52results about How to "Optimizing Model Parameters" patented technology

The invention discloses a fault detection and identification method for civil aircraft system based on LSTM-AE depth learning framework and relates to the technical field of condition monitoring and fault diagnosis of complex civil aircraft system, and can be used to realize the monitoring and identification of flight faults. The invention comprises the following steps: selecting time series dataof multi-state parameters of an aircraft in flight under a certain stable condition, and according to the characteristics of the monitored object, the time series data of state parameters under suitable conditions are selected for the training of the system reconstruction model, then the fault-free state of civil aircraft system is modeled and reconstructed by making full use of the long-time series-dependent memory ability of LSTM model. The fault monitoring and identification are realized by further analyzing the reconstruction error of its state parameters. The invention solves the problemof insufficient fault monitoring means of civil aircraft system, utilizes the advantage of deep learning in big data analysis to mine massive operation and maintenance data of civil aircraft, and provides important support for fault monitoring of civil aircraft system and route fault isolation.

The invention relates to a segmentation-free off-line handwritten Chinese character text recognition method. The method comprises the steps that (S1) an off-line handwritten Chinese character text image is preprocessed; (S2) a spatial transformation network model is constructed; (S3) a deep convolutional neural network model is constructed; (S4) a recurrent neural network model is constructed through depth features extracted by the deep convolutional neural network model; (S5) probability distribution of sequence tags is output through a classifier CTC; and (S6) greedy search and search basedon dictionary rules are adopted to obtain a final text recognition result. According to the method, by the adoption of a model combining a spatial transformation network, a deep convolutional neural network and a recurrent neural network, correction processing and segmentation-free recognition can be performed on text lines with large offset, and the accuracy and robustness of recognition of complicated text lines are improved; the whole model framework is solved based on an iterative algorithm without the need for complicated excessive segmentation preprocessing, therefore, losses brought byan excessive segmentation method can be well reduced, entire model parameters can be optimized in a united mode, and recognition accuracy is improved.

The invention discloses a variable selection and forecast method of the silicon content in molten iron of a blast furnace. According to the method, blast furnace process parameters of a prediction model of the silicon content in molten iron of a blast furnace are used as input variables; after normalization preprocessing is performed on sample data of the input variables, variable selection is performed on the sample data of the input variables by using a multivariate correlation analysis method and a Spearman rank correlation analysis method, to remove the correlation between production process parameters; and a support vector machine algorithm is used to establish the forecast model of the silicon content in molten iron of the blast furnace, and particle swarm optimization algorithm is introduced to optimize model parameters. The variable selection and forecast method of the silicon content in molten iron of a blast furnace has universal applicability in forecasting the silicon content in molten iron during smelting of a blast furnace, and can achieve desirable forecast accuracy and improve the forecast hit rate of the silicon content in molten iron.

The invention relates to a voice separation model training method and device, a computer readable storage medium and a computer device. The method comprises the steps of obtaining a first audio and asecond audio; wherein the first audio comprises a target audio and correspondingly has a marked audio; wherein the second audio comprises noise audio; obtaining a coding model, an extraction model andan initial estimation model; performing unsupervised training on the coding model, the extraction model and the estimation model according to the second audio, and adjusting model parameters of the extraction model and the estimation model; performing supervised training on the coding model and the extraction model according to the first audio and the annotated audio corresponding to the first audio, and adjusting model parameters of the coding model; and continuously carrying out unsupervised training and supervised training, so that the unsupervised training and the supervised training arecarried out in an overlapping manner, and ending the training until a training stopping condition is met. According to the scheme provided by the invention, the model training cost can be reduced.

The invention provides a distributed group robot cooperative clustering algorithm based on an improved gene regulation network. Through embedding a diamond mesh distribution equation and a trajectoryfollowing equation in a gene regulation network model based on the Turing reaction diffusion mechanism, the moving vector speed of each robot is controlled, thus each robot whose initial state is in random distribution always gathers at a preset cluster trajectory position at a time t, the robots are arranged in a self-organized way to be in a diamond mesh distribution and can avoid obstacles in adynamic environment and repair formation by themselves. Parameter values in the improved gene regulation network are given by an NSGA II optimization algorithm. The distributed group robot cooperative clustering algorithm has the advantages of low computational complexity and good expansibility, for any robot, only the collection of the location information of a neighboring robot is needed, so arequired communication range is small, and the communication burden is effectively reduced. In addition, if some robots fail in operation, a system can still work normally, and the algorithm has goodrobustness and a great application prospect.

