32results about How to "Less prone to overfitting" patented technology

The invention discloses a Chinese text sentiment analysis method based on deep learning, and belongs to the technical field of natural language processing. The defects of an unsupervised sentiment analysis method based on English are overcome. The method comprises the following steps: after converting an obtained corpus text into pinyin, pre-training a constructed language model to obtain a pre-trained language model; obtaining a small amount of text data which is in the same field as the corpus text and has emotion categories, converting the text in the text data into pinyin, and training a constructed emotion classification model based on a pre-trained language model to obtain a trained emotion analysis model; and carrying out sentiment classification on the unlabeled text by utilizing the trained sentiment analysis model to obtain a corresponding sentiment category label. The method is used for Chinese text sentiment analysis.

The invention discloses a random forest-based tunnel operation state sensing model building method. The method comprises the steps of randomly selecting ntree new self-service sample sets, constructing ntree decision trees, randomly selecting mtry features at each node of the decision tree, selecting out one feature to perform branch growth to obtain unbiased estimation of a random forest generalization error, and calculating program running time; and iteratively running all ntree and mtry parameter combinations, outputting unbiased estimation and running time corresponding to all the parameter combinations, determining an optimal ntree and mtry parameter combination value in a random forest, and building a tunnel operation state sensing model. According to the method, a capability of analyzing complex related relationship data can be improved and an over-fitting phenomenon does not easily occur; an actual prediction result shows that the average sensing precision, the recall rate andthe F measure are all superior to those of a comparison model; tunnel operation state change requirements can be better met; and accurate real-time sensing and prediction can be provided for tunnel operation states.

The invention discloses a sequence characteristic analysis method for forecasting a miRNA target gene. The method comprises the steps of constructing related characteristics of 27 miRNA-target point pairing sequences on the basis of a CLASH experiment data set, and forming a characteristic set comprising 84 characteristic values by combining traditional characteristic; and performing machine learning by using a random forest model, and constructing a miRNA target gene forecast model to perform miRNA target gene recognition. The model constructed according to the method has the advantages of high accuracy, sensitivity, specificity and precision, and the miRNA target gene can be relatively and accurately forecasted.

The invention discloses a method and a device for determining an application scene of a mobile terminal. The method is executed in the mobile terminal. One or more sensors are arranged in the mobile terminal and suitable for acquiring state data of the mobile terminal. The method comprises steps of acquiring state data in preset duration according to a preset frequency; determining state feature information of the mobile terminal according to the acquired multiple state data; and according to the state feature information, using a classifier to determine the application scene of the mobile terminal. In addition, the invention also discloses a device capable of employing the method, and the mobile terminal including the above device. According to the invention, the application scene of the mobile terminal can be identified, so it is facilitated that personalized services are provided for users.

The invention discloses a human face key point-based prediction system and method in an Android platform. The method comprises the following steps of 1, collecting a human face sample picture set and calibrating human face key points to form a training sample set; 2, obtaining a human face key point initial shape set S<0>; 3, performing training to obtain a global binary characteristic phi 1 during first-time cascading; 4, training linear regression W<1>; 5, obtaining a predicted deformation increment delta S<0> and a human face key point shape set S<1>=S<0>+W<0>.phi 0(I,S<0>) during first-time cascading; and 6, returning the trained global binary characteristic and linear regression device, and obtaining a regression model and final human face key point shapes S<T> until a maximum layer number of cascading is T. According to the system and the method, the algorithm running efficiency can be improved; a relatively small memory is consumed in a mobile platform; and human face key points are accurately located at a high speed, so that human face key parts can be quickly beautified.

The invention discloses a spectrum denoising method which comprises the following steps: acquiring a plurality of groups of spectrum signal samples; setting an order number and a regularization coefficient of a self-adaptive filter, selecting a minimum mean square error function as an optimal target function of the filter, and taking the samples as input signals of the filter so as to obtain output signals; based on a minimum mean square error function corresponding to a same position n of k samples, acquiring a weight coefficient vector W of the self-adaptive filter according to an Adam algorithm; calculating a signal to noise ratio of the self-adaptive filter; within a preset range of the order number and the regularization coefficient, updating the order number and the regularization coefficient of the self-adaptive filter, repeating the step of acquiring the signal to noise ratio of each self-adaptive filter, and selecting a self-adaptive filter corresponding to the maximum singleto noise ratio; performing filtering denoising on a same type of spectrum signals under a same environment condition by using the selected self-adaptive filter. Compared with a conventional standard LMS algorithm, the method disclosed by the invention is optimal in denoising effect, and rapid in convergence rate.

The invention provides an MIMO user detection and channel estimation device and method. The MIMO user detection and channel estimation device comprises a pilot frequency sequence generation module, achannel estimation module and a user detection module. The pilot frequency sequence generation module generates a pilot frequency sequence of a user by using a single-layer complex fully-connected neural network; the pilot frequency sequence is distributed and the pilot frequency sequence is sent to a user served by the base station; a neural network model based on an AMP algorithm form is built in the channel estimation module; the channel estimation module takes a base station receiving signal and a known pilot frequency sequence as inputs and takes a channel matrix as an output, the outputend of the channel estimation module is connected with the user detection module, and the user detection module takes the channel matrix as an input and takes a user activity vector as an output to obtain a user detection result. According to the method, the neural network based on the AMP form is adopted for channel estimation, the MIMO user detection and channel estimation device and method havefewer parameters compared with a common neural network, are easier to train, have higher accuracy and convergence compared with the AMP and have the lower computation complexity.

