CN113840297A - Frequency spectrum prediction method based on radio frequency machine learning model drive - Google Patents
Frequency spectrum prediction method based on radio frequency machine learning model drive Download PDFInfo
- Publication number
- CN113840297A CN113840297A CN202111169587.5A CN202111169587A CN113840297A CN 113840297 A CN113840297 A CN 113840297A CN 202111169587 A CN202111169587 A CN 202111169587A CN 113840297 A CN113840297 A CN 113840297A
- Authority
- CN
- China
- Prior art keywords
- network
- model
- data
- training
- prediction
- 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
- 238000000034 method Methods 0.000 title claims abstract description 106
- 238000001228 spectrum Methods 0.000 title claims abstract description 90
- 238000010801 machine learning Methods 0.000 title claims abstract description 27
- 238000012549 training Methods 0.000 claims abstract description 62
- 230000008569 process Effects 0.000 claims abstract description 34
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000012360 testing method Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 19
- 238000013135 deep learning Methods 0.000 claims description 12
- 230000003595 spectral effect Effects 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 11
- 238000005562 fading Methods 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 239000000654 additive Substances 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 4
- 230000000996 additive effect Effects 0.000 claims description 4
- 238000005457 optimization Methods 0.000 claims description 4
- 230000000644 propagated effect Effects 0.000 claims description 4
- 230000003247 decreasing effect Effects 0.000 claims description 3
- 230000001902 propagating effect Effects 0.000 claims description 2
- 238000012216 screening Methods 0.000 claims description 2
- 238000013528 artificial neural network Methods 0.000 description 11
- 238000004891 communication Methods 0.000 description 9
- 230000015654 memory Effects 0.000 description 7
- 238000004088 simulation Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 4
- 230000006403 short-term memory Effects 0.000 description 4
- 230000007787 long-term memory Effects 0.000 description 3
- 238000013508 migration Methods 0.000 description 3
- 230000005012 migration Effects 0.000 description 3
- 238000007476 Maximum Likelihood Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 229910003460 diamond Inorganic materials 0.000 description 1
- 239000010432 diamond Substances 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
- H04W16/22—Traffic simulation tools or models
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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; CALCULATING OR 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/391—Modelling the propagation channel
- H04B17/3913—Predictive models, e.g. based on neural network models
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Electromagnetism (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses a frequency spectrum prediction method based on radio frequency machine learning model driving, which comprises the following steps: s1, collecting frequency spectrum data, and preprocessing the collected frequency spectrum data; s2, determining the step length M of the input data for the order of the autoregressive model according to the Chichi information criterion; s3, expanding the linear combination process of the autoregressive model into an M-layer network structure, and introducing new trainable parameters into the M-layer network structure to construct an M-layer frequency spectrum prediction network model driven by a radio frequency machine learning model; s4, training the spectrum prediction network model by using the training set data; s5, judging whether training is finished or not, if so, inputting the test set data into the trained frequency spectrum prediction network model, outputting a prediction result, and ending the process; if not, after adding one to the training iteration number, returning to the step S4 until the maximum iteration number is reached. The invention not only endows the network with interpretability, improves the prediction performance, but also accelerates the convergence speed of the network.
Description
Technical Field
The invention relates to the technical field of communication, in particular to a frequency spectrum prediction method based on radio frequency machine learning model driving.
Background
The spectrum prediction is a vital technology in a cognitive radio network, and the use state of a future spectrum is deduced by learning the rule among historical spectrum data, so that a large amount of time and energy consumed in the spectrum sensing process are reduced, and the spectrum sensing efficiency is improved. Meanwhile, the spectrum prediction is helpful for improving the performance of spectrum management and control, and is an essential part in spectrum decision. Existing spectrum prediction methods can be divided into two categories, model-based prediction methods and data-based prediction methods. Model-based methods, such as autoregressive models, are based on the ideal assumption that data at the current point in time can be predicted from a linear combination of past data, which is often impractical in real scenarios, and therefore can lead to poor prediction performance, especially at low signal-to-noise ratios.
