613results about How to "Strong generalization" patented technology

The invention provides a double-motor skidproof differential drive axle used in electric automobiles, which comprises two driving motors, a speed reducer and an output axle which is connected with theThe invention provides a double-motor skidproof differential drive axle used in electric automobiles, which comprises two driving motors, a speed reducer and an output axle which is connected with the speed reducer. The two driving motors are arranged face to face in the middle of the driving axle and connected into a whole through the shell of the speed reducer. The power output by the two drivinspeed reducer. The two driving motors are arranged face to face in the middle of the driving axle and connected into a whole through the shell of the speed reducer. The power output by the two driving motors converges together and flows into the speed reducer; and then the power further drives the left and the right wheels of the electric automobile to rotate through the output axle which is conng motors converges together and flows into the speed reducer; and then the power further drives the left and the right wheels of the electric automobile to rotate through the output axle which is connected with the speed reducer. The driving motors and the speed reducer can be arranged in an engine chamber to achieve compact structure and save space. Simultaneously, in order to solve the problem oected with the speed reducer. The driving motors and the speed reducer can be arranged in an engine chamber to achieve compact structure and save space. Simultaneously, in order to solve the problem of skidding, a skidproof differential device is also arranged in the speed reducer.f skidding, a skidproof differential device is also arranged in the speed reducer.

The invention discloses a motion recognition method based on a three-dimensional convolution depth neural network and depth video. In the invention, depth video is used as the object of study, a three-dimensional convolution depth neural network is constructed to automatically learn temporal and spatial characteristics of human body behaviors, and a Softmax classifier is used for the classification and recognition of human body behaviors. The proposed method by the invention can effectively extract the potential characteristics of human body behaviors, and can not only obtain good recognition results on an MSR-Action3D data set, but also obtain good recognition results on a UTKinect-Action3D data set.

The invention relates to a rolling bearing fault diagnosis method. The method uses a learning algorithm of a CNN (Convolutional Neural Network) theory to complete a feature extraction task needed by fault diagnosis, and does not rely one manual selection, intrinsic features of input data are extracted automatically from simple to complex and from low-level to high-level, and abundant information hidden in known data can be dug out automatically; and a support vector regression method is used to identity a test sample in a classifying manner, support vector regression with a high generalization capability can be used to identity an unknown new sample in higher precision, and the disadvantage that a default classifier of deep learning tends to be low in the generalization capability can be overcome when support vector regression serves as a classifier to identify samples in the classified manner. The rolling bearing fault diagnosis accuracy and validity can be improved, a new effective approach is provided for solving problems in rolling bearing fault diagnosis, and the method of the invention can be widely applied to fault analysis of complex mechanical systems in the fields of chemical industry, metallurgy, electric power, aviation and the like.

The invention discloses a fire video detection and early warning method based on image multi-feature fusion, and the method comprises the steps: firstly carrying out the preprocessing after an image sequence of a video is obtained; extracting a foreground region, and obtaining a detected candidate region; secondly, extracting static features and dynamic features from the candidate areas, judging whether flames are contained or not by taking the static features and dynamic features as input of an SVM classifier when the flames are detected, and obtaining whether smoke is contained or not afterlogic combination selection and calculation are conducted on feature judgment results when the smoke is detected; finally, if it is detected that flames or smoke exists, carrying out fire judgment according to the growth trend of the flames or smoke; when it is judged that a fire occurs, carrying out fire alarm on the monitoring site, and otherwise carrying out fire alarm only on the background. The method can be combined with an existing monitoring system to be applied to places such as shopping malls and warehouses, the fire detection and early warning cost is reduced, the detection method is good in generalization ability and applicability, reliable fire detection and early warning functions can be provided, and the method has practical value.

