47results about How to "Reduce Model Complexity" patented technology

The invention relates to the technical field of computer application engineering, in particular to a method for using an improved neural network model based on particle swarm optimization for data prediction. The method includes the steps of firstly, expressing data samples; secondly, pre-processing data; thirdly, initiating the parameters of an RBF neural network; fourthly, using the binary particle swarm optimization to determine the number of neurons of a hidden layer and the center of the radial basis function of the hidden layer; fifthly, initiating the parameters of the local particle swarm optimization. By the method for using the improved neural network model based on particle swarm optimization for data prediction, the number of the neurons of the hidden layer of the RBF neural network model can be determined easily, RBF neural network performance is improved, and data prediction accuracy is increased. In addition, the improved neural network model based on particle swarm optimization is low in model complexity, high in robustness and good in expandability.

The invention discloses an EMD based fiber-optic gyroscope temperature drift multi-scale extreme learning machine training method. The EMD based fiber-optic gyroscope temperature drift multi-scale extreme learning machine training method comprise the following steps of (1) adopting a bounded ensemble empirical mode decomposition (EEMD) method to respectively decompose drifting output data of a fiber-optic gyroscope in different temperature-changing-rate environments into a series of intrinsic mode functions; (2) adopting a sample entropy (SE) measurement theory to calculate SE values of the intrinsic mode functions (IMF) in the step (1); (3) determining an IMF set led by noise and an IMF set having different self-similarity features according to the fluctuation trend and sizes of the SE values; (4) superposing the IMF sets determined in the step (3) and having the similar self-similarity features to serve as ELM model training inputs, using temperature gradients at the temperature change rates corresponding to the group of output data as another input training ELM model, similarly, using different superposed self-similarity IMF and corresponding temperature gradients to generate different ELM models through training; (5) accumulating the multiple ELM models generated in the step (4) to obtain a final integrated multi-scale model.

The invention discloses a hybrid-system modeling and simulation platform and method based on a HSTPN model. The hybrid-system modeling and simulation platform comprises a human-computer interaction module, a logic expression module, a simulation analysis module, a file storage module and an equivalent rule library. The platform can carry out modeling and simulation on a hybrid system with five hybrid characteristics of dispersing, continuation, delaying, random and decision at the same time, the hybrid characteristics are only defined in a five-type library, the modeling complexity can be reduced, the transitional moment theory is guaranteed, and the modeling expandability is improved; a hybrid state of the HSTPN model is jointly described through a library identification and the continuous state; the platform provides the equivalent function of the HSTPN model by establishing equivalent rules of a hybrid Petri net and a derivative model of the hybrid Petri net. The hybrid-system modeling and simulation platform and method can be applied to modeling and simulation of hybrid systems such as an information physical fusion system, a flexible manufacturing system, a logistic system and defense industry.

The invention relates to a risk grade evaluation method based on prison prisoner effective influence factors and an implementation system thereof. The method comprises the following steps: (1) acquiring prisoner feature information and performing data mining; (2) training an RF classification model; (3) querying: generating an evaluation level, namely a criminal danger level label, of each prisoner, and storing the evaluation levels into a database; and (5) performing extremely high risk early warning: displaying the feature information and the risk grade score corresponding to the sentences with the highest risk grade score of the first 5%, and sending out an early warning for police officers to check the states of the dangerous persons and take corresponding control measures for the dangerous situations. The risk levels of the prisoners are evaluated, evaluation feedback of police officers is received in real time, the RF classification model is continuously updated and optimized, and the evaluation precision and effectiveness are improved.

The invention discloses a method and a system for predicting an instantaneous value of airport noise based on time series analysis. In the method, an analysis research is carried out in allusion to the instantaneous value of the airport noise, the characteristics of the airport noise are explored from a time series angle and then, predication model establishment and predication are performed in sequence. The system comprises a noise acquisition module, a data processing module, a storage module and a computer processing module, wherein in the noise acquisition module, noise information acquired by a sound sensor is amplified by an amplification circuit and then is input into the data processing module via an analog-to-digital conversion module, input into a noise information database module in the computer processing module via the storage module, and input into a predication model module after being processed by the input module, and the predication model module processes the noise information to obtain prediction data. The method disclosed by the invention has the advantages of developing new concepts and research fields related to airport noise predication, enhancing the learning capability and generalization capability of the models while reducing the modeling complexity, and greatly improving prediction precision.

