64results about How to "Good noise robustness" patented technology

The invention discloses a radar one-dimensional range profile target recognition method based on a depth convolution neural network, includes the following steps: a data set is collected, the collected data is preprocessed, features are extracted from the preprocessed data, the HRRP signal is divided into two parts: low SNR and high SNR, A feature enhancement algorithm based on robust Boltzmann isconstructed, and a HRRP target recognition model based on convolution neural network and bidirectional loop neural network based on LSTM is constructed. The parameters of the network model are fine-tuned by using gradient descent algorithm, and an effective target recognition model is obtained. A radar HRRP automatic target recognition technology with small sample robustness and noise robustnessconstructed by the invention has strong engineering practicability, and a radar one-dimensional range profile target recognition model based on a convolution neural network and a cyclic neural networkis proposed from the aspects of feature extraction and the design of a classifier.

The invention relates to a rolling bearing fault diagnosis method and system based on particle swarm optimization, and a medium. The method comprises: obtaining an initial vibration signal of a rolling bearing in the operation process, and obtaining an initialization parameter of the initial vibration signal by adopting a particle swarm optimization method; performing variation mode decompositionon the initial vibration signal according to the initialization parameters to obtain a plurality of intrinsic mode components; selecting one inherent mode component containing most fault characteristic information from the plurality of inherent mode components as a most sensitive mode component, and carrying out band-pass filtering processing on the most sensitive mode component to obtain a faultcharacteristic vibration signal; and analyzing the fault characteristic vibration signal, extracting the fault characteristic information, and identifying the fault according to the fault characteristic information. The fault diagnosis accuracy and the fault identification precision of the rolling bearing can be effectively improved, the method can be widely applied to the technical field of signal fault diagnosis, and normal operation of mechanical equipment is guaranteed.

The invention relates to an ecological voice recognition method based on multiband signal reconstruction. The ecological voice recognition method comprises the steps of: firstly, using OMP (Orthogonal Matching Pursuit) sparse decomposition as a first-stage reconstruction, and reserving a main body structure of foreground voice; secondly, allocating remained components decomposed in the former stage according to bands, and carrying out adaptive compensation on reconstruction signals according to the frequency distribution of the foreground voice and background noise to complete a second-stage reconstruction; finally, extracting compound noise-proof characteristics according to atom time-frequency information and frequency-domain information in a support set, and carrying out classification and recognition on ecologic voice by using a high-credibility network under different environments and signal to noise ratio conditions. According to the ecological voice recognition method, the noise can be inhibited by adopting two times of reconstruction, and the reconstruction precision of the foreground voice is improved; better noise robustness is achieved under a natural environment.

The invention discloses a phase retrieval based 4f mirror surface detection imaging system and a phase retrieval based 4f mirror surface detection imaging method. The system comprises a laser, a neutral density filter, a microobjective, a pinhole, a measured mirror surface, a 4f imaging unit and a computer, wherein the 4f imaging unit comprises a lens 1, a space light modulator, a lens 2 and a charge coupled device (CCD) camera. The light emitted by the laser irradiates the measured mirror surface after passing through the neutral density filter, the microobjective and the pinhole. The CCD camera arranged in the 4f imaging unit is used for acquiring a plurality of times of a light wave modulation image, and then the image is sent into the computer for sparse constraint phase recovery treatment. Based on the acquired light wave intensity image of the measured mirror surface, the method utilizes the sparse constraint phase recovery treatment to obtain the phase position of the light wave on the measured mirror surface, thus realizing the error detection for the measured mirror surface. The invention has the advantages of being high in accuracy, good in stability, simple in operation and good in noise robustness.

The invention discloses an isolated digit speech recognition classification system and method combining a principal component analysis (PCA) with a restricted Boltzmann machine (RBM). First of all, a Mel frequency cepstrum coefficient (MFCC) is employed for combination with a one-order difference MFCC, and a voice dynamic characteristic of an isolated digit is preliminarily drawn off; then, linear dimension reduction processing is carried out on an MFCC combination characteristic by use of the PCA, and dimensions of a newly obtained characteristic are unified; accordingly, nonlinear dimension reduction processing is performed on the obtained new characteristic by use of the RBM; and finally, finishing recognition classification on a digit voice characteristic after nonlinear dimension reduction by use of a Softmax classifier. According to the invention, PCA linear dimension reduction, unification of the dimensions of the characteristic and RBM nonlinear dimension reduction are combined together, such that the characteristic representation and classification capabilities of a model are greatly improved, the isolated digit voice recognition correct rate is improved, and an efficient solution is provided for high-accuracy recognition of isolated digit voice.

