54results about How to "Implement feature extraction" patented technology

The invention discloses a rolling bearing fault diagnosis method based on improved variational model decomposition and an extreme learning machine. The method comprises: vibration signals of a rollingbearing under different types of faults are collected, the vibration signals are filtered by means of maximum correlation kurtosis deconvolution, parameter optimization is carried out on the maximumcorrelation kurtosis deconvolution method by using a particle swarm algorithm, and an enveloped energy entropy after signal deconvolution is used as a fitness function; the mode number of variationalmodel decomposition is improved by an energy threshold and improved variational model decomposition of the filtered vibration signals is realized to obtain mode matrixes of the corresponding vibrationsignals; singular value decomposition is carried out on the mode matrixes to obtain a singular value vector and a rolling bearing fault feature set is constructed; and the fault feature set is trained by using an extreme learning machine and a rolling bearing fault diagnosis model is established. Therefore, stable feature extraction of the complex vibration signal of the rolling bearing is realized, so that the diagnostic accuracy is improved.

Disclosed is a hyperspectral data classification method which is combined spectrum and spatial information. The steps comprises (1) reading the hypersectral data, (2) confirming the minimum size of structural element, (3) calculating differentiation between picture elements in neighborhood of each structural element by extended mathematical morphology expansion and corrosion operation, (4) obtaining exponential value of morphology eccentricity by the extended expansion and the corrosion operation of step (3), (5), constantly repeating the above steps with the adding of the size of the structural element to achieve the maximum size of the structural element, (6), constantly updating the exponential value MEI of morphology eccentricity in iteration process via the obtained new value, and generating a final exponential value MEI of morphology eccentricity after the iteration process is finished, (7) realizing the extraction of the data characteristic by the image of the exponential value MEI of morphology eccentricity, namely generating ground object type information, and realizing sophisticated category of the ground object by a minimum-distance classifier. The method is an unsupervised classification method for hyperspectral ground object with strong stability, high reliability and high accuracy.

The invention discloses a transformer fault diagnosis method based on improved principal component analysis, which belongs to the technical field of transformer fault diagnosis. The method adopts the sum of absolute values of sample indexes to carry out standardized treatment on sample index values, and thus, the difference of each sample index value in magnitude can be eliminated, and the information difference features between samples can also be kept; and sample principal components are selected according to cumulative contribution rates of the principal components, Euclidean distances among the sample principal components are clustered, and a transformer fault type is judged. The method of the invention can effectively improve the accuracy for diagnosing inner hidden fault of the transformer.

The invention relates to hydraulic cylinder leakage monitoring and inner leakage level classification, in particular, a hydraulic cylinder inner leakage fault diagnosis and evaluation method and belongs to the equipment health monitoring field. According to the method, wavelet decomposition and a BP neural network are used in combination; pressure signals of the inlet of a hydraulic cylinder are segmented through adopting a wavelet analysis method, and time-domain features of segments are extracted; and a hydraulic cylinder leakage level evaluation method is built through adopting a BP neural network method. As indicated by an experiment, the method is effective and can accurately realize diagnosis on the inner leakage of a hydraulic cylinder. According to the method, inner leakage of the hydraulic cylinder can be detected accurately through an inlet pressure sensor which is normally arranged in the hydraulic cylinder; and the level of the inner leakage of the hydraulic cylinder can be detected. The hydraulic cylinder inner leakage fault diagnosis and evaluation method is simple and practical, and can realize the detection of the inner leakage of the hydraulic cylinder. With the method adopted, with no extra sensors adopted, diagnosis on the inner leakage of hydraulic cylinders commonly used in the industry can be realized, and accidents caused by the inner leakage of the hydraulic cylinders can be avoided.

The invention discloses a self learning-based coal rock recognition method. The method comprises steps: a high-layer structure feature matrix D is firstly learnt from auxiliary data in an offline mode, wherein the auxiliary data are non-labeled non-coal rock natural images and are acquired more easily; the coal rock images are represented as a linear combination of a plurality of high-level characteristic atoms, and the coefficient of the linear combination forms a new eigenvector as the eigenvector of the coal rock image; the extracted coal rock eigenvector is then used to train a classifier;and in a recognition process, an eigenvector of an unknown category of a coal rock image is extracted and inputted to the classifier which completes training, and the category of the coal rock imageis finally outputted. The method uses non-labeled non-coal rock natural images which are acquired easily as training samples, the overhead of marking a large amount of coal rock samples is saved, theextracted coal rock eigenvector has strong discriminability and robustness, and good recognition effects are achieved.