In recent years, data-driven methods have attracted attention because of their powerful feature learning capabilities. The differential moving average autoregressive model, the time-delay neural network and the long-short term memory neural network are compared in a published paper of "Experimental results on the prediction of memory in network for spectrum prediction in mobile radio bases" (IEEE trans. Cogn. Commun. Net., vol.6, No.2, pp.771-782,2020) of Ozyegen O, Mohammadja S, Kavurmacolog E and the like, and experiments prove that the data-driven method obtains higher precision than the model-driven prediction method. But network parameters are difficult to determine when addressing specific prediction problems, as the selection of these parameters is not based on a specific theory. Yu L, Chen J, Ding G et al, in its published paper, "Spectrum Prediction Based on Taguchi Method in Deep Learning With Long Short-Term Memory" (IEEE Access, vol.6, pp.45923-45933,2018), analyzed the predicted performance under different network configurations and analyzed the effect of network hyper-parameters on the predicted performance. Meanwhile, experiments prove that the long-term and short-term memory neural network can obtain better prediction performance than the traditional multilayer perceptron. A cross-band spectrum prediction method based on deep migration learning is disclosed in a patent application 'a cross-band spectrum prediction method based on deep migration learning' (application number CN201910995721.3 application publication number CN110730046A) proposed by China people liberation army engineering university. The method utilizes the similarity of each channel between frequency bands, and digs out the internal rules and the association among the frequency spectrum data through historical frequency spectrum data measured on the channels of other frequency bands, and migrates to the current frequency band so as to predict the frequency spectrum state of the current frequency band at the future moment, thereby solving the problem of insufficient training data to a certain extent, but the migration learning method has more limitations. Firstly, a pre-trained model is needed which already achieves good predictive performance on a large data set, and secondly, the prediction problem aimed at by the pre-trained model must have high similarity with the problem to be solved, which has certain limitations in practical application.
These data-only approaches all assume that there is sufficient tag data to complete network training, which is not possible in practical communication scenarios. The reason for this is that it is difficult to acquire a large amount of high-quality tag data in practice. In addition, the convergence rate of the algorithm is low, and real-time spectrum prediction is difficult to realize, so that the prediction performance cannot accurately reflect the use state of the real spectrum, and the real-time requirement of a future wireless communication network cannot be met. Therefore, there is a need to develop new spectrum prediction schemes.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a frequency spectrum prediction method based on the driving of a radio frequency machine learning model, and not only is network interpretability endowed, but also the prediction performance is improved and the convergence speed of the network is accelerated by introducing the knowledge in the communication field and providing a lightweight design method.
In order to achieve the purpose, the invention adopts the following technical scheme:
a spectrum prediction method based on radio frequency machine learning model driving, the spectrum prediction method comprises the following steps:
s1, collecting frequency spectrum data, and preprocessing the collected frequency spectrum data;
s2, determining the step length M of the input data for the order of the autoregressive model according to the Chichi information criterion;
s3, expanding the linear combination process of the autoregressive model into an M-layer network structure, and introducing new trainable parameters into the M-layer network structure to construct an M-layer frequency spectrum prediction network model driven by a radio frequency machine learning model;
s4, training the spectrum prediction network model by using the training set data;
s5, judging whether training is finished or not, if so, inputting the test set data into the trained frequency spectrum prediction network model, outputting a prediction result, and ending the process; if not, the training iteration number is increased by one, and the process returns to the step S4.
In order to optimize the technical scheme, the specific measures adopted further comprise:
further, in step S1, the process of collecting the spectrum data includes:
when the master user is busy, the received signal is a fading signal subjected to channel fading and additive white Gaussian noise interference, xt(n)＝ht(n)st(n)+εt(n); wherein s ist(n) denotes an nth sampling point, x, of a signal transmitted at the t-th time slott(n) represents the nth sample point of the received signal in the t-th time slot, h (n) represents the channel fading function, epsilon (n) represents the mean value of 0 and the variance of sigma2Additive white gaussian noise of (1);
when the primary user is idle, the received signal contains only noise, xt(n)＝εt(n)。
Further, in step S1, the preprocessing of the collected spectrum data includes the following steps:
and S11, calculating the power value of the spectrum signal by adopting the following formula:
wherein N represents the number of signal sampling points in each time slot, | · | represents the modulo operation, x (t) is the spectral power data obtained for t time slot, xt(n) represents an nth sampling point of a received signal at a t-th slot;
s12, normalizing the frequency spectrum signal by adopting the following formula:
where x (t) represents the spectral data at time t, xminRepresents the minimum value, x, of all spectral datamaxRepresenting the maximum of all spectral data.
Further, in step S2, the process of determining the step size M of the input data according to the akachi pool information criterion for the auto-regressive model includes the following steps:
s21, setting the initial model order p to 1 for the p-order autoregressive model;
s22, calculating and recording the AIC value of the p-th order autoregressive model according to the following formula:
yAIC＝-2ln(L)+2k
wherein L represents the maximum likelihood function of the model, k is the number of variables in the model, and yAICRepresenting the corresponding AIC value of the model;
s23, let the model order p be p +1, return to step S22, recalculate and record the AIC value until p is pmaxSo far, proceed to step S24;
and S24, screening out the model order corresponding to the minimum AIC value as the step length M of the input data.
Further, p ismaxIs 10.
