switched reluctance motor is used as the control object for simulation and experimental verification, which proved that the proposed method could effectively suppress the torque ripple of the switched reluctance motor in a wide speed range.

Abstract:Recently, the adversarial training framework of maximizing and minimizing the discrepancy between bi-classifier has been proved effective in unsupervised domain adaptation (UDA). Classical UDA approaches usually choose to use some simple intra-class discrepancies to measure the difference between the bi-classifier, such as L-1 norm and Kullback-Leibler divergence. From a geometric point of view, this work designs a novel European dual difference by considering the distribution of dual classifiers in the European Space and combining the defects of the classical dual classifier algorithm, and combines it into this adversarial UDA framework. This novel discrepancy can effectively distinguish the two probabilities predicted by the bi-classifier whether they are close in determinacy or in uncertainty. In addition, we also provide theoretical support to prove the upper bound of the theoretical error of the metric. Experiments on the public UDA dataset show that the average accuracy of the European-style dual adversarial algorithm in the small-scale dataset Digits, the medium-scale dataset Office-31, and the large-scale dataset VisDA are 98.3%, 87.8%, and 81.7%, which outperforms other 2-classifier UDA methods with intra-class variance and achieves results comparable to state-of-the-art methods.

Abstract:With the concept of energy conservation and environmental protection and the concept of sustainable green development deeply rooted in the hearts of people, how to do a good job in energy conservation and emission reduction of blast furnaces has become one of the main problems facing the steel industry at present. In order to achieve energy conservation and emission reduction of blast furnace, a multi-objective optimization scheme with the lowest fuel ratio and the highest coal ratio is proposed by combining artificial intelligence technology with blast furnace production data. In terms of fuel ratio and coal ratio prediction, random forest (RF), long short memory network structure (LSTM), and long short memory combined with attention mechanism (LSTM Attention) are used for comparative analysis, and LSTM Attention model, which is the most accurate for fuel ratio and coal ratio prediction, is selected as the prediction model. On the basis of LSTM Attention prediction model, combined with multi-objective particle swarm optimization (MOPSO) and non dominated sorting genetic algorithm (NSGA-II), the pareto optimal solution is found and compared, and MOPSO with better effect is selected for result analysis. The results show that under a certain production condition of the blast furnace, the energy consumption can be reduced by 4.06% by controlling the parameter values of the decision variables such as pressure difference, oxygen content, coal injection volume and air volume×1011 kJ/year, reducing CO2 emissions by 25.91 t/year, providing technical support for energy conservation and emission reduction of blast furnace.

Abstract:Aiming at the problem that it is difficult to accurately estimate the dynamic state of the generator in the electromechanical transient process, a dynamic state estimation method based on continuous-discrete cubature information filtering is proposed in this paper. First, a continuous-discrete state estimation model that can accurately describe the actual operating dynamics of the generator is established, then the stochastic differential equation (SDE) is converted into a stochastic difference equation by using 1.5th-order Taylor expansion, and the state prediction value is accurately calculated according to the third-order spherical radius cubature rule, and finally the predicted state is corrected by the measurements to obtain an accurate state estimate. The simulation results of the four-machine two-zone generator system show that, compared with the traditional generator state estimation methods, the proposed method in this paper not only has higher estimation accuracy, stronger robustness and acceptable computational overhead, but also has the flexibility to be easily extended to distributed power system state estimation.

Abstract:The non-contact heart rate estimation method based on FMCW radar has the advantages of comfort and convenience. Due to the interference of noise and respiratory harmonics, the accuracy of current methods is still limited. To solve this problem, this paper proposes a heart rate estimation method. On the basis of phase analysis, the phase difference operation is performed on the phase signal, which can enhance the heartbeat signal. Then, an improved wavelet threshold denoising method is used for denoising to eliminate noise interference. During frequency estimation, this paper processes the heartbeat signal with the Hamming window function, and then uses the multiple signal classification (MUSIC) algorithm for frequency estimation, which may solve the spectral leakage problem caused by signal interception, and improve the resolution and stability of MUSIC. Thus, the influence of respiratory harmonics and noise on heart rate estimation may be reduced, and accurate heart rate values may be obtained. Simulation results show that the accuracy of heart rate estimation is improved.