The invention discloses a low signal-to-noise ratio ISAR envelope alignment method based on piecewise linear movement estimation. The method mainly solves the problem that in the prior art, ISAR echo envelopes with a low signal-to-noise ratio can not be effectively or quickly aligned. According to the implementation scheme, envelope errors of target echoes are modeled in a high-order polynomial form based on the continuity of target movement, a full-aperture time is divided into multiple periods of sub-aperture time, and envelope errors in each period of sub-aperture time are approximated to be linear. Linear coefficient estimation is carried out on each sub-aperture through a traditional envelope alignment algorithm, and estimation on envelope errors of a full aperture is achieved through the least square method based on linear coefficient estimation results of the sub-apertures. According to the method, efficient and accurate alignment of the target echo envelopes can be achieved under the condition of a low signal-to-noise ratio, and the method can be used in the target detection and recognition and ISAR imaging fields.

The invention discloses a high-voltage power equipment partial discharge feature extraction method and apparatus. According to the method, after sample signals of partial discharge of high-voltage power equipment are obtained, multiple modal components of the sample signals are obtained by performing VMD on the sample signals. Accordingly, effective modal components are obtained by performing denoising on the modal components on the basis of an ICA algorithm and a threshold denoising method, and features of the partial discharge are represented by use of the feature quantity PMSE of the effective modal components. The VMD can help to realize adaptive decomposition of signals, such disadvantages of modal mixing and sampling frequency influences and the like do not exist, components of close frequencies can be well distinguished and separated, the noise robustness is better, an optimal decomposition effect can be realized, and the precision of feature extraction of partial discharge signals is improved.

The invention discloses a multi-channel chaotic synchronization communication system in a simple topological structure. The output of a semiconductor laser which is operated in a chaotic state is uniformly divided into two parts, and the two parts are injected into two mutual injection lasers which are connected by n different-time-delay bidirectional optical fiber branches respectively through two same unidirectional optical fiber branches; under an injection-locking effect of unidirectional injection, stable and high-quality chaotic synchronization can be realized between the two multiple-time-delay mutual injection lasers; and accordingly, each mutual injection laser can synchronously send n pieces of information and receive n pieces of information, so that the multi-channel chaotic synchronization communication is realized. By the multi-channel chaotic synchronization communication system, the chaotic synchronization is extremely easily realized in a multiple-time-delay mutual injection system in a simple structure, and the problem that the injection and feedback are required to conform to a strict mathematical relation for realizing the synchronization in the conventional multiple-time-delay mutual injection system is solved. Moreover, the multi-channel chaotic synchronization communication system supports multi-information simultaneous bidirectional transmission between the two lasers and has a guiding significance for chaotic network communication.

The invention discloses a heartbeat and breathing feature monitoring method based on an ultra wide band radar sensor and relates to the technical field of heartbeat signal detection. The method comprises following steps of using detected breathing signals and detected heartbeat signals as echo signals; preprocessing the echo signals and selecting a distance door to obtain a one-dimensional signal;using a classic mode decomposition algorithm to carry out modal decomposition on the obtained one-dimensional signal so as to obtain the modal number K and K specific modals; according to the modal number K and the K specific modals, carrying out modal decomposition on the one-dimensional signal in the step 2 through a variation modal decomposition algorithm, and extracting breathing modals and heartbeat modals; and carrying out frequency spectrum conversion on the extracted breathing modals and the heartbeat modals to obtain respective frequency domain information and through the frequency domain information, directly obtaining the breathing times and the heartbeat times of each target in every minute. According to the invention, problems of poor anti-interference ability, low accuracy and modal aliasing in the current method are solved.