The invention discloses an on-line prediction method for residual life of rolling bearing, As that roll bearing move from a healthy state to a damaged state, The original signal samples and corresponding degeneration energy indexes are extracted from the running process of the bearing, and the original signal samples are used as the input of the five-layer convolution neural network model, and thedegeneration energy indexes are used as the output of the convolution neural network model, and the degeneration energy state model is obtained by training. Real-time acquisition of the original running signals of the rolling bearings to be tested; The original running signal of the rolling bearing to be tested is input into the degradation energy state model, and the degradation energy index isestimated. Then the estimated energy degradation index is used to predict the residual life of the rolling bearings to be tested. The prediction process of the invention only needs to collect the original operation signal of the bearing, and does not need to extract and screen the features, thus overcoming the technical problems that the prior art adopts the methods of feature extraction, featurescreening and regression prediction, which have the characteristics extraction difficulty and the precision is limited.

A large-scale scene three-dimensional reconstruction method for fusion of additional information includes: extracting SIFT (scale invariant feature transform) points of all images, performing image matching, and structuring external-pole geometric graphs to obtain trajectories corresponding to all three-dimensional spots; according to inertial measurement unit information or compass angles, obtaining initial camera rotation matrixes of all images, iteratively searching currently reliable connecting edges from the external-pole geometric graphs and performing global optimization by the aid of the edges; initializing the center of a camera to be a GPS (global position system) corresponding to the images to obtain initial projection matrixes of the images according to image initializing focus information, the rotation matrixes and the center of a camera, and iteratively triangulating and adjusting in bundle according to the projection matrixes and the trajectories of the three-dimensional spots. The large-scale scene three-dimensional reconstruction method is rapid in calculation, the obtained three-dimensional spots are reasonable and reliable, image mismatching sensitiveness is low, generalization performance is high, and the method is applicable to both orderly and disorderly image sets.

The invention relates to the technical field of mode recognizing, in particular to a real-time human body action recognizing method and device based on depth image sequence. The method comprises the steps of S1, extracting target action sketch from a target depth image sequence and extracting a training action sketch from a training depth image set; S2, performing gesture clustering on the training action sketch and performing action calibrating on the clustered outcome; S3, computing the gesture characteristics of the target action sketch and training action sketch; S4, performing the gesture training based on a Gauss mixing model by combining the gesture characteristics of the training action sketch and constructing a gesture model; S5, computing the transferring probability among all gestures of the clustered outcome in each action and constructing an action image model; and S6, performing action recognizing on the target depth image sequence according to the gesture characteristics of the target action sketch, the gesture model and the action image model. The real-time human body action recognizing method disclosed by the invention has the advantages of improving the efficiency of action recognizing and the accuracy and the robustness of the action recognizing.

The invention discloses a convolutional neural network classification-based container number recognition method. The method comprises the following steps of: S1, obtaining RGB images of a container indifferent directions, carrying out a series of preprocessing on the RGB images so as to locate container number areas in the images; S2, carrying out perspective transformation and binarization on the located container number areas in the step S1, and correctly segmenting various container number characters through a character border and projection method combined processing method; and S3, inputting the segmented container number characters in the step S2 into a preset five-layer convolutional neural network model according to a sequence of positions of the container number characters in thecontainer number, combining recognition results, and carrying out post-processing to obtain a correct container number. The method is capable of correctly solving various problems such as character deformation, incompletion and adhesion, and realizing higher recognition correctness and speed.

The invention provides a flame detection method based on image target detection, which belongs to the field of image processing, fire detection and video monitoring. The method comprises steps: firstly, a flame detection data set containing flame images and annotation information of each image is built, and the flame detection data set is divided to a training set and a test set; a deep convolution neural network model is built, the training set is used to carry out iterative updating on the model, the test set is used to calculate a loss function for the updated model, and if the loss function for the current model does not drop any more, the model completes training; and a real-time video is photographed, the model which completes training is used to detect each frame of image, if flameexists, the coordinate position of the flame in the image is outputted by the model, and a rectangular frame is used for marking. The phenomenon that features are manually designed for generating a candidate area for suspected flame is not needed, the deep convolution neural network model can be directly used for carrying out flame detection on the whole image, the position information of the flame is obtained, early warning of a fire is thus carried out, and hazards brought by fire can be reduced maximally.

