Abstract:The four-dimensional information of space and angle can be obtained by one-time imaging of light field. The existing methods are mostly used for the light field image of small baseline scene in the angular super-resolution reconstruction, and there are some phenomena such as blur when reconstructing the large baseline scenes. At the same time, the reconstruction effect of the occlusion area is poor in the process of light field reconstruction, and the long-distance spatial relationship is difficult to capture. To solve this problem, a sparse light field intensive reconstruction method combining depth clues and geometric structure is proposed. This method uses spatial pyramid pool to extract multi-scale features, which can preserve the texture details and high frequency information of images better. By introducing void convolution and dense connection in the depth estimation module, the receptive field is expanded and the accuracy of depth estimation of large baseline scene is improved. The view refinement module is used to optimize the image and reconstruct the occlusion area while preserving the parallax structure. Experimental results show that the proposed method solves the problem of largebaseline scene optical field reconstruction well, and exceeds other algorithms in the large-baseline scene data set, with PSNR increased by 2 dB and SSIM increased by 0. 018. The quality of reconstructed images is superior to the existing algorithms.

Abstract:Speckle noise in optical coherence tomography (OCT) images obscures important morphological details, and hinders the clinical observation and diagnosis of retinal lesions. A structural similarity constrained generative adversarial network ( SSGAN) is proposed for retinal OCT image denoising. The proposed SSGAN utilizes the residual strategy to improve the structure of original generative adversarial network, and incorporates the structural similarity index measure loss into the objective function to achieve more structural constrains while suppressing speckle noise. The experiments on the SD-OCT public dataset published by Duke University show that the peak signal-to-noise ratio and edge preserve index of the proposed method are 28. 08 and 0. 960 respectively, outperforming the other denoising comparison methods. Further experiments demonstrate that the proposed method can be easily applied to other public datasets from the A2A SD-OCT study.

Abstract:To solve the problems of blurred edge information and poor accuracy of small targets feature extraction when the lightweight semantic segmentation algorithm is applied to the UAV high-resolution traffic scenes image segmentation, a lightweight semantic segmentation algorithm combining attention mechanism and ghost feature mapping is proposed. Firstly, the hybrid attention module is embedded in the semantic branch 8-fold and 16-fold down-sampling process of the BiSeNet V2 to redistribute the weights of the deep feature maps and enhance the local key feature extraction ability. Then the ghost feature mapping unit is used to optimize the traditional convolution layers to further reduce the computational cost. Finally, the dynamic threshold loss function is applied to supervise the training, adjusting the training weights of the high-loss difficult samples. Using the UAVid dataset to train and test the improved algorithm, it is found that the mIoU is 52. 7%, which is 7. 8% higher than the BiSeNet V2. When the input images size is 1 280×736, the inference speed can reach 73. 6 FPS, meeting the real-time segmentation requirements. The results show that the algorithm can be well adapted to complex traffic scenes, and can effectively improve the problems of blurred edge information and poor accuracy of small objects.

Abstract:Polarization is one of the important characteristics of light. Polarization imaging technology can obtain the intensity information and polarization information of the target in the scene. Polarization information can reflect the material characteristics of the target surface. In this paper, two image enhancement schemes based on polarization information are proposed to meet the accuracy requirements of common target recognition results in road scenes under haze weather conditions. First of all, the polarization data set is constructed through multiple acquisition experiments, data cleaning and image labeling, with a total of 4 649 images and 31 877 tags. For the scene with slight haze pollution, the sky region in the polarization intensity image is segmented by the region automatic growth algorithm, and the reflected light of the target is reversely generated according to the polarization degree and polarization angle information of the sky region and the atmospheric physical scattering model, so as to realize the image defogging. For the heavily polluted scene of haze, wavelet transform is used to enhance the image, and the degree of polarization image is used to enhance the target contour in the intensity image. The image gray variance and image information entropy are used as image quality evaluation indicators, and the YOLO v5s deep learning network is used for object detection. The results show that in the case of light haze pollution, the image quality and object detection accuracy have been improved, the image information entropy has increased by 3. 36%, the gray variance has increased by 40. 27%, and the object detection mAP has reached 76. 40%, increased by 12. 69%. In the case of heavy smog pollution, the object detection mAP increased by about 1. 69%.

