227 results about "Feature mining" patented technology

The invention discloses a high-efficiency anti-DDoS security gateway system, which can effectively detect and defend a DNS denial-of-service attack. The system comprises two core components, namely a detector and a filter. The system supports two deploying ways, namely serial connection and bypass. A detecting method is based on the self-learning network flow state statistic and characteristic disinterment, achieves the detection and positioning of the network abnormal flow by setting network performance parameter thresholds, and can effectively identify the suspected attack flow. A defending method carries out a thought of deep defense, and the system is deployed with two defense steps of attack characteristic defense and baseline defense so as to ensure the attack defending effect of thesystem under normal network condition and the basic defense capacity of the system in the individual and severe attack environment. The methods can effectively improve the security and attack resista nce of a DNS server and can ensure the normal operation of the DNS service.

The invention discloses a deep learning-based multi-target locating method for a power device infrared image. The method comprises the steps of 1) obtaining the standard power device infrared image through a substation device detection apparatus; 2) establishing a power device infrared image sample library, and extracting a training set, a verification set and a test set; 3) establishing a FASTER-RCNN deep target detection neural network, training the established FASTER-RCNN deep target detection neural network by using the training set of the sample library, and verifying the over-fitting degree of a model by using the verification set; and 4) by utilizing the network model built by training, performing multi-target identification and locating on the infrared image in the test set, and generating an identification result. According to the method, the input infrared image is subjected to deep feature mining by utilizing a deep learning algorithm, without depending on manual extractionof feature parameters, and the regions and positions of various power main devices can be effectively and accurately identified, so that the labor amount is reduced to a certain extent.

The embodiment of the invention provides a liver tumor segmentation method and device based on multi-stage CT image guidance, and the method comprises the steps: obtaining an abdomen CT image with enhanced contrast, inputting the abdomen CT image with enhanced contrast into a preset single-channel full convolutional neural network, and obtaining a liver region-of-interest image; inputting the liver region-of-interest image into a preset tumor segmentation network to obtain a tumor segmentation result corresponding to the contrast-enhanced abdominal CT image; wherein the tumor segmentation network is obtained by performing multi-channel fusion training according to liver region-of-interest image samples in different periods and image samples with tumor region marks corresponding to the abdominal CT image samples. According to the embodiment of the invention, the liver tumor CT images in different periods are effectively subjected to feature mining by adopting a multi-channel fusion training network, so that the trained network has higher segmentation precision and better robustness on liver tumors.

The invention discloses a construction method of a general medical special disease data system. According to the method, clinical text analysis and knowledge extraction are carried out to integrate two kinds of heterogeneous clinical data sources including unstructured medical records and structured clinical data; and then a general medical special disease data system for disease characteristic analysis is constructed. The method is characterized in that on the basis of a special disease segmentation dictionary, the analysis is carried out by combing contexts of medical records, the structuralization of the electronic medical records is completed; with the pattern matching and feature mining technologies, characteristic value of diseases in the medical records are obtained to complete analysis and digging of the semantics of medical records, so that disease condition features and disease condition prognosis conditions of diseases in the unstructured electronic medical records are provided and dug. The analysis and mining of the disease characteristics, disease outcomes, and the treatment process of the medical special diseases are completed by integrating the structured clinical data and thus a comprehensive complete medical data system is built.

Methods and systems of detecting tampering in a digital image includes using hybrid large feature mining to identify one or more regions of an image in which tampering has occurred. Detecting tampering in a digital image with hybrid large feature mining may include spatial derivative large feature mining and transform-domain large feature mining. In some embodiments, known ensemble learning techniques are employed to address high feature dimensionality. detecting inpainting forgery includes mining features of a digital image under scrutiny based on a spatial derivative, mining one or more features of the digital image in a transform-domain; and detecting inpainting forgery in the digital image under scrutiny at least in part by the features mined based on the spatial derivative and at least in part by the features mined in the transform-domain.

