Abstract:The renewable energy sources, such as fuel cells, photovoltaic (PV) panels, and battery pack, produce a low and variable dc output voltage, and their power generation efficiency and lifespan are affected by the output current ripple. Therefore, a high-gain converter with continuous input current is required as an interface to satisfy the requirements of front-end current ripple rate and rear-end voltage level. To address the common issue of insufficient boosting ability in the existing non-isolated high-gain converter, by combining the quadratic boosting, switched-inductor and charge pump technologies, a new triple-inductor single-switch high-gain boost converter is proposed in this paper. A 200W/50kHz experimental prototype is designed and manufactured to verify its feasibility. The experimental results show that the proposed converter can achieve the ultra-high voltage gain of 2/(1-D)2, which is much higher than the existing schemes. Moreover, it has the advantages of continuous input current, common ground, small amount of devices and ease of control.

Abstract:To solve the issue of complicated structure and low real-time performance of the two-step multiple object tracking by detection algorithm, a real-time multiple object tracking algorithm founded on modified YOLOv4-Tiny and DeepSORT algorithm is proposed. The depthwise separable convolution is employed in YOLOv4-Tiny, to reduce the calculation of the model; The detection branches are increased to 3, and multi-scale feature fusion structure is built to decrease the missed ratio of tiny objects; The modified GC attention module is used to extract the global context features of the model. In the tracking part, the pedestrian motion model of DeepSORT is optimized and the appearance model is reconstructed, the detection and tracking algorithms are combined and experimented in MOT16 test sequences finally. The results show that the total parameters of the improved algorithm are 4.2M, 52% less than the original algorithm and 5.2% more MOTA, faster processing speed under GPU, and the tracking speed of an average of 11 frames per second can be achieved under a single CPU, which can fulfill the requests of precision and speed for multiple object tracking mission in low-calculation devices.

Abstract:Microstrip antenna is widely used in telemetry system due to its small size, easy conformal and suitable for metal environment. Aiming at the narrow bandwidth of traditional microstrip telemetry antenna, a design to expand bandwidth by slotting on antenna patch and floor is proposed, simulation and test verification are carried out and their results are consistent with each other. The HFSS simulation results show that the relative impedance bandwidth(S11<-10dB) of the antenna is 5.6%(2.409~2.545GHz) which basically covers the ISM frequency band by slotting at appropriate positions and the highest gain can reach 4 dBi. Test results show that its relative impedance bandwidth is 6%(2.41~2.56GHz) and the transmission coefficient of S21 is about -22 dB when the distance between two antennas is kept at 10 mm, which meets the requirements of the telemetry system.

Abstract:Real time fluorescence quantitative polymerase chain reaction is an important means of rapid detection of New Coronavirus COVID-19.After sampling a series of complex chemical reactions, the virus results were obtained through the RT-qPCR temperature control system. The RT-qPCR temperature controller is actually an accurate temperature controller and spectral detection electronic equipment. This paper adopts the design architecture includedARM processor,H-bridge driverandPeltier, meanwhile realizes the rapid rise and fall of temperature through normalized PID algorithm.So as to build a low-cost and miniaturized temperature control system. The system can select nuclear acid detection suitable for other viruses according tothe different temperature settings and fluorescent agents, and the temperature control can achieve the heating rate of not less than 15℃/s The cooling rate shall not be less than 8℃/s, the temperature accuracy shall be within0.1℃, the overshoot of temperature rise and fall shall be less than 4℃, and the stability time shall be 3s.