The invention discloses a weight window self-adaptation non-local mean image denoising method. According to the weight window self-adaptation non-local mean image denoising method, the sizes of weight windows can be controlled in a self-adaptation mode according to image local structure characteristics, noise is suppressed while an edge structure is protected, and therefore the image quality is remarkably improved. The method includes the following steps that first, a frame of noise image is initialized and read in; second, a structure tensor matrix is built; third, according to the built structure tensor matrix, edge structure indicators are built, and the characteristics of the area where pixel dots are located are positioned; fourth, the areas of the image are classified through the edge structure indicators; fifth, according to the type of the area to which each pixel dot belongs, the size of the adjacent area of each pixel dot is determined; sixth, according to the determined size of the adjacent area of each pixel dot, a similarity metric function between the adjacent areas is built; seventh, S dots with highest similarity are screened; eighth, a denoising model is built, and a denoised image is acquired.

A method for analyze time-frequency of high-order synchronous extracted transform signal includes such steps as expanding high-order Taylor series, carrying out short-time Fourier transform, obtaininglocal instantaneous frequency, obtaining high-order instantaneous frequency, and obtaining high-order synchronous extracted operator, carrying out high-order Taylor series expansion, carrying out short-time Fourier transform, and obtaining high-order synchronous extracted operator. Finally, taking the frequency set after the short-time Fourier transform as the center frequency set, combined withthe high-order synchronous extraction operator, it 'extracts' the value of each time-frequency point corresponding to the instantaneous frequency in the vicinity of each center frequency, that is, only the energy near the time-frequency ridge in the time-frequency plane is reserved, and all the other divergent energy is eliminated to obtain the high-order synchronous extraction transform value. High-order synchronous extraction transform not only inherits the advantages of high time-frequency focusing and good noise robustness of synchronous extraction transform, but also has higher time-frequency precision than short-time Fourier transform and synchronous extraction transform, so it is a new time-frequency analysis method with high precision.

The invention discloses a speech recognition method based on a domain-invariant feature, and the method applies a speech domain-invariant feature extraction model to an end-to-end speech recognition model. The feature extraction model used in the speech recognition method aims at the robustness problem, and by adding more types of speech data to train the speech feature extraction model, better parameters can be obtained and a better domain-invariant feature extraction model can be obtained. The speech recognition method based on the domain-invariant feature uses unlabeled pure speech data totrain the feature extraction model, uses a small number of speeches with text annotation to train an end-to-end acoustic model, and provides important technical support for improving the robustness ofthe end-to-end acoustic model. Compared with the prior art, the speech recognition method based on the domain-invariant feature has higher recognition accuracy in different noise environments, smaller task quantity of speech annotation tasks, and faster training and testing speed of the models.

The invention relates to a de-noising method for an electric vibration signal of a wind driven generator and a storage medium. The method comprises the following steps: decomposing the vibration signal of the wind driven generator into a group of modal components through a variational modal decomposition algorithm; respectively calculating a multi-scale permutation entropy value of each modal component; judging whether the multi-scale permutation entropy value of each modal component exceeds a preset entropy value or not; if the multi-scale permutation entropy exceeds the preset entropy, performing noise reduction processing on the modal components with the multi-scale permutation entropy exceeding the preset entropy; and combining and reconstructing the modal component subjected to the noise reduction processing and the modal component not subjected to the noise reduction processing to obtain a vibration signal subjected to the noise reduction processing. Vibration signals are denoised through combination of variational mode decomposition and multi-scale permutation entropy, signal distortion can be effectively reduced, and the denoising effect is good.

The invention discloses a carbon emission price combination prediction method. The method includes the steps that 1, an original carbon emission price sequence is decomposed into a series of intrinsic function components through a variation mode decomposition algorithm; 2, an output variable is given, the input variable of each IMF component is determined through statistical tools comprising a partial autocorrelation function and a corresponding partial autocorrelation diagram thereof; 3, each IMF component is predicted through a Spiking neural network; 4, prediction results of all the IMF components are superimposed to obtain a predicted value corresponding to an original carbon emission price. By means of the method, prediction precision is effectively improved, and carbon emission price prediction can be well achieved.

