85 results about "Unsupervised learning algorithm" patented technology

A method for learning a classification model by a computer system is disclosed. One or more positive class data and one or more negative class data are prepared. The classification model is trained based on the positive class data to adjust one or more parameters of the classification model so that the positive class data is reconstructed by the classification model. The classification model is trained based on the negative class data to adjust the one or more parameters so that the negative class data is prevented from being reconstructed by the classification model. For the negative class data, changes in the one or more parameters with gradient of an objective function may be calculated using an unsupervised learning algorithm. The one or more parameters may be updated based on the changes in an opposite manner to the training based on the positive class.

The present invention provides a portable multi-channel depression tendency evaluation system based on emotional stimulation task. In a mood stimulation task, Electroencephalogram (EEG), electrodermatogram (ECG), electrocardiogram (ECG), eye movement (EOM), speech and image information were collected by portable equipment. Supervised learning and unsupervised learning algorithms were used to extract features from multi-channel physiological signals and behavioral data, to screen features, to train and validate models, to integrate the results of multi-channel analysis, to calculate depressionpropensity index (DPI) and to evaluate depression propensity.

A calibration method for calibrating at least one gamma radiation detector includes a monolithic scintillation crystal. The calibration method comprises obtaining event data for a plurality of scintillation events. The event data for each scintillation event includes a plurality of location sensitive signals observed by the at least one gamma radiation detector to be calibrated, applying an unsupervised learning algorithm to embed the event data on a low-dimensional manifold, and obtaining calibration data considering the low-dimensional manifold embedding.

The embodiment of the invention discloses a machine learning method and a related device, and the method comprises the following steps: obtaining a training set and a testing set, enabling the training set to be unlabeled data, and enabling the testing set to be labeled data; mapping all the training samples in the training set to an embedding space through an unsupervised learning algorithm, clustering all the training samples by adopting a preset clustering algorithm to obtain a plurality of categories, and enabling each category to correspond to one pseudo label; configuring a plurality oftarget tasks for the training samples according to the pseudo tags; running a meta-learning algorithm based on the plurality of target tasks to train the training samples corresponding to the plurality of categories to obtain a target neural network model with a target learning mechanism; inputting the test set into a target neural network model for training to obtain updated parameters; optimizing the model parameters of the target neural network model according to the updated parameters to obtain the optimized target neural network model. By adopting the embodiment of the invention, the performance of the neural network model can be improved.

Embodiments of the present specification disclose a data processing method, apparatus and device for insurance fraud identification, and a server. The method may comprise: acquiring relationship association data of a target population, wherein the target population comprises a set of selected applied claimants and insurants; constructing multi-degree relationship network map data for the target population based on the relationship association data; clustering the multi-degree relationship network map data by using an unsupervised learning algorithm to determine a classified communities of thetarget population and corresponding community information; and calculating community characteristics of the classified communities based on the community information, and marking the classified community that meets the insurance fraud decision-making condition as a fraud group. According to the embodiments of the present specification, a group of fraudsters can be effectively and efficiently identified from a multi-degree relationship.

The invention discloses a group fraud recognition method and device of internet finance small / micro loans. The method includes the following steps: acquiring feature information of loan application users; defining relationships among the users through comparison of similarity among the feature information of the users; building a relationship network of the users on the basis of the defined relationships among the users, wherein one user represents one node, and users with a relationship are connected through a line; using an unsupervised learning algorithm to find all connected branches in the relationship network; distinguishing a group as an abnormal suspicious group if the number of users in the group exceeds a set threshold value; and detecting the abnormal suspicious group to furtheranalyze entity features of the abnormal suspicious group, and carrying out further investigation thereon if it is analyzed that fraud suspicion of the abnormal suspicious group is higher. The invention aims to build the relationship network of the users on the basis of features of the users of the internet finance small / micro loans, and effectively identify a group with the fraud suspicion according to the relationship network.

