64 results about "Generalization error" patented technology

The invention relates to a rapid target detection method based on a convolutional neural network, and relates to the computer vision technology. The rapid target detection method comprises the following steps: training convolutional neural network parameters by utilizing a training set; solving the problem of max-pooling losing feature by using an expander graph and generating a discriminative complete feature graph; regarding the full-connection weight of the convolutional neural network as a linear classifier, and estimating the generalization error of the linear classifier on the discriminative complete feature by using a probable approximately correct learning framework; estimating the required number of the linear classifiers according to the generalization error and the expected generalization error threshold value; and finally, completing the target detection on the discriminative complete feature graph by using the linear classifiers on the basis of a smooth window. The detection efficiency and the target detection precision are obviously improved.

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.

Time lapse, characteristic of aging, is a complex process that affects the reliability and security of biometric face recognition systems. Systems and methods use deep learning, in general, and convolutional neural networks (CNN), in particular, for automatic rather than hand-crafted feature extraction for robust face recognition across time lapse. A CNN architecture using the VGG-Face deep (neural network) learning produces highly discriminative and interoperable features that are robust to aging variations even across a mix of biometric datasets. The features extracted show high inter-class and low intra-class variability leading to low generalization errors on aging datasets using ensembles of subspace discriminant classifiers.

The invention discloses a power transformer fault diagnosis method. The power transformer fault diagnosis method comprises the following steps of determining N fault types of a transformer, and determining corresponding fault characteristic quantities for diagnosing the N fault types; taking the fault characteristic quantities corresponding to the N fault types as a testing sample, and performing normalizing processing of testing sample data; combining any two of the N fault types, establishing X (described in the specification) SVM ( Support Vector Machine) secondary classifiers, and training the X SVM secondary classifiers, and optimizing a SVM kernel function using a method based on combination of K-fold cross validation and an artificial bee colony algorithm; calculating generalization errors of each SVM classifier according to the K-fold cross validation method; and diagnosing the N fault types using an improved reordering adaptive directed acyclic graph support vector machine method. The invention has a capability of well diagnosing the fault type of a transformer, can greatly improve the accuracy of transformer fault diagnosis, and provides a reliable basis for transformer maintenance.

The invention relates to an RBF-neural-network-based atmospheric pollutant concentration prediction method. The RBF-neural-network-based atmospheric pollutant concentration prediction method includesthe steps: dividing experimental data according to the actual situation of the predicted area, and pre-processing the atmospheric pollutant concentration data; using the MMOD improved K-means++ algorithm to solve the center of clustering, and calculating each kernel function width based on the variance; sampling the experimental data, wherein data subsets taking part in creation of RBF neural networks are IOB, and the remaining data that are not drawn are OOB data; evaluating learners to screen out the RBF neural network with the smallest generalization error, training an integrated RBFNN model; and by means of the weighted integrated RBFNN algorithm, based on weighted Euclidean distance, training single parameter through the center of clustering, the width and the weight to optimize RBFNN, and applying the single parameter to the integrated RBFNN to predict data. The RBF-neural-network-based atmospheric pollutant concentration prediction method is applied to atmospheric pollutant concentration prediction, and can greatly improve accuracy of atmospheric pollutant concentration prediction.

Disclosed is a visualized optimization processing method for a random forest classification model. The method comprises: for a random forest classification model which has been constructed, estimating the degree of correlation between various decision trees of the random forest classification model via out-of-bag data; constructing a correlation matrix using the degree of correlation between various decision trees of the random forest classification model; according to the correlation matrix, by means of the dimension reduction technology, acquiring a visual pattern of the random forest classification model in a space with dimensions fewer than three; and according to the visualized pattern of the random forest classification model, conducting optimization processing on the random forest classification model, so that the upper limit of a second generalization error of the processed random forest classification model does not go beyond the upper limit of a first generalization error of the random forest classification model prior to processing. By means of the above-mentioned method, the present invention can reduce the number of decision trees in the random forest classification model and reduce the memory space required by the random forest classification model, and can also improve the prediction speed and accuracy at the same time.

