Abstract:Datadriven fault diagnosis (DFD) refers to applications of machine learning and deep learning theories as machine fault diagnosis. This paper systematically summarizes the state of art theories and methods of DFD as well as their applications in spacecraft following the progress of machine learning, then offers a future perspective. DFD could be divided into three categories, traditional machine learningbased methods, deep learningbased methods, and transfer learningbased methods, according to the development of technology. Traditional machine learning based methods adopt advanced signal processing method and feature extraction, which requires a lot of contribution from human labor and extensive expert experience. Although it has excellent performance on small sample data, it is not suitable for processing big data. Over the recent years, the advent of deep learning, which encourages to construct an endtoend diagnosis model, further releases the human labor. Four deep learning models: Stack autoencoder, deep belief network, convolutional neural network, and recurrent neural network are introduced, their applications in diagnosing spacecraft faults are also summarized. Aiming to release the challenge that deep learning relies heavily on labeled data, transfer learning which attempts to use the diagnosis knowledge from one or multiple diagnosis tasks to other related ones, is introduced, and scenarios adapted to spacecraft applications are proposed, picturing a roadmap for the engineering application of DFD.

Abstract:The digital twin of flexible production line can realize the realtime perception of the operation state of flexible production process, and optimize production by using twin data. For this reason, a state aware method of flexible production line based on digital twin is proposed. The architecture of the method is established firstly, and three key technologies in the system implementation are discussed in detail: the construction of digital twin model based on unity 3D, the realtime acquisition of heterogeneous equipment data based on OPC UA, and the state perception and evaluation based on twin model. Finally, based on a flexible production workshop, the state perception and fault diagnosis are realized, and the feasibility and effectiveness of the proposed method are verified, which provides an effective solution for the realization of the state perception of the flexible production line.

Abstract:Aiming at the problem that it is difficult to evaluate the overall health status of the crane rotary system under realtime conditions, a health evaluation method for the rotary system combining Laplacian Eigenmaps and KullbackLeibler distance is proposed. After collecting the multidimensional signal of rotary system, the Laplacian eigenmaps and Random Forest are used to reduce noise and dimensionality of the signal. Then combined with the working principle of the rotary system, the health performance of the rotary system is characterized by Gaussian kernel density estimation. The KullbackLeibler distance between different rotary system is calculated by probability density to characterize the health performance of the rotary system. The test results show that this method can avoid the noise interference of the original data and the health assessment results of the rotary system are consistent with the expert assessment results.

Abstract:Intelligent fault diagnosis technology can effectively guarantee the safe operation of mechanical equipment. Traditional bearing fault diagnosis generally assumes that the labeled source and unlabeled target domain data follow the same distribution. However, the conditional and marginal distributions of bearing data usually do not satisfy the same distribution assumption in actual diagnosis scenarios. Moreover, feature distortions are difficult to eliminate when performing adaptive distribution alignment in the original Euclidean space, which affects the fault diagnosis performance. In this paper, an adaptive bearing fault diagnosis model based on manifold feature learning and dynamic distribution alignment is proposed to address these challenges. First, we construct a geodesic flow kernel in the Grassmann manifold and extract the inherent manifold feature representation associated with the bearing fault information, avoiding data feature distortions. Second, a crossdomain adaptive factor is defined by distance to dynamically evaluate the conditional and marginal distributions of manifold features. Finally, a crossdomain classifier is solved iteratively to predict the target domain samples under the principle of structural risk minimization. The experimental analysis of multiple indicators shows that the model can effectively avoid feature distortions and use dynamic weights to adjust the relative importance of conditional and marginal distributions between crossdomain data, which verifies the effectiveness of the proposed method.

Abstract:In order to improve the decomposition ability and decomposition accuracy of complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN）, and solve the problem of noise residual in CEEMDAN, an improved CEEMDAN method called meanoptimized ensemble empirical mode decomposition with adaptive noise (MEEMDAN) is proposed. MEEMDAN introduces different weights in the process of iteration screening. Based on the minimum orthogonality, the optimal IMF is selected from the decomposition results under different weights as final decomposition result to ensure that the IMFs of each order are globally optimal. The simulation results show that MEEMDAN is superior to CEEMDAN in decomposition ability and accuracy. At the same time, a new fault diagnosis method for rolling bearings combining MEEMDAN with maximum correlation kurtosis deconvolution (MCKD) is proposed and applied to the simulation and measured data analysis. The results show that, compared with the existing methods, the proposed method can extract fault characteristic frequency more accurately, and has more advantages in decomposition ability and interference suppression frequency.