The invention discloses a monitoring region segmentation method for PM2. 5 inversion and a PM2. 5 inversion method. The monitoring region segmentation method comprises the following steps: with geographic static indexes and PM2. 5 measured concentration in a monitoring region being as samples, summarizing a correlation coefficient of each static index and the PM2. 5 measured concentration; with the correlation coefficient being as weight, carrying out index normalization on each static index to obtain a normalized parameter N_index, wherein the normalized parameter N_index is displayed in theform of raster data; and carrying out multi-scale segmentation on the raster data of the normalized parameter N_index, and determining an optimal segmentation scheme and dividing the monitoring regioninto a plurality of subregions according to the optimal segmentation scheme. The monitoring region is divided through the multi-scale segmentation algorithm, thereby reducing interference of spatialheterogeneity on parameter estimation; and for each of different research subregions, a specific particle concentration inversion model is established.

The invention discloses an online learning method of an artificial intelligence auxiliary OFDM receiver. The method comprises the following steps: carrying out offline training on a neural network inthe artificial intelligence auxiliary OFDM receiver; Inserting known online training bit data of the receiver into the bit data to be demodulated by the transmitter communicating with the OFDM receiver according to a fixed interval, and carrying out OFDM modulation and transmission; enabling The artificial intelligence auxiliary OFDM receiver to receive the signals and perform OFDM demodulation, and separating the signals through the two data collectors according to the same sequence as the transmitter to obtain receiving frequency domain data and frequency domain training data; Carrying out on-line training on the neural network in the artificial intelligence auxiliary OFDM receiver to obtain the neural network after the network parameters are updated on line; And inputting the frequencydomain received data into the neural network after the network parameters are updated online, outputting the estimation of the bit data to be demodulated, and performing judgment to recover the bit stream. By introducing neural network online learning, the robustness and receiving bit error rate of the receiver in different environments are improved.

The invention discloses a small sample terahertz image foreign matter detection method based on integrated deep learning, and mainly solves the problems that an existing method needs to manually design image features, the training process is complex, and foreign matter detection cannot be carried out on a small sample terahertz image with a small sample number. The method comprises the following specific steps: (1) making a small sample terahertz image data set; (2) amplifying the image training set; (3) constructing an integrated deep learning network; (4) training the integrated deep learning network; and (5) detecting the image test set. According to the method, the image features can be automatically extracted, the training process is simple, the situation that a certain type of samples in actual samples are particularly few is considered, foreign matter detection can be conducted on the terahertz images of the small samples, and the detection accuracy of the certain type of samples in the small samples is particularly few can be improved.

The embodiment of the invention discloses a method and a device for predicting scenic spot passenger flow volume based on a random forest algorithm, and the method comprises the steps: establishing arandom forest algorithm model with optimized parameters, debugging model parameters according to goodness of fit and an average standard error, and searching an optimal random forest algorithm model based on a grid search algorithm; and inputting the feature data set into the optimal random forest algorithm model to obtain the predicted future daily passenger flow volume of the scenic spot. According to the method and the device provided by the embodiment of the invention, the accuracy of the prediction model can be effectively improved, the timeliness of prediction is ensured, and managementof cities and scenic spots is facilitated.

The invention provides a facility light environment regulation and control method through fusion of a random forest algorithm. A photosynthesis regulation and control model through fusion of the random forest algorithm is established by using a photosynthesis rate model optimization method of the improved fish swarm algorithm by aiming at the problems of low fitting degree and complex fitting formula existing in the commonly used photosynthesis rate model (multivariate regression, linear fitting, etc.) at present; and a raspberry pie system framework and platform system capable of realizing algorithm transplanting is designed by aiming at the problems that the conventional embedded light environment control system cannot directly load the intelligent algorithm model, the equipment reliability is low and system response is slow, the equipment is mainly composed of a raspberry pie master control node, sensor monitoring nodes and LED light regulation nodes, and all the nodes realize information interaction through the ZigBee wireless technology. The deficiency of the light supplementing system in the conventional facility agriculture can be effectively compensated so as to have the advantages of great algorithm transplantability, fast light supplementing process response, high equipment reliability and convenient system upgrading in facility light environment regulation and control.

The invention discloses a method for screening sensitive variables influencing product quality in a complex industrial process based on step-by-step reduction, which belongs to the technical field ofsoft measurement and comprises the following steps of: selecting auxiliary variables influencing the product quality through expert knowledge and collecting data samples; calculating an auxiliary variable sensitivity index by comprehensively considering the variable correlation and the sensitivity of the variable to the working condition change, and preliminarily screening the sensitive variable influencing the dominant variable; and constructing a weighted cosine Mahalanobis system, and accurately screening key sensitive variables influencing the product quality. According to the method, thecorrelation and the working condition information of the variables can be accurately reflected, meanwhile, the redundancy of the variables is well reduced, the product quality prediction precision canbe improved, the complexity of the prediction model can be effectively reduced, and the method also has important significance for maintenance of the soft sensor model.