Further, in step S3, the process of developing the linear combination process of the autoregressive model into an M-layer network structure includes the following steps:
expanding the calculation process of the autoregressive structure into a multi-layer network in deep learning, taking linear operation of each order as one layer of the network, and training by a deep learning method to obtain optimized network parameters;
the m-th network computing process comprises the following steps:
in the formula, xT+m-MRepresenting the incoming data for the m-th layer network,
weight parameter of input data of mth layer, M is total number of network layers, T is current time slot, ymIndicating the output data of the m-th layer network.
Further, in step S3, the process of introducing new trainable parameters into the M-layer network structure to construct an M-layer spectrum prediction network model driven by the radio frequency machine learning model includes the following steps:
for the mth layer, the calculation formula is transformed into:
where T denotes the current time slot, xT+m-MInput data representing a network of m-th layers, ymOutput data representing an m-th layer network;
and alphamTwo learnable parameters representing the m-th tier network,
is weighting of current input dataParameter, αmThe weighting parameter is output by the upper layer network and is used for selectively forgetting the past partial output; the nonlinear activation function σ (-) is used to avoid linear input data combinations.
Further, in step S4, the process of training the spectrum prediction network model by using the training set data includes the following steps:
s41, randomly initializing trainable parameters of the network according to a monotone decreasing rule;
s42, setting the maximum iteration times and the learning rate, and taking an Adam optimization algorithm as a network training optimizer;
s43, making the iteration time epoch equal to 1;
s44, inputting the training data into the spectrum prediction network model in batches for training, and reversely propagating the training error of each batch to optimize the network parameters; the error loss function adopts the most common mean square error loss function, and the calculation formula is as follows:
wherein, yiRepresenting the true result of the ith set of samples,
the prediction result of the ith group of samples is shown, and Q represents the total number of samples.
And S45, when all the data of all the batches in the training data are completely propagated reversely, making epoch equal to epoch +1, and returning to the step S44 until the epoch reaches the maximum iteration number.
The invention discloses a frequency spectrum prediction method based on radio frequency machine learning model driving, which mainly solves the problems of low prediction accuracy, poor real-time performance, network lack of interpretability and the like of the existing frequency spectrum prediction method. The method comprises the following implementation steps: collecting frequency spectrum data; preprocessing frequency spectrum data; determining the order of the model according to the Chichi information criterion so as to determine the step length of input data; expanding the linear combination structure of the autoregressive model into a multilayer network; adding new trainable parameters in a network to construct a frequency spectrum prediction framework driven by a radio frequency machine learning model; training a network model by using the training data; judging whether the network training is finished; inputting the test set data into a network; and outputting a prediction result. The invention introduces the knowledge in the communication field, overcomes the black box effect of the traditional deep learning network and increases the interpretability of the network; the network structure is simple, trainable parameters are few, the network training speed is accelerated, and the method can be applied to actual communication scenes; compared with pure model driving and pure data driving methods, the method improves the spectrum prediction performance.
The invention has the beneficial effects that:
first, the invention introduces an autoregressive structure as expert knowledge of a new prediction network framework, overcomes the black box effect of the network and increases the interpretability of the network compared with the traditional deep learning framework.
Second, the invention designs a network framework in which the network elements of each layer contain only two trainable parameters. Compared with a pure data driving method, the method greatly reduces the network scale, improves the training speed of the network, and provides guarantee for the real-time requirement of the actual communication scene.
Thirdly, compared with a pure model-driven method, the method does not need ideal assumption in the model, improves the spectrum prediction performance by utilizing the method based on the radio frequency machine learning model drive, and can be applied to a future actual wireless communication system.
Drawings
Fig. 1 is a flowchart of a spectrum prediction method based on radio frequency machine learning model driving according to an embodiment of the present invention.
FIG. 2 is a diagram of a predictive network framework according to an embodiment of the invention.
Fig. 3 is a comparison graph of prediction accuracy of the spectrum prediction method according to the embodiment of the present invention and other existing spectrum prediction methods under different snr conditions.
Fig. 4 is a graph comparing the training convergence rates of the spectrum prediction method according to the embodiment of the present invention and other existing spectrum prediction methods.
Detailed Description
The present invention will now be described in further detail with reference to the accompanying drawings.
It should be noted that the terms "upper", "lower", "left", "right", "front", "back", etc. used in the present invention are for clarity of description only, and are not intended to limit the scope of the present invention, and the relative relationship between the terms and the terms is not limited by the technical contents of the essential changes.
Fig. 1 is a flowchart of a spectrum prediction method based on radio frequency machine learning model driving according to an embodiment of the present invention. Referring to fig. 1, the spectrum prediction method includes the steps of:
and S1, collecting the frequency spectrum data and preprocessing the collected frequency spectrum data.
And S2, determining the step length M of the input data for the autoregressive model according to the Chichi information criterion.
