32results about How to "Avoid "negative transfer"" patented technology

The present invention provides a rolling bearing fault diagnosis method in various working conditions based on feature transfer learning, and relates to the field of fault diagnosis. The objective ofthe invention is to solve the problem that a rolling bearing, especially to various working conditions, is low in accuracy of diagnosis. The method comprise the steps of: employing a VMD (VariationalMode Decomposition) to perform decomposition of vibration signals of a rolling bearing in each state to obtain a series of intrinsic mode functions, performing singular value decomposition of a matrixformed by the intrinsic mode functions to solve a singular value or a singular value entropy, combining time domain features and frequency domain features of the vibration signals to construct a multi-feature set; introducing a semisupervised transfer component analysis method to perform multinuclear construction of a kernel function thereof, sample features of different working conditions are commonly mapped to a shared reproducing kernel Hilbert space so as to improve the data intra-class compactness and the inter-class differentiation; and employing the maximum mean discrepancy embedding to select more efficient data as a source domain, inputting source domain feature samples into a SVM (Support Vector Machine) for training, and testing target domain feature samples after mapping. Therolling bearing fault diagnosis method in various working conditions has higher accuracy in the rolling bearing multi-state classification in various working conditions.

The invention discloses an adversarial-learning-based multi-source-domain adaptive migration method and system. The method comprises: step one, pre training is carried out by using all-source-domain data and a representation network and a classifier of a target model are initialized; step two, multi-path adversarial adversarial processing is carried out on multi-source-domain data and target-domain data and a representation network and a multi-path discriminator of the target model are updated; step three, adversarial scores between the source domains and the target domain are calculated; stepfour, target domain classification is carried out based on the classifiers and the adversarial scores of all source domains; step five, a target domain pseudo sample with a high confidence coefficient is selected for fine tuning of the representation network and the classifier of the target model; and step six, the steps from the step two to the step five are carried out again until model convergence is realized or a maximum iteration number of times is reached, and then training is stopped. According to the invention, reliance on the hypothesis of consistency of the single-source-domain tagset and the target domain is eliminated; and a negative migration phenomenon existing in the multi-source domain adaptation process is avoided effectively.

The invention discloses a far infrared pedestrian detection method for changed scenes. A sample extension target data set is screened out of auxiliary data on the basis of the Boosting-style inductive transfer learning algorithm DTL Boost. At first, a sample importance measurement model based on the k-nearest neighbor is utilized for evaluating the similarity between the auxiliary data and the target data, and corresponding initial weights are distributed for different samples in the auxiliary data. In the training process, the prediction inconsistency degree of member classifiers is explicitly defined, iterative updating is carried out on the current weights of the auxiliary data and the target data sample through the prediction error rate of the current member classifiers, a sample extension training set with the positive transfer ability is screened out of the auxiliary data, and the different member classifiers are encouraged to learn different parts or aspects of the target data. In this way, an integrated classifier with the stronger generalization ability is obtained through training, and the robustness of pedestrian detection in the new scene is enhanced.

The invention relates to a task splitting and mitigating learning prediction method based on multi-source domains. The goal of this method is to divide the domains involving the characteristic items of a target domain into corresponding sub-domains under the condition that a complex target task is difficult to train and learn with an extremely limited amount of target domain data so as to divide out the sub-tasks corresponding to the sub-domains. Through the domain characteristic amount corresponding to the sub-tasks, the sub-task model weights are calculated; through the use of the characteristic items required in the sub-domains corresponding to the sub-tasks, the characteristic mapping manner is utilized to establish the sharing characteristic space so that in ample multi-source domain sample, the sub-task models could be trained out; the associated characteristic item sharing parameter models are extracted to integrate initial target task models to be rapidly fit in combination with a gradient raising method to realize the task splitting and mitigating prediction. The invention serves as a cross-discipline model mitigating prediction method using the task splitting manner.