Abstract:Local characteristicscale decomposition (LCD) has improved the empirical mode decomposition (EMD) method, but it also inherits the mode mixing of EMD. Noise assisted analysis is one of the important methods to solve the mode mixing, however, LCD is more sensitive to noise. If we adopt the white noise used in ensemble empirical mode decomposition (EEMD) as an auxiliary signal, the mode mixing cannot be effectively resolved and more false components will produce. An improved LCD method termed uniform phase local characteristicscale decomposition (UPLCD) is proposed to solve the above problem. UPLCD uses narrow wave signal with uniform phase as auxiliary signal, which can suppress the mode mixing and avoid the increase of false components caused by white noise. After the validity of UPLCD is verified by the simulation signal analysis, the proposed method is applied to the mechanical fault diagnosis by comparing it with EEMD, LCD and uniform phase empirical mode (UPEMD). The results show that the proposed UPLCD can effectively decompose the modes of rotating machinery signals and has more advantages in decomposition accuracy and interference suppression than the above three methods.

Abstract:Based on the research background of the CNC spindle bearing, this paper proposes a fault diagnosis method combining wavelet packet mixing feature and support vector machine（SVM）, aiming at the problem that bearing fault information is complex and difficult to obtain and fault data samples are few. First, carry out wavelet packet decomposition and reconstruction of the bearing vibration signal, and extract the mixed features of the signal to construct a joint feature space. Then use tdistributed stochastic neighbor embedding (tSNE) method to observe the distribution of sample data and observe the data distribution of the mixed feature sample set. Finally, a nonlinear SVM is used for fault classification. The Experimental results show that the accuracy is 100% for the fault identification of the spindle bearing inner ring, outer ring and ball. Compared with the fault classification effect of linear SVM and BP neural network, this method has achieved good results in the application of fault diagnosis of spindle bearing of CNC.

Abstract:Aiming at the problem that the vibration signal of the check valve is easily overwhelmed by noise and the fault characterization is not obvious. A fault diagnosis method for the check valve based on the total variation denoising and recurrence quantification analysis (TVDRQA) was proposed. Firstly, the total variation denoising method was used to denoise vibration signals and improve their signaltonoise ratios; Then, draw a recurrence plot on the denoised signal, extract the nonlinear characteristic parameters in the recurrence plot through the recurrence quantification analysis method, and perform sensitivity analysis on the extracted feature parameters to find out the feature vectors with higher sensitivity; Finally, the obtained feature vector is input into the weighted Knearest neighbor classifier (WKNN) to complete the identification of the check valve failure type. Experimental results show that the method has obvious effects in reduce background noise, digging fault information, and ensuring the accuracy of fault diagnosis, and has certain engineering application value.

Abstract:Aiming at the problem that rolling bearing fault signals are easily disturbed by environmental noise and it is relatively difficult to obtain fault feature information, a rolling bearing fault feature extraction method based on VMD and fast spectral kurtosis is proposed. First, the bearing signal is decomposed into several IMF components, and then the maximum correlation kurtosis deconvolution algorithm is used to calculate the modal components of each order, and several IMF components with relatively large correlation kurtosis values are selected as the most prominent study of the fault information object and perform fast spectral kurtosis analysis on it; finally, set the filter frequency range according to the results of the fast spectral kurtosis map, and perform square envelope spectrum analysis on the filtered signal to obtain the fault characteristic information of the bearing. Public data and experimental analysis show that this method can successfully diagnose bearing faults.

Abstract:Bearings are one of the important components in the rotating machinery; it is significant to assess its performance degradation and predict the remaining useful life of the bearings by using sensors data to improve the reliability and decrease the maintenance costs. For the traditional datadriven approaches that rely on prior knowledge or expert experience in feature extraction and do not fully model middle and longterm dependencies hidden in timeseries data, we propose an endtoend deep framework for bearing performance degradation prognosis based on convolutional neural network（CNN）and bidirectional long shortterm memory（BLSTM）. The model adopts threelayer structure, the neural network firstly uses CNN to extract feature vectors directly from raw sensor data, then, the feature vector is constructed in a time series manner, and the BLSTM layer is introduced to capture temporal feature, finally, fullyconnected layers and the linear regression layer are built on top of BLSTM to predict the target value. The results of bearing accelerated life experiments show that the RMSE and MAPE of the proposed method are 127% and 171% lower than the traditional methods, indicating that the method can effectively improve the prediction accuracy of bearing performance degradation.

