31results about How to "Reduce generalization error" patented technology

The invention relates to a vehicle flow predicting method based on an integrated LSTM neural network. On the basis of historical data obtained by vehicle flow detection, an integrated LSTM neural network vehicle flow prediction model is established to carry out vehicle flow prediction, so that the generalization error of the prediction model is reduced and the accuracy is improved. The method comprises the following steps that: data preprocessing is carried out; according to a preprocessed vehicle flow time sequence value, a vehicle flow matrix data set is constructed and the vehicle flow of an (n+1)th period of time is predicted by using first n periods of time, wherein each period of time is delta t expressing the time length and the unit is min; a plurality of different LSTM neural network models are constructed by using different initial weights; on the basis of a bagging integrated learning method, a training set and a verification set are constructed; a plurality of LSTM neural networks are trained to obtain an optimized module; a weighting coefficient of the single LSTM model is calculated by using the verification set; and inverse transformation and reverse normalization are carried out on a predicted vehicle flow value to obtain a predicted vehicle flow and integrated weighting is carried out to obtain a vehicle flow value predicted finally by the model.

The invention relates to a social media user demographic attribute prediction method based on multi-model stack fusion, and three demographic attributes including gender, age and region of a user arepredicted. The prediction of the three demographic attributes comprises the steps of (S1) user feature extraction, (S2) model training, and (S3) multi-model fusion to obtain a prediction result. The feature extraction in the present invention is not only for the text content of user Weibo, but also relates to statistical features, time information features and social relationship features, and theaccuracy of prediction is ensured. The mode of multi-model stack fusion is used to fuse logistic regression, random forest, and XGBoost models, a generalization error can be effectively reduced, andthe prediction accuracy is greatly improved.

The invention discloses an AC / DC power distribution network load prediction method based on ensemble learning, and the method comprises the steps of carrying out the load data filling and normalization on the original load data, sampling a load sample input vector and a sample label through a sliding time window, and constructing a training data set; establishing a gradient lifting model, settingthe number of weak learners, and establishing a plurality of shallow neural networks to fit the negative gradient of the gradient lifting algorithm to obtain a combined prediction model; and selectinga load vector before a to-be-predicted time point as an input vector by using the sliding time window, and determining a load prediction value in combination with an ensemble learning model. According to the invention, the load prediction is carried out by fusing the strong learners of multiple models, so that the precision of load prediction is improved.

The invention discloses an active learning multi-label social network data analysis method based on graph data, concretely comprising the steps of social network data acquisition, type marking and training, model building and social network user data analysis. The invention provides a multi-label graph data classification method, and is combined with an active learning method based on error bound minimality. A series of objective equations are obtained through multi-label classification and LLGC, and are applied to transductive Rademacher complexity. The method aims to minimize the experience transductive Rademacher complexity and to obtain a minimized generalization error bound, and thereby obtains a few nodes containing vast information on graphs. The method can classify massive multi-label graph data so as to provide support for subsequent decisions.

The invention relates to the technical field of speech signal recognition and processing, and provides a speech separation and tracking method for public security criminal investigation and monitoring. The speech separation and tracking method includes the following steps that initial speech is imported according to timing sequence, the initial speech is subjected to framing and windowing processing, and a windowed speech signal is obtained; the windowed speech signal is time-frequency decomposed, and a time-frequency two-dimensional signal is obtained by the short-time Fourier transform; an endpoint of the time-frequency two-dimensional signal is detected in a frequency domain, and a corresponding speech signal segment of an empty language segment is filtered; a bidirectional long and short time memory network structure is used for performing speech separation of the two filtered dimensional time-frequency signal, and a great deal of speech waveform of a target speaker are output; anda target speaker model based on GMM-UBM is established and trained, the speech waveform of the target speaker are taken as models and input, a GMM model of the target speaker is acquired through an adaptive method and the speech waveform are recognized, a sequence number of the target speaker is outputted, that is, a speech tracking result.

