CN116910699A - A dual-frequency index prediction method and system based on generative adversarial network - Google Patents
A dual-frequency index prediction method and system based on generative adversarial network Download PDFInfo
- Publication number
- CN116910699A CN116910699A CN202310942061.9A CN202310942061A CN116910699A CN 116910699 A CN116910699 A CN 116910699A CN 202310942061 A CN202310942061 A CN 202310942061A CN 116910699 A CN116910699 A CN 116910699A
- Authority
- CN
- China
- Prior art keywords
- data
- training
- model
- regression model
- frequency
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/27—Regression, e.g. linear or logistic regression
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The application discloses a dual-frequency index prediction method and a system based on a generated countermeasure network, comprising the steps of obtaining a dual-frequency index database; the dual-frequency index database comprises a training data set for training the model and a test data set for testing the model; based on the ACGAN structure, carrying out structural improvement on the ACGAN structure so as to establish and obtain a regression model; training the regression model based on the training data set to obtain a trained regression model; calculating a test data set based on the trained regression model to obtain error data; evaluating the error data based on built-in regression evaluation standard data to judge the judging result of the trained regression model on the double-frequency index value; and if the error data is smaller than the regression evaluation standard data, judging the network depth of the regression model to determine the optimal prediction model. The method has the effect of improving the accuracy of the double-frequency index prediction.
Description
Technical Field
The application relates to the field of double-frequency index prediction, in particular to a double-frequency index prediction method and a double-frequency index prediction system based on a generated countermeasure network.
Background
The dual-frequency index is one of reference indexes for monitoring the brain at different sedation levels, and the electroencephalogram is one of important information sources for researching brain activities, and the electroencephalogram dual-frequency index can be used for monitoring the anesthesia depth, however, the sealing performance and the complexity of the traditional dual-frequency index calculation cause a certain limit on the electroencephalogram monitoring level, and the traditional dual-frequency index calculation needs to be improved.
In the related art, an original signal is converted into a time-frequency image through short-time-interval Fourier transform, and then images of two channels are combined to be used as input to train a neural network so as to convert an electroencephalogram signal time domain signal into a time-frequency image, so that the time-frequency image is analyzed.
For the related art described above, although the time domain signal is converted into the time-frequency image by using the frequency characteristic of the time domain brain wave signal, the two-dimensional representation of the time-frequency image will lose the time characteristic and the spatial characteristic, so that there is an error in the brain wave identification judgment, and there is an improvement.
Disclosure of Invention
In order to improve the accuracy of dual-frequency index prediction, the application provides a dual-frequency index prediction method and a system based on a generated countermeasure network.
In a first aspect, the present application provides a dual-frequency index prediction method based on generation of an countermeasure network, which adopts the following technical scheme:
a dual-frequency exponential prediction method based on generating an antagonism network, comprising:
acquiring a double-frequency index database; the dual-frequency index database comprises a training data set for training the model and a test data set for testing the model;
based on the ACGAN structure, carrying out structural improvement on the ACGAN structure so as to establish and obtain a regression model;
training the regression model based on the training data set to obtain a trained regression model;
calculating a test data set based on the trained regression model to obtain error data;
evaluating the error data based on built-in regression evaluation standard data to judge the judging result of the trained regression model on the double-frequency index value;
and if the error data is smaller than the regression evaluation standard data, judging the network depth of the regression model to determine the optimal prediction model.
According to the technical scheme, the data in the double-frequency index database is divided into two parts by acquiring the double-frequency index database as a reference basis for model establishment, a training data set for training the established model and a test data set for testing the model. The structure of the ACGAN model is improved, so that the improved model can be used for carrying out more fitting on the prediction of the double-frequency index, a regression model is obtained, meanwhile, the established regression model is trained according to a training data set to improve the identification performance of the regression model on the electroencephalogram signal, the trained regression model is subjected to test data set detection to judge the prediction effect of the regression model, error data is output, the regression model prediction result evaluation is carried out according to the error data and built-in regression evaluation standard data to judge whether the trained regression model can carry out accurate prediction on the double-frequency index value, if the error data is smaller than the regression evaluation standard data, the regression model is judged to be capable of carrying out prediction on the double-frequency index value, the network depth of the regression model is judged again to determine the optimal network depth of the regression model, the optimal prediction model is obtained, and the error data for the prediction of the double-frequency index value is reduced by calculating different network depths of the regression model, and the prediction accuracy is improved. .
Preferably, the regressor replaces an auxiliary classifier in the original ACGAN structure to process and regress the tagged data;
the regressor model comprises two generators, two discriminators and two regressors in the time domain or frequency domain simultaneous training; the regressor is the same as other network layers of the arbiter, and the activation functions on the full connection layer are different;
taking the Sigmoid function as a classification function of an output layer in the discriminator; the Tanh function is used as an output layer activation function of the generator; the Relu function acts as the output layer function of the regressor.
Through the technical scheme, the original ACGAN structure is improved, so that the improved structure type is more fit with the prediction of the double-frequency index value. The regressor replaces an auxiliary classifier in the original structure so as to process the labeled data and perform regression operation on the processed data to determine the double-frequency index value. Meanwhile, the actual digital range can be mapped to the (0, 1) interval by adopting the Sigmoid function so as to be convenient for classification operation, namely judging the authenticity of the data, and the convergence rate of the data can be increased by adopting the Tanh function so as to reduce the iteration times, thereby reducing the calculated amount and improving the operation efficiency; all negative values can be changed into 0 by adopting the Relu function, and the positive values are unchanged, so that the calculated amount is reduced, the operation efficiency is further improved, and meanwhile, the convergence of the regression model is more stable.
Preferably, acquiring electroencephalogram data and preprocessing the electroencephalogram data to obtain time domain data of the electroencephalogram;
performing Fourier transform on the time domain data according to the time domain data to obtain corresponding frequency domain data;
transmitting the time domain data and the frequency domain data to a regression model to obtain time domain characteristic data and frequency domain characteristic data;
directly splicing the time domain characteristic data and the frequency domain characteristic data by adopting a convolutional neural network to obtain corresponding double-frequency index values;
the directly splicing the time domain feature data and the frequency domain feature data specifically comprises the following steps:
the method comprises the steps of performing the same dimension setting on time domain data and frequency domain data input by a generator and the same structural parameter setting to obtain time domain feature data and frequency domain feature data with the same dimension;
and directly splicing the time domain characteristic data and the frequency domain characteristic data with the same dimension.