The invention relates to a speech recognition method based on neural network stacking autoencoder multi-feature fusion. Firstly, the original sound data is framed and windowed, and the typical time-domain linear predictive cepstrum coefficient feature and the frequency-domain Mel frequency cepstrum coefficient feature are respectively extracted from the framed and windowed data; then the extractedfeatures are spliced, the initial feature representation vector of acoustic signals is constructed and the training feature library is created; then the multi-layer neural network stacking autoencoder is used for feature fusion and learning; the multi-layer autoencoder adopts the over-limit learning machine algorithm to learn training; and finally the extracted features are trained using the over-limit learning machine algorithm to get the classifier model; and the constructed model is finally used to test sample classification and identification. The method adopts multi-feature fusion basedon the over-limit learning machine multi-layer neural network stacking autoencoder, which has higher recognition accuracy compared with the traditional single feature extraction method.

The invention belongs to the technical field of computer image recognition, and specifically relates to a pedestrian re-identification method based on attitude normalized image generation. The methodcomprises steps of: predicting a pedestrian average attitude and attribute features; constructing, training and testing an attitude normalized image generation model and generating eight pedestrian images having different attitudes; constructing, training and testing pedestrian re-identification feature extraction network to obtain a pedestrian re-identification feature; and finally, performing the pedestrian re-identification feature fusion, and obtaining the features of a pedestrian target to be detected and all candidate pedestrian targets. The method of the invention has the advantages ofhigh speed, high precision, good robustness, good generalization ability and good expandability, and is very suitable for practical applications such as video pedestrian monitoring and video pedestrian information retrieval.

The invention relates to a laser welding process parameter optimization method based on a Bagging integrated prediction model and a particle swarm optimization algorithm. The optimization consists ofa laser welding process parameter prediction model and a multi-target particle swarm optimization method, the prediction model is obtained by fusing a plurality of basic learners through a Bagging model fusion algorithm, and the prediction model establishes a nonlinear mapping relationship between the laser welding process parameters and weld quality evaluation parameters. The optimized laser welding process parameters are finally obtained through the multi-target particle swarm optimization algorithm, the prediction accuracy is higher, guidance can be better provided for the formulation of the laser welding process parameters, and the formulation efficiency of the process is improved.

The invention discloses a target tracking method, system, device and medium based on a dual-stream convolution neural network. The method comprises the following steps: a spatial stream two-dimensional convolution neural network is constructed to extract the characteristic information of an image block in a current frame; a three-dimensional convolutional neural network is constructed to extract the motion information of the objects between frames in a video sequence within a certain time range; additive fusion of feature information of spatial flow two-dimensional convolution neural network and timing sequence flow three-dimensional convolution neural network; according to the fused feature information, a fully connected sub-network is constructed to extract the image blocks that meet therequirements. Boundary box regression is used to get the predicted position and size of the object in the current frame. Two-dimensional convolution neural network of spatial flow and three-dimensional convolution neural network of temporal flow are trained off-line before target tracking. In the process of target tracking, the on-line fine tuning of all-connected subnetworks is carried out. Theinvention achieves good tracking effect.

A migration learning lung lesion tissue detection system based on an MaskScoring R-CNN network comprises a storage module for storing four lung diseases including lung cancer, pneumonia, pulmonary tuberculosis and emphysema and further comprises a diagnosis module, and the diagnosis module is in communication connection with the storage module and comprises the following steps of 1) preprocessinga medical image; 2) constructing the MaskScoring R-CNN network model, wherein the step 2) specially comprises 1, constructing a shared convolutional neural network backbone (for feature extraction); 2, carrying out transfer learning on a shared convolutional neural network; 3, constructing an FPN network; 4, constructing an RPN network; 5, constructing an ROIAlign layer; 6, adding the MaskIoU head; and 3) identifying the lung medical image lesion tissue, inputting a to-be-detected lung CT image into the constructed MaskScoring R-CNN network, outputting and obtaining an identified image by thenetwork, framing out and masking the identified lesion tissues, and marking the lesion categories. According to the method, the requirement for high precision of medical image segmentation is met, andthe network can have the good generalization.

