139results about How to "Guaranteed generalization ability" patented technology

The invention provides a wind-field power combination predicting method based on a fuzzy neural network and a support vector machine, which comprises the following steps of: acquiring and pre-processing data; setting up a fuzzy neural network model by using a normalized training sample set and a prediction sample set and predicting; setting up a support vector machine model and predicting; linearly combining the prediction results of the two algorithms to obtain a wind speed prediction value; and setting up a wind speed power expert table via historical data, and inquiring the expert table according to the predicted wind speed value so as to obtain a power prediction result. By the method provided by the invention, the short-term prediction of a wind speed sequence can be effectively realized, the power prediction precision is improved, and fewer computing resources are consumed.

The invention discloses a convolutional neural network-based photovoltaic glass defect classification method and a convolutional neural network-based photovoltaic glass defect classification device. The method comprises the following steps of carrying out multi-angle and variable-illumination image acquisition on a plurality of defect samples to obtain a plurality of images; preprocessing the images to fuse the multi-channel information and generate a multi-channel-fused defect sample image; adopting a convolution neural network model which meets a preset condition, extracting a network according to defect category design features and constructing a feature extraction convolutional neural network; obtaining the number of layers of all-connected neural networks and the number of neurons ofeach layer, and constructing a feature classification network; under the condition that the cross entropy loss function is minimized, completing the training of the convolutional neural network; according to an input sample image, outputting a prediction result of a defect category through the trained convolutional neural network. Based on the method, the situation that training sets are insufficient can be effectively solved while the generalization ability and the prediction precision of the model are guaranteed. Meanwhile, the high classification precision can be achieved for a small amountof glass defect samples.

The invention provides a gas concentration real-time prediction method based on a dynamic neural network. Firstly, the neural network is trained by means of data in a mine gas concentration historical database, activeness of hidden nodes of the network and learning ability of each hidden node are dynamically judged in the network training process, splitting and deletion of the hidden nodes of the network are achieved, and a network preliminary prediction model is built; secondly, mine gas concentration information is continuously collected in real time and input into the prediction model of the neutral network to predict the change tendency of gas concentration in the future, and the network is trained timely through predicted real-time data according to the first-in first-out queue sequence to update a neutral network structure in real time, so that the neutral network structure can be adjusted according to real-time work conditions to improve gas concentration real-time prediction precision. According to the method, the neural network structure can be adjusted timely on line according to the real-time gas concentration data, so that gas concentration prediction precision is improved, and the technical requirements of a mine gas concentration information management system are met.

The invention provides a pipeline defect magnetic flux leakage inversion method based on Adaboost-RBF synergy, relating to the technical field of magnetic flux leakage detection of pipelines. The method comprises the following steps: carrying out magnetic flux leakage detection on standard defects, and carrying out feature extraction; measuring defect shape parameters of front several meters of a pipeline on which to-be-tested defects are located; carrying out the magnetic flux leakage detection on the pipeline on which to-be-tested defects are located, and carrying out feature extraction; determining sample data and to-be-tested data; establishing an Adaboost-RBF neural network initial model; correcting the Adaboost-RBF neural network initial model; and inputting the to-be-tested data into the final model, so as to obtain the shape parameters of the to-be-tested defects, thereby finishing the inversion. By inverting the pipeline defects by virtue of an Adaboost-RBF neural network model, the rapid defect shape reconstitution can be realized, the learning speed is high, the precision is high, the generalization performance is good, and the severity of the defects can be judged, so that the pipeline leakage is prevented, and the loss is avoided.

The invention discloses a distributed optical fiber sensing signal mode identifying method and system. The distributed optical fiber sensing signal mode identifying method includes steps of acquiring samples, and acquiring several groups of samples for excitation signals of each behavior, so as to set up a sample library; roughly screening samples, screening the acquired sample library to remove samples high in discreteness; finely screening the samples, training to generate a classifier used for signal mode identification by means of the screened sample library and storing; identifying signal modes, calculating characteristic vectors for a newly triggered and read section of signal data and subjecting the characteristic vectors to calculation in the stored classifier to obtain mode identification results. The distributed optical fiber sensing signal mode identifying method and system can effectively identify invasion or damage signals in case that signals are carried with noise or distortion, and in the meantime, mistaken reports are reduced.