According to the technical scheme, in the process of establishing the regression model, according to the change of the ACGAN structure, the algorithm is confirmed, the time domain data corresponding to the electroencephalogram data are judged through analysis processing of the electroencephalogram data, then Fourier transformation is carried out on the time domain data to obtain the corresponding frequency domain data, feature extraction is carried out on the time domain data and the frequency domain data according to the regression model to obtain time domain feature data and frequency domain feature data, and then the time domain feature data and the frequency domain feature data are spliced through a convolutional neural network to obtain the corresponding double-frequency index value. The splicing process is to set the same dimension of input time domain data and frequency domain data of the generator, set the same structural parameters of the generator for analyzing time domain data and the generator for analyzing frequency domain data, so that the structural dimensions of the time domain feature data and the frequency domain feature data are the same, different feature data can be spliced directly, meanwhile, the regression result in the same domain is more visual through splicing the feature data in different domains, and the result in the double domain and the result in the single domain can be compared and analyzed, so that the prediction on the double-frequency network can be clearly judged, and the judging accuracy is improved.
Preferably, different training strategies and nonlinear activation functions are adopted to establish a regression model;
participating the loss of regressors in generating countermeasure training based on training patterns similar to ACGAN structure;
the relu function is used as an activation function of a convolution layer, and rated quantity normalization processing is carried out on the data.
Through the technical scheme, the regression model is built by adopting different training strategies and nonlinear activation functions, so that the regression model can cope with various different situations, the identification performance of the regression model on the electroencephalogram signals is improved, and the identification accuracy is improved. Meanwhile, the loss of the regressor is also involved in the generation countermeasure training, so that the average error of the regression model is reduced, the accuracy is improved, the stability of convergence of the model is improved by taking the Relu function as an activation function of a convolution layer, and meanwhile, the rated quantity normalization processing is carried out on the data, so that the data is prevented from being subjected to transition fitting, the integrity of the data is improved, and the data obtained through analysis can be more accurate.
Preferably, a time domain training data set and a preset training frequency are input;
constructing a GAN generator neural network of a time domain and a frequency domain, and constructing a discriminator neural network and a regression neural network by combining the time domain and the frequency domain;
Performing Fourier transform on the time domain training data set to obtain a frequency domain training data set;
training the regression model, stopping training when the training times are larger than the preset training times, and outputting a training result model.
According to the technical scheme, the data is built in the regression model, the time domain training data set is used for analyzing time domain data to obtain frequency domain data, the preset training times are input to determine whether the training of the regression model reaches the designated times or not so as to ensure that the training result can reach the designated effect, meanwhile, the GAN generator neural network of the time domain and the frequency domain, the time domain and the frequency domain are combined to construct the discriminator neural network and the regression neural network, the time domain training data set is subjected to Fourier transformation to convert the time domain training data set into the frequency domain training data set, finally, the time domain training data set and the converted frequency domain training data set are input into the regression model, and meanwhile, the training times of the regression model are judged so that the training of the regression model reaches the designated times, the accuracy and the reference of the test result in detection of the regression model are guaranteed, the test error is reduced, and the reliability of the output result is improved.
Preferably, training the arbiter neural network and the regression neural network;
inputting rated quantity and optimizer super parameters, and obtaining time domain training data and frequency domain training data of batch size quantity;
according to the GAN generator neural network of the time domain and the GAN generator neural network of the frequency domain, generating rated amount of time domain noise and frequency domain noise correspondingly;
and setting an Adam optimizer based on the optimizer super parameters to train and update the neural network model parameters of the time domain and frequency domain discriminator neural network and the regression neural network so as to obtain model loss data.
Preferably, the GAN producer is trained;
respectively inputting the time domain noise, the frequency domain noise and the model loss data of the discriminator and the regressor into a GAN generator neural network of a time domain and a GAN generator neural network of a frequency domain;
the GAN generator model is trained and updated according to Adam optimizers.
According to the technical scheme, in the training process of the GAN generator, the time domain noise, the frequency domain noise and the model loss data of the regressor are input into the GAN generator neural network of the time domain and the GAN generator neural network of the frequency domain, and the Adam optimizer is set according to the super-parameters of the optimizer, and the GAN generator model is trained and updated. In the training process, the GAN generator is enabled to generate noise more accurately and effectively, and meanwhile, the model loss data error is enabled to be smaller through feedback to the discriminator upgrading network and the regressor neural network, so that the dual-frequency index value judged by the regressor neural network is enabled to be more accurate.
Preferably, when the error data is smaller than the regression evaluation standard data, the trained regression model is judged to be capable of judging the double-frequency index value and judging the depth of the regression model so as to determine the optimal prediction model;
performing error data calculation of different network depths on the regression model to obtain corresponding depth error data;
comparing the depth error data to obtain difference data between the depth error data of different depths, and comparing the difference data;
if the difference data is smaller than the preset difference, the network depth corresponding to the corresponding difference data is read to obtain the optimal network depth, and the optimal network depth is used as the optimal network depth of the regression model to obtain the optimal prediction model.
According to the technical scheme, when the obtained error data is smaller than the set regression evaluation standard data, the trained regression model is indicated to be capable of predicting the double-frequency index value, and prediction accuracy is required to be improved, so that the network depth of the regression model is required to be judged to reduce the size of the error data, and accordingly prediction accuracy and accuracy are improved. However, the deeper the network depth, the smaller the corresponding error data, and the more data complexity and data parameters the higher the device requirements, so the network depth needs to be selected.
