45results about How to "Reduce generalization" patented technology

The invention discloses a grounding grid corrosion rate level prediction method which comprises the following steps: (1) inputting training sample data; (2) randomly sampling training samples according to a bootstrap sampling principle in a Bagging algorithm, forming training sample bootstrap subsets with the number of M, and constituting training sample bootstrap subset data sets; (3) structuring a weak classifier model according to a k-nearest neighbor (KNN) algorithm, sequentially training the training sample bootstrap subsets with the number of M, and obtaining weak classifiers with the number of M; (4) structuring a strong classifier model according to an Adaboost algorithm; (5) inputting to-be-tested sample data, predicting a grounding grid corrosion rate level, obtaining a predicting result, and displaying the predicting result through a displayer. The grounding grid corrosion rate level prediction method has the advantages of being novel and reasonable in design, convenient and fast to use and operate, high in predicting precision, capable of achieving an accurate prediction to the grounding grid corrosion rate level by means of a small amount of data samples which are measured in the prior art, low in implementation cost, strong in practicability and high in value of popularization and application.

The invention provides a landslide displacement prediction method based on wavelet transform-rough set-support vector regression (WT-RS-SVR) combination. By means of the method, according to the characteristics of the influence factors of landslide displacement, the complex displacement process and landslide displacement monitor data measured in real time, accumulative displacement of a typical monitor point is decomposed into trend-term displacement and periodic-term displacement through WT, and a trend-term displacement prediction function is obtained through curve fitting; screening is conducted on the influence factors of the landslide displacement through an RS algorithm, and selected factor sets are used as input factor sets of an SVR machine, accordingly a landslide displacement optimization prediction model based on WT-RS-SVR combination is established, and the precision of a prediction result is analyzed and evaluated. The prediction result of the landslide displacement prediction method can well embody the development and change tendency of the landslide displacement. The landslide displacement prediction method has high prediction capacity, and is accurate, effective and practical.

The invention discloses a privacy protection method in multi-sensitive-attribute data release, and solves the problem of poor quality of quasi-identifier data in multi-sensitive-attribute data release. The basic thinking of the invention is as follows that: firstly, clustering is executed on data sets, the data sets of which quasi-identifiers are similar are aggregated into one aggregate, and a plurality of data aggregates are generated; secondly, a multi-dimension bucket structure is constructed on the basis of sensitive attributes, and data records are mapped into the multi-dimension bucket structure according to values of the sensitive attributes; and then on the basis of multi-dimension buckets, grouping is carried out, i.e., main sensitive attributes are selected, dimension capacity of the main sensitive attributes is calculated, L (L is greater than or equal to 2) main sensitive attributes with the maximum dimension capacity are selected, one data record is respectively selected from the L main sensitive attributes, whether the data records meet the multi-sensitive-attribute L-diversity is judged, and if not, each bucket is sequentially traversed according to the capacity from big to small until the data records meet the multi-sensitive-attribute L-diversity. The process is repeated until the data in the buckets do not meet the multi-sensitive-attribute L-diversity. Finally, all groups are subjected to anonymization processing.

The invention discloses an intelligent fault diagnosis method under a small sample based on an attention mechanism element learning model. According to the intelligent fault diagnosis method, an attention mechanism and a meta-learning method are used for establishing an association network model; short-time Fourier transform is carried out on mechanical signals to obtain a time-frequency spectrogram of the mechanical signals; feature extraction and operation state recognition are further carried out on the time-frequency spectrogram; and rich fault information contained in the mechanical signals can be effectively mined. According to the intelligent fault diagnosis method, a pseudo distance can be trained adaptively to evaluate the similarity between related data; clear mathematical formula definition is not needed; and high mechanical equipment fault diagnosis accuracy can be obtained. Therefore, the dependence of a feature extraction process on artificial experience and the dependence of an existing intelligent fault diagnosis algorithm on a large amount of training data in a traditional diagnosis method are eliminated, and the problem of mechanical equipment fault diagnosis under the condition of small sample data is practically solved.