The invention discloses a ship flow prediction method based on VMD-FOA-GRNN. The method comprises the following steps: 1, preprocessing ship flow data; 2, performing mutation inspection on the preprocessed ship flow data, and selecting non-mutated ship flow data; 3, performing VMD on the ship flow data which are not mutated, generating a series of IMFs with different frequency scales, and obtaining the decomposed ship flow data; 4, constructing a GRNN based on the FOA, and predicting the decomposed ship flow data to obtain a predicted value; and 5, based on the taste concentration judgment function, performing error analysis on the predicted value and the true value to obtain an average absolute percentage, and completing prediction of the ship flow data. According to the method, the problems that an existing prediction method is not high in prediction precision and does not have universal applicability are solved, the prediction precision of the ship flow is improved based on the generalized regression neural network of variational mode decomposition and fruit fly optimization, the problem of universal applicability of time sequence prediction of complex nonlinear time is solved,and the stability is improved.

The invention discloses a video sequence expression recognition system and method based on a self-attention enhanced CNN. The system comprises a feature enhancement CNN module, a self-attention mechanism module and a full connection layer. A video sequence is input into a feature enhancement CNN module, feature vectors output by the feature enhancement CNN module are input into the self-attentionmechanism module, feature vectors output by the self-attention mechanism module are input into the full connection layer, and the full connection layer maps the feature vectors into a sample marking space to realize classification; the feature enhancement CNN module adds a plurality of convolution layers in a backbone network, leads out a feature enhancement branch from a middle layer of the backbone network, fuses the output of the feature enhancement branch with the output of the backbone network, and replaces a full connection layer in the network with a global flat pooling layer. The system provided by the invention is lower in complexity, can effectively improve the accuracy of video sequence expression recognition, and has a wide application prospect in the fields of human-computer interaction, wisdom education, patient monitoring and the like.

The invention discloses an emotion recognition method, system and device based on multi-modal feature fusion and a medium, and the method comprises the steps: obtaining preset first voice information and corresponding first visual information, and carrying out the feature extraction of the first voice information and the first visual information, and obtaining a voice feature image and an expression feature image; performing feature fusion on the voice feature image and the expression feature image to obtain a first multi-modal feature, and constructing a training data set according to the first multi-modal feature; inputting the training data set into a pre-constructed convolutional neural network for training to obtain a trained multi-modal feature recognition model; and identifying the emotion of the person to be tested according to the multi-modal feature identification model. On one hand, the model complexity is reduced, the model training and emotion recognition efficiency is improved, on the other hand, the influence of the voice features and the expression features on the emotion recognition result of the model is considered, the emotion recognition accuracy is improved, and the emotion recognition method can be widely applied to the technical field of emotion recognition.

Embodiments of the invention disclose a power distribution network overload risk assessment method and apparatus. The method comprises the steps of obtaining historical change ranges of a running parameter and an environment parameter of a target primary device in a power distribution network; obtaining running data and environment data of the target primary device; and according to the historicalchange ranges of the running parameter and the environment parameter of the target primary device and the running data and the environment data of the target primary device, judging whether the powerdistribution network is overloaded or not. An assessment model is built based on the historical change ranges of the running parameter and the environment parameter of the target primary device of the power distribution network; the power distribution network is assessed based on the assessment model and the running data and the environment data, collected in real time, of the target primary device; and compared with the prior art, the assessment method and apparatus has the advantages of lowering modeling complexity and improving assessment efficiency.

The invention discloses a novel multi-head attention mechanism. According to a method employing local attention, compared with a global attention method adopted by a traditional multi-head attention mechanism, the complexity of the model is reduced, the sizes of all matrixes in the operation process are only in direct proportion to the length of the sequence, and compared with a matrix in direct proportion to the square of the sequence in the traditional attention mechanism, the storage space consumption of the model is reduced to a large extent. In calculation process, compared with a solution in Transformer-XL, blocking processing is not carried out on the sequence, the sequence characteristics of the original sequence are reserved to a great extent, softmax is used for establishing global semantics, and compared with a cross-block connection mode used in Transformer-XL, the parallelism degree of the model is improved.

The invention relates to a prisoner dangerous behavior prediction method and system based on effective influence factors. The method comprises the following steps: (1) structured processing: 1) data cleaning, and 2) extracting of effective influence factors; (2) dangerous behavior prediction based on weight scores: firstly, classifying all to-be-assessed personnel, and secondly, in combination with prison administration business, managers often pay attention to a small part of groups with extremely high dangerous behaviors and dangerous behavior prediction and early warning based on weight scores; and (3) carrying out online optimization on the Random Forest model. According to the method, analysis and screening are carried out from two aspects of data correlation of evaluation personnel and information entropy provided by fields, and important factors for dangerous behaviors of the prisoners are mined. The method and the system can reduce the calculation complexity and model complexity, and improves the evaluation and prediction accuracy and effectiveness.