And S3, expanding the linear combination process of the autoregressive model into an M-layer network structure, and introducing new trainable parameters into the M-layer network structure to construct the M-layer frequency spectrum prediction network model driven by the radio frequency machine learning model.
And S4, training the spectrum prediction network model by using the training set data.
S5, judging whether training is finished or not, if so, inputting the test set data into the trained frequency spectrum prediction network model, outputting a prediction result, and ending the process; if not, the training iteration number is increased by one, and the process returns to the step S4.
The foregoing spectrum prediction method is further described with reference to the drawings.
Spectral prediction requires decisions to be made in two cases based on samples of the received signal.
First, when the primary user is busy, the received signal is a fading signal subjected to channel fading and additive white gaussian noise interference, xt(n)＝ht(n)st(n)+εt(n) of (a). Wherein s ist(n) denotes an nth sampling point, x, of a signal transmitted at a t-th time slott(n) denotes the nth of the received signal in the t-th slotSample points, h (n) denotes the channel fading function, ε (n) denotes the mean 0, and the variance σ2White additive gaussian noise.
Second, when the primary user is idle, the received signal contains only noise xt(n)＝εt(n)。
In the embodiment of the present invention, a common BPSK signal is used as the transmission signal. The active mode of the primary user is modeled as a process of switching between switch states and the on and off durations are modeled as two random processes. Meanwhile, we assume that the two random processes of turning on and off are independent of each other.
And 2, preprocessing the frequency spectrum data.
According to the spectrum data obtained in step 1, we first calculate the signal power value:
wherein, N represents the number of signal sampling points in each time slot, | · | represents the modulo operation, and x (t) is the spectrum power data obtained for the t-th time slot. In the example of the present invention, we set N1000.
For a deep neural network structure, although the deeper network can further extract features, the deeper network inevitably brings about the problems of gradient disappearance or gradient explosion. Thus, after completing the signal power calculation, we perform an initial normalization operation on the input data:
where x (t) represents the spectral data at time t, xminRepresents the minimum value, x, of all spectral datamaxRepresenting the maximum of all spectral data.
Step 3, determining the step length M of input data according to the model order of the Chichi information criterion (AIC):
we propose a spectrum prediction framework based on the driving of a radio frequency machine learning model by using knowledge in the field of an autoregressive model, so that the autoregressive model needs to be firstly ordered to determine the input step length of the network.
AIC is a standard for measuring the fitting superiority of a statistical model, is built on the concept of entropy, and provides a standard for balancing the complexity of an estimated model and the superiority of fitting data. Usually, it is a weighted function of the fitting accuracy and the unknown number of parameters, yAIC-2ln (l) +2 k. Wherein, L represents the maximum likelihood function under the model, k is the number of variables in the model, yAICRepresenting the corresponding AIC value of the model. When there is a large difference between the two models, this difference is mainly reflected in the likelihood function term, and when the difference is not significant, the second term plays a major role, so that the model with a small number of parameters is considered to be a more appropriate model. In general, the aim of AIC is to find a model that can best interpret data but contains the fewest parameters.
For a p-order autoregressive model, setting initial p to be 1, setting maximum p to be 10, calculating AIC values corresponding to different model orders p, wherein p corresponding to the minimum AIC value is the most appropriate model order, namely the step length M of input data.
And 4, developing the linear combination process of the autoregressive model into an M-layer network structure.
The formula for the autoregressive structure is calculated as follows:
wherein x istRepresenting the value of the sequence at time t, p being the order of the autoregressive model,
representing a weight parameter, c being a constant term, εtIs an error term.
The autoregressive structure adopts the traditional least square method to estimate the weight parameters
But instead of the other end of the tubeThe precision of the parameters obtained by the method is low. The conventional prediction scheme based on deep learning has a large-scale network architecture, needs to train a large number of parameters, and has a low convergence rate. Furthermore, its black box architecture makes the network unexplainable. Aiming at the problems, the calculation process of the autoregressive structure is expanded into a multi-layer network in deep learning, linear operation of each order is used as one layer of the network, and optimized network parameters are obtained through training by a deep learning method. The calculation process of the m-th network is as follows:
wherein x isT+m-MRepresenting the incoming data for the m-th layer network,
weight parameter of input data of mth layer, M is total number of network layers, T is current time slot, ymIndicating the output data of the m-th layer network.
And 5, introducing new trainable parameters into the network to construct a spectrum prediction framework driven by a radio frequency machine learning model.
In order to better learn the distribution of the spectrum data, the invention introduces new trainable parameters on the basis of an autoregressive structure to selectively forget the output result of the previous layer, and simultaneously introduces a nonlinear activation function to avoid linear input data combination. The complete network structure consists of M cascaded layers. Each layer has the same architecture but different trainable parameters. For the m-th layer network, the calculation process is as follows:
where T denotes the current time slot, xT+m-MInput data representing a network of m-th layers, ymIndicating the output data of the m-th layer network.
