44results about How to "Improve model accuracy" patented technology

The invention discloses a high sulfur natural gas purifying process modeling and optimizing method based on an extreme learning machine. The method comprises the steps of determining the input variable of a model; acquiring production process data; preprocessing the production process data; conducting data normalization; conducting data modeling by means of the extreme learning machine to obtain a model of technological operation parameters to H2S and CO2 content; designing a preference function according to two output variables of the extreme learning machine model, and optimizing the input variable by means of the multi-objective genetic algorithm; applying input variable optimal solution sets to the extreme learning machine model in sequence to calculate two output values, namely the content of H2S and the content of CO2, of the model at the moment, comparing the output values with an average sample value, and observing the optimization effect. By the adoption of the method, an accurate and reliable high sulfur natural gas purification and desulfurization industrial process model can be established quickly, the yield of finished gas can be increased on this basis, energy consumption during desulfurization can be reduced, and the method has important practical significance in guiding actual industrial production.

A newly established crossing traffic flow prediction method based on a generating type deep belief network belongs to the technical field of short-period traffic flow prediction. The newly established crossing traffic flow prediction method settles the problems of small amount of data and low prediction precision in traffic flow prediction for a newly established crossing. The newly established crossing traffic flow prediction method comprises the steps of establishing a generating type deep belief network regression model with a 144 input structure and a 144 output structure based on a deep learning theory and a restricted Boltzmann machine; performing pre-training on the deep belief network regression model through mature crossing data of a city to which the newly established crossing is affiliated, and obtaining a deep belief network regression pre-training model; performing fine adjustment on the deep belief network regression pre-training model by means of prestored actual traffic flow data of the newly established crossing, and obtaining a final deep belief network regression model; and acquiring the current actual traffic flow data of the newly established crossing, and performing online prediction on the traffic flow by means of the final deep belief network regression model. The newly established crossing traffic flow prediction method is used for predicting the traffic flow of the newly established crossing.

The invention discloses a 3D game engine LOD system achievement method. The achievement method mainly includes the steps that a scene to be drawn is imaged to be a large terrain block, and the terrain block is divided into terrain block bodies in the specification of M*N, wherein M and N are both natural numbers; the vertex height value of each terrain block body is raised or reduced, namely the Y value of the vertex of each terrain block body is changed, the uneven terrain block is produced, and in the terrain coordinates XOZ, check terrain block bodies, namely terrain checks, corresponding to the plane where the terrain block is located are overlooked; a scene graph with a low resolution ratio is drawn according to different scenes, produced along with moving of a camera, of each terrain block body. According to the 3D game engine LOD system achievement method, the defects that in the prior art, model precision is low, terrain graphs are difficult to draw and rendering efficiency is low are overcome, and the advantages that the model precision is high, the terrain graphs are not difficult to draw and the rendering efficiency is high are achieved.

The present invention discloses a satellite telemetry data intelligent interpretation method based on a BP neural network. The method comprises an offline autonomous learning module and a real-time online interpretation module. The offline autonomous learning module performs autonomous learning based on a telemetry data sample in a historical telemetry database and a new obtained telemetry data sample to obtain a neural network model for telemetry data interpretation; and the real-time online interpretation module performs online real-time interpretation on telemetry data according to the neural network model obtained by the offline autonomous learning module. According to the method provided by the present invention, the telemetry data sample in the historical telemetry database is utilized to perform algorithm model learning and establishing, and the new obtained telemetry data sample is utilized to perform relearning in a telemetry data interpretation process; and in the entire telemetry data intelligent interpretation process, the precision of the neural network model for the telemetry data interpretation is increased gradually with time extending and telemetry data volumes increasing.