Abstract:The interaction of hands and tools is the key information to distinguish the behavior of workers. To prevent the errors and omissions of the process in the assembly of pumps and achieve the purpose of real-time monitoring, a workshop tool detection algorithm based on spatial feature fusion is proposed. First, in order to improve the localization ability and detection accuracy of the object, the hand motion region in the foreground is segmented based on the frame difference method to obtain a texture map with hand spatial information, which is combined with RGB images of the assembly process to form a dual channel inputs to the object detection network. The spatial perception module is designed to realize the spatial feature fusion of the dual channel inputs and obtain the global spatial information. The feature enhancement module is proposed to mix the global spatial information and deep semantic information to enhance the feature response at salient locations. Then, the ESNet (enhance shuffleNet) is used to reconstruct the backbone network and form a multi-scale feature extraction module by deep separable convolution to improve the detection speed. Finally, in view of the local elements change both in the foreground and the background, the CutOut data enhancement method is used to improve the anti-interference capability. The experimental results show that the proposed algorithm can effectively reduce the false detection rate, and improve the mAP by 6. 4% compared with the traditional YOLOv5s. The method can quickly and accurately detect the tools used by shop workers.

Abstract:Hazardous gas detection technology is one of the necessary guarantees of safety in many industries such as petrochemical enterprises, and plays an important role in industrial production. To make up for the shortage of single-frame gas leak detection algorithm which based on infrared imaging technology and improve the detection accuracy, a method of optical flow enhanced infrared imaging gas leak detection was proposed in this paper. The motion features in the video are extracted with the optical flow network in the first step. Then the motion features are fused with the original image to generate an optical flow enhanced gas leak image to be sent to the YOLO network for detection. Finally, whether there is a leakage in the video was determined and its location was obtained. After a comparison with the data before enhancement, the method keeps the recall rate in an acceptable range while the false alarm rate is reduced from 17. 87% to 0. 60%, and precision is improved from 77. 21% to 99. 99%. The detection speed is about 13 fps, which satisfies the needs of real-time detection.

Abstract:As the fingerprint image collected by the fingerprint sensor tends to become miniaturized, the fingerprint image contains less and less fingerprint feature information. Aiming at the problems of large calculation, unsatisfactory accuracy and poor anti-interference ability of traditional template matching algorithms when processing small-area sliding fingerprints, this paper proposes an improved sliding fingerprint tracking algorithm based on MOSSE. The improved MOSSE algorithm uses multiple inputs, weighted fusion of grayscale features and HOG features at the response layer, and introduces the Fourier-Mellin algorithm and Hanning window to process the fingerprint of rotation. The results of tracking small-area fingerprints are compared by a variety of algorithms, which shows that this algorithm inherits the advantages of the original MOSSE algorithm, and improves the fingerprint matching accuracy, the matching accuracy of normal images is 99%, the matching accuracy of noisy images is 90. 3%, and the average calculation time of each frame is 0. 103 6 s, which ensures the real-time and robust nature of fingerprint tracking. It can also track deformed and rotated fingerprint images well.

Abstract:The automatic reading of pointer instrument by camera is easily affected by complex environment, different camera angles and other factors, and it is difficult to balance the detection speed and detection accuracy in practical applications. Therefore, this paper proposes a pointer instrument reading algorithm based on key point detection. ResNet18 is used as the backbone network, the residual blocks in the last two stages and subsequent fully connected layers are abandoned, and a lightweight feature fusion network is designed according to the characteristics of the pointer meter panel, while introducing a pose refine machine ( PRM) that improves model performance. Finally, using the obtained three key point information of the dial circle center, the zero scale line, and the current pointer scale, the reading calculation is completed by the angle method. The experimental results show that, the reading error of the algorithm in this paper is only 0. 506%, and the speed can reach 53 frames/ second, which is more accurate than the traditional algorithm; compared with other similar algorithms, the proposed algorithm can still achieve high accuracy prediction of pointer key points with fewer parameters and computational complexity, fully proving the effectiveness of the proposed algorithm.

Abstract:There is a problem of low positioning accuracy when using Apriltags to measure tags in long distance. In this paper, we analyze the causes of the problem, and propose a method to improve the longitudinal measurement accuracy and range of Apriltags based on relative positioning technology. This method solves the problem by dividing long distance into few short distance and obtains the space position by calculating with ensuring the measurement accuracy. When both direct measurement and indirect measurement results exist, combine the respective advantages of the two methods through the way of data fusion. This method can effectively improve the longitudinal distance measurement accuracy of tags at medium and long distances while ensuring the measurement accuracy of transverse distance and height. We verified the effectiveness of the method by conducting the simulation experiments and real experiments, achieved the reduction of 1 dm absolute measurement error and 0. 45% relative measurement error at a distance of 3 m ( equal to 12. 5 times the tag size).