The invention discloses a method for digging identification characteristics of application layer protocol. The method comprises the following steps of: A, filtering firstly and coding a training data packet set, extracting standard protocol identification characteristic data information; B, performing a first digging to the extracted standard protocol identification characteristic data information to obtain a multistage frequent set; C, performing the first digging to the multistage frequent set, and correcting frequent degree of the rest multistage frequent set after the first digging, performing a second digging to obtain final protocol identification characteristics; D, if byte identification rate of all the final protocol identification characteristics meets the demand, or the total identification rate of the data packet meets the demand, no longer digging the data of the second and the subsequent data packets; otherwise, circularly digging the second and the subsequent data packets until the total identification rate meets the demand. The invention can analyze, dig the data packet set, and extract all the identification characteristics of the corresponding application layer protocol, which greatly improves characteristic extraction efficiency and the total identification rate.

The invention discloses a power grid monitoring alarm event identification method based on convolution and a long-term and short-term memory network, and the method comprises the steps: generating aninformation vector through historical monitoring alarm information and time marks in a power grid monitoring system, extracting event samples from the collected historical monitoring alarm information, and constructing an alarm event sample library; secondly, establishing a deep learning recognition model based on the combination of a long-term and short-term memory network and a convolutional neural network, and training the model by utilizing an alarm event sample; and finally, identifying the monitoring alarm information by using the trained deep learning model, and outputting the event category with the maximum probability as an identification result. According to the method, the excellent performance of the long-term and short-term memory network in time sequence problem processing and the excellent performance of the convolutional neural network in short text local feature mining are combined, the combined model is established, rapid identification of the power grid alarm event can be realized, the screen monitoring pressure of monitoring service personnel is effectively reduced, and the working efficiency of daily monitoring and accident exception handling is improved.

The invention discloses an IaaS (Infrastructure as a Service) cloud platform network fault positioning method and system based on log analysis. The IaaS cloud platform network fault positioning system comprises a fault injection module, a log acquisition and analysis module, a knowledge generation module and a fault detection and positioning module. Firstly, by injecting various typical network faults, various corresponding fault logs are formed; then aiming at various faults, log information related to network faults of each layer of physical resources, an operation system, a virtual machine, an OpenStack and the like is respectively acquired, and fault feature mining is carried out on the acquired network fault log information by using an Apriori algorithm; on such basis, according to a maximal frequent item set and parameters, such as a supporting degree, a confidence degree and the like, association rules and knowledge, which correspond to the specific network faults, are generated by utilizing a bayes formula; and finally, when a system has a network fault again, the network fault can be compared with the association rules of a knowledge base and analyzed according to an acquired fault log, so that the layer on which the network fault occurs is positioned.

The invention discloses an automatic webpage type identification method based on Web structure characteristic mining. The automatic webpage type identification method comprises the following steps that S1, a webpage source code set is obtained through a crawler system; S2, webpage source codes are preprocessed; S3, webpage characteristics are extracted; S4, a classifier is established by applyinga classification algorithm used in machine learning, and automatic webpage type identification is completed through the classifier. Before a webpage characteristic set is extracted, a depth-first traversal search strategy is adopted to search noise labels needing to be removed, the volume of a webpage is decreased, the number of labels to be processed is decreased, and the performance of extracting the webpage characteristic set is improved. An HTML document characteristic set is extracted from four aspects closely bound up with a webpage structure through Web structure mining, and then the classification algorithm used in machine learning is applied to establish the classifier so as to complete automatic webpage type identification. Compared with other webpage type identification methods,the automatic webpage type identification method has the advantages of being simple in concept, easy to achieve, convenient to popularize, good in universality and high in accuracy rate.

The invention discloses a wireless signal detection and electromagnetic interference classification system and method based on deep learning. The method comprises the steps of obtaining the observation data by using the frequency spectrum monitoring nodes deployed in a distributed manner; executing two types of signal feature mining in parallel based on the complex value observation data to obtaina wireless signal detection data set and an electromagnetic interference classification data set, training two groups of convolutional neural networks in parallel based on the two types of data sets,and then detecting wireless signals and executing electromagnetic interference classification by using the two groups of trained convolutional neural networks respectively. The system and the methodhave the beneficial effects that the accuracy of the wireless signal detection and the electromagnetic interference classification can be improved; the generalization singular value decomposition andthe space division are performed on two types of data sets, so that the additive noise can be eliminated, the crosstalk from adjacent channels can be inhibited, and the authenticity of data can be enhanced; and the wireless signal detection and the electromagnetic interference classification are carried out concurrently, so that the efficiency is high, and the response is fast.