The invention discloses a polarization SAR image classification method based on a Wishart deep network and mainly aims at solving problems that current feature extraction needs a lot of prior knowledge and a manual labor intensity is high. The method comprises the following steps of (1) inputting a polarization SAR image and carrying out filtering processing; (2) constructing a multilayer Wishart RBM learning feature to the image after filtering; (3) using a learned feature to train softmax classifier; (4) using the multilayer Wishart RBM and the softmax classifier to construct a deep network DBN and training the deep network DBN; (6) using the deep network DBN to classify the polarization SAR image and outputting a result. Compared to a classic classification method, by using the method of the invention, a classification correct rate is high; a classification-result homogeneous area is complete; area consistency is good and classification performance is good too. The method is suitable for carrying out terrain classification and target identification on the polarization SAR image.

The invention discloses a pipeline positioning method. The pipeline positioning method comprises the steps that original signals are subjected to CEEMD; then noise-containing signals which are formedafter decomposition is accomplished are subjected to noise removal; cross-correlation calculation is conducted on the signals, and invalid redundancy components are removed according to the cross-correlation degree; and finally, the signals which are subjected to decomposition and noise removal are rebuilt, single leakage source signals are isolated through FastICA, and leakage points are positioned by use of a time difference positioning formula. The experiment result shows that the pipeline positioning method is suitable for detection and positioning of multi-point leakage of a metal pipeline, can improve the accuracy of pipeline leakage positioning to a large extent, and provides the powerful theoretical basis and the practical experience for practical engineering application.

The invention provides a pipeline multi-point leakage accurate positioning method. Firstly, original signals are decomposed into the sum of several IMFs by using an improved CEEMD; then envelope signals of the IMFs in all orders are calculated, and a sample entropy of each envelope signal is calculated; then according to the size of sample entropy values, the IMFs in all orders are optimized, andthen the preferred IMFs are reconstructed and are subjected to dimension raising; single leak source signals are separated through a blind source separation and signal sparsity combined manner; and finally, a time difference positioning formula is used for positioning leakage points. The experimental results show that the method is suitable for pipeline multi-point leakage detection and location,the pipeline leakage positioning accuracy can be improved to a large extent, and the powerful theoretical basis and practical experience are provided for practical engineering application.

The invention discloses a face recognition system and method based on deep learning and relates to the technical field of deep learning face recognition. The system includes an image acquisition module and a face recognition module. The face recognition module includes a control unit. The control unit is electrically connected with a face detection module, a deep learning training module, a feature extraction module, a matching identification module, a storage unit and a display screen. The storage unit is electrically connected with the deep learning training module, the feature extraction module and the matching identification module. The feature extraction module is electrically connected with the face detection module, the deep learning training module and the matching identification module. A face feature model library is disposed in the storage unit. According to the face recognition system and method based on deep learning, a convolutional neural network model is obtained through training of the deep learning training module, feature extraction in the process of face recognition by the convolutional neural network model, and the problems of low accuracy and low efficiency ofthe existing face recognition are solved.

The invention provides a remote sensing image road extraction method based on D-LinkNet, and the method comprises the following steps: S1, inputting a feature map into a D-LinkNet network, and completing the processing in an encoder sub-network based on a residual network and transfer learning; S2, inputting the feature map output in the step S1 into a feature extraction sub-network based on an expansion convolution and convolution block attention module for feature extraction; S3, enabling the feature map obtained after processing of the first two sub-networks to enter a decoder sub-network based on transposed convolution to realize image recovery. According to the invention, downsampling can be carried out on the road features in the remote sensing image, the problem of network degradation is well avoided, and the extraction of the road features is enhanced; the receptive field can be amplified by using the expansion convolution, the road characteristics in a larger range are sensedand the characteristics are extracted without increasing down-sampling, and the problem that the proportion of the road part in the remote sensing image is too small can be well solved.