Abstract:For the classification of metal defects, the mainstream classification methods represented by deep learning are mainly statistical learning methods based on large datasets. However, when applying deep learning, not only many quality labeled samples are needed, but also the result may suffer poor generalization. A classification approach with few samples is proposed, which embeds the hierarchical and concise knowledge of humanoid into deep learning. First, a CNN is built as the backbone of the classification model, and a humanoid learning module is designed, which uses the features of human classification to classify. To improve the generalization, robustness and better fusion effect of the model, a mathematical integration model based on logarithmic function is designed. The mathematical integration model in the module couples the outputs of backbone network and humanoid learning module by using the idea of integrated learning. The experimental results show that for the metal defect data of small training set and large test set, the classification performance and the amount of training parameters are better than the deep learning method. Humanoid learning module and mathematical integration model are embedded in different backbone, and good performance is achieved, which shows that the proposed method is suitable for various deep convolution neural networks.

Abstract:To address the problems in the application of lower limb exoskeleton mechanical equipment, XGBOOT Algorithm-based research on gait phase recognition is carried out, only using motion attitude data measured by a single IMU. Firstly, foot motion data of six different gaits are collected, and each gait is divided into four phases. On this basis, XGBOOT algorithm optimized is applied to analyze the gait phase recognition with the foot motion data as the training set. In the process of establishing the model, the parameters involved in the model are further optimized by the Bayesian optimization algorithm (BOA). Through calculation, the results show that the average accuracy of the model is 89. 26% in the verification set, the precision of the model is 89. 64% in the verification set, the recall rate of the model is 89. 26% in the verification set, F1 value of the model is 89. 10% in the verification set, which indicates that the model can achieve better gait phase recognition.

Abstract:This study proposes a teleoperation system based on a six-degree-of-freedom robot arm, aiming to design a non-wearable and intuitive control method. The system uses a Kinect V1 camera and a UR3 robot arm, with the Microsoft skeleton recognition library as the basic method for human pose recognition. By mapping the human arm to the robot arm joints, the robot arm can track the motion of the human arm in real-time. Meanwhile, the nonlinear model predictive control (NMPC) algorithm is used to optimize the robot arm motion control, and fuzzy rules are set to achieve adaptive adjustment of NMPC parameters. The experimental results show that under the optimization of NMPC, the robot arm has significantly reduced average displacement and rotation errors in both the x and z translation directions and the three rotational directions, as well as average error in joint angle changes. The test results also demonstrate that the overall motion tracking performance of the robot arm is good, verifying the accuracy of the proposed mapping rules and kinematic model, as well as the effectiveness of the fuzzy NMPC controller.

Abstract:Fault area detection is one of the important contents of circuit board fault diagnosis. In recent years, a lot of research work has been devoted to explore circuit fault diagnosis methods through theoretical simulation, but the difference between simulation conditions and actual measurement environment reduces the feasibility of such methods in practical applications. Combined with the characteristics of the measured circuit board data, this paper proposes a circuit board fault area detection method based on near-field scanning and time series similarity measurement. This method obtains the electromagnetic radiation data of the circuit board under normal and fault conditions through near-field scanning, and uses the variational mode decomposition (VMD) method to reduce the noise of the original data. After that, the data in the two states are regarded as two types of time series, the improved time series similarity measurement algorithm is used to calculate the distance value of the two types of time series, and the fault area of the circuit board is determined according to the distance value. According to the experimental results of data sets, the similarity measurement algorithm in this paper shows better measurement ability in processing time series than other algorithms, and the classification accuracy of distance value is also 6. 3%, 8. 4% and 4. 2% higher than the three comparison algorithms. At the same time, the consistency between the experimental results of the measured data and the theoretical simulation results verifies the reliability and practicability of the method in this paper. This method provides a new way to realize the circuit board fault diagnosis.