The invention discloses a mean influence value data transformation-based k-nearest neighbor fault diagnosis method. The method includes the following steps that: S1, a data set X is collected; S2, standardization processing is performed on the data set X; S3, a BP (Back Propagation) neural network is constructed; S4, the mean influence value (MIV) of the data set is calculated; S5, a weighted dataset X' is calculated; and S6, the weighted data set X' is inputted into a k-nearest neighbor classifier for fault diagnosis, so that a fault result is obtained. According to the mean influence valuedata transformation-based k-nearest neighbor fault diagnosis method of the present invention, the standardized data are processed by the BP neural network, so that the mean influence value (MIV) of data change can be obtained; the MIV can reflect the change condition of the weight matrix of the BP neural network and is the best index for evaluating the correlation of the input parameters of the BPneural network, and therefore, the MIV can determine the weight of the influence of the input neurons of the neural network on the output neurons; and the inputted data set is processed according tothe MIV, so that effective features in the data set can be highlighted, and therefore, the dimensions of the data can be reduced, and correlation between the inputted data set and the output can be enhanced.

The invention relates to a portrait key point detection method based on feature fusion, and the method comprises the steps: S1, sending a portrait image into a face detection network for face detection and cutting, and converting coordinate information in a training data set into thermodynamic diagram information; s2, the portrait image is sent to a regression network based on Transform and Convotion feature fusion to be trained, the regression network is of a parallel structure, low-level semantic features of the portrait image are captured through Convotion, high-level semantic features in the portrait image are captured through Transform, obtained feature maps are subjected to jump connection, and a thermodynamic diagram containing coordinate information is jointly coded; s3, combining the N thermodynamic diagrams of the N key points in the same channel on the basis of a Convoltion and Transform feature fusion regression network, generating a thermodynamic diagram with boundary information, and outputting the thermodynamic diagrams of N + 1 channels; and S4, decoding the first N thermodynamic diagrams of the output thermodynamic diagrams to obtain accurate coordinate information of N key points. The method and the system are favorable for improving the detection precision and the operation speed.

The invention discloses a template-based quick parametric modeling method, which comprises the following steps of: 1) analyzing a model composition relationship which comprises basic composition elements and characteristics of a model; 2) generating template data according to the parameterized model, wherein the template data comprises various replacement items; 3) defining names of various replacement items and storing a template; 4) in the model editor, carrying out self-defined assignment on various replacement items of the template; 5) emptying the sub-models of all the current scenes by the model editor, and applying the sub-models and expressions in the template; and 6) displaying the 3D model generated through the template in the three-dimensional view. According to the method, theparameterized template is created through an existing model, modeling personnel can complete modeling of the complex model by filling in template replacement items, the modeling complexity is greatlyreduced, the modeling time of the complex model such as the cabinet body can be shortened to be within 3 minutes from 30 minutes, and the time cost of the modeling personnel is greatly saved.

The invention relates to a voice synthesis method and device, a readable storage medium and electronic equipment. The method comprises the steps of determining acoustic feature information corresponding to a to-be-synthesized text; obtaining audio information corresponding to the to-be-synthesized text through a first vocoder according to the acoustic characteristic information, wherein the firstvocoder is obtained by carrying out knowledge distillation on a second vocoder, the first vocoder and the second vocoder are vocoders based on a neural network model, and the model complexity of the first vocoder is lower than that of the second vocoder. Thus, the first vocoder learns the excellent data processing capability of the second vocoder, has a simple model structure, and has two advantages of accuracy and speed. Thus, speech synthesis is performed on the to-be-synthesized text based on the first vocoder, and a speech synthesis result can be quickly obtained on the basis of ensuring speech synthesis accuracy.

The invention discloses a distributed calculation and deep learning-based electrocardio beat classification method. The method comprises the following steps of: firstly obtaining an electrocardio beatsignal, dividing a sample set, and carrying out local regionalization on electrocardio data in a training set; constructing a distributed deep learning field, carrying out training by utilizing dataof the training set, and realizing data parallelization by adoption of a software synchronization method in the training; and finally classifying electrocardio data of a test set by utilizing the trained deep learning field. By utilizing the method, potential information in data can be discovered, so that the problem that sign description in traditional electrocardio beat classification process iseasy to cause deviations and wrong classification is easy to occur when electrocardio data features are not obvious are solved, and the problem that single-machine training consume too much time is solved; and the method can be applied to the classification of mass EGG data and remarkably improve the calculation efficiency.