The invention provides an electrical load prediction method and system based on big data. The method comprises the steps of acquiring weather forecast data of an electrical region based on predictiontime; and substituting the prediction time, the electrical region and the weather forecast data into a pre-established prediction training model to obtain a historical prediction electrical load within the period of prediction time, wherein the prediction training model is determined by a training characteristic dataset of the electrical load, the time and the weather data based on a GBDT (Gradient Boosting Decision Tree). According to the electrical load prediction method and system based on the big data, an operation of writing intermediate data into a memory is implemented through the big data, the computational efficiency is greatly improved for the system for implementing the effect that the data are processed in real time, and the problems that the historical mass electrical load data cannot be rapidly acquired, processed, analyzed and stored in a conventional method can be well remedied through a big data analysis technology.

The invention discloses a user electricity consumption prediction method based on a Prophet-LSTM model, and the method comprises the following steps: S1, obtaining the historical data of the electricity consumption of a user through an intelligent electric meter, wherein the historical data comprise time series data, weather and temperature data, and holiday and festival data; S2, preprocessing and normalizing the historical data, wherein the original power consumption data is X={x1, x2,..., xn}, and the preprocessing of the original data comprises the processing of missing values, abnormal values, repeated values and invalid values; S3, constructing a Prophet prediction model, inputting the processed historical electricity consumption data X'={x'1, x'2,..., x'n} into the Prophet model, and performing Prophet prediction; S4, in order to prevent prediction overfitting, performing combined prediction in combination with an improved long and short term memory (LSTM) network model; and S5, measuring and verifying the fitting degree and the prediction effect of the combined model, and using common evaluation indexes. According to the method, the characteristics and rules of the power consumption data are analyzed, the accuracy of the prediction model is improved, and the method has important guiding significance for making effective power supply services by the state grid and each power supply company.

The invention relates to a process industry system prediction module based on a crossed relevance time-lag gray correlation analysis. The process industry system prediction module based on the crossedrelevance time-lag gray correlation analysis comprises the steps that the degree of association between all candidate variables and target variables is calculated; the variables are arranged in an order descending mode, the variable with the degree of association greater than the degree of association threshold value is obtained, and feature variable set is obtained. The feature variable set is taken as an input variable of an index prediction module, and the relative delay time of the feature variable set is mixed into the process of mould establishment. The index change tendency is predicted through an artificial neural network, the prediction mould is trained, the minimum prediction error is taken as a target, an optimal input feature is selected, and a prediction module is established. The time series mixing delay time in different sections of the feature variables in an optimal input feature subset is taken as the input of the index prediction model, the model is tested, the result is compared with the real value of the target variable, and the predictive performance is quantitatively evaluated. According to the process industry system prediction module based on the crossed relevance time-lag gray correlation analysis, the precision of the whole model is improved, and the effective prediction on a process industry key index is finally realized.

The invention relates to a method for predicting a corrosion rate of buried pipelines based on an APSO-optimized LSSVM. The method is applied to the prediction of the corrosion rate of the buried pipelines based on an adaptive particle swarm optimization (APSO)-optimized least squares support vector machine (LSSVM). The method has a significant predictive effect on improving the corrosion rate ofthe buried pipelines. An LSSVM model has a faster learning speed in the modeling process. At the same time, the APSO is used for parameter optimization, which improves the prediction accuracy and generalization ability of the model. The availability in predicting the corrosion rate of the buried pipelines is very high.

The invention provides an internet information risk prompting method and system. The internet information risk prompting method comprises the steps that internet information serves as original text data, word segmentation is conducted on the original text data, words are converted into digital identifiers, and segmented word arrays consisting of the digital identifiers are formed; according to theassociation relation of corresponding words of the digital identifiers in the segmented word arrays and context, the segmented word arrays are mapped into corresponding word vector sequences; classified identification is conducted on the word vector sequences through a neural network model to obtain classification result data; risk analysis is conducted on the classification result data, and riskprompts are sent according to risk analysis results. The internet information risk prompting method can understand human languages in the mode of thinking closer to human thinking, and a variety of information can be quickly and intelligently identified. Thus, when much information is processed, workers can be replaced to organize huge complicated service texts and rapidly identify risk event information, and lots of labor and time are saved.