The invention relates to a vehicle color recognition method based on target recognition area interception, and belongs to the technical field of vehicle color recognition, and the method comprises thesteps: obtaining a picture containing a to-be-detected vehicle; performing target detection on the to-be-detected picture to obtain an image of the to-be-detected vehicle; extracting vehicle window area information of the to-be-tested vehicle to obtain coordinate values of four corners of a front vehicle window; removing a part of detection pictures with too low resolution; intercepting a vehicleengine hood area as a target identification area of the picture by utilizing the colinearity and parallelism of the boundaries of the vehicle window and the vehicle engine hood; carrying out saturation enhancement processing on the extracted vehicle engine hood area image; and performing color recognition on the saturation-enhanced vehicle engine hood image by using the RGB color recognition model and the HSV color recognition model, and outputting a final recognition result. According to the method, the problem that in traditional vehicle color recognition, interference areas such as backgrounds and vehicle windows influence vehicle body color recognition is solved, and the accuracy and robustness of vehicle color recognition are improved.

The invention discloses a wireless cellular network flow prediction method based on deep transfer learning and cross-domain data fusion, and belongs to the technical field of intelligent communication. Similarity among three services of short messages, telephones and the Internet and the similarity among different regions are analyzed, a plurality of cross-domain data sets is fused, and a space-time cross-domain neural network model is adopted to predict the wireless cellular traffic; a cross-service and region fusion transfer learning strategy based on a space-time cross-domain neural networkmodel (STC-N) is provided, and the prediction precision of a target domain is improved according to data characteristics of a source domain. The method can verify that the more comprehensive the considered data set is, the higher the prediction precision of the model is; in addition, the proposed transfer learning strategy can reduce the training data, calculation capability and generalization capability required for constructing the deep learning model.

The invention discloses a false positive gene mutation filtering method for targeted capture of gene sequencing data. The false positive gene mutation filtering method comprises the following steps: preprocessing gene mutation detection data; selecting three different supervised learning algorithms based on a triple training method to construct three different initial classifiers H1, H2 and H3, namely selecting three different supervised learning automatons and a learner generated based on an initial training set; training the H1, H2 and H3 to obtain an extended training set, and updating themodel; and marking the unmarked sample set U by using the trained model, and filtering according to a marking result. The method solves the problem that the traditional method cannot effectively copewith the batch difference.

The invention belongs to the field of computer vision, particularly relates to a mark detection model training and mark detection method, system and device based on multi-stage transfer learning, andaims to solve the problem of low detection accuracy of an existing mark detection model due to few mark samples. The system model training method comprises the steps of pre-training a mark detection model based on a sample selected in an ImageNet data set to obtain a first model; performing fine adjustment training on the first model based on the synthetic mark sample to obtain a second model; training the second model based on the real mark sample to obtain a third model; and taking the third model as a trained mark detection model. The detection method comprises the following steps: acquiring a to-be-detected mark image; and carrying out target mark detection on the mark image through the mark detection model obtained by the model training method. According to the invention, the number of mark samples is increased, and the detection accuracy of the mark detection model is improved.

The present invention relates to a pile hole construction method. Said method includes the following 9 drilling and slag-discharging processes: slurry circulation clash rotary grab drilling; slurry circulation impact rotary grab drilling; slurry circulation rotary grab driling; slurry circulation hydraulic force horizontal static pressure grab drilling; slurry circulation hydraulic force active feeding submerged grab drilling; impact rotary grab drilling; rotary grab drilling; slurry circulation rotary grab drilling; hydraulic force horizontal static pressure grab drilling; slurry circulation hydraulic force horizontal static pressure drilling; hydraulic force active feeding submerged yrab drilling and slurry icrculation hydaulic force active feeding submerged grab drilling. said invented various drilling methods and slag-discharge methods and their combination can be used for various strata, and have the advantages of quick hole-forming speed, small drilling resistance and small slurry pollution.