Abstract:Deformable mirror (DM) is the most important wavefront corrector in adaptive optics. It can compensate wavefront aberration by changing the reflective mirror’s surface frequently. However, the commercial piezoelectric deformable mirror can’t obtain an ideal flat surface under the initial voltage. The optical aberration caused by this initial distortion will seriously weaken the performance of closed-loop control of deformable mirror. Therefore, it is necessary to flatten the surface of deformable mirror before using it for optical aberration correction. In the paper, a deformable mirror control system based on embedded platform is designed to realize surface flattening and mirror deformation. The embedded platform is Jetson nano development board of NVIDIA company. In order to properly control the optical figure of the deformable mirror we have to obtain an response matrix which is the response of optical surface to the DM actuator’s stroke. Here, the multithreading algorithm is used to improve the computing speed of the response matrix. Because of the hysteresis effect of compressed electrical materials, the flatteing of deformable mirror needs to be carried out iteratively. After several close-loop iterations, the surface error RMS of the deformable mirror is reduced from 0.92 λ to 0.03 λ, The deformable mirror has a good flattening effect. The system is not only conducive to the commercialization of the deformable mirror system, but also greatly improves the closed-loop control ability of the deformable mirror and eliminates the initial aberration of the deformable mirror itself in the adaptive optical system.

Abstract:The variation of operating conditions will lead to the variation of circuit health characterization parameters, so it is impossible to judge whether the health characterization parameters are caused by the degradation of circuit performance or the variation of operating conditions. Aiming at this key problem, aviation inverter is taken as the research object. Firstly, a multi-evaluation index optimization model was used to select the relevant sensitive health characterization parameters. Then, based on extreme learning machine, the mapping model of health representation parameters under the condition of working condition and no fault was established. Finally, based on the relative changes between the current health representation parameters and the health representation parameters output by the mapping model, the circuit health indicators considering working conditions were constructed to realize the health assessment of the aviation inverter under different working conditions. The experimental results show that the proposed method can effectively reduce the influence of working conditions on health indicators, and the MAE and RMSE of the proposed method are 64.4% and 66.8% lower than those of the method directly based on Euclide distance.

Abstract:According to the situation that a small high-speed wind tunnel at a low level of informatization and the flow field of the wind tunnel needs to be manually adjusted to achieve stability.Based on the flow field control process, this paper designs a LabWindows/CVI-based upper computer control system and applies it in a small high-speed wind tunnel. The system communication uses the OPC communication protocol based on DataSocket, The upper computer can efficiently and reliably realize the information interaction with PLC by receiving the data of Ni OPC server;the upper computer software uses multi-threading, asynchronous timer and other technologies to realize the functions of real-time data display, historical data record and storage, and friendly human-machine interface which is convenient for operators to use.Through a lot of test and analysis, the newly designed flow field control system has fast response, stable operation and high accuracy, which meets the design requirements of high-speed wind tunnel control system. It has certain reference value for reducing energy consumption, improving information technology and automation level of wind tunnel.

Abstract:Borehole ground penetrating radar (BGPR) can effectively solve the contradiction between detection depth and accuracy of traditional ground penetrating radar. In order to realize orientated detection and accurate location of geologic anomalous body with BGPR, the data acquisition and control system of BGPR for ultra-deep orientated detection is designed in this paper. Firstly, the orientated detection functional module including control unit, control algorithm and mechanical structure is researched. The closed-loop control unit is composed with gyroscope, one-chip computer and stepper motor. The control algorithm is PID, which can determine the motor control parameters according to the deviation between the real-time angle acquired from gyroscope and the setting detection angle. Then the motor drives the “double tubes” mechanical structure to finally realize orientated detection. Secondly, the radar transmit-receive pulse control circuit, radar reflected wave signal acquisition circuit and communication circuit are designed, whose main control chip is STM32F407. Finally, the key signals of the above system such as the transmit and receive trigger pulses are tested and analyzed, and the field model and coal mine working face detection are carried out using the orientated detection BGPR, which is equipped with the above-designed system. The signal test results and detection results show that the novel system can fully meet the functional requirements of orientated detection BGPR, and the maximum deviation between the measured and theoretical values of key signal evaluation indexes is only 9.8%. In addition, the orientated detection BGPR can accurately detect the positional information such as the depth and orientation of the target, and can also accurately locate the interface between coal seam and rock stratum.

