206results about How to "Improve generalization ability" patented technology

The invention provides a business risk assessment method and device, and a risk control system. The method comprises steps that to-be-assessed business data is received, and a risk behavior rule set matched with the business data is identified according to a preset risk rule engine; for multiple rules of the rule set, a risk value of each corresponding rule is assessed according to a pre-established risk assessment model; the risk values of the corresponding rules are accumulated, the accumulated risk value sum is compared with a total risk score threshold of each preset risk grade, and a risk grade of the business data is determined. The method is advantaged in that generalization capability and processing efficiency of a risk system can be improved.

Provided are a training method and device and an identification device for intent recognition. The method includes: acquiring a corpus text vector corresponding to a corpus in a corpus text; constructing a joint loss function formula of a training model; acquiring training data; performing segmentation processing on the training data and then mapping to the corresponding corpus text vector, recorded as a training vector; using the training model to predict a training sample vector, calculating a loss function value of each training model according to a prediction result; calculating the jointloss function value based on the loss function value of each model, judging whether the joint loss function value is less than a set threshold, if yes, ending the training, if not, updating each modelparameter, and continuing iterative training. Therefore, the generalization capability of intent recognition can be improved as much as possible, and semantic ambiguity and fault tolerance problems can be solved.

Provided is a method of forecasting the wind speed of a wind power station. According to the method, firstly, attribute reduction is conducted on various factors affecting the wind speed of the wind power station in a fuzzy rough set method, redundant information is removed, and input variables of neural network forecasting models are obtained; then, by means of the clustering method of conducting an improvement through the weighting Euclidean distance, data having high similarity are extracted to serve as training samples of the neural network forecasting models, and various forecasting models are trained by using the clustered data; finally, the matched forecasting model is selected for forecasting the wind speed according to a current attribute value. On the basis of traditional neural network forecasting models, the method optimizes the input variables and the training samples, affecting forecasting performance of neural networks, of the models, and the generalization ability of the models is enhanced greatly. Test results show that the method can greatly enhance the forecasting performance of the neural networks and effectively enhance forecasting accuracy of the wind speed of the wind power station.

The invention relates to an ancient writing identification method and system based on deep learning, which comprises the following steps of: obtaining an ancient writing image data set and making a classification label; preprocessing an image in the image data set, and expanding the data set through image augment; extracting feature on the image data set which has been preprocessed and image augmented through a deep neural network, and classifying extracted features through a classifier so as to realize evolution identification of the seal script of different times; and using artificial designfeatures for performing auxiliary training by using template matching method so as to further classify classification results of the classifier. According to the ancient writing identification methodand system based on deep learning, correct classification can be achieved more accurately according to improved deep convolutional neural network model, and performing auxiliary training by virtue ofthe traditional artificial design features, classification accuracy is further improved, and better experience is provided for the user group with ancient writing identification needs.

The invention discloses a magnetic resonance reconstruction method based on deep learning and data consistency, and relates to the field of magnetic resonance reconstruction methods. The method comprises the steps: 1, collecting K-space data and integrating the K-space data into a network, formed by the sequential stacking of a convolutional neural network and a data consistency layer, for completing the network construction; 2, converting the undersampled data in the K-space data into a roll pleat image as the input of the built network, converting the full-sampled data of the K-space data into a complete image as the tag data of the built network, and obtaining a mapping relation between the network input and output through a back propagation training network; 3, inputting the corresponding images of a test set into the trained network, and performing the forward propagation to obtain an output image to complete the magnetic resonance reconstruction. The method solves the problems ofpoor reconstruction performance and stability caused by a condition that a conventional magnetic resonance reconstruction method based on deep learning does not fully utilize the collected data and can only deal with a single channel, achieves implementation supervision, improves the learning ability, and improves reconstruction performances.

The invention discloses a cross-database micro-expression recognition method and a device based on an optical flow attention neural network. The method comprises the following steps: (1) acquiring twodifferent micro-expression databases as a training set and a test set; (2) converting the micro-expression video into a face image sequence; (3) extracting a starting frame, a peak frame and a termination frame from each face image sequence, calculating the starting frame and the peak frame to obtain a first single-channel optical flow graph, and calculating the peak frame and the termination frame to obtain a second single-channel optical flow graph; (4) forming a fusion feature graph by the first single-channel optical flow graph, the second single-channel optical flow graph and the peak frame of each face image sequence; (5) establishing an optical flow attention neural network, and taking the fusion feature graphs corresponding to the training set and the test set as input for training; and (6) processing the micro-expression video to be identified to obtain a fusion feature map, and inputting the fusion feature map into the optical flow attention neural network to obtain a micro-expression category. The method is high in generalization capability and high in recognition accuracy.

