253results about How to "Avoid fit problems" patented technology

The invention discloses a radar radiation source recognition method based on a deep learning strategy and a multitask learning strategy and mainly aims to solve the problem that recognition accuracy is low in the prior art. The method comprises the implementation steps that 1, an original radar radiation source signal is subjected to data preprocessing; 2, envelope features, fuzzy function features, slice features, cyclic spectrum features and frequency spectrum features of the preprocessed radar radiation source signal are extracted, and values of the features are linearly transformed into [0,255] and saved as an image set; 3, a convolutional neural network (CNN) is designed, and the multitask learning strategy and a random inactivation strategy are introduced into the CNN; and 4, four feature training sets are used to train the CNN, then four trained CNN models are utilized to classify four feature test sets, and a radar radiation source recognition result is output. The method is high in recognition accuracy and can be applied to electronic intelligence reconnaissance, electronic support reconnaissance and radar threat warning systems.

The invention discloses a text detection method of a document image in a natural scene. Commonly used Chinese characters are selected to make Chinese character pictures, a dataset 1 is formed, randomrevolving and cropping operations are carried out on labeled document images, then a manner of Poisson cloning is used to fuse different background images, and a dataset 2 is formed; the dataset 1 isadopted to carry out training of a text classification model on a VGG16 network, and after the model converges, obtained parameters are used to initialize a fully convolutional neural network model, and the dataset 2 is used to train the model; the trained fully-convolutional neural network model is used to process the image, a classification situation of each pixel point is obtained according toa maximum probability method, and a text-non-text binary image is formed; a method of connected regions is used to obtain text regions, the original image is binarized, and only text information in the text regions in the text-non-text region binary image is extracted to obtain a text binary image; the image is corrected through a maximum variance method; and projection is carried out again on thecorrected image, and the text-non-text region binary image is refined.

The invention discloses a rolling bearing fault diagnosis method based on a convolutional neural network (CNN). By aiming at problems of rolling bearing characteristic components such as easy submergence and difficulty in extraction and combining with rolling bearing signal own and large monitoring data quantity and other characteristics, the CNN is introduced in the rolling bearing fault diagnosis. By short time Fourier Transform, a motor vibration signal is converted into a time frequency spectrogram to be adapted to a CNN network training sample format, and then mass sample data having labels used to express different faults is established, and therefore sample diversity is guaranteed, and network overfitting is prevented. The CNN network having a proper layer number is established, and parameters are initialized, and then the preprocessed samples are input in the CNN for forward propagation. By combining with predetermined label calculation errors, a network weight is adjusted by using an error reverse propagation algorithm, and then after a plurality of times of iterations, the network used for the interconnection between the signal and equipment is established, and therefore the rolling bearing fault accurate diagnosis is realized.

The invention belongs to the technical field of image processing, and specifically discloses a block-based shielded face recognition algorithm. The algorithm comprises the following steps of: in an offline training stage, carrying out preprocessing operations such as face detection, geometric face normalization, illumination normalization, extracting four blocks such as a left eye, a right eye, anose and a mouth from a face image on the basis of the preprocessing result, training a network model of each block, extracting corresponding features, training a shielding judgement network of each block to obtain a shielding judgement result, finally fusing the features according to the shielding judgement result of each block, and constructing a K-D tree feature index; and in an online recognition stage, extracting face image features by adoption of a method same as the offline training stage, and carrying out feature query through a K-D tree index manner so as to obtain a recognition result. Experiment results prove that the algorithm has better correctness.

The invention discloses a CNN and selective attention mechanism based SAR image target detection method. An SAR image is obtained; a training data set is expanded; a classification model composed of the CNN is constructed; the expanded training data set is used to train the classification model; significance test is carried out on a test image via a simple attention model (a spectral residual error method) of image visual significance to obtain a significant characteristic image; and morphological processing is carried out on the significant characteristic image, the processed characteristic image is marked with connected domains, target candidate areas corresponding to different mass centers are extracted by taking the mass centers of the connected domains as the centers, and the target candidate areas are translated within pixels in the surrounding to generate an target detection result. According to the invention, the CNN and the selective attention mechanism are applied to SAR image target detection in a combined way, the efficiency and accuracy of SAR image target detection are improved, the method can be applied to target classification and identification, and the problem that detection in the prior art is low in detection efficiency and accuracy is solved mainly.