The embodiment of the invention discloses a voiceprint recognition method and device based on empirical mode decomposition and the mel frequency cepstrum coefficient (MFCC), and relates to the technical field of speech signal processing and computers. According to the method, the instantaneous feature parameter of a speech signal after empirical mode decomposition is extracted and fused with the traditional Meier feature parameter to form an improved feature parameter for voiceprint recognition. The device includes a data acquisition module, a high-speed data transmission module, an algorithmimplementation module, a data storage module and an user interface module. The device and the method can improve the efficiency and accuracy of identity authentication, the noise robustness of a system is improved, the response time is shortened, and meanwhile better user experience is brought.

The invention discloses a game theory-based MIMO channel tracking method, which comprises the following steps: transmitting a pilot signal, performing channel estimation through the pilot signal and taking a channel estimation value as an initial value of tracking; modeling MIMO channel noise serving as a game opponent; resolving a tracking error and determining a tracking target function; and solving saddle points of the target function through the game theory to realize channel tracking. The method combines the game theory and the MIMO channel tracking method, thereby having high tracking accuracy.

The invention discloses a video time-space super-resolution reconstruction method based on a robust optical flow and Zernike invariant moment. The method comprises the following steps: performing motion analysis on a video sequence in a time-space domain, constructing a robust optical flow motion estimation model of the video sequence, and obtaining a motion vector; according to the motion vector, performing bidirectional time-space motion compensation on the video sequence to obtain a compensated video sequence; and by use of a cross-scale fusion strategy of a Zernike invariant moment based rapid non-local fuzzy registering mechanism, performing time-space super-resolution reconstruction on the compensated video sequence to obtain a video sequence with high time-space resolution. According to the invention, the method is not dependent on accurate sub-pixel motion estimation, can be applied to various complex motion modes, such as angle rotation, local motion and the like, can provide clear and smooth video information for accurate identification and tracking of a motion object, and has a quite high actual application value.

The invention belongs to the data recognition and data representation technical field and discloses a local invariant gray feature-based image registration method and an image processing system. According to the local invariant gray feature-based image registration method, feature extraction descriptors are constructed; feature points between registration images are searched, the nearest neighborprinciple is used to find matched key points; and an affine transformation matrix H between the registration images is calculated, and six parameters of the affine transformation matrix H are obtainedthrough singular value decomposition. The descriptors are constructed; sampling points are divided into an odd part and an even part, and therefore, dimensionality during the construction of the descriptors is significantly lowered, operating time is reduced, the accuracy and accuracy of registration are improved; when being constructed, descriptor vectors are sequenced according to gray values,and therefore, rotation invariance can be realized. The method of the invention has high detection precision, good noise robustness and low computational complexity, which mainly benefits from the great reduction of the dimensionality of the original descriptors and insensitiveness to illumination transformation.

The invention discloses a method for an underdetermined DOA estimation based on a partially calibrated and nested array under an amplitude and phase error. The method comprises the following steps of:S1, selecting the array output data to construct a fourth-order cumulant vector; S2, combining a threshold decision and a multiplier function, to complete the estimation of the amplitude and phase error of the array, which is under the timely work of the reference information source; S3, correcting the fourth-order cumulative vector by utilizing the estimation result of the amplitude and phase error, and obtaining a Toeplitz non-singular matrix based on the strategy of a sub-array division; S4, accomplishing the DOA estimation by the ESPRIT algorithm. According to the method for the underdetermined DOA estimation based on the partially calibrated and nested array under the amplitude and phase error, which effectively estimates 2L<2> sources by using 2L array elements, and has good robustness on the amplitude and phase error of the array, the Gaussian white noise / color noise, and unknown non-uniform noise. Meanwhile, the method avoids an angle raster search with a large amount of computation.

The invention discloses a voltage fluctuation and flicker signal detection algorithm comprising the specific steps of mathematically modeling a voltage signal including flicker; converting detection of a fundamental wave signal into a simple mathematical optimization equation solving problem by constructing a proper target function according to an orthogonality characteristic of a cosine function; building an optimization equation for recovering an envelope from a modulation signal by using an L0 norm regularization method; and acquiring detection estimation of the fundamental wave signal and a flicker signal by solving the two optimization equations. The algorithm provided by the invention is high in detection precision and good in noise robustness which comes from the noise inhibition effect of a penalty term with a norm form in a flicker signal recovery equation.