The invention discloses a die frame assembly which comprises an upper fixing plate, a female die plate, a female die core, a male die core, a male die plate, a lower fixing plate and a die opening / closing auxiliary part. The male and female die cores are respectively fixed on the male and female die plates through supporting blocks. The die opening / closing auxiliary part comprises a sliding block, a support block, a binding block and an angle pin. The sliding block is horizontally connected with the support block side by side. The support block is arranged on the male die plate in a slideablemode and can drive the sliding block to slide on the male die plate. The binding block is fixedly arranged on the female die plate and is positioned above the support block. The angle pin is selectively arranged on one of the support block and the binding block and the other one of the support block and the binding block is provided with a chute. The angle pin can draw the support block in the process of closing a die so that the sliding block moves towards the male die core, and limits the support block to displace towards the direction far away from the male die core after the die is closed. By the die frame assembly, the size of a die frame can be effectively reduced so as to reduce the used material of the die frame and the machining cost, the assembling difficulty of the die also canbe greatly reduced and the correction of assembling the die is improved.

The invention discloses a Saastamoinen model-based BP nerve network troposphere delay correction method. The method is characterized by comprising the following steps: S1, according to a Saastamoinen model, calculating a troposphere mostire delay value ZWDSAAS of a station; S2, establishing a BP nerve network representing a moisture delay at the station, and representing nonlinear rations between the moisture delay of the station and meteorological parameters and a Saastamoinen model moisture delay; S3, training the BP nerve network by use of high-precision IGS troposphere delay product data; S4, calculating the moisture delay at the station through the BP nerve network; and S5, calculating a troposphere zenith delay after modification. The precision of the method is quite high.

The invention provides a modeling method of a microwave high-power transistor. The modeling method comprises steps as follows: S1, a non-linear equivalent circuit model of a small-size unit-cell transistor is established; S2, electromagnetic simulation software is used for simulating microwave transmission characteristics of a passive component of a large-size transistor, and an S parameter of an input structure and an S parameter of an output structure are acquired; S3, thermal simulation software is used for simulating thermal transmission characteristics of the large-size transistor, parameter values of a thermoelectric coupling parameter network are extracted according to thermal simulation data, and the thermoelectric coupling parameter network is acquired; S4, the non-linear equivalent circuit model of the small-size unit-cell transistor, the S parameter of the input structure, the S parameter of the output structure and the thermoelectric coupling parameter network are connected according to a port corresponding relationship, and a large-size transistor model is obtained. Electromagnetic simulation data are used for describing the parasitic effect of an input-output structure, a gold wire, an isolation resistor and the like of the large-size transistor, thermal simulation data are used for extracting thermoelectric coupling parameters, the modeling precision is high, and the parameters are easy to extract.

The invention discloses a real-time equipment abnormality detection device and method based on a time sequence prediction model. The device comprises a data layer, a logic control layer, a model center and a display layer. The data layer comprises a real-time database, a data buffer, a historical database and a time sequence data processing module. The logic control layer comprises a trainer and apredictor. The model center comprises a machine learning model. And the display layer comprises a result display module. Compared with a traditional equipment abnormality detection scheme, the methodhas the advantages that the time sequence prediction model is utilized, information of a certain time window can be utilized in the equipment operation process, the model precision of the system is higher, and the result better conforms to the reality. The incremental training mode is supported, so that the model can capture the operation conditions of different devices in time, and the machine learning model of the model center can be updated in time.

The invention relates to a shield-tunnel lining segment T-shaped seam two-sealing-pad waterproof test device and method, and the device is wholly shaped like T, and comprises a lower test die, an upper left test die, and an upper right test die. The lower test die, the upper left test die and the upper right test die are respectively sequentially provided with an internal trench and an external trench from the inside to the outside on the corresponding contact surface. The interior of each internal trench and the interior of each external trench are respectively provided with an elastic sealing pad. The inner ring of the upper left external trench is a plane, or is provided with a rectangular groove or a rectangular boss. The inner ring of the upper right external trench is a plane or is provided with a rectangular groove or a rectangular boss. The lower external trench is a plane or is provided with a rectangular groove or a rectangular boss. Compared with the prior art, the device is advantageous in that the device can truly simulate the waterproof test of two sealing pads of a T-shaped seam under different stress conditions, is high in model precision, and truly simulates the waterproof capability test under different misplaced conditions.