The present invention discloses a face recognition algorithm configuration based on unbalance tag information fusion. The face recognition algorithm configuration comprises two layers of configurations (L1 and L2); the L1 is configured to train the face data and corresponding tag information to obtain an initial face recognition model 1 through adoption of a supervised learning algorithm, train no-tag data through adoption of an unsupervised method, alternately optimize the face data tag information and the parameters of the model 1, and obtaining a final face recognition model 1 through calculation after multiple iterations; the L2 has a thought opposite to the L1, the L2 is configured to initiate parameters of a face recognition model 2 at random, then perform unsupervised training to update the parameters of the model 2, input tag data, continuing training through adoption of the supervised learning algorithm, and finally obtaining the face recognition model 2; and the model 1 and the model 2 are fused to obtain a final face recognition model. Through combination of the advantages of a supervised learning algorithm and an unsupervised learning algorithm, the face recognition algorithm configuration based on unbalance tag information fusion gives full play to mass of no-tag data to allow the algorithm to have excellent recognition capability in a special scene and adapt different scenes.

The invention discloses an image classification method based on a multilayer spring convolutional neural network, and relates to the field of image processing, and the method comprises the steps: converting images in a training set into pulse sequences; setting a spring neuron parameter, and constructing a convolutional neural network by using a spring neuron; taking the pulse sequence as input, training the convolutional neural network layer by layer, obtaining visual features of the pulse sequence, and obtaining a classification result, the training method being an unsupervised learning algorithm based on a Hebbian rule; converting the to-be-identified image into a pulse sequence, and inputting the pulse sequence into the trained convolutional neural network to obtain a classification result of the to-be-identified image. The technical problem of sparking neuron membrane voltage redundancy calculation caused by increased neuron scale is solved, and meanwhile, the technical problem oflearning non-convergence caused by the fact that an SNN activation function cannot be derived and a back-propagation calculation residual error cannot be used is solved.

The invention discloses an unsupervised learning scene feature rapid extraction method fusing semantic information, which belongs to the technical field of image processing. The method mainly solves the technical problem of image feature description in the scene recognition problem. In view of the problem of serious interference of unstable information in an image on a scene matching effect and the problem of poor robustness of binarization feature descriptors to severe environmental changes, accurate scene semantic features are extracted through a semantic segmentation model obtained througha weighted model fusion strategy to guide the detection of a key region containing specific information; on the basis of the region, a screening strategy based on pixel point position clues and an unsupervised learning algorithm are respectively adopted to extract a binarization feature descriptor with high discrimination capability, so that the scene matching precision can be improved while the calculation complexity is reduced.

The related unsupervised learning algorithm to on-line treat speech signal for speech interactive front-end process comprises: with sufficient respect to time-domain relevance and non-stationary of natural speech signal, constructing a group of adaptive feed-forward filter networks to separate speech signal from multiple sources. This invention enhances speech signal, reduces algorithm complexity for hardware implementation, and has wide application.

The invention provides a position recommendation method, comprising the following steps S1: acquiring user access data and generating a data matrix; 2, executing a supervised learning algorithm on that data matrix; 3, executing an unsupervised learning algorithm on that data matrix and generating prediction result data; 4, generating position pre-recommendation data according to that predict result data and the character portrait data. The invention has the beneficial effects that the system automatically learns, completes user analysis, and accurately recommends work information according tothe visitor's equipment information, access content, operation behavior, etc. through the machine learning method of post pre-recommendation. It enables users to find more highly relevant and interesting position information, and improves user viewing content, so as to have a higher user conversion rate.

The invention discloses an online train identification and speed estimation method based on optical fiber vibration signals, and belongs to the field of the data-driven fault diagnosis. The method comprises the following steps of: (1) obtaining the eigenvalue energy entropy for data classification by short-time Fourier transform after performing smooth filtering on the fiber vibration signals collected by each sampling point; (2) classifying eigenvalue of each sampling point online to determine whether there is a suspected train signal after calculating a threshold offline by the unsupervisedlearning algorithm; (3) finding signals that satisfy a train characteristic model in the suspected train signals by modeling characteristics of the train; and (4) obtaining the real-time position andthe speed of the train by performing the improved global piecewise polynomial fitting on the train signals. According to the online train identification and speed estimation method based on optical fiber vibration signals, the problem of fault diagnosis and tracking for unlabeled data under an interference condition is solved based on the fiber vibration signals.

A method of unsupervised learning of a metric representation and a corresponding system for a mobile device determines a metric position information for a mobile device from an environmental representation. The mobile device comprises at least one sensor for acquiring sensor data and an odometer system configured to acquire displacement data of the mobile device. An environmental representation is generated based on the acquired sensor data by applying an unsupervised learning algorithm. The mobile device moves along a trajectory and the displacement data and the sensor data are acquired while the mobile device is moving along the trajectory. A set of mapping parameters is calculated based on the environmental representation and the displacement data. A metric position estimation is determined based on a further environmental representation and the calculated set of mapping parameters.