The invention discloses a circular RNA recognition method based on a machine learning strategy. The method comprises the steps: inputting data, positioning each candidate circular RNA on a reference genome, and extracting Reads features nearby circular RNA regions; training a supervised machine learning model by using the extracted features; and performing true and false positive classification onthe candidate circular RNA set by using the trained model, and outputting final circular RNA. The method belongs to a machine learning filtering strategy, has the advantages of the machine learning filtering strategy, and can remarkably save the cost, time and the like in clinical practice.

The invention provides an SSH man-in-the-middle attack detection system based on session similarity analysis, which selects a suspicious SSH session pair from network traffic data, and then discriminates the similarity between encrypted session pairs through a neural network technology so as to complete the detection of a man-in-the-middle attack event in the network traffic data. The method specifically comprises the following steps: designing an SSH man-in-the-middle attack detection process framework based on session similarity analysis, and defining composition modules and detection stepsof a detection scheme; designing an SSH suspicious session pair selection algorithm, so as to effectively reduce the session pair scale needing similarity discrimination; providing a sequence data representation method of the SSH session, and effectively identifying the similarity and uniqueness of the SSH session; constructing a session pair similarity judgment module based on an LSTM neural network and a full connection layer neural network, achieving prediction of SSH session pair similarity, and then completing determination of man-in-the-middle attack events.

The invention discloses an oil delivery pump rolling bearing state evaluation method based on a convolutional neural network and a long-term and short-term memory network. The method comprises the following steps: acquiring vibration data; dividing the state of the rolling bearing according to the existing full life cycle data of the rolling bearing; acquiring a time domain feature, a frequency domain feature and a time-frequency domain feature of the vibration data; preprocessing the data, and constructing a training set and a test set; constructing a convolutional neural network-long short-term memory network model; performing forward propagation on the training network, and performing back propagation to update network parameters; judging whether the model precision meets requirements or not, wherein the output model is used for state evaluation. According to the invention, on the basis of a large number of experiments, it is found that the convolutional neural network is high in accuracy, the long-term and short-term memory network is high in generalization ability, and the models of the two methods are fused to obtain a model with the final accuracy reaching 95% and the generalization ability reaching 78%. And a one-dimensional convolutional neural network is applied, so the process of converting data into images is omitted, and the efficiency is improved.

The embodiment of the invention provides a loss function optimization method and device of a classification model and a sample classification method. The optimization method comprises: generating a filter vector corresponding to a classification label vector, wherein the classification label vector and the filter vector both comprise a dimension corresponding to a first type of classification anda dimension corresponding to a second type of classification, and a dimension value corresponding to the second type of classification in the filter vector is zero; Generating an original loss function according to the classification label vector and the output result of the classification model; Filtering the original loss function by using the filter vector to remove a component of a second class of classification in the original loss function to obtain a loss filtering function; And performing post-processing on the loss filtering function according to a preset rule to obtain a loss optimization function. Therefore, the optimized loss function can improve the learning weight of the classification model to the text features of the first classification, does not learn the text features ofthe second classification, reduces the generalization of the classification model, and improves the text classification accuracy.

The invention discloses a data processing packet modeling method for a decoupling mode of a lightweight design of a car body. The data processing packet modeling method comprises the following steps of: decoupling a response function into uncoupling terms and first-order coupling terms; preliminarily judging the number of terms required for constructing a model; constructing every uncoupling term, and judging whether the uncoupling terms are nonlinear or not; repeating the step of constructing every uncoupling term until all the uncoupling terms are constructed to obtain a preliminary an approximation model formed by the pure uncoupling terms and comparing the approximation model with a true model; identifying whether the first-order coupling terms exist or not; if the first-order coupling terms exist, identifying associations of variables coupled with each other, and constructing corresponding coupling terms by using the approximation model technique; repeating the step of identifying whether the first-order coupling terms exist or not until all the first-order coupling terms are identified to obtain a global approximation model and optimizing the global approximation model; and entering an iteration step if a condition of convergence is not satisfied. The data processing packet modeling method disclosed by the invention has the advantages that the principle is simple; the constructing requirement of the high-dimensional approximation model required by engineering can be met; the solving precision is ensured; and the efficiency of an optimization algorithm of the approximation model is improved.

