40 results about "Incremental learning algorithm" patented technology

The invention discloses a face recognition method based on improved incremental non-negative matrix factorization. An incremental non-negative matrix solution algorithm based on Fisher discriminant analysis is adopted; prior information trained by an initial sample is utilized in the algorithm; the column vector of a newly increased coefficient is assigned to a corresponding means vector through an index matrix; therefore, the convergence speed of the algorithm is increased; in addition, an INMF algorithm is improved into a batch incremental learning algorithm; initialized assignment is carried out by utilization of the index matrix; furthermore, restriction having the maximum interclass divergence and the minimum in-degree is imposed onto a target function; and thus, the best sub-space projection is obtained. The scheme has relatively high recognition rate and relatively rapid recognition speed, and is applied to occasions, such as a face recognition scene.

The invention discloses a rolling bearing fault detection method based on an actual measurement signal. The rolling bearing fault detection method comprises the steps of: firstly, converting a rollingbearing fault time domain vibration signal to an angle domain through employing an order tracking technology; secondly, carrying out parameter optimization on variational mode decomposition through adpopting a longicorn beard search algorithm, and decomposing all state vibration signals of a rolling bearing to obtain a series of intrinsic mode functions, wherein frequency band energy in differentintrinsic mode functions can change when different faults happen to the bearing; thirdly, extracting Renyi entropy features from modal components containing main fault information, and constructing afeature subset; and finally, using normal state vibration signals easy to obtain for training, extracting fault characteristic quantities, establishing fault data samples and incremental learning data samples, acquiring a fault recognition model through training by adopting a single-class support vector machine incremental learning algorithm, judging whether the rolling bearing breaks down or notaccurately, and achieving fault early warning.

The invention relates to an sEMG self-adaptive mode recognition method based on an on-line SVM and application of the method on an intelligent wheelchair and belongs to the technical field of surface electromyogram signal recognition and control. According to the method, the incremental learning algorithm is adopted to perform on-line training on an SVM model, and meanwhile head information is introduced in the on-line learning process and serves as correction information to jointly form a classifier of an sEMG; effective recognition is achieved on the sEMG at different muscle states, influences on the stability of a man-machine interactive system by muscle fatigue are reduced, and finally the method is successfully applied to the intelligent wheelchair. The method effectively improves the self-adaptive capacity of the system in the long-term man-machine interaction process, and the interaction can be more natural and friendlier.

The invention discloses a compressor outlet parameter prediction modeling method based on a width learning system. The method comprises the steps of collecting data and preprocessing the data; constructing an outlet parameter prediction model of the compressor by utilizing a width learning system, setting the number of characteristic node groups of the width learning system as n, the number of k characteristic nodes in each group, the number of enhanced node groups as m and q enhanced nodes in each group, and performing modeling; when a new training data input model exists, using an incremental learning algorithm for carrying out online updating on an original model, an input data matrix of new training data being set as Xa belonging to Ra * M, an output data matrix being set as Ya belonging to Ra * C, a representing the number of newly-added training data, and performing model updating; and verifying the constructed model by using the root-mean-square error. According to the method, the performance prediction model of the large compressor can be quickly established, and the development time and cost of the model can be effectively saved.

The invention discloses a representative data reconstruction-based incremental SVR (support vector regression) load prediction method. The method includes the following steps that: electric load data are acquired; multiple-input-single-output pattern data are obtained through using a phase-space reconstruction theory; a support vector regression model is established by using the obtained pattern data and a particle swarm algorithm; newly-increased electric power load prediction data are obtained in real time; an optimal representative data subset is updated through using an incremental learning algorithm; model parameters are updated by using a nested particle swarm method; a support vector regression model is established by using the updated model parameters and optimal the representative data subset; and incremental load prediction is determined, and an incremental load prediction value is outputted. According to the method of the invention, support vectors of support vector regression are applied to the knowledge understanding research of massive data. With the method adopted, newly increased data-caused representative data reconstruction can be realized; the problems of high calculation complexity of massive data and difficulty in knowledge extraction can be effectively solved; the updating of the model parameters is realized in a nested manner; and references can be provided for the planning and operation of an electric power system.