The invention belongs to the technical field of sleep state monitoring, and discloses an intelligent sleep phase detection and sleep quality evaluation system and method. The intelligent sleep phase detection and sleep quality evaluation system comprises an eeg signal acquisition module, a background server and an on-line service platform, wherein the eeg signal acquisition module is used for acquiring sleep eeg data, and is used for completing acquisition of the sleep eeg data through eeg signal detecting components; the background server is used for receiving acquired original sleep eeg datato complete the treatment of sleep eeg signals, so that sleep eeg automatic phase separating is realized, and sleep quality is evaluated; and the on-line service platform is used for tracking the sleep trend of a user and providing services including sleep psychological consultation, expert suggestion and the like. The intelligent sleep phase detection and sleep quality evaluation method comprises the following steps of obtaining an eeg signal, performing pretreatment, performing feature extraction, performing mode recognition and classification and outputting the sleep phase. The evaluationresult can be pushed to a mobile terminal so that professional sleep psychological consultation and expert advice are provided for the user, are completely independent of the working experience of anoperator, and can be used as beneficial supplement and explanation for clinical diagnosis.

The invention discloses a polarization SAR image classification method based on a complex contour wave convolution neural network, and a problem of low classification accuracy in the prior art is mainly solved. The method comprises the steps of (1) inputting and normalizing a polarization coherent matrix T of a polarization SAR image to be classified, (2) according to the normalized matrix, constructing characteristic matrixes of a training data set and a test data set, (3) constructing a complex convolution neural network, and thus obtaining a complex contour wave convolution neural network, (4) training the complex contour wave convolution neural network by using the training data set, and obtaining a trained model, and (5) inputting the characteristic matrix of a test data set into the trained model to carry out classification, and obtaining a classification result. According to the method, the convolution neural network is extended to a complex domain to carry out operation and extract image characteristics of multiple scales, multiple directions and multiple resolution characteristics, the classification precision of the polarization SAR image is effectively improved, and the method can be used for target detection and identification.

The invention relates to an optic fiber gyroscope random drift modeling method based on a locally variable integrated neural network. The method comprises the following steps: (1) establishing an integrated neural network model; (2) obtaining a learning sample; (3) training the integrated neural network; and (4) modeling to the random drift of an optic fiber gyroscope. The method firstly uses theFLP/wavelet filtering method to denoise the original output signals of the optic fiber gyroscope and form a data pretreatment system with good denoising effect and wide application range; secondly the locally variable integrated neural network is used in modeling, the integrated neural network overfits the training samples, the defects of low precision and poor stability of the sub-neural networkare overcome; the locally variable weight method based on temperature is introduced to ensure that the network can adjust the integrated weight matrix according to the change of temperature; and the optimum network modeling precision can be ensured at different temperatures, the generalization ability of network is increased, and the dynamic modeling of optic fiber gyroscope random drift is realized.

The invention discloses a partial discharge pattern identification method for a crosslinked polyethylene cable, and belongs to the technical field of online monitoring and fault diagnosis of electrical equipment. The partial discharge pattern identification method comprises the following steps: adopting online or offline monitoring data of a cable, performing partial discharge signal feature extraction based on the natural mode singular value decomposition technology, and selecting the sample data to realizing learning of a cable partial discharge detection limit learning machine model, so as to realize partial discharge detection of the XLPE (Crosslinked Polyethylene) power cable. The partial discharge pattern identification method disclosed by the invention aims to running state of seven cables, based on the natural mode singular value decomposition technology, performs feature extraction on the corresponding raw signals corresponding to different running states of the cables, and determines the target vector form of the sample data, so as to complete partial discharge sample data set construction of the XLPE power cable, and learn and test the constructed partial discharge detection ELM model of the cable. The partial discharge pattern identification method disclosed by the invention can accurately and quickly identify the insulation defect and partial discharge mode of the power cable, so as to ensure safe and healthy operation of cable equipment, and provide the basis for the power cable repair scheme.