The present invention discloses a distributed photovoltaic electricity-stealing supervising method based on multi-time scale output estimation. The method comprises: first, establishing a photovoltaic output calculation model based on historical statistical data of a power station; by using historical power generation information of the power station and contemporaneous meteorological information, generating a fitting curve of meteorological information and photovoltaic output; loading real-time meteorological information so as to obtain real-time output of the power station; and then based on multi-time scale photovoltaic output estimation, giving out an electricity-stealing identification method with a three-layer screening structure: real-time electricity-stealing estimation, short-term electricity-stealing estimation and medium and long-term electricity-stealing estimation, and giving out a corresponding inspection scheme according to an electricity-stealing suspect determination result. According to the distributed photovoltaic electricity-stealing supervising method based on multi-time scale output estimation, a statistical method is adopted, so that independent modeling of each constituent element of the photovoltaic power station is avoided, a calculation error is small, and model precision is improved; and by designing the three-layer screening structure of electricity-stealing identification, the electricity-stealing identification rate of distributed photovoltaic power generation is greatly improved, evidence is provided for effective supervision of distributed photovoltaic power generation, and the inspection efficiency is improved.

An engine remaining service life prediction method based on an improved stacked sparse auto-encoder (SSAE) and an attention echo state network (Attention-ESN) comprises the following steps: firstly, removing original noise by using a 3sigma criterion to obtain high-quality original data and realize data reconstruction, extracting features of each cycle of an engine by using the improved SSAE, and performing feature dimension reduction; adopting a BN layer and a Dropout layer in an encoder to solve the problems of gradient disappearance and overfitting, creating extracted engine features into an HI value to acquire an HI curve representing the degradation trend of an engine, finally introducing an attention mechanism into an ESN network, processing different types of features in a self-adaptive mode, optimizing network parameters and finally acquiring an RUL value. And the residual service life prediction of the turbofan engine is realized. The residual service life is predicted by adopting a combined model of feature extraction and a network prediction structure, and the prediction precision is improved. The abstract drawing is as shown in figure 1.

The invention discloses a Boolean simulation method for a sandbody with the complicated morphology. The method comprises the steps that a blank template is firstly built, the shape of the transverse section of the river channel sandbody is drawn according to the width and depth of the river channel section to obtain a shape template, all nodes of the shape template form a two-dimensional matrix K, and when the nodes of the shape template are located in the shape of the transverse section of the river channel sandbody, the matrix element km,n=1; when the nodes of the shape template are located outside the shape of the transverse section of the river channel sandbody, the matrix element km,n=0; a work area grid G with the dimensionlity of W*H is defined, all nodes of the work area grid G are sequentially accessed through a Boolean algorithm, values of all the nodes, corresponding to the nodes of the work area grid G, in the shape template are assigned to the nodes, and a section model of the sandbody with the complex geometrical morphology is obtained. According to the Boolean simulation method for the sandbody with the complicated morphology, determination of the shape of the transverse section of the river channel sandbody is not restricted by an elliptical shape parameter, a sandbody image with the complex geometrical morphology is built and Boolean simulation is performed, and therefore the simulation result is truer.

The invention provides a constructing method of a noise model for millimeter waves FET. The model comprises following steps: dividing millimeter waves FET into a passive structure area and an active structure area; equalizing the width direction of the active structure area along grid electrodes into N first units, equalizing the first units into parts equal to the number of fingers of millimeter waves FET in order to obtain multiple second units; setting up a low-frequency noise model of millimeter waves FET and utilizing the low-frequency noise model to obtain an intrinsic parameter network containing intrinsic parameters and noise source expressions of the second units; calculating transmission features of millimeter wave signals in input electrodes and output electrodes of millimeter waves FET in order to obtain input electrode S parameters and electrode S parameters; and connecting an intrinsic parameter network, input electrode S parameters and output electrodes S parameters of the second unit according to port corresponding relations in order to obtain a noise model of millimeter waves FET. The noise model has higher precision in millimeter waves and high frequencies.

The embodiment of the invention discloses a model training method and device based on risk identification and electronic equipment. According to the specific scheme, the method comprises the steps ofobtaining a first data set without a sample label, wherein the first data set comprises sample data expected to have a first type of sample label, and the sample data expected to have the first type of sample label is doped with sample data with a second type of sample label; and pre-configuring a first type of sample tags for the first data set, and operating the target model configured with thefirst model parameters by utilizing the first data set to generate a prediction value; utilizing a loss function to judge the loss amount of the predicted value relative to a target value reflected bythe first data set; and estimating a statistical center estimated value of the first data set corresponding to the loss amount, converting the statistical center estimated value into a statistical center expected value, and adjusting the first model parameter by using the loss amount and the statistical center expected value corresponding to the loss amount until the loss amount reaches a presetcondition.