The invention discloses a convolutional neutral network-based worker absence-from-post detection method. According to different characteristics extracted from monitoring images, the images are classified via an image classification algorithm, and whether workers are absent from posts can be detected via a convolutional neutral network training model. Monitoring workers in a company can be helped to process monitoring data efficiently via the worker absence-from-post detection method, whether the workers are at the posts can be determined, and whether the workers are absent from the posts can be accurately detected in working environment; a conventional convolutional neutral network structure can be improved so as to be applied to indoor working scenarios such as companies, enterprises and the like; worker absence-from-post detection can be realized, and the worker absence-from-post detection method is well applied to solving the problem.

The invention discloses a self-coding network and a training method thereof, and a method and a system for detecting abnormal power consumption, wherein the training method comprises the following steps: splicing the sample power data by a sliding window to obtain a training sample set, and marking the training sample set containing the surveyed user to obtain a labeled sample, wherein the unlabeled training sample is a non-labeled sample; carrying out unsupervised training on the self-coding network by unlabeled samples, and obtaining the initialization parameters of the self-coding network,then taking the discrete class tags obtained from the coding layer of the coding network as classifiers, carrying out supervised training on the classifiers by the labeled samples, and updating the parameters of the coding layer to obtain the trained self-coding network; then, utilizing the trained self-coding network to detect the power data of the user to be measured, and judging whether the user to be measured abnormally uses power. The invention can mine abnormal information in low-density electric power data, avoid noise data interference, and improve abnormal detection accuracy.

The invention discloses an image classification method and system based on migration learning. The method comprises the steps that 1, a feature similarity known training set A is utilized to make a training set B of a migration network through a support vector machine; 2, a migration learning network is constructed; 3, the training set B classified in the step 1 is used as a training learning setof the migration learning network, and a migration learning network model with high robustness and good accuracy is obtained through training; and 4, a to-be-classified dataset is introduced into thetrained migration learning network model, and a final classification result is obtained and marked with a tag. Through the image classification method and system, the requirement that a big sample dataset is needed to serve as input when a common RGB image is trained through deep learning is overcome, the problems of overfitting and locally optimal solutions in the training process are avoided, and classification precision is improved to some extent compared with a traditional classification algorithm.

The invention discloses a significant guidance and unsupervised feature learning based image classification method and belongs to the field of machine learning and computer vision. The image classification method comprises significant guidance based pixel point collection, unsupervised feature learning, image convolution, local comparison normalization, spatial pyramid pooling, central prior fusion and image classification. With the adoption of the classification method, representative pixel points in an image data set are collected through significant detection; the representative pixel points are trained with the sparse self-coding unsupervised feature learning method to obtain high-quality image features; features of a training set and a test set are obtained through image convolution operation; convolution features are subjected to local comparison normalization and spatial pyramid pooling; pooled features are fused with central prior features; and images are classified by adopting a liblinear classifier. According to the method, efficient and robust image features can be obtained and the classification accuracy of various images can be significantly improved.

The invention discloses a remote sensing image classification method based on a deep fusion convolutional neural network, and the method comprises the steps: constructing an original remote sensing image into a data set, carrying out the preprocessing of the original remote sensing image, dividing the preprocessed image into a training set, a test set and a verification set, and carrying out the data augmentation of the training set; constructing a deep fusion convolutional neural network; training to obtain an optimal network model; and classifying the actually measured remote sensing imagesby using the optimal network model. The invention provides a new classification method. A new deep fusion convolutional neural network is constructed; an improved encoder-decoder model is combined with a VGG16 model to obtain a VGG16 model; the model fuses the deep features and the middle-layer features of the remote sensing image, so that the defect of low classification precision caused by single or redundant feature extraction of the remote sensing image in the prior art is effectively overcome, the advanced feature expression capability of the target is obtained by establishing the novel network model, and the classification accuracy of the remote sensing image is improved.