The invention discloses an adaptive adversarial learning-based urban traffic scene semantic segmentation method and system. The method comprises steps of obtaining training data of the semantic segmentation model, and preprocessing the game synthesis data set GTA5 to obtain a new synthesis data set SG-GTA5 which is close to urban scene real data set Cityscapes in distribution; constructing a generative adversarial network model for semantic segmentation; based on the training data set, performing self-adaptive confrontation learning on the generative adversarial network model, using self-adaptive learning rates in confrontation learning of different feature layers, adjusting loss values of the feature layers through the learning rates, and then dynamically updating network parameters to obtain an optimized generative adversarial network model; and carrying out verification on the city scene real data set CityScapes. According to the method, the semantic segmentation precision of a complex urban traffic scene which lacks annotation information and has more scale targets can be improved, and the generalization ability of a semantic segmentation model is enhanced.

The invention discloses an air quality prediction method based on an extreme learning machine. The air quality prediction method comprises the steps that step 1, air quality data requiring prediction analysis are acquired and divided into a training data asset and a test data set; step 2, CNN is trained based on the training data set and a model performing in each validation set is selected to act as a feature extraction model; step 3, the activation value of the first layer of the CNN full connection layers acts as the input of GBELM, the GBELM is trained and the GBELM performing the best in each validation set is selected to act as a prediction model; and the GBELM is utilized to replace the CNN full connection layers obtained in the step 2 so as to obtain a final air quality prediction model; and step 4, the test data are inputted to the air quality prediction model to calculate the abstract features of the test data and then inputted to the GBELM to obtain the output value of each ELM to perform summation so as to obtain the prediction result. With application of the technical scheme, the air quality prediction method has high prediction accuracy.

The invention relates to the processing technology of image information, in particular to a method and a device for classifying sheet media quickly according to the image information. The method comprises the following steps of: acquiring the image information of the sheet media to be identified; performing normalization processing on the acquired image information according to a preset standard image dimension so as to obtain the image information to be identified; performing feature extraction on the image information to be identified to form a feature vector; and sending the feature vector into a classifier which is trained by standard sheet media to obtain a classification result, wherein the normalization processing comprises the following steps of: identifying whether the dimension of the acquired image information is smaller than that of the preset standard image, if so, directly returning an unknown type of the sheet media; and otherwise, calculating the geometric center of the acquired image information; and intercepting the image information with the same dimension as that of the preset standard image by taking the geometric center as the center to form the image information to be identified. By the method, the discrimination degree of the feature vectors in different classifications can be improved effectively and the generalization capability of the classifier can be enhanced.

The invention discloses a social-network-oriented extraction method of specific events; on one hand, representation learning of text is achieved via a neural network, and characteristics of great generalization ability suitable for mass data are extracted; on the other hand, a supervised learning method used herein is more specific than a non-supervised learning method, event representations of good semantic clarity can be provided, mass event-irrelevant tweets are filtered using a keyword-based primary filter algorithm, and algorithm speed is increased; event sorting for the tweets is achieved using the neural network, artificial design characteristics are not required, and the algorithm has better generalization ability; hierarchical sorting from event type to event instance is constructed to form a systematic event system; an event instance cluster is subjected to information extraction, namely, event information is comprehensively extracted from different tweets in the same event,and the information is more comprehensive.

The invention relates to a fiber-optic gyroscope temperature drift error compensation method and belongs to the technical field of precision instruments. The effect of temperature, temperature variation rate and temperature field gradient on fiber-optic gyroscope temperature drift is fully considered, temperature piecewise linear fitting is performed on temperature coefficient, temperature variation rate coefficient and temperature gradient coefficient, and precision of a model is remarkably improved on the premise of not adding the number of orders of the system model. The number of temperature piecewise points is optimally selected by adopting a genetic algorithm, the piecewise points are selected as few as possible on the basis of giving consideration to the precision of the model, andgeneralization of the model can be improved; the number of the orders is smaller than or equal to 2-order, simpleness in parameter calculation procedure is achieved, training time like neural networksis not needed, order determination of the polynomial model is not needed, and difficulty in modeling is remarkably reduced; meanwhile, the calculated amount in real-time compensation and storage costare greatly reduced.