The invention relates to an optimal scheduling method for a multi-objective optimization military transportation process, and belongs to the technical field of intelligent optimal scheduling of vehicle scheduling. A scheduling model and an optimization objective of the multi-objective optimization military transportation process are determined, and an optimal scheduling method based on a quantum population incremental learning algorithm is used, so that the optimization objective is optimized. The scheduling model and the optimization objective of the multi-objective optimization military transportation process are put forward, and the structure is clear and accurate. An optimal individual of a current population is obtained according to algorithm steps, and information about the optimal individual is fed back to a quantum bit observation model for updating a next generation population, so that global search through the algorithm can be effectively guided. The operation of Ejection Pool and Operate Assemble is executed on the optimal individual of the population, so that the local development capability of the algorithm is remarkably enhanced, and the solution quality is further enhanced.

The invention discloses a knowledge base construction and partial order structure graph generation method based on incremental learning. The incremental learning idea based on the covering principle is adopted to guide the covering operation of attribute and object set, and the special set such as attribute database is constructed by special covering relation, and then the formal background is generated, and the redundant pattern is deleted. On the basis of formal background, the algorithm of computer generation is to get the precise coordinate location of each concept node by computing the hierarchical pattern matrix, and to complete the dynamic construction of partial order structure graph by synthesizing the elements of hierarchy, node, line and so on, and introducing incremental learning algorithm. Compared with the prior art, the invention combines the incremental learning based on the covering principle and the construction algorithm of the partial order structure graph to realize the dynamic construction of the knowledge concept base and the automatic generation of the hierarchical pattern matrix. The progressive construction of partial order structure graph is realized.

The invention provides a method for controlling standby restart of a centrifugal-type water cooling unit. The method includes step S1, acquiring running parameters before standby, and comparing the running parameters with preset parameters to determine a standby reason; step S2, when the standby reason is water temperature standby, setting restart target parameters according to an incremental learning algorithm; step S3, controlling restart of the unit according to the restart target parameters. According to the method, when the standby reason is water temperature standby, the restart target parameters are set according to the incremental learning algorithm, and restart of the unit is controlled according to the restart target parameters, so that running load of the unit is enabled to be matched with working conditions, the unit is enabled to run under small-load working conditions, potential safety hazards caused by frequent start and stop of the unit are lowered, and the unit is enabled to be more energy saving. In addition, high-low pressure difference of the unit can be lowered, so that the unit is enabled to be away from a surge area, and potential surge possibility of the unit is lowered.

The invention discloses a logistics customer loss prediction method and system, which has high fault tolerance, accuracy and timeliness, is rapid and reliable in prediction, and is suitable for predicting customer loss of dynamic data flow. According to the technical scheme, the method comprises the steps of collecting logistics industry customer behavior preference big data; extracting conditionfactors with relatively strong correlation with the customer loss from the collected logistics industry customer behavior preference big data by adopting a rough set theory, thereby forming an original decision table; performing information entropy-based discretization processing on continuous attribute values in the formed original decision table to obtain an initial decision table; reducing redundant condition attributes in the initial decision table to obtain a secondary decision table, which serves as an input of a BP neural network; and by using an incremental learning algorithm, adaptively determining the number of neurons in a hidden layer by using an ELM as a basic classifier, and verifying the accuracy of the decision table; and adjusting a weight and a threshold value of an output layer, and training a data set until an optimal solution is output.

The invention discloses a face recognition tracker based on Haar-like features and an incremental learning algorithm, and mainly relates to the field of computer vision and image processing. Accordingto the method, Haar-like feature evaluation is accelerated by using an integral graph, strong classifiers for distinguishing human faces and non-human faces are trained by using an AdaBoost algorithm, and the strong classifiers are cascaded together by using screening type cascading, so that the accuracy is improved. And the face tracking part predicts the position of the central point of the current time frame according to the position of the central point of the previous frame of image tracking frame. And main features of the image in the frame are extracted by using a PCA algorithm, and acorresponding dimension-reduced graph is predicted according to the position of the center point of the frame at the moment. A forgetting factor is introduced, and image data is updated once every five frames. The incremental algorithm does not need to train a model, so that the efficiency is improved. Theoretics and practices show that the method can automatically recognize a human face, when thedirection of the human face changes greatly, for example, when the front face becomes a side face, recognition and tracking can be continued, continuous recognition is kept, and interruption is avoided.