The invention provides a construction method of a domain self-adaptive classifier, a construction device for the domain self-adaptive classifier, a data classification method and a data classification device, wherein the construction method comprises the following steps that: a combined penalty objective function for constructing the domain self-adaptive classifier is determined, wherein the domain self-adaptive classifier is a classifier for classifying the data of a target domain and a source domain; the domain self-adaptive generalization error upper limit is determined on the basis of the combined penalty objective function; and on the basis of the domain self-adaptive generalization error upper limit, more than two classifiers are subjected to coordination training, and the domain self-adaptive classifier is constructed. The problem of distribution unconsistency of a source domain and a target domain in the prior art is solved; the more accurate classification can be realized on the premise of ensuring the convergence; the computation complexity is greatly reduced; and the problem of cross-domain information processing which cannot be handled by ordinary mode identification is solved.

Provided is a method of optimizing a deep neural network based on learning automata (LA). In the training phase of a deep neural network, by starting from a fully connected initial network structure, weak connections in the network are continuously found and removed in the process of iterative parameter update through gradient descent. Thus, a more sparsely connected network structure with smaller generalization error is obtained, and image classification can be carried out on test samples more accurately. The weak connections are judged by LA through continuous interaction with the neural network in the process of training. By using the idea of reinforcement learning for reference, introducing a learning automaton algorithm to improve the traditional back propagation algorithm and removing redundant connections to reduce network parameters, the classification accuracy of test samples is increased, and the method has stronger ability to prevent over-fitting.

The invention provides a satellite anomaly detection method based on an improved Gaussian process regression model. The method comprises the steps of determining a prediction variable according to thecorrelation distance from a response variable and the change of the correlation distance before and after abnormality, then establishing a GPR model, calculating posterior distribution of a responsevariable set corresponding to the prediction variable, determining a prediction range according to generalization errors of the model, and judging whether abnormality occurs or not according to the prediction range.

The invention relates to a method for classifying electroencephalogram (EEG) signals based on a multi-scale brain function network. The method comprises the steps that data are acquired, specifically,the EEG signals in a resting state are collected and preprocessed; multi-scale time series are calculated, specifically, multi-scale processing is conducted on preprocessed time series of all leads,and thus the generalized multi-scale coarse-grained time series are obtained; the multi-scale brain function network is built, specifically, the amplitude correlation degree and the phase correlationdegree of all the leads of the EEG signals are taken as the quantitative standards to calculate multi-scale weighted brain function networks; a multi-scale convolutional neural network capable of learning the multi-scale brain function network is built; and neural network training is conducted, specifically, the cross-correlation degree of decision errors generated by all convolutional neural network paths is taken as a penalty term and a penalty loss function to accelerate decision-making of the neural network and reduce the generalization error.

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 discloses an ensemble learning method for legal text information mining, involving the fields of information mining and ensemble learning, By extracting different features from the preprocessed legal texts and building corresponding feature engineering models, using linear SVM classifier to learn the text vectors from different feature engineering models, The learning linear SVM classifier is used to predict the pre-processed legal texts, and the Stacking method is used to integrate the predicted results. At the same time, the ensemble learning model is trained and constructed tooutput more comprehensive and more accurate predicted results for the legal texts to be processed. This method can better synthesize the existing information, discover the relevance of the context inthe information, so as to form a stronger non-linear division ability, reduce the generalization error, and have a higher accuracy in the prediction of charges, laws, sentences and other contents than the prediction of a single model. In addition, the invention also discloses an integrated learning system for legal text information mining.

The invention discloses a garbage classification method and system based on transfer learning and model fusion and a storage medium. In the method, the method includes classifying the junk image datathrough a classifier, and outputting a corresponding classification result; moreover, the generation process of the classifier adopted by the invention comprises the step of constructing two differentclassification networks into two strong classifiers by using an Adaboost algorithm, so that the generalization error rate is low, over-fitting is avoided, and the high accuracy of industrial requirements can be achieved. And the two strong classifiers are fused in a decision-making layer fusion mode, so that the feature diversity is fully utilized, and the classification accuracy is further improved.