The invention discloses a shale gas well staged fracturing effect evaluation and yield prediction method based on a random forest. The method comprises the steps of firstly, finding main fracturing and geological influence factors influencing the section yield through a Pearson correlation coefficient and a secondary dimension reduction strategy of a recursive feature elimination method; establishing a random forest model based on optimized influence factors, analyzing the gain degree of the main fracturing factors to the section yield by using the model, and finishing fracturing effect evaluation and yield prediction. The calculation method is simple, and the method is advanced. The microstructure of reservoir rock is very complex and irregular, large-scale fracturing is carried out, it is difficult for a traditional theory to combine complex and numerous fracturing parameters and geological parameters together to establish a nonlinear equation, and the historical fitting difficulty in a numerical simulation method is large. And important yield influence factors can be identified by adopting the secondary dimension reduction strategy and a random forest algorithm, and yield prediction can be well carried out.

The invention relates to the technical field of data mining, and particularly relates to an e-commerce discount coupon use probability prediction method based on an integrated model. According to thee-commerce discount coupon use probability prediction method based on the integrated model, the integrated model comprising a primary learner and a secondary learner is constructed, and in combinationwith the advantages of a specific single model in the prediction of discount coupon use probability, the defects of single model prediction are avoided. In the case of large sample data, the prediction accuracy is improved, and the generalization error of the prediction model is reduced. The e-commerce can build the integrated model through the technology to more accurately predict the probability value of a user to use the discount coupon and to decide whether to issue the discount coupon to the user according to the probability value, thereby reducing the operation cost of the merchant andbetter achieving a promotion effect desired by the merchant.

The invention relates to a method of soft measurement for concentration of reactant in an unsaturated polyester resin reacting kettle. The method uses a data acquisition module based on OPC (object linking and embedding for process control) technology, an abnormal data eliminating and pretreating module, a data normalization pretreating module, a digital filtering and wavelet treating module, a main factor analyzing module, and a self-adapting genetic algorithm optimized BP (back propagation) neural network soft measurement module. The method with high measurement precision for the soft measurement of the concentration of reactant in the reacting kettle during a batch process is provided, and the difficulty in measuring the reactant concentration in a batch reacting kettle is solved.

The invention discloses a research and development method of siRNA for COVID-19 virus drug therapy, which comprises: a first part of preliminarily screening potential high-efficiency siRNA based on multiple indexes, which specifically comprises the following steps: step 1, selecting an S gene sequence as a target sequence; step 2, obtaining a corresponding siRNA double strand; step 3, screening asiRNA sequence with a concentration of 36 to 53%; step 4, screening out siRNA with free energy; step 5, defining and calculating an index I; step 6, screening out siRNA of the first 50%; step 7, defining and calculating an index II; step 8, screening out siRNA with the index II being equal to 5; step 9, defining and calculating an index III; step 10, ranking the first 50% of siRNA; step 11, directly taking all the candidate siRNAs selected in the step 10; step 12, specifically targeting a target sequence; step 13, taking the remaining siRNA as the siRNA subjected to preliminary screening; anda second part of performing interference efficiency prediction and optimization by using a machine learning model. The method has the beneficial effect of realizing safe, reliable and high-interference-efficiency siRNA design.

The invention discloses a training method and a detection method of a network traffic anomaly detection model. The network traffic anomaly detection model comprises a feature extraction network and aclassification network, and the training method comprises the following steps: determining the number of hidden layers and the number of neurons in each hidden layer according to a training sample; constructing an initial feature extraction network according to the number of the hidden layers and the number of neurons in each hidden layer; training the initial feature extraction network by using atraining sample to obtain a trained feature extraction network; extracting abstract feature data of a training sample by using the trained feature extraction network, and training a classification network by using the abstract feature data so as to complete training of a network traffic detection model. The network structure can adapt to network flow data, the situation that the structure of a detection model is too complex and too simple is avoided, and therefore, generalization errors are reduced, the detection time can be obviously shortened, and the detection accuracy can be obviously improved.

The invention relates to an intelligent tooth brushing process evaluation method, which comprises the following steps that: 1) tracking a tooth brushing process of a standard Bass tooth brushing method, and collecting a denoising signal output by a GY-521 gyroscope in an electric toothbrush; 2) checking a data interval preprocessed and labeled in S1), and if the interval time of data is shorter than t seconds, according to the time interval of t seconds, solving a mean value in the period of time; 3) training an evaluation model based on a random forest; and 4) preprocessing the data of a person who brushes teeth according to a sensor data processing method in S1), and using a trained random forest model to classify the toothbrush state of the person who brushes teeth at different moments.The method has the beneficial effects that a machine learning method is adopted to classify the tooth state of the toothbrush in each time slice in a user tooth brushing process, the tooth brushing state of the user at different moments in the tooth brushing process can be predicted, and the tooth cleaning state and strength of the user at different moments can be accurately judged.