The present invention solves problems of cold start, first rater, sparsity and scalability for recommendation. A recommendation system according to the present invention finds association rules through data mining. Then, the recommendation system integrates a rough-set algorithm and a statistical analysis prediction for recommendation. The recommendation is dynamically made from a result of the rough-set algorithm and a result of the statistical analysis prediction by setting a standard deviation as a threshold.

The invention discloses a weak supervised image semantic segmentation method based on spatial pyramid concealment pooling, which comprises the following steps: selecting a convolution neural network H, processing the input image X through the convolution neural network H to obtain a classification characteristic map; the spatial pyramid pooling module is established according to the classificationcharacteristic map, and then the spatial pyramid is concealed to obtain the output characteristic map. According to the output characteristic graph, the category activation vector and the category probability vector are calculated, and then the competitive spatial pyramid masking pooling loss function is established. The convolutional neural network H is trained according to the pooling loss function of competitive spatial pyramid concealment and the segmentation feature map is extracted. The invention realizes a weak supervised image semantic segmentation model with richer local features, more perfect region feature mining and more robust target size and posture, improves the extraction ability of local semantic information and strengthens the recognition ability of local targets or parts in the weak supervised semantic segmentation.

The invention discloses a remote sensing image ground object semantic segmentation method, and aims to improve remote sensing image ground object segmentation accuracy and solve the problem that edgerecognition is not fine enough. According to the technical scheme, the method comprises the steps that a pyramid scene analysis network is constructed, a network model with high image feature mining capacity is migrated to a semantic segmentation network model from the related field, and information contained in a remote sensing image is mined from the channel dimension; spectral information contained in the remote sensing image is mined in combination with a channel attention mechanism, and a data correlation type up-sampling module carries out up-sampling on feature maps of different spatialscales to the size of an original feature map and splices the feature maps with the original feature map; the risks of gradient disappearance and gradient explosion are effectively reduced by adopting a loss function tower, and the prediction effect of the image edge is further improved by adopting an IoU-based loss function; and a network model is trained by using the labeled training data, testset data is inputted into the optimized semantic segmentation network model, and different ground objects in the image are identified.

The invention relates to a distribution transformer fault diagnosis method with automatic feature mining and automatic parameter optimization. The distribution transformer fault diagnosis method comprises the following steps of: installing a vibration signal acquisition device on a distribution transformer, and acquiring a vibration waveform of the distribution transformer during operation from the vibration signal acquisition device; constructing a distribution transformer fault feature extraction model based on a stack auto-encoder after secondary tuning; extracting a vibration signal feature vector yn by using the stack auto-encoder after secondary tuning, labeling corresponding features, and establishing a database containing normal and various faults; segmenting a data set, namely splitting the data set into a training set and a test set according to the proportion of X1: X2; training a random forest classifier by using the training set; and based on the network parameters of thetrained random forest classifier, establishing a distribution transformer fault diagnosis model to realize distribution transformer fault diagnosis. According to the distribution transformer fault diagnosis method, the accuracy of fault diagnosis of the distribution transformer can be remarkably improved, and the method has good robustness and outstanding diagnosis performance.

The invention discloses a holographic city big data model and knowledge graph enterprise portrait construction method. The method comprises the steps of enterprise holographic data model construction,enterprise knowledge graph construction and enterprise label automatic extraction. Enterprise characteristics can be finely positioned through holographic portraits, potential enterprise relations are mined, any existing enterprise is described by constructing the enterprise portraits, and a way for sufficient cognition and comprehensive understanding of enterprises is provided for enterprise information demanders. The enterprise holographic portrait establishes an enterprise all-information database, and the natural person language processing and data mining technology is used to extract thelabel from the enterprise holographic data and the knowledge graph to draw the enterprise portrait, so that the problems of enterprise data dispersion and loss can be alleviated, and the enterprise all-dimensional information can be displayed.