And alphamTwo trainable parameters representing the m-th network,
is a weighting parameter, alpha, of the current input datamIs a weighting parameter for the output of the network of the previous layer, and is used for selectively forgetting the past partial output. A non-linear activation function σ (-) is introduced to avoid linear input data combinations.
From the above formula, it can be seen that, unlike the conventional deep learning black box structure, the whole network structure of the present invention is modified from an autoregressive structure, which gives interpretability to the network. Compared with the traditional autoregressive model in which the least square method is used for parameter estimation, the network provided by the invention obtains the optimized parameters through deep learning method training, so that more accurate parameter estimation can be obtained. Finally, only two trainable parameters per network element can accelerate the network training process.
And 6, training the network model by using the training data.
Firstly, the trainable parameters of the network are initialized randomly, and since the sequence is always related to the data point closest to the current time to the highest degree, and the distance is longer, the relevance is lower, the initial trainable parameters of the network are initialized randomly according to a monotonous decreasing rule. The number of initial network training iterations epoch is equal to 1, the maximum number of iterations is 100, the learning rate is 0.001, and an Adam optimization algorithm is used as a network training optimizer.
And secondly, inputting the training data into the network for training in batches, wherein the batch size can be adjusted, and the training error of each batch is propagated reversely so as to optimize the network parameters. The error loss function adopts the most common mean square error loss function, and the calculation formula is as follows:
wherein, yiRepresenting the true result of the ith set of samples,
the prediction result of the ith group of samples is shown, and Q represents the total number of samples.
When all the data of all the batches in the training data are completely propagated reversely, the data are 1 epoch.
And 7, judging whether the network training is finished.
And judging whether the current epoch reaches the maximum iteration number, if so, performing the step 7, otherwise, adding one to the epoch, and continuing to perform the second step of network training of the step 6.
And 8, inputting the test set data into a network.
And 9, outputting a prediction result.
The effect of the present invention will be further explained with the simulation experiment.
1. Simulation conditions and parameter setting:
the simulation experiment of the invention is carried out on a simulation platform of Python3.6, Pytroch 1.5.1. The computer CPU is of type Inter core i7, and is equipped with independent display card of type Inviad gel GTX 1660 SUPER.
The sampling point of the signal in each time slot is 1000 points. The switching duration of the primary user is modeled as two random processes that follow an exponential distribution. The input step size of the network is 7, ordered by the AIC criterion. The maximum iteration number of the network training is 100, the learning rate is 0.001, an Adam optimization algorithm is selected as a network training optimizer, and the data volume of each batch is 100.
2. Simulation content:
FIG. 3 is a comparison graph of prediction accuracy under different SNR conditions using the present invention and the prior art. The abscissa in fig. 3 represents different signal-to-noise ratios (dB) and the ordinate represents prediction accuracy. The polygonal line marked by a rectangle represents a prediction accuracy curve adopting the method, the polygonal line marked by a circle represents a prediction accuracy curve of a traditional frame based on the long-term and short-term memory neural network under different signal-to-noise ratios, and the polygonal line marked by a diamond represents a prediction accuracy curve of a traditional autoregressive model under different signal-to-noise ratios. The signal-to-noise ratio varies from-1 dB to 14 dB. By comparison, the prediction accuracy of the method is obviously higher than that of other existing methods. From the general trend, the prediction accuracy increases as the signal-to-noise ratio increases. When the signal-to-noise ratio is 0dB, the prediction accuracy of the method reaches about 74%, which exceeds the accuracy of the method based on the long-short term memory neural network by about 10% and exceeds the accuracy of the method based on the autoregressive model by about 21%; at a signal-to-noise ratio of 4dB, the prediction accuracy of the method reaches about 81%, which exceeds the accuracy of the method based on the long-short term memory neural network by about 8% and exceeds the accuracy of the method based on the autoregressive model by about 18%. In addition, under the condition of a lower signal-to-noise ratio, the prediction accuracy of the autoregressive model is poor, but the accuracy is obviously improved along with the increase of the signal-to-noise ratio, and when the signal-to-noise ratio reaches 8dB, the prediction accuracy based on the autoregressive model exceeds the method based on the long-short term memory neural network, but both are lower than the method provided by the invention.
Fig. 4 is a graph comparing training speeds of networks using the present invention and the prior art. In fig. 4, the abscissa represents the number of training sessions (times), and the ordinate represents the loss function value. The broken lines marked by rectangles represent the convergence curves of the method of the invention, and the broken lines marked by circles represent the training convergence curves of the framework based on the long-short term memory neural network. By comparing the training speed convergence curves obtained by the two methods, the training speed of the method is obviously higher than that of the existing method based on the long-term and short-term memory neural network. The inventive method achieves convergence only at about 10 training sessions, whereas the long-short term memory neural network-based method achieves convergence at about 30 x training sessions.
