CN103209417B - Based on Forecasting Methodology and the device of the spectrum occupancy state of neural net - Google Patents

Abstract

The invention discloses a neural network-based prediction method and device for spectrum occupancy state. The prediction method of the spectrum occupancy state based on the neural network includes the following steps: Step S1: Construct the neural network and use a pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron; Step S2: Input the set {b For each element in _i , b _i-1 ,..., b _im }, compare the output result of the neural network with the representation variable h _i+1 representing the occupancy state of the preset frequency band at the i+1th time sequence; modify the initial Parameters to obtain prediction parameters; b _i is the signal strength of the i-th time series preset frequency band, i is a positive integer, and m is a positive integer number smaller than i; step S3: input the set {b _i+1 , b _i , to the neural network at the same time, ..., each element in b _i-m+1 }, the neural network outputs the predicted characterization variable H _i+2 representing the occupancy state of the i+2th time sequence preset frequency band, and the i+2 time sequence is determined according to the predicted characterization variable H _i+2 The occupancy state of the preset frequency band has the advantages of low computational complexity and simple implementation.

Description

Method and device for predicting spectrum occupancy state based on neural network

technical field

The present invention relates to the field of communication technologies, in particular to a method and device for predicting spectrum occupancy status.

Background technique

In the process of dynamic spectrum allocation, the secondary user needs to temporarily use the idle spectrum of the primary user to communicate in order to achieve the purpose of improving spectrum utilization. The premise of communication based on the foregoing dynamic spectrum allocation is that the spectrum occupancy state can be accurately predicted, and an idle spectrum is found for communication of secondary users. Existing spectrum occupancy state prediction methods such as Kalman filter model, hidden Markov model and other prediction methods are all based on current and/or historical time series spectrum occupancy state and spectrum distribution information. However, in the specific implementation process, it is difficult for the secondary user end to obtain spectrum occupancy distribution information, and problems such as large amount of calculation, high computational complexity, and inaccurate prediction results are caused by the spectrum occupancy distribution involved.

Contents of the invention

(1) Purpose of the invention

The present invention provides a method and device for predicting spectrum occupancy status based on a neural network based only on the required frequency band occupation status in historical time series, with small calculation amount, low calculation complexity and high accuracy.

(2) Technical solutions

For reaching above-mentioned purpose, the prediction method of the spectrum occupancy state based on neural network of the present invention comprises the following steps successively:

Step S1: Construct a neural network and use a pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

Step S2: Input each element in the set {b _i , b _i-1 ,..., b _im } into the neural network, and compare the output result of the neural network with the representation variable h used to represent the occupancy state of the i+1th time sequence preset frequency band _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th timing preset frequency band, i is a positive integer, and m is a positive integer less than i;

Step S3: Simultaneously input each element in the set {b _i+1 , b _i ,..., L, b _i-m+1 } to the neural network, and the neural network outputs a prediction representing the occupancy state of the i+2th time sequence preset frequency band The characteristic variable H _i+2 , according to the predicted characteristic variable H _i+2 , determines the occupancy state of the i+2 time sequence preset frequency band.

Preferably, a step S0 is also included before the step S1;

The step S0: collect {bi, bi _-1 , ..., bi _im } and h _{i +1} .

Preferably, between the step S0 and the step S1, a step of normalizing the data collected in the step S0 is further included.

Preferably, both the initial parameters and the predicted parameters at least include weight coefficients between two adjacent layers of the neural network;

In the step S2, the initial parameters are corrected with a backpropagation algorithm according to the comparison result.

Preferably, in the step S3, the occupancy state of the preset frequency band at time i+2 corresponding to the H _i+2 is determined by a threshold value determination method.

In order to achieve the above object, the present invention is based on the prediction device of the spectrum occupancy state of neural network comprising:

The neural network building block is used to construct the neural network and use the pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

The neural network training module is used to input the elements in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output results of the neural network to represent the occupancy state of the i+1th timing preset frequency band The characteristic variable h _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th time series preset frequency band, i is a positive integer, and m is a positive integer smaller than i;

The spectrum occupancy state prediction module is used to simultaneously input the elements in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the output of the neural network represents the occupancy of the i+2th timing preset frequency band The predictive characterization variable H _i+2 of the state is used to determine the occupancy state of the preset frequency band at the i+2 time sequence according to the predictive characterization variable H _i+2 .