A composite power quality disturbance identification method based on the segmented improved S transformation and a random forest comprises the following steps of firstly, segmenting an improved S transformation frequency domain based on the disturbance signal characteristics, and assigning different window width adjustment factor values to each segment; secondly, extracting the disturbance signalcharacteristics according to the frequency distribution characteristics of different frequency bands, constructing an RF classifier based on a CART algorithm by means of the extracted disturbance signal characteristics, and classifying the signals to be detected. According to the composite power quality disturbance identification method based on the segmented improved S transformation and the random forest, the method has higher classification precision for the most single power quality disturbance signals and the common double composite power quality disturbance signals and has better noise robustness, and the generalization error of the constructed classifier is lower.

The invention discloses a method for determining a transfer learning boundary of heterogeneous relation data in public opinion data role recognition, and relates to the technical field of transfer learning. The problem that in the prior art, data in two fields are not combined for learning and then applied to a target domain, and the classification effect is inaccurate is solved. The method comprises the following steps: defining divergence for measuring the difference between two heterogeneous domains according to the formula (1), and defining the divergence for measuring the difference between the two heterogeneous domains, solving Empirical distances from two fields of the same abstract hypothesis class A by using the method, and giving an algorithm for converting the two classes into the same feature space; giving A difference boundary between an empirical distance and a real distance , a boundary for minimizing an error of a target domain, a giving generalization error which is strongest in generalization ability and combines training data of a source domain and the target domain, and obtaining a boundary of the error of the target domain by minimizing a joint error. And obtaining the obtained boundary ensures that a reasonable boundary value under the condition that the labeled data of the target domain is very few. The method is suitable for various identification problems in public big data and new media data platforms.

The invention discloses a face recognition optimization method based on an SGASEN algorithm. First, the features of a face image are extracted. Then, recognition and classification are carried out using an SGASEN algorithm. In view of the problem that there are a large number of invalid cross operations on population in the SGASEN algorithm, a regression tree is used as a base classifier, and computationally simple Jaccard similarity detection is used before cross operation, so that cross operations on similar individuals are reduced, and the population is more diverse. In view of the fact that the number of base classifiers is not limited in the SGASEN algorithm, generalization error and the number of base classifiers are considered in a fitness function, and genetic evolution is carried out using an optimal retention strategy. Finally, a strong classifier with strong generalization ability and fewer base classifiers is obtained, and the accuracy of face recognition is improved.

The invention provides an off-line embedded abnormal sound detection system. The off-line embedded abnormal sound detection system comprises a sound acquisition module, a sound audio feature extraction module and a neural network module. The sound audio feature extraction module processes sampling data obtained by the sound acquisition module through a digital microphone in a frequency domain by using fast Fourier transform, and inputs the sampling data to the neural network module to complete anomaly classification. The neural network module comprises a CNN feature extraction layer, an LSTM long-term and short-term memory layer, a full connection and classification layer and a trigger judgment layer. The number of network layers of the CNN feature extraction layer is dynamically adjustable, the network structure of the full connection and classification layer is dynamically variable, and a trigger decision layer is used for eliminating generalization errors generated by the neural network. The invention further comprises a method for carrying out anomaly detection by utilizing the off-line embedded abnormal sound detection system. The method works in an off-line environment, has less dependence on a network, is high in performance and reliable in work, and can adapt to a changing abnormal diagnosis working environment.

