49results about How to "Reduce the amount of training" patented technology

The invention provides a method for training a wake-up model and a device thereof, and the method comprises the steps: obtaining a first training set and a second training set when the model trainingis triggered; respectively inputting the first training set into an initial acoustic model and a current acoustic model, and determining a first difference parameter by comparing output results of theinitial acoustic model and the current acoustic model; inputting the second training set into a current acoustic model, and determining a second difference parameter by comparing an output result ofthe current acoustic model with a one-hot code corresponding to wake-up voice which can be recognized by the current acoustic model; and adjusting model parameters of the current acoustic model according to the first difference parameter and the second difference parameter. By utilizing the method provided by the invention, the current acoustic model can be compatible with the initial voice underthe condition of ensuring that the current acoustic model adapts to the current scene, the risk of unstable performance of the wake-up model caused by updating is reduced, and the trained acoustic model is well compatible with the previous initial wake-up scene on the premise of adapting to a more complex scene.

The invention discloses an image classification method based on an observation matrix transformation dimension, which comprises the following steps of: performing sparse coding on an image by using perceptual compression to obtain a data set consisting of low-dimension images, and dividing the data set containing label labels into a training set and a test set; Constructing an image classificationnetwork comprising an input layer, a hidden layer and an output layer, wherein the hidden layer is a perceptron unit; providing At least two image classification networks,wherein each image classification network comprises different node number perceptron units; Taking the training set as input, and carrying out training under the supervision of a label to obtain a corresponding neural network image classification model after the training is completed; Verifying the image classification accuracy of the neural network image classification model by using the test set, and selecting the neural network image classification model with the highest accuracy as a final neural network image classification model; And inputting an image to be detected, and outputting a prediction probability of an image classification result. According to the image classification method provided by the invention, the model efficiency can be greatly improved under the condition that the image classification precision is not reduced.

The invention relates to a die profile local springback compensation method based on elliptical surface mapping drive. The die profile local springback compensation method based on the elliptical surface mapping drive comprises the following steps that firstly, a local springback distribution cloud map of a part is obtained; secondly, a drive coordinate system is established according to the cloud map, an elliptical sketch is drawn, an interior area of the elliptical sketch is defined as a springback compensation area, and a reference curved surface and a target curved surface are established according to the springback compensation area and the springback compensation amount; and finally, the die profile is driven to deform in a mapping manner with the reference curved surface and the target curved surface as the deformation criterion and with the Z-axis direction of the drive coordinate system as a deformation direction, and the local springback compensation of the product is completed. Compared with the prior art, the die profile local springback compensation method based on the elliptical surface mapping drive has the beneficial effects that the efficiency of local springback compensation of the profile of a stamping die is improved, the local springback compensation precision of the profile of the die and the curved surface quality of the compensated profile of the die are effectively guaranteed, the die research and study amount is reduced, and the manufacturing cycle of the die is shortened.

The present invention relates to a training machine for strength training and rehabilitation comprising pull or press means (42, 43), which are arranged to be moved backwards and forwards while a pre-determined number of weights (18, 72) in a weight package is arranged to be lifted and lowered, alternatively, by connection means (22, 19), and means (42, 80) to lift said pre-determined number of weights in a continuous movement by means of a first user intended powered force against said pull or press means (42, 43) and to lower said pre-determined number of weights by means of a second user intended powered force against said pull or press means (42, 43), said first power being less than said second power. The training machine comprises a frame carrying said weights and along which said weights are slidable and which is turnably journalled about a substantially horizontal turning axis (9, 10).