A soft measuring meter based on bionic intelligence comprises on site intelligent meter connected with the industrial process object, storage device and upper position machine, intelligent meter, data storage device and the upper position machine connected sequentially, the said upper position machine being soft measurement intelligent processor, which comprises standardized handling module, radial basic function neural network establishment module, module parameter optimization module based on chaotic genetic algorithm, signal sampling module and soft measurement module. It also puts forward a failure diagnostic method. It is convenient, extensive in application, fine in measuring result, high in precision.

The invention discloses electromagnetic sounding constraint inversion methods based on a resistivity equivalence principle and belongs to the field of geophysical exploration. A method is characterized by comprising the following steps of a, assuming that a one-time inversion result is m, and carrying out a forward action on the m to obtain the derivative of response with respect to layer thickness, thereby obtaining deltam; b, taking the deltam as a coefficient corresponding to epsilon in an S equivalence convergence algorithm of each layer; c, carrying out low resistivity thin layer convergence through utilization of S equivalence, thereby obtaining converged thin layer resistivity and thin layer thickness; and d, carrying out high resistivity thin layer thickness compensation through utilization of H equivalence, adding 1 / 2 of low resistivity thin layer reduced thickness to an upper high resistivity thin layer, and solving resistivity of the high resistivity thin layer. According tothe method, under the condition that no any known condition exists, relatively accurate high resistivity thin layer and low resistivity thin layer information can be obtained at the same time, so inversion explanation precision is improved; computing speed is fast; and adaptability is high.

The invention provides a sewage treatment effluent parameter prediction method based on a grey neural network composition model, the method comprises the steps: a historical influent parameter sequence x (t) is acquired and is accumulated to form an accumulation sequence y(t); according to the accumulation sequence y(t), a grey neural network is established and is utilized to predict historical effluent parameters, and in the prediction process, an improved particle swarm optimization is adopted to optimize the grey neural network; according to the optimized neural network, present effluent parameter prediction is performed; a Markov chain is adopted to correct obtained present effluent parameters.

The invention provides an application layer malicious request detection method based on a Transformer model. The method comprises the following steps: firstly, constructing an application layer user request data sample set; constructing a dictionary according to the occurrence frequency of lexical elements in the application layer user request data, and vectorizing the application layer user request data according to the dictionary; performing data embedding and position encoding on a vectorized result to obtain a feature matrix; predicting the category of the user request data of the corresponding application layer for the feature matrix based on a Transformer model; calculating a loss function value according to a model prediction category and a real category, and optimizing network parameters; and finally, vectorizing to-be-detected application layer user request data, and performing feature extraction and category prediction on the to-be-detected application layer user request databased on the optimized network to realize application layer malicious request detection. The method is good in effect and high in efficiency.

The invention discloses a lake-reservoir algal bloom generating mechanism time varying model optimization and prediction method based on a taboo searching algorithm and a genetic algorithm. The method comprises the steps that first, a water bloom generating mechanism time varying model is established; second, an influence factor function model base is established; third, based on the genetic algorithm, water bloom generating mechanism time varying model parameters are optimized; fourth, based on the taboo searching algorithm, a water bloom generating mechanism time varying model structure is optimized, and influence factors are analyzed; and fifth, optimum water bloom generating mechanism time varying model prediction is carried out. According to the method, a time variable is introduced into the water bloom generating mechanism model, the water bloom generating mechanism time varying model is established, so that the method is suitable for simulating a water bloom generating process and can be used for water bloom prediction, and the problem that water bloom prediction based on a data driving model is not accurate enough, and a mechanism driving model cannot predict water bloom is solved.

The embodiment of the invention provides a text labeling method and device based on teacher supervision, and belongs to the technical field of natural language processing. The method comprises: usinga character labeling model to label a to-be-labeled text, and generating a character labeling result containing labeling words; carrying out word segmentation processing on the to-be-labeled text through a preset word segmentation model to generate a word segmentation result containing word segmentation words; and according to the similarity between each annotated word and each segmented word, carrying out character annotation on the character annotation result again based on the segmented words to obtain a fusion annotation result, and outputting the fusion annotation result. By adopting themethod and the device, the text labeling accuracy and recall rate can be improved.