Abstract:An intelligent fuzzy monitoring algorithm for concrete slurry concentration considering multiple influencing factors is proposed. Firstly, the influence factors, including production plan, traffic restriction, precipitation and temperature, which will affect the monitoring cycle of the muddy commercial concrete concentration, are analyzed and described with mathematical functions. Secondly, the fuzzy domain set of four factors is established by combining the fuzzy theory with the expert scoring weight. Accordingly, the trapezoid (semi-trapezoid) membership functions of the influence factors are designed respectively. Finally, a fuzzy decision calculation formula for the slurry concentration monitoring cycle is given based on fuzzy theory. The test results of a stirring station plant experimental platform show that the proposed algorithm can adaptively decide the monitoring period. Result analysis indicates that the proposed method can meet the monitoring requirements and effectively reduce the working time of concentration sensor by 16. 7% when compared with other automatic monitoring method, revealing the practical application value of the proposed algorithm.

Abstract:In view of the complex structure and severe working conditions of gas turbine, a multi-channel convolutional neural network (MCCNN) deep transfer learning gas turbine rotor fault diagnosis was proposed for the problem that it was difficult to obtain the rotor system fault samples during operation and the fault diagnosis accuracy was low due to the small sample size. The method firstly, took the one-dimensional raw vibration signal of the bearing as the input, then rearranged and combined the data to simulate the converted twodimensional image, effectively avoiding the tedious operation of the actual converted image. The MCCNN model was trained with the public bearing data from Case Western Reserve University (CWRU) and Xi′an Jiaotong University (XJTU) to update the weights and diagnose. The fault classification accuracy is up to 99. 95% ~ 100%. CWRU bearing fault datasets were used as the source domain and the gas turbine rotor fault datasets were used as the target domain, the model parameters obtained from the source domain training were retrained by using transfer learning method for the target domain datasets and the classification accuracy for the gas turbine fault data was 97. 78%. The experimental results demonstrated that multi-channel convolutional neural networks and transfer learning model is suitable to the task needs and can solve the problem with a small sample size of rotor system.

Abstract:With the popularity of renewable energy generators in the power grid, the traditional doubly-fed induction generator (DFIG) vector control strategy did not provide additional active power support for the power grid due to the decoupling of inertia and power grid frequency fluctuation. To solve this problem, an improved virtual synchronous generator (VSG) control strategy is introduced into the rotor-side converter (RSC) control of doubly-fed wind turbines. Aiming at the frequency deviation problem existing in the traditional VSG control strategy, a Washout filter is applied in the P-f droop link, which can keep the frequency at the stable power frequency of 50 Hz without adding additional secondary control loop. At the same time, in order to solve the problems of limited dynamic response and poor robustness of the inner loop of voltage and current control based on internal proportional integral ( PI) control in traditional VSG, the port controlled hamiltonian with dissipation (PCHD) is used for the first time in this paper. The interconnection and damping assignment passivity based-control ( IDA-PBC) method of PCHD model was applied to the inner loop current control. A passive controller for rotor side converter (RSC) inverter is designed. Finally, a simulation system is built for experimental verification, and the results show the effectiveness and superiority of the proposed control strategy.

Abstract:Aiming at the nonlinear bifurcation and chaos behavior of the second-order inverter under the linearized model, the control method is added to increase the stability domain and disturbance immunity. Firstly, established the second-order coefficient linear iteration model of inverter, recreated the chaotic behavior of system by using bifurcation diagram and folding figure, sought Jacobian matrix and equilibrium point of the system. Secondly, the delay feedback control is added and the limiting conditions of the control parameters are given based on the Juri criterion. Finally, numerical modeling experiments are unfolded to declaration the meliority of the chaos control method. The results indicate that the bifurcation point of the proportional parameter K of the system moves from 0. 573 to 0. 973 after the delay feedback control is applied, and the system still runs stably when the input voltage of dc power supply changes abruptly. The results show that the stability domain of the second-order inverter is increased and the overall disturbance immunity is enhanced after the delay feedback control is applied.

Abstract:In order to improve the average accuracy of loudspeaker abnormal sound classification, a convolutional neural network plus gated current unit (CNN-GRU) and sparrow search algorithm optimization variational modal decomposition ( SSA-VMD) model was proposed to classify loudspeaker abnormal sound. In the aspect of feature extraction, the SSA-VMD model was used to determine the optimal value of the second penalty factor (α) and modal decomposition number (k) in VMD, so as to improve the accuracy of feature extraction and reduce the extraction time. Finally, the VMD was used to extract the characteristics of the loudspeaker response signal. In terms of classification network, the CNN-GRU network was used to classify the abnormal sound of loudspeakers, the CNN-based feature extraction network was used, and the GRU network was used for deeper feature extraction to achieve the goal of improving the average classification accuracy of loudspeakers. The experimental results show that after optimizing the parameters of SSA-VMD model, VMD can extract features more effectively, and the decomposition time was reduced by 59. 8%. The CNN-GRU model has a higher and more stable recognition rate, with an average classification accuracy of 99. 2%.