The invention provides a user intention recognition method and device for an intelligent voice robot, and computer equipment. The method comprises the following steps: extracting candidate feature words from a historical question and answer text between the intelligent voice robot and a user, and establishing a feature database based on the candidate feature words; constructing a plurality of intention recognition models, wherein the plurality of intention recognition models comprise model parameters updated in the training process; obtaining a to-be-processed user voice text, and determining a corresponding intention recognition model; and outputting an intention prediction value of the to-be-processed user voice text by using the determined intention recognition model. The specific feature words can be effectively extracted, the feature database for user intention recognition is established, and ambiguity resolution and semantic uniformity of the specific feature words can be effectively realized; a plurality of intention sub-models for optimizing model parameters are constructed, so that the model precision can be improved; and the user intention can be identified more accurately, and finer-grained user intention mining can be realized.

The invention discloses a method and system for detecting imperfect grains of granular crops. Crops are arranged in order; upper and lower side image information of crops is collected, the upper and lower side image information is preprocessed, and image classification is performed through a classification model to obtain a corresponding crop imperfect grain classification report; a correspondingimperfect grain sorting prompt array image is generated according to an image classification result, and crop imperfect grain sorting collection is performed according to the imperfect grain sorting prompt array image. According to the invention, an artificial intelligence technology based on deep learning is utilized; construction of classification model is performed, the accuracy, repeatabilityand generalization of imperfect classification results of granular crops are improved, the method and system can be used for detecting imperfect grains of various granular crops, greatly shortening the sorting time of detectors, avoiding the disadvantage that the screening standard varies from person to person, and also solving the problem that the error rate is increased due to the work fatigue of the detectors.

The invention discloses a liquid medium aluminum melting furnace model prediction control method based on data driving, and belongs to the field of industrial process control, and the method comprises the following steps: collecting and storing data such as current and voltage signals of a three-phase electrode on site; selecting a controller according to the deviation between the actual value and the set value of the electrode current, selecting a prediction controller when the deviation is larger and lasts for 10 seconds, and selecting a PI controller when the deviation is smaller; when the prediction controller is used, establishing an LSSVM (least square support vector machine) prediction model; optimizing key parameters in the LSSVM by using a POS; and designing a model prediction control algorithm according to the obtained prediction model. Compared with a traditional PID control system, the method is more accurate, the model identification accuracy is high, a good control effect is achieved, and engineering requirements can be met.

The invention discloses a method for discriminating the shelf life of an apple through near infrared spectroscopy based on JADE and ELM. The method comprises the following steps: 1, collecting a sample, acquiring the spectrum of the sample to obtain sample near infrared diffuse reflection spectrum data, and packing the original near infrared spectrum data by using discrete wavelet transform; 2, decomposing the packed spectrum data by using a characteristic matrix joint approximate diagonalization algorithm to obtain independent component matrixes and a demixing matrix; 3, establishing an extreme learning machine analysis model by using an extreme learning machine technology with the demixing matrix as model input and corresponding shelf life as output; and 4, carrying out quality evaluation on the model, and determining the shelf life of the sample to be identified. The method has the advantages of realization of rapid identification of the shelf life of apples, enrichment of chemometrics methods, and good application prospect.

The invention provides a big data risk control model and online system configuration technology, which comprises risk control model construction and risk control model configuration, and is characterized in that the risk control model construction comprises the steps of sample acquisition, data processing, data table generation, data division, feature engineering and model establishment, wherein the model configuration comprises feature configuration, model configuration and monitoring configuration. By implementing the technical scheme of the invention, on one hand, a plurality of feature construction modes are combined together, and a variable cross combination operation mode is added, so that the combination relationship between variables is comprehensively mined, and the model precision is improved; on the other hand, on the basis of traditional feature screening, correlation inspection between variables is added, so that excessive variables entering the model can be prevented, thecalculation cost can be prevented from being increased, and model over-fitting can also be prevented; projects, characteristic variables, models, user roles and the like can be managed in a unified mode.