The present invention relates to a Web user access path prediction method based on a recurrent neural network. In the method provided by the present invention, a user access path is taken as a research target, a recurrent neural network is introduced into a path prediction problem, and a network model for path prediction is studied and designed; based on the simple recurrent neural network, a feature layer is added, and a Long-Short Term Memory (LSTM) unit is used in a hidden layer; and the method can effectively utilize the context information of the user session sequence, learn and memorizethe access rules of the user, obtain good model parameters through training data learning, and then predict the next access path of the user. The theoretical analysis and experimental results show that the method disclosed by the present invention is relatively high in path prediction efficiency, relatively accurate in prediction result, and applicable to solving the problem of Web user access path prediction.

The invention belongs to the field of computer aided diagnosis and medical image processing, and relates to a skin disease image focus segmentation method based on a deep convolutional neural network,which is used for improving the quality of a skin disease image and further improving the focus segmentation accuracy so as to obtain more accurate focus information. The method comprises the specific steps that data preprocessing is responsible for carrying out noise reduction processing on a skin disease image, and removing artificial and natural noise which hinders focus position determinationin the image; the data expansion is responsible for expanding a data set by deforming and rotating the image subjected to the noise reduction processing; a segmentation model is constructed to perform first feature extraction on the image, encoding is performed to obtain more detail features, and the features obtained at the first time are fused to obtain a prediction graph;.

The invention discloses a WT-KPCA-SVR coupling model based gas emission quantity prediction method. The method comprises the following steps: firstly, performing data preparation, namely collecting gas emission quantity monitoring data and corresponding factors, extracting gas emission quantity subsequences according to wavelet transform, and separating out a trend item subsequence and a fluctuation subsequence; determining the influence factors of each subsequence according to a gray relative analysis method, performing kernel principal component dimensionality reduction on the influence factors of each subsequence to reconstitute the principal component of each subsequence; combining the reconstituted principal components of all the subsequences and values of all gas emission quantity subsequences into a sample set; respectively establishing support vector machine regression models of the trend item subsequence and the fluctuation subsequence according to a training sample; synthesizing the two models to obtain a final gas emission quantity prediction model; performing model precision detection by using a detection sample, wherein the model can be used if passing the detection. The model is reliable in design principle, simple in prediction method, high in prediction accuracy and friendly in prediction environment.

The embodiment of the invention provides a fine-grained classification method and device for a target object and electronic equipment, and the method comprises the steps: extracting a feature vector which represents the feature of the target object by utilizing a convolutional neural network model based on an image of the target object; On the basis of the feature vectors, retrieving a standard feature vector set corresponding to a standard image library, and obtaining a fine-grained classification result of the target object, Wherein the convolutional neural network model is obtained by training based on a cross entropy loss function and a triple loss function in advance. According to the embodiment of the invention, the convolutional neural network is trained based on the cross entropy loss function and the triple loss function, and the trained convolutional neural network is adopted to realize the extraction process of the image features, so that the generalization ability of the classification algorithm can still be ensured under the condition of small data volume, and the classification accuracy is improved.

The invention discloses a face recognition neural network training method, system and device and a storage medium, and the method comprises the following steps: obtaining a face image as a training set and a test set, and combining a loss function of a face recognition neural network with an adaptive additional loss function; inputting the preprocessed training set into a face recognition neural network for training; inputting the test set into the trained face recognition neural network, and verifying the recognition accuracy of the trained face recognition neural network. According to the invention, when the face recognition neural network is trained, the loss function is combined with an adaptive additional loss function to obtain a final loss function; the intra-class distance when theface images are classified is shortened through the final loss function, the inter-class distance when the face images are classified is increased, meanwhile, balance of multi-sample classes and few-sample classes is considered, when sample distribution is unbalanced, the generalization performance of the face recognition neural network can be guaranteed, and the accuracy and reliability degree of face recognition are improved.