The invention discloses a method for predicting the parameters of a lithium battery impedance model. The method comprises the steps of: establishing a least squares support vector machine model for predicting target parameters; obtaining a voltage response value of a lithium battery to be tested; obtaining the target parameters of the lithium battery to be tested according to the least squares support vector machine model and the voltage response value. Therefore, the target parameters can be obtained according to the pre-established least squares support vector machine model and the obtainedvoltage response value; and the use of an artificial neural network to predict the parameters of the lithium battery impedance model is avoided, and the least squares support vector machine uses the principle of structural risk minimization, when applied to the prediction problem, the upper limit of a generalization error of the model is reduced while minimizing the sample point error, and the generalization ability of the model is improved, so that the parameters of the lithium battery impedance model can be predicted more accuracy.

The invention discloses a classification method for an imbalance problem based on a cost local generalization error. The classification method comprises the following steps of performing normalizationprocessing on each-dimensional feature of data; according to the statistical law of the data set, calculating the misclassification cost of different types of samples; constructing a cost-sensitive multilayer perceptron neural network model; for each sample, calculating random sensitivity (ST-SM) corresponding to a local generalization error of the sample; and completing model training by combining cost sensitivity and random sensitivity of a local generalization error. According to the method, the disadvantage of excessive influence of a data level method on data set distribution is avoided,the advantage of improving the generalization ability of a local generalization error and the advantage of faster training and testing time of an algorithm-level-based method are effectively combined, so that the classification stability of the unbalanced data set and the accuracy of a classification result of each category are improved, and a more reasonable classification decision boundary is obtained.

The invention discloses a support vector machine based automatic focusing method of ultrasonic phased arrays of ring welds. The method includes: establishing a rectangular coordinate system, taking a least squares support vector regression (LSSVR) machine as the regression model and Gaussian radial basis function as the kernel function, performing training with input and output which are subjected to standardized processing, and establishing a prediction model of LSSVR initial function; taking a generalization error calculated with a k-fold cross validation method as the target function, optimizing the LSSVR initial function by adopting an optimization method combining a coupling-simulated annealing algorithm with a grid searching method to acquire an optimal hype-parameter and an LSSVR optimization model, predicting tested data to acquire the output, performing inverse normalization processing to obtain the optimal acoustic beam path plan of a corresponding partition, performing reverse solution according to the Fermat theorem to obtain time delay, and applying ultrasonic signals to perform focusing detection on ring weld defects. By the arrangement, the method has the advantages that intelligent ring weld defect detection is realized, and detection precision and efficiency are effectively improved.

The invention discloses a software BUG classification method based on a sparse cost matrix. The method comprises the following steps of S1, obtaining a software BUG report; s2, encoding the software BUG report data; S3, initializing a sparse cost matrix; s4, training a weighted extreme learning machine by using the encoded data to enable the weighted extreme learning machine to output a correct report classification result; s5, solving a local generalization error of the weighted extreme learning machine by using the trained weighted extreme learning machine; s6, generating a new software BUGsparse weight matrix by using a crossover mutation strategy in a differential evolution algorithm; s7, training a new weighted extreme learning machine on the same unbalanced data set by using the newsparse weight matrix; S8, until a lower local generalization error cannot be obtained; and S9, predicting an unknown software BUG report by using a weighted extreme learning machine capable of obtaining the lowest local generalization error to obtain a corresponding report classification result.

The invention discloses a new-drilled well workload prediction method based on ensemble learning. The method is characterized in that a new-drilled well workload prediction model based on a random forest is built, a key hyper-parameter combination of the new-drilled well workload prediction model is optimized through a particle swarm optimization method, and a weighted voting mechanism is added ina decision stage; by adjusting the weight value of the weak classification decision tree, the generalization error of new-drilled well workload prediction is reduced, and the precision of the new-drilled well workload prediction model is improved. The key hyper-parameter combination in the new-drilled well workload prediction model is optimized by adopting the particle swarm optimization method,the influence of noise data in oil reservoir development historical data is reduced, and the stability and the operation speed of the random forest method are improved; a weighted voting mechanism isadded in the decision stage, wherein the weight proportion of a high-score decision tree is increased and meawhile the negative influence of a low-score decision tree on a prediction result is reduced, and thereby the precision of a new-drilled well workload prediction model is improved.