Abstract:Brain-computer interface (BCI) is a system that converts brain activity information directly into artificial output, allowing users to directly control external devices through thinking activities. Electroencephalogram ( EEG) technology could obtain real-time neurophysiological electrical signals generated by brain activity. EEG, which has the advantages of non-invasiveness, low cost, and high time resolution, is one of the mainstream methods for BCI to obtain brain activity information. The EEG-based BCI system (EEG-BCI), which provides functions of acquiring signal, processing signal and outputting results, has the ability to evoke characteristic EEG and control external devices. And it has great application value in rehabilitation, diagnosis and neuroscience research. With the everincreasing application demands of EEG-BCI, the technologies that can ensure it rapid and efficient deployment and application are increasingly important. According to the research and application of the EEG-BCI in recent years, this article reviewed the currently technologies of the hardware and software platforms for building EEG-BCI, summarized current status, and evaluated future trends, to promote the development of EEG-BCI.

Abstract:In order to improve the effect of muscle fatigue detection, a dual-sensor fusion method is proposed to make up for the shortcoming that information is easily lost in single-sensor mode. The method realizes a new dual-sensor fatigue detection mode by integrating the time-frequency domain features of the surface EMG signal with the muscle thickness feature of the A-type ultrasound signal in multiple dimensions. Using support vector machine and neural network multi-model training, the detection accuracy of surface EMG and A-type ultrasonic dual-sensor fusion in three fatigue states can reach 85%. Compared with using only the time-frequency domain features of surface EMG signals ( 76. 99%) and the muscle thickness of A-mode ultrasound ( 74. 87%) for fatigue detection, the accuracy is increased by 8% ~ 13%. For fatigue detection, the results show that the dual-sensing fusion mode of surface EMG signal and ultrasonic signal is more accurate and effective than the single-sensing mode.

Abstract:Aiming at the problem of fast and accurate detection of bad information under complex environment background, a humanmachine collaboration system for bad information detection based on rapid serial visual presentation (RSVP) is proposed. Firstly, using the fast-wearing portable acquisition system collected the EEG data of 12 subjects; then the Mallat algorithm was used to extract the lower-dimensional time-frequency features of the EEG data, and EEG signal classification uses artificial neural network (ANN) and support vector machine (SVM). Finally, different times of superimposed average data are introduced in the training set to improve the classification performance of the model. The experimental results show that at least 2 targets are correctly output on average in 60 images containing 3 targets, and the AUC value reaches 0. 9. The system has good performance in the detection of small batch data sets and bad image information with complex environmental changes, and has improved efficiency compared with manual detection.

Abstract:Transcranial alternating current stimulation (tACS) is a widely used noninvasive brain stimulation method. However, due to nonlinear electrical stimulation artifacts interference, it is difficult to obtain the real neural activity during stimulation directly. Therefore, an adaptive variational mode decomposition (AVMD) method is proposed to remove the nonlinear tACS artifacts. In this method, the envelope of artifacts is extracted by Hilbert transform (HT), then, the VMD modes is obtained by WFT spectrum analysis. VMD is used to decompose the recorded data to obtain multiple intrinsic mode signals. According to the amplitude characteristics, the artifact components are selected, and the effective EEG components are recovered. AVMD algorithm were tested on the synthetic data and the public experimental data. The correlation coefficient between reconstructed EEG and real EEG was used to measure the artifact removal effect for the synthetic data. The mean absolute error (MAE) of the statistical characteristics between recovered EEG and sham EEG was used to evaluate the artifact removal effect for the experimental data. For the synthetic data, under the conditions of amplitude modulation depth ma∈ [0. 001, 0. 01], phase modulation depth mp∈ [0. 001, 0. 01] and stimulation frequency f arti∈ [10, 100],the average correlation coefficients between reconstructed EEG and real EEG are 0. 988 5, 0. 893 5, 0. 948 4, respectively. The MAE of the statistical characteristics between recovered EEG and sham EEG are 0. 989 6 ( kurtosis), 2. 991 8( root mean square amplitude), 0. 175 1 (sample entropy) for the experimental data with the stimulation frequency 11 Hz, and are 0. 940 7 (kurtosis), 2. 473 1 (root mean square amplitude) and 0. 084 1 (sample entropy) for the experimental data with the stimulation frequency 62 Hz. AVMD method shows more stable and better nonlinear tACS artifact removal performance compared with superposition of moving averages ( SMA), adaptive filtering (AF) and empirical mode decomposition (EMD). This method provides support for the development of closed-loop tACS instrument.