Abstract:Optical voltage sensors face the problem of temperature stability. Based on the Pockels effect model of BGO crystal and the thermo-optic effect and so on, the temperature response model of the optical voltage sensor under the action of multi-physical fields is derived and the spectrum analysis of the output signal is carried out to obtain the effect law of temperature on the output of the sensor, i. e. , the output drift caused by temperature belongs to the low frequency component. Based on the Kalman filtering noise reduction, a high-pass filtering temperature compensation method based on spectral analysis is proposed to improve the temperature stability by filtering out the low frequency components, and calibration experiments and temperature response characteristics are conducted. The experimental results show that the accuracy of the output voltage measurement is better than ±1. 79% in the temperature range of [0 ℃ , 50 ℃ ], and the method is easy to implement and effectively suppresses the influence of temperature drift when compared with the BP neural network temperature compensation method in the same platform.

Abstract:To address the problems that the existing 1DCNN method for aero-engine fault diagnosis lacks the multi-scale feature extraction capability of fault frequency and the insufficient extraction of time-domain features of the original vibration signal, improved 1DCNN aero-engine fault diagnosis method is proposed by fusing embedded multiscale layers to dual-channel 1DCNN. The method of amplitude change rate is proposed for the time domain feature enhancement of vibration signals, and the amplitude change channel is added as the second channel on the basis of single-channel 1DCNN to build a dual-channel 1DCNN to strengthen the time domain feature extraction capability of 1DCNN, then the multi-scale module is improved to an embedded multi-scale layer and applied to the first channel of 1DCNN to extract multi-scale features of aero-engine fault frequency. Finally, the improved 1DCNN is applied to the diagnosis of aero-engine transient static rubbing, blade fracture and other faults, and the superiority, noise resistance, generalization of the improved 1DCNN detection and the feasibility of the improvement points are proved through comparative experiments.

Abstract:Fault diagnosis methods for rotating machine parts include traditional methods and deep learning, and the former often requires a lot of expert experience and the diagnosis accuracy is poor. A multi-scale attention deep convolutional neural network (MADCNN) is proposed to improve the fault diagnosis method. The MADCNN method provides three convolutional channels, and the principle of differential kernel size of each channel effectively widens the network to achieve multi-scale feature extraction of the original time-domain data. At the same time, CBAM further assigns weights to the extracted features to enhance the differentiation of the model for different types of faults. The accuracy of the validation set was improved by 7. 76% compared with the traditional deep convolutional model by using the bearing failure data from Case Western Reserve University (CWRU) and the planetary gearbox test bench failure data. The experimental results show that the method has high diagnostic accuracy and good generalization performance.

Abstract:For the problem that it is difficult to achieve model-free optimal tracking control in the dynamic system of wheeled mobile robot, a new online integral reinforcement learning control algorithm based on actor-critic framework is proposed in this paper. Firstly, the critic neural network based on RBF is constructed to fit the quadratic tracking control performance index function and the weight updating law of the network is designed based on the approximate Behrman error. Secondly, the RBF actor neural network is constructed to compensate the unknown terms in the dynamic system and the weight updating law is designed to minimize the performance index function. Finally, it is proved by Lyapunov theory that the proposed integral reinforcement learning control algorithm can make the value function, the critic and actor neural network weights error uniformly and finally bounded. Simulation and experimental results show that the algorithm not only realizes the tracking of constant or time-varying velocity, but also can be implemented on the embedded platform.

Abstract:Aiming at the problems of long and unsmooth paths in the path planning of mobile robots traversing multiple target points, this paper proposes a multi-point traversal path planning method based on improved SMA. Firstly, the standard slime mold algorithm (SMA) is improved by combining Singer mapping and small hole imaging reverse learning strategy. Then, the map is preliminarily constructed, and the improved SMA is used to plan the path to determine the optimal value of the maximum side length of the triangular mesh. Finally, the triangular grid map is reconstructed based on the optimal value of the maximum edge length of the triangular mesh, the improved SMA is used to generate the path, and the path is smoothed by the B spline function to improve the smoothness of the path. The benchmark function test results show that the improved SMA converges faster and has higher optimization accuracy. Path planning experiments on triangular grid maps show that the path length and smoothness of improved SMA planning are significantly better than those of SMA, SSA and WOA, and compared with SMA, SSA and WOA, the length of the improved SMA generated path in complex scene is reduced by 6. 31%, 18. 76% and 19. 74%, which verifies the effectiveness of the improved SMA method.