Abstract:Printed Circuit board is an indispensable part of electronic products, and its market demand is increasing day by day. Therefore, it is of great significance to manufacture PCB without defects. In the PCB defect detection, the defect targets to be detected are small and most of the detection targets are easily confused with the background, so the improved algorithm introduces the Coordinate Attention mechanism into the backbone network of the native YOLOv5 algorithm. A Transformer Encoder was introduced into the neck network and a high-resolution detection head suitable for small targets was added. The Intersection over Union algorithm of selected anchor frames was changed to a more advanced E-IoU. Compared with the original YOLOv5 algorithm, the performance of the improved algorithm is significantly improved according to the results of Precision, recall and mean Average Precision of the algorithm evaluation index, and the mean Average Precision is 98.46%. It can meet the precision requirement of PCB defect detection in industrial field.

Abstract:In order to improve the efficiency of automobile fault diagnosis and reduce the cost of diagnosis, this paper proposes a fault diagnosis method of automobile system based on Dynamic Fault Tree (DFT). Firstly, the failure mode model of the system is the DFT is established by using DFT, and the DFT is transformed into a sequential binary decision diagram (SBDD) to calculate the minimum cut set (MCS). Secondly, the diagnostic importance (DIF) of the MCS and the components in the MCS is calculated. On this basis, a diagnosis method for automobile based on the Unified Diagnostic Service (UDS) protocol is proposed. The MCS corresponding to the diagnostic trouble codes are filtered and sorted, and MCS and components with large DIF are prioritized for diagnosis. Finally, the effectiveness of the proposed method was verified by the case study of an Active Rear Steering system. The case analysis results show that, the proposed method can calculate the fault diagnosis sequence and can effectively guide the vehicle fault diagnosis work.

Abstract:Under the background of "carbon peak and carbon neutral", the future energy development will be transformed from single energy system to integrated energy system. With the development and application of electric vehicle network interaction technology and the significant improvement of flexible load ratio, the dispatchable resources in integrated energy system will become more abundant. In this context, the integrated energy system including electric vehicle battery swapping station and multi-type flexible load is taken as the research object. Firstly, the integrated energy system model including wind power, energy storage, combined heat and power unit, gas boiler and other equipment is constructed.Secondly, the charging and discharging characteristics of battery swapping station are taken into consideration in the integrated energy system optimization scheduling based on the flexible load responses of electricity, heat and gas. Taking the minimum operating cost of the system as the objective function, the integrated energy system optimization scheduling model considering flexible load and orderly charge and discharge of battery swapping station was established. Finally, the problem is modeled by Yalmip and solved by Cplex. The results show that the system operation cost is reduced by 13.04%, the wind power consumption rate is increased by 8.65%, and the load peak valley difference is reduced by 24.58%. The effectiveness of the proposed model in reducing the system operation cost, peak shaving and valley filling and wind power consumption is verified.

Abstract:Self-driving vehicles are an important part of intelligent transportation and a trend of future transportation. Improving the reliability of autonomous driving technology requires extensive testing of autonomous driving vehicles. However, conducting real-world road tests is costly and risky. It is especially important to build models to generate diverse and realistic traffic scenarios for testing autonomous driving techniques. A traffic scenario generation generative adversarial network model called TSG-GAN is proposed for the generation of traffic scenarios. The TSG-GAN model uses Generative Adversarial Networks to rapidly generate realistic and diverse traffic scenarios by using rich traffic scenario data (e.g., lane geometry, crosswalks, traffic signals, surrounding vehicles, etc.). With reasonable driving intentions of vehicles, the TSG-GAN model can precisely generate realistic traffic scenarios that are not observed in practice. The effectiveness of the proposed model is verified by testing on a publicly available dataset.

Abstract:In the process of gas discharge tube production, the uniformity of electronic powder coating on the electrode surface is the key to the quality of gas discharge tube products, it is mainly detected by human eyes now. Aiming at the problems of low efficiency and poor accuracy of manual detection, an uneven electronic powder coating detection algorithm based on improved YOLOv5 is proposed. Firstly, the collected images of electron powder coating on the electrode surface are made into data sets, and data enhancement was performed. Secondly, the STDC module is used to optimize the backbone feature extraction network, to improve the detection accuracy of uneven surface defects of hard-to-recognize metal electrodes, and two feature layers are generated for adapting to the dataset size. Finally, Kmeans++ clustering is used to optimize the computation of adaptive anchor boxes. The experimental results show that the mAP@50 of the improved YOLOv5 algorithm proposed reaches 99.22%, which is 6.84% higher than that of the original YOLOv5 network, greatly improving the detection accuracy, and is more efficient than manual detection.