The invention provides a cerebral infarction classification method based on deep learning. According to the method, the concept of deep learning is utilized, different tasks are calculated in parallelthrough the network, feature fusion is carried out on features extracted by the network in a subsequent feature layer, and therefore classification in the aspect of cerebral infarction is carried outthrough the fused features. According to the method, the information is fully utilized by utilizing the characteristic that the CT information and the MR information can complement each other, and meanwhile, the information is fused on a subsequent characteristic layer, so that redundant information is removed, computing resources of a computer are reduced, and the cerebral infarction detection precision is improved.

The invention relates to a miRNA-mRNA target prediction method based on sequence statistical characterization learning, and belongs to the field of bioinformatics. Multi-scale and multi-granularity feature extraction is carried out on structural features of miRNA and mRNA sequences by adopting a mode of combining a neural network and an attention mechanism, so that the obtained features not only contain local and global multi-scale features of each base character of the sequences and a target region sequence, but also contain fine-granularity and coarse-granularity multi-granularity semantic information feature relation. And a variational automatic encoder structure is used as an overall framework, the prediction accuracy is ensured by utilizing probability distribution of input data, and the interpretability of the model is improved. According to the miRNA-mRNA target prediction method based on sequence statistical characterization learning, target prediction of miRNA-mRNA can be effectively completed.

The invention discloses a classifier updating method. The classifier updating method comprises the steps of firstly, collecting a wrongly-classified training sample and an incremental wrongly-classified sample; secondly, collecting all abnormal samples in an incremental wrongly-classified sample set by utilizing a basic wrongly-classified sample set; finally, updating a classifier by utilizing an abnormal sample set and the learning of an incremental machine. According to the classifier updating method disclosed by the invention, the incremental wrongly-classified sample set is screened by utilizing the basic wrongly-classified sample set, and thus the phenomenon that the generalization performance of a harmful image classifier is reduced as some helpful wrongly-classified training samples are used for updating can be avoided.

The invention discloses a semi-supervised image classification method based on random region interpolation, and the method comprises the steps: selecting a small number of images with real labels froma training set, and taking other remaining images as images without real labels; simultaneously sending the two types of images to a random region interpolation module; wherein the interpolation process is different, the image with the real label can directly generate a new augmented image through interpolation, but each image without the real label cannot be normally interpolated, so that high-confidence label information can be obtained through a teacher network to serve as a temporary label of each image without the real label, and then interpolation operation is carried out; and trainingthe network by using the new augmented image until the network model is trained to a preset number of times. Random region interpolation is carried out on two types of images at the same time, a new augmented image is generated for training a classification network, and the generalization performance of a training model is improved.

The invention discloses a bionic intelligent control method based on a multi-connotation self-adjusting neural network. The method comprises: step one, carrying out application restriction design countermeasure for a universal approximation theorem; step two, on the basis of the method, designing a multi-connotation self-adjusting neural network; step three, establishing a high-order non-affine system; step four, designing an MSAE-NN-based controller; and step five, applying a controller u to the non-affine system established at the step two to enable a system state x1 to track an expected track xd (t) precisely with modeling uncertainty and external interference existence. Therefore, a multi-connotation self-adjusting neural network having a time-varying ideal weight, a smoothing self-increasing neuron and a diversified basis function is constructed based on the working principle of the neural network and the multi-connotation self-adjusting neural network is applied to control of the uncertain high-order non-affine system, so that problems that are common and are ignored of the NN controller designed based on the universal approximation theorem are solved.