The invention relates to a construction method and application of a lightweight gesture detection convolutional neural network model, and the method comprises the steps: constructing a lightweight gesture detection convolutional neural network framework based on a SquezeNet convolutional neural network framework and an SSD multi-target detection convolutional neural network framework; Acquiring agesture picture and a background picture, and performing image data enhancement and picture synthesis processing on the gesture picture based on the background picture to obtain a gesture data set; And based on the public data set and the gesture data set, training a lightweight gesture detection convolutional neural network framework to obtain a lightweight gesture detection convolutional neuralnetwork model. According to the invention, a small amount of gesture data is expanded into the gesture data set containing a large amount of picture data at a high speed; The technical problem that alarge amount of high-quality gesture picture data is difficult to obtain is solved, in addition, by combining the SquezeNet convolutional neural network architecture and the SSD multi-target detectionconvolutional neural network architecture, the constructed lightweight gesture detection convolutional neural network model occupies few computing resources, and can be applied to various detection platforms.

The invention belongs to the technical field of machine vision detection, and particularly relates to a defect classification method based on an improved particle swarm wavelet neural network. The problems that a traditional BP neural network algorithm is prone to convergence and prematurity, and cause a local minimum value and the like are solved. The method comprises the following steps: loadingan original image, carrying out graying and median filtering processing, segmenting the image, calculating a defect feature vector, initializing a particle swarm, calculating a target fitness value,evaluating each particle, updating the position and speed of each particle, checking whether the requirement is met, outputting an optimal solution, and finally carrying out defect classification on the image. According to the method, a variation factor is added, so that the generalization capability of the algorithm is ensured. A nonlinear weight factor is set, and a target of flexible adjustmentof global search and local search is realized. A global extreme value of Gaussian weighting is introduced, convergence of the global extreme value to the optimal solution direction is facilitated, defects can be classified quickly and accurately, the classification result is more accurate, and the efficiency is higher.

The invention discloses an age estimation method based on deep learning. The age estimation method based on deep learning comprises steps of (1) constructing an age database, (2) performing pre-processing on images of the constructed age database, (3) performing unification and normalization on sizes of aligned images to obtain images with a size being 64*64, (4) taking the obtained images and corresponding labels as inputs of a deep model, using a CNN convolution deep network to train an age estimation model, (5) inputting an age estimation model into a tested image to obtain similarity values of tested images on various kinds of labels, (6) multiplying the obtained corresponding label with the obtained similarity value and then adding obtained corresponding label and the obtained similarity value to obtain a final age estimation result. The age estimation method based on deep learning can obtain a smaller deep model and is fast in operation time and high in an age estimation identification rate. The database comprises massive samples of children and senior people and can effectively identify age of a special group.

The invention provides a control method and device of a refrigerating unit, electronic equipment and a computer readable storage medium. The method comprises the following steps of carrying out training of a linear regression model based on historical operation parameters and historical power of each piece of refrigerating equipment in a refrigerating unit, and obtaining a trained linear regression model; according to the historical operation parameters and historical power of the refrigeration equipment, training a hybrid model fusing the linear regression model and a neural network model ofthe refrigeration equipment to obtain a power prediction hybrid model; constructing a target function according to the power prediction hybrid model; and controlling the corresponding refrigeration equipment according to the refrigeration control parameter in the operation parameters corresponding to each refrigeration equipment when the function value is minimum. According to the technical scheme, the objective function is constructed based on the power prediction hybrid model corresponding to each refrigeration equipment, and the refrigeration control parameters in the operation parameters are used for controlling the refrigeration equipment when the function value of the objective function is minimum, so that the energy consumption of the refrigeration unit is accurately optimized.