Abstract:Cell imaging and detection are of great significance in the field of biomedical research and clinical diagnosis, while label-free and high-throughput detections are particularly challenging. On the basis of dynamic scattering theory, this study built a dynamic scattering imaging system, proposed a spectral decomposition-based dynamic signal extraction algorithm, and achieved label-free and high-throughput cell classification by combining machine learning algorithms. Blood cells, EG7-OVA tumor cells and A549 lung cancer tumor cells are used to verify the current method. Experimental results show 98% accuracy for binary classification of blood cells and tumor cells, and 91% accuracy for the three-type classification of blood cells, EG7-OVA and A549. In summary, the proposed method provides high-throughput, label-free cell detection and classification, and is potential for clinical application.

Abstract:In order to solve the shortcomings of low accuracy and low efficiency of artificial arch index measurement. It is proposed to use the dynamic information of the plantar support surface to measure the arch index. Obtain the trajectory of the ground reaction force on the heel, the outside of the sole, and the inside of the sole from the developed distributed force plate. The area SΔ ABC formed by the three points is used to express the support surface information. The mean value MSΔ ABC of SΔ ABC for five 30-second bipedal standing posture tasks was obtained as a key parameter of the foot arch height index (FAI). The combined caliper arch height index measurement system (AHIMS) was used to classify the foot arch types of 30 subjects, and the FAI value was analyzed by variance. The results showed that the FAI value gradually increased from the low arch group, the normal foot group to the high arch group. There was a significant difference between the low arch group and the high arch group (P<0. 001), and between the normal foot and the high arch group (P< 0. 01). The results show that the dynamic information SΔ ABC of the plantar support surface can express the arch height index of the foot.

Abstract:The design of electrical impedance tomography (EIT) electrode array is one of the key factors affecting the performance and imaging effect of the system. At present, the electrode array is optimized under the premise of regular shape field and equal spacing distribution which is not suitable for irregular lung boundaries. In this paper, an optimization method of electrode array based on deep learning network is proposed for lung EIT. The optimization goal of the network is electrode position. The relative error of the reconstructed image, the image correlation coefficient, the distribution uniformity and the condition number of the Hessian matrix for the sensitive field are used as the network inputs. The positions of the electrodes are taken as the network output. The optimization model is constructed based on DNN network. The experimental results show that, for end-expiration and end-inspiration states, the ICC, SSIM and PSNR of images reconstructed based on measured data obtained from optimized electrode increased by 33. 17% and 33. 86%, 14. 5% and 14. 39%, 26. 3% and 28. 27%, respectively, compared with the traditional electrode array with equal-distance distribution. Therefore, it can be concluded that optimizing electrode positions for lung EIT using deep learning is more suitable than traditional methods.

Abstract:In gait recognition, the discriminative gait feature cannot be extracted due to the occlusion of clothing and backpack, which leads to the low recognition accuracy. A gait recognition method combining ResNet and multi-level block structure is proposed in this paper. First of all, the gait energy map is divided into different scales in the horizontal direction to extract the fine-grained features of different regions, which reduce the impact of local occlusion on other regions. At the same time, in order to better learn the characteristics of the region with the highest motion frequency, the Inception module is added. Secondly, in order to improve the recognition accuracy of the network model, cross-entropy loss, triple loss and L2 regularization are utilized to constrain the weight of the residual network. Finally, experiments were processed in the public gait data set CASIA-B and OU-ISIR Treadmill B, and the recognition rate reached 87. 5% and 82. 6% under different clothing or backpack conditions. It is indicated that under these conditions, the method could obtain favorable veracity and good robustness.