And performing error data calculation on different network depths to obtain error data of corresponding depths, namely depth error data, comparing the depth error data of two adjacent depths to obtain difference data between the depth error data of different depths, and comparing the difference data, if the difference data is smaller than a preset difference value, judging that the accuracy of the dual-frequency index value prediction corresponding to the network depth corresponding to the difference data meets the requirement, and correspondingly selecting the depth data of the two network depths corresponding to the difference data if the data complexity and the equipment requirement are lower, and taking the network depth corresponding to the smaller depth in the two network depth data as the network depth of the regression model, namely increasing the accuracy of the regression model on the dual-frequency index value prediction, and further enabling the complexity of the equipment and the data to be smaller and increasing the feasibility of the dual-frequency index value prediction.
In a second aspect, the present application provides a dual-frequency index prediction system based on generation of an countermeasure network, which adopts the following technical scheme:
a dual-frequency exponential prediction system based on generating an antagonism network, comprising:
the data acquisition module is used for acquiring a double-frequency index database; the dual-frequency index database comprises a training data set for training the model and a test data set for testing the model;
The regression model building module is used for obtaining the ACGAN structure and improving the ACGAN structure to build a regression model;
the regression model training module is configured to be in signal connection with the data acquisition module and the regression model establishment module and is used for receiving the training data set and the regression model, inputting the training data set into the regression model and training the regression model to obtain a trained training result model;
the model test module is internally provided with regression evaluation standard data and preset difference values, is configured to be in signal connection with the regression model training module and the data acquisition module and is used for receiving the training result model and the test data set, and inputting the test data set into the training result model to obtain error data; comparing the error data with regression evaluation standard data, and if the error data is in the regression evaluation standard data, judging that the training result model can predict the double-frequency index value; and meanwhile, carrying out error data calculation on the training result model at different network depths to obtain depth error data, carrying out difference calculation on the depth error data at different depths to obtain difference data, comparing the difference data with a preset difference value to judge the optimal network depth, and determining an optimal prediction model according to the optimal network depth.
Preferably, the regression model building module comprises a generator, a discriminator and a regressor, wherein the regressor is the same as other network layers of the discriminator, and the activation functions on the full connection layers are different;
the generator is in signal connection with the data acquisition module and is used for receiving the double-frequency index database and carrying out time domain analysis and frequency domain analysis on data in the double-frequency index database so as to obtain error data; when the frequency domain analysis is carried out on the double-frequency data, the double-frequency data is subjected to fast Fourier transform to obtain corresponding frequency domain data, and the frequency domain analysis is carried out on the frequency domain data;
the discriminator is in signal connection with the data acquisition module and is used for receiving the double-frequency index database, performing fast Fourier transform on the double-frequency index data to obtain frequency domain data, performing data analysis on the time domain data and the frequency domain data to judge whether the data are true or false, and outputting corresponding discrimination data;
the regressor is in signal connection with the generator and the discriminator and is used for receiving the error data and the discrimination data and analyzing the error data and the discrimination data to obtain a regression model which is established correspondingly.
In summary, the present application includes at least one of the following beneficial technical effects:
1. the method comprises the steps of dividing data in a double-frequency index database into two parts by acquiring the double-frequency index database to serve as a reference basis for model establishment, a training data set for training the established model, and a test data set for testing the model. The structure of the ACGAN model is improved, so that the improved model can be used for carrying out more fitting on the prediction of the double-frequency index, a regression model is obtained, meanwhile, the established regression model is trained according to a training data set to improve the identification performance of the regression model on the electroencephalogram signal, the trained regression model is subjected to test data set detection to judge the prediction effect of the regression model, error data is output, the regression model prediction result evaluation is carried out according to the error data and built-in regression evaluation standard data to judge whether the trained regression model can carry out accurate prediction on the double-frequency index value, if the error data is smaller than the regression evaluation standard data, the regression model is judged to be capable of carrying out prediction on the double-frequency index value, the network depth of the regression model is judged again to determine the optimal network depth of the regression model, the optimal prediction model is obtained, and the error data for the prediction of the double-frequency index value is reduced by calculating different network depths of the regression model, and the prediction accuracy is improved. The method comprises the steps of carrying out a first treatment on the surface of the
2. The feasibility of predicting the double-frequency index value by the established regression data is determined by judging the error data, and if the feasibility is feasible, the network depth of the regression model is required to be judged so as to reduce the size of the error data, thereby improving the prediction precision and accuracy. However, the deeper the network depth is, the smaller the corresponding error data is, and the more the corresponding data complexity and data parameters are, the higher the requirement on the equipment is, so that the network depth needs to be selected; performing error data calculation on different network depths to obtain error data of corresponding depths, namely depth error data, comparing the depth error data of two adjacent depths to obtain difference data between the depth error data of different depths, and comparing the difference data, if the difference data is smaller than a preset difference value, judging that the accuracy of dual-frequency index value prediction corresponding to the network depth corresponding to the difference data meets the requirement, and correspondingly selecting the two network depth data corresponding to the difference data if the data complexity and the equipment requirement are lower, and taking the network depth corresponding to the smaller depth in the two network depth data as the network depth of the regression model, namely increasing the accuracy of the regression model on the dual-frequency index value prediction, and further enabling the complexity of the equipment and the complexity of the data to be smaller, thereby increasing the feasibility of the dual-frequency index value prediction;
3. The regression model is built by comprehensively utilizing different training strategies and nonlinear activation functions, so that the regression model can cope with various different situations, the recognition performance of the regression model on the electroencephalogram signals is improved, and the recognition accuracy is improved. Meanwhile, the loss of the regressor is also involved in the generation countermeasure training, so that the average error of the regression model is reduced, the accuracy is improved, the stability of convergence of the model is improved by taking the Relu function as an activation function of a convolution layer, and meanwhile, the rated quantity normalization processing is carried out on the data, so that the data is prevented from being subjected to transition fitting, the integrity of the data is improved, and the data obtained through analysis can be more accurate.