The invention provides a method for soft measurement of the effluent total phosphorus (TP) in the sewage disposal process based on the neural network, and belongs to the field of sewage disposal field. The mechanism is complex in the sewage disposal process, and to enable a sewage disposal system to be in a good running working condition and to obtain the higher effluent quality, the procedure parameters and the water quality parameters in the sewage disposal system need to be detected. The invention provides a soft measurement model established based on the self-organization radial-based neural network to solve the problem that the effluent total phosphorus of a current sewage disposal plant cannot be obtained in real time. The initial structure and the initial parameters of the neural network are determined according to the self-organization method, the structure of the neural network is simplified, and real-time soft measurement is carried out on the effluent TP. According to the soft measurement result, the related control link in the sewage disposal process and materials in the biochemical reaction are adjusted, the quality of the effluent obtained after sewage disposal is improved, and a theoretical support and a technological guarantee are provided for safe and stable running in the sewage disposal process.

The invention discloses a hyper-spectral remote sensing image classification method based on self-adaptive hierarchical multi-scale. The method comprises the following steps that step 1, the irregular neighborhood structure of pixels is calculated according to the spectral angle; step 2, the scale parameters of each layer are determined according to the Ka measure hierarchy in the irregular neighborhood structure, the corresponding kernel matrix of each layer is calculated layer by layer and the weight of the kernel function of each layer is obtained by using the maximum projection variance so that a self-adaptive hierarchical multi-scale kernel function is obtained; and step 3, a hyper-spectral image is mapped to the kernel space of the self-adaptive hierarchical multi-scale kernel function obtained in the step 2, and the pixels to be measured are linearly represented by a dictionary formed on the basis of the known training sample pixels so that a reconstruction sparse matrix is obtained and the pixels to be measured are allocated to the optimal reconstruction category. The hyper-spectral remote sensing data can be rapidly and accurately classified.

The invention belongs to the technical field of irrigation forecasting. The invention discloses a real-time irrigation forecasting system based on regional soil moisture monitoring and remote sensingdata. The real-time irrigation forecasting system comprises an irrigation area monitoring module, a crop image acquisition module, an environment data acquisition module, a central control module, a soil moisture prediction module, an irrigation area identification module, an irrigation prediction module, an actual water consumption module, a watering module, a water level depth measurement module, an alarm module, a wireless communication module, a data storage module, a data management module, a terminal module, a power supply module and a display module. According to the real-time irrigation forecasting system, the prediction precision can be improved and the generalization ability can be ensured through the soil moisture prediction module, so that large-scale deployment and expansion are facilitated, and better expandability and transportability are achieved; the irrigation area recognition module meets the requirement for extracting irrigation information of small plots, so that the precision of an irrigation area monitoring result is improved; and the real-time irrigation forecasting system can be widely applied to remote sensing identification and extraction of irrigation areas with high spatial resolution.

The invention provides an intelligent detection and yield optimization method for an HDPE (high density polyethylene) cascade polymerization reaction course, which solves the problems that the HDPE cascade polymerization reaction course is complicated in technology, on-line measurement of a key quality variable is difficult, and the operating cost of a production course is high. The method comprises the steps of preprocessing by adopting data correction and data excavation techniques, searching a rule from production and analysis data, adopting an artificial neural network technique to establish an intelligent soft measuring instrument and a polyethylene product unit consumption model, and adopting an extension engineering technique to optimize an artificial neural network structure and improve the neural network modeling accuracy. The method has the characteristics of high response time, high modeling accuracy, high inferential capability and convenience in management, and assists in ensuring safe production of HDPE, improving the quality of a polymer product, and saving the production cost.

The invention discloses a construction machinery hidden danger detection method for a power transmission line. The construction machinery hidden danger detection method comprises the following steps:acquiring images in a power transmission line channel and around the power transmission line channel in real time; making a data set of large construction machinery existing in a power transmission line channel, and randomly distributing according to a ratio of a training set to a test set of 4: 1; performing image preprocessing on the training set; processing the training set by adopting a multi-sample image synthesis method to obtain new training set sample data; training the new training set sample data by using a Faster R-CNN + FPN model to obtain a detection model of the hidden danger ofthe power transmission line channel; carrying out image target detection on the test set, updating detection model parameters, and carrying out secondary verification on the test set; and detecting the image acquired in real time by using the updated detection model, and detecting whether a large construction machinery hidden danger exists in the power transmission line channel or not. According to the construction machinery hidden danger detection method, the generalization problem of the detection model is solved, and the false alarm rate and the missing report rate are reduced.