The invention discloses a data classification method based on depth-width variable multi-kernel learning. The method comprises the steps of 1, preparing a data set; step 2, providing an algorithm structure of data set classification; 3, carrying out the first classification of the data of the DWS-MKL algorithm in the step 2 through employing an SVM as a classifier; 4, after the data in the step 3is classified for the first time, performing kernel parameter learning; 6, performing data training by utilizing the steps; and 7, processing test set data by using the classification model obtained by training in the step 6, and obtaining the classification accuracy of the algorithm. According to the invention, the nonlinear mapping capability of the kernel method is brought into full play, the structure is flexibly changed according to the data, and the parameters are optimized by using the level-one-out error boundary, so that the classification accuracy of the method is improved.

The invention discloses a pulmonary nodule intelligent diagnosis method based on mixed characteristics, and the method comprises the steps: obtaining CT image data which comprises a chest medical image file and a corresponding diagnosis result lesion mark; carrying out resampling, smoothing processing and normalization processing on the acquired CT image data; learning a 3D CT image by using a 3Dresidual-tight connection network to obtain high-dimensional depth features, and obtaining LBP-based texture features and HOG-based shape features for describing characterization features of pulmonarynodules; and classifying benign and malignant pulmonary nodules by adopting a GBM gradient elevator based on the high-dimensional depth features, the LBP-based texture features and the HOG-based shape features. According to the method, benign and malignant pulmonary nodules can be quickly distinguished in CT medical images, the misjudgment rate is reduced, the classification precision is improved, and the generalization ability of a diagnosis network model is enhanced.

The invention discloses a wireless sensor network interference model estimating method based on Sigmoid function; the method is characterized in that an interference model is constructed by introducing stretch coefficient and translation coefficient based on standard Sigmoid function model, and each node initializes the stretch coefficient and translation coefficient of the interference model according to an experience value; sample value of the interference model is collected continuously during normal data transmission; the interference model is adjusted in real time by using linear regression estimation model parameters according to the sample value. The wireless sensor network interference model is simulated by introducing stretch coefficient and translation coefficient and using Sigmoid function as a prototype, calculating complex Gaussian error function is avoided, the model precision is adjusted by means of parameter estimation, and modeling complexity is decreased; the stretch coefficient and translation coefficient in the model are estimated by collecting samples in a network during normal data transmission, and modeling expense is decreased.

The invention discloses a traffic target detection method and system based on improved YOLOv4, the detection system comprises a MobileViT-S backbone network, an SPP feature pyramid network, a PANet feature enhancement network and a target detection head, and convolution used in the PANet feature enhancement network and the target detection head is deep separable convolution. The method can be used for detecting pedestrians, vehicles and traffic light targets in an intelligent traffic scene, high detection accuracy is achieved on the basis of light weight, the omission ratio is low, and the detection effect is good.

The invention discloses a roof modeling method and device. The method comprises: preprocessing an input outer wall contour line, dispersing an arc in the outer wall contour line, calculating and generating a straight skeleton of the roof according to the preprocessed and discretized polygons, calculating a Z coordinate of a straight skeleton node, generating a wireframe and roof polyhedron model based on the generated straight skeleton and coordinates thereof, generating a conical surface when the input outer wall contour line contains an arc, generating a roof lower surface model, and generating eaves. The situations of arcs, gable walls and different slope angles are processed through a reinforced straight skeleton algorithm, the roof model can be efficiently and accurately generated, and a roof modeling mode which is stable, reliable and capable of meeting different requirements of users can be achieved.

The invention discloses a real image denoising method based on a pseudo 3D self-correlation network, and the method comprises the steps: constructing a pseudo 3D self-correlation module P3AB based on one-dimensional fast convolution so as to extract the self-correlation features of elements at each position of an input feature map in horizontal, vertical and channel directions through the one-dimensional fast convolution, and after the traversal of all positions is completed, respectively obtaining pseudo 3D self-correlation features in three directions; performing channel cascading and adaptive feature fusion on the pseudo 3D self-correlation features in the three directions to obtain global self-correlation features including spatial domain self-correlation information and channel domain self-correlation information; adding the global self-correlation feature and the input feature map through residual connection to serve as the output of the P3AB; constructing a pseudo 3D autocorrelation network P3AN, the P3AN comprises a shallow feature extraction unit, a stacked P3AB and a tail convolutional layer, and is provided with two layers of jump connection; training the P3AN; and de-noising an input real noise image by using the trained P3AN, and outputting a de-noised image.