Disclosed is a method for acquiring nuclear fuel assembly resonance parameters. The method includes the following steps: building a commonly-used nuclide multi-group database and a changing curve of resonance peaks along with energy; distinguishing isolated peaks and dense peaks, wherein the interval where the isolated peaks are located is a resonance energy region low energy section with obvious resonance interference, and the interval where the dense peaks are located is a resonance energy region high energy section without obvious resonance interference; solving a subgroup total cross section, a subgroup partial cross section, subgroup probability and corresponding multi-group data coupling in the resonance energy region high energy section by using a subgroup method; coupling calculation of the resonance energy region high energy section and calculation of the resonance energy region low energy section through inter-group scattering from the high energy section to the low energy section, acquiring the resonance parameters of the high energy section, and then calculating a scattering source and a fission source from the high energy section to the low energy section; acquiring a multi-group neutron energy spectrum in a low energy region by using a wavelet expansion method; and merging and continuing the total cross section and the partial cross section by using the multi-group neutron energy spectrum of the resonance region low energy section to acquire the nuclear fuel resonance parameters of the low energy section. By means of the method, resonance parameters of nuclear fuel assemblies with any resonance material composition, any geometry and any quantity of resonance regions can be effectively acquired.

The invention discloses a modulation identification method based on a convolutional neural network. The method comprises the following steps: S1, selecting a modulation signal data set and designing aconvolutional neural network model structure; s2, constructing a residual error unit in the convolutional neural network model in a residual error connection mode; s3, performing batch normalizationon the data in the network layer in batches in the convolutional neural network model; s4, setting convolutional neural network parameters; s5, training the convolutional neural network, and randomlylosing data in the training set; s6, substituting the signal into the trained convolutional neural network and performing modulation identification; the parallel network is provided by combining the time sequence feature extraction capability of the time convolution network and the feature expression enhancement capability of the attention mechanism, so that the spatial features extracted from theconvolution neural network and the time sequence features extracted from the time convolution network are fused, and the modulation recognition performance is further improved.

The invention provides an electric power equipment infrared image real-time detection and identification method based on artificial intelligence. The method comprises the following steps of S1, acquiring infrared images of various types of electric power equipment through an infrared thermal imager; S2, preprocessing an acquired image to form an effective power equipment infrared image data set; S3, performing target label processing on the obtained data set; dividing the data set into a training set and a test set; S4, constructing an improved YOLOv4 real-time detection model for detecting and identifying an infrared image target of the power equipment; S5, training and parameter adjustment of the model are carried out by using a training set in the data set; and S6, performing target detection and identification on the trained model by using a test set in the data set to prove the effectiveness of the model; through the above steps, automatic detection and identification of infraredimages of various types of power equipment are realized; the accuracy degree of identification can be greatly improved, detection and identification efficiency is improved, and operation resources areeffectively utilized.

The invention provides a named entity recognition model training method and device. The named entity recognition model training method comprises the steps of obtaining labeled training data and unlabeled training data; training a target named entity recognition model according to the labeled training data; inputting the unlabeled training data into the target named entity recognition model to obtain at least one entity word output by the target named entity recognition model and a confidence score corresponding to each entity word; determining a target entity word according to the confidence score corresponding to each entity word, and labeling the unlabeled training data according to the target entity word to generate newly added labeled training data; training the target named entity recognition model contineusly according to the newly added labeled training data. According to the method, through a weak supervised learning mode, the number of the labeled training data is increased, model training overfitting is effectively prevented, and meanwhile the cost of manually labeling the labeled training data is reduced.

The invention relates to a sorting method of secondary batteries. The sorting method comprises the following steps: selecting featured charge states corresponding to to-be-sorted batteries, acquiring electrochemical alternating-current impedance spectroscopy of the to-be-sorted batteries, extracting two-dimensional feature parameters of the to-be-sorted batteries, and then determining battery monomers combined to form a battery pack. According to the internal features of the parts, from a charging platform to a charging cut-off voltage, of the secondary batteries, the consistency of the essentially sorted batteries is better than that of batteries sorted by a conventional method, the integral performance of the battery pack is improved, and the service life is prolonged.