The invention discloses a traffic intersection congestion prediction method based on machine learning, and belongs to the technical field of machine learning. According to the method, the congestion degree of a corresponding traffic intersection is predicted through multiple structural features, the requirement for equipment is low, and the speed is high; meanwhile, the degree of intersection congestion within a long time can be predicted, and an unstable model is adopted as a base learner for training; and finally, the deviation and variance of the model are reduced through integrated learning and model fusion, so that the generalization ability of a prediction result of the model in an actual scene is ensured.

The invention discloses an abnormal power consumption detection method and system and a storage medium, and belongs to the technical field of intelligent power grid abnormal power consumption behaviorrecognition. The abnormal power consumption detection method comprises the following steps: collecting data of a to-be-detected intelligent electric meter; converting the intelligent electric meter data into multi-time sequence data; inputting the multi-time sequence data into a trained abnormal power consumption detection model to obtain a predicted power consumption type label; and taking the power consumption type with the maximum probability in the predicted power consumption type label as an abnormal power consumption detection result of the to-be-detected intelligent electric meter data. According to the invention, big power utilization data can be fully and comprehensively utilized, advanced features of abnormal power utilization behaviors are deeply mined, the generalization performance of abnormal power utilization detection is improved, the misjudgment ratio is reduced, and abnormal power utilization is efficiently and accurately detected.

The invention discloses a SVM (Support Vector Machine) solar wing unfolding reliability evaluation method based on kernel optimization. The method comprises the following steps of: establishing a solar wing unfolding reliability comprehensive evaluation index system according to expert knowledge; obtaining a weight vector of the evaluation index system by a matter element method and an analytic hierarchy process; grading measured values of each factor influencing the unfolding of the solar wing by experts, and taking the grading result as sample data; automatically selecting the SVM kernel and parameter values thereof by a program so as to form a training model; performing cross validation to check whether the kernel and the parameters thereof need to be regulated finely; and validating the formed model by a test sample, and evaluating the reliability of unfolding of the solar wing. The method has the following advantages that: the evaluation result is objective and credible under circumstances of zero failure, small sample, nonlinearity, high dimensionality and the like.

Because the generalization of a flame detection model based on image features is not strong, and the requirement of a deep neural network model for the number of training samples is high, the invention provides a flame target detection method based on digital images and convolution features, and the method comprises the steps: firstly making a data set comprising video dynamic features; replacingthe standard convolution of the VGG16 in the classic Faster R-CNN with the depth separable convolution, and reducing the number of convolution layers; cutting 256 image blocks from the original imageaccording to a candidate box generated by the RPN, and extracting LBP features of each image block; reducing the size of an output feature map of the ROI pooling layer and the number of neurons of a full connection layer through convolution, and further reducing network parameters; and finally, combining the extracted LBP features, the dynamic features in the data set and the pooled tiled featurevectors, and sending the combined feature vectors to a full connection layer for classification and regression. The flame target detection model constructed by the patent has relatively high detectionprecision, is convenient to improve for overcoming the defects of a test result, and is high in flexibility.

The invention discloses an online biochemical oxygen demand (BOD) soft measurement method based on a dynamic feedforward neural network. The method comprises the following steps of: designing a dynamic feedforward neural network topological structure for BOD soft measurement of a sewage aeration tank, determining an input sample of the dynamic feedforward neural network, and performing online normalization processing on the input sample; calculating the variation condition of an ownership connection value connected with a hidden node in the neural network in each training process by employing a standardized data training neural network, judging the activeness of the hidden node, and splitting the hidden node with high activeness; judging the capacity of learning information of the hidden node by calculating the absolutely output variation condition of the hidden node in the training process, and deleting a hidden node without the learning capacity; adjusting parameters of the neural network; and determining the BOD of effluent of the aeration tank after the training process of the dynamic feedforward neural network is ended. The method has the advantages of high real-time property, high stability, high precision and high neural network generalization ability.