Abstract:Aiming at the problem of the low accuracy of grid supply load forecasting in high permeability areas of small hydropower, a grid supply load forecasting model based on artificial intelligence and residual correction is proposed to predict and correct the periodic and random components contained in the grid supply load. Ensemble empirical modal decomposition (EEMD) is used to decompose and extract the components of different frequency bands in the network load, and a multi-level gated recurrent unit (GRU) network model based on modal components is constructed, the accuracy of the prediction results on the test set is improved by increasing the complexity of the network model. In addition, the Fischer value is introduced to characterize the cumulative effect of rainfall on the output of small hydropower, and the fisher information-weighted Markov Chain (FI-WMC) residual correction step is added in the output of prediction results, reduce the deviation of prediction result caused by the uncertainty of small hydropower output. The results of simulation verification show that the multi-level EEMD-GRU-FIWMC model can be better applied to the grid load forecasting in areas with high permeability of small hydropower. In areas where the penetration rate of small hydropower is above 20%, compared with the traditional GRU model and the no-residual correction model, its prediction accuracy is increased by 7.61% and 3.85%, respectively.

Abstract:At present, several technical schemes of vector atomic magnetometer have been reported at home and abroad, and a vector atomic magnetometer based on magnetic field rotation-modulation method can realize the continuous measurement of vector magnetic field. In this paper, the working principle of the vector atomic magnetometer based on the magnetic field rotation-modulation method is described, and the design, generation and calibration method of the "rotating magnetic field" are introduced. Then, the superposition magnetic field of the rotating magnetic field and the vector magnetic field is experimentally measured by a pump-probe atomic magnetometer. The coincidences of the average value and peak-to-peak value of the measured magnetic field with the theoretical calculation results are verified, and the origins of the deviation between the average values of the measured magnetic field and the theoretical results are analyzed. At last, the continuous measurement ability of vector atomic magnetometer is verified by experiments. The research content of this paper lays a technical foundation for the development of continuous measurement vector atomic magnetometer.

Abstract:Aiming at the problem that the capacity estimation method of lithium-ion battery for vehicle is not high precision, a method of residual capacity estimation of lithium-ion battery based on BP neural network optimized by genetic algorithm is proposed in this paper. First, after collating NASA's lithium-ion battery data set, the peak value of battery capacity increment curve under different health conditions was obtained. Secondly, the health factor was analyzed by principal component analysis to reduce its dimension, and the connection weight of BP neural network was optimized by genetic algorithm to predict the capacity of lithium ion battery. Finally, the model was validated on different NASA batteries. The results show that the proposed method can accurately estimate the capacity of four kinds of lithium ion batteries under different training amounts, and the square mean error of the estimation is less than 2%, and the proposed method has higher prediction accuracy than the prediction results without genetic algorithm optimization.

Abstract:Aiming at the problem that the existing harmonic current detection algorithms are difficult to accurately detect subsynchronous harmonic current and supersynchronous harmonic current, an improved ip-iq harmonic current detection algorithm is proposed. Firstly, it is clear that the low pass filter gentle attenuation characteristics and insufficient dynamic adjustment ability are the fundamental reasons for the low harmonic current detection accuracy. Secondly, the low pass filter is replaced by a self-tuning filter with steep attenuation characteristics near the fundamental frequency. The self-tuning filter can enhance the detection effect of subsynchronous harmonic current and supersynchronous harmonic current without reducing the filtering effect of the integer harmonic current; The dynamic characteristics of ip-iq algorithm are ameliorated through the adaptive filtering algorithm based on the improved hyperbolic sinusoidal function. The adaptive filtering algorithm establishes the functional relationship between the step size parameter and the harmonic compensation current, and realizes the dynamic adjustment of the step size parameter. This algorithm overcomes the adverse influence of the time-varying quasi steady state characteristics of subsynchronous harmonic current and supersynchronous harmonic current on harmonic detection. In this paper, the detection effect of the improved algorithm is evaluated by current distortion rate and harmonic content rate. Simulation experiments verify the effectiveness of the improved algorithm. Compared with the traditional ip-iq harmonic current detection method, this method can improve the detection accuracy of harmonic current by 23.59%, and realize the accurate detection of harmonic current in the environment of subsynchronous harmonic and supersynchronous harmonic.

