Abstract:Flight data are series of parameters related to flight and operation status. The flight data anomaly detection technique is aimed at monitoring the health status of key components of the aircraft and discovering the crew flight control problems, which benefits to the daily maintenance, eliminating potential safety hazard and ensuring aviation safety. However, the lack of labeled data, high accuracy requirement, computing resource constraint and other issues also pose serious challenges to practical use. This paper describes the basic connotation and research status of flight data anomaly detection, and on this basis, discusses the potential problems and possible development directions, and strives hard to provide feasible research ideas for the development of flight data anomaly detection.

Abstract:This study aims to enhance the accuracy and efficiency of copper burr detection for large motor rotor wires used in industrial environment. The problems of optical illumination, variety of burr species, and low recognition rate of the detection system are considered. To solve these problems, one kind of defect feature extraction method based on the copper burr growth region of the motor is proposed. The burr defect automatic recognition system is consistedof hardware module and image processing module. Based on analysis of burr features, the standard image of the detected objectis obtainedby using the maskbased image optimization algorithm. Then, the image region to be detected and segmented by morphological algorithm is constructed. Finally, the judgement results through classification algorithm and threshold denoising scheme for various burrs are achieved.The automatic detection of the copper burr defect is realized. Experimental results show that the algorithm can detect burr defects quickly and accurately. It has high robustness for detectingthe burrs generated during copper processing. The detection rate, leakage rate and false detection rate are close to 98%, 0%, and 1471%,respectively. The proposed method can meet the requirements of industrial test.

Abstract:When the circumferential scanning for thickwalled pipe is carried out with ultrasonic time of flight diffraction (TOFD), the ray path of lateral wave is inconsistent with the curved surface, which leads to the quantitative errors of length and angle for inclined cracks. To achieve deep crack detection, the probe center spacing (PCS) may be increased, which further increases the measurement errors. In this paper, considering the geometrical relationship among the curvature radius of pipe outer wall, PCS and crack tip depth, the accurate quantification study of the inclined cracks in thickwalled pipes is conducted. The quantitative formulas for the crack length and inclined angle are deduced for the detection of thickwalled pipes with circumferential scanning of TOFD, and the measurement errors before and after optimization are compared. The simulation results show that the quantitative errors of crack length and inclined angle are reduced by 010 mm and 158°, respectively for the crack with crack length of 40 mm and inclined angle of 10°~50° in the carbon steel pipe with wall thickness of 300 mm and curvature radius of outer wall of 1480 mm. Experiment on a carbon steel pipe specimen was conducted. In the experiment, the quantitative errors of crack length and inclined angle are reduced from 030 mm and 274° down to 027 mm and 028°, respectively for the crack with crack length of 40 mm and inclined angle of 30°. The proposed method is suitable for the quantitative detection of inclined cracks in pipes with different curvature radii, and has wide application fields.

Abstract:In the magnetic mixing frequency detectionof ferromagnetic materials, the excitation parameters of high and low frequency magnetic fields directly relate to the intensity and sensitivity of magnetic mixing frequency effect. Aiming at the magnetic mixing frequency detection issue of the hardness of ferromagnetic materials, the uniform design method is used to study the influence of excitation parameters on magnetic mixing frequency effect. Based on the fuzzy theory, a fuzzy evaluation model of mixing frequency effect with two characteristic indexes of variation coefficient and correlation coefficient is established. On this basis, the influence law of magnetic mixing frequency excitation parameters on the mixing frequency effect is obtained through regression analysis.The excitation parameter combination that can obtain strong magnetic mixing frequency effect is optimized. The optimization results show that the selection of the mixing frequency excitation parameters has a great influence on the magnetic mixing frequency effect.Using the optimized excitation parameters can obtain stronger mixing frequency effect and greatly improve the sensitivity of mixing frequency detection.The research work effectively promotes the application of magnetic mixing frequency detection technology in the field of nondestructive testing.

Abstract:near field cross spectral method; focused ultrasonic field; sound field reconstruction; multiparameter

Abstract:The magnetic interference will destroy the attitude information contained in the magnetic sensor output, and thus reduce the attitude estimation accuracy of the fusion estimation algorithms based on magnetic, angular rate and gravity (MARG) measurements. In order to improve the antimagnetic interference ability of attitude estimation, a method of correcting the magnetic sensor output under the condition of static or small motion acceleration is studied. In this method, the normalized magnetic sensor output vector is corrected by the gravity acceleration measurements. The correction method is based on the principle of a constant angle between the geomagnetic field and the gravity acceleration and the principle of minimum correction angle. The fusion model and the correction principle are analyzed, and the correction equations and formulas are deduced. The effectiveness of the correction method is verified by static and dynamic experiments. Experimental results show that the RMSE of pitch and roll estimation can be reduced by 26°and 16° respectively. In addition, the correction does not change the fusion process, and thus is compatible with other improvement measures.