The invention discloses a polarized SAR image classification method based on a dual-channel convolutional network. The method comprises the steps of filtering a to-be-classified polarized SAR image; extracting a multi-dimensional feature vector from a coherence matrix of each pixel point of the filtered polarimetric SAR image; performing spatial weighting on the polarized SAR image; randomly selecting a training sample and a test sample for each surface feature type of the polarized SAR data according to the real surface feature mark; constructing a multilayer convolutional network model; inputting the training sample into a multilayer convolutional network model to obtain a trained convolutional network model; inputting the test sample into the trained convolutional network model to obtain a classification result of each pixel in the test sample; comparing the classification result with a real surface feature mark, and calculating a correct rate; outputting the classification results.The method has higher classification accuracy on the ground objects, the homogeneous region is more complete, the region consistency and the classification performance are better, and the method is suitable for the ground object classification and target recognition on the polarized SAR images.

The invention relates to an air quality prediction method in gridding monitoring. The method comprises the following steps: firstly, performing data cleaning on position information and historical air pollutant concentration information of each monitoring station in gridding monitoring input by a user, then inputting the processed data into a GCN to extract space correlation information among the monitoring stations, and inputting the data with the space information into an LSTM to extract time features; and finally, integrating the features extracted by the GCN and the LSTM through a linear regression layer, generating a prediction result, and returning the prediction result to a user. The effectiveness of the method is verified through related experiments.

The invention belongs to the technical field of human body heart rate health monitoring, and particularly relates to a wearable device real-time heart rate monitoring method based on sensor fusion. Interference factors in original photoelectric pulse signals and acceleration signals are removed by constructing a multi-sensor fusion least square adaptive filtering denoising algorithm model, a decision tree classification model is established by extracting peak-to-peak values and root-mean-square features of three-axis acceleration signals, and meanwhile, a heart rate classification interval is defined according to the model; then a PPG signal output by a filter is input into a classification model to determine a frequency spectrum interval of a current heart rate value, and the maximum frequency spectrum peak value is searched in the interval to determine the heart rate frequency; and finally, three paths of denoised photoelectric pulse wave signals are fused to calculate a heart rate value, and the heart rate value is used as a final heart rate value. Compared with the prior art, according to the invention, noise interference in the calculation process is effectively removed, the calculation process is simple, and the accuracy of the obtained heart rate value is higher.

The invention relates to a signal processing technology for electromagnetic excitation acoustic emission. The signal processing technology comprises the following steps of: performing feature extracting and positioning based on fast fourier transform by a signal acquisition system, a preamplifier, a data acquisition card and a band-pass filtering system; amplifying an acoustic emission signal collected by the high-sensitivity, high-speed and distortion-less signal acquisition system by the preamplifier, and then inputting into the data acquisition card for converting into a digital signal; and after passing by the band-pass filtering system, performing feature extraction on the digital signal by adopting fast fourier transform, thereby realizing the judging and positioning on the defects. According to the signal processing technology provided by the invention, the quick treatment for the electromagnetic acoustic emission signal of the crack surface defect of the metal sheet made of a non-ferromagnetic material is realized by adopting the fast fourier transform; the signal processing technology has the characteristics of high processing speed, obvious signal feature and high reliability; according to the technology, the noise jamming to the acoustic emission technology caused by external environment is reduced; the usability and portability demands on an engineering application are met; and the industrial application value is excellent.

The invention provides a method and system for extracting the static stability boundary characteristics of a large power network based on a spatio-temporal sequence, belonging to the field of safe andstable operation of the large power network. The method of the invention comprises the following steps of collecting static stable boundary holographic sample data; selecting static and stable boundary sample data. mapping the static stable boundary sample data to three sample images; calculating the Gaussian difference scale space of that three sample picture by extracting the features of the source sample picture, the net sample picture and the charge sample picture based on the scale invariant characteristic transformation method; carrying out the feature matching between the boundary loadsample image and the boundary net sample image; extracting the sample charge-network static stable boundary characteristics. The method of the invention avoids the shortcomings of less actual operation data and small working condition coverage, and the obtained static stable boundary characteristics are more convincing.