The invention discloses an airport gas detection system based on hyperspectrum. The airport gas detection system comprises a hyperspectral image data acquisition module, a data preprocessing module and a spectral analysis module, wherein the hyperspectral image data acquisition module comprises a plurality of single-waveband hyperspectral cameras with different shooting angles; the single-wavebandhyperspectral camera is used for acquiring hyperspectral image data of airport gas in real time; the data preprocessing module is used for preprocessing the collected airport gas hyperspectral imagedata; and the spectral analysis module is used for analyzing and processing the preprocessed airport gas hyperspectral image data, a bidirectional recurrent neural network model is arranged in the spectral analysis module, and the bidirectional recurrent neural network model inputs the airport gas hyperspectral image data and outputs content information of target gas in an airport. The invention further discloses an airport gas detection method based on the hyperspectrum. Large-scale real-time detection can be carried out on gas in an airport area, and simultaneous measurement of various target gases is realized.

The invention provides a small sample medical relationship classification method based on a multilayer attention mechanism, and relates to the technical field of relationship classification. The method comprises the steps: constructing a relation classification model based on a neural network, wherein the relation classification model comprises a word embedding layer, two position embedding layers, a coding layer and a full connection layer, sentences in a support set and a query set are input, and relation categories to which the sentences in the query set belong are output; obtaining a public relationship extraction data set, setting training times, training the relationship classification model by utilizing a training set of the relationship extraction data set, and randomly extracting a support set and a query set which are required for training the relationship classification model each time from the training set; for a support set containing any N relationships and a query set in which sentences contained in the support set belong to the N relationships, utilizing the trained relationship classification model to predict a relationship category in which the sentences in the query set belong to the support set. The influence of noise on the accuracy of the model is reduced from different aspects, and the relationship between entities is mined more accurately.

The invention discloses an enterprise core competitiveness evaluation method based on multilevel model fusion and a storage medium. The method comprises the steps of collecting enterprise data, performing dimension division on an enterprise, and obtaining and quantifying enterprise features; calculating subjective weight values of the indexes by adopting an analytic hierarchy process; calculating objective weight values of the indexes by adopting an entropy evaluation method; determining a subjective weight combination proportionality coefficient and an objective weight combination proportionality coefficient of the index by adopting a variable coefficient method and a Lagrange extreme value method; establishing a combined weight model; constructing a linear weighting and evaluation model according to an evaluation index system and the combined weight model, evaluating the capability of the to-be-analyzed enterprise according to the linear weighting and evaluation model, and respectively establishing GBDT models for evaluation results; inputting the enterprise information and the features into a model; and realizing model fusion on different levels through a Stacking technology, and outputting an enterprise core competitiveness evaluation result. According to the invention, through the Stacking integrated learning thought, the fault-tolerant and anti-disturbance capabilities of the model are enhanced, and the model precision is effectively improved.

The invention discloses a DNA binding protein prediction method based on local evolution information. The method comprises the specific steps: extracting the evolution information of a protein, segmenting the evolution information into local evolution information, and obtaining a feature vector for prediction; sorting the feature vectors according to contribution degrees of the feature vectors tothe model by using an SVMRFE + CBR feature extraction method, and removing irrelevant features; dividing the feature vectors without the irrelevant features into five parts by adopting a five-fold cross validation method, and inputting the four parts as a training set into an SVM model to train the SVM model; and after processing the protein, inputting the feature vector of the protein into an SVMmodel to obtain a prediction result. In the invention, the characteristics of multiple protein sequences are combined, wherein the local evolution information, original evolution information, amino acid composition and dipeptide information of the protein are combined, so that local and overall information of the protein is all contained, and the precision of a DNA binding protein prediction calculation model is improved.

The invention provides an animal identification method, system and device based on tourism images and a storage medium, and the method comprises the steps: carrying out the animal classification of animage through a trained animal classification network, and obtaining the animal information of the image; and the animal information of the image is compared with a preset animal set, operation information corresponding to animal types hit by the animal information is executed, the animal set comprises multiple animal types, and each animal type corresponds to at least one piece of operation information. The method can achieve the automatic recognition of an open scene, reduces the labor cost, improves the efficiency, can recognize the types of animals in an image through the detection and classification, and carries out the corresponding operation.