The embodiment of the invention provides a driving event recognition method and device, a training method and device, equipment and a storage medium. The method comprises the following steps: collecting data representing a driving event as a driving parameter, and dividing the driving parameters in the first time period into a first target parameter and a second target parameter; in an event recognition model matched with the type of a vehicle, searching an event recognition model suitable for processing the second target parameter as an original event recognition model, training an original event recognition model by taking the first target parameter as a sample for identifying urgency and the second target parameter as a sample for identifying non-urgency; obtaining a target event recognition model, wherein the target event recognition model is called to recognize the emergency driving event from the driving parameters in the second time period, training is continued on the basis ofthe previous event recognition model, the training amount is small, the real-time requirement is met, the event recognition model conforming to the driving style of the user is learned, and the personalized driving event is recognized.

The invention discloses a method for effectively acquiring event-related magnetic field information in magnetoencephalogram signals. The method includes the steps of firstly, collecting and preprocessing magnetoencephalogram data; secondly, building a time-frequency atom database; thirdly, using a single-channel matching pursuit algorithm to build a linear combination; fourthly, forming a multi-channel matching pursuit algorithm; fifthly, determining iteration termination by the total residual energy of all channels so as to obtain atoms after signal decomposition; sixthly, removing the atoms representing artifact noise, and rebuilding the signals. The method has the advantages that the magnetoencephalogram signals are post-processed by the method, stimulation times can be reduced greatly, and test results are prevented from being affected by the fatigue, which is caused by long-time and repeated stimulation, of a scanned person; the training amount of a to-be-tested person is reduced, the requirements of the to-be-tested person are lowered, and the selection range of to-be-tested persons of clinical researches is expanded; data collecting time is reduced, research cost is lowered, and the clinical actual researches and popularization and application of event-related magnetic fields are benefited.

The invention discloses a head posture detection method and device and computer equipment. The method comprises the steps that event time sequence signals and RGB video data in the moving process of a to-be-detected object are acquired; respectively screening out an event data stream and a key RGB video stream at the head posture change moment of the to-be-detected object from the event time sequence signal and the RGB video data, and carrying out framing processing to obtain an event image sequence and a key RGB image sequence; outputting the event image sequence and the key RGB image sequence into a trained fusion model, respectively extracting event modal features and image modal features, and fusing the event modal features and the image modal features to obtain a head posture feature image of the to-be-detected object; predicting a head posture angle of the to-be-detected object according to the head posture feature image; according to the method, the head posture is estimated by using the visible light-event bimodal image, the change moment of the head posture can be effectively screened and captured, and accurate estimation can also be realized under the conditions that the illumination condition is not ideal, partial shielding exists and the like.

The invention relates to the field of image processing, and provides a pipeline weld defect detection method and a related device, and the method comprises the steps: obtaining an auxiliary picture set and a detection picture set; inputting the auxiliary picture set and the detection picture set into a feature extraction network to obtain a first feature map corresponding to the auxiliary picture set and a second feature map corresponding to the detection picture set; processing the first feature map and the second feature map by using a regional suggestion network to obtain a defect part predicted in the detection picture set; determining the similarity between the defect part marked in the auxiliary picture set and the defect part predicted in the detection picture set; and outputting a detection result according to the similarity. According to the technical scheme of the embodiment of the invention, the pipeline weld defect detection efficiency and detection effect can be improved.

The invention discloses a forced oscillation hierarchical positioning method based on multi-stage transfer learning. The method comprises an offline training part and an online positioning part, and comprises the steps of firstly, partitioning a power system according to the generator correlation, visualizing the smooth pseudo Wigner-Ville distribution of the oscillation principal components of all partitions, and forming an interval WVD image; performing the first-stage transfer learning on a pre-trained convolutional neural network to obtain a first-layer partition positioning model; and inputting a WVD image in a positioning subarea, and performing the second-stage transfer learning on the subarea positioning model to obtain a second-layer unit positioning model, and finally, verifyingthe offline positioning accuracy of the method. According to the present invention, the online positioning of a disturbance source is achieved by sequentially inputting an interval where the forced power oscillation actually occurs and the WVD image in the interval into a partition positioning model and a unit positioning model respectively. The method not only has the higher positioning accuracy,but also has the characteristics of high positioning speed, high adaptability, strong robustness and the like.