The invention discloses a dense stacking target detection method based on automatic labeling and transfer learning. The dense stacking target detection method comprises the following steps: establishing a labeled training image set by high-resolution image segmentation; inputting the labeled training image set into a pre-trained target detection model YOLOv3, optimizing a priori frame size and a loss function of the YOLOv3 model, and performing fine adjustment on the model by using the training image set; and finally, inputting a to-be-detected image into the fine-tuned YOLOv3 model, outputting the classification of the target sub-regions and the positions of the sub-regions, splicing output result images to recover the output result images into an original image, and counting a total counting result. The dense stacking target detection method provided by the invention has strong anti-interference performance and robustness, and has low requirements on an image photographer and a photographing illumination condition; through an unsupervised learning method, quasi-automatic annotation of the image is realized, and the workload of manual annotation is greatly reduced, and the model training efficiency is improved; and the dense stacking target detection method can be used for image recognition of a large number of densely stacked targets which are mutually shielded, and is suitable for automatic counting scenes of various densely stacked targets.

The invention relates to a video target tracking method based on twin network fusion multi-template features, and provides a semi-supervised template online updating strategy, when a to-be-tracked target in a video sequence has complex conditions such as occlusion, deformation and illumination change, the target change and the occluded condition are evaluated by calculating an APCE value and template similarity, when the appearance of the target is greatly changed, the features extracted from the previous frame of picture are fused with the original template features to obtain a new template with higher expression capability, so that the method is favorable for adapting to various complex conditions; in order to improve the generalization ability of the model and adapt to multiple types oftargets, a regularization technology is adopted in the training process to prevent model overfitting; in order to further improve the speed of the algorithm, only an original template is adopted fortracking in a non-complex situation, so that the calculated amount is greatly reduced, and the method provided by the invention achieves a higher running speed than other methods under the condition of obtaining better tracking performance.

A method and an apparatus for general image classification base on a semi-supervised generative adversarial network are disclosed, relating to the image classification technology. The method includes:1, obtaining a deep convolutional generative adversarial network DCGAN through training, wherein that DCGAN comprises a generation network and a discrimination network, and the discrimination networkcomprises a convolutional neural network and a Softmax multi-classifier connected in turn; 2, inputting the image to be classified into the convolution neural network to obtain image features; 3, inputting the obtained image features into the Softmax multi-classifier to obtain a classification result. As the discriminant network is train in the DCGAN, the generalization performance and the classification accuracy rate of the discriminant network are improved. After image features are obtained through the discriminant network and Softmax multi-classifier is combined, the accuracy rate of imageclassification can be greatly improved.

The invention discloses an artificial intelligence-based contract review method, device and equipment and a storage medium, and the method comprises the steps: obtaining a contract review request which comprises a to-be-reviewed contract and a contract type; If the file format of the to-be-examined contract is a picture format, performing fuzzy detection on the to-be-examined contract, if the fuzzy detection is passed, identifying the to-be-examined contract by adopting an OCR identification technology, and obtaining a to-be-examined text corresponding to the to-be-examined contract; If the file format of the to-be-examined contract is a text format, taking the to-be-examined contract as a to-be-examined text; And carrying out intelligent review on the to-be-reviewed text by adopting a target contract review model corresponding to the contract type to obtain a contract review result. The contract review method can realize intelligent review, does not need manual intervention, and improves the review efficiency.

The invention discloses an automatic adjustment method and system for a cavity filter, relates to the technical field of industrial robots, and aims at solving the technical problems that existing automatic adjustment method and system for the cavity filter cannot achieve fine waveform adjustment, and are poor in extensiveness and stability. According to the technical scheme, the automatic adjustment method comprises the following steps: (1) building a corresponding target vector field from a current waveform to a standard waveform; (2) respectively pre-adjusting each screw with the fine adjustment amount with fixed size and direction to obtain a corresponding vector field; (3) fitting a coefficient of the vector field corresponding to the change of each screw according to the vector field generated by pre-adjusting and the target vector field; and (4) obtaining the to-be-adjusted amount of each screw through the coefficient of each screw vector field and making the overall action of screw adjustment close to the direction of the standard waveform all the time.