The invention belongs to the technical field of artificial intelligence, and relates to an unsupervised domain adaptive method combining deep attention features and conditional adversarial. The methodcomprises the following steps: dividing a to-be-processed image data set into a source domain and a target domain; designing a network capable of migrating attention and conditional confrontation; preprocessing the image source domain and the target domain before the image source domain and the target domain are inputa network capable of migrating attention and conditional adversarial; importingthe preprocessed source domain and the preprocessed target domain into the designed network in batches in sequence, obtaining weighted feature maps through a migratable attention network, inputting the weighted feature maps into a conditional adversarial network for training, and finally performing probability operation through a full connection layer; respectively calculating the image classification accuracy of the source domain and the target domain; and finally, directly applying the network which is trained on the source domain and can migrate attention and conditional adversarial to thetarget domain to perform image classification through iteration and back propagation training. According to the method, the generalization ability of the unsupervised domain adaptive network is greatly improved.

The invention discloses a face beauty prediction method and device based on adversarial transfer learning. The face beauty prediction method comprises the following steps: screening a face beauty prediction model with the highest correlation from a plurality of auxiliary tasks for recognizing face factors through similarity measurement, and constructing the first human face beauty prediction modelaccording to the face beauty prediction model; migrating universal characteristic parameters formed after adversarial network pre-training to a second face beauty prediction model; and inputting a to-be-detected face image to realize recognition. The training cost of pre-training is reduced, and negative migration caused by auxiliary tasks with unrelated factors is reduced; and through adversarial transfer learning, the calculation amount of retraining of the second face beauty prediction model is reduced, and the effect of obtaining a more accurate model by using fewer training images is achieved.

The invention belongs to the technical field of side-scan sonar image recognition, and particularly relates to a side-scan sonar image target classification method based on style migration. Accordingto the method, a saliency detection method and a style migration network are used, a conventional optical image is converted into a side-scan-imitating sonar image, the distance between a source fieldand a target field is shortened, the number of basic features capable of being directly migrated is increased, and migration learning efficiency can be effectively improved; meanwhile, a migration learning method is used for migrating a fully-trained deep learning network, and the characteristic that the basic characteristics of the image have universality is utilized, so the number of optimization parameters can be effectively reduced, and the problem that a deep learning technology cannot be applied due to insufficient training samples is avoided. According to the method, a style migrationand transfer learning method is used for migrating a convolutional neural network trained by using the artificially generated side-scan-imitating sonar image, so transfer learning efficiency is improved, and the phenomenon of negative migration is prevented.

The invention discloses a personalized federated meta-learning method for data isomerism. The method comprises the following steps: determining an automatic encoder structure in an initialization stage of each client and a meta-model structure in a personalized stage; initializing parameters of a federation training stage; grouping the clients according to the local data distribution vectors uploaded by the clients; aggregating the client models in each group, and issuing the aggregated client models to the clients in the group to carry out the next round of iteration; and after federation training is finished, the client performs fine tuning on the meta-model in the group and local data thereof to generate a personalized model. According to the method, when the clients participate in federation training, the clients with approximate data distribution are dynamically divided into the same group according to the local data distribution vectors uploaded in each round, and the corresponding meta-model is set for each group, so that the problems of slow model convergence and low accuracy caused in a highly heterogeneous data environment are solved.

The invention relates to a method for quickly mixing a high-order attention domain adversarial network based on transfer learning. The method comprises the steps of designing a quick mixing high-orderattention and domain adversarial adaptive network for a to-be-processed image data set; preprocessing the source domain and the target domain; importing the preprocessed source domain and the preprocessed target domain into the designed network in batches in sequence, obtaining weighted feature maps through a fast mixing high-order attention network, then inputting the weighted fine feature mapsinto a domain adversarial adaptive network for training, and finally performing probability operation through a full connection layer; respectively calculating average image classification accuracy ofthe source domain and the target domain; enabling a gradient inversion layer in back propagation to take a reverse gradient direction to form adversarial training, then performing iterative training,and applying a fast mixed high-order attention and domain adversarial adaptive network trained on a source domain directly to a target domain to carry out image classification. According to the invention, the recognition rate and migration capability of the unsupervised domain adaptive network in migration learning are improved.