The embodiment of the invention provides a hydroelectric generating set vibration trend prediction method and system. The method comprises the steps: collecting an original signal feature set of a hydroelectric generating set; analyzing the correlation between each intrinsic mode component and the environment variable based on a maximum information coefficient (MIC), and extracting the environment variable of which the correlation is greater than a preset correlation threshold as the environment feature of each intrinsic mode component; forming a feature input sequence by each intrinsic mode component and the corresponding environment feature, taking the feature input sequence as the input of a neural network according to a time sequence, and performing neural network training to obtain a vibration trend prediction model; and enabling each intrinsic mode component of the current hydroelectric generating set to be predicted and the corresponding environmental characteristics to form a characteristic input vector, inputting the characteristic input vector into the vibration trend prediction model for prediction, and fusing all prediction results to obtain a future trend prediction value of the hydroelectric generating set, so that the precision and stability of the vibration trend prediction result are greatly improved.

The invention provides a bearing fault diagnosis method, terminal equipment and a computer storage medium. The method comprises the steps: designing a corresponding differential operator for a bearingvibration signal model, and extracting fault vibration signal subcomponents from a vibration signal through a null space tracking algorithm; extracting instantaneous energy components in the vibration signal sub-components through a Teager energy operator, and further realizing data demodulation work; and finally, displaying the fault characteristic component by adopting a 1.5-dimensional energyspectrum. The method provided by the invention can accurately detect the fault frequency of the fault signals of the inner and outer ring bearings, and has good noise robustness.

The invention relates to a method for extracting local textural roughness of an image. The method comprises the steps of firstly calculating the average gray value of pixels in a multi-scale movable window in the image, respectively calculating average gray difference, between two eccentric overlapping windows, of each pixel in the horizontal direction and the vertical direction, then calculating the optimum size by utilizing the size corresponding to the maximum average gray differential value, and calculating the local roughness of each pixel point according to the optimum size of each pixel point in the image. According to the method disclosed by the invention, the noise robustness is good, and the local textural roughness of the image can be accurately measured.

The invention provides an SAR target recognition method based on non-negative least square sparse representation. According to the method, a frequency spectrum characteristic based on an SAR image is taken as a recognition characteristic, a test sample is projected on a training set, non-negative constraint is added in a sparse projection process, influence of non-practical mathematic description on radar target recognition caused by positive and negative sparse coefficients in sparse representation can be avoided, moreover, a low dimension structure of a target in a high dimension space can be effectively reflected through a sparse solution, category of a test sample can be determined through a sparse reconstruction process, radar target recognition is realized, the recognition rate is improved, moreover, azimuth estimation on SAR image targets and influence of factors such as defocusing or a signal to noise ratio on target recognition can be avoided, excellent noise robustness is realized, radar target recognition accuracy can be effectively improved.

The invention discloses a complex component point cloud splicing method and system based on feature fusion, and belongs to the field of computer vision, and the method comprises the steps: collecting the local three-dimensional point cloud of a component, and carrying out the positioning, wherein the adjacent local three-dimensional point clouds are partially overlapped; converting the local three-dimensional point clouds to a world coordinate system according to a positioning result, and extracting key points after uniform downsampling to obtain corresponding key point clouds; inputting the key point clouds into a feature fusion matching network to obtain a matching corresponding relationship between any two adjacent key point clouds, wherein the feature fusion matching network takes two adjacent point clouds as input and is used for extracting multi-scale features or multiple features of points in the point clouds and fusing the multi-scale features or the multiple features into feature descriptors of corresponding points so as to calculate a matching corresponding relation of the adjacent point clouds; and carrying out point cloud registration by using a matching corresponding relation of adjacent key point clouds, and splicing the key point clouds into an integral three-dimensional point cloud model based on a point cloud registration result. According to the invention, the precision of point cloud splicing of complex components can be improved.