The invention discloses a road scene segmentation method based on a residual network and expanded convolution. The method comprises: a convolutional neural network being constructed in a training stage, and a hidden layer of the convolutional neural network being composed of ten Respondial blocks which are arranged in sequence; inputting each original road scene image in the training set into a convolutional neural network for training to obtain 12 semantic segmentation prediction images corresponding to each original road scene image; calculating a loss function value between a set formed by12 semantic segmentation prediction images corresponding to each original road scene image and a set formed by 12 independent thermal coding images processed by a corresponding real semantic segmentation image to obtain an optimal weight vector of the convolutional neural network classification training model. In the test stage, prediction is carried out by utilizing the optimal weight vector of the convolutional neural network classification training model, and a predicted semantic segmentation image corresponding to the road scene image to be subjected to semantic segmentation is obtained. The method has the advantages of low calculation complexity, high segmentation efficiency, high segmentation precision and good robustness.

The invention discloses a mid-air gesture recognition method based on an inertial sensor. The method includes the steps that a mid-air gesture signal sequence is extracted aiming at sensing signals acquired by the inertial sensor, then data preprocessing is carried out, then a training sample set, a verification sample set and a test sample are acquired, an LSTM-RNN model is subjected to parameter initialization, the training sample set is used for training the LSTM-RNN model, in the training process, verification samples in the verification sample set are input in the LSTM-RNN model trained in the iteration process, the iteration frequency is controlled according to the recognition error rate of the verification sample set, and a final LSTM-RNN classifier is obtained; finally the test sample is input in the final LSTM-RNN classifier, and a gesture corresponding to the test sample is recognized through the final LSTM-RNN classifier. The method has the advantages of being higher in mid-air gesture recognition precision and accuracy.

The invention discloses a camera intrinsic parameter calibration method, and relates to the field of computer vision, in particular occasions for calibrating cameras by ranging through a board camera. The technical scheme is that: a regular octagonal template is adopted as a calibration template; the template is shot by the camera in at least three directions, and at least three images of the template can be acquired; coordinate values of nine characteristic points in a single template image are acquired by a corner detection method which can be accurate to a sub-pixel level, and coordinates of four end points in directions of four diagonals are solved so as to solve coordinates of two ring points; coordinates of at least six ring points are obtained according to the at least three images of the template in different directions so as to obtain a camera intrinsic parameter matrix, and the camera intrinsic parameter is obtained. The camera intrinsic parameter calibration method self-calibrates the camera intrinsic parameter by adopting a single regular octagonal template and is a practical calibration method.

The invention discloses an automatic searching and shimming method based uneven magnetic field fitting linearity. The method includes: using a gradient echo pulse sequence to measure the uneven image of a spatial magnetic field, using the same to fit and represent the virtual spectrogram line and related performance indexes of the current magnetic field evenness; using the related performance indexes of the virtual spectrogram line as the evaluation standard for obtaining good or bad magnetic field evenness according to the multidimensional spatial vectors built by a shimming coil, performing multidimensional descending simplex method iteration search according to the evaluation standard, and returning the optimal shimming current. The method has the advantages hat during shimming, the performance indexes such as half-width and symmetry of spectrum can be fully considered; compared with the gradient shimming using gradient echo imaging, the shimming method does not need to consider measurement and fitting the shimming coil field graphs, and the shimming difficulty and limit requirements by the aid of gradient imaging are lowered; automatic fitting and searching and convergence conditions allows the method to get rid of complex manual configuration, and shimming speed and efficiency are increased.

The invention provides an impact localization method based on phase-sensitive optical reflection and deep learning of a convolutional neural network. The impact localization method includes the following steps: 1) two optical fiber monitoring network topological structures for impact load localization, based on a phase-sensitive optical time domain reflection principle; 2) structure design of a phase-sensitive optical time domain reflection sensing probe for impact load monitoring; 3) construction of a distributed phase-sensitive optical time domain reflection sensing impact monitoring systemand meshing of a thin plate monitoring area; 4) phase-difference-based phase-sensitive optical time domain reflection technique for impact load localization; 5) generation of an impact response samplelibrary based on a Phi-OTDR sensor; 6) data pre-processing and deep learning convolutional neural network design; and 7) using the trained deep learning convolutional neural network to identify the impact response data of the Phi-OTDR sensor.