The invention discloses a method and a device for predicting effective wave height of sea waves based on 20CR data. The method comprises the steps of acquiring original data in the 20CR data, and carrying out data preprocessing on the original data; selecting a sea level pressure field meeting a preset condition; correcting a prediction model according to the data related to the sea level pressurefield in the data of a first age section in the 20CR data; assessing the prediction model according to the data of a second age section later than the first age section in the 20CR data; and after the assessment is passed, predicting the effective wave height of the sea waves by adopting the prediction model. According to the method and the device, the 20CR data jointly published by the United States Department of Energy and the National Oceanic and Atmospheric Administration in the United States is adopted; the time span exceeds the hundred years; and after the sea level pressure field meeting the preset condition is selected, the prediction model is corrected according to the data related to the selected sea level pressure field in the data of the first age section in the 20CR data, andmodel parameters are optimized, so that the accuracy of predicting the effective wave height of the sea waves is improved.

The embodiment of the invention discloses a neural network model training method and device for sensing data fusion, a sensing data fusion method and device based on a neural network model, electronic equipment and a storage medium. The neural network model training method for sensing data fusion comprises the following steps: constructing training data based on a sample image and a sample three-dimensional point cloud data set; and training an initial neural network model based on the training data to obtain a trained neural network model. According to the embodiment of the invention, the neural network model can be better and continuously iteratively optimized based on the sample data of the long-tail scene.

The embodiment of the invention discloses a linguistic model training method. The method comprises the steps that a terminal determines a first linguistic model; the terminal trains the first linguistic model through historical input data locally generated so that a trained first linguistic model can be obtained; and the terminal sends first model parameters of the trained first linguistic model to a server, wherein the first model parameters are used for determining second model parameters of a second linguistic model, and the second linguistic model is used for displaying candidate items obtained through association for the terminal. Therefore, in the process of obtaining the second linguistic model, the historical input data containing the privacy of a user are trained on the local terminal through the first linguistic model, the original historical input data do not need to be uploaded to the server, and thus the risk of exposure of the private information of the user is reduced; and meanwhile, the terminal only uploads the first model parameters of the trained first linguistic model to the server, thus the uploaded data volume is relatively small, and therefore, the uploadingefficiency of the terminal during the process of obtaining the second linguistic model is improved.

The invention belongs to the technical field of security and protection monitoring, and particularly relates to a safety protection equipment identification method based on an autonomous learning strategy and storage medium. The method comprises the following steps: S1, collecting a training picture set; S2, preprocessing the training picture set; S3, training the constructed deep network model according to the training picture set; S4, inputting a to-be-detected picture set into the deep network model for identification to obtain an identification result set; S5, classifying the recognition result set into a successful recognition set and an unrecognizable set; S6, outputting an identification success set; and S7, taking the to-be-detected picture set corresponding to the unrecognizable set as a new training picture set, and skipping to S2 to continue execution. The method does not need to occupy a large number of artificial resources and computing resources, generates new training samples in a semi-automatic manner, and is suitable for different complex scenes. By introducing the weight of the features and a label smoothing mechanism, the accuracy of obtaining the network features is ensured, and overfitting is effectively prevented.

The invention discloses a lithium battery SOP estimation method based on an electrochemical model in an electric automobile, and the method comprises the steps: calculating a fitness value in a genetic algorithm according to a model voltage and a real-time voltage, finding out an optimal solid-phase diffusion coefficient of positive and negative electrode particles, calculating an optimal lithium ion concentration value of the outermost layer of the positive and negative electrode particles, solving peak charging current and discharging current limited by the optimal concentration value of lithium ions on the outermost layers of the positive and negative electrode particles, and calculating to obtain peak charging and discharging power; a lithium ion concentration limit is adopted for SOP prediction, the power capacity of a battery can be better reflected from the electrochemical principle, the method is more sensitive to the front state change of the battery, the SOP prediction of the battery under a dynamic working condition is higher in precision and reliability, the genetic algorithm is adopted for carrying out real-time parameter identification on the model, model parameters can be optimized in real time, the influence of temperature change and battery aging on a prediction result is reduced, and the prediction precision is improved.