Abstract:As the aging of the population continues to deepen, the number of elderly people living alone is also increasing. In the application of solving the problem of elderly living alone, indoor positioning is the most basic and critical issue. Aiming at the needs of indoor positioning, a binocular visual odometer positioning algorithm is proposed, which has been studied from four aspects: Camera imaging model, feature extraction, feature matching and motion estimation. First, the binocular camera is used as the sensor for image acquisition; then the matching relationship between adjacent images is completed by extracting ORB feature points, and the matching relationship between the corresponding feature points of the left and right images is obtained according to the stereo matching algorithm based on the BRIEF descriptor, the camera movement is estimated. The hardware and software platforms are designed to experiment with the proposed method. Experiments have proved that the indoor positioning technology based on visual odometer can accurately locate the position of the elderly indoors, and can safely monitor the elderly in real time.

Abstract:Aiming at the problem that when applying deep learning for gas turbine fault diagnosis, the fault signal data is difficult to obtain, resulting in many normal operation samples and few fault samples, which affect the accuracy of fault diagnosis. A method for augmenting gas turbine fault samples using deep convolutional generative adversarial learning is proposed. According to the characteristics of the gas turbine vibration signal, the fault signal is preprocessed by using fast Fourier transform, empirical mode decomposition and demodulation, and the fault frequency domain features are extracted and the eigenvalue index is selected, and the vibration signal is converted into a two-dimensional gray image. The orthogonal gradient penalty algorithm is used to train the deep convolutional generative adversarial fault sample generation model. The example results show that the test accuracy rate of CWRU bearing dataset obtained is 98. 01%, and the test accuracy rate of a certain type of gas turbine’s fault samples generated by the proposed method is 97. 43%, which are better than other mainstream fault sample generation methods under the same conditions. The effectiveness and superiority of the proposed fault sample generation method are verified.

Abstract:In view of the inaccurate and incomplete fault data in the accumulated telemetry data of space power system, it is difficult for the ground long-time management system to comprehensively select and evaluate the effectiveness of fault detection model according to actual fault data. This paper focuses on the research on the optimal selection method of twin data-driven fault detection model for space power system. Based on the full analysis of the composition, working principle and input-output relationship of the power system, the digital twin model of each component unit of the spacecraft power system is constructed by Simulink. Combined with the analysis of fault mechanism, typical faults are injected into the twin model to enrich the types and quantity of fault data, and the effectiveness of various fault detection models is evaluated based on the twin data. Experiments show that the twin data generated based on this framework are more than 90%, which is similar to the real telemetry data, and six typical failure modes can be injected, where the step-type and gradient-type fault detection ability of the fault detection model can be effectively evaluated. The research of this method can effectively serve the actual ground long-time management system and provide an important model and data basis for the selection of effective fault detection model.

Abstract:To solve the problem of image quality degradation caused by specular highlights, a specular highlight removal method based on spatial constrained clustering analysis is proposed in this paper. Firstly, after projecting the component image into the minimummaximum chromaticity space, the fixed clustering center is introduced to realize the separation of chromatic pixels and achromatic pixels while ensuring the similar chromaticity in a cluster. Then, the intensity ratio adjustment and brightness histogram statistics are used to determine the specular reflection components in the clustering of chromatic pixels and achromatic pixels respectively. Finally, combined with dichromatic reflection model, specular highlight removal is realized. Experimental results show that the entropy value and structure similarity of the image are 5. 750 and 0. 998 8 after highlight removal by proposed method. The proposed method can effectively remove specular highlights in chromatic and achromatic regions and obtain high quality images.

Abstract:Motorized spindle is an important functional part of CNC machine tool, and its advantages and disadvantages directly affect the quality of parts. A support vector machine regression model ( SVR) optimized by chaos genetic algorithm (CGA) is used for spindle fault diagnosis. The principle of the method is to use principal component analysis ( PCA) to reduce the dimensionality of the timefrequency characteristic vector of the vibration signal of electric spindle wear fault, and input the dimensionality reduced characteristic vector into the SVR model optimized by CGA parameters for training and testing. The results show that the accuracy of training and testing is 99. 272% and 95. 249% respectively, which can diagnose the wear fault of motorized spindle accurately.