Theembodiment of that present application provide a neural network system, Among them, the neural network system, includes a clustering module for clustering a plurality of simulation results to determine a plurality of simulation clustering results, and determining a corresponding plurality of action sets according to the plurality of simulation clustering results, wherein each action set includes at least one action sample, and the simulation result is a simulation result of a target entity after the action sample is executed; The reinforcement learning module is configured to select an action set to be executed by the target entity from the plurality of action sets according to a state set of the target entity and output the selected action set to be executed by the target entity. The scheme provided by the embodiment adopts the reinforcement learning module of the action set training neural network system, and the order of magnitude of the action set is greatly reduced compared with the action sample, so that the training time of the neural network system can be shortened, the speed of finding the optimal solution in the training process can be increased, and the accuracy of the training result can be improved.

The invention discloses a platform patent value evaluation method based on an information ecological theory and an RF-GA-BP neural network. The platform patent value evaluation method comprises the following steps: analyzing network platform patent value influence factors based on the information ecological theory; obtaining a training sample and a test sample of the GA-BP neural network; performing dimension reduction processing on the platform patent value influence factors by adopting an RF-based feature selection method to obtain an optimal feature subset; designing a BP neural network topological structure, and setting related parameters; optimizing the weight and threshold of the BP neural network by adopting a GA algorithm; and taking the optimal feature subset selected by the RF asan input signal of a GA-BP neural network model, training and checking the GA-BP neural network, and finally constructing a platform patent value evaluation model based on an RF-GA-BP neural networkalgorithm.

The invention discloses a classification model training method, which is applied to the technical field of artificial intelligence and is used for solving the technical problems of high model training difficulty and inaccurate prediction result when an existing prediction model is used for predicting a low-risk probability event. The method provided by the invention comprises the following steps: acquiring risk training samples and non-risk samples; determining the number of base classifiers; determining a first training sample for training a first base classifier; binning the non-risk samples relative to the current base classifier according to the loss of each non-risk sample, and calculating the weight of each box under the current base classifier; according to the weight of each box and the number of the risk training samples; carrying out sampling from the corresponding boxes according to the determined number, and obtaining training samples of the current base classification; and training the first base classifier through the first training sample, and training the corresponding base classifiers through the training sample of the current base classification to obtain a trained classification model.

The invention discloses an image segmentation quality evaluation network system, method and system based on sorting constraints, belongs to the technical field of image segmentation, and achieves the precise prediction of the quality of segmented spectrums, especially the precise prediction of the quality relation between the segmented spectrums. The network system comprises a parameter-shared quality evaluation sub-network Q1 and a parameter-shared quality evaluation sub-network Q2, and the quality evaluation sub-network Q1 is a twin network and comprises two parameter-shared feature extraction branches C1 and C2, a feature conversion module and a quality prediction module; the feature conversion module fuses the first features extracted by the branches C1 and C2 and converts the first features into second features; the quality prediction module maps the second feature into a quality prediction value; and the quality evaluation sub-network Q2 and the quality evaluation sub-network Q1 have the same structure.

The invention discloses a lithium battery residual life prediction method comprising the following steps: 1, offline modeling, collecting lithium battery offline data, extracting a health factor sample set, using a random forest algorithm to carry out weight analysis on the health factor sample set, determining a selected health factor sample, and carrying out BiLSTM network model training to obtain a health factor model; optimal hyper-parameters of the model are selected through Bayesian optimization, and a prediction model is constructed; 2, on-line prediction: obtaining a health factor sample set through lithium battery on-line data and feature selection corresponding to an off-line stage; and predicting the service life of the lithium battery by using the prediction model in the step 1. According to the invention, while the prediction accuracy of the neural network is maintained, the number of parameters is reduced, the complexity of parameter training is reduced, the loss caused by failure of the lithium battery is reduced, the safety of the lithium battery is improved, and the problems of redundancy and insufficiency in selection of health factors of the lithium battery and selection complexity of different hyper-parameters of the neural network are solved.