The invention discloses a short-term generalized load prediction method based on transfer learning, and the method comprises the following steps: constructing a short-term load prediction integrated model, and carrying out the analysis of a prediction error of the short-term load prediction model; solving the weight by using an algorithm based on iteration and cross validation; constructing a short-term load prediction model based on load time series decomposition and instance migration; based on the hidden variable model, constructing a public model for the target problem and the source problem; and designing a hidden variable extraction module based on the load affine curve. According to the method, the target of transfer learning is introduced into the short-term load prediction problem, the similarity between the source problem and the target problem is ingeniously utilized, the source problem data set is introduced to assist the training process of the target problem, and the target of improving the prediction effect of the target problem can be achieved; the prediction precision can be improved by utilizing the hidden variable model; through a hidden variable extraction module designed based on a load affine curve and based on the hypothesis, the calculation complexity can be reduced.

The present invention provides a navigation path planning method based on strategy reuse and reinforcement learning, and belongs to the technical field of navigation path planning, which solves the problem of insufficient reuse of the source strategy in the existing method. According to the method provided by the present invention, a function representing state importance is introduced to assist strategy selection, strategy reuse and strategy library reconstruction, so that the purpose of rapidly planning a navigation path in a road network map is achieved; compared with the existing traditional path planning method, a reinforcement learning algorithm based on strategy reuse is used in the algorithm ARES-TL, the complete strategy library is updated in real time, the algorithm time is savedby occupying some space storage strategy library, and the reinforcement learning algorithm can cope with the online micro-updated map; compared with the same type of strategy reuse method, by using the algorithm ARES-TL of the present invention, the negative migration caused by the reuse of the irrelevant source strategy is avoided with respect to PRQL and OPS-TL, and exploration efficiency is improved and navigation tasks are accurately completed; and the method provided by the present invention can be applied to the technical field of navigation path planning.

The invention discloses a cross-subject EEG fatigue state classification method based on generative adversarial domain self-adaptation. The method comprises the following steps: firstly, acquiring andpre-treating data, and removing artifacts; secondly, performing EEG feature extraction through PSD, and acquiring a two-dimensional sample matrix from a three-dimensional EEG time sequence; distinguishing a source domain data set and a target domain data set to obtain a training set and a test set which are not overlapped; training a classification model GDANN by using part of label-free target domain data and random data conforming to Gaussian distribution; and finally, evaluating the accuracy of the classification result by using a confusion matrix. The generative adversarial network and the domain invariant thought are further combined, the problem that EEG signal data sets are rare and difficult to obtain is solved, the problem that source domain data and target domain data are not matched is balanced, negative migration is avoided to a certain extent, a high-precision cross-subject fatigue detection classifier is trained, and the method is expected to have a wide application prospect in actual brain-computer interaction.

The invention discloses a multi-virtual power plant decentralized self-discipline optimization method. The method comprises the following steps: S1, initializing parameters; S2, classifying tasks andforming an initial knowledge matrix; S3, acquiring information; S4, determining an optimization individual action; S5, calculating an objective function value of each agent; S6, calculating a rewardfunction; S7, updating the knowledge matrix; S8, information feedback: each agent returns the current optimal solution to an information center; and S9, judging whether the maximum number of iterations is reached or not, and if yes, outputting an optimal knowledge matrix of the corresponding task, otherwise, returning to S3. By applying the multi-virtual power plant decentralized self-discipline optimization method, the technical problems that the existing distribution network regulation and control cannot meet the requirement that a plurality of virtual power plants participate in the power market in real time for profit-by-profit, and the grid connection behavior of distributed equipment is effectively controlled to support the safe and effective operation of the distribution network aresolved.

The invention discloses a domain adaptive support vector machine generation method, belongs to the field of domain adaptive learning, and solves the technical problem of how to fully mine generality information of data between domains in domain adaptive learning and overcome dependency on tags of source domain data and target domain data. The method comprises the following steps of generating synthetic data through an over-sampling algorithm based on the source domain data and the target domain data; based on the synthetic data and the target domain data, mining a shared implicit characteristic space between the synthetic data and the target domain data according to a probability distribution integral mean square error minimum principle; and based on the target domain data, performing training on an expanded characteristic space consisting of the shared implicit characteristic space and an original characteristic space to generate a domain adaptive support vector machine. The generality information of the source domain data and the target domain data can be fully mined. According to the method, the generality information of the source domain data and the target domain data can be mined without depending on the tags of the source domain data and the target domain data.