By integrating the simulation results and analysis, the frequency spectrum prediction network framework based on the radio frequency machine learning model drive can achieve higher prediction accuracy than the existing method, is low in network complexity and faster in convergence speed, and can be better applied to actual communication scenes.
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the invention may be made by those skilled in the art without departing from the principle of the invention.
Claims (8)
1. A frequency spectrum prediction method based on radio frequency machine learning model driving is characterized by comprising the following steps:
s1, collecting frequency spectrum data, and preprocessing the collected frequency spectrum data;
s2, determining the step length M of the input data for the order of the autoregressive model according to the Chichi information criterion;
s3, expanding the linear combination process of the autoregressive model into an M-layer network structure, and introducing new trainable parameters into the M-layer network structure to construct an M-layer frequency spectrum prediction network model driven by a radio frequency machine learning model;
s4, training the spectrum prediction network model by using the training set data;
s5, judging whether training is finished or not, if so, inputting the test set data into the trained frequency spectrum prediction network model, outputting a prediction result, and ending the process; if not, after adding one to the training iteration number, returning to the step S4 until the maximum iteration number is reached.
2. The method for predicting frequency spectrum driven by radio frequency machine learning model according to claim 1, wherein the step S1, the collecting of frequency spectrum data includes:
when the master user is busy, the received signal is a fading signal subjected to channel fading and additive white Gaussian noise interference, xt(n)＝ht(n)st(n)+εt(n); wherein s ist(n) denotes an nth sampling point, x, of a signal transmitted at the t-th time slott(n) represents the nth sample point of the received signal in the t-th time slot, h (n) represents the channel fading function, epsilon (n) represents the mean value of 0 and the variance of sigma2Additive white gaussian noise of (1);
when the primary user is idle, the received signal contains only noise, xt(n)＝εt(n)。
3. The radio frequency machine learning model-driven spectrum prediction method according to claim 1, wherein in step S1, the preprocessing of the collected spectrum data includes the following steps:
and S11, calculating the power value of the spectrum signal by adopting the following formula:
wherein N represents the number of signal sampling points in each time slot, | · | represents the modulo operation, x (t) is the spectral power data obtained for t time slot, xt(n) represents an nth sampling point of a received signal at a t-th slot;
s12, normalizing the frequency spectrum signal by adopting the following formula:
where x (t) represents the spectral data at time t, xminRepresents the minimum value, x, of all spectral datamaxRepresenting the maximum of all spectral data.
4. The radio frequency machine learning model-driven spectrum prediction method according to claim 1, wherein in step S2, the autoregressive model is ranked according to the Chi information criterion, and the process of determining the step size M of the input data comprises the following steps:
s21, setting the initial model order p to 1 for the p-order autoregressive model;
s22, calculating and recording the AIC value of the p-th order autoregressive model according to the following formula:
yAIC＝-2ln(L)+2k
wherein L represents the maximum under the modelLikelihood function, k is the number of variables in the model, yAICRepresenting the corresponding AIC value of the model;
s23, let the model order p be p +1, return to step S22, recalculate and record the AIC value until p is pmaxSo far, proceed to step S24;
and S24, screening out the model order corresponding to the minimum AIC value as the step length M of the input data.
5. The radio frequency machine learning model-driven spectrum prediction method according to claim 4, wherein p ismaxIs 10.
6. The radio frequency machine learning model-driven spectrum prediction method according to claim 1, wherein the process of developing the linear combination process of the autoregressive model into the M-layer network structure in step S3 includes the following steps:
expanding the calculation process of the autoregressive structure into a multi-layer network in deep learning, taking linear operation of each order as one layer of the network, and training by a deep learning method to obtain optimized network parameters;
the m-th network computing process comprises the following steps:
in the formula, xT+m-MRepresenting the incoming data for the m-th layer network,
weight parameter of input data of mth layer, M is total number of network layers, T is current time slot, ymIndicating the output data of the m-th layer network.
7. The method according to claim 6, wherein in step S3, the process of building an M-layer radio frequency machine learning model-driven spectrum prediction network model by introducing new trainable parameters into the M-layer network structure includes the following steps:
for the mth layer, the calculation formula is transformed into:
where T denotes the current time slot, xT+m-MInput data representing a network of m-th layers, ymOutput data representing an m-th layer network;
and alphamTwo learnable parameters representing the m-th tier network,
is a weighting parameter, alpha, of the current input datamThe weighting parameter is output by the upper layer network and is used for selectively forgetting the past partial output; the nonlinear activation function σ (-) is used to avoid linear input data combinations.