The invention provides a distributed group robot cooperative clustering algorithm based on an improved gene regulation network. Through embedding a diamond mesh distribution equation and a trajectoryfollowing equation in a gene regulation network model based on the Turing reaction diffusion mechanism, the moving vector speed of each robot is controlled, thus each robot whose initial state is in random distribution always gathers at a preset cluster trajectory position at a time t, the robots are arranged in a self-organized way to be in a diamond mesh distribution and can avoid obstacles in adynamic environment and repair formation by themselves. Parameter values in the improved gene regulation network are given by an NSGA II optimization algorithm. The distributed group robot cooperative clustering algorithm has the advantages of low computational complexity and good expansibility, for any robot, only the collection of the location information of a neighboring robot is needed, so arequired communication range is small, and the communication burden is effectively reduced. In addition, if some robots fail in operation, a system can still work normally, and the algorithm has goodrobustness and a great application prospect.

The invention belongs to the technical field of satellite navigation positioning and meteorological crossing and discloses a novel high-temporal-spatial-resolution global ZTD vertical section grid model construction method. The method comprises the steps of dividing the whole world into regular windows of 5 degrees * 4 degrees based on a sliding window algorithm, expressing a ZTD vertical profilefunction model of each window of the whole world by utilizing a negative exponential function, and estimating a ZTD vertical profile model coefficient of each window of the whole world based on MERRA-2 reanalysis data; and based on the five coefficients of each window beta factor in the world, storing in the form of a grid with the plane resolution of 5 degrees * 4 degrees to obtain a global ZTD vertical section grid model with the plane resolution of 5 degrees * 4 degrees and considering the fine seasonal change of the elevation scaling factor. According to the global ZTD vertical section grid model established by the invention, the model parameters are better optimized, and the practicability of the model is improved.

The invention relates to the technical field of Internet of Things, in particular to a neural network monitoring model for Internet of Things edge node security. The neural network monitoring model comprises a hybrid memory type neural network model. The neural network monitoring model is characterized in that the hybrid memory type neural network model comprises a sparse automatic encoder, a deepbelief network module, a classifier and a memory module. The model comprises the following steps: S1, the input information of the sparse automatic encoder is x, output is hW, b (x), and the objective function expression of the sparse automatic encoder is shown in the specification, wherein W and b represent a parameter and a bias term respectively, and J (W, b) represents an objective function of a conventional sparse automatic encoder. The problem that the classification judgment capability of the Internet of Things fog end gateway for the data flow security condition is insufficient when facing large-flow node data can be effectively solved, the dimensionality reduction capability can greatly reduce the dimensionality and complexity of data processing, the memory capability can effectively memorize the type and the condition of node security, and finally the accuracy and the high efficiency of real-time security monitoring are obviously improved.

The invention provides a method for manufacturing a multi-curved surface lens light distribution device. A multi-curved surface lens of an axisymmetric structure composed of eight curved surfaces is designed and manufactured to serve as the light distribution device to conduct light distribution on an LED street lamp, so that a rectangular lighting light spot is obtained. The method for manufacturing the multi-curved surface lens light distribution device is characterized by comprising the following steps that step 1, the length L, width D and height H of the multi-curved surface lens and the installing height h of the LED street lamp are determined primarily; step 2, the cross section of the multi-curved surface composed of eight curved lines is designed through the SolidWorks and the structural parameter of the cross section is obtained; steps 3, the cross section is stretched longitudinally to be at the length L to obtain a complete multi-curved surface lens model; step 4, the multi-curved surface lens model is induced to optical design software Lighttools, stimulation optimization is conducted, and the optimized structural parameter of the multi-curved surface lens light distribution device is obtained; step 5, the multi-curved surface lens light distribution device is manufactured according to the optimized structural parameter of the multi-curved surface lens light distribution device.