The invention provides a human respiration state monitoring system and method based on flexible sensing and deep learning, and the system comprises a flexible tension sensor which is used for converting human respiration data into an analog signal for output; a microcontroller which is connected with the flexible tension sensor through an analog-to-digital converter and used for receiving an analog signal output by the flexible tension sensor, converting the analog signal into a digital signal and outputting the digital signal through a wireless transmission module; a terminal which is connected with the microcontroller through a wireless transmission module and is used for receiving and outputting the digital signal output by the microcontroller; and a cloud server which is used for receiving the digital signal output by the terminal and monitoring and analyzing the respiration state of the human body by using a deep learning network model. According to the system and the method provided by the invention, the monitoring accuracy is improved based on deep learning, the cost is relatively low, and the use requirements can be well met.

The invention discloses a model building method and system, a paragraph label obtaining method and a medium, and relates to the field of natural language processing transfer learning, and the method comprises the steps: collecting all judgment document data from a database, and obtaining pre-training data; defining paragraph labels of different types of judgment documents; marking paragraph labels of different types of judgment documents to obtain training data; constructing a judgment document structured model; pre-training the model; training a pre-trained judgment document structured model by utilizing the training data; and debugging the trained judgment document structured model to obtain a final judgment document structured model, wherein the input of the judgment document structured model is judgment document text data, a task prefix is added to a paragraph of the input judgment document, and the output of the judgment document structured model is paragraph label text data of the judgment document. The model established by adopting the method can predict any type of judgment document paragraph label after being trained.

The invention relates to the technical field of face detection and recognition and aims to improve classification precision effectively and reduce generalization errors to realize effective classification screening on human face false detection result. The invention particularly relates to a semantic feature-based face false detection screening method, which comprises the following steps of performing face detection and alignment on an original image through a face detection and alignment algorithm by taking original image data as input of the stage, and zooming a detection alignment result to112 * 112 dimensions; performing pixel-level face semantic segmentation on the input face detection alignment result by adopting a BiSeNet-based real-time face semantic segmentation method to obtaina semantic segmentation result; processing the semantic segmentation result by adopting a feature engineering technology, and constructing and selecting a semantic feature with the highest representation capability; calculating the input semantic features by adopting a Stacking model integration framework to obtain a final human face false detection classification result and complete false detection screening; and realizing effective classification and screening of human face false detection results, thus the performance and robustness of the overall detection algorithm are improved.

The invention provides a double-channel adaptive correction network optimizing system based on a feature generalization layer. The system comprises a generalization channel, a correction channel, an error calculation unit and an adaptive correction unit. The generalization channel is used for generalizing original features maps and extracting features of weighting-corrected feature maps layer by layer. The correction channel corrects data in the generalization channel according to the error among the feature maps. The error calculation unit calculates the degree of difference of output featuremaps of some feature extraction layer in the generalization channel and the correction channel. The adaptive correction unit weights the feature map output by some feature extraction layer in the correction channel and the feature map output from the corresponding position of the generalization channel. Error of mean square calculated from all feature extraction nodes in the generalization channel and the correction channel is added into a target function as bound terms; through multiple times of iteration in a training process, a generated feature map is closer to original data, so that generalization error is reduced gradually.

The invention discloses a method for predicting the service life of an underground structure in an acid environment based on data driving, and the method comprises the steps of building a life sample simulation database of multiple influence factors according to different influence factors of the service life of an underground building structure in the acid environment; determining GP parameters needing to be optimized in a genetic programming program and an objective function form, taking given sample data as a drive, and using Python to compile a Harris-Eagle algorithm to optimize the GP parameters; and outputting the GP parameter, the objective function and the objective function value after the HHO algorithm optimization. The invention can be applied to prediction of the service life of the underground structure in the acid environment and has guiding significance on repairing and reinforcing of the underground structure in the acid environment; the established service life prediction model has appropriate training errors and small generalization errors, and the engineering practicability is high.