The invention discloses a remote sensing image unsupervised change detection method based on a Siamese network structure. The remote sensing image unsupervised change detection method comprises the following steps: (1) initializing parameters; (2) obtaining a difference image; (3) performing pixel-level fusion on the two information complementary difference images in the step (2) by using an adaptive local energy weighting algorithm to obtain a new difference image; (4) adopting a clustering algorithm to realize pre-classification; (5) taking a pre-classification result as a label, and realizing the precise detection of an SAR image change region through a DFF-Siamese network; according to the method, unsupervised change detection of the SAR image is realized, priori knowledge is introduced into the deep convolutional neural network, feature mining is deeper by adding a layer-by-layer difference measurement module in the Siamese network, the learning ability of the network is effectively improved, and a more ideal change detection result can be obtained.

The invention discloses an abnormal telephone recognition method and system based on feature selection and integrated learning. The method comprises the steps of constructing a mixed data set; miningsample characteristics through historical call behaviors of the user in a window from the starting time to the ending time; combining and optimizing the characteristics based on the user call behavior, and mining the characteristics with behavior information from the aspects of time, frequency, short message, flow, position and contact person; performing oversampling based on user call behavior samples, increasing the number of few samples, and reducing the influence of sample imbalance on the model; performing feature dimension reduction processing on the user call behavior sample; and establishing a model by using the integrated learning training data set, and carrying out abnormal telephone identification. According to the method, the original information of the sample is fully restoredin a feature mining combination and dimensionality reduction mixing mode, so that the prediction precision is improved.

The invention relates to a fan blade early icing fault detection method based on a deep neural network pre-trained by an auto-encoder. The neural network comprises a fault feature mining and classification network, is used for solving the problem that the early icing state of a fan is difficult to detect in the prior art, and adopts the specific technical scheme that the method comprises the following steps: (1) acquiring a fan icing original data set; (2) preprocessing the original data set to obtain a training set and a test set; (3) pre-training the DNN layer by layer by using an auto-encoder, (4) determining a network structure, training a deep neural network model by using the training set, and optimizing and finely adjusting the model; and (5) carrying out early icing fault detectionon the fan blade by utilizing the trained model. According to the method, the influence of all data collected by the SCADA system on freezing of the fan blades is fully considered, fault detection ofearly freezing of the fan is achieved, and the detection accuracy reaches 98% or above.

The invention discloses a webpage text extracting method based on text tag feature mining. The webpage text extracting method comprises the following steps: S1, preprocessing webpage tags and repairing Html tags; S2, selecting and extracting Html tag features; S3, clustering and mining tag features and selecting a text cluster; S4, adjusting tags in the text cluster empirically; S5, extracting a tag text of the text cluster. In the webpage text extracting method, tags of webpage source codes are mined, the webpage tags are clustered by a hierarchical clustering algorithm, a cluster in which the text tag is positioned is extracted, the tags in the tag cluster is adjusted according to experience, and text is extracted according to the adjusted text cluster feature. Compared with other news webpage text extracting methods, the webpage text extracting method has the characteristics of higher universality, higher accuracy, easiness in use and no need of special settings for specific webpages.

The invention discloses a dynamic process monitoring method based on a latent variable autoregression model, and aims to establish the latent variable autoregression model and implement dynamic process monitoring on the basis of the latent variable autoregression model. Specifically, the method comprises the steps of defining a least square objective function of an autoregression model of a latentvariable, inferring a corresponding feature mining algorithm, and then establishing a fault monitoring model so as to implement online fault monitoring. According to the method disclosed by the invention, the dynamic autocorrelation latent variable is mined by establishing the target of the latent variable autoregression model, and the autoregression model meeting the least square condition is given correspondingly. Through the latent variable autoregression model, autocorrelation characteristics in original training data can be mined, and the influence of latent variable autocorrelation canbe eliminated. Therefore, the method provided by the invention is obviously different from the traditional dynamic process monitoring method, and the interpretability of the model is stronger. In other words, the method provided by the invention is a more preferable dynamic process monitoring method.