The invention discloses a vibration signal denoising method and system based on independence. The method comprises the following steps: obtaining a phase mark starting point position and a mark lengthparameter of the signal; According to the reference signal, creating the reference data at the starting point position of the phase mark; A phase shifting data set is created according to the phase mark starting point position and the mark length parameters of the comparison signal. Independent component analysis (ICA) was used to process the datum and phase-shifted dataset to obtain the processseparation signal. Acquiring a phase marker factor matrix of the process separation signal; Determining the phase information according to the outlier information of the phase marker factor matrix; Adjusting the phase of the comparison signal according to the phase information, and constructing an adjustment data set together with the reference signal; The adjusted data set is processed by independent component analysis, and the final separated signal is obtained. The denoised signal is determined according to the time-frequency characteristics of the final separated signal. The invention caneffectively remove the noise of the vibration signal and realize the feature extraction of the vibration signal.

To reduce complexity of face recognition, the invention provides an augmented reality device low in power consumption. The augmented reality device can realize processing of face recognition data andfeature extraction of face texture information accurately. Meanwhile, an algorithm is relatively easy, and low-power-consumption distributed AR data transmission on the basis of face recognition can be achieved.

The invention relates to the technical field of neural networks, and provides a CNN model compression method and device based on a DS structure, and a storage medium, wherein the method comprises thesteps: S110, enabling DW convolution and SE Module to form DS convolution blocks through common convolution and batch standardization BN operation, wherein the DS convolution blocks comprise a convolution Conv-1, a batch standardization BN, an activation function, a DW convolution, a batch standardization BN, an activation function, an SE Module, a convolution Conv-2 and a batch standardization BN; S120, stacking the DS convolution blocks in order to form a neural network structure; and S130, adding an input layer, a pooling layer, a full connection layer and a classification layer to the neural network structure to form a neural network model. According to the method, the DS convolution block structure is applied to the neural network, and the number of parameters of the neural network isgreatly reduced while the picture feature extraction capability is ensured.

A correlation filter tracking method based on saliency detection and image segmentation is disclosed. An improved correlation filter method combining saliency detection and image segmentation is proposed, which can improve the tracking accuracy by destroying the background in an image, highlighting the target features, and weakening the background features. The method comprises the steps of acquiring a video stream; detecting a target; distinguishing the foreground and background in the image by saliency detection; performing image contrast enhancement and image segmentation; carrying out andoperation on the obtained segmentation map and the original image to obtain a rectangular frame containing only the target information. Through saliency detection, the contrast degree is improved and,through the image segmentation, the background information in the original rectangular frame is destroyed and the target information is preserved, which makes the expression of the target informationfeature stronger than the background information, and solves the problem of correlation filter tracking failure in complex background.

According to the traffic sign recognition method provided by the invention, the detection effect on the road traffic sign is remarkably improved, and the requirement on real-time performance is met. A weighted bidirectional feature pyramid network with more balanced accuracy and efficiency is adopted to replace a path aggregation network, and channel features of road traffic signs are better fused. Secondly, common convolution is replaced by the cavity convolution, and the cavity convolution is combined with the space pooling pyramid, so that the receptive field is further expanded. Meanwhile, the detection scale is increased to four types, and the small target detection effect is improved; and random cutting is added in a data enhancement technology, so that the model learns more detail features. Finally, digital image operation technique is used to increase the number of instances for low precision categories.

The invention discloses a near-infrared spectrum feature extraction method and device. The method comprises the following steps: acquiring to-be-detected N samples; acquiring near-infrared spectrum data of the N to-be-detected samples by using a spectrograph; preprocessing the near-infrared spectrum data to obtain two-dimensional near-infrared spectrum smooth data; arranging arrangement and conversion on the two-dimensional near-infrared spectrum smooth data to obtain four-dimensional spectrogram data; carrying out feature extraction on the four-dimensional spectrogram data; and then carryingout feature arrangement on the four-dimensional spectrogram data after feature extraction to obtain two-dimensional feature data. The near-infrared spectrum feature extraction method and device have the following advantages: the data integrity can be guaranteed; feature extraction can be carried out in a full-spectrum interval; and the information is protected from being lost.

The embodiment of the invention provides a water pollution prediction method and device and an electronic device. The method comprises the steps of obtaining to-be-tested data of a first station and to-be-tested data of at least one second station in a predetermined area corresponding to the first station; obtaining a first influence parameter and a second influence parameter by using an attentionmechanism of a first convolutional neural network model; obtaining a prediction result aiming at the water quality data of the first station, wherein the prediction result is a change trend of the water quality data of the first station in a future predetermined time period; and processing the first influence parameter, the second influence parameter and the prediction result by using a preset second convolutional neural network model to obtain an adjusted prediction result. According to the embodiment of the invention, the problem of low prediction result accuracy caused by the fact that theinfluence of other water flow information near the tested point is not considered in the prior art is solved.