The invention belongs to the field of collaborative filtering recommendation, and provides a cross-regional cross-score collaborative filtering recommendation method and system, and the method comprises the steps: dividing all users in a target domain score matrix and a source domain score matrix into active users and inactive users, and dividing all items into hot items and non-hot items; decomposing the target domain scoring matrix and the source domain scoring matrix, and extracting user implicit vectors and item implicit vectors in the target domain and the source domain; aiming at active users and hot items, respectively learning a mapping relation of user implicit vectors and item implicit vectors corresponding to a target domain and a source domain under two scoring systems; utilizing the mapping relation between the user implicit vectors and the item implicit vectors of the active users and the hot items to obtain non-active user and non-hot item characteristics on the target domain; and constructing a limited matrix decomposition model according to the inactive user and non-hot item characteristics on the target domain, predicting the score of any user on any item, and selecting the item with the highest predicted score as a recommendation result of the user.

The invention provides a cross-domain medical care equipment recommendation method and system with a privacy protection function. The method comprises the following steps: acquiring ID data of a target domain about users and items of medical health care equipment; performing one-hot coding representation on the user ID data and the project ID data of the target domain; expanding the one-hot coding representation of the target domain by using the auxiliary domain user domain dependence feature and the auxiliary domain user domain non-dependence feature to obtain an expanded feature; and using the extended features and the trained factorization machine model to predict the score of the target domain user to the item, and recommending the medical health care equipment according to the score result. The method can ensure that other platforms cannot reversely derive the original score information according to the disclosed user potential factor vector, and can effectively solve the problem of original data privacy leakage of the user in the auxiliary domain data migration process.

The invention discloses a method and a device for training a transfer learning model and a recommendation model. The method comprises the following steps of clustering a source domain sample and a target domain sample to obtain a clustering result, determining the weight of the source domain sample according to the clustering result, wherein the weight of the source domain sample is used for representing the similarity between the source domain sample and the target domain sample, determining similar samples of the target domain samples from the source domain samples according to the weights of the source domain samples so as to form training samples containing the similar samples and the target domain samples, and training the transfer learning model according to the training sample.

A multi-source community label development trend prediction method based on the A-TrAdaboost algorithm, comprising the following steps: (1) constructing a label network in the question-and-answer community; (2) calculating the vector representation of the community structure in the source domain and the target domain (3) Calculate the structural similarity between the source domain and the target domain; (4) Construct a multi-source community tag popularity prediction model based on the A‑TrAdaboost algorithm. The present invention uses the network graph representation method to obtain the vector representation of each network, and then calculates the similarity between networks as the domain distance between different question-and-answer communities, and uses the similarity between network structures as a multi-source migration learning algorithm The initial weight of TrAdaboost can better avoid the negative migration problem in multi-source migration when predicting the popularity of new tags across communities, and improve the training time and accuracy of the model.

The invention discloses a crowd gathering sensitive image detection method, which comprises the steps of: collecting image data in the gathering crowd, respectively obtaining image data sets of the original domain and the target domain; The transfer model of the common subspace and the sparse expression reconstruction matrix are learned at the same time, and the supervised discriminant regularization term is added to the transfer model; the common subspace of the classification task in the original domain and the target domain is obtained by using the image sample label information, and in the common subspace The optimization process is realized through transfer learning; the alternate optimization strategy and ADMM algorithm are used to alternately solve the optimization variables; sensitive images are identified. The present invention utilizes the establishment of a supervision and discrimination sparse transfer model to improve the accuracy of classification tasks in the case of small samples, thereby greatly improving the accuracy and discrimination of sensitive image recognition in complex environments such as crowd gatherings.