The present invention discloses a CO emission forecasting system and method for a garbage incineration boiler with a circulating fluidized bed. On the basis of using a running mechanism of a garbage incineration boiler with a circulating fluidized bed and knowledge hidden in historical running data, a Gamma Test algorithm and a random forest integrated modeling method are adopted, a rapid, economic and adaptive updating system and method are constructed to perform real-time forecasting on CO emission of flue gas at the tail of a boiler, and cumbersome and complicated mechanism modeling work is avoided. A dynamic change characteristic of CO emission is characterized by using a non-linear mapping ability, a generalization ability and a real-time forecasting ability of a random forest algorithm, which provides a new approach for operators and designers to understand the change characteristic of CO emission; and the Gamma Test algorithm is used to acquire an optimal training sample, so that over-fitting and under-fitting situations of the model during training are avoided. The whole modeling process has a clear logic, needs to set fewer parameters, has high modeling automation and is easy to master and promote.

The invention discloses a construction method of a pulmonary thromboembolism risk prediction model based on single nucleotide polymorphism, an SNP locus combination and application. The construction method of the prediction model comprises the following specific steps of S1, carrying out sample collection and gene detection; S2, carrying out data quality control and whole genome association analysis (GWAS); S3, carrying out meta analysis by combining the genome data of an external population; S4, screening an SNP locus combination with a prediction value; and S5, building a regression model, and carrying out training and testing. According to the prediction model, 48 SNP site combinations are obtained, at least one SNP site combination can be applied to pulmonary thromboembolism risk assessment or screening products, pulmonary thromboembolism risk prediction of Asian people, especially Chinese people can be achieved, and the detection method is simple, convenient and easy to implementand convenient to use clinically.

The invention particularly relates to a hybrid cell type identification method based on fine granularity identification, comprising the following steps: a fine granularity identification convolutionalneural network model and a cell image database are established in advance; the cell image database comprises a hybrid cell image; the hybrid cell image is an image including a plurality of types of cells; the hybrid cell type identification method comprises the following steps of: 1, collecting mixed cell images; 2, inputting the mixed cell image into a fine-grained recognition convolution neuralnetwork model to obtain a cell type thermogram; 3, performing threshold that mixed cell image to obtain a binary image of the cell region; 4, combined with binary image of cell region and thermogramof cell species, the cell species identification results being obtained. The invention accurately identifies cell species according to the specificity of cell morphological characteristics, and avoidsthe shortcomings of the traditional cell species identification method that takes a long time and the process is tedious. The model can learn the morphological characteristics of fine-grained cells and identify cell types through texture information, which has high recognition accuracy and robustness.

The present invention is applicable to the field of information push, and provides a method and device for estimating the click-through rate of an advertisement. The method includes: obtaining user characteristics and advertisement characteristics of advertisement information to be pushed, and encoding to obtain user characteristic vectors and advertisement characteristic vectors; The limited BFGS algorithm uses interactive fixed parameters to iteratively calculate the weight vector in the logistic regression algorithm; estimates the click-through rate of the advertisement information according to the weight vector, user feature vector and advertisement feature vector. Compared with the prior art, the method of iterating through interactive fixed parameters in the present invention can effectively constrain the parameters and avoid the over-fitting problem in the logistic regression algorithm, thereby improving the efficiency of estimating the click-through rate of advertisements and improving the accuracy of estimating .

The invention discloses a bridge handling ship machine, which comprises the headframe, the crossbeam, the hanger and the handling device. The crossbeam crosses the headframe transversely; the handling device is under the crossbeam and the headframe. The handling device includes the grab, the funnel, the feed machine, the bifurcation bucket, the feed belt machine and the telescopic chute. The grab is above the crossbeam, the outlet of the funnel connects the feed machine, the bifurcation bucket is behind the feed machine, one outlet of the bucket connects the feed belt machine, and the other outlet connects the inlet of the telescopic chute. The invention resolves the bias running of the belt machine, and realizes the integration of loading and unloading work.