Abstract:The collaborative filtering recommendation algorithm framework based on graph convolutional neural network is the most advanced recommendation algorithm framework at present. The framework does not pay attention to the timing of interaction occurrence in the feature learning of user-item interaction embedding vector, but in actual situations, users- Item interaction generally has obvious timing characteristics and is an important factor affecting recommendation performance. Based on this, a graph convolution collaborative filtering recommendation algorithm based on time series features is proposed, which redo multiple data sets, retain the original information of the data sets, especially the time series features, and summarize the historical time series information of user-item interaction in the data set. It is parameterized and put as an important feature input to the high-order cooperative signal transmission of graph convolutional network model training. Set up multiple sets of ablative experiments on three publicly available official datasets-Gowalla, Yelp and Amazon-book, and use recognized evaluation indicators-ndcg and recall to evaluate the performance of the recommendation algorithm. The experimental results show that under the same parameter Settings, the figure convolution collaborative filtering recommendation algorithm based on temporal characteristics performance beyond the existing same type figure convolution, collaborative filtering recommendation algorithm to verify the timing characteristics are recommended to improve effect of the positive role, improve the efficiency of model training and prediction, shooting more efficiently solve the problem of network information overload, to satisfy the higher application requirements.

Abstract:The large-scale access of distributed power sources in the distribution network aggravates the power fluctuation of the power grid. In order to stabilize the power of the power grid and absorb the power generated by new energy, it is necessary to reasonably allocate energy storage in the power grid. In this paper, the energy storage configuration is divided into the planning layer and the operation layer based on the two-layer programming model, and a hybrid algorithm for energy storage optimization configuration combining the global harmony search algorithm based on trend movement and the second-order cone relaxation theory is proposed. The planning layer uses the minimum sum of the annual investment cost of energy storage and the annual operating cost of the distribution network as the objective function, and studies the economic benefits of energy storage configuration location, power, capacity, etc.; The minimum sum is the objective function, and the optimal charging and discharging power of energy storage in each period is studied. Based on the actual data of a certain place, a 33-node distribution network is taken as an example to conduct a simulation analysis, and the configuration methods of energy storage in different scenarios are studied and the excellence of the hybrid algorithm in this paper is discussed. The results show that the hybrid algorithm proposed in this paper can successfully solve the energy storage configuration problem in this scenario, which proves the effectiveness of the hybrid algorithm in this paper. At the same time, the energy storage configuration method in multiple scenarios is discussed, and it is proved that the use of the two-tier planning model to configure the energy storage is beneficial to reduce the comprehensive cost of the distribution network. Finally, the hybrid algorithm in this paper is compared with the traditional intelligent optimization algorithm to solve such problems, which proves the speed and accuracy of the hybrid algorithm in this paper in solving such problems.

Abstract:A speed-up method of spring test on automatic measure for time series complexity is proposed, including the spring graph in 5-threshold compressed space and the construction creep (CC) rate. For its slow speed, the specific method uses peak and valley values fitting to simplify the spring diagram and to speed up the self-similarity measuring . We select wavelet mapping chaos, unified equation and Rossler hyperchaos equation as test equations. To take the bifurcation diagram as the basic comparison rule of geometric characteristics and to adopt the spectral entropy complexity (SEC) based on FFT as the gold standard. The results of spring self-similarity measure CC before and after the improvement are compared; The calculation time and stability of each test method is given.The results show that the improved method can distinguish between period and chaos and elevate the resolution ratio of hyperchaos and chaos significantly with 2 to 20 times hoisting in terms of time. This work contributes an improved spring test method for the signal complexity measure featured with automatic parameter selection for distinguishing periodic, chaotic and random states. The essence is to extract the asymmetric nonlinear statistical invariants of three-dimensional phase diagram (the spring diagram).