Abstract:With the wide application of highend equipment in the industrial field, its operating state has a great impact on the safety of equipment and the performance of products, sudden failures often cause huge loss of people's lives and property and affect the safety and stability of society. The electromechanical system is in the state of variable speed operation, and its state characteristic information is difficult to obtain, which makes it difficult to diagnose and predict the fault of the electromechanical system. In view of this problem, a fault classification, recognition and diagnosis model of electromechanical system based on deep learning is proposed. Firstly, the vibration signals of the key parts are converted into timefrequency graphs by timefrequency transformation to form the input samples；Secondly, the samples were input into the deep learning neural network for feature learning and state recognition, the method of combining different transformations and deep learning convolutional neural networks is studied, which is applied to a mechanical and electrical system test bench to compare the fault state classification performance. The experimental results show that this method provides a new way for the fault diagnosis of electromechanical system.

Abstract:In order to analyze the modulation characteristics of vibration response signals caused by different rotorstator rubimpact faults, based on the Jeffcott rotor model, this paper utilized the variational mode decomposition and Hilbert transform to extract and analyze the intrawave modulation characteristics of different rub impact fault simulation signals, and reveals the relationship between the dynamic characteristics of rubimpact rotor and the intrawave modulation characteristics of fault signals. The simulation results show that the periodk rubimpact fault can cause the vibration response signal in the medium and low frequency band to oscillate its instantaneous frequency at the center of 1/k octave, and the vibration frequency is still 1/k octave, which results in the 1/k side frequency band in the original frequency spectrum. In the quasiperiod rubimpact fault, the internal frequency modulation is irrational octave and integer octave close to 1/k octave, which results in the irrational number side band in the original signal spectrum. The rotor fault diagnosis test of centrifugal pump showed that the intrawave modulation characteristics can effectively be used to identify the severity degree of rotor rubimpact fault.

Abstract:FPGA is a programmable logic controller, which not only has the operating speed and stability of hardware circuits, but also has the programmability and flexibility of software. FPGA can easily perform multichannel highspeed parallel fast Fourier transform. Therefore, FPGA becomes more and more popular in online condition monitoring, vibration signal processing and edge computing. In order to make full use of the computation power of FPGA, as well as improve the accuracy in signal processing, this paper uses a rotor system and FPGAbased vibration monitoring device to experimentally study the key steps that affect the accuracy of the vibration signal processing, in order to find the optimal parameters for filtering, digital integration, and peak detection. The results show that the integration cutoff frequency greatly affects integrated velocity. The cutoff frequency should be set reasonably according to the speed of the equipment to improve the acquired RMS value. Implementing a bandpass 4th order Butterworth filter can eliminate the influence of high frequency noise and improve the stability of vibration peaks. Calculated peaks from an analog capacitor charge and discharge algorithm is stable, which is suitable for calculating the original signal peaks with “burrs”. It can be used for vibration protection. The onesecond peaktopeak value is very sensitive, which is suitable for early warning/fault detection.

Abstract:This paper presents a method of bearing remaining useful life prediction based on KPCA and long short memory neural network (LSTM) with dropout. Firstly, 14 timedomain indexes of vibration signal, such as effective value, maximum value, peakpeak value and kurtosis, are extracted. KPCA method is used to fuse the timedomain characteristic indexes of bearing vibration signal and extract the first principal component to evaluate the degradation of bearing performance, and multiple principal components that meet the requirements are used as the input of the prediction model. Then an improved LSTM prediction model based on dropout strategy is established finally, the bearing data are used to verify the proposed method. The results show that the proposed method can effectively predict the remaining useful life of the bearing, and has a good prediction effect. The prediction accuracy reaches 9592%.

Abstract:The Flow excitation fault is a common fault of gas turbine due to the working medium. Aiming at the flow excitation fault of the gas turbine, deep belief network (DBN) model is established to realize fault diagnosis based on the peak hold down sampling (PHDS) algorithm and particle swarm optimization (PSO) algorithm. The vibration data which is compressed by the PHDS algorithm is used as the input of the DBN to reduce the training time of the model. The PSO algorithm is adopted to optimize the structure parameters of the DBN to find the model with the best diagnostic effect. The results of example show that the optimized model not only reduces the training time of the model, realizes the intelligent optimization of network structure parameters, but also diagnoses the flow excitation faults of gas turbine effectively and the accuracy of the test was about 998%.