Abstract:Aiming at the problems of low efficiency and poor accuracy of glue mark detection caused by poor lighting conditions, a multi-threshold segmentation method of airport runway glue marks based on improved sparrow search algorithm was proposed. Firstly, the lens imaging reverse learning is used to improve the diversity of the initialized population, and then the optimized performance level and adaptive factor are introduced to improve the individual quality and search ability of the discoverer. Secondly, the firefly algorithm is introduced to assist the traditional sparrow search algorithm to jump out of the local optimum. Finally, use the improved sparrow algorithm to optimize the Tsallis relative entropy metric function to achieve automatic and accurate segmentation of glue traces. The experimental results show that the detection accuracy of this method is much higher than that of the traditional algorithm, its FSIM values are all greater than 0.8, and the SSIM values are close to 1, and it shows a good segmentation effect in the case of poor lighting conditions and the mixture of pavement, marker lines and glue marks.

Abstract:An automatic tagging robot needs to be provided the 3D coordinates of a bar’s bottom center to weld a tag on the bundle. A method based on binocular vision was proposed to select and localize bars’ bottom centers. Virtual image camera model was adopted in the binocular vision system. The extrinsic parameters of two cameras were calibrated from a planar calibration pattern which was put parallel to the common end plane of bars. The virtual images of two cameras were created according to the calibrated results. A method using SVM and connected region was adopt to extract the center point features of bars in both virtual images from two cameras. A group of candidate features pairs were selected using epipolar constraint to the features and coplanar constraint to the recovered physical points. The 3D coordinates of corresponding physical points were recommended to the robot to try to weld the tag. Simulation results showed all recommended points to the robot were from correct matched pairs. It demonstrated effectiveness of the features matching method presented. In real experiment, the maximum depth displacement error of the recommended points was 0.20 mm, the average error was 0.09 mm. It demonstrated the effectiveness of bar bottom center extraction method presented.

Abstract:The flame image of the sintering machine tail section contains many features related to the sintering endpoint. It is feasible and practical to make full use of the feature information of the sintering flame image to judge the state of the sintering endpoint online. Aiming at the problems of difficult extraction of flame image feature information of sintering machine tail section, low recognition accuracy, and difficulty meeting realtime requirements, an improved MobileNetV3 sintering section flame image recognition algorithm is proposed. MobileNetV3 is taken as the basic model for feature extraction of flame state at the sintering endpoint, and the attention mechanism is introduced; it improves the attention structure of the SE channel to solve the problem of weak resolution of features extracted from the original model; The introduction of Spatial Attention (SA) mechanism and the design of Two Branch Channel Spatial Attention (TBCSA) module accurately capture the position and content information of the red fire zone in the flame image of the sintering section; The data enhancement and cosine annealing learning rate are introduced to improve the generalization ability of the model, and the freezing training strategy is used to accelerate the model convergence. The experiment on the sintering flame data set shows that the algorithm can fully use the feature information in the sintering flame image, and the recognition accuracy reaches 97.54%, which is 6.41 percentage points higher than before.

Abstract:To solve the problem of poor classification effect of convolution neural network on tomato leaf type diseases, a tomato disease identification method based on MobileViT lightweight network was proposed in this paper. Firstly, feature fusion of input and global representation is deleted and local and global representation are fused to make local representation more closely related to global representation. Secondly, in order to avoid the large increase of parameters and FLOPS when the model is scaled, the 3×3 convolution layer is replaced by 1×1 convolution layer in the fusion block. Then, the residual structure of input and fusion blocks is added to optimize the deeper level in the network model. Finally, the model accuracy is further improved by replacing the ReLU6 activation function with the H-Swish activation function. The experimental results showed that the improved MobileViT model can well recognize tomato diseases, with an average recognition accuracy of 99.16%. Compared with other convolution neural network models, it has higher recognition accuracy.