The invention discloses a method for carrying out quantitative evaluation on the soup hue quality of tea, and the method comprises the following steps: obtaining the final sensory soup hue evaluation values of selected tea samples by more than three tea-tasters; respectively measuring the soup hue measuring values of multiple batches of selected tea samples by using a color difference meter and calculating derivative index values, and carrying out principal component analysis on the tea soup hue parameter variables such as the soup hue measuring values and the derivative index values so as to obtain the previous k principal component load data of the selected tea samples; on the basis of taking the previous k principal component load data of the selected tea samples as the input of a BP (back propagation) neural network, and the final sensory soup hue evaluation values of the selected tea samples as the output of a BP neural network model, carrying out repeated training, obtaining the BP neural network model; and obtaining the previous k principal component load data of a to-be-detected tea sample by using the same method, then inputting the previous k principal component load data of the to-be-detected tea sample into the BP neural network model to predict the quantitative value of the soup hue quality of the to-be-detected tea sample. By using the method disclosed by the invention, the quantitative values of the soup hue quality of tea can be given scientifically and effectively; and the method disclosed by the invention has the advantage of extremely good consistency with an artificial sensory evaluation method.

The invention discloses a time sequence recommendation algorithm based on generation sorting, and belongs to the technical field of recommendation systems. The method comprises the specific steps of 1, randomly sampling a data sample, transmitting the data sample to a recommendation model, and scoring the data sample through the recommendation model; 2, converting the data into corresponding recommendation scores through one-hot coding, an embedding layer, a generation layer and a conversion layer; 3, updating recommendation model parameters; 4, judging whether the accuracy of the recommendation model is improved or not; and 5, after the training of the recommendation model is finished, scoring by using the recommendation model, sorting according to a scoring result, and recommending the project to the user. Aiming at the problem that data sparsity and user preferences change along with time due to a large number of users and products in recommendation system training data, negative sampling and a Gaussian distribution-based generation method are used for training the model, so that the time sequence model has generalization ability and can identify changes of user preferences; finally, the accuracy of recommendation is improved.

The invention discloses a travel purpose identification method based on mobile phone signaling data, and the method comprises the steps: carrying out map matching of the signaling data, starting from the identification of a stop point, carrying out the identification of the stop point based on an ST-DBSCAN space-time density clustering algorithm, and combining with a heuristic algorithm. Parameters of a spatio-temporal clustering algorithm are mined depending on user mobile phone signaling data with labels, meanwhile, speed characteristics of travelers are considered, and the fineness of stay point recognition is improved. Trajectory spatial-temporal features, personal attributes and traffic facility built environment features of user travel are obtained through feature extraction, and the features are abstracted as nodes. A directed arc is obtained through a constraint-based Bayesian network structure learning algorithm, Bayesian network modeling is preliminarily completed, and a Bayesian network probability model is perfected through a rule heuristic modeling method by taking a travel purpose and commuting characteristics as deductive reasoning objects. When the travel purpose identification is carried out, the travel characteristics are obtained through the mobile phone signaling data of the user, and the travel purpose probability result of the traveler can be obtained.

The invention discloses a document classification method and device and electronic equipment, and relates to the technical field of data processing, in particular to the document classification method and device and the electronic equipment. According to the specific implementation scheme, the method comprises the following steps: acquiring image blocks and word blocks according to a to-be-identified document; inputting the image blocks and the word blocks into a pre-training migration model, and obtaining visual representation and text representation; and acquiring the category of the to-be-identified document according to the visual representation and the text representation. According to the embodiment of the invention, the visual representation and the text representation of the to-be-identified document are extracted, so that the classification of the text in the document is realized. According to the embodiment of the invention, the uncertainty of manual labeling can be avoided, and the accuracy of document classification is improved.