Abstract:This paper aims at the problem of internal covariate displacement in the training process of extracting transferable features based on the convolutional neural network (CNN) domain adaptive technology. A multidomain adaptive rolling bearing fault diagnosis method is proposed. First, use CNN to extract the migratable features of the original vibration data; Secondly, multilayer domain adaptation and weight regularization terms are used to constrain CNN parameters to further reduce the distribution difference of migratable features, thereby solving the problem of domain shift; Finally, the rolling bearing data set of Case Western Reserve University was used for experimental verification. The results show that this method can effectively reduce the difference in feature distribution between the source domain and the target domain, and improve the diagnostic performance of the CNN model on the target domain dataset, compared with the adaptive fault diagnosis method at the highest level, the proposed method can achieve higher classification and recognition results in the fault diagnosis of migration between two data sets.

Abstract:The healthy condition of the drive motors bearing is an important premise to realize the reliable operation of the pure electric vehicles and avoiding safety accidents.Due to the lack of the state monitoring methods of the rolling bearing,a new method based on sparse autoencoder (SAE)and support vector machine (SVM) for rolling bearing of pure electric vehicles condition monitoring is proposed. In terms of feature extraction, the time domain, frequencydomain and timefrequencydomain feature sets of rolling bearing vibration signals are used to construct highdimensional data sets, and the data fusion with multilayer SAE is performed to eliminate feature redundancy, which obtains more robust concise features.In terms of condition monitoring,the characteristic representation of bearing conditionis input into SVM for training to obtain a bearing condition monitoring model. Finally, the effectiveness of the method is evaluated by designing a bearing of pure electric vehicle motor condition experiment.The results show that comparing with the traditional feature + SVM, the monitoring method of rolling bearings of pure electric vehicles based on SAESVM is more accurate and reliable.

Abstract:Vibroseis is an important equipment in the field of geophysical exploration. Its vibration isolation performance in low frequency band is not ideal, which affects the exploration effect to a certain extent. In this paper, KZ28 vibroseis vehicle is taken as the research object, and the negative stiffness vibration isolation system is established. By reducing the stiffness, the isolation efficiency of the original vibration isolation system in low frequency band is improved. The simulation results show that the system can be applied to the vibration isolation of vibroseis vehicle. The negative stiffness system is introduced into the vibroseis vehicle model to test the vibration isolation effect. The results show that the negative stiffness vibration isolation system can improve the isolation efficiency by more than 15%. It is feasible to apply the negative stiffness vibration isolation system to the vibroseis vehicle, which lays the foundation for the followup engineering application.

Abstract:A new type of ceramicbased porous composite material is used to bond structural members. When the ultrasonic Cscan system is used to detect the debonding defects of the components, the ultrasonic data produces strong acoustic noise due to the porosity and anisotropy of the material, resulting in blurred edges on reconstructed images. Aiming at this problem, this paper proposes to use the method of wavelet packet timefrequency domain analysis and energy spectrum analysis to obtain the spectrum and energy spectrum characteristics of ultrasonic echo signals. In order to enhance the characteristics of the echo signal, an adaptive wavelet packet threshold algorithm is used for noise reduction. Compared with the traditional ultrasonic denoising method, the signaltonoise ratio is increased by nearly 20%, which effectively reduces the interference of the noise signal to the original ultrasonic signal. On this basis, the image reconstruction of the tiny adhesion defects on the near surface of the ceramic substrate is performed, which achieved a clear characterization of debonding defect with a diameter of 2 mm. The results show that the effectiveness of the method proposed in this paper, the quality of the reconstructed image is greatly improved, and the edges of small defects are more clearly reflected, thereby improving the accuracy of quantitative representation.

Abstract:Aiming at the common uncertainty in the diagnosis and testing of practical complex systems, an optimal selection method of test points based on artificial immune clone selection algorithm (AICS) is proposed under the unreliable condition. In this model, a fitness function reflecting the performance of the test points is constructed by comprehensively considering the performance indexes such as fault detection rate, isolation rate, false alarm rate and total test cost, and an unreliable test point optimization scheme is designed based on AICS, which effectively reduces the complexity computing. As a result, the time cost is reduced to 0496 seconds, which demonstrates the improvement efficiency of proposed model. Finally, this model is verified by a test utilized with the consumption component in the fuel consumption measurement system. The results show that this method can obtain a set of test points with the lowest test cost while meeting the performance requirements of fault detection rate, isolation rate, false alarm rate, and its comprehensive performance index is better than that of genetic algorithm and simulated annealing particle swarm optimization algorithm.