Abstract:Recently, the series arc fault detection uses the main line current as decision object, which is easy to be affected by the singularity of normal working current in nonlinear load and results in misjudgment. And the arc fault characteristic in the small power branch is easily ‘submerged’ by the main circuit current, which leads to missed judgment. In order to solve this problem, according to the electromagnetic coupling mechanism of high frequency arc current, a detection method based on asymmetrical distribution of LN lines and highorder cumulant identification is presented in this paper. Through analyzing the coupling signal of high frequency residual flux during the process of arcing extinction and reignition, the kurtosis of the coupling signal is calculated by means of the highorder cumulant statistical tool. In this paper, the arc faults in different conditions such as combined load main circuit arc fault and combined load branch circuit arc fault are analyzed and judged, and the unified kurtosis threshold is obtained. The results show that the method can be effectively used to detect and identify series arc faults.

Abstract:Rehabilitation robot and assistive robot are expected to improve the life quality of patients, which has received extensive attention in basic researches and clinical fields. One of the significant purposes is to help disabled patients perform their daily activities. The humanmachine shared control robot system fully takes the advantage of human in upperlevel plan and machine in the fine control, which has achieved excellent control ability. First, this study briefly presents the composition and analyzes the advantage of humanmachine shared control robot system. Then, based on the review of the domain and abroad works, different types of humanmachine shared control robot systems are introduced in detail. Finally, this study looks forward to the future development direction of this technology through two aspects of new humanmachine interaction and shared control.

Abstract:Generative adversarial network(GAN) is an active branch of deep learning field, which has become a popular research direction in the field of artificial intelligence. GAN adopts an unsupervised learning method and automatically learns from the source data, which can produce amazing effects without artificially labeling data. In this paper, we present the background, basic idea of GAN and comb its related theory, training mechanism and stateoftheart applications. We also summarize the common network architectures, training skills and model evaluation metrics, and compareGAN with other generative model VAE and GAN variants. Finally, we point out the advantages and disadvantages of the GAN and look forward to the further development direction.

Abstract:Aeronautical blades are complex profiled surfaces. As one of the essentialparts of aeroengine, highprecision shape and structure detection is indispensable. Due to the complexedge curvature of complex profiled surfaces, the current point cloud registration technology can not satisfy the highprecision registration requirements. The highprecision registration technology is neededforaeroengine blade complex profiled surfaces detection. Thus, a new registration algorithm based on multiscale descriptor is proposed, which combines multiscale features with spatial angle difference threshold filtering algorithm to achieve highprecision registration of specialshaped surfaces. This method can effectively reduce mismatching points and noise points. Experimental results show that the registration algorithm has high registration accuracy and good robustnessfor complex profile surfaces of aeronautical blades.

Abstract:To explore the sensing performance of fiber bragg grating (FBG) embedded in the organisms, one kind of fiber FBG flexible sensor is developed. Its performance of sensing temperature and force is studied. First, the structure and manufacturing method of the FBG flexible sensor are introduced. Then, the temperature and pressure sensing model of the FBG flexible sensor are formulated. The pressure of flexible sensor is theoretically simulated by finite element analysis software Abaqus. Finally, comparison experiments between the FBG flexible sensor and the bare FBG sensor are implemented in terms of temperature and force perception. Experimental results show that the sensitivity of the flexible sensor during heating and cooling are 1314 and 1106 pm/℃, which are 13 times higher than those of the bare FBG sensor. The static pressures loading process and unloading process are 0318 1 and 0345 3 pm/g, respectively. These metrics are three times higher than those of bare FBG sensor. The linear correlation coefficient is 995%, which is better than that of bare FBG sensor. It has a stronger signal than bare FBG sensor for detecting the traffic load. The strain value of 20 g solid ball height from 50 cm is 38 με, which illustrates the obvious tracking effective ness. The sensor has high sensitivity and good repeatability, which has a significant reference for the study of the bionic body tactile slip sense.