The invention discloses an auto-encoder training method and assembly, and an abnormal image detection method and assembly. In a certain iteration process, the same sample image is utilized to train an auto-encoder, a vector discriminator and a reconstruction discriminator respectively, so that the image reconstruction capability of the auto-encoder can be improved. The vector discriminator is enabled to constrain sample vectors to be approximately uniformly distributed. A reconstruction discriminator is enabled to improve the ability of the reconstruction discriminator to discriminate the original occlusion region and the occlusion region obtained through reconstruction, thereby reducing the possibility of the occurrence of the identical mapping and the generalization ability of the auto-encoder. Finally, the auto-encoder is enabled to have a good reconstruction ability only for a normal image. Therefore, the detection accuracy of the auto-encoder on the abnormal image is improved. Correspondingly, the invention provides an auto-encoder training assembly, an abnormal image detection method and an abnormal image detection assembly, which also have the above technical effects.

The invention discloses a high-power LED high-voltage constant-current driving module which comprises a plurality of parallel-connecting single lamp strips. The single lamp strips comprise a plurality of LED lamps in a series mode. The LED lamps of each single lamp strip are connected in parallel with an external zener diode and then are connected in series with an independent driving module. The concept that the zener diode is externally arranged solves the problems of non-reworking performance caused by an internal traditional LED anti-static zener diode and non-reworking performance after secondary optical lens using at middle-high power. Accordingly, LED production is easy, cost is lowered, lamp strip maintaining is easy, single lamp strip independent driving is achieved, power supply manufacturing is easy, and universalization is achieved.

The invention provides a network model fine adjustment method and system adapting to a target data set, a terminal and a storage medium. The system comprises an original neural network, a target neural network, and a fine adjustment module. The original neural network comprises a unified neural network layer and an output layer, and the original neural network is used for pre-training the originalneural network by using an original data set to obtain parameters of each layer of the pre-trained original neural network and marking the parameters. The target neural network comprises network layers matched with those of the original neural network, the target neural network also comprises an output layer and a unified neural network layer, and the parameters of the unified neural network layer in the target neural network adopt the parameters of the unified neural network layer in the original neural network. The fine adjustment module is used for finely adjusting the characteristic parameters of all the layers except the output layer in the target neural network so as to adapt to a recognition target.

The invention provides a refrigeration house, and relates to the technical field of refrigeration equipment. The technical problems of long construction period and high cost of a fixed refrigeration house are solved. The refrigeration house comprises a refrigeration house body, a bottom frame and a movable assembly, wherein the refrigeration house body is connected with the movable assembly through the bottom frame, free movement of the refrigeration house body can be realized through the movable assembly, and the refrigeration house further comprises a refrigeration system arranged on the refrigeration house body; assembly type connection modes are adopted between the bottom frame and the refrigeration house body, between the bottom frame and the movable assembly, and / or between the refrigeration system and the refrigeration house body; and at least two temperature zones are arranged in the refrigeration house body, and the temperature in each temperature zone adopts an independent control mode. According to the refrigeration house, the refrigeration house body is arranged on a transportation vehicle or rolling wheels, the refrigeration house can be moved, transportation and workof the equipment under different conditions can be met, the use place can be changed conveniently, the carrying link is reduced, manpower is saved during distribution, the construction period is short, and the production cost is reduced.

The invention discloses a modeling and analysis method and device for predicting tunnel extrusion deformation, and a storage medium. The modeling method comprises the following steps of establishing ahistorical data set of tunnel extrusion deformation, wherein the historical data set is selected from a historical case and comprises multiple features used as training database parameters and used for describing data; performing parameter optimization on the historical data set to obtain an optimized database; and taking tunnel extrusion deformation degrees as classification standards, and performing multi-classification SVM training by utilizing the optimized database to obtain a multi-classification SVM model. The classifier training time can be shortened; tunnels are classified accordingto the deformation degrees; relatively good performance is achieved in prediction precision; and the seriousness of a potential extrusion problem can be estimated according to a predicted extrusion type.