Drawings
Fig. 1 is a flowchart mainly showing a dual-frequency index prediction method based on generation of an countermeasure network in the present embodiment;
fig. 2 is a flowchart mainly showing steps of performing structural improvement on ACGAN in a dual-frequency index prediction method based on generation of an countermeasure network in the present embodiment;
FIG. 3 is a flowchart mainly showing steps of processing data by a regression model in a dual-frequency index prediction method based on generating an countermeasure network according to the present embodiment;
FIG. 4 is a flowchart mainly showing the steps of the method for predicting the running form of the regression model in the dual-frequency index based on the generation of the countermeasure network according to the present embodiment;
FIG. 5 is a flowchart of the sub-steps of the method for generating a dual-frequency index prediction based on an countermeasure network according to the present embodiment;
FIG. 6 is a flowchart mainly showing the sub-steps of S3400 in the dual-frequency index prediction method based on generating an countermeasure network according to the present embodiment;
FIG. 7 is a flowchart mainly showing the sub-steps of S600 in the dual-frequency index prediction method based on generation of an countermeasure network according to the present embodiment;
fig. 8 is a block diagram of a dual-frequency index prediction system mainly embodying the present embodiment based on generation of an countermeasure network.
Reference numerals: 1. a data acquisition module; 2. a regression model building module; 21. a generator; 22. a discriminator; 23. a regressing device; 3. a regression model training module; 4. and a model test module.
Detailed Description
The application is described in further detail below with reference to fig. 1-8.
The embodiment of the application discloses a dual-frequency index prediction method and a system based on a generated countermeasure network.
Examples: as shown in fig. 1, the dual-frequency index prediction method based on the generation of the countermeasure network of the present application includes:
s100, acquiring a double-frequency index database; the dual-frequency index database comprises a training data set for training the model and a test data set for testing the model;
S200, based on the ACGAN structure, carrying out structural improvement on the ACGAN structure so as to establish and obtain a regression model;
s300, training the regression model based on the training data set to obtain a trained regression model;
s400, calculating a test data set based on the trained regression model to obtain error data;
s500, evaluating error data based on built-in regression evaluation standard data to judge the judgment result of the trained regression model on the double-frequency index value;
and S600, if the error data is smaller than the regression evaluation standard data, judging the network depth of the regression model to determine the optimal prediction model.
In this embodiment, the data in the dual-frequency index database is divided into two parts by acquiring the dual-frequency index database as a reference basis for model establishment, a training data set for training the established model, and a test data set for testing the model. The structure of the ACGAN model is improved, so that the improved model can be used for carrying out more fitting on the prediction of the double-frequency index, a regression model is obtained, meanwhile, the established regression model is trained according to a training data set to improve the identification performance of the regression model on the electroencephalogram signal, the trained regression model is subjected to test data set detection to judge the prediction effect of the regression model, error data is output, the regression model prediction result evaluation is carried out according to the error data and built-in regression evaluation standard data to judge whether the trained regression model can carry out accurate prediction on the double-frequency index value, if the error data is smaller than the regression evaluation standard data, the regression model is judged to be capable of carrying out prediction on the double-frequency index value, the network depth of the regression model is judged again to determine the optimal network depth of the regression model, the optimal prediction model is obtained, and the error data for the prediction of the double-frequency index value is reduced by calculating different network depths of the regression model, and the prediction accuracy is improved.
Referring to fig. 2, in step S200, the ACGAN structure is structurally modified based on the ACGAN structure to build a regression model, which includes the following steps:
s211, replacing an auxiliary classifier in the original ACGAN structure with a regressor 23 to process and regress the labeled data;
s212, the regressor 23 model comprises two generators 21, two discriminators 22 and two regressors 23 in the simultaneous training in the time domain or the frequency domain; the regressor 23 is the same as the other network layers of the arbiter 22, while the activation functions on the fully connected layers are different;
s213, taking the Sigmoid function as a classification function of the output layer in the discriminator 22; the Tanh function acts as an output layer activation function for the generator 21; the Relu function acts as an output layer function of the regressor 23.
In this embodiment, an improvement is performed on the original ACGAN structure, so that the improved structure type is more consistent with the prediction of the dual-frequency index value. The regressor 23 replaces the auxiliary classifier in the original structure to facilitate processing of the tagged data and regression operation of the processed data to determine the double frequency index value. Meanwhile, the actual digital range can be mapped to the (0, 1) interval by adopting the Sigmoid function so as to be convenient for classification operation, namely judging the authenticity of the data, and the convergence rate of the data can be increased by adopting the Tanh function so as to reduce the iteration times, thereby reducing the calculated amount and improving the operation efficiency; all negative values can be changed into 0 by adopting the Relu function, and the positive values are unchanged, so that the calculated amount is reduced, the operation efficiency is further improved, and meanwhile, the convergence of the regression model is more stable.
Referring to fig. 3, in step S200, the ACGAN structure is structurally modified based on the ACGAN structure to build a regression model, which includes the following steps:
s221, acquiring brain electricity data and preprocessing the brain electricity data to obtain time domain data of brain electricity;
s222, carrying out Fourier transform on the time domain data according to the time domain data to obtain corresponding frequency domain data; the brain electrical data is regarded as a whole and fast Fourier transformation is carried out to obtain complete frequency domain data. The data integrity is beneficial to the accurate extraction of the neural network to the characteristics, so that a better regression effect is achieved.
S223, conveying the time domain data and the frequency domain data to a regression model to obtain time domain characteristic data and frequency domain characteristic data;
s224, directly splicing the time domain feature data and the frequency domain feature data by adopting a convolutional neural network to obtain corresponding double-frequency index values;
the directly splicing the time domain feature data and the frequency domain feature data specifically comprises the following steps:
s2231, performing the same dimension setting and the same structural parameter setting on the time domain data and the frequency domain data input by the generator 21 to obtain time domain feature data and frequency domain feature data with the same dimension;
S2232, directly splicing the time domain feature data and the frequency domain feature data with the same dimension.
In this embodiment, in the process of establishing the regression model, according to the change on the ACGAN structure, the algorithm is confirmed at the same time, the time domain data corresponding to the electroencephalogram data is determined by analyzing and processing the electroencephalogram data, then the time domain data is subjected to fourier transform to obtain the corresponding frequency domain data, the time domain data and the frequency domain data are subjected to feature extraction according to the regression model to obtain the time domain feature data and the frequency domain feature data, and then the time domain feature data and the frequency domain feature data are spliced through the convolutional neural network to obtain the corresponding double-frequency index value. The splicing process is to set the same dimension of the input time domain data and the frequency domain data of the generator 21, and set the same structural parameters of the generator 21 for analyzing the time domain data and the generator 21 for analyzing the frequency domain data, so that the structural dimensions of the time domain feature data and the frequency domain feature data are the same, different feature data can be directly spliced, and meanwhile, the regression result in the same domain is more visual by splicing the feature data in different domains, and the result in the double domain and the result in the single domain can be compared and analyzed, so that the prediction on the double-frequency network can be clearly judged, and the accuracy of judgment is further improved.