The invention discloses a vehicle collision detection method based on machine learning. The vehicle collision detection method based on machine learning comprises the following steps: acquiring a velocity vector, an acceleration vector, an angular velocity vector and longitude and latitude; preprocessing the acquired speed data respectively; calculating new vector data through the preprocessed data and acquiring a road scene label; composing the calculated data into an input vector; calculating collision probabilities under different models through input vectors and calculating comprehensive collision probabilities; acquiring the category of the input vector through a preset method and judging whether the comprehensive collision probability value and the category of the input vector are abnormal or not; and marking the input vector as a collision vector or a non-collision vector. By integrating supervised learning and unsupervised learning algorithms, the cost of machine learning is reduced; collision detection is carried out by using a plurality of trained classifiers, and collision detection under a plurality of collision scenes is covered by adopting a mode of deep mining of a plurality of dimensions and a plurality of collision scenes, so that the accuracy rate and the recall rate of collision detection are improved.

In example implementations of the present disclosure, processing of a specification and / or other parameters generates a NoC with flows that meet specification requirements. In example implementations, the specification is processed to determine the characteristics of the NoC to be generated, the characteristics of flow (e.g. number of hops, bandwidth requirements, type of flow such as request / response, etc.), flow mapping decision strategy (e.g., limit on number of new virtual channels to be constructed, using of existing VCs, yx / xy mapping), and desired strategy to be used for how the flows are to be mapped to the network. In such processing, the machine learning algorithm can provide a determination as to if a flow is acceptable or not in view of the specification (e.g., via a Q score). In example implementations, the machine learning decisions can be applied on a flow by flow basis, and can involve supervised learning and unsupervised learning algorithms.

The invention discloses an intelligent law article recommendation auxiliary system based on a FastText algorithm. The intelligent law article recommendation auxiliary system comprises a judgment document data set, a word vector text library, law word vectors, a document vector text library and a text classification model, wherein the judgment document data set is used for storing judgment documents; the word vector text library is used for storing text segment categories, and the document vector text library is used for storing text segment categories; each text segment category is the contentin the judgment document; the law word vector is obtained by taking a pre-trained general word vector as an initial vector and training the word vector text library by using a FastText unsupervised learning algorithm; and the text classification model is obtained by taking a trained law word vector as an initial vector and performing text classification on a document vector text library by usingthe FastText supervised learning algorithm. Applicable laws are comprehensively and accurately recommended for case description, a new thought is developed for providing judicial assistance work for artificial intelligence, and text classification models are established for various types of crimes such as cheating crimes, robbery, economic crimes or marriage dispute cases.

The invention provides an unsupervised learning detection method for receiving signal change caused by an underwater invasion target. Data needing to be tested are preprocessed to serve as training data, a binary tree is sequentially established for the training data from a root node, the number of nodes which the testing data averagely pass through is counted, abnormal scores of the testing pulsedata are calculated, and whether receiving signal abnormal changes caused by forward scattering signals exist or not is judged; an unsupervised learning algorithm is adopted, weak receiving signal abnormity caused by forward scattering of the underwater invasion target is detected, various kinds of known information are fully utilized while the time-varying environment influence is weakened, andgood detection performance is achieved; and compared with an existing method, the method has the advantages that the forward scattering signals can be detected in real time, and compared with a conventional iForest method, the method is more remarkable in effect of detecting sound field changes caused by forward scattering of a target.

The invention discloses a model training method, a model-based red envelope material laying prediction method and a device thereof. The model training method comprises the following steps: acquiring at least one scanning code receiving red envelope data and at least one scanning code payment data; according to the related information of at least one setting merchant, each piece of acquired data being respectively extracted with features to obtain a training sample corresponding to the setting merchant, wherein the training sample takes the extracted eigenvalue as an input value; a prediction model being obtained by training the obtained training samples with an unsupervised learning algorithm, wherein the prediction model takes the eigenvalues of the code scanning red packet data and the eigenvalues of the code scanning payment data as input values and the probability of laying the red packet materials by a business as output values.