The invention discloses an electrocardiogram data pathological feature quantitative analysis method and device, relates to a method and device for quantitative extraction of dynamic multiple pathological features of a cardiac electrical activity system and an abnormal analysis method and device for cardiac electrical signals, and belongs to the field of cardiac disease data feature mining. The electrocardiogram data pathological feature quantitative analysis method and device are used for solving the problem of mining more and richer dynamic pathological features in electrical activities of acardiac nonlinear system. The key point of the electrocardiogram data pathological feature quantitative analysis method and device is that the step of extracting quantification index of inherent cardiac dynamic pathological characteristics of dynamic data of a cardiac electrical activity nonlinear system by the method of heterogeneity analysis is achieved, wherein the heterogeneity analysis refersto non-uniformity and complexity analysis of the dynamic data of the cardiac electrical activity nonlinear system in the process of spatial distribution and time deduction. The effect of the electrocardiogram data pathological feature quantitative analysis method and device is that the obtained information can present the dynamic pathological features which are difficult to measure by traditionalmethods of the cardiac electrical activity nonlinear system.

The invention relates to a multi-sensor fusion convolutional neural network aero-engine bearing fault diagnosis method. The method comprises the steps of S1, performing data acquisition; s2, performing data preprocessing; s3, taking the data acquired by the analogue simulation test platform as source domain data, and taking the data acquired by the online monitoring system as target domain data; s4, building a multi-sensor information fusion 1D-CNN model, and putting the source domain data into the source domain 1D-CNN model for training; s5, carrying out target domain bearing online diagnosis; s6, generating a fault diagnosis result. Vibration signals of different positions of aero-engine bearing in different fault states are collected, multi-channel input 1D-CNN model is adopted, data collected by vibration acceleration sensors at different positions are fused and put into model for training, and target domain bearing online diagnosis is carried out. Therefore, the fault diagnosis and analysis are carried out on the bearing of the rotating mechanical part of the aero-engine, so that fault type identification is accurately completed, and the process of manual feature mining in a traditional method is omitted.

The invention provides a trend prediction method based on a depth quantum neural network, which comprises the following steps: 1. constructing an initial depth quantum neural network; 2, periodicallycollecting vibration signals of the bearing, and carrying out feature mining on the vibration signals by using wavelet packet decomposition; 3, training the depth quantum neural network model by usingthe train set data, and evaluating the performance of the model by using the verification set data; preprocessing the collected signals and dividing the processed feature parameters into training data set and testing data set. 4, adjusting the parameters of the depth quantum neural network model, and selecting an optimal prediction model for performance evaluation through continuously training the model; 5, predicting the trend of the bearing by using the prediction model. Through the above steps, the trained depth quantum gods network can realize the trend prediction of the bearing, and thebearing can be maintained in time through the trend prediction of the bearing, the repair time can be shortened, the repair cost can be reduced, and the mechanical failure problem caused by the bearing maintenance delay can be solved.

The invention discloses a welding spot quality identification method fusing a knowledge graph and a graph convolutional neural network, and the method comprises the steps: photographing a welding spot, and obtaining a welding spot appearance image; the welding spot appearance image comprises a welding spot and a position visual feature of the welding spot; cutting the welding spot appearance image to obtain a welding spot cutting image; the sizes of all the welding spot cutting images are the same, and each welding spot cutting image only comprises one welding spot and the position feature of the welding spot; importing the cutting image of the welding spots into a fine-grained network for feature mining to obtain a visual feature matrix of the welding spots; establishing a knowledge graph according to the quality of the welding spots and the position relationship between the welding spots, and performing feature mining on the knowledge graph by using a graph convolutional neural network to obtain a high-dimensional spot type spatial feature matrix of the welding spots; and carrying out vector inner product on the visual feature matrix and the high-dimensional spot type spatial feature matrix to obtain a classification detection result of the welding spot quality.

The invention relates to the field of finance technology, and discloses a feature mining method and device for a classification label. The method comprises the steps: obtaining a trained target classification model, and enabling the target classification model to carry out the training through a feature variable of a training object and a classification label of the training object, and obtaininga corresponding model parameter; obtaining a characteristic variable of the target object; determining a classification result contribution degree of each feature by utilizing an interpretation modelaccording to the feature variable of the target object and a trained target classification model; and determining classification interpretation information according to the classification result contribution degree, thereby solving the problem that the identification features and the identification results of the high-income crowd cannot be interpreted and analyzed.