The invention provides an automatic Parkinson's disease identification method based on multimode hyperlinks network modeling. The method includes: the DTI structure connection is used as the constraint and fused into the building process of an fMRI brain function network to build a multimode hyperlinks network model; node degree, edge degree and fit degree are extracted according hypernet featuresto serve as the original feature set, a multitask feature selection method (semi-M2TFS) is used to perform optimal feature subset screening on the original feature set to obtain the feature subset indicating the maximum difference degree between a Parkinson's disease patient and a normal person; a multi-core support vector machine pattern classifier is trained according to the optimal feature setand applied to Parkinson's disease patient classification diagnosis. Compared with an existing single-mode hyperlinks network modeling method, the method has the advantages that the multimode hyperlinks network can truly reflect the brain function connection mechanism and is excellent in classification identification accuracy and significant to the assisting of Parkinson's disease clinical diagnosis and automatic identification.

The invention provides a phishing website detection method and system based on an adaptive heterogeneous multi-classification model. The method is characterized by for a multiple-base classification algorithm, through linear addition, constructing the adaptive heterogeneous multi-classification model; training the multi-classification model, wherein a model input is the input of each base classification algorithm and an output is a sample label, and each base classification algorithm extracts a corresponding characteristic from a sample record and is taken as the input; and using a machine learning algorithm to solve a model parameter, adopting a test set to test and optimize, and finally acquiring the detection model of the type of a phishing website. The system comprises a domain name morpheme characteristic classifier, a subject index characteristic classifier, a content similarity characteristic classifier, a structural style characteristic classifier, a visual rule characteristicclassifier, a linear addition training module, an integrated classifier, a training data set management module, and a detection and alarm module. In the invention, the phishing website can be detectedin real time, and the accuracy and the stability of phishing website detection are increased.

The invention discloses a pedestrian re-identification model, method and system for adaptive difficulty mining. The identification method comprises the steps of: randomly dividing sample pictures intoa training set used for each iteration, inputting the training set into a convolutional neural network, obtaining the probability that each sample pair belongs to a positive or negative sample pair by using a softmax function, and then obtaining the loss of each sample pair by using a multinomial logistic function; obtaining a difficult sample pair by using the loss of each sample pair; and training the convolutional neural network by using the difficult sample pair until the current number of iterations reaches the upper limit of the number of iterations, thus obtaining the pedestrian re-identification model. The pedestrian re-identification model is used to extract features of each picture in a picture set to be identified, and then a similarity order of the sample pairs in the pictureset to be identified is obtained. The pedestrian re-identification model, method and system avoid over-fitting and under-fitting, and have high recognition accuracy.

The invention provides a diabetic retinopathy grade classification method based on deep learning. The diabetic retinopathy grade classification method comprises the steps of: constructing a sample library; removing backgrounds and noise of ophthalmoscope photographs in the sample library; normalizing the images of different brightness and different intensity to the same range by adopting a local mean value subtracting method; adopting random stretching and rotating methods for different samples for data augmentation, and constructing a training set and a test set; training an initial deep learning network model by establishing an input portion architecture, a multi-branch feature transformation portion architecture and an output portion architecture separately; and inputting samples to betested into the trained initial deep learning network model for diabetic retinopathy grade classification. Compared with the traditional processing method, the diabetic retinopathy grade classification method gets rid of the dependence on prior knowledge, and has good generalization ability; and by adopting the designed multiple grades, a small-sized convolution kernel can be used for extracting very tiny lesion features, thereby making the classification results more reliable.