The invention discloses a laser welding method and device and a storage medium thereof. The method comprises the steps that a bad factor sample set in the laser welding process is collected as a target data set; an existing sample set with the highest similarity with the target data set is determined as a source data set; pre-training is continuously conducted on a deep learning model by using thesource data set to obtain a first pre-training model; structure adjustment is performed on the first pre-training model by using the target data set to obtain a second pre-training model; and parameters of the first pre-training model are migrated to the second pre-training model to obtain a final model. The source data set which has the highest similarity with and is matched with the target dataset can be confirmed, and negative migration is avoided; and the defect that a small amount of sample data is prone to over-fitting is overcome in a transfer learning mode, and the model convergencespeed is increased.

The embodiment of the invention discloses a figure image recognition method and device, electronic equipment and a computer readable medium. A specific embodiment of the method comprises the following steps: acquiring a figure image; performing feature extraction on the figure image through a feature extraction network included in a preset image recognition model, obtaining a global feature map, wherein the preset image recognition model further comprises a joint recognition network, the joint recognition network comprises a deconvolution branch network group, and the deconvolution integral branch network group is used for generating a figure joint thermodynamic diagram group; inputting the global feature map into the joint recognition network to obtain a figure joint thermodynamic diagram group; and generating a figure image recognition result based on the figure joint thermodynamic diagram group, and sending the figure image recognition result to a display terminal for display. According to the embodiment, the accuracy of generating the character image recognition result can be improved.

Embodiments of the present disclosure disclose a person image recognition method, device, electronic equipment and computer readable medium. A specific implementation of the method includes: acquiring a person image; performing feature extraction on the above-mentioned person image through a feature extraction network included in a preset image recognition model to obtain a global feature map, wherein the above-mentioned preset image recognition model also includes joints A recognition network, wherein the joint recognition network includes a deconvolution branch network group, and the deconvolution branch network group is used to generate a character joint heat map group; the above global feature map is input into the joint recognition network to obtain a character joint heat map group; generating a person image recognition result based on the above-mentioned person joint heat map group, and sending the above-mentioned person image recognition result to a display terminal for display. This implementation manner can improve the accuracy of generating a person image recognition result.

The invention discloses an image classification method based on unsupervised domain adaptation, which includes: selecting a labeled source domain image data set and an unlabeled target domain image data set, and performing data augmentation; constructing an anti-transfer network to reduce source domain images The difference between the conditional distribution of the data set and the target domain image data set; constructing a time series integration network to regularize the prediction results of the image labels in the target domain image data set; combining the confrontation migration network and the time series integration network to construct an image classification model; The final image data set is used as the training set, and the image classification model is trained by means of meta-learning; the trained image classification model is used to identify the target image to be classified and complete the target image classification. In the process of classifying massive image data on the Internet, the present invention greatly reduces manpower labeling data without affecting the accuracy of image classification, and users can quickly and accurately search for desired images from massive image data.

The invention provides a discussion mining system and method based on multi-task sparse shared learning, and the system comprises an encoder module which is used for learning context information through employing a bidirectional long-short-term memory neural network; a double-path attention coding module used for carrying out feature extraction on word vectors by using self-attention and external attention in parallel to obtain word semantic attention degrees from different angles and strengthening relation modeling between words; a sparse shared learning module used for carrying out multi-task learning on the encoding module for obtaining the sentence vector, generating a task-specific sparse parameter matrix for different tasks so as to solve the negative migration influence of multi-task learning, and obtaining sentence-level encoding representation; and a multi-task label output module used for completing classification result prediction of different tasks by using a task-specific classifier. A sparse shared structure of a plurality of tasks can be automatically learned, and the specific sub-networks of the respective tasks are utilized to perform joint training, so that the negative migration phenomenon of multi-task learning is effectively avoided.