Abstract:Aiming at the problem that it is difficult to realize effective fault diagnosis for weak signals of rolling bearings in the complex environment of strong noise and variable working conditions, a bearing fault diagnosis method based on improved convolutional Lenet-5 neural network is proposed. Firstly, the collected one-dimensional time-domain bearing vibration signals are preprocessed and converted into two-dimensional grayscale images which are convenient for convolution operation. Secondly, the continuous one-way traditional convolutional layers in the most basic Lenet-5 model are improved into Block1 module, Block2 module and Block3 module to extract more concrete and accurate feature information. Finally, L2 regularization and Dropout optimization are used to avoid overfitting. In order to verify the robustness and generalization performance of the proposed method in complex working conditions, experimental validation was carried out using the rolling bearing dataset and the gearbox experimental dataset. The experimental results of the bearing dataset show that the average accuracy of the proposed method in the variable noise experiments is 99. 3%. In the variable load experiments, the average accuracy of fault diagnosis is higher than 90. 26%. In the variable operating conditions experiments, the average accuracy of fault diagnosis is higher than 89. 01%. In the gearbox dataset experiments, the fault diagnosis accuracy of anti-noise is up to 96. 3%. The improved Lenet-5 method has the better ability of fault diagnosis for 12 fault types of rolling bearings, and has better anti-interference and generalization performance under variable working conditions.

Abstract:To overcome the shortcoming of existing contact resistance threshold detection algorithms that rely too much on prior knowledge and result in high false negative rate, a relay contact system fault detection algorithm based on interval evidential reasoning (IER) was presented. First, the prior threshold and contact resistance test data are converted into interval belief structures. Then, the prior threshold and test data are fused by IER based on the modified weight. Finally, the fused detection threshold is calculated. This proposed method considers the uncertainty contained in the contact resistance measurement model and the consistency of weights in the fusion process of IER, avoiding the shortage of adaptive ability of existing algorithms. Experimental verification was carried out and the results show that compared with the empirical threshold algorithm and the conventional IER method, the proposed algorithm has better adaptive ability and can effectively improve the fault detection accuracy of electromagnetic relay contact system. Considering the data imbalance problem, the proposed algorithm can further improve the performance of fault detection after resampling the original data.

Abstract:Aiming at the inherent constraints between tracking accuracy and disturbance suppression performance of H∞ repetitive control system for uninterruptible power supplies (UPS) inverter, an H∞ repetitive control strategy with equivalent input disturbance (EID) compensation is proposed, which realizes the rapid suppression of load sudden change and other disturbances and the high-precision tracking of reference voltage. Firstly, the dynamic mathematical model of UPS inverter is established by using the state space average method, and the overall control block diagram is given. Based on the H∞ repetitive controller, a set of state feedback gains that can maintain stability when the load changes suddenly are obtained. Secondly, the EID estimator is constructed by a state observer, and the specific design steps are given. The experimental results show that the proposed control strategy can shorten the recovery time of UPS inverter after suffering from strong nonlinear interference by about one cycle, reduce the maximum offset of output voltage RMS to 0. 09%, and improve the dynamic performance of UPS inverter repetitive control system for disturbance suppression.