Abstract:When the doublewavelength method is applied to detect nitrate nitrogen, its prediction accuracy is affected by the superposition of absorbance of organic pollutants. To solve this problem, one kind of method based on agent parameters is proposed. The six integral area values contained in the hybrid spectrum are used as the agent parameters. The demarcation point of the agent parameters is determined by the dynamic segmentation algorithm of the fourth derivative spectrum. Each agent parameter at the demarcation point is obtained by the area integral method. According to the ratio of organic pollutants agent parameters, their contribution degree for the hybrid absorption spectrum before the demarcation point is obtained. By using the preprocessed data, the least square method is used to formulate the prediction model for nitrate concentration and its integral area. This model is applied to the practical water sample detection. Compared with the traditional dualwavelength method, experimental results show that the determination coefficient of the model is 09998 and the average standard deviation is less than 3%. It can also realize lower detection limit (≥0009 mg/L), better recovery (95%~104%) and wider linear range, which provides reliable technical support for the accurate detection of nitrate nitrogen in water samples.

Abstract:With the development of advanced drilling technologies, more and more downhole data need to be transmitted to groundsurface in real time, however the transmission rate of existing logging while drilling (LWD) data transmission techniqueis too low to meet the requirement of actual drilling technology.This paper tries to apply the mature underwater acoustic communication technology to the LWD data transmission. The theoretical model of LWD data acoustic transmission was established, the influence of pipeline acoustic signal long transmission condition andacoustic frequencyon acoustic signal attenuation was analyzed. Finite element software COMSOL was used to simulate the logging while drilling acoustic data transmission system.Under the conditions of well depth of 2 000 m, acoustic frequency of 1, 4 and 8 kHz,and drilling fluid viscosity of 20 mPa·s, the influence of acoustic frequency on acoustic attenuation was simulated. The maximum relative error of simulation result and theoretical analysis is less than 85%, which verifies thecorrectness of the theoretical model. The influence of the drilling pipe diameter changing and bending on the acoustic transmission was analyzed. Finally, the experiment test was carried out, which demonstrates the technical potential of acoustic data transmission method in drilling process, and lays a foundation for the practical application of this technology.

Abstract:In this paper, the failure features for electrical life prediction of AC contactor under mechanical vibration environment are studied, and a method based on optimized wavelet denoising and kernel principal component analysisPearson correlation coefficient methods is proposed to obtain effective failure features. Firstly, the life cycle tests for AC contactor were carried out under normal and vibration environment respectively, and several degradation parameters were selected as the analysis objects. Secondly, aiming at the measured data, the multiindex comprehensive evaluation method was used to optimize the wavelet denoising and then the degradation parameters were denoised and smoothed with the optimized method. Considering the degradation trend of the parameters, the influence of vibration on the electrical life of the AC contactor was analyzed. Finally, the full information features were constructed for the degradation parameters and their kernel principal components, and Pearson correlation coefficients between the features and the residual electrical life were calculated. In the vibration environment, using the proposed method the cumulative arcing energy, first kernel principal component, closing voltage and contact resistance are finally extracted as the failure features for electrical life prediction of AC contactor.

Abstract:In the application of object tacking,there are some typical problemsin complex scenario,such as light changing, shadow, occupancy and moving background. To overcome light changing and shadow, the algorithm of shadow tolerant local binary similarity pattern is proposed, which is based on the local binary pattern background model. Then, the distance discriminator which is calculated between the current target bounding box and the target bounding box in history is utilized to solve the problems of occupancy and moving background. The distance discriminator is also adopted to detect target fragments. Based on this detection result, the merging procedure can be achieved. After merging, the optical flow of each target bounding box is calculated.The differencesin a block and amongblocks are calculated. The bounding box segmentation procedure is processed based on the calculated similarity information. Finally, the structure support vector machineis used to design the target association procedure. Experimental results on multiple target tracking benchmark show that the proposed algorithm can achieve excellent tracking precision and robustness.

Abstract:Distributed fiber optic sensor(DFOS) hasadvantage of leak detection in the buried thermal pipeline. However, limited by the spatial resolution, the distributed sensors have low accuracy to detect the local temperature variation caused by small leak, which makes the temperature measurement quite different from the practicaltemperature field. To solve this problem, one kind ofmethod is proposed to establish the correspondence between thesetwo temperature fields. Laboratory experiments and field tests are designed to simulate the actual leakage, where Raman optical time domain reflectometer(ROTDR) and thermistors are employed to measure the temperature. The extracted features based on the mapping relation between the actual temperature field and the DFOS measurements aredetermined using artificial neutral network(ANN). Experimental results show that the proposed ANN model is effective to establish the relation between the actual and measured temperatures. The measurement accuracy of ROTDR is improved from the aspects of data processing which provides valuablereference for small leak warning.