Referring to fig. 4, in step S200, the ACGAN structure is structurally modified based on the ACGAN structure to build a regression model, which includes the following steps:
s231, building a regression model by adopting different training strategies and nonlinear activation functions;
s232, participating the loss of the regressor 23 into the generation of countermeasure training based on the training mode similar to the ACGAN structure;
s233, using the Relu function as an activation function of a convolution layer, and performing rated number normalization processing on the data.
In the embodiment, the regression model is built by adopting different training strategies and nonlinear activation functions, so that the regression model can cope with various different situations, the identification performance of the regression model on the electroencephalogram signals is improved, and the identification accuracy is improved. Meanwhile, the loss of the regressor 23 is also involved in the generation countermeasure training, so that the average error of the regression model is reduced, the accuracy is improved, the stability of the convergence of the model is improved by taking the Relu function as the activation function of the convolution layer, and meanwhile, the rated quantity normalization processing is carried out on the data, so that the data is prevented from being subjected to transition fitting, the integrity of the data is improved, and the data obtained through analysis can be more accurate.
Referring to fig. 5, in step S300, training a regression model based on a training data set to obtain a trained regression model includes the steps of:
s310, inputting a time domain training data set and preset training times;
s320, constructing a GAN generator 21 neural network of a time domain and a frequency domain, and constructing a discriminator 22 neural network and a regression neural network by combining the time domain and the frequency domain;
s330, carrying out Fourier transform on the time domain training data set to obtain a frequency domain training data set;
s340, training the regression model, stopping training when the training times are greater than the preset training times, and outputting a training result model.
In this embodiment, the data is built in the regression model, the time domain training dataset is used for analyzing the time domain data to obtain the frequency domain data, the preset training times are input to determine whether the training of the regression model reaches the designated times, so as to ensure that the training result can reach the designated effect, meanwhile, the GAN generator 21 neural network of the time domain and the frequency domain, the time domain and the frequency domain are combined to construct the discriminator 22 neural network and the regression neural network, the time domain training dataset is then subjected to fourier transformation to convert the time domain training dataset into the frequency domain training dataset, finally, the time domain training dataset and the converted frequency domain training dataset are input into the regression model, and the training times of the regression model are judged, so that the training of the regression model reaches the designated times, the accuracy and the reference of the test result when the regression model is detected are ensured, the test error is reduced, and the reliability of the output result is improved.
Referring to fig. 6, in step S340, training is performed on the regression model, and when the training frequency is greater than the preset training frequency, the training is stopped, and a training result model is output, including the steps of:
s341, training a discriminator 22 neural network and a regression neural network;
s342, inputting rated quantity and optimizer super parameters, and acquiring time domain training data and frequency domain training data of batch size quantity;
s343, correspondingly generating rated amount of time domain noise and frequency domain noise according to the GAN generator 21 neural network of the time domain and the GAN generator 21 neural network of the frequency domain;
and S344, setting an Adam optimizer based on the super parameters of the optimizer to train and update the neural network model parameters of the arbiter 22 neural network and the regression neural network of the time domain and the frequency domain so as to obtain model loss data.
S345, training a GAN producer;
s346, inputting the time domain noise, the frequency domain noise and the model loss data of the discriminator 22 and the regressor 23 to the GAN generator 21 neural network of the time domain and the GAN generator 21 neural network of the frequency domain respectively;
s347, training and updating the GAN generator 21 model according to the Adam optimizer.
In this embodiment, the neural network of the arbiter 22 and the regression neural network are trained, and the GAN generator 21 is trained, so that the trained regression model is more mature, and the calculation result is more accurate. In the training process of the discriminator 22 neural network and the regression neural network, time domain training data and frequency domain training data of the batch size are obtained, corresponding time domain noise generation and frequency domain noise generation are carried out according to the generator 21 neural network of the time domain and the frequency domain, meanwhile, an Adam optimizer is set according to the super-parameters of the optimizer, and the neural network model parameters of the discriminator 22 neural network and the regression neural network of the time domain and the frequency domain are updated, so that model loss data are obtained.
In training the GAN generator 21, time domain noise and frequency domain noise and model loss data of the regressor 23 are input to the GAN generator 21 neural network of the time domain and the GAN generator 21 neural network of the frequency domain, and Adam optimizer is set according to optimizer super parameters and the GAN generator 21 model is trained and updated. In the training process, the GAN generator 21 can generate more accurate and effective noise, and meanwhile, the noise is fed back to the arbiter 22 upgrading network and the regressor 23 neural network, so that the obtained model loss data error is smaller, and the dual-frequency index value judged by the regressor neural network is more accurate.
Referring to fig. 7, in step S600, if the error data is smaller than the regression evaluation criterion data, the network depth of the regression model is determined to determine the optimal prediction model, which includes the following steps:
s610, when the error data is smaller than the regression evaluation standard data, judging that the trained regression model can judge the double-frequency index value and judge the depth of the regression model so as to determine the optimal prediction model;
s620, performing error data calculation of different network depths on the regression model to obtain corresponding depth error data;
S630, comparing the depth error data to obtain difference data between the depth error data with different depths, and comparing the difference data;
and S640, if the difference data is smaller than the preset difference, reading the network depth corresponding to the corresponding difference data to obtain the optimal network depth, and taking the optimal network depth as the optimal network depth of the regression model to obtain the optimal prediction model.
In this embodiment, when the obtained error data is smaller than the set regression evaluation standard data, it indicates that the trained regression model can predict the dual-frequency index value, and the prediction accuracy needs to be improved, so that the network depth of the regression model needs to be judged to reduce the size of the error data, thereby improving the prediction accuracy and the accuracy. However, the deeper the network depth, the smaller the corresponding error data, and the more data complexity and data parameters the higher the device requirements, so the network depth needs to be selected.