A convolutional neural network-based single word pattern character recognition method disclosed by the present invention comprises the following steps of determining a target image on which the character recognition is going to be carried out; positioning the character information in the target image, and extracting the characters; according to the extracted characters, obtaining a plurality of single word patterns of the target image; separately inputting each single word pattern in a pre-trained target convolutional neural network, propagating forwards, and obtaining a character recognition result of each single word pattern. By applying the technical scheme provided by the embodiment of the present invention, the target convolutional neural network is obtained by training a lot of single word pattern samples, and the character recognition rate of the single word patterns and the generalizable ability are improved. The present invention also discloses a convolutional neural network-based single word pattern character recognition device which has the corresponding technical effect.

The invention relates to the technical field of big data application, discloses a method for obtaining a fault solution based on big data, and solves the problem that a user cannot obtain a recommended solution when equipment has a fault. The method comprises the steps of collecting a fault model of the equipment in history and a corresponding fault solution, and loading collected data as trainingdata to a big data platform; for the current equipment, measuring the efficiency of a decision tree algorithm and the efficiency of a random forest algorithm, and performing adjustment and optimization on the algorithms and parameters of a decision tree and a random forest; performing multi-time sampling with replacement for the training data to generate multiple required samples, and performingtraining on the samples to generate the corresponding decision tree; and performing voting on the current fault solution of the equipment by utilizing a decision forest to obtain a voting result, anduploading the current fault model and the corresponding fault solution to the big data platform. The method is suitable for fault solving of household appliances.

The invention discloses a Wishart and SVM (support vector machine)-based polarimetric SAR (synthetic aperture radar) image classification method. The technical problems of low classification accuracy and low classification efficiency of a conventional polarimetric SAR classification method in case of fewer training samples are mainly solved. The method is implemented by the following steps: inputting an image; performing filtration; performing Cloude decomposition; calculating a similar matrix of a characteristic set F; calculating a similar matrix of a coherence matrix characteristic set T; calculating a final similar matrix; performing classification by virtue of an SVM; calculating accuracy. When the method is used for classification, the problem of more misclassified points caused by noise in the prior art is solved, crosstalk between polarization channels can be avoided, polarimetric information and counting correlation can be maintained, the contour and edge of a polarimetric SAR image are clearer, the quality of the classified image is improved, higher polarimetric SAR image classification accuracy and higher noise adaptability are achieved, and the method can be used for the target identification and tracking of the polarimetric SAR image.

The invention relates to voice semantics, and provides a voice endpoint detection method for emotion recognition. The method comprises: acquiring an audio signal; performing processing operation on the audio signal, comprising adding pure noise sections and human voice noise sections in various scenes to the audio signal and randomly setting a signal-to-noise ratio; extracting an MFCC feature anda second-order difference feature of the processed audio signal; inputting the features into a neural network model, and extracting high-dimensional information of the audio signal and front and backassociated features of the audio signal; inputting the extracted high-dimensional information and the associated features of the audio signal into a full-connection network model to obtain a detectionresult of each frame of the audio signal, the detection result comprising human voice and non-human voice; and segmenting the audio signal into a human voice part and a non-human voice part accordingto the detection result of the audio signal. The invention further provides electronic equipment and a storage medium. Accurate voice endpoint detection can be realized in a low signal-to-noise ratioenvironment and a non-stationary environment.

The invention relates to the technical field of network information, in particular to a private encryption protocol message classification method based on sparse representation and a convolutional neural network, which comprises the following steps: obtaining and preprocessing network traffic data to obtain a data set file and a label file; importing the data set file into a sparse auto-encoder for unsupervised feature learning to obtain input data with smaller dimension; and training the two-dimensional convolutional neural network by using the training set after sparse representation and thetraining set label, performing convolution and pooling, and minimizing errors to obtain a classifier. According to the classification method disclosed by the invention, the classification characteristics of the private encryption protocol message are automatically learned from the original network flow, and classification identification is realized; the method does not depend on the IP address and port number information of the header of the network traffic data packet, and the generalization ability of the classification model is high; sparse representation is used for learning local features of private encryption protocol messages, a two-dimensional convolutional neural network is used for learning global features of the private encryption protocol messages, and the recognition precision of the classifier is improved.