Abstract:In this paper, a compact triple-notch ultra-wideband multiple-input multiple ( UWB-MIMO) antenna is proposed, which combines a semicircle and a regular hexagon as a radiating patch, and a grounding plate is introduced into the belt. “Comb” slot-like Tshaped branches to achieve higher isolation, the antenna size is 36 mm×18 mm×1. 6 mm. WiMAX (3. 3~ 3. 6 GHz) and WLAN parts are realized by etching the inverted “ Ω”-shaped groove on the radiation patch, etching the “ U”-shaped groove on the feeder and introducing the “U”-shaped branch next to the patch. Frequency band (5. 725~ 5. 825 GHz) and X-band downlink frequency (7. 25~ 7. 75 GHz) are notched in three frequency bands. Both simulation and actual measurement results show that the working bandwidth of the UWB-MIMO antenna is 1. 9 ~ 10. 6 GHz, the relative bandwidth reaches 139%, and the three notch frequency bands are 2. 9 ~ 3. 7 GHz, 5. 6~ 6. 0 GHz and 7. 05~ 7. 76 GHz, the degree of isolation is greater than 20 dB, and the envelope correlation coefficient ECC is less than 0. 003, indicating that the antenna has good characteristics in all aspects and can meet the requirements of UWB-MIMO antennas.

Abstract:With the improvement of the manufacturing process, static power consumption has gradually become one of the main power consumption sources of routers. Aiming at the high power consumption of multi-core systems based on the on-chip network architecture, this paper proposes a power gating strategy based on packet classification, and modifies the router architecture. In this scheme, first classify data packets, and then use bypass to perform different processing on classified data packets to bypass the dormant router, thereby reducing the extra data packet delay and power consumption caused by the application of power gating. Bypass can ensure the transmission of data packets when the router is sleeping, increase the sleep time of the sleeping router, and reduce static power consumption. The classification of data packets makes the sleep and wake-up of the router more reasonable and efficient. The simulation results show that the solution in this paper can significantly reduce the static power consumption of the router and the delay of the network. Compared with the traditional power gating technology, it can reduce the static power consumption by 72. 4% and the average packet delay by 16. 8%; compared with the classic power gate, the controlled bypass router reduces the static power consumption by 12. 4% and the average packet delay by 4. 7%. The additional hardware overhead of the router has only increased by 3. 2%.

Abstract:To handle the problems of difficulty in extracting fault features and low identification accuracy of traditional transformer fault identification methods, a novel identification method is proposed by fusing standard deviation-based multiscale fuzzy entropy ( SDMFE) and Harris hawks algorithm ( HHO) optimized support vector machine ( SVM). Firstly, multiscale analysis method based on fuzzy entropy is employed to quantify the complex dynamic characteristics of transformer vibration signals, and then extract fault features under multi-time scales. Subsequently, the fault features obtained by SDMFE are input into SVM for identifying transformer different faults. At the same time, to improve SVM recognition performance, an optimization strategy integrating HHO algorithm is introduced to select SVM parameters adaptively and accurately. Finally, a comparative experiment is carried out using the measured vibration signal of the transformer. Compared with different information entropies, different optimization strategies and different classifiers, the proposed method achieves the highest identification accuracy of 98. 56% and identification stability. The results show that the proposed method can effectively extract fault sensitive features and accurately identify transformer fault status.

Abstract:As the obstacle detection technology has become an important part of ADAS, many kinds of obstacles are difficult to cover. Therefore, a method of detecting small obstacles during car parking based on panoramic vision system is proposed. First, the real-time pictures in four directions map to the panoramic view using perspective transformation and image mosaic. A detection model of the ground feature point is proposed, which can be accurately extracted and matched with the right angle intersection of the parking line. After obtaining the ground feature points in the two frames,self-vehicle motion estimate is calculated by the SVD decomposition method and the simulated current frame of the previous frame is obtained, and the dynamic background is eliminated. Finally, a detection method based on color segmentation is proposed to determine whether it is an obstacle part. To verify the feasibility of the algorithm, various small obstacles were placed in the parking for testing, with a total of 864 obstacles in the three video sequences, with an average true positive rate of 94. 7% and an average false alarm Rate of 7. 3%. The results show that the algorithm can detect small obstacles in parking with certain accuracy and robustness.