Abstract:The fixed pointing multi-beam antenna consists of the main reflector fixed and the feeder system which can move along the inclined orbit to match different satellites, and the main beam point is unique. In contrast to full motion or limited motion antennas, the main reflector surface has high precision surface accuracy, ensure the main beam pointing accuracy and realize the precise positioning of the feed system.This paper proposes a variety of spatial measurement technology including static GPS measurement system, total station measurement system,laser tracker measurement system and digital industrial photographic measurement system.During the installation of a multi-beam antenna, the main reflection surface accuracy is better than ±0.15mm, the main beam pointing accuracy is better than ± 0.0122°, and the feeder system positioning accuracy is better than ±1.5mm, which meets the requirements of the structural installation index. By testing the electrical performance index meets the requirements through the multi-beam antenna system, the characteristics of the feasibility, reliability, high efficiency and high efficiency of the scheme are fully verified.

Abstract:To address the problem that the low contrast and visibility of medical CT images are not conducive to human eye observation and post-processing, this paper proposes a contrast enhancement algorithm for medical CT images based on a multi-scale exposure fusion framework to achieve enhancement of medical CT images. Firstly, the original image is decomposed and reconstructed by Laplace pyramid to reduce the interference of image noise and enhance the image details. The image is then enhanced by calculating the weight estimation matrix, the exposure rate and the brightness conversion function of the reconstructed image using the exposure combination algorithm, which enhances the image contrast and improves the image visibility at the same time. Experiments show that the method is significantly better than other traditional image enhancement algorithms and has a significant enhancement effect on medical CT images.

Abstract:At this stage of industrial production line, many types of circuit board soldering can not be automated instruments. In order to reduce the loss of manpower and material resources due to rework in factories for the phenomenon of missing solder in manual soldering of SMD components, automatic detection of soldering of SMD components by machine vision technology is proposed. Using the improved ResNet-FPN structure, the shallow feature information is fused with multi-scale channels, thus increasing the richness of feature information of tiny and occluded targets, reducing the training parameters, and speeding up the forward speed of network training; the number of classification samples is balanced and the loss value is reduced by introducing the Focal loss (FL). The experimental results show that the improved Cascade RCNN algorithm trains slightly faster than the original model, with a small increase in recall and an average mean accuracy (mAP) of 90.9%, which is 2.2 percentage points higher than the original model, and achieves better detection results.

Abstract:At present, fiber optic sensing technology is popularly used in the monitoring fields of bridges, tunnels and dams and has achieved good results, while the application of FBG sensing technology based on weak reflection grating is still in the initial stage. In view of the limitations of BORDR and FBG technologies in geological disaster monitoring, this paper introduces the principle of weak-reflection fiber grating sensor array monitoring technology and designs a kind of external fixed point weak-reflection grating sensing optical cable based on the characteristics of weak reflecting grating sensor array with large capacity and high precision. And then constructs a real-time landslide monitoring system by combining with weak-reflection grating sensor array monitoring technology and Internet of Things technology. This paper also carries out a calibration experiment of weak-reflection grating sensing optical cable and finally presents a field application experiment using the real-time landslide monitoring system. The monitoring system measured that the deformation of 4 consecutive weak reflection gratings from 158#-161# reached 61.91mm, which indicates that the system can accurately locate the location of the landslide displacement and monitor the size of the landslide displacement, realizing the high-density automatic quasi-distributed monitoring of the landslide geological hazard body.