Abstract:Aiming at the deviation of the recognition accuracy for pipeline system blocking recognition model under complicated working conditions, a method is proposed for recognizing blockage and lateral connection in pipeline individually based on timefrequency image and convolution neural network algorithm. Firstly, acoustic wave detection method is used to obtain lowfrequency sound pressure signals under different working conditions, smooth pseudo WignerVille timefrequency analysis method is performed to the filtered signal to obtain the timefrequency distribution map. Then, the Otsu threshold segmentation method is applied to adaptively segment the timefrequency distribution images to obtain timefrequency images of blockages and lateral connection under single and complicated working conditions. At last, the timefrequency images of light blockage, heavy blockage, lateral connection and pipe end under a single working condition are entered into the CNNSVM model for training, the trained parameter model is applied to the automatic recognition of blockages and pipe components under complicated working conditions. The experimental results show that the recognition rate of the proposed method for four kinds of targets under complicated conditions is over 96%, and the recognition accuracy is higher than that of the traditional artificial feature extraction model, which verified that the influence of the blockage on the acoustic wave under different working conditions is common and different from that of the lateral connection. individual analysis of different degrees of blockage and lateral connection under complicated working conditions individually, can effectively avoid the deviation of model recognition accuracy owing to the difference of working condition distribution.

Abstract:Due to the attenuation and scattering of light in the water, the image captured underwater has the problems of color deviation, low contrast, poor definition and uneven illumination. In this paper, an underwater image enhancement method based on color correction and improved 2D gamma function is proposed. Firstly, MSRCR is used to correct the color deviation problem to obtain an input image; secondly, the improved twodimensional gamma function is used to reduce the influence of uneven illumination on the underwater image, and BEASF is used to enhance the image contrast to obtain another input image; finally, the four weights of contrast, saliency, saturation and exposure are combined for multiweight fusion to obtain the final enhancement image. Experimental results show that this algorithm can effectively improve the problem of underwater image color deviation, and enhance the details and contrast of the image.

Abstract:The complexity of the underwater environment poses a challenge to underwater acoustic communication, the high rate and reliable transmission of underwater communication is achieved by optimizing channel estimation technology. In this paper, the filter bank multicarrier/offset quadrature amplitude modulation (FBMC/OQAM) modulation technology is introduced underwater, the channel estimation technology in the FBMC/OQAM system is studied, and the discrete pilot channel estimation algorithm is studied in detail. This paper analyzes the performance of discrete pilot channel estimation under the influence of different channels, and improves the problem of excessive power consumption of the auxiliary pilot method. An improved auxiliary pilot method (IAP) is proposed to reduce the power consumption, it also improves the estimated performance. Theoretical analysis and simulation results show that the IAP algorithm proposed in this paper is a good choice for the practical application of underwater acoustic communication.

Abstract:In order to improve the accuracy of indoor positioning and reduce the cost of onsite investigation, an indoor localization method based on Gaussian process regression and WiFi fingerprint is proposed. Firstly, during the offline phase, Gaussian process regression model (GPR) is used to expand the WiFi fingerprint database. In other words, by training different GPR kernel functions, the best GPR prediction model is obtained, and then the signal strength (RSS) of unknown region is predicted by using the limited known data. Then in the fingerprint matching stage, the weighted nearest neighbor algorithm (WKNN), maximum likelihood estimation (MLE) and multilayer perceptron (MLP) methods are used to locate the unknown points according to the RSS database obtained in the offline phase. Specifically, in order to further improve the positioning accuracy, an error correction model is proposed and applied to the above different positioning algorithms. The experimental results show that the kernel function combination of kRBF+kMatern+kRQ is the best GPR prediction model, and the average RSS estimation error is 459 dBm. In addition, compared with the results which is obtained by using the original survey map, the location algorithm based on GPR has higher positioning accuracy. Among them, the GPRWKNN algorithm has the highest positioning accuracy, with 80% positioning error of 132 m. The above results indicate that the method of using GPR to expand the map and further predict the location is accurate and effective, and can meet the high requirements of positioning accuracy in new application scenarios such as commodity recommendation and material dynamic management, emergency rescue, intelligent parking, infectious disease tracking and so on.