Methods and devices for long term, cuffless, and continuous arterial blood pressure estimation using an end-to-end deep learning approach are provided. A deep learning system comprises three modules: a deep convolutional neural network (CNN) module for learning powerful features; a one- or multi-layer recurrent neural network (RNN) module for modeling the temporal dependencies in blood pressure dynamics; and a mixture density network (MDN) module for predicting final blood pressure value. This system takes raw physiological signals, such as photoplethysmogram and/or electrocardiography signals, as inputs and yields arterial blood pressure readings in real time.

The invention discloses a UUV (unmanned underwater vehicle) dynamic planning method based on an LSTM-RNN (long short term memory-recurrent neural network), and belongs to the field of unmanned underwater vehicles. The UUV dynamic planning method includes the steps: (1) selecting a geometric model to build an obstacle environment model; (2) building a UUV dynamic planner for acquiring a data set byan ant colony algorithm; (3) designing an LSTM-RNN model for dynamic planning; (4) acquiring the data set; (5) training the LSTM-RNN by data of a training set in the data set to obtain the dynamic planner based on the LSTM-RNN; (6) inputting sonar detection information and target point information to the dynamic planner based on the LSTM-RNN to obtain the navigational direction and the navigational speed of a UUV at a next time. The method has strong learning capacity and further has quite strong generalization capacity, so that the implemented dynamic planner is applicable to complex environments. The requirement of real-time performance is met, and planned routes conform to movement characteristics of the UUV.

The invention discloses a network intrusion detection method based on space-time feature fusion. The method comprises: establishing a convolutional neural network with a plurality of convolution kernels of different scales to be connected with a fusion detection model of a long-term and short-term memory network; carrying out spatial domain feature extraction on the network traffic data to be detected by adopting a convolutional neural network with a plurality of different scale convolution kernels, and then continuously carrying out time domain feature extraction by adopting a long and shortterm memory network; and then carrying out pooling operation, and finally carrying out classification processing on the to-be-detected network traffic data fused with the spatial domain features and the time domain features by combining a classifier. The method has higher accuracy, lower false alarm rate and missing report rate and excellent generalization capability when processing a public dataset with high dimension, nonlinearity and large data volume.

The invention provides a remote sensing image thin and weak target segmentation method. Firstly, data enhancement and corresponding preprocessing are carried out on an original remote sensing image, U-net is improved by means of the dense connection thought of DenseNet, and a Dense-Unet network structure is provided. Dense convolution is used in a network structure, the cascade relation between convolution channels is enhanced, through a symmetric structure and a jump connection thought, the connection between features of all layers is further tighter, and thin and weak target features can belearned more effectively. In order to ensure the real-time performance of final network identification and reduce the parameter quantity, a bottleneck layer and a batch normalization layer are introduced behind each dense block. And the objective function is adjusted by using the cost-sensitive vector weight, so that the problem of unbalanced segmentation target categories is solved, and the segmentation precision is further improved. And finally, a plurality of independent models are trained by using an ensemble learning method, the independent models are combined, and target category information is jointly predicted in the picture.

Disclosed is a diagnostic method of space gridding structure node bolt loosening injury. The diagnostic method of space gridding structure node bolt loosening injury comprises that a bolt ball node is modeling elaborately and a structural combination unit model is set up. The diagnostic method of space gridding structure node bolt loosening injury is divided into a plurality of substructures and is numbered according to a way of geometric position continuity according to composition characteristics of the space gridding structure. A structure testing point is optimizely arranged in a sensor mode and sensitivity analysis of a rod piece is carried out by frequency. Numerical value of bolt loosening injury is simulated. A training sample, an input parameter and an output parameter of a neural network is confirmed. An injury sample is input to a generalized regression neural network (GRNN) network B which is training completed. The output is an injury index of the node, namely the injury position is positioned to the existing node. The number of a training sample when the injury is accurately positioned can be greatly reduced. Practicality that the space gridding structure node bolt loosening injury is diagnosed by taking advantage of neural network technology is strengthened. Especially the diagnostic method of space gridding structure node bolt loosening injury has a prominent advantage to a large-scale space gridding structure with a plurality of the nodes.