The method for analyzing composite material panels based on the structural genome technology of the present invention includes: extracting a representative structural genome from the structural form of the composite panel, establishing a geometric model based on the structural genome, and performing a network on the geometric model. Lattice discretization; using the mechanical properties of the matrix phase and reinforcement phase of the internal structure of the structural genome to obtain the material properties of the structural genome through homogenization theoretical analysis; according to the macrostructure of the composite material siding to establish an analysis model of the composite siding; The material properties of the siding are given an analytical model for the analysis of composite siding. Based on the structural genome technology, the invention establishes the structural genome of the wall plate structure from the periodicity of the composite material wall plate structure, reduces the complexity of modeling the composite material wall plate structure, and greatly shortens the time period on the basis of ensuring the accuracy of the analysis results. Pre-analytical processing time and analysis time for composite siding.

The invention discloses a power station equipment fault monitoring model generation method, system and device, and the method comprises the steps: presetting the maximum capacity of an initial training sample set, and carrying out the standardization processing of a large amount of historical normal data stored in a database based on the maximum capacity, and obtaining a training sample set containing training data; determining an optimal cluster number by calculating and comparing Calinski-Harabasz score values under training data under different clusters, taking the determined optimal cluster number as a cluster number of a K-means algorithm, and performing algorithm training on the training data to determine a clustering center set of the data; and taking the determined clustering center set as an initial clustering center of an FCM algorithm, performing algorithm training on training data to determine a cluster membership degree of the data, establishing an FCM model, classifying the training data according to the maximum membership degree, establishing corresponding PCA models according to different categories, and completing a training process.

The invention discloses a teaching quality evaluation prediction method based on PSO optimized RBF model. A teacher teaching quality evaluation model based on a PSO optimized RBF neural network modelovercomes difficulty in a traditional teacher teaching quality evaluation process. Compared with other evaluation methods, the teaching quality evaluation prediction method has advantages of making anevaluation simulating result and precision accord with reality, effectively reducing influence of artificial factors in index weight determining, and promoting enthusiasm of teachers. The teaching quality evaluation prediction method based on the particle swarm optimization algorithm improved neural network model can dynamically optimize a central point position and a weight matrix of an RBF neural network model hidden layer, thereby improving performance of the RBF neural network, and improving data prediction accuracy. Furthermore a least square method is used for calculating a weight matrix, thereby reducing small number of required points and high convergence speed.

The invention discloses a blast furnace gas cabinet position prediction method based on multi-factor analysis. The blast furnace gas cabinet position prediction method comprises the following steps: 1, obtaining blast furnace gas flow and historical data related to a gas cabinet position in a database; 2, performing data preprocessing on the historical data; 3, carrying out multi-factor analysis on cabinet position fluctuation, and determining main influence factors of the blast furnace gas cabinet position fluctuation by utilizing an absolute flow ratio and grey correlation degree analysis method; and 4, constructing a blast furnace gas cabinet position prediction model based on the influence factors analyzed in the step 3, and realizing future multi-step prediction of the blast furnace gas cabinet position. According to the method, the absolute flow proportion analysis method and the grey correlation degree analysis method are combined to determine the main influence factors of the cabinet level fluctuation, model input is simplified, and modeling complexity is reduced; and the characteristic that the fluctuation of the cabinet position is influenced by multiple factors is fully considered, so that the accurate prediction of future multi-step numerical values of the cabinet position can be realized when the production scene changes and the production consumption fluctuates, and reference and basis are provided for the reasonable scheduling of a blast furnace gas system.

The invention discloses an acacia honey authenticity identification method based on feature selection and a machine learning algorithm. The acacia honey authenticity identification method comprises the following steps: collecting true and false honey samples and generating acacia honey data; performing true and false labeling on the acacia honey data to obtain an acacia honey data set; obtaining a low-dimensional acacia honey data set through feature selection; constructing a honey true and false identification model RF-XGBoost; performing parameter optimization and model verification on the model; and carrying out authenticity identification on to-be-detected honey by utilizing the trained model. According to the method, the authenticity of the black locust honey can be effectively and accurately identified, errors caused by manual checking of a spectrogram for authenticity identification are avoided, the accuracy, the root mean square error and the AUC value of the authenticity identification of the black locust honey are effectively improved, the data feature dimension, the model training time, the model complexity and the over-fitting risk are reduced, and the method is an effective method for identifying authenticity of acacia honey.