And performing error data calculation on different network depths to obtain error data of corresponding depths, namely depth error data, comparing the depth error data of two adjacent depths to obtain difference data between the depth error data of different depths, and comparing the difference data, if the difference data is smaller than a preset difference value, judging that the accuracy of the dual-frequency index value prediction corresponding to the network depth corresponding to the difference data meets the requirement, and correspondingly selecting the depth data of the two network depths corresponding to the difference data if the data complexity and the equipment requirement are lower, and taking the network depth corresponding to the smaller depth in the two network depth data as the network depth of the regression model, namely increasing the accuracy of the regression model on the dual-frequency index value prediction, and further enabling the complexity of the equipment and the data to be smaller and increasing the feasibility of the dual-frequency index value prediction.
Based on the description of the embodiment of the dual-frequency index prediction method based on the generation of the countermeasure network, the embodiment of the invention also discloses a dual-frequency index prediction system based on the generation of the countermeasure network:
as shown in fig. 8, a dual-frequency index prediction system based on generation of an countermeasure network includes: the system comprises a data acquisition module 1, a regression model establishment module 2, a regression model training module 3 and a model test module 4.
The data acquisition module 1 is used for acquiring a double-frequency index database, wherein the double-frequency index database comprises a training data set for training a model and a test data set for testing the model;
the regression model building module 2 comprises a generator 21, a discriminator 22 and a regressor 23, wherein the regressor 23 is the same as other network layers of the discriminator 22, and the activation functions on the full-connection layer are different;
the generator 21 is in signal connection with the data acquisition module 1, and is used for receiving the double-frequency index database, and performing time domain analysis and frequency domain analysis on the data in the double-frequency index database to obtain error data; when frequency domain analysis is carried out on the double frequency data, fast Fourier transformation is carried out on the double frequency data to obtain corresponding frequency domain data, and the frequency domain analysis is carried out on the frequency domain data;
The discriminator 22 is in signal connection with the data acquisition module 1, and is used for receiving the double-frequency index database, performing fast Fourier transform to obtain frequency domain data, performing data analysis on the time domain data and the frequency domain data to determine whether the data are true or false, and outputting corresponding discrimination data;
the regressor 23 is in signal connection with the generator 21 and the arbiter 22 for receiving the error data and the discrimination data and analyzing to obtain a corresponding regression model.
The regression model training module 3 is in signal connection with the data acquisition module 1 and the regression model establishment module 2, and is used for receiving a training data set and a regression model, inputting the training data set into the regression model, and training the regression model to obtain a trained training result model;
the model test module 4 is internally provided with regression evaluation standard data and preset difference values, and the model test module 4 is in signal connection with the regression model training module 3 and the data acquisition module 1 and is used for receiving a training result model and a test data set, and inputting the test data set into the training result model to obtain error data; comparing the error data with regression evaluation standard data, and if the error data is in the regression evaluation standard data, judging that the training result model can predict the double-frequency index value; and meanwhile, carrying out error data calculation on the training result model at different network depths to obtain depth error data, carrying out difference calculation on the depth error data at different depths to obtain difference data, comparing the difference data with a preset difference value to judge the optimal network depth, and determining an optimal prediction model according to the optimal network depth.
Compared with the existing dual-frequency index prediction method and system based on the generation countermeasure network, the method and system provided by the application have the advantage that the accuracy of dual-frequency index prediction is improved.
The above embodiments are not intended to limit the scope of the present application, so: all equivalent changes in structure, shape and principle of the application should be covered in the scope of protection of the application.
Claims (10)
1. A method for generating a dual-frequency index prediction based on an antagonism network, comprising:
acquiring a double-frequency index database; the dual-frequency index database comprises a training data set for training the model and a test data set for testing the model;
based on the ACGAN structure, carrying out structural improvement on the ACGAN structure so as to establish and obtain a regression model;
training the regression model based on the training data set to obtain a trained regression model;
calculating a test data set based on the trained regression model to obtain error data;
evaluating the error data based on built-in regression evaluation standard data to judge the judging result of the trained regression model on the double-frequency index value;
and if the error data is smaller than the regression evaluation standard data, judging the network depth of the regression model to determine the optimal prediction model.
2. A method of generating a dual-frequency index prediction for an countermeasure network as claimed in claim 1, wherein: the step of performing structural improvement on the ACGAN structure based on the ACGAN structure to establish and obtain a regression model specifically comprises the following steps:
replacing an auxiliary classifier in the original ACGAN structure with a regressor (23) to process and regress the tagged data;
the regressor (23) model comprises two generators (21), two discriminators (22) and two regressors (23) in the simultaneous training of the time domain or the frequency domain; the regressor (23) is identical to the other network layers of the arbiter (22), while the activation functions on the fully connected layers are different;
taking the Sigmoid function as a classification function of an output layer in the discriminator (22); the Tanh function acts as an output layer activation function for the generator (21); the Relu function acts as an output layer function of the regressor (23).
3. A method of generating a dual-frequency index prediction for an countermeasure network according to claim 2, wherein: the step of performing structural improvement on the ACGAN structure based on the ACGAN structure to establish and obtain a regression model further comprises the following steps:
acquiring electroencephalogram data and preprocessing the electroencephalogram data to obtain time domain data of electroencephalogram;
Performing Fourier transform on the time domain data according to the time domain data to obtain corresponding frequency domain data;
transmitting the time domain data and the frequency domain data to a regression model to obtain time domain characteristic data and frequency domain characteristic data;
directly splicing the time domain characteristic data and the frequency domain characteristic data by adopting a convolutional neural network to obtain corresponding double-frequency index values;
the directly splicing the time domain feature data and the frequency domain feature data specifically comprises the following steps:
the time domain data and the frequency domain data input by the generator (21) are subjected to the same dimension setting and the same structural parameter setting so as to obtain time domain feature data and frequency domain feature data with the same dimension;
and directly splicing the time domain characteristic data and the frequency domain characteristic data with the same dimension.