A method of detecting tampering in a compressed digital image includes extracting one or more neighboring joint density features from a digital image under scrutiny and extracting one or more neighboring joint density features from an original digital image. The digital image under scrutiny and the original digital image are decompressed into a spatial domain. Tampering in the digital image under scrutiny is detected based on at least one difference in a neighboring joint density feature of the digital image under scrutiny and a neighboring joint density feature of the original image. In some embodiments, detecting tampering in the digital image under scrutiny includes detecting down-recompression of at least a portion of the digital image. In some embodiments, detecting tampering in the digital image includes detecting inpainting forgery in the same quantization.

The invention provides an equipment information extraction method for a transformer substation, and the method comprises the steps: receiving voice operation information in a voice operation ticket ofan electric power maintainer, building a double-layer neural network model for the search and matching of operation ticket information, and extracting a DO object matched with the operation ticket information from an SCD file through similarity calculation and sorting; training the depth target detection neural network by using the sample library, verifying the fitting degree in the training process, performing multi-target identification on the power equipment infrared image in the sample library by using the connection weight and the bias parameter of the verified depth target detection neural network, and extracting equipment information of the transformer substation; and correcting the extracted equipment information by means of the DO object. The deep learning algorithm is used for carrying out deep feature mining on the input infrared image, feature parameters are not extracted manually, information of various kinds of power equipment can be effectively and accurately recognizedand obtained, and the accuracy of carrying out information extraction by means of the infrared image is improved.

The invention discloses an aboveground biomass estimation and scale conversion method oriented to a mean regional spectral unit, which mainly comprises the following steps: expression of remote sensing geoscience cognitive knowledge; Feature mining and screening; Homogeneous regional spectral unit segmentation; Biomass estimates at various scales. The invention can solve the problem that the current multi-scale biomass estimation mostly depends on the multi-resolution remote sensing image and the sampling of the field sample points under the multi-scale, and the introduction of the multi-source data makes up for the limitation of using only the optical data, and improves the operability of the fast estimation and conversion of the biomass under the different scales.

The invention discloses an advertisement click classification method based on a multi-scale stacking network. According to the advertisement click classification method, combined features are automatically constructed through an MSSP structure for constructing multi-scale features based on different receptive fields; by constructing a plurality of observers with different angles and different visual fields, multi-scale features are stacked bidirectionally from two angles of depth and width, and high-order and low-order features in different local visual fields are mined, and the diversity of extracted features is ensured; in addition, the structure learns parameters through factorization, thus guaranteeing that high-order features can be effectively learned in sparse data. According to theadvertisement click classification method, the defect that LR, Wide & Deep excessively depend on manual construction of combined features is overcome; meanwhile, compared with traditional Poly2 and FM models, characteristics of different scales can be mined from multiple angles to guarantee the diversity of information learned by the model; and compared with the characteristic that the time complexity of models such as FFM is too high, the time complexity can be kept at the linear level, and the high requirement of online advertisements for time response can be met.

The invention relates to a credit assessment technology, and aims to provide a personal credit assessment method and system based on fusion neural network feature mining. The method comprises the steps that behavior data of an individual user are preprocessed and checked and then subjected to matrix processing, and the obtained data serve as input of an LSTM model and a CNN model at the same time; in the LSTM model, sequentially processing by an embedding layer, a bidirectional long short-term memory neural network and an attention mechanism layer, and outputting a time sequence behavior feature vector extracted from the data; in the convolutional neural network model, processing is carried out through a convolutional layer and a pooling layer in sequence, and local behavior feature vectors extracted from the data are output; and carrying out vector splicing on the two types of feature vectors, taking the spliced feature vectors as input of an XGBoost classifier, and carrying out training to finally obtain a personal credit evaluation result. Compared with the prior art, the method has the characteristics of comprehensive index coverage, wide processing index source, advanced modeling mode, flexible model expansion, complete and effective feature extraction and accurate result.