8. The method for spectrum prediction driven by a radio frequency machine learning model according to claim 1, wherein the step S4 of training the spectrum prediction network model by using the training set data includes the following steps:
s41, randomly initializing trainable parameters of the network according to a monotone decreasing rule;
s42, setting the maximum iteration times and the learning rate, and taking an Adam optimization algorithm as a network training optimizer;
s43, making the iteration time epoch equal to 1;
s44, inputting the training data into the spectrum prediction network model in batches for training, and reversely propagating the training error of each batch to optimize the network parameters; the error loss function adopts the most common mean square error loss function, and the calculation formula is as follows:
wherein, yiRepresenting the true result of the ith set of samples,
representing the prediction result of the ith group of samples, and Q represents the total number of samples;
and S45, when all the data of all the batches in the training data are completely propagated reversely, making epoch equal to epoch +1, and returning to the step S44 until the epoch reaches the maximum iteration number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111169587.5A CN113840297B (en) | 2021-10-08 | 2021-10-08 | Frequency spectrum prediction method based on radio frequency machine learning model drive |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111169587.5A CN113840297B (en) | 2021-10-08 | 2021-10-08 | Frequency spectrum prediction method based on radio frequency machine learning model drive |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113840297A true CN113840297A (en) | 2021-12-24 |
| CN113840297B CN113840297B (en) | 2022-08-30 |
Family
ID=78968372
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111169587.5A Active CN113840297B (en) | 2021-10-08 | 2021-10-08 | Frequency spectrum prediction method based on radio frequency machine learning model drive |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113840297B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114978304A (en) * | 2022-05-12 | 2022-08-30 | 上海交通大学 | Method, system and device for measuring gain spectrum of optical amplifier |
| CN115118365A (en) * | 2022-06-22 | 2022-09-27 | 中国人民解放军国防科技大学 | Frequency spectrum prediction interpretation method and system based on mask |
| CN115242336A (en) * | 2022-08-01 | 2022-10-25 | 北京航空航天大学 | Nonlinear output frequency spectrum prediction method for receiver radio frequency front-end device |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO1996035259A1 (en) * | 1995-05-05 | 1996-11-07 | Unto Kalervo Laine | Method of modelling the spectrum of a signal and apparatus for realizing the method |
| US6032113A (en) * | 1996-10-02 | 2000-02-29 | Aura Systems, Inc. | N-stage predictive feedback-based compression and decompression of spectra of stochastic data using convergent incomplete autoregressive models |
| CN107528649A (en) * | 2017-08-30 | 2017-12-29 | 华南师范大学 | A kind of Forecasting Methodology of the cognitive radio networks frequency spectrum based on Bagging |
| JP2018182405A (en) * | 2017-04-05 | 2018-11-15 | 株式会社国際電気通信基礎技術研究所 | Channel state predicting device, channel state predicting method, radio communication device, and radio communication method |
| CN109194423A (en) * | 2018-08-13 | 2019-01-11 | 中国人民解放军陆军工程大学 | The single-frequency point spectrum prediction method of shot and long term memory models based on optimization |
| CN110502806A (en) * | 2019-07-31 | 2019-11-26 | 电子科技大学 | A kind of wireless frequency spectrum degree prediction technique based on LSTM network |
| CN110730046A (en) * | 2019-10-18 | 2020-01-24 | 中国人民解放军陆军工程大学 | Cross-frequency-band spectrum prediction method based on deep migration learning |
| CN113067653A (en) * | 2021-03-17 | 2021-07-02 | 北京邮电大学 | Spectrum sensing method and device, electronic equipment and medium |
| CN113283576A (en) * | 2021-01-18 | 2021-08-20 | 上海应用技术大学 | Spectrum sensing method for optimizing LSTM based on whale algorithm |
-
2021
- 2021-10-08 CN CN202111169587.5A patent/CN113840297B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO1996035259A1 (en) * | 1995-05-05 | 1996-11-07 | Unto Kalervo Laine | Method of modelling the spectrum of a signal and apparatus for realizing the method |
| US6032113A (en) * | 1996-10-02 | 2000-02-29 | Aura Systems, Inc. | N-stage predictive feedback-based compression and decompression of spectra of stochastic data using convergent incomplete autoregressive models |
| JP2018182405A (en) * | 2017-04-05 | 2018-11-15 | 株式会社国際電気通信基礎技術研究所 | Channel state predicting device, channel state predicting method, radio communication device, and radio communication method |
| CN107528649A (en) * | 2017-08-30 | 2017-12-29 | 华南师范大学 | A kind of Forecasting Methodology of the cognitive radio networks frequency spectrum based on Bagging |