Abstract:With the increase using of millimeterwave radar in advanced driverassistance systems (ADAS), there will be a trend of largescale application in the next few years, and therefore the probability of automobile millimeterwave radar interfering with each other on the same traffic road increases. To this end, this paper proposes an interference reduction method for automotive millimeterwave radar based on Wiener filtering. It is known that Wiener filtering has excellent suppression performance on conventional stationary noise interference signals, however, radar interference signals are nonstationary and nonGaussian. In order to make the classic Wiener filter suitable for interference suppression of automobile millimeter wave radar, this paper first makes statistics on the noise floor level of the echo signal, and then distinguishes the interference part of the echo signal from the noninterference part; in the distance to Fourier transform In the domain, a shortlength window is used to perform Wiener filtering on the radar echo signal containing interference, and the filter coefficients are adaptively and dynamically updated to suppress the interference echo. Based on the existing radar system parameters, the simulation experiment of 77 GHz automotive millimeter wave radar and the actual measurement experiment of radar hardware interference were completed. Experimental results show that the method in this paper can effectively suppress the interfering signal, and successfully recover the target drowned in the interfering signal, and the signaltointerference ratio of the target is improved by 91 dB.

Abstract:Aiming at the poor performance of timedelay estimator (TDE) under low signaltonoise ratio (SNR) of leakageinduced vibration signal in watersupply pipelines, the performance of six TDE (crosscorrelation, ROTH, SCOT, WIENER, PHAT and ML) for watersupply pipelines leakageinduced vibration signal was studied by combining simulation analysis with experimental verification. Firstly, the variation rule of six TDE with SNR is studied under different noise types. Then, the noise suppression performance of the TDE under different SNR is analyzed. Finally, six TDE are applied in leak location experiments of watersupply pipelines to analyze the relative errors of leak location caused by different TDE to compare their noise suppression performance. Simulation analysis and experimental verification show, the SCOT has the best performance in noise suppression and the average relative error of SCOT based leak location in watersupply pipelines is 348%, the Wiener has the worst performance and the average relative leak location error is 1823%.

Abstract:Aiming at the poor performance of timedelay estimator (TDE) under low signaltonoise ratio (SNR) of leakageinduced vibration signal in watersupply pipelines, the performance of six TDE (crosscorrelation, ROTH, SCOT, WIENER, PHAT and ML) for watersupply pipelines leakageinduced vibration signal was studied by combining simulation analysis with experimental verification. Firstly, the variation rule of six TDE with SNR is studied under different noise types. Then, the noise suppression performance of the TDE under different SNR is analyzed. Finally, six TDE are applied in leak location experiments of watersupply pipelines to analyze the relative errors of leak location caused by different TDE to compare their noise suppression performance. Simulation analysis and experimental verification show, the SCOT has the best performance in noise suppression and the average relative error of SCOT based leak location in watersupply pipelines is 348%, the Wiener has the worst performance and the average relative leak location error is 1823%.

Abstract:Visualized nondestructive testing (NDT), with the development of deep learning technology, is leading a huge opportunity in data processing. However, obtaining sufficient labeled data sets is a big challenge. The realization of the expansion of the nondestructive inspection image data set is conducive to improving the ability of deep learning in defect detection. Therefore, this article has effectively expanded the existing data by studying the characteristics of nondestructive testing image data, combined with CycleGANs (CycleGANs) method. Improved the deep convolutional neural network (DCNN) to effectively use the expanded data to improve the ability to recognize defective images. Finally, through comparative experiments, it is shown that this expanded data has an important role in improving the training of the defect detection network.

Abstract:Visualized nondestructive testing (NDT), with the development of deep learning technology, is leading a huge opportunity in data processing. However, obtaining sufficient labeled data sets is a big challenge. The realization of the expansion of the nondestructive inspection image data set is conducive to improving the ability of deep learning in defect detection. Therefore, this article has effectively expanded the existing data by studying the characteristics of nondestructive testing image data, combined with CycleGANs (CycleGANs) method. Improved the deep convolutional neural network (DCNN) to effectively use the expanded data to improve the ability to recognize defective images. Finally, through comparative experiments, it is shown that this expanded data has an important role in improving the training of the defect detection network.