The present invention provides a rolling bearing fault diagnosis method in various working conditions based on feature transfer learning, and relates to the field of fault diagnosis. The objective ofthe invention is to solve the problem that a rolling bearing, especially to various working conditions, is low in accuracy of diagnosis. The method comprise the steps of: employing a VMD (VariationalMode Decomposition) to perform decomposition of vibration signals of a rolling bearing in each state to obtain a series of intrinsic mode functions, performing singular value decomposition of a matrixformed by the intrinsic mode functions to solve a singular value or a singular value entropy, combining time domain features and frequency domain features of the vibration signals to construct a multi-feature set; introducing a semisupervised transfer component analysis method to perform multinuclear construction of a kernel function thereof, sample features of different working conditions are commonly mapped to a shared reproducing kernel Hilbert space so as to improve the data intra-class compactness and the inter-class differentiation; and employing the maximum mean discrepancy embedding to select more efficient data as a source domain, inputting source domain feature samples into a SVM (Support Vector Machine) for training, and testing target domain feature samples after mapping. Therolling bearing fault diagnosis method in various working conditions has higher accuracy in the rolling bearing multi-state classification in various working conditions.

The invention discloses a classification model training method combining active learning and transfer learning. The classification model training method mainly comprises the following important steps: 1) transmitting source task knowledge to a target task model in a mode of selecting training samples for the target task model by adopting a source task model; 2) the source task model and the target task model actively select a certain proportion of samples for training the target task model; and (3) the source task model selects samples with high certainty, the target task model selects samples with high uncertainty, and the relative proportion of the number of the samples selected by the source task model and the number of the samples selected by the target task model is dynamically determined by the relative advantages and disadvantages of the classification performance of the two models. According to the method, negative migration is avoided, the method is suitable for the field needing data safety / privacy protection, the collected target task training sample set is high in quality, and learning is more efficient. Meanwhile, the number of training samples needed for training the target task model is reduced, the problem of imbalance of the training samples is relieved, and knowledge migration between heterogeneous models can be achieved.

The present invention discloses a multi-source domain adaptive migration method and system based on confrontational learning. The method includes the following steps: step 1, using each source domain data for pre-training and initializing the representation network and classifier of the target model; step 2 , use multi-source domain data and target domain data to perform multi-way confrontation, and update the representation network and multi-way discriminator of the target model; Step 3, calculate the confrontation score between each source domain and target domain; Step 4, based on each source domain domain classifier and adversarial score to classify the target domain; Step 5, select high-confidence target domain pseudo-samples to fine-tune the representation network and classifier of the target model; Step 6, return to Step 2, and proceed to Step 2-5 until the model Stop training when it converges or reaches the maximum number of iterations, the present invention can no longer rely on the assumption that the label set of a single source domain is consistent with the target domain, and can effectively avoid the negative transfer phenomenon existing in the multi-source domain adaptation process.

This application discloses a decentralized self-discipline optimization method for multiple virtual power plants, including: S1, parameter initialization; S2, task classification and formation of an initial knowledge matrix; S3, information acquisition; S4, determination of optimal individual actions; S5, calculation of each agent S6, calculating the reward function; S7, updating the knowledge matrix; S8, information feedback: each agent returns the current optimal solution to the information center; S9, judging whether the maximum number of iterations is reached, and if so, output The optimal knowledge matrix for the corresponding task; otherwise, return to S3. This application discloses a distributed self-discipline optimization method for multi-virtual power plants, which solves the problem that the existing distribution network regulation is difficult to meet the real-time participation of multiple virtual power plants in the power market for profit, and effectively control the grid-connected behavior of distributed equipment to support the distribution network. Technical aspects of safe and efficient operation.

The present invention provides a navigation path planning method based on strategy reuse and reinforcement learning, and belongs to the technical field of navigation path planning, which solves the problem of insufficient reuse of the source strategy in the existing method. According to the method provided by the present invention, a function representing state importance is introduced to assist strategy selection, strategy reuse and strategy library reconstruction, so that the purpose of rapidly planning a navigation path in a road network map is achieved; compared with the existing traditional path planning method, a reinforcement learning algorithm based on strategy reuse is used in the algorithm ARES-TL, the complete strategy library is updated in real time, the algorithm time is savedby occupying some space storage strategy library, and the reinforcement learning algorithm can cope with the online micro-updated map; compared with the same type of strategy reuse method, by using the algorithm ARES-TL of the present invention, the negative migration caused by the reuse of the irrelevant source strategy is avoided with respect to PRQL and OPS-TL, and exploration efficiency is improved and navigation tasks are accurately completed; and the method provided by the present invention can be applied to the technical field of navigation path planning.