| CN109194423A (en) * | 2018-08-13 | 2019-01-11 | 中国人民解放军陆军工程大学 | The single-frequency point spectrum prediction method of shot and long term memory models based on optimization |
| CN110502806A (en) * | 2019-07-31 | 2019-11-26 | 电子科技大学 | A kind of wireless frequency spectrum degree prediction technique based on LSTM network |
| CN110730046A (en) * | 2019-10-18 | 2020-01-24 | 中国人民解放军陆军工程大学 | Cross-frequency-band spectrum prediction method based on deep migration learning |
| CN113283576A (en) * | 2021-01-18 | 2021-08-20 | 上海应用技术大学 | Spectrum sensing method for optimizing LSTM based on whale algorithm |
| CN113067653A (en) * | 2021-03-17 | 2021-07-02 | 北京邮电大学 | Spectrum sensing method and device, electronic equipment and medium |
Non-Patent Citations (2)
| Title |
|---|
| 杨丰瑞: ""基于频谱预测的认知无线接入研究"", 《中国新通信》, 5 April 2009 (2009-04-05) * |
| 赵晓东: ""一种融合EMD分解和LSTM网络的频谱占用度预测模型"", 《计算机科学》, vol. 47, 15 June 2020 (2020-06-15), pages 2 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114978304A (en) * | 2022-05-12 | 2022-08-30 | 上海交通大学 | Method, system and device for measuring gain spectrum of optical amplifier |
| CN114978304B (en) * | 2022-05-12 | 2023-08-25 | 上海交通大学 | Method, system and device for measuring gain spectrum of optical amplifier |
| CN115118365A (en) * | 2022-06-22 | 2022-09-27 | 中国人民解放军国防科技大学 | Frequency spectrum prediction interpretation method and system based on mask |
| CN115118365B (en) * | 2022-06-22 | 2023-07-07 | 中国人民解放军国防科技大学 | Spectrum prediction interpretation method and system based on mask |
| CN115242336A (en) * | 2022-08-01 | 2022-10-25 | 北京航空航天大学 | Nonlinear output frequency spectrum prediction method for receiver radio frequency front-end device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113840297B (en) | 2022-08-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113840297B (en) | Frequency spectrum prediction method based on radio frequency machine learning model drive | |
| CN110675623A (en) | Short-term traffic flow prediction method, system and device based on hybrid deep learning | |
| Ma et al. | Deep-esn: A multiple projection-encoding hierarchical reservoir computing framework | |
| CN107730003A (en) | One kind supports more high-precision NILM implementation methods of appliance type | |
| CN112766496B (en) | Deep learning model safety guarantee compression method and device based on reinforcement learning | |
| CN114039870B (en) | Deep learning-based real-time bandwidth prediction method for video stream application in cellular network | |
| CN107704426A (en) | Water level prediction method based on extension wavelet-neural network model | |
| CN112183742A (en) | Neural network hybrid quantization method based on progressive quantization and Hessian information | |
| CN115018193A (en) | Time series wind energy data prediction method based on LSTM-GA model | |
| CN113935513A (en) | CEEMDAN-based short-term power load prediction method | |
| CN113988086A (en) | Conversation processing method and device | |
| CN114596726A (en) | Parking position prediction method based on interpretable space-time attention mechanism | |
| CN108631817B (en) | Method for predicting frequency hopping signal frequency band based on time-frequency analysis and radial neural network | |
| CN116595356B (en) | Time sequence signal prediction method and device, electronic equipment and storage medium | |
| CN116709409A (en) | Knowledge distillation-based lightweight spectrum prediction method | |
| CN117313795A (en) | Intelligent building energy consumption prediction method based on improved DBO-LSTM | |
| Peng et al. | Hmm-lstm for proactive traffic prediction in 6g wireless networks | |
| CN115348182A (en) | Long-term spectrum prediction method based on depth stack self-encoder | |
| CN114970674A (en) | Time sequence data concept drift adaptation method based on relevance alignment | |
| CN114298166A (en) | Spectrum availability prediction method and system based on wireless communication network | |
| Chockalingam et al. | Extending World Models for Multi-Agent Reinforcement Learning in MALMÖ. | |
| CN114444654A (en) | NAS-oriented training-free neural network performance evaluation method, device and equipment | |
| Jerome et al. | Forecasting and anomaly detection on application metrics using lstm | |
| CN113469013A (en) | Motor fault prediction method and system based on transfer learning and time sequence | |
| CN117636183A (en) | Small sample remote sensing image classification method based on self-supervision pre-training |
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 | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20240510 Address after: No. 57 Wangqiao Road, Xiongzhou Street, Liuhe District, Nanjing City, Jiangsu Province, 211500 Patentee after: Nanjing Xingpuzhi Information Technology Co.,Ltd. Country or region after: China Address before: No. 29, Qinhuai District, Qinhuai District, Nanjing, Jiangsu Patentee before: Nanjing University of Aeronautics and Astronautics Country or region before: China |