The invention provides a twin network change detection model based on deep learning, which comprises a double-branch calculation model used for acquiring a difference image, the double-branch calculation model comprises a twin network, a second branch convolution network and an up-sampling convolution network, the twin network is used for respectively extracting time phase feature maps of two time phases, and the up-sampling convolution network is used for carrying out up-sampling on the time phase feature maps of the two time phases; the second branch convolutional network is used for calculating a difference feature map according to the two time-phase feature maps and a difference feature map of the two time-phase feature maps, and the up-sampling convolutional network is used for carrying out up-sampling and / or deconvolution operation on the difference feature map to obtain a difference image. According to the method, the ResNet18 model is transformed to establish a twin network ResAtNet for a change detection scene, the difference feature extraction capability is improved through a double-branch difference feature graph generation method, the model can be suitable for target learning high-dimensional change features, suitable feature expression does not need to be selected by expert knowledge, the method is adaptive to various change scenes, and compared with other existing models, the method has the advantages that the method is simple and convenient to implement, and the method is suitable for large-scale popularization and application. The method has an obvious precision advantage.

Embodiments of the present disclosure provide an initialization method and an initialization system for hybrid modeling, the initialization method includes: generating a plurality of initial points for hybrid modeling; estimating the first effective number of iterations based on model complexity as the selected The number of iterations of the hybrid modeling; the hybrid modeling is performed on the training data set from each of the initial points to generate a corresponding hybrid model; and the corresponding hybrid model with the best accuracy is selected. The initial point is used as the best initialization point.

The invention relates to the field of large generator bar performance simulation analysis, in particular to a multi-factor optimization design method for a motor bar surface electric field. Accordingto the method, a neural network and a genetic algorithm are combined, and design key points of the motor bar can be optimized in a targeted manner according to performance requirements, for example, the conductivity and the length of each section of material of the anti-corona layer of the bar are optimized according to requirements of the bar on electric field homogenization capacity. According to the method, the problems that in the motor bar design process, the motor bar performance influence factors are excessive, and the optimal design scheme is not easy to obtain are solved, the multi-factor optimization capacity is achieved, and the capacity of improving the motor bar design quality is achieved.

The invention discloses a pumping well multi-well working fluid level depth prediction method based on a dynamic and static information feature fusion neural network, and belongs to the field of soft measurement, the method comprising the steps of collecting historical data of a plurality of sucker-rod pumping wells on site, and obtaining factors highly related to the depth of the underground working fluid level; constructing a prediction model structure, and dividing oil well historical operation data into a training set, a verification set and a test set in proportion; and taking Huber loss as a loss function of the neural network, and optimizing parameters of the dynamic and static information feature fusion neural network by adopting a gradient descent method to obtain an optimal value. According to the invention, the multi-well working fluid level depth prediction of the sucker-rod pumping well in different underground environments is realized, the prediction precision is high, and the stability is high.

The invention discloses a GaN HEMT equivalent circuit topological structure based on a novel resistance model. The GaN HEMT equivalent circuit topological structure comprises a transistor GH, whereina grid G, a source S and a drain D of the GH are respectively connected with one ends of inductors L1, L3 and L2; the other end of the L1 is connected with one end of a capacitor C1, one end of a resistor R1 and one end of a capacitor C2; the other end of the resistor R1 is respectively connected with one end of a capacitor C3 and one end of a capacitor C4; the other end of the C3 is connected with one end of a resistor R2; the other end of the capacitor C4 is connected with one end of a resistor R3; the other end of the resistor R3 is respectively connected with one end of a current source Ids, a capacitor C5, a resistor R5 and one end of a resistor R4; and the other end of the resistor R2 is respectively connected with one end of the resistor RS, the other end of the current source Ids,the other end of the C5 and the other end of the R5. According to the GaN HEMT equivalent circuit topological structure, a novel parameter Rs model of the resistor RS is adopted, so a problem that theparameter Rs of the resistor RS changes along with the change of drain-source current and temperature can be solved.