Abstract:Aiming at the problem of unstable pose measurement of roadheader fuselage in a low-illumination and non-uniform dust environment in coal mines, a vision measurement technology for roadheader fuselage with three laser spots as point features is proposed. According to the characteristics of strong laser penetrability, the laser target image collected by the explosion-proof industrial camera is processed. Through the combination of the smallest inscribed rectangle and the ellipse fitting of the spot area, the spot characteristics of the three laser pointers are obtained. It adopted the P3P monocular vision positioning algorithm to calculate the spatial pose of the tunneling machine body through the transformation of the spatial coordinate matrix. According to the posture measurement experiment of the roadheader’s fuselage, under the influence of dust and stray light, the error of the position information of the roadheader’s fuselage obtained by this method is within 30 mm, and the attitude error is within 0. 5°. Accurate extraction of laser spots and stable measurement of body pose under complex backgrounds such as stray light, dust and water fog, basically meet the accuracy requirements of roadway excavation construction.

Abstract:The ground surface is not an ideal conductor of a homogeneous medium. The traditional one-dimensional atmospheric electric field component cannot accurately locate the thunderstorm cloud. When the three-dimensional field intensity component of the atmospheric electric field is used to locate the thunderstorm cloud, the field strength of the electric field and the altitude angle of the thunderstorm cloud are often measured. The electric field will be affected by the dielectric constant of the ground and air, resulting in positioning errors. In order to solve the problem of low positioning accuracy of thunderstorm clouds, this paper analyzes the sensitive characteristics of the three-dimensional atmospheric electric field to the relative permittivity of the surface, combined with the mirror image method, and establishes a thunderstorm cloud positioning model using the principles of electric field distribution in the air and the charge structure of thunderstorm clouds. Analyze the relationship between the atmospheric electric field strength, altitude angle and the relative permittivity of the environment. The experimental results show that the greater the relative permittivity of the surface, the greater the measured electric field horizontal component and the height angle of the thunderstorm cloud. The correlation coefficient between the relative permittivity of the surface and the horizontal component of the atmospheric electric field is -9. 5, showing a negative correlation. Therefore, to obtain the accurate height and azimuth of the thunderstorm cloud, the relative permittivity of the ground surface must be corrected in real time.

Abstract:In the surface geometric features measurement of automobile structural parts, the traditional point cloud simplification method will destroy the geometric features in the point cloud at a high simplification rate, reducing the integrity and dimensional accuracy of geometric features. To solve this problem, a point cloud simplification method for geometric feature preservation was proposed. Firstly, the K-means clustering of point clouds was carried out based on the idea of spatial region segmentation, and geometric feature descriptors were constructed to extract feature region point clouds by calculating information entropy in the cluster. Secondly, the iterative clustering reduction based on fuzzy C-means (FCM) & K-means was carried out for the point cloud of feature region, and octree reduction was carried out for the point cloud of non-feature region. Finally, the simplified point clouds of different regions are spliced to achieve the purpose of simplification. The results show that the proposed method can completely retain the geometric features of the model surface and avoid the appearance of holes. At the simplification rate of 94. 30%, the maximum error between the simplified point cloud and the original point cloud is 0. 912 mm, and the root mean square error is 0. 041 mm, which is more accurate than the traditional method.

Abstract:A person re-identification method based on feature fusion and multi-scale information was proposed to solve the problem of low accuracy of person re-identification due to the large difference of human image background and similar global appearance of human body. Firstly, the global feature map of human body image is extracted by ResNet50. Secondly, the branch structure is designed. In the first branch, the spatial transformation network is used to align the global feature images adaptively, and the local feature images are obtained by horizontal segmentation of the global feature images. The correlation between the global feature and each local feature is mined by fusing the global feature and each local feature separately. The second branch adds four convolution layers of different scales to extract multi-scale features from global images. Finally, in the reasoning stage, the features of the first branch and the second branch are connected in series as the comparative features of person. Experiments on the Market-1501 and DukeMTMC datasets show that the proposed method has better performance than the AlignedReID and EA-NET feature alignment and local feature extraction methods. In the Market-1501 dataset, mAP and Rank-1 reach 86. 77% and 94. 83%, respectively.