Abstract:The traditional digital instrument recognition method has a large amount of computational amount, not enough real-time, low accuracy. This paper studies a meter identification method is studied in combination with deep learning and image processing. In order to reduce the amount of computation, YOLOv4 target detection network is used and GhostNet is adopted as YOLOv4 basic network. At the same time, the Depthwise Separable Convolution and Ghost module can be introduced in YOLOv4 to reduce the amount of parameters, and the H-Swish activation function is applied to increase the accuracy. In order to highlight color information in the image binarization process, a digital binarized method is studied based on color model multi-threshold segmentation. The main color of RGB image is enhanced, and then converts to an HSI image, and then the pixel point satisfying the condition will be reserved by multi-threshold processing, thereby obtaining a binarized image. Digital information can be better extracted in comparison with traditional image pretreatment algorithms. Experimental results show that the proposed method of reaches 87.98 mAP on the test data set, and detection speed is increased to 37.2 FPS, and then the effect is significant in digital instrument positioning and digital inspection.

Abstract:In industrial production and traffic engineering, safety helmets and reflective clothing are important safety protection. Aiming at the problem that the traditional helmet reflective clothing recognition method can only detect single color reflective clothing and low detection efficiency, we proposed a helmet reflective clothing detection method based on improved YOLOX-S network model. The simplified BiFPN module is used to replace the original enhanced feature extraction network to improve the feature extraction ability of the network for targets with different scales. The mosaic method is used for model training to improve the detection ability of the network in complex scenes. The GIoU loss function is used to further improve the recognition accuracy of the model. Experiments on the expanded helmet reflector data set show that the proposed algorithm can achieve 83.74% map while maintaining a high detection speed. Compared with the original YOLOX-S, the detection AP of wearing helmets, reflective clothing and pedestrians is improved by 1% ~ 3%, and there is no dependence on the color of reflective clothing, which effectively realizes the rapid and accurate detection of helmets and reflective clothing.

Abstract:The traditional image restoration problem mainly adopts the divide-and-conquer method, image restoration problem is divided into different sub-problems, and the optimal solution is obtained by processing the sub-problems. Due to the connection and loss between different processing links, the optimal solution of the sub-problem cannot be the global optimal solution. To solve this problem, an end-to-end dual control network is proposed, which uses a control module to control the degenerate branch and the processing branch through parameters. The network uses a special Encode-Decode structure to deal with the feature problem under the fixed scale factor subnet, the loop skip connection structure is used to eliminate the stacking blocks of the convolutional layer and enhance the feature display at the output end. Experiments show that the peak signal-to-noise ratio (PSNR) value of the image restored by the proposed method and the comparison method is above 30, and the structural similarity index measure (SSIM) is above 0.90, which effectively improves the visual effect.

Abstract:In order to ensure the pointing accuracy of radar, shipborne radar was usually fixed on the vessel by screw connection. This connection mode will make the radar suffer severe impact in the underwater explosion impact environment, which was seriously affects the safety of radar equipment. For this reason, an array buffer system composed of cut-off excitation buffer was proposed, which has the characteristics of rigidity under low impact and flexibility under strong impact. The contradiction between keeping accuracy under low impact and keeping strength under strong impact was effectively alleviated. Then take the array buffer system as the research object, the mechanical properties of the newly developed buffer system were systematically analyzed through some important tests such as the stiffness test of elastic elements, buffer dynamic characteristic test, impact test and bump test of the array buffer system, so as to provide reference for the subsequent experimental research of the newly developed buffer system.

Abstract:The hunting motion generated by high-speed trains can seriously affect the safety of trains, so predicting hunting motion can provide early warning. The current research on hunting motion is mainly about the prediction of hunting instability, but there is a small-amplitude hunting intermediate state between normal and hunting instability during train operation, and the prediction of the small-amplitude hunting state can provide early warning of hunting instability. To this end, a prediction method for imbalanced data based on a one-dimensional convolutional neural network 1D-CNN and a conditional generative adversarial network CGAN is proposed for the extreme imbalanced case of high-speed train hunting motion data using the bogie lateral acceleration signal as the standard. The adversarial learning mechanism of CGAN method first utilised to optimise the update parameters through a game between the generator and the discriminator. The well-trained CGAN model is then used to generate samples, feed the enhanced data into a 1D-CNN classifier, and output the prediction results. Experiments are conducted on actual high-speed train operation data, and the results show that CGAN can fit the data distribution of high-speed train hunting fault motion and enhance the dataset, and the prediction accuracy based on the proposed method is 97.5%, which is substantially better than the comparison method. Thus the CGAN-1DCNN-based minor hunting prediction method can predict minor hunting under data imbalance and achieve early warning of hunting instability.