Abstract:A prediction model based on Wasserstein conditional generative adversarial network-gradient penalty ( WCGAN-GP) and convolution long and short-term memory network (ConvLSTM) is proposed to address the problem of unbalanced data caused by the difficulty of collecting fault data during the operating cycle of an aero-engine. First, a WCGAN-GP model is used to learn the deep distribution characteristics of the pre-processed time-series data; then, a generator is used to generate fault samples and mix them with real samples as a training set to input into the prediction model based on the ConvLSTM network for training. Through testing with CMAPSS data set, the results show that compared with the single real sample training prediction model, the performance indexes RMSE and score of the model using mixed data are reduced by 12. 65% and 48. 95% on average.

Abstract:In order to solve the mode-mixing problem of empirical mode decomposition (EMD) and overcome the difficulty of separating the noise components and signal components, a novel denoising method based on secondary complementary ensemble empirical mode decomposition (CEEMD) and time domain feature analysis is presented in this paper. In the proposed method, CEEMD is employed to solve the mode-mixing problem, then the boundary of noise dominant intrinsic mode function (IMF) components and the signal dominant IMF components is determined based on the time domain feature analysis of IMFs returned by CEEMD, whereby the noise components and the signal components are distinguished. Secondary CEEMD decomposition is performed on the noise dominant IMF component and signal dominant IMF component at the boundary to further filter the residual noise while maintain as much useful signal as possible. The experimental results of denoising the actual airborne platform data with impulse noise interference show that the proposed method can better suppress the noise interference and significantly improve the signal-to-noise ratio by effectively separating the noise and signal components.

Abstract:In order to solve the problem that the peak sidelobe of the traditional time delay algorithm is relatively low when it faces impulse noise, and there are many false positives that are difficult to judge, a new weighted Gaussian correlation entropy time delay estimation method is proposed and applied to the simulation model of cable fault location. The simulation results show that, compared with the existing methods, it not only obtains a good time delay estimation effect in the impulsive noise environment, but also maintains a high positioning accuracy in the strong impulsive noise interference. In the background of pulse noise with different intensities, the calculation results show that compared with the other three methods, the absolute value of the main peak to side lobe ratio increases by more than 0. 020 3 dB, the ratio of misjudged peak to fault point peak decreases by more than 0. 053 9, and the mean square error decreases by more than 1. 863 6 m.

Abstract:With the rapid development of industrial technology, large and complex machinery equipment are used for a large number of long periods of time in production, so it is inevitable that there will be wear of parts, and the debris generated by wear will be doped in the lubricating oil, so the detection of debris in the oil is particularly important for the condition monitoring of the equipment. Based on the principle of electromagnetic induction, the use of alternating current drive avoids the heating problem of DC drive sensor. Maxwell software is used to model the internal magnetic field distribution of the oil debris sensor, analyze the influence of various structural parameters on the output signal of the sensor, simulate the core size of the strengthened magnetic field and analyze the magnetic field change inside the sensor, also design the corresponding signal processing circuit and the excitation signal generation circuit. The sensor finally realizes the detection of 200 μm ferromagnetic abrasive grains and 500 μm non-ferromagnetic abrasive grains in the 8 mm flow channel.

Abstract:Inclination sensors are susceptible to measurement errors due to ambient temperature changes, namely temperature drift. Aiming at this problem, a temperature drift compensation model based on the improved genetic algorithm ( IGA) optimized back propagation neural network (BPNN) was designed. A new selection strategy and crossover mutation operator are used in the genetic algorithm, and a mechanism for jumping out of the local optimal solution is added. The experimental results show that the mean square error (MSE) of the IGABP compensation model is 0. 003 28, and the average temperature drift after the compensation model correction is 0. 039°, which is far better than the average temperature drift of 0. 190° without correction. The results show that, the IGABP compensation model has faster convergence speed and higher compensation accuracy compared with the traditional neural network model, which can effectively compensate the measurement error caused by temperature and improve the stability and accuracy of the inclination sensor.