Abstract:Flow measurement is the most important factor affecting the test accuracy of hydro turbine efficiency. Ultrasonic method is mainly used in the flow measurement of large diameterpipe. However, there is no effective calibration method for its measurement accuracy and error composition. Based on the principle of ultrasonic time difference flow measurement, the comprehensive flow error is deduced and the flow measurement error description model is established. A quantization method for error analysis based on ideal flow measurement system is proposed, which provides an analytical approach for error analysis and control of ultrasonic flow measurement system. The influence of ideal flow measurement system on the accuracy of ultrasonic flow measurement is simulated. The influence of the measurement errors of various parameters on the comprehensive errors of the system is analyzed. The relevant correction methods are proposed for the major influential factors, and the control of the comprehensive error of the system is analyzed. Finally, the experimental system is set up, and the experimental results preliminarily verify the effectiveness of the proposed method.

Abstract:In order to improve the repairing effect of vibration signal, Bayesian compressed sensing (BCS) theory is introduced, and a Bayesian compressed sensingbased repairing method with empirical mode decomposition (EMD) is proposed to solve the problem of continuously missing signal restoration. For the randomly missing signals, Bayesian compressed sensing algorithm is designed to repair them based on the principle of compressed sensingbased repairing. While for the continuously missing signals, empirical mode decomposition is firstly performed on them, and then all the basic mode components obtained by decomposition are repaired by multitask Bayesian compressed sensing algorithm. Finally, all the repaired mode components are accumulated to get the whole signal. Experiments on open bearing data from Case Western Reserve University show that the proposed method is superior to orthogonal matching pursuit and regularized orthogonal matching pursuit in timefrequency domain, error, signaltonoise ratio and peak signaltonoise ratio. From the perspective of repairing effect, it is found that this method successfully restores the fault feature frequency in the basic mode components of the outer ring fault signal, and achieves the purpose of repairing.

Abstract:Theinstantaneous surface temperaturemonitoringis always challenging for the aeroengine hot components composed of C/SiC composite materials.An measurement method of instantaneous surface temperature monitoring is studied for the aeroengine hot components composed of C/SiC composite materials. The NiCrZrO2 composite transition layer isdeposited on the surface of C/SiC composite by magnetron sputtering and electroplating.Insulating film SiO2, NiCr/NiSithin film thermocouple and SiO2 protective filmareprepared sequentially on the NiCrZrO2composite metal transition layer on the surface of C/SiC composite prepared by magnetron sputtering.The insulation properties of SiO2 insulating films with different thicknesses areinvestigated. The results indicatethat the insulation resistance of SiO2 films with a thickness of 3 μm is about 164×109 Ω. The experimental results of the static performance of the sensor show that the Seebeck coefficient is 421 μV/℃ and the nonlinear error is 152% in the range of 50～600℃. The dynamic performance of the sensor is studied by the combination of theoretical calculation and experiment. The results show that the dynamic response time of the sensor is in the microsecond range, which can realize the transient temperature test. The temperature measurement experiment shows that the sensor can meet the requirements of transient temperature detection ranged from room temperature to 600℃.

Abstract:The current indicator diagram sensing system of rod pumping well is not suitable for the pumping well at a low pumping speed, and the displacement accuracy is low. To overcome the shortcomings of the existing measuring equipment, an automatic indicator diagram sensing system with high precision displacement is designed. By analyzing a large amount of the data from the indicator diagram, it can be concluded that the motion trajectoriesof the polish rod are similar when the strokes are identical. An automatic acquisition method of high precision indicator diagram is proposed by using position sensing and displacement multiplexing, and the hardware composition and the key software of the sensing system are introduced in detail.Through analysis of the real data collected by the sensing system in the oil field, the results show that the sensing system is stable and reliable, and the lowest stroke can reach 02 times per minute.Moreover, the error of the stroke and the displacement is less than 1%. Compared with other types of indicator diagram sensing system, it has obvious advantages, and it completely satisfies the requirements of the indicator diagram acquisition system with widerange pumping speed and high precision for the oil field.