Abstract:The application of dynamic photogrammetry to the dynamic monitoring of wind turbine blades is of great significance to optimize the design of wind power system and ensure the safe and reliable operation of the system. Automatic matching of image feature points is a key technology in dynamic photogrammetry system. The quality of matching method directly affects the final measurement accuracy. In the actual dynamic photogrammetry of wind turbine blades, the shimmy of the blades during operation will aggravate the non-planarity of the blades, which makes it more difficult to directly use the epipolar or homography matching technology to match the binocular image features. Therefore, a binocular image matching method based on pre-screening and local homography is proposed, which can be used in the measurement scene of wind turbine blades characterized by a smooth continuous surface. Firstly, the point images of the blade are obtained through the binocular photogrammetry system. For the point image captured by the left camera, the candidate matching point set of each point in the left image is filtered from the right point image through the epipolar constraint. Secondly, a local approximate plane region centered on this point is extracted for each point in the current point to be matched and its candidate matching points and all local approximate plane regions are combined accordingly. Then, for each group of approximate plane regions, the improved random sample consensus algorithm is used to estimate the corresponding optimal homography matrix model, Finally, the final point that matches the current point to be matched is determined in multiple candidate points through the local homography transformation and reasonable error threshold. This method has been successfully applied to the shimmy experiment of the wind turbine blade. The image point matching rate is not less than 96% and the matching accuracy is not less than 97%. The method can satisfy the requirements of actual wind turbine blade image matching for accuracy.

Abstract:In order to improve the accuracy of the detection and positioning of the number of ladle carrier tank based on computer vision, reduce the detection error in the case of contamination, reduce the missing detection problem caused by the small area of the number of tank and improve the detection speed, a detection and recognition method of the number of ladle carrier tank based on improved YOLOv5 network was proposed. The feature extraction capability of the model was enhanced by adding attention mechanism into the feature extraction network. By upgrading the backbone network to lightweight GhostBottleNeck, the reasoning speed of the model is accelerated. By performing Affine Transformation on the target character, the distorted character is converted into a near-positive perspective, and then the improved ResNet network is used for single-character recognition. The results show that the accuracy of the improved network is 90.3%, the recall rate is 87.3%, and the final number identification accuracy is 97.7%, indicating that the method can effectively achieve the accurate location and identification of the number of the ladle carrier tank, and provide reliable data support for intelligent management.

Abstract:In order to solve the problem of maximum pooling loss of information and average pooling fuzzy features, improve the time-frequency image recognition efficiency of the model and reduce the model complexity, A CNN network model using a deep detachable small Convolutional kernel for down-sampling and CBAM is proposed for fault diagnosis of bearings. Firstly, in the pooling layer except the last layer, the depth separable small convolution layer is used to replace the pooling layer to realize the down-sampling function of the pooling layer. Secondly, CBAM is introduced in the last pooling layer to pay more attention to the fault features represented by time-frequency images to improve the computational efficiency of the model. Thirdly, global average pooling is used instead of traditional full connection layer to further reduce the number of model parameters. Finally, CWRU bearing vibration data and self-made experimental platform data were used to verify the validity and feasibility of the proposed method in rolling bearing fault diagnosis. Experimental results show that the fusion depth separable small convolution kernel and CBAM improved CNN model can effectively reduce the training parameters and computation required by the model, and achieve better performance in recognition accuracy.