The invention discloses a target detection model training method and a rapid target detection method. The training method comprises the following steps: adding a target region feature enhancement layer during training; performing feature extraction on the training sample by using a feature extraction unit, and averaging channels of a feature map output by a previous-level feature extraction unit to obtain a first feature value matrix with a normalized channel number; traversing each pixel point in the training sample to generate a second eigenvalue matrix; multiplying the element values of thefirst eigenvalue matrix by the element values of the second eigenvalue matrix to obtain a third eigenvalue matrix; multiplying the third eigenvalue matrix by a preset adjustment function, then performing element value addition on each channel eigenmatrix of the eigenmap to obtain a target enhanced eigenmap, and inputting the target enhanced eigenmap into a next eigenextraction unit. According tothe invention, the network is fully trained, the relationship between the network depth and the detection precision is balanced, the background perception capability of the feature map is enhanced, the detection precision is high, the calculation is simple, and hardware platform transplantation is facilitated.

The invention belongs to a hydraulic control technology, and particularly relates to a load maintaining control valve. The load maintaining control valve is capable of, through designing a relative conical surface structure of a main valve element and a control valve element and isolating a main valve element oil way and a pilot control oil way, realizing extremely low leakage performance, using alarge action area ratio structure of a pilot push rod and the control valve element to realize low-pressure control of a high-pressure main valve, and adequately using multi-functional integration ofan internal structure and each part structure of the main valve element to complete structure simplification and a compact design, reducing the number of parts and avoiding matching a high-precisionslide valve, thereby reducing processing cost. So a low-leakage requirement design and a cost reducing and structure compact design of the load maintaining control valve are completed well, and the load maintaining control valve has extensive application prospect.

The invention discloses a training method and device for a multi-fault prediction network model of a power information system, and the method comprises the steps: obtaining an alarm data set of a timesequence, carrying out the data enhancement of the alarm data set, and obtaining an enhanced training sample set; obtaining an input sample for model training and a target output sample correspondingto the input sample based on the training sample set; and performing iterative training on a preset neural network model based on the input sample, the target output sample and a preset network modelloss function to obtain a multi-fault prediction network model. According to the method, data feature equalization is realized by performing data enhancement processing on the original data set, andthe multi-fault prediction network model obtained by performing model training fitting based on the training sample set after data enhancement has higher prediction precision and a more stable prediction effect.

The invention discloses an adjustable clamp for a rubber tensile set test. The adjustable clamp comprises a track bottom plate, a fine adjustment sliding block and a rough adjustment sliding block, wherein the fine adjustment sliding block and the rough adjustment sliding block are arranged on the track bottom plate. The fine adjustment sliding block and the rough adjustment sliding block are provided with cover plates used for clamping a rubber piece sample, an adjusting screw penetrates through a step hole of a base plate and then is screwed into and penetrates out of a baffle, the base plate is fixed to the outer side of the fine adjustment sliding block, and the baffle is fixed to the track bottom plate. The adjustable clamp has the advantages that the rubber piece sample can be rapidly and quickly stretched to the distance 1.5 times of the original distance, and thus the accurate deformation amount can be obtained. By means of the clamp, test data (the rubber deformation amount) can be accurately obtained, and the positioning speed of test clamping can be increased; meanwhile, multiple adjustable clamps for the rubber tensile set test and samples can be put in a test box, so that the utilization rate of test equipment is greatly increased, the test efficiency is effectively improved, and a large amount of test cost is also saved.

The invention relates to the technical field of fault distinguishment in an electric power system, in particular to a method for distinguishing transient state instability of the electric power system in real time based on voltage track after disturbance. The method comprises five steps, namely, selecting a voltage observation node, identifying voltage drop after fault clearance, complexly integrating the voltage track, distinguishing the transient state instability of the electric power system and identifying the fact that local voltage is over-low. The method for distinguishing the transient state instability of the electric power system in real time based on the voltage track after disturbance, disclosed by the invention, aims to solve a problem of system transient state instability caused by large disturbance out of a two-defense-line control range based on a measurement configuration of the existing phasor measurement unit ( PMU ) and a wide area measurement system ( WAMS ) in a power grid, and has the advantages of simple algorithm, low calculated quantity, less dependence on data precision and high reliability.