4. A method of generating a double frequency index prediction for an countermeasure network according to claim 3, wherein: the step of performing structural improvement on the ACGAN structure based on the ACGAN structure to establish and obtain a regression model further comprises the following steps:
different training strategies and nonlinear activation functions are adopted to build a regression model;
participating the loss of the regressor (23) in generating the countermeasure training based on a training pattern similar to the ACGAN structure;
The relu function is used as an activation function of a convolution layer, and rated quantity normalization processing is carried out on the data.
5. A method of generating a dual-frequency index prediction for an countermeasure network as claimed in claim 1, wherein: the step of training the regression model based on the training data set to obtain a trained regression model specifically comprises the following steps:
inputting a time domain training data set and preset training times;
constructing a GAN generator (21) neural network of a time domain and a frequency domain, and constructing a discriminator (22) neural network and a regression neural network by combining the time domain and the frequency domain;
performing Fourier transform on the time domain training data set to obtain a frequency domain training data set;
training the regression model, stopping training when the training times are larger than the preset training times, and outputting a training result model.
6. The method for generating a double-frequency index prediction based on an antagonism network according to claim 5, wherein: the step of training the regression model, stopping training when the training times are greater than the preset training times, and outputting a training result model specifically comprises the following steps:
training a discriminator (22) neural network and a regression neural network;
Inputting rated quantity and optimizer super parameters, and obtaining time domain training data and frequency domain training data of batch size quantity;
according to the GAN generator (21) neural network of the time domain and the GAN generator (21) neural network of the frequency domain, generating rated amount of time domain noise and frequency domain noise correspondingly;
the Adam optimizer is set based on the optimizer hyper-parameters to train and update neural network model parameters of the time-domain and frequency-domain discriminators (22) neural network and the regression neural network to obtain model loss data.
7. The method for generating a double-frequency index prediction based on an antagonism network according to claim 6, wherein: the step of training the regression model, stopping training when the training times are greater than the preset training times, and outputting a training result model, and the step of training the regression model further comprises the following steps:
training a GAN producer;
the model loss data of the time domain noise, the frequency domain noise and the discriminators (22) and the regressors (23) are respectively input into a GAN generator (21) neural network of the time domain and a GAN generator (21) neural network of the frequency domain;
the GAN generator (21) model is trained and updated according to Adam optimizers.
8. A method of generating a dual-frequency index prediction for an countermeasure network as claimed in claim 1, wherein: if the error data is smaller than the regression evaluation standard data, judging the network depth of the regression model to determine the optimal prediction model, which specifically comprises the following steps:
When the error data is smaller than the regression evaluation standard data, judging that the trained regression model can judge the double-frequency index value and judge the depth of the regression model so as to determine the optimal prediction model;
performing error data calculation of different network depths on the regression model to obtain corresponding depth error data;
comparing the depth error data to obtain difference data between the depth error data of different depths, and comparing the difference data;
if the difference data is smaller than the preset difference, the network depth corresponding to the corresponding difference data is read to obtain the optimal network depth, and the optimal network depth is used as the optimal network depth of the regression model to obtain the optimal prediction model.
9. A dual-frequency exponential prediction system based on generation of an countermeasure network, characterized in that: comprising the following steps:
the data acquisition module (1) is used for acquiring a double-frequency index database; the dual-frequency index database comprises a training data set for training the model and a test data set for testing the model;
the regression model building module (2) is used for obtaining an ACGAN structure and improving the ACGAN structure to build a regression model;
The regression model training module (3) is configured to be in signal connection with the data acquisition module (1) and the regression model establishment module (2) and is used for receiving the training data set and the regression model, inputting the training data set into the regression model and training the regression model to obtain a trained training result model;
the model test module (4) is internally provided with regression evaluation standard data and preset difference values, is configured to be in signal connection with the regression model training module (3) and the data acquisition module (1) and is used for receiving the training result model and the test data set, and inputting the test data set into the training result model to obtain error data; comparing the error data with regression evaluation standard data, and if the error data is in the regression evaluation standard data, judging that the training result model can predict the double-frequency index value; and meanwhile, carrying out error data calculation on the training result model at different network depths to obtain depth error data, carrying out difference calculation on the depth error data at different depths to obtain difference data, comparing the difference data with a preset difference value to judge the optimal network depth, and determining an optimal prediction model according to the optimal network depth.