The invention relates to a normalization possibilistic fuzzy entropy clustering method based on a Gaussian kernel hybrid artificial bee colony algorithm. The method comprises: (1), carrying out normalization pretreatment to obtain a new sample set X_new; (2), carrying out parameter initialization; (3), carrying out calculation to obtain a distance to an initial clustering center and carrying out calculation on a membership matrix U and a possibilistic matrix T to obtain an initial fitness value fitness (i); (4), entering a honey gatering bee stage; (5), entering a following bee stage; (6), entering a scout bee stage; and (7), obtaining a final optimal clustering center Vbest, obtaining a corresponding membership matrix U by the Vbest, and then obtaining a final clustering unit according to a formula: ci=argmax(uij). The provided method has the great noise robustness; the human dependence of parameters is reduced to a certain extent; and after artificial bee colony algorithm introduction, the global characteristic of the algorithm is improved and a parameter initial value sensitivity problem is solved. The feasibility and effectiveness are improved.

The invention relates to a noise reduction algorithm for a transient electromagnetic detection signal based on variational mode decomposition. The decomposition process of a goaf strong-interference transient electromagnetic signal is converted into solving of a variational problem. A constraint variation problem is converted into a non-constraint problem by carrying out variation construction onstrong interference transient electromagnetism and introducing a secondary penalty factor and a Lagrange penalty operator, limited modal components are obtained through a multiplication operator alternating direction algorithm, effective separation of signals is achieved, and noise and effective signals are separated. According to the invention, signals can be decomposed in a self-adaptive manner,priori knowledge is not needed, and the actual operation is simpler and more convenient; the variational mode decomposition algorithm is simple, the computer operation time is short, and the storagespace and the calculation time are saved; the modal aliasing problem in empirical mode decomposition is effectively solved, the extracted signals are more accurate, and better noise robustness is shown.

The invention provides a multi-scale U-shaped convolutional neural network MACS Net in order to solve the problems that target segmentation regions in overlapped chromosome images are different in size, not obvious in distinguishing and the like. Multi-layer cavity convolution and synchronous long pooling technology is introduced at the bottommost layer of UNet to realize detection of target segmentation regions of different sizes and extraction of features; convolution block connection is introduced between UNet codecs, so that semantic information difference is relieved. An intersection-to-parallel ratio (IoU) of chromosome overlapping regions is used as an evaluation index; the result shows that the segmentation IOU of the MACS Net at the chromosome overlapping part reaches 0.9860, which is improved by 2.78% compared with 0.99593 of UNet, and the MACS Net respectively shows more ideal noise robustness in the data set polluted by spiced salt, Gauss and Poisson noise.

The invention discloses a distribution network fault positioning method. The method comprises the following steps: acquiring a traveling wave signal of a fault current; carrying out Karenbauer phase-mode transformation processing on the traveling wave signal to generate a traveling wave component; carrying out variational mode decomposition processing on the traveling wave component to generate anIMF component; obtaining a high-frequency modal component in the IMF component; calculating a Teager energy operator according to the high-frequency modal component; detecting arrival time of a faulttraveling wave according to the Teager energy operator; and calculating a fault position according to the arrival time. The invention also discloses a distribution network fault positioning method and system. Variational mode decomposition and the Teager energy operator are combined to form a single-ended traveling wave distance measurement method, which can be complementary with a double-ended distance measurement method; and meanwhile, refraction and reflection waves can be accurately distinguished, and the positioning precision is improved.

The invention provides an end-to-end speech recognition method and system. The method comprises the following steps: training an initial speech feature extraction model by using a source corpus based on a VGGNet model; removing a full connection layer in the initial speech feature extraction model and freezing a preset number of convolutional layer parameters, and training the removed and frozen initial speech feature extraction model by using a target corpus to obtain a frequency domain feature extraction network; constructing an end-to-end speech recognition framework, wherein the framework comprises an encoder and a decoder; and training the end-to-end speech recognition framework by using a target corpus, and performing end-to-end speech recognition based on the trained end-to-end speech recognition framework. The method can effectively solve the problem of model overfitting under the condition of limited data, improves the accuracy of speech recognition, and has good noise robustness.