The invention discloses a workpiece recognition device and method based on an image recognition-SVM learning model. The device comprise an image collection unit, a recognition unit and a robot, and the image collection unit is used for obtaining a to-be-detected workpiece image and is in data connection with the recognition unit; The recognition unit is used for classifying the workpieces to be detected by adopting an SVM learning model classifier after extracting the feature vectors of the workpieces in the images, outputting a classification result and being in control connection with the robot; And the robot is used for classifying the workpieces to be detected according to the classification result. The device can identify and grab a predetermined target in a relatively complex environment, is not influenced by translation, scale and rotation geometric changes, and has relatively high stability and real-time performance. Another aspect of the application also provides a method of the device.

The invention discloses a multi-round dialogue intention identification method, which comprises the following steps of: 1) determining a dialogue intention to be identified according to an applicationscene, obtaining a large amount of dialogue data, manually finding out dialogue blocks in the dialogue data and intentions corresponding to the dialogue blocks, and carrying out corpus annotation; 2)removing stop words in the dialogue data according to a manually arranged stop word table; 3) constructing a sentence vector model for intention recognition, namely BiLSTM- A CRF model and a StarSpace model are established; and 4) obtaining dialogue data in real time through the sentence vector model, the BiLSTM-CRF model and the StarSpace. The method has the beneficial effects that intention recognition is efficient, the accuracy is high, the generalization ability is high, and the manual corpus annotation cost is low.

The invention provides a pancreatic cancer pathological image classification method based on self-attention feature fusion, and the method comprises the steps: firstly, employing a convolutional neural network model to extract the features of an input image, and carrying out the feature embedding of a feature map outputted by the final stage of the convolutional neural network model; secondly, feature maps output by the convolutional neural network model at different stages are subjected to attention analysis to obtain attention guidance information; then, a Transform model based on self-attention feature modeling and a self-attention feature fusion network model are constructed; and finally, training the self-attention feature fusion network model for multiple rounds, measuring and determining the model corresponding to the round with the optimal result by using the pathological image of the verification set, thereby constructing a pancreatic cell cancerization classification diagnosis system, and judging whether the pancreatic cell pathological image is a pancreatic cancer cell image or a normal cell image through the system. According to the invention, global modeling is carried out on the convolutional neural network features by applying a self-attention technology and an attention analysis mechanism so as to realize high-precision rapid on-site evaluation of pancreatic cancer.

The invention discloses a crowd density map estimation method and system, and the method comprises the following steps: obtaining a scene image, carrying out the preprocessing of the scene image, andgenerating a crowd density label map; performing data augmentation on the scene images and the crowd density label graphs to obtain a plurality of scene images and corresponding crowd density label graphs; training a crowd density map estimation model according to the plurality of scene images and the corresponding crowd density label map; and receiving a scene image, and performing crowd densityestimation based on the trained crowd density map estimation model. Aiming at the problem of head size difference under a complex background, effective features are extracted by using a multi-scale module and features of a feature enhancement unit, and crowd density map estimation is carried out by using a coarse-to-fine strategy.

The invention discloses a coronary artery region segmentation method and system based on improved 3D Unet, and the method comprises the steps: inputting a to-be-segmented DICOM image, and carrying outthe preprocessing of the image; inputting the preprocessed image into a pre-trained coronary artery segmentation model, and obtaining and outputting an image in which a coronary artery region is segmented, wherein the coronary artery segmentation model is obtained by training a historical DICOM image, the coronary artery segmentation model is an improved 3D Unet model, each layer of down-samplingof the improved 3D Unet model is finally added into an incomplete module, and the number of residual blocks is increased by one layer as the network depth is increased by one layer; and the up-sampling process adopts a deconvolution process to amplify the picture. Compared with a traditional method in which a deconvolution process relates to a weight updating learning process, the method of the invention has higher adaptability to different tasks.