The invention discloses a swimming athlete training load prediction method based on PCA-PNN. The method comprises the steps: step 1, carrying out feature index selection and database construction; step 2, carrying out principal component analysis to construct a PNN network input quantity; and step 3, constructing a PNN swimming athlete training load prediction model. The invention provides the swimming athlete training load prediction method based on PCA-PNN; the method constructs the swimming athlete training load prediction model through the research of the training load impact fusion features of an athlete, so as to evaluate the training arrangement of one swimming athlete.

The present invention disclosed a method and system of image depth estimation. The implementation of the methods includes: building an in -depth estimation network, the depth estimation network includes: coding part, convolution connection part and decoding part, the anti -volume layer and encoding part of the decoding part, the decoding partThe last layer of convolutional layers of the same convolutional block of the same scale is connected to form the final anti -convolutional layer; select the two continuous images in the sample image and the depth diagram of one of the images as the training sample, and use the training sample to use the training sampleTraining in -depth estimation network, get training in depth estimation network; collect testing images, extract the current frame image of the test image and the previous frame of the current frame image; enter the current frame image and the color channel of the previous frame image input training well.Estimate the network in depth to obtain the depth diagram of the current frame image.The method of the present invention is high efficiency, and the depth diagram has high accuracy and strong denseness.

The invention discloses a neural network generation method and device and a computer readable storage medium, and the neural network generation method comprises the steps: obtaining an optimal micro-unit, constructing a first network through the optimal micro-unit, enabling the first network to have enough powerful performance, and meeting the actual application demands; and training the first network by using a preset training data set to obtain a second network, establishing third networks, respectively training each micro-unit of all the third networks by using the second network to obtain a first training data set, training each micro-unit of all the third networks by using the second training data set, and constructing the neural network model according to the micro-units of the trained third network, so that the functions of all the micro-units of the third network correspond to the functions of the second network, and compared with a traditional method of training the micro-units one by one, the training number can be reduced, the computing power demand can be effectively reduced, and then the cost of generating the neural network model is reduced.

The invention particularly relates to an interference signal identification method based on a knowledge graph and Softmax regression, and the method comprises the steps: building a knowledge graph related to the identification of an interference signal, and embedding the knowledge graph of various interference types into a low-dimensional vector space; the knowledge contained in the knowledge graph is reserved, entities and relations in the knowledge graph are converted into vectors, the knowledge graph serves as priori knowledge to provide auxiliary information for the Softmax regression method, the model training speed is higher, the number of needed samples is smaller, and the recognition performance of interference signals under the low interference-to-signal ratio is further improved.

The invention discloses a forced oscillation hierarchical positioning method based on multi-stage transfer learning. The method includes two parts: offline training and online positioning. First, the power system is partitioned according to the generator correlation, and the smooth pseudo Wigner-Ville distribution of the oscillation principal components of each partition is visualized to form an interval WVD image. Then, the first stage of migration learning is performed on the pre-trained convolutional neural network to obtain the first layer partition localization model. Input the WVD image in the area of ​​the positioning partition, and then perform the second stage transfer learning on the partition positioning model to obtain the second-level unit positioning model. Finally, the offline positioning accuracy of this method is verified. The on-line location of the disturbance source is realized by successively inputting the WVD image of the section and the WVD image in the area where the forced power oscillation actually occurs into the partition location model and the unit location model. The present invention not only has high positioning accuracy, but also has the characteristics of fast positioning speed, high adaptability, strong robustness and the like.