The invention provides an electronic examination system comprising an examination question storage module, an identity confirmation module, a wrong question analysis module, an examination point obtaining module, an examination question extraction module and a compliance detection module. The identity confirmation module receives and verifies the examinee's identity information. The wrong questionanalysis module obtains the error-prone examination points corresponding to the identity information according to the body information. The examination point obtaining module obtains the preset examination points corresponding to the examination paper according to the identity information. The examination question extraction module uses the unsupervised learning algorithm to cluster the preset examination points into different groups according to similarity based on the error-prone examination points and extracts the corresponding examination questions from the examination question database to generate the paper for different groups. The compliance detection module performs statistics of the difficulty of the examination paper of different examinees in the same group and re-extracts the examination questions to generate the test paper according to the test paper whose difficulty deviates from the average examination paper difficulty and exceeds the preset value so as to at least partially solve the problems of high complexity in automatic generation of the examination paper, the lack of personalized assessment, one-sided assessment points no compliance detection.

The present application discloses an automatic test paper generation method, an electronic device and a storage medium. The method includes receiving and verifying the identity information of the examinee; Obtaining an error-prone test point corresponding to the identity information according to the identity information; Acquiring a preset examination point corresponding to the test paper according to the identity information; According to the error-prone test points, the preset test points are clustered into different groups according to similarity by using an unsupervised learning algorithm.According to the different grouping, the corresponding test questions are extracted from the test question bank to generate test papers. In order to at least partially solve the problems of high complexity in the automatic generation of test papers, lack of personalization of the assessment, and one-sided assessment points.

The invention relates to an automatic enterprise data clustering method and system. The automatic enterprise data clustering method comprises the steps that relevant data of enterprises is obtained; the relevant data of the enterprises is preprocessed and stored; text vectors are obtained from the processed relevant data of the enterprises; the text vectors are classified and clustered according to a K-Means algorithm. The accuracy and reliability of data sources are ensured by obtaining massive reliable relevant data of the enterprises, the enterprises are preprocessed by utilizing an unsupervised learning algorithm based machine learning technology, data storage is performed by utilizing HAFS, and automatic and reliability clustering is ensured. A natural language is adopted for data analysis, the text vectors and weights corresponding to the text vectors are obtained, the text vectors are automatically clustered from the massive data based on K value optimization and classificationmodels of specialists, it is ensured that the clustering supports the demand meeting actual applications, and automatic and available clustering is ensured.

The embodiment of the invention provides a submarine motion model simplification method and device. The method comprises the steps of constructing a six-degree-of-freedom motion model of a submarine;selecting a maneuvering motion performance index corresponding to a certain operation condition, and calculating a sensitivity index of the maneuvering motion performance index to each hydrodynamic coefficient in the six-degree-of-freedom motion model of the submarine; clustering each hydrodynamic coefficient in the six-degree-of-freedom motion model of the submarine by using an unsupervised learning algorithm according to the sensitivity index to obtain a clustering result of each hydrodynamic coefficient; and carrying out hydrodynamic coefficient simplification on the six-degree-of-freedom motion model of the submarine according to the clustering result to obtain a submarine motion simplification model. According to the embodiment of the invention, the unsupervised learning algorithm isadopted to cluster each hydrodynamic coefficient in the submarine motion model, and the simplification of the submarine motion model is realized based on the clustering result, so that the operation difficulty and complexity of measurement of the hydrodynamic coefficients can be reduced, and the simplified model has higher accuracy.

The invention provides an underwater intrusion-induced sound field abnormality real-time detecting method. The underwater intrusion-induced sound field abnormality real-time detecting method comprisesperiodically transmitting single-frequency pulse signals through a transmitting end, preprocessing received data at a remotely arranged receiving end into training data, establishing a binary tree for the training data to calculate abnormality scores, and according to the abnormality scores, determining whether sound field abnormality caused by forward scattered signals exists to achieve receiving mutation detection on signals received when a target penetrates through a receiving-transmitting connecting line. The underwater intrusion-induced sound field abnormality real-time detecting methodapplies an isolated forest unsupervised learning algorithm to achieve detection of weak sound field abnormality caused by forward scattering of an underwater intruding target and meanwhile to be highin detecting performance. Compared with existing methods, the underwater intrusion-induced sound field abnormality real-time detecting method can achieve real-time detection on forward scattering signals, require no prior information except the training data; by updating the training data, the underwater intrusion-induced sound field abnormality real-time detecting method can adapt to different application environments without adjusting the algorithm.