The invention provides a screening method and device for a blind area detection training set, a server and a storage medium. The method comprises the steps: inputting an image into a semantic segmentation neural network; acquiring an output matrix of a feature recognition deconvolution layer in the semantic segmentation neural network, wherein the output matrix is used for representing feature intensity corresponding to each pixel of the output image; mapping the output matrix to a basic image template to form a feature image, wherein the basic image template is used for highlighting the intensity of feature intensity; and calculating an image feature value mean value of a preset region of interest in the feature image, comparing the feature value mean value with a preset mean value threshold value, and when the feature value mean value is not greater than the preset mean value threshold value, determining that the image is added into a blind area detection training set. A unified, objective and visual prediction result of a data set can be obtained, and targets with poor feature extraction and uncertainty can be added to next training.

The invention discloses a lake-reservoir algal bloom generating mechanism time varying model optimization and prediction method based on a taboo searching algorithm and a genetic algorithm. The method comprises the steps that first, a water bloom generating mechanism time varying model is established; second, an influence factor function model base is established; third, based on the genetic algorithm, water bloom generating mechanism time varying model parameters are optimized; fourth, based on the taboo searching algorithm, a water bloom generating mechanism time varying model structure is optimized, and influence factors are analyzed; and fifth, optimum water bloom generating mechanism time varying model prediction is carried out. According to the method, a time variable is introduced into the water bloom generating mechanism model, the water bloom generating mechanism time varying model is established, so that the method is suitable for simulating a water bloom generating process and can be used for water bloom prediction, and the problem that water bloom prediction based on a data driving model is not accurate enough, and a mechanism driving model cannot predict water bloom is solved.

The invention provides a privacy calculation method and system based on a feature engineering IV value and a readable storage medium, and the method comprises the steps: a participant A generates a public and private key pair of the participant A, and publishes a public key to a participant B; the participant A encrypts the label value of each piece of sample data by using a public key, and sends the ciphertext label value to the participant B; the participant B groups the plurality of sample data based on the characteristic values, and calculates a ciphertext IV value of each group in combination with the ciphertext label value of each sample data and the public key of the participant A; the participant B accumulates the ciphertext IV values of the groups to obtain a final ciphertext IV value; the participant B scrambles the final ciphertext IV value and then sends the final ciphertext IV value to the participant A; the participant A decrypts the final ciphertext IV scrambling value by using a private key to obtain a plaintext IV scrambling value and sends the plaintext IV scrambling value to the participant B; and the participant B performs descrambling on the plaintext IV scrambling value to obtain the plaintext IV value of the feature. According to the method, the privacy calculation of the multi-party feature engineering IV value can be realized.

A target detection method based on CARLA simulation mainly solves the problems that in existing automatic driving, target detection research is high in cost, poor in precision, low in speed and incapable of dealing with complex road conditions, and the implementation scheme is that the method comprises the steps: calling a sensor provided by a CARLA simulator to acquire image information, and generating a large number of training samples used for training a target detection model; taking the images as a training set, a verification set and a test set of a target detection algorithm; obtaining a model by training a target detection algorithm based on a convolutional neural network, verifying and improving model parameters, repeating the process to obtain a final model, obtaining the related index information of the final model through testing, and finally, applying the target detection model to a CARLA simulator to achieve high-precision, high-speed, efficient and stable target detection in automatic driving.

The invention provides an OPLC thermal circuit modeling method based on a superposition principle, and belongs to the technical field of power cable detection. The method includes the steps of using COMSOL software to achieve the simulation of an OPLC temperature field; building thermal circuit models under the situations of OPLC four-core heat emitting and single-core heat emitting, and accordingto the superposition principle, building a thermal circuit model of OPLC three-core heat emitting; using a particle swarm algorithm to conduct parameter identification on the thermal circuit model ofOPLC three-core heat emitting; according to an identification result, building an OPLC optical fiber position thermal circuit model. According to the method, by using the superposition principle, superposition is conducted on the thermal circuit models under the two different situations of OPLC four-cable core heat emitting and single-cable core heat emitting, the accurate thermal circuit model under the situation of OPLC three-cable core heat emitting is obtained, and the building reasonability of the OPLC thermal circuit model is effectively improved. By using the particle swarm algorithm,parameters in the thermal circuit model are identified, the model parameters are optimized, the accuracy of temperature calculation values of all layers inside OPLC is effectively improved, and the method is of great significance for mastering the operation state of the OPLC, improving the reliability of OPLC operation and achieving the engineering application of the OPLC.