Abstract:In order to solve the problem that the diagnostic model cannot assign different weights to the features in the one-dimensional vibration signal according to their importance, which leads to the failure to extract representative features and thus affects the accuracy and robustness of the diagnostic model. A feature highlighting method based on reverse attention amplification mechanism ( RA) is proposed to reduce the proportion of non-important features by reversing attention and pruning the features, so as to highlight the important features. The long short term memory (LSTM) network is used to learn the temporal information between the features and to classify the fault types through the fully connected layer. The optimal data interception length, pruning hyperparameters and the stability of the model after adding noise to the signal are selected experimentally. The optimal data interception length, pruning hyperparameters are experimentally selected and verified the stability of the model after adding noise to the signal. The experimental results show that the proposed RA-LSTM bearing fault diagnosis method has excellent fault diagnosis performance, with fault diagnosis accuracy reaching 100% and excellent robustness even after the addition of noise.

Abstract:An automated detection system based on machine vision is proposed for the task of on-line detection of the crack defects on tongue spatula surface and the removal of inferior products. Firstly, based on the analysis of tongue spatula and its crack feature, a hardware device consisting of two groups of visual detection mechanisms is designed. The device is based on the chain-type conveyor belt as the basic transmission mechanism of the spatula. The specific assembly mode of chain-type conveyor belt and reflective photoelectric proximity switch is proposed to generate pulses and provide the timing sequence of the system. Based on multilevel caching mechanism, the system control architecture is designed for the collaborative allocation of the timing pulse triggering and the calling and enabling of each hardware component. In the aspect of crack detection algorithm, a method based on direction-space significance is adopted. Firstly, the preprocessing of OTSU algorithm, area screening and morphological operation is used to locate the spatula region. Then the crack feature points are extracted based on the direction-space significance. Furthermore, the candidate crack lines are generated based on the double threshold connection restriction. Finally, the cracks are accurately identified based on the characteristics of elongation angle, starting position and so on. The system performance is tested on the actual production site. The result shows that with the detection efficiency of 11 sticks per second, false positive rate (FPR) is as low as 4. 17% and false negatives rate (FNR) is 2. 68%, which are reduced by 6. 66% and 5. 36% respectively compared with the current manual detection method. It shows superior performance and has strong practical application value.

Abstract:Global warming is becoming more and more serious, and carbon dioxide, as the main component of greenhouse gases, needs to be precisely controlled. Tunable semiconductor laser absorption spectroscopy is widely used in gas detection and other fields due to its high sensitivity and high resolution. In order to further improve the measurement accuracy of the TDLAS system, the denoised TDLAS second harmonic signal was analyzed in the frequency domain on the basis of wavelet denoising, and the frequency domain characteristic signal related to the change of CO2 concentration was extracted by discrete wavelet transform. And establish a regression model to invert the gas concentration. The correlation coefficients of the time domain regression model calibration set and prediction set are 0. 998 5 and 0. 997 3, the root mean square error (RMSE) values were 0. 045 9% and 0. 017 9%, respectively, and the maximum relative error of the prediction set is 4. 62%. The correlation coefficients of the frequency domain regression model calibration set and prediction set were 0. 999 3 and 0. 999 7, the RMSE values were 0. 032 0% and 0. 006 9%, respectively, and the maximum relative error of the prediction set was 1. 54%. The experiment results show that the prediction ability and measurement accuracy of the TDLAS system were effectively improved, which verifies the feasibility of the method.

Abstract:Aiming at the shortcomings of the traditional total variational denoising methods, such as low peak SNR and low iteration efficiency, a new adaptive total variational denoising model is proposed in this paper. Firstly, the regular exponent of the total variational equation is improved by using differential curvature to distinguish noise points. Then, combined with the properties of level set curvature and gradient mode, the smooth region and edge region can achieve different denoising effects, so that the new model can preserve both edge and smooth noise. Experimental results show that compared with the current three mainstream models, the new model improves the peak signal to noise ratio (PSNR) by more than 1. 4 dB, reduces the mean absolute error by more than 2. 5, improves the iteration efficiency by at least 1. 6 times, and increases equally the structural similarity by 0. 13, which is more beneficial to practical application.