Abstract:In order to studythe effect of excitation direction on the output performance of vibration energy harvester(VEH), the dynamic model and the output model of piezoelectric effect are obtainedfor a single cantilever beam energy harvester under vibration excitation in any direction. The direction of vibration is defined by two angle parameter φ and θ from spherical coordinate system.According to the theoretical analysis and simulation, it is found that the absorption efficiency of the energy harvester varies ascosine function with the increase of φ when the vibration excitation direction varies, and varies assine function with the increase of θ.But the influence of φ on the collection efficiency is much greater than the influence of θ. Experimental results show that when θ=0° and φ increases from 0° to 90°, the efficiency of the VEH decreases from 9353% to 288%. When φ=90° and θ increase from 0° to 90°, the efficiency increases from 288% to 388%. The relationship of VEH′s output with the vibration direction is obtained, which can providecertain reference value for the research of VEH with cantilever beam.

Abstract:The sparse unmixing model is improved based on the spatial constraint which represents the similarity and difference between the adjacent pixels.Thus, the accuracy of the hyperspectral image unmixing is increased.The unmixing spectrum library is generated by compressing the primary spectrum library to increase the influence of the spatial constraint on the unmixing model, and reduce the sparsity of hyperspectral image in the unmixing spectrum library.According to unmixing spectrum library, the improved unmixing model is constructed by sparse unmixing and manifold regularization, which represent the similarity and difference between the adjacent pixels.The improved unmixing model is solved by means of the convex optimization algorithm such as alternating directions method of multipliers.Experimental results show that the proposed algorithm has high spectral unmixing accuracy and strong performance.

Abstract:Recognizing the target face from the surveillance video using the given singletarget image is still a challenging issue in practical applications. Therefore, a lowresolution singletarget face recognition algorithm with singlesample is proposed in this work. Two significant limitations are taken into consideration, the sample numbers of the target objects and nontarget objects are seriously unbalanced in the probe set, and the single object problem cannot utilize the mutually exclusive relationship between different categories. In this paper, firstly, to increase the robustness of the openface face recognition, the clustering algorithm is utilized to transform the single object recognition problem into a multi classification recognition problem. Furthermore, the iterative label propagation algorithm is applied to optimize the attribution probability of the probe sample continuously. During the iteration, the face recognition threshold of each object is estimated according to the confidence probability. Hence, the single sample is capable to train the classifier. Finally, experimental results on multiple face datasets show that the proposed algorithm can achieve good performances in both accuracy and recall rate.

Abstract:The measurement technology of line structured light vision is utilized for the rail wear measurement due to its noncontact and high precision features. The measurement of rail full profile and field calibration of line structured light visual measurement system are two typical problems. To enhance its performance, a binocular calibration method of line structured light based on free plane targets is proposed for field application. Firstly, the vision measurement system is established based on line structured light binocular visual measurement model. The chessboard images containing line structured light and the chessboard images without line structured light at any positions in the public viewing angle of both sides of the camera are collected respectively. Then, the internal parameters of both sides of the camera are obtained by using the chessboard plane calibration method. The feature points which are generated by intersection of line structured light plane and target plane at different positions are used to fit the planar equations of the line structured light plane in both sides of the camera coordinate system. According to the Rodriguez transform principle, the external parameters between the line structured light plane and two cameras are solved. Finally, the internal parameters and the external parameters are utilized to realize the measurement of rail full profile. The field test is also carried out. Experimental results show that the calibration error of the cameras is about 003 pixels, the fitting degree of line structured light plane is 0999, and the total measurement error of rail profile data is 054mm, which meets the requirement of measurement accuracy.

Abstract:To achieve linear motor mover position measurement, a subpixel displacement image detection method based on extended sampling phase correlation method (ESPCM) is proposed to improve measurement accuracy and antiinterference performance. Firstly, a linear motor position detection system is formulated. The stripe image sequences are captured in real time by a highspeed camera. Secondly, the edge feature extraction is performed on the stripe image. The crosspower spectrum of the adjacent stripe image is obtained based on the phase correlation, which is the integer pixel displacement of mover position. Then, the upsampling phase correlation is calculated based on the crosspower spectrum of the integer pixel neighborhood. The highprecision subpixel displacement measurement is realized and the actual displacement value is further obtained by system calibration. Compared with the traditional phase correlation algorithm, the proposed method can improve the measurement accuracy and noise suppression performance. Finally, an experimental detection platform is established to verify the effectiveness of the proposed algorithm.