A method for learning a classification model by a computer system is disclosed. One or more positive class data and one or more negative class data are prepared. The classification model is trained based on the positive class data to adjust one or more parameters of the classification model so that the positive class data is reconstructed by the classification model. The classification model is trained based on the negative class data to adjust the one or more parameters so that the negative class data is prevented from being reconstructed by the classification model. For the negative class data, changes in the one or more parameters with gradient of an objective function may be calculated using an unsupervised learning algorithm. The one or more parameters may be updated based on the changes in an opposite manner to the training based on the positive class.

A method includes sampling input / output requests from a host to generate sampled input / output requests; classifying the sampled input / output requests into clusters using an unsupervised learning algorithm; determining a hot data range based on a characteristic of the clusters; and incorporating the determined hot data range into a hot data table.

The invention discloses a method for normalizing a URL (Uniform Resource Locator), and aims to solve the problems existing in an existing URL normalization method. The method comprises the following specific steps: step 1, encoding an original URL into a numerical vector through a deep learning method, so that the distances of URLs with the same path and different parameters in a vector space after encoding are very close; and step 2, combining URLs with similar numerical vectors so as to achieve the purpose of normalization. According to the method, a complex regular expression does not needto be compiled, the parameter part can be accurately identified no matter how long the parameter part is, and the URL can be accurately normalized; according to the method, an Autoencoder method is adopted, the Autoencoder method is an unsupervised learning algorithm, and manual annotation is not needed; according to the method, a URL mapping table or a directory structure does not need to be maintained, and better stability is achieved when new URLs appear when small-scale revision is conducted on the website.

The embodiment of the invention provides a method and system for detecting the abnormal state of a railway overhead line system bolt. The method comprises: acquiring an image of a to-be-detected railway contact network bolt; and based on the trained image generation model and the multiple spatial mapping model, carrying out anomaly detection analysis on the to-be-detected railway contact network bolt image to obtain a railway contact network bolt abnormal state detection result. According to the embodiment of the invention, based on an unsupervised learning algorithm, the problems of poor generalization, difficulty in defining abnormal categories and great reduction of accuracy of unseen abnormal samples of an existing detection algorithm are solved, and the unsupervised learning algorithmdoes not need to define the abnormal categories in detail, so that the data annotation amount is greatly reduced, and the working intensity of detection personnel is reduced.

The invention provides an inhaul cable time-variant cable force process recognition data driving method based on acceleration monitoring. According to the method, by applying an efficient unsupervised learning algorithm (complexity degree tracing) and utilizing the monitoring information of multi-channel acceleration sensor distributed on an inhaul cable, real-time identification of cable force time interval can be realized; The complexity degree tracing algorithm can be used for automatically decomposing an acceleration response of the inhaul cable into a single-mode response of the inhaul cable, the real-time frequency of the inhaul cable is identified by virtue of the acceleration information of a quick short time period, and the cable force time interval is calculated by virtue of a tensioning chord theory; proven by simulation analysis on a stayed-cable bridge with actually-measured cable force and actually-measured wind speed and the model test of the inhaul cable, the provided complexity degree tracing algorithm can be used for performing accurate real-time recognition on a time-variable cable force process. The method provided by the invention is a direct and effect time-variable cable force process recognition method, and is simple and easy to use, high in cable force recognition accuracy, high in aging property and capable of realizing online real-time recognition.

The invention relates to the technical field of complex pipe network leakage positioning, in particular to a complex pipe network leakage positioning method based on a deep belief network, and the method is implemented according to the following steps: 1, obtaining a training sample and a test sample comprising a monitoring point pressure value and a leakage position coordinate; 2, constructing a complex pipe network leakage positioning model based on a deep belief network; 3, pre-training the complex pipe network leakage positioning model by adopting a layer-by-layer unsupervised learning algorithm according to the test sample; 4, performing parameter optimization on the pre-trained complex pipe network leakage positioning model by adopting a BP algorithm; step 5, performing leakage positioning on the test sample in the step 1 by using the complex pipe network leakage positioning model obtained in the step 4, and outputting a positioning result; and calculating the diagnosis accuracy of the model. According to the method, monitoring of the whole pipe network can be completed only by establishing the complex pipe network leakage database training network, and the diagnosis efficiency of pipe network leakage is greatly improved.