The invention discloses an entity relationship extraction method for wind tunnel fault text knowledge. The method comprises the following steps: 1, defining a knowledge structure; 2, dividing a training set and a test set; 3, performing entity labeling; 4, performing relation labeling; 5, performing data preprocessing; 6, inputting the training set into a model word embedding layer, and training a word embedding matrix; 7, inputting the word embedding matrix into a bidirectional GRU layer of the model, and extracting character-level features; 8, inputting a character-level feature set into a multi-head attention layer of the model, generating a weight vector, and multiplying the weight vector by the character-level features to obtain a sentence-level feature; 9, inputting the sentence-level feature into a model output layer to obtain a relation category; 10, performing iterative training; and 11, testing and evaluating the model; According to the wind tunnel fault entity relation extraction method based on a bidirectional GRU and a multi-head attention mechanism, knowledge is extracted from a wind tunnel fault text, conversion from unstructured fault data to structured data is achieved, and the utilization efficiency of the text knowledge in the wind tunnel health monitoring and fault diagnosis process is improved.

A data-driven unit commitment intelligent decision-making method based on an E-Seq2Seq technology, comprising the steps: 1, carding the types and structures of input and output sequences of a unit commitment model, and forming a unit commitment elastic multi-sequence mapping type sample; 2, constructing a unit commitment deep learning model based on an E-Seq2Seq technology by taking the GRU as a neuron; and 3, carrying out deep learning on the unit combination deep learning model. Compared with the existing intelligent decision-making method, the data-driven unit commitment intelligent decision-making method disclosed by the invention can consider the influence of multi-type and multi-dimensional input factors on the unit commitment decision-making at the same time, and can adapt to the elastic change of sample types and dimensions, so that the decision-making precision is higher.

The invention relates to an air conditioner load prediction method and system.The method comprises the steps of acquiring a sample set, dividing data of the sample set into a training data set and a prediction data set, wherein the sample set at least comprises sample data and label data corresponding to the sample data; extracting influence factors in the sample data, and screening a plurality of characteristic indexes from the influence factors as an input sequence by adopting grey correlation analysis; obtaining a basic model, wherein the basic model comprises a bidirectional LSTM model and a bidirectional GRU model which are arranged in sequence; inputting a corresponding input sequence in the training data set into the basic model for training to obtain a prediction model; and inputting a corresponding input sequence in the prediction data set into the trained prediction model for load prediction to obtain a predicted value of the air conditioner load. Effective dimensionality reduction is performed on a large amount of data by adopting grey correlation analysis, and the selected basic model is matched, so that the prediction precision can be improved, and the training time can be shortened.

The invention relates to a thyroid malignant nodule detection method based on deep learning, the method is used for the auxiliary diagnosis of thyroid malignant nodules, and the automatic labeling ofa region of interest can be achieved after the training of a large amount of data under the characteristics of low medical image resolution, low precision and low target and background identificationdegree. Errors caused by subjective factors can be effectively reduced, and radiologists can be helped to diagnose quickly and accurately. The detection precision and speed are considerable, and the method has the potential of clinical application.

The invention discloses an implementing method of a sliding directional drilling simulator. The implementing method comprises the following steps that a sliding directional training data set is obtained, and meanwhile, random noise is generated; data processing and data arrangement are carried out, and meanwhile, a generation model generates data samples through the random noise; data sources arejudged through a discrimination model, and a generative adversarial network (GAN) forms a sliding directional data model is formed; and multiple categories of sliding directional data are automatically generated, and effective amplification data are obtained to form an amplified sliding data set. According to the implementing method, effective and real data are provided for the discrimination model by providing the sliding directional training data set and carrying out data processing and data arrangement, the data samples are generated through the generation model, and the generated data areprovided for the discrimination model. The discrimination model judges the data sources, the GAN forms the sliding directional data model, the multiple categories of sliding directional data are generated through the sliding directional data model, and thus the purpose of simulating sliding directional parameters through the adversarial network is realized.

The invention relates to a load evaluation method of a roof steel plate structure based on ResNet. The method comprises the following steps that Ansys finite element analysis software is used for conducting finite element simulation on a steel plate structure of a roof, a strain collection position is determined, and strain values corresponding to different loads are obtained; performing data screening and cleaning on the obtained strain values, establishing a roof panel mechanism load-strain database, and dividing a training set and a test set; training the training set data by using ResNet34, and testing in the test set to obtain a test result; and the average relative error is used as an evaluation index. After data preprocessing is completed, the training set is used for training the ResNet34 model, and the model is evaluated on the test set; and then the trained ResNet model is used for load prediction of the roof steel plate structure. According to the method, the roof steel plate structure load is accurately detected, the practicability is high, and the accuracy is higher than that of a traditional load calculation method. The method can be widely applied to the field of roof steel plate structure load calculation.