The invention discloses a method for solving similar mathematical problems based on equations and questions, which comprises the following steps: (1) analyzing a corresponding relationship between numbers in the equations and numbers in the questions to form a universal equation of a single question; (2) marking question types, querying general equations of other single questions under the questions of the same type, and performing intersection calculation to form the general equations of the questions of the same type; and (3) inputting mathematical problems of the same kind, matching a universal equation, and solving a result. According to the invention, the general equation used by the question of the sentence pattern is determined by analyzing the relationship between the equation and the numbers in the question, and the question of the same kind of sentence pattern is solved by utilizing the general equation, so that a large amount of data is not required to be used as a training sample for training, and a general equation can be generated only by a small number of training samples; solving of similar questions is achieved, the data training amount is reduced while excessive computing resources are not consumed, the speed is high, efficiency is high, and implementation is easy.

The disclosure relates to a method and device for synthesizing singing audio frequencies, and belongs to the field of audio processing. The method comprises the following steps: acquiring a pre-storedaverage singing model; obtaining a target feature value text and a target acoustic parameter of target singing data; according to the target feature value text and the target acoustic parameter, determining a feature correspondence table of the feature value text of the target speaker and the acoustic parameters; on the basis of the target feature value text and the target acoustic parameters, training the average singing sound model to obtain a target singing sound model; and when a to-be-synthesized music score is received, extracting a feature value text of the to-be-synthesized music score, and on the basis of the feature value text of the to-be-synthesized music score, the target song model and the feature corresponding table, obtaining a target song audio frequency corresponding tothe to-be-synthesized music score. By adopting the method and the device, the singing synthesis efficiency can be improved.

The invention discloses an identity confirmation method based on the overall end-to-end unsteady-state iris cognitive recognition. The method is as follows: step 1, collecting 2000 iris grayscale images; step 2, setting the eye image of the template tester Feature; Step 3, setting the iris quality feature of the template tester; Step 4, converting the template iris normalized image; Step 5, setting 32-bit template features; Step 6, repeating Steps 1 to 5; Step 7, Collect 1 iris grayscale image; Step 8, extract eye image information; Step 9, extract iris quality information; Step 10, extract template iris feature information; Step 11, test iris normalized image; Step 12, test iris features information; step thirteen, the extracted matching value of the template features of the template tester; beneficial effects: improve the integrity of the scheme, improve the fit between data feature expression and recognition, improve the accuracy of recognition and the usability of data.

The invention discloses an image classification method based on transforming dimensions of an observation matrix, which includes: using perceptual compression to sparsely encode images to obtain a data set composed of low-dimensional images, and dividing the data set containing labels into a training set and a test set Constructing an image classification network comprising an input layer, a hidden layer and an output layer, the hidden layer is a perceptron unit; the image classification network is at least two, respectively comprising different node number perceptron units; the training set is used as Input, train under the supervision of the label, and obtain the corresponding neural network image classification model after training; use the test set to verify the accuracy of the neural network image classification model for image classification, and select the one with the highest accuracy as the final neural network image classification Model; input the image to be tested, and output the predicted probability of the image classification result. The image classification method provided by the present invention can greatly improve model efficiency without reducing image classification accuracy.

The invention provides a static gait debugging method and system for a robot, an electronic device and a storage medium, and the method comprises the steps: a non-contact force simulation step: carrying out the non-contact force simulation debugging of a robot model through employing a pre-constructed initial motion node instruction of a static gait, and obtaining a first debugging instruction; a contact force simulation step: performing contact force simulation debugging on the robot model by using the first debugging instruction to obtain a second debugging instruction; and a real debugging step: carrying out real debugging on a real robot by using the second debugging instruction to obtain a target action node instruction of the static gait. According to the technical scheme, the technical problems that in the prior art, when gait debugging is conducted on the robot, the workload is large, the gait planning accuracy is low, and the debugging cost is high are solved.

The invention relates to a die profile local springback compensation method based on elliptical surface mapping drive. The die profile local springback compensation method based on the elliptical surface mapping drive comprises the following steps that firstly, a local springback distribution cloud map of a part is obtained; secondly, a drive coordinate system is established according to the cloud map, an elliptical sketch is drawn, an interior area of the elliptical sketch is defined as a springback compensation area, and a reference curved surface and a target curved surface are established according to the springback compensation area and the springback compensation amount; and finally, the die profile is driven to deform in a mapping manner with the reference curved surface and the target curved surface as the deformation criterion and with the Z-axis direction of the drive coordinate system as a deformation direction, and the local springback compensation of the product is completed. Compared with the prior art, the die profile local springback compensation method based on the elliptical surface mapping drive has the beneficial effects that the efficiency of local springback compensation of the profile of a stamping die is improved, the local springback compensation precision of the profile of the die and the curved surface quality of the compensated profile of the die are effectively guaranteed, the die research and study amount is reduced, and the manufacturing cycle of the die is shortened.

The invention relates to the technical field of intelligent traffic, in particular to a traffic flow prediction method based on similar time sequence comparison, which comprises the following steps of: 1, acquiring traffic flow data of a certain region according to a public data website, and processing the traffic flow data; 2, mining a regional flow period; 3, pre-training an encoder; coding is carried out by using a deep ResNet network from a spatial angle, flow graph features are captured from the spatial angle, and due to the fact that regional flow distribution of the same city function is similar, features of similar regions are drawn close to each other by using a multi-instance contrast learning method, so that the features are far away from features of dissimilar regions; 4, the pre-trained encoder is put into the flow prediction model for fine adjustment; and 5, the model is stored. Compared with the conventional traffic flow prediction, the method has the characteristics of less parameter quantity and training cost, obvious modeling effect, good prediction result and the like.

The invention relates to a video classification method and device, equipment, a medium and a product, and relates to the technical field of computers, and the method comprises the steps: generating a first-class classification result of a first-class video based on a classification network, and generating a second-class classification result of a second-class video based on the classification network; wherein the first type of video belongs to a first video set, the second type of video belongs to a second video set, and the video number of the first video set is greater than that of the second video set; determining a first loss function between the first-class video tag and the first-class classification result, and determining a second loss function between the second-class video tag and the second-class classification result; and training a classification network based on the first loss function and the second loss function, and classifying the target second-class videos to be classified based on the trained classification network. In this way, the classification precision of the classification network for the second-class videos with the small sample size can be improved.

The invention discloses a method for constructing an ICA‑CNN classification frame of resting-state complex fMRI data of patients and healthy people, belonging to the field of biomedical signal processing. The present invention takes the functional network of interest separated from resting-state complex fMRI data by ICA as the research object, uses 2D CNN learning features with fewer parameters, and realizes the classification of patients and healthy people; Data augmentation was performed on ICA results to solve the problem of fMRI data shortage. Compared with the existing 3D CNN network, it not only reduces the amount of training, but also improves the accuracy. For example, for the complex fMRI data collected in the resting state of 82 subjects, applying the DMN components extracted by ICA, the slice recognition accuracy rate is higher than that of 3D CNN (0.728vs 0.701), and the subject recognition obtained after the subjects' decision-making is accurate The rate was further improved (0.914vs 0.701).

The invention discloses an edge detection filter optimization method based on deep learning, and the method comprises the steps: collecting a plurality of to-be-detected object images, and recording the to-be-detected object images as a training image set; selecting to-be-detected features in each image in a framing manner and labeling the to-be-detected features to obtain labeled images; taking the first training image as an input image; carrying out the convolution on the input image, calculating the gradient of each pixel point and inputting the gradient into a Sigmid function for activation processing, and acquiring an output result graph; recording the output result image as a new input image, and repeating; obtaining a normalized result graph by utilizing a softmax function, and calculating loss matrixes MLoss and LOSS values of the normalized result graph and the annotated image; performing back propagation by using the loss matrix MLoss to obtain corrected edge detection filters of each layer; taking the next training image as an input image, and continuing repeating by using the corrected edge detection filters until the LOSS value converges. The method is better in edge detection stability, high in robustness and small in calculated amount.