The invention discloses a method and system for identifying a falling fault of a rail wagon hook lifting rod, and belongs to the technical field of freight train detection. The objective of the invention is to solve the problem of low accuracy of truck hook lifting rod falling fault detection by use of an existing method. The specific implementation process of the method comprises the following steps: step 1, collecting a truck image, determining a hook lifting rod region from the collected truck image, extracting a hook lifting rod region image, and constructing a training set by utilizing the extracted hook lifting rod region image; 2, constructing a deep learning model, and training the constructed deep learning model by using the training set to obtain a trained deep learning model; and step 3, inputting a to-be-detected image into the trained deep learning model to obtain an output result of the deep learning model, and processing the output result by using an image processing method to obtain a judgment result of the falling fault of the hook lifting rod. The method and system are mainly used for detecting falling of the hook lifting rod.

The invention discloses an aviation three-level AC power generator rotary rectifier online fault diagnosis method based on an optimization extreme learning machine, and belongs to the technical field of power generator state monitoring and fault diagnosis. The method comprises the following steps that 1) an overall simulation model of an aviation three-level AC power generator is established and the fault mode and fault testable points of a rotary rectifier are determined; 2) corresponding voltage or current information under various fault modes is acquired by utilizing the testable points, feature extraction is performed on fault information and normalization processing is performed, and a training sample set and a test sample set are determined; 3) the parameters of a mind evolutionary algorithm, the number of nodes in each layer of the extreme learning machine and an excitation function are set, the extreme learning machine is optimized, the optimal input weight and threshold are outputted, and an extreme learning machine model is established and the model is verified by using the test sample set; and 4) the model passing verification in the step 3) is used for online intelligent fault diagnosis, and a fault signal and feature extraction method is maintained to be consistent with that of step 2).

The invention provides an unmanned aerial vehicle aeromagnetic holoaxial gradient magnetic disturbance compensation method based on a feedforward network. The method includes the steps that calibrated flight data is subjected to preprocessing operation of wavelet transform and normalization, normalized feedforward network input vectors are transmitted into the established feedforward network to be trained, a regularized cost function is adopted, stable feedforward network parameters are iteratively obtained through an error back-propagating algorithm, and a finally convergent network serves as a compensation network; exploration flight data is subjected to identical preprocessing operation and transmitted to the obtained compensation network to be calculated, estimation of an airplane disturbing magnetic field is obtained, and magnetic disturbance compensation is obtained. The stable network parameters are obtained by training the feedforward network, the network serves as the compensation network for compensating for the exploration flight data, the problems of inverse matrix instability and over-fitting in a least squares algorithm are effectively avoided, the generalization performance of the feedforward network is expanded, and unmanned aerial vehicle aeromagnetic holoaxial gradient magnetic disturbance compensation is achieved.

The invention discloses a medical image pulmonary nodule detection method based on a cyclic feature pyramid, and mainly solves the problem of low detection precision caused by high extraction difficulty of pulmonary nodules and ground glass nodules in the prior art. The method comprises the following implementation steps: 1) acquiring a medical image; 2) preprocessing the medical image, and expanding a sample data set; 3) constructing a circular feature pyramid detection model by combining the feature map; 4) training the detection model by using the extended data set sample to obtain a targetdetection model; and 5) inputting the test set in the data set into the trained detection model to carry out pulmonary nodule detection. The method constructs a new feature extraction network, accelerates the network training speed, enhances the sensitivity to nodules of different sizes, improves the detection precision of pulmonary nodules of medical images, and can be used for a computer-aidedmedical diagnosis system.

The invention discloses an electricity load type identification method, which is realized through an electricity load identification device consisting of an information collection module, an information processing module and a communication module. The electricity load type identification method simultaneously adopts electricity load starting current characteristics including starting process time, a starting current maximum value, and starting current maximum value time and a load current frequency spectrum characteristic of the electricity load as identification characteristics for the electricity load, and the characteristic information is rich. The electricity load type identification method adopts a combination classifier comprising a support vector machine classifier and a Bayes classifier to perform identification classification, performs comprehensive identification in consideration of characteristics of two classifiers, and thus has high identification accuracy. The provided methods for obtaining starting current characteristics and load current frequency spectrum characteristics are simple and reliable. The electricity load identification device can be used in some collective public places like a students 'dormitory, a large-scale pedlars' market, etc, where the electricity load management is needed, and can also be used in other places where need to perform electricity load type statistics and electricity appliance management.