10. A method of generating a dual-frequency index prediction for an countermeasure network as claimed in claim 1, wherein: the regression model building module (2) comprises a generator (21), a discriminator (22) and a regressor (23), wherein the regressor (23) is the same as other network layers of the discriminator (22), and the activation functions on the full connection layers are different;
the generator (21) is in signal connection with the data acquisition module (1) and is used for receiving the double-frequency index database and carrying out time domain analysis and frequency domain analysis on data in the double-frequency index database so as to obtain error data; when frequency domain analysis is carried out on double frequency data, fast Fourier transformation is carried out on the double frequency data to obtain corresponding frequency domain data, and frequency domain analysis is carried out on the frequency domain data;
the discriminator (22) is in signal connection with the data acquisition module (1) and is used for receiving the double-frequency index database, performing fast Fourier transform on the double-frequency index data to obtain frequency domain data, performing data analysis on the time domain data and the frequency domain data to judge whether the data are true or false, and outputting corresponding judging data;
the regressor (23) is in signal connection with the generator (21) and the discriminator (22) and is used for receiving the error data and the discrimination data and analyzing the error data and the discrimination data to obtain a corresponding regression model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310942061.9A CN116910699B (en) | 2023-07-29 | 2023-07-29 | Dual-frequency index prediction method and system based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310942061.9A CN116910699B (en) | 2023-07-29 | 2023-07-29 | Dual-frequency index prediction method and system based on generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116910699A true CN116910699A (en) | 2023-10-20 |
| CN116910699B CN116910699B (en) | 2025-10-17 |
Family
ID=88354688
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310942061.9A Active CN116910699B (en) | 2023-07-29 | 2023-07-29 | Dual-frequency index prediction method and system based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116910699B (en) |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200176117A1 (en) * | 2017-08-11 | 2020-06-04 | Vuno, Inc. | Method for generating prediction result for predicting occurrence of fatal symptoms of subject in advance and device using same |
| CN111428648A (en) * | 2020-03-26 | 2020-07-17 | 五邑大学 | Electroencephalogram signal generation network, method and storage medium |
| CN111833893A (en) * | 2020-06-16 | 2020-10-27 | 杭州云嘉云计算有限公司 | Speech enhancement method based on artificial intelligence |
| US20220129735A1 (en) * | 2019-05-20 | 2022-04-28 | Institute of intelligent manufacturing, Guangdong Academy of Sciences | Semi-supervised Hyperspectral Data Quantitative Analysis Method Based on Generative Adversarial Network |
| KR20220064851A (en) * | 2020-11-12 | 2022-05-19 | 재단법인대구경북과학기술원 | Human-machine recognition system and method based on mind map database and method of the same |
| CN115062728A (en) * | 2022-07-12 | 2022-09-16 | 杭州央宏信息科技有限公司 | Emotion identification method and system based on time-frequency domain feature level fusion |
| CN116204817A (en) * | 2023-03-03 | 2023-06-02 | 杭州电子科技大学 | Emotion electroencephalogram data enhancement method integrating space-time characteristics |
| CN116226717A (en) * | 2023-03-09 | 2023-06-06 | 中国科学院空天信息创新研究院 | A hardware-accelerated recognition system for EEG signals based on a hybrid neural network |
-
2023
- 2023-07-29 CN CN202310942061.9A patent/CN116910699B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200176117A1 (en) * | 2017-08-11 | 2020-06-04 | Vuno, Inc. | Method for generating prediction result for predicting occurrence of fatal symptoms of subject in advance and device using same |
| US20220129735A1 (en) * | 2019-05-20 | 2022-04-28 | Institute of intelligent manufacturing, Guangdong Academy of Sciences | Semi-supervised Hyperspectral Data Quantitative Analysis Method Based on Generative Adversarial Network |
| CN111428648A (en) * | 2020-03-26 | 2020-07-17 | 五邑大学 | Electroencephalogram signal generation network, method and storage medium |
| CN111833893A (en) * | 2020-06-16 | 2020-10-27 | 杭州云嘉云计算有限公司 | Speech enhancement method based on artificial intelligence |
| KR20220064851A (en) * | 2020-11-12 | 2022-05-19 | 재단법인대구경북과학기술원 | Human-machine recognition system and method based on mind map database and method of the same |
| CN115062728A (en) * | 2022-07-12 | 2022-09-16 | 杭州央宏信息科技有限公司 | Emotion identification method and system based on time-frequency domain feature level fusion |
| CN116204817A (en) * | 2023-03-03 | 2023-06-02 | 杭州电子科技大学 | Emotion electroencephalogram data enhancement method integrating space-time characteristics |
| CN116226717A (en) * | 2023-03-09 | 2023-06-06 | 中国科学院空天信息创新研究院 | A hardware-accelerated recognition system for EEG signals based on a hybrid neural network |
Non-Patent Citations (2)
| Title |
|---|
| HONGNA LI, ET AL.: "An Improved Neural Network Bispectral Index Monitoring Algorithm in Time and Frequency Domain", EITCE \'23: PROCEEDINGS OF THE 2023 7TH INTERNATIONAL CONFERENCE ON ELECTRONIC INFORMATION TECHNOLOGY AND COMPUTER ENGINEERING, 17 April 2024 (2024-04-17), pages 645 * |
| 荣芬奇: "基于生成对抗网络的脑电信号研究", 中国优秀硕士学位论文全文数据库-基于生成对抗网络的脑电信号研究, 15 September 2021 (2021-09-15), pages 006 - 166 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116910699B (en) | 2025-10-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113160190A (en) | Method, equipment, device and storage medium for detecting abnormity of fastener of contact network | |
| CN118393005A (en) | Ultrasonic detection method and system | |
| Duan et al. | Health indicator construction and status assessment of rotating machinery by spatio-temporal fusion of multi-domain mixed features | |
| CN117235661A (en) | AI-based direct drinking water quality monitoring method | |
| CN119623531B (en) | Supervised time series water level data generation method, system and storage medium | |
| CN112200238A (en) | Hard rock tension-shear fracture identification method and device based on sound characteristics | |
| CN108491931B (en) | Method for improving nondestructive testing precision based on machine learning | |
| CN118277842A (en) | Substrate classification method and system based on deep sea multi-beam water body bottom echo information | |
| CN120011713A (en) | A data collection and analysis system for network big data information analysis | |
| CN116738259B (en) | Method and device for fingerprint extraction and identification of electromagnetic leakage radiation sources based on multi-harmonics | |
| CN117041847B (en) | Adaptive microphone matching method and system for hearing aid | |
| Wang et al. | Self‐supervised representation learning of metro interior noise based on variational autoencoder and deep embedding clustering | |
| Pillai et al. | Sonic Signatures: Sequential Model-driven Music Genre Classification with Mel Spectograms | |
| CN114302301A (en) | Frequency response correction method and related product | |
| Huang et al. | Signal frequency estimation based on RNN | |
| CN115310499B (en) | Industrial equipment fault diagnosis system and method based on data fusion | |
| CN116910699B (en) | Dual-frequency index prediction method and system based on generation countermeasure network | |
| CN119780638B (en) | Cable oscillation ultralow-frequency voltage withstand and partial discharge integrated test method and system | |
| Säilynoja et al. | Posterior SBC: Simulation-Based Calibration Checking Conditional on Data | |
| CN120317526A (en) | Method, device, equipment and storage medium for characteristic analysis of distribution transformer load data | |
| CN118937915A (en) | A method for detecting arc fault | |
| CN114782211A (en) | A method and system for obtaining information on the distribution range of seamounts | |
| CN120802030B (en) | General-purpose generator electrical monitoring system with fault self-diagnosis function | |
| Kunte et al. | A Domain Knowledge Informed Approach for Anomaly Detection of Electric Vehicle Interior Sounds | |
| CN115223538B (en) | Method, device, equipment, medium and program product for training vocoder model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |