122 results about "Model migration" patented technology

In large-scale deployments of speaker recognition systems the potential for legacy problems increases as the evolving technology may require configuration changes in the system thus invalidating already existing user voice accounts. Unless the entire database of original speech waveform were stored, users need to reenroll to keep their accounts functional, which, however, may be expensive and commercially not acceptable. Model migration is defined as a conversion of obsolete models to new-configuration models without additional data and waveform requirements. The present disclosure investigates ways to achieve such a migration with minimum loss of system accuracy.

The invention discloses a crowd counting method based on a convolutional neural network. The method comprises the following steps of (1) after a training picture is marked, performing convolution operation with a Gaussian kernel to obtain a real crowd density diagram, and taking the real crowd density diagram as a label for model training; (2) inputting the training picture and the corresponding real crowd density diagram into a convolutional neural network model for performing training, and optimizing iterative updating parameters every time until the model is converged; (3) creating a new scene data set, and finely adjusting the obtained model by utilizing a model migration method, thereby finishing the model training; and (4) performing performance evaluation and testing on the trainedmodel. By utilizing the method, the number of parameters needed to be trained for the model is reduced; the model structure is simplified; the real-time performance of the model is improved on the premise that the accuracy is guaranteed; and the requirements of actual application are met.

The invention relates to an infrared image target detection method based on deep transfer learning and an extreme learning machine. The infrared image target detection method comprises the following steps: training a visible light image target detection model, training on a visible light sample set D by using a maskrcnn two-stage multi-task detection architecture, inputting a mask mask into a neural network, and redefining a loss function of an overall network structure; based on a sample migration method, obtaining a data set of migration learning by expanding the distribution of a target domain, namely an infrared sample set T; based on a model migration method, taking a target detection model with high precision based on a visible light image as a pre-training model of the generated migration learning data set, and carrying out training by adopting the same framework as visible light target detection; adopting an extreme learning machine to replace a network full connection layer, so as to overcome the over-fitting phenomenon of small sample model migration training.

The present invention discloses a coronary artery calcified plaque detection method based on model migration. The method comprises: reading a coronary artery CT image in a training set, and extractinga candidate calcified plaque in the coronary artery CT image according to a medical imaging standard; carrying out a data enhancement operation on a candidate calcified plaque image; inputting the data enhanced candidate calcified plaque image into a full convolutional network model that has been trained by the natural image for training so as to obtain a detection model; reading the coronary artery CT image in a test set, and according to the medical imaging standard, extracting a candidate calcified plaque in the coronary CT image of the test set; and taking the candidate calcified plaque image as the input of the detection model, and obtaining a detection result whether each pixel belongs to the calcified plaque in an end to end manner. According to the method of the present invention,a small amount of training samples are used to train the model for detecting the coronary artery calcified plaques according to the characteristics that the convolutional neural network model can bemigrated in different areas.

The invention provides a comment emotion classification method and system based on deep hybrid model transfer learning. The comment emotion classification method comprises the following steps: Step S1, collecting a commodity comment and preprocessing a source domain data sample set of the commodity comment; Step S2, mapping the preprocessed data into a word vector; Step S3, pre-training the sourcedomain data sample set of the commodity review with the depth mixing model; Step S4, fine-tuning the depth mixing model for the target domain data sample set of the commodity review; and Step S5, classifying the emotion of the commodity comment in the target domain. Tthe training speed is fast and the training difficulty is low, the invention only need several rounds of training to obtain high classification accuracy, and also can obtain good classification effect when the data set with more noise or less quantity is trained, and has little dependence on the data set and good robustness. Theinvention also effectively improves the transferability, and achieves the purpose of improving the classification accuracy after the transfer learning.

The invention provides a face beauty prediction method based on multi-task transfer learning, and the method comprises the steps: building a multi-task face database, carrying out the feature learning, carrying out the feature fusion, and building a face beauty prediction model. The method improves the accuracy of face beauty prediction through improving the expression recognition and age recognition. In order to avoid over-fitting of a deep network trained by a small amount of sample data and insufficient computing equipment, a VGG, ResNet and GoogleNet backbone deep convolutional network isused as a shared feature learning network structure, model migration is used, and the trained convolutional network is used to train a migratable shared feature. Network parameters are shared among tasks in the training process, and shared characteristics are learned, so that the accuracy of single task learning of the network is improved. Through using the deep learning network for multi-task learning, the shared representation layer can enable the tasks with universality to be better combined with the correlation information, and the task specific layer can independently model the task specific information.

The invention discloses a rolling bearing fault diagnosis method for improving a model migration strategy, and belongs to the technical field of rolling bearing fault diagnosis. The method is providedfor solving the problem of large distribution difference of data in the same state in a source domain and a target domain, and comprises: obtaining time-frequency spectrums of vibration signals of different types of bearings through wavelet transform, and constructing an image data set; selecting data of a certain model as a source domain, and selecting data of other models as a target domain; training a ResNet-34 deep convolutional network by using the source domain data to obtain a source domain data classification model; adaptively determining a migration knowledge level and knowledge content by using implicit gradient meta-learning to realize improvement of a model migration strategy and avoid a phenomenon that a gradient in a heterogeneous system structure is not easy to converge; introducing the migrated knowledge into a target domain ResNet-152 convolutional neural network data training process, and realizing model migration through parameter transmission; and optimizing network parameters by adopting a stochastic gradient descent algorithm when the source domain network and the target domain network are trained, and establishing fault diagnosis models of different types ofrolling bearings.

The invention discloses a local model migration learning-based gear fault recognition method. The method comprises time-frequency domain feature extraction, auxiliary data set selection in migration learning and migration learning on the basis of a local model. The method comprises the following steps of: calculating a similarity between target data and auxiliary data through establishing a Wilcoxon signed rank test and chi-square test combined model on the basis of a given time-frequency domain extraction characteristic, and screening the auxiliary data; and migrating useful common parametersof the screened auxiliary data to the target data by utilizing a local migration model taking an SVM as a core so as to realize fault recognition of a gearbox. According to the method, the diagnosisprecision of machine learning when less target data exists is enhanced, the diagnosis cost is reduced, the environmental suitability and universality of gear fault diagnosis are strengthened, and potential economic value is provided.

The invention provides a mobile robot visual following method based on deep reinforcement learning. A framework of "supervised pre-training of simulated images + model migration + RL" is adopted. Themethod comprises the following steps: firstly, collecting a small amount of data in a real environment, and automatically expanding a data set by adopting a computer program and an image processing technology, so as to obtain a large amount of analog data sets capable of adapting to a real scene in a short time for supervising and training a direction control model of a following robot; secondly,building a CNN model for robot direction control, and carrying out supervised training on the CNN model by using an automatically constructed analog data set to enable the CNN model to serve as a pre-training model; and then, migrating knowledge of the pre-training model to the DRL-based control model, so that the robot executes a following task in a real environment. By combining a reinforcementlearning mechanism, the robot can follow while improving the direction control performance in an environment interaction process, so that the robustness is high, and the cost is greatly reduced.

The invention discloses a model-transfer-based large-sized new compressor performance prediction rapid-modeling method, which comprises the following steps: determining a rated value of each parameter and a stable running interval on the basis of a performance prediction model for an existing similar compressor by utilizing the prior experience knowledge of a new / old compressor; designing an experiment to acquire a small number of experimental data samples, performing normalization processing on the acquired samples according to rated running parameters of a new compressor, establishing a performance prediction model for the new compressor by utilizing an ELM (Extreme Learning Machine) neural network, performing transfer learning, and performing model transfer training by using experimental sample input data and a predicted output value of the basic model as input variables of the new model and using experimental sample output data as the output of the new model; testing the effectiveness of the new model by using the experimental samples. According to the method, the performance prediction model for the new compressor can be rapidly developed under the condition of less experimental data information by virtue of the performance prediction model for the existing similar compressor and the prior knowledge of the new compressor, so that the modeling efficiency and accuracy are improved.

A video action segmentation method based on a hybrid time convolution and cycle network presented by the invention mainly includes a model structure, model migration and variation, and model parameter setting. The method includes the following steps: an encoder composed of a convolutional layer, an activation function and a pooling layer, a decoder composed of an up-sampling layer and a long short term memory network, and a Sofmax classifier are designed; an original video frame signal is processed by the encoder to get an intermediate layer result; and the result is input to the decoder, processed and transmitted to the classifier to segment, identify and classify a video action. Video signals compressed at different degrees can be processed. A hybrid time network is provided to solve the problem of video action segmentation. The accuracy and efficiency of action content identification are improved.

The invention relates to an industrial data classification method based on model migration. The method comprises: collecting source domain data and target domain data respectively; performing data enhancement on the source domain data; constructing a convolutional neural network with a residual structure; establishing a loss function to minimize the difference of cross-domain learning feature covariances, and minimizing domain displacement by aligning second-order statistics of distribution of source domain data and target domain data on a feature level; and training and predicting the model.According to the method, other similar data are used for learning to carry out feature migration under the conditions that the target data are few and the data are difficult to obtain, and then the target domains are classified, so that the method has a relatively high application value.

The invention discloses a model configuration method and device, an electronic device and a readable storage medium. The method comprises the steps of acquiring a first machine learning model, and acquiring one or more second machine learning models, wherein each second machine learning model is constructed according to corresponding equipment and / or platform; migrating the capability of the firstmachine learning model to each second machine learning model to obtain one or more trained second machine learning models; and configuring the equipment and / or the platform corresponding to each trained second machine learning model according to the trained second machine learning model, so that the equipment and / or the platform run the trained second machine learning model. According to the method, the mass data and the machine learning ability learned by a large machine learning model are embedded into various different devices by using a model migration technology, so that the performanceof a plurality of large machine learning models can be migrated to a small machine learning model, and the actual problem in life is solved.

The invention discloses a soft measurement method for the temperature of a rotary kiln burning zone. The method comprises the following steps that: taking a rotary kiln head temperature model and a kiln tail temperature model as basic models, and carrying out model migration to obtain a rotary kiln burning zone temperature model to predict the temperature of the burning zone.

The invention discloses a named entity identification method and device, equipment and a storage medium, and relates to the fields of natural language processing, semantic analysis and understanding,artificial intelligence and the like. The method comprises the steps of performing entity identification on new domain text data to obtain new domain seed entity words; labeling the new domain text data according to the new domain seed entity words to obtain labeled new domain text data; training a named entity recognition model by using the labeled new domain text data to obtain a named entity recognition model suitable for the new domain; and identifying entity words in other text data of the new domain by utilizing the named entity identification model suitable for the new domain. Accordingto the embodiment of the invention, the data annotation workload can be reduced, the model migration training threshold is reduced, and the field universality of the algorithm is improved.

The invention discloses a model migration method and electronic equipment, and relates to the technical field of machine learning. The specific implementation scheme is as follows: analyzing a first data file of a learning model on a first platform to obtain a network structure of the learning model; wherein the network structure comprises N nodes, and N is a positive integer; according to the topological sequence of the N nodes, mapping the N nodes to a second platform in sequence to obtain M nodes, M being a positive integer, and M being larger than or equal to N; and generating a second data file of the learning model on the second platform according to the M nodes. According to the technical scheme, the learning model of the first platform can be migrated to the second platform, codesdo not need to be rewritten, the learning model migrated to the second platform does not need to be retrained, the recoding and retraining time is saved, and the migration efficiency of the learning model is improved.

The invention discloses an intelligent vehicle end-to-end decision method and system orienting expressway scenes. The system concretely comprises the following modules including a pre-training decision module, an intelligent vehicle end-to-end decision system framework module and an intelligent vehicle end-to-end decision system test module. An initial training network model is trained by using atraining sample set; a pre-training initial model is obtained; a trained decision model is loaded in the end-to-end decision system for calculating a steering wheel rotating angle value; the system has high stability; the predicted steering wheel rotating angle value is stable; the stable running of an intelligent vehicle on the practical expressway is ensured; during turning, an intelligent vehicle can well fit with a reference curve; great deviation occurrence like a convolutional network cannot occur; and a pre-trained space-time characteristic fusion network has certain capability of predicting the steering wheel rotating angle. By using a model migration method, the decision network is directly migrated into the expressway scene; much time can be saved; and the network model trainingfrom the beginning is not needed.

The invention provides a mixed domain Fourier finite difference prestack depth migration imaging method based on the coefficient optimization in order to improve the earthquake migration imaging accuracy of a complex high steep structure region and forcefully guide accurate prediction and exploitation of oil-gas resources. According to the method, wave field extrapolation operators are expanded through a pade approximation rational function, expanded coefficients are optimized through a Chebyshev polynomial, new wave field extrapolation operators are obtained trough deduction, relative errors between the new wave field extrapolation operators and wave equation accurate wave field extrapolation operators are reduced, the approach degree of the wave field extrapolation operators is increased, and the earthquake migration imaging accuracy of the complex high steep structure region is improved while computational efficiency is guaranteed. The imaging effect of the improved mixed domain Fourier finite difference migration method on the Marmousi model migration section is compared with that of a conventional Fourier finite difference migration method on the Marmousi model migration section, and it is proved that the mixed domain Fourier finite difference prestack depth migration imaging method obtained based on coefficient optimization has higher migration imaging accuracy. The method has important significance in improving the earthquake migration imaging accuracy of the complex high steep structure region.

The invention provides a capsule model Chinese word segmentation method based on multi-regularization combination. The capsule model migration is applied to a natural language processing NLP sequencelabeling task, namely a Chinese word segmentation task, by adding a capsule sliding window capsule split window, so that the technical problem that the capsule model is not suitable for the sequence labeling task is solved; A plurality of regularization items are combined to realize simple field migration, and a capsule model is adapted to a sequence labeling task to complete Chinese word segmentation with higher accuracy and help a more complex natural language processing task; Through combination of multiple regular items, the generalization ability of the model is improved, certain field migration is achieved, manual corpus labeling can be reduced, and the labor and time cost of manual corpus labeling during natural language processing research is reduced.

The invention discloses a brain-computer interface lead optimization method based on independent component analysis. The method includes the following steps: building ICA based on a single test sampledesign ICA-MIBCI system for selecting high quality single test data from the EEG training set to ensure subsequent ICA-MIBCI system design reliability. On the basis, a new EEG lead optimization strategy is adopted to automatically optimize EEG leads for different BCI users. The invention is applied to ICA-MIBCI system can not only complete EEG lead selection based on specific person quickly and accurately, but also alleviate the impact of low-quality training data on MIBCI performance. At the same time, the advantages of ICA spatial filtering method in real EEG source acquisition and location, model migration and training data acquisition can be better utilized, and the stability and practicability of ICA-MIBCI system can be further effectively improved.

The invention discloses a specific pedestrian clothing analysis method and system based on fashion picture migration. The method comprises the steps that first, a model is initialized by training a fashion dataset; second, the model is used for analyzing a monitoring image containing pedestrians, and stripe constraint is introduced to optimize the analysis result; and last, a weak tag in a monitoring dataset is adopted to solve a model migration problem through an example-based migration learning method. Through the method, clothing information in a fashion picture set can be obtained and applied to the monitoring field through migration learning in combination with weak supervision, the information acquired through clothing analysis in the fashion field can be reserved, meanwhile, the clothing analysis effect can be improved and monitored, and data annotation workload is greatly reduced. In quantitative and qualitative tests of the actual monitoring dataset, the validity of the methodis proved.

The invention discloses a model migration training method, device and equipment and a storage medium, and relates to the field of artificial intelligence. The specific implementation scheme is as follows: taking network parameters of at least two migration layers in a source model as initial parameters of associated migration layers in a target model; constructing a target function according to the distances between the training parameters associated with the at least two migration layers and the initial parameters; and training a target model including initial parameters based on the target function. According to the embodiment of the invention, the target function is constructed; the distance between the training parameters and the initial parameters of the migration layers is introduced, so that the model migration and training conditions of each migration layer are considered in the model training process, the inheritance of information of the source model and the self-adaptation of the target model are realized, the over-fitting phenomenon in the model migration training process is avoided, and the generalization ability of the target model is improved.

The invention discloses a battery life prediction method, a system and a device based on model migration. The method comprises the following steps: establishing a battery accelerated aging model according to the experimental data in the accelerated aging test process of the battery; Bayesian Monte Carlo method was used to establish the battery slow aging model according to the accelerated aging model and the experimental data of the battery slow aging process. Input battery aging time to predict the battery life; the system includes battery accelerated aging model building module, battery slowaging model building module and prediction module. The apparatus includes a memory and a processor. The invention reduces the time cost of battery life prediction, improves the accuracy of predictionresult, and can be widely applied in the battery testing technical field.

The invention relates to a text verification code recognition method and device based on cross-domain element learning and a storage medium. The method comprises the steps of (1) an element training stage: firstly, generating a large number of verification code pictures with different security features as basic training data; then, carrying out character segmentation, and inputting segmented characters into the ResNet neural network model for feature extraction; and finally, obtaining a loss value of the pre-estimated category; and (2) a fine tuning stage: marking a small number of verification code pictures of different types from the basic training data in the meta-training stage, and performing fine tuning on the ResNet neural network model to obtain a final recognition result. The method has the characteristics of extremely small marked sample size, high model training speed, strong generalization ability and high recognition accuracy, solves the problems that an existing verification code recognition method needs a large amount of labeled data and the model migration difficulty is large, can meet the industrial requirements, and has wide application prospects.

The invention belongs to the field of spectrum detection and spectrograph model transmission, and provides a spectrum model transfer method based on a CNN-SVR model and transfer learning. The method comprises the following steps: acquiring and preprocessing spectral data, dividing the processed data into a training set and a test set, constructing a main instrument CNN-SVR model, inputting the main instrument training set into the model for training and optimizing to obtain an optimal main instrument CNN-SVR model and hyper-parameters thereof; migrating the model to a slave instrument, freezing a CNN network hyper-parameter value, inputting a slave instrument training set to train and finely adjust SVR parameters, obtaining a migration model based on the CNN-SVR network, and inputting a slave instrument test set into the migration model to predict the model transmission performance. The invention can automatically extract the essential characteristics of the high-dimensional wavelength variables, is suitable for small sample spectrum prediction, and realizes the transmission of the spectrum model among different spectrum instruments by using the characteristics of transfer learning.

The invention discloses a machine tool response modeling method, a modeling system and a response prediction method based on transfer learning. The modeling method comprises the following steps: training a source domain response prediction model by utilizing source domain data; a self-adaptive layer is added to the source domain response prediction model, parameters of the source domain response prediction model are reversely adjusted with the target that a loss function is smaller than a preset value, a domain adaptation initial model is obtained, and the loss function comprises classification loss and domain adaptation loss; inputting the target domain data into the domain adaptation initial model for fine tuning to obtain a domain adaptation model; inputting the source domain data into the domain adaptation model to obtain auxiliary training data; and training the target domain response prediction model by using the auxiliary training data and the target domain data. According to the method, model migration and sample migration are combined, multiplexing of source domain data is achieved, the demand quantity of model establishment for new data under the new working condition is reduced, and therefore the experiment cost of data collection for various different working conditions in actual production is reduced.

The invention discloses a text classification method based on semi-supervised transfer learning. The method comprises the following steps of: (1) data set and data preprocessing: acquiring a small amount of marked data sets and a large amount of unmarked data sets, cleaning and denoising, vectorizing data set samples through a word2vec method, and selecting 100 vector dimensions; (2) data enhancement: performing K times of text enhancement on each sample in the unmarked data in an anti-translation mode; (3) pseudo label pre-judging: inputting a labeled sample into a pre-training model Bert, and carrying out model migration by adopting a fine tuning method; (4) mixing samples; and (5) text classification: using the trained best model for carrying out text classification prediction on data in a test set. According to the text classification method based on semi-supervised transfer learning, semi-supervised learning is combined, transfer learning is used for solving the problem that annotation data are difficult to obtain in the field of text classification, and meanwhile the performance of a text classification model can be improved.