The invention provides a voiceprint recognition method based on negative correlation incremental learning. The method comprises: step one, preprocessing and feature extraction are carried out on an inputted voice signal; step two, network integration is initialized; and if network integration already exists, current all networks are copied; step three, the network integration is trained; step four, structure adjustment is carried out on each network in the network integration; step five, the current networks are screened and the part of optimal networks are selected; and step six, the currently obtained networks are applied; and if new data arrive, steps are executed circularly by starting with the step one. According to the voiceprint recognition method provided by the invention, with the incremental learning method, voiceprint recognition is studied, so that the efficiency and identification accuracy under the data incremental arrival scene can be improved; and an incremental problem can be solved by using the negative-correlation-learning-based incremental learning algorithm. Improvement is carried out from perspectives of model training and model selection and thus a novel algorithm is put forward to solve problems; and then the novel method is applied to incremental learning.

The invention discloses an annotation system suitable for substation equipment inspection images. The system comprises an application module and an annotation model, wherein the application module comprises three functional modules including an import module, a regional operation module and a detection and analysis module; the annotation model comprises four parts including an incremental learningalgorithm, a semi-automatic annotation strategy, an annotation functional module and an annotation process design. According to the method, efficient and accurate annotation of a large number of inspection images generated by a transformer substation scene every day can be achieved, and compared with a time-consuming and labor-consuming manual annotation method and a low-accuracy full-automatic annotation method, the annotation time can be greatly shortened, and extremely high accuracy is achieved. Further, the work of data set acquisition, classification, annotation and the like can be closely linked, and various work systems are efficiently and clearly displayed in front of an operator, so that the work efficiency is remarkably improved.

In view of how to analyze the operating state of a transformer and perform fault diagnosis based on a vibration method, the embodiment of the invention discloses an online abnormality detection methodfor transformer vibration. The method comprises: obtaining new data samples about a vibration signal of the transformer operating state within a preset time period; extracting the characteristic parameters of the new data samples based on wavelet packet analysis; completing the learning of the new data samples based on a fast convex hull algorithm, training and updating a single-class abnormalitydetector model; and calling the updated single-class abnormality detector model to perform online abnormality detection of the current transformer vibration. According to the online abnormality detection method for transformer vibration, since the incremental learning algorithm is used to learn only the new data samples within the preset time period, the new data can be processed in time and effectively, the new training samples can be learned online in real time, and rapid upgrade of detection models can be achieved. At the same time, time and space requirements of the model update can be reduced.

The mobile application malicious software detection method for the power enterprise is characterized by comprising the steps that to-be-detected software is acquired, and decompilation is conducted onthe to-be-detected software to obtain a source code of the to-be-detected software; Extracting a feature vector of the source code, inputting the feature vector of the source code into a pre-constructed SVM classification model for comparison, and determining whether the to-be-detected software is malicious software or not; Wherein the SVM classification model comprises an SVM classifier, and theSVM classifier performs iterative updating based on a double weight increment method. According to the technical scheme, the problem that a support vector machine is not suitable for classified learning of a large number of samples is solved, the SVM incremental learning algorithm based on the double-weight function is provided for learning and classifying the application samples, and the methodcan reduce the learning time to the maximum extent and improve the learning efficiency on the premise that it is guaranteed that the application classification precision is not reduced.

The embodiment of the invention provides a method and device for training a risk identification model. According to the method of the embodiment, firstly, determining a newly-added task; then determining the similarity between the newly added task and an existing task corresponding to a first risk identification model obtained by training; if the similarity meets a preset condition, modifying a structure in the first risk identification model according to the newly added task, the modified structure comprising a newly added structure and/or a deleted structure; and finally, updating model parameters of the first risk identification model containing the modified structure by using the newly added task by adopting an incremental learning algorithm to obtain a second risk identification model.