Abstract:Aiming at the low accuracy of common target recognition and tracking algorithms caused by camera high frequency sloshing under dynamic view, this paper proposes an adaptive window target tracking algorithm based on Canny and GrabCut. Firstly, the speeded up robust features (SURF) algorithm is used to learn the picture library and remember the picture features. The memory target recognition algorithm based on SURF algorithm is designed. Then, GrabCut adaptive optimization algorithm is used to segment the region of interest to achieve rough tracking of the target. Finally, a windowed algorithm based on Canny is achievedtoaccurately track the target. The experimental results show that the algorithm designed in this paper can quickly identify the target and accurately outline its contour. Moreover, the target can be stably tracked. Compared with other algorithms, the algorithm has obvious improvement in computation efficiency and accuracy.

Abstract:The development of machine vision technology provides an effective method for automatic measurement of particle size distribution. However,the overlapping particles in theimageis still difficult to be segmented. To solve this problem, one kind ofpellet image segmentation algorithm based on morphological reconstruction and Gauss mixture model is proposed. To achieve unsupervised clustering, morphological reconstruction combined with clustering validity index isused to obtain the optimal number of clusters.EM algorithm isutilizedto solve this problem. Finally, the missing particle contours arereconstructed by circle fitting method.The segmentation of overlapping pellets is realized. Experimental results show that the algorithm can effectively segment overlapping pellets.The segmentation accuracy evaluation index AC is 936%, which is obviously superior to the compared algorithms. The measurement of particle size distribution based on machine vision is founded.

Abstract:There are serious false alarm and poor detection performance of the local anomaly detection operator for hyperspectral image. To solve these problems,one kind of improved anomaly detection algorithmbased on background discrimination and neighborhood compensation by kernel spectral angle is proposed.The background pixels screening and detecting results compensation are taken into account.In the termof background pixel processing, thealgorithm based on kernel spectral Angle distance similitude is proposed.The kernel spectral angle with stronger spectral resolution is introduced into the background difference discrimination process. In this way,the optimization of local background pixel accurately and reliably is realized.Meanwhile, to solve the problem of low detection accuracy, the joint compensation mechanism of spacespectral characteristics is introduced into neighborhood weighting.The dynamic template convolution compensation algorithm based on kernel spectral Angle distance similarity is proposed, which significantly enhances the separability of background and target. Compared with other abnormal detection algorithms (e.g., RX, LRX, KRX and CRD), the proposed algorithm shows strong detection performance and achieves good effectiveness in suppressing false alarms and improving detection accuracy.

Abstract:Permanent magnetic levitation system is a typical nonlinear system. As the air gap between the magnetic conductor and suspension changes, the control characteristics change which affects the stability of the system. In order to prevent the external interference from causing the suspended matter falls or adsorbs on the magnetic conductor, as the system deviating from the equilibrium point. An antifall and antiadsorption control method based on air gap variation is proposed. The multiplepoint preset levitation force is selected in the entire suspended air gap, the control parameters are calculated based on the preset levitation force of each design point, and the complete antifall and antiadsorption control law is obtained according to the interpolation principle. This method can effectively compensate for the uneven variation of levitation force, and improve the controllability, stability and robustness of the system. Simulation and experimental results show that the antifall and antiadsorption control method can mitigate impact and improve antifall performance. When the air gap is enlarged by the external disturbance, the levitation force increases rapidly to prevent the suspended object from falling. When the external disturbance causes the air gap to become smaller, the levitation force rapidly becomes small, and the suspended matter is prevented from being adsorbed on the magnetic conductors.

Abstract:Aromatic hydrocarbon yield is considered as one of the important product quality indicator in catalytic reforming production process. Aiming at the difficulty of the aromatic hydrocarbon yield online measurement, a soft sensor modeling method of aromatic hydrocarbon yield is proposed based on mutual informationimproved gravitational search algorithm (MIIGSA) optimized extreme learning machine (ELM). Firstly, the MI method is used to extract the most relevant process feature quantities and perform dimension reduction processing, and the auxiliary variables of the soft sensor model are determined. Secondly, through introducing the successive quadratic programming (SQP) method and chaos mutation strategy, the IGSA with good global optimization performance is constructed. The IGSA algorithm is then applied to optimize the hidden layer threshold parameters and input weight parameters of ELM, and the optimization target considers the minimization of both the root mean squared error (RMSE) of the model output and the number of conditions of the hidden layer output matrix. Finally, the aromatic hydrocarbon yield soft sensor model is established based on the IGSA optimized ELM method. The proposed model was applied in the prediction study of the aromatic hydrocarbon yield of the catalytic reforming equipment in a certain refinery plant, the simulation results show that the proposed soft sensor model possesses promising prediction accuracy and reliability.