The invention discloses a method for enhancing a crack magneto-optical image through an improved GAN, and the method comprises the steps: collecting and preprocessing a crack magneto-optical image, expanding the preprocessed magneto-optical image, obtaining training data for training an improved GAN model, carrying out the training of the improved GAN model, and finally inputting the magneto-optical image of a to-be-tested piece into the improved GAN model. Image blocks are output through an internal generation network and spliced, so that a crack magneto-optical image with high contrast is obtained, and defect enhancement is realized.

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.

According to the power load frequency domain prediction method and system based on the IRF and the ODBSCAN, the technical problem that an existing method is large in error can be solved. The inventionprovides an improved random forest IRF (Implanted Random Field) and ODBSCAN (Open Distributed Broadcast System Controller Area Network)-based method. The improved random forest IRF and ODBSCAN-basedmethod is based on an improved random forest IRF (Implanted Random Field) and an improved ODBSCAN (Open Distributed Broadcast System Controller Area Network). The invention relates to a frequency domain combination prediction method based on frequency domain combination prediction (ions width Noise). The method comprises the following steps of: firstly, decomposing a load by adopting EWT (EnhancedWavelet Transform) to obtain different intrinsic mode parts (IMFs); secondly, predicting by adopting a reasonable method according to the characteristics of each part; wherein IRF prediction is adopted for the low-frequency part and the intermediate-frequency part; the high-frequency parts have uncertainty, the ODBSCAN is used for clustering according to the temperature and humidity of meteorological factors, and then a processing method is selected according to the characteristics of each type of samples. And finally, superposing the prediction values of the parts to obtain a total prediction result. An experiment is carried out according to field load data of a city; the prediction results are compared with the prediction results of an EWT-IRF model, an EWT-RF (Random Forest) model andan EMD (Empirical Mode Decomposition)-IRF model respectively, so that a better prediction effect can be obtained, and the change rule of an actual load is reflected.

The invention relates to the technical field of raw material oil product analysis, in particular to a raw material oil product molecular characterization method and system based on a limit gradient boosting tree model. The method comprises the following steps: S1, selecting an input feature vector from macroscopic physical property parameters according to preparation process characteristics, and selecting an output feature vector from phase components; s2, preprocessing the initial preparation process data set; s3, dividing into a training set, a verification set and a test set according to a specified proportion; s4, constructing a raw material oil product molecular characterization model based on a limit gradient boosting tree model; s5, obtaining a raw material oil product molecular characterization model conforming to the evaluation index; and step S6, carrying out molecular characterization on the raw oil product in the preparation process. The limit gradient lifting tree is adopted as a model, the prediction precision is high, and the calculation speed is high.

The invention relates to the field of computer signal processing, in particular to a brain wave signal analysis method. The invention discloses a brain wave analysis method based on Hilbert-Huang transform and optimization of an artificial bee colony algorithm and optimization of a support vector machine. The method comprises the following steps: collecting brain wave signal data; wherein the brain wave data are brain wave signals corresponding to imagination of different motion states; decomposing the original brain wave signal by adopting an empirical mode decomposition method to obtain a series of intrinsic mode functions; extracting brain wave features from the intrinsic mode function; and taking the extracted brain wave features as input vectors, and classifying the input vectors by using a classifier so as to distinguish motion states corresponding to the brain wave signals. According to the method, Hilbert-Huang transform is used to extract features, and artificial bee colony algorithm is used to optimize support vector machine classification, so that the method has stronger adaptability, better classification capability and higher calculation efficiency, and is helpful to improve the accuracy of brain wave classification.

The invention designs an online learning input degree identification method and system based on deep learning, and the method comprises the steps: firstly, carrying out the student face detection through YOLOv4 in order to guarantee that an image is not affected by an irrelevant background; secondly, aiming at the problems that the VGG16 network is huge in parameter quantity, time-consuming in training and the like, an improved VGG16 model is provided, and meanwhile, in the model training process, a depth deterministic information bottleneck method DIB is adopted to make up for the deficiency of a traditional loss function so as to obtain relatively compact feature expression, reduce generalization errors and improve the universality and stability of the model, and achieving precise identification of the learning input degree in a complex online learning scene; and finally, through comparison and analysis with various methods such as traditional machine learning and other deep learning, the effectiveness of the method is verified.