The invention relates to an intelligent warehousing sorting method and system based on incremental learning. The method comprises the following steps: identifying object types of storage object images in a storage object sample image set by adopting an open set identification algorithm; when the storage object image is a newly-added storage object image, generating a newly-added storage object image data set; optimizing the warehousing system classification model to obtain an incremental learning algorithm model, wherein the warehousing system classification model is a warehousing object classification model originally implanted in the intelligent warehousing system; training an incremental learning algorithm model by adopting a newly added storage object image data set; and inputting the to-be-detected storage object image to the trained incremental learning algorithm model to obtain the object type of the to-be-detected storage object. According to the invention, the incremental learning model is applied to the intelligent warehousing system, so that the recognition accuracy of old-class warehousing objects can be improved, and the recognition efficiency of old-class and new-class warehousing objects through autonomous learning can be improved.

The invention discloses a vehicle sending information authenticity real-time detection method based on width learning. The vehicle sending information authenticity real-time detection method comprisesthe following steps: (1) collecting and sharing vehicle data information; (2) performing feature extraction on the acquired information, standardizing the proposed features, and standardizing the authenticity label of the message by using a two-dimensional vector; (3) establishing a false message detection model; (4) training a false message detection model; (5) detecting the authenticity of themessage sent by the vehicle; and (6) updating the model: when new data collected in the system is accumulated to a certain quantity, triggering the model to automatically utilize the newly added datato update by using an incremental learning algorithm. According to the method disclosed by the invention, the requirements on real-time performance and expandability of improper behavior detection ofthe vehicle can be well met. Meanwhile, the resource consumption is reduced, a more reliable basis is provided for traffic safety analysis, and potential safety hazards in traffic are reduced.

The invention provides a drug response prediction method based on incremental width and deep learning, a medium and equipment. The method comprises the following steps: performing text coding and position coding on a drug sequence to construct a drug information code; inputting the drug information code into a Transform encoder to excavate the structural features of the drug, inputting gene expression data into the feature representation of a multilayer perceptron learning gene, and splicing the drug features and the gene features together to form a drug-gene feature pair; and inputting the feature pair into a width learning system to obtain a predicted drug sensitivity regression value. The method can solve the problem of poor drug expression; drug expression and gene expression characteristics are fused by adopting a width learning system, so that the accuracy of a drug sensitivity prediction result is improved; the network weight is updated through an incremental learning algorithm, the model performance is improved, and the whole model does not need to be retrained.

The invention discloses a lithium battery temperature field online modeling method and system. The method comprises the following steps: firstly, carrying out space-time separation on space-time data by using a KL method to respectively obtain a space basis function and a time coefficient; secondly, training a time coefficient by using a width learning network to obtain an offline model of the lithium battery temperature field; and then updating the network weight of the off-line model by using an incremental learning algorithm of width learning, establishing an on-line model of the lithium battery temperature field, and finally reconstructing an obtained prediction time coefficient and a space basis function to obtain a final prediction model to predict data. According to the method, modeling is carried out on the temperature field of the lithium ion battery, and online updating is carried out on the model by using an incremental learning algorithm with excellent width learning, so that the model adapts to the time-varying behavior of the system, and the problem of temperature field prediction with unknown, nonlinear and time-varying characteristics of the system is solved.

The invention relates to a crop disease and pest identification method and device based on incremental learning and a storage medium, and the method comprises the steps: obtaining the existing disease and pest image data which comprise the known disease and pest image data; preprocessing the known disease-like pest image data, and adding a first disease-like pest category label to the known disease-like pest image data to obtain a known disease-like pest sample set; taking the known disease and pest sample set as input, and training a neural network by using a cross entropy loss function to obtain a standard disease and pest image classification model; modifying a structure in the standard disease and pest image classification model by adopting an incremental learning algorithm, and modifying a total loss function in the incremental learning algorithm by adopting a dynamic parameter correction method to obtain an incremental disease and pest image classification model; and inputting the obtained new disease and pest image data into the incremental disease and pest image classification model for identification and classification. According to the method, new disease and pest categories can be accurately detected, and the identification result is more accurate.