The invention relates to a method and a device of recognizing ventricular tachycardia heart rhythm based on transfer learning. The method includes the following steps: imputing a multi-lead electrocardiogram signal into a trained SCNN neural network; and judging the type of the multi-lead electrocardiogram signal according to the output result of the SCNN neural network. During the training of theSCNN neural network, firstly plenty of non-ventricular tachycardia electrocardiogram signals of known types and collected and copied ventricular tachycardia electrocardiogram signals are adopted to train the SCNN neural network to confirm the parameters of convolutional layers and pooling layers; and then the SCNN neural network is trained by the collected ventricular tachycardia electrocardiogram signals to confirm the parameters of fully connected layers. Through training twice, the parameters of the convolutional layers and the pooling layers are confirmed by the first training and the parameters of the fully connected layers are confirmed by the second training. Therefore, the invention provides the method of recognizing the ventricular tachycardia heart rhythm based on the transfer learning with high recognition accuracy without collecting too many ventricular tachycardia electrocardiogram signals.

The invention discloses a laser ultrasonic detection method for metal surface microcrack depth, which comprises three steps of laser excitation, laser detection and signal processing, and comprises the following steps: based on a phase grating, combining a grating modulated laser pulse signal simulated by a finite element method with an experimental signal to construct a data set; adding a random Gaussian noise amplification data set, and carrying out wavelet transform processing on the simulation signal and the experiment signal; and subsequent signals are processed by using the deep neural network. According to the laser narrow-band ultrasonic micro-crack identification method based on deep learning, a quantitative characterization device for the depth of the micro-crack on the metal surface is designed, and the micro-crack on the surface of a metal material can be reliably and flexibly characterized; and the method can be further expanded into a universal fully contactless automatic detection method for representing surface and subsurface defects of metal and semiconductor materials of dozens of microns or even submicrons.

The invention discloses an ampoule bottle printed word defect detection method based on image registration, which comprises the following steps of: 1, acquiring an ampoule bottle printed word image by using a linear array industrial camera to obtain a template image and a to-be-registered image; 2, performing feature point extraction and feature point description on the template image by using an SURF algorithm, and performing feature point extraction and feature point description on the to-be-registered image; step 3, performing feature point matching by using an FLANN matching algorithm; 4, according to the matched feature point pairs, calculating a transformation matrix of the to-be-registered image mapped to the template image; 5, calibrating the to-be-registered image through the transformation matrix, and eliminating image distortion; and 6, carrying out image difference on the template image and the calibrated to-be-registered image to obtain a difference image, and judging whether the printed characters have defects or not according to the difference image. According to the method, the feature extraction, matching, correction and detection of the ampoule bottle printed character image can be quickly and effectively realized, and the detection process takes less time.

The invention discloses a biological sequence feature extraction method based on word embedding and auto-encoder fusion. The method comprises the steps: constructing a representation model and a compression model, wherein the representation model comprises a word embedding network, and the compression model is an auto-encoder model and comprises an encoder and a decoder; taking minimization of a set total loss function as an optimization target, jointly training a representation model and a compression model, taking a short sequence Kmer set as input by a word embedding network, shielding part of the short sequence Kmer, carrying out context association on the Kmer in the sequence, learning an embedding vector of each Kmer in the sequence, and obtaining embedding information corresponding to the Kmer forming the sequence; enabling an encoder of the compression model to convert the embedded information into a low-dimensional feature vector, decoding Kmer embedding of a reconstruction sequence through a decoder, and outputting a reconstruction vector; and using the reconstruction vector to classify the shielded Kmer in the sequence. According to the method, efficient characterization of the biological sequence is realized, and the accuracy of subsequent classification is ensured.

The invention provides a distributed AR data transmission method, and aims at reducing the complexity of face recognition. Processing of the face recognition data and feature extraction of the face texture information can be accurately realized. Meanwhile, multiple defects existing in the prior art can be overcome and the algorithm is relatively simple and easy to implement.