Further, the neural network-based prediction device for spectrum occupancy status further includes a data collection module for collecting {bi, bi _-1 , ..., bi _{im }} and hi ₊₁ .

Further, the data collection module also includes a normalization sub-module for normalizing the collected data.

Further, both the initial parameters and the prediction parameters at least include weight coefficients between two adjacent layers of the neural network;

The neural network training module includes a parameter modification sub-module for modifying initial parameters with a backpropagation algorithm according to the comparison result.

Further, the spectrum occupancy state prediction module includes a determination submodule for determining the occupancy state of the preset frequency band at time i _{+2 corresponding to H i+2} through a threshold determination method.

(3) The beneficial effects of the neural network-based spectrum occupancy state prediction method and device of the present invention

First: The present invention is based on the prediction method and device of the spectrum occupancy state of the neural network. The neural network determined by the pruning algorithm and adjusted by training is used to represent the signal strength and signal strength of the preset frequency band in the m+1 time sequence before the prediction in wireless communication. The mapping relationship between the predicted timing spectrum occupancy states, so that when the secondary user predicts the spectrum occupancy state, it does not need to obtain the occupancy distribution information of the frequency band like the traditional method, so there is no problem that the secondary user is difficult to obtain the frequency band distribution information .

Second: Since the frequency band occupancy distribution usually satisfies formulas such as Gaussian distribution and Poisson distribution, it is inevitable to perform complex operations such as differentiation and integration. The prediction method and device of the spectrum occupancy state based on the neural network of the present invention no longer involve frequency band occupation distribution information in the prediction process, and the neural network is an input composed of a nonlinear network formed by simulating the human brain and composed of simple calculation units The output mapping model, so the calculation is usually a relatively simple function operation and the weight operation between two adjacent layers of the network, which has the advantages of simple calculation, small calculation amount and accurate calculation.

Third: The present invention is based on the prediction method and device of the spectrum occupancy state of the neural network. Due to the strong learning ability of the neural network and the strong generalization ability, it can be applied to the prediction of spectrum occupancy in different frequency bands and different timings. And the prediction parameters are adjusted in real time so that the obtained results are more accurate, so compared with the traditional prediction methods and devices, it has stronger applicability, and the provided prediction information is more practical.

Description of drawings

FIG. 1 is a flow chart of a method for predicting a spectrum occupancy state based on a neural network according to Embodiment 3 of the present invention;

Fig. 2 is a flow chart of the apparatus for predicting spectrum occupancy status based on a neural network according to the second embodiment of the present invention.

detailed description

The neural network-based spectrum occupancy state prediction method of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

Embodiment one:

The signal strength of a certain frequency band can be used to measure the occupancy status of the frequency band, so the signal strength can be used as an input to obtain the occupancy information of the frequency point after related calculations such as de-attenuation of the neural network; and then by observing and abstracting several The occupancy state information of the frequency band in the next time sequence of the several consecutive time sequences can be predicted through corresponding calculation and transformation of the change trend of the occupancy state of the continuous time series.

The method for predicting the spectrum occupancy state based on the neural network in this embodiment is characterized in that the prediction method for the spectrum occupancy state based on the neural network includes the following steps:

Step S1: Construct a neural network and use a pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron; the neural network includes an input layer, an intermediate layer, and an output layer in turn; each layer is composed of several neurons, and the same layer There is no data interaction between the neurons of each layer, and the data is transmitted between the neurons between the layers to represent the mapping relationship between the input and the output; the pruning algorithm can effectively eliminate unnecessary neurons and/or connections of neurons, so as to achieve a simple mapping relationship between the input and output of the constructed neural network, accurate mapping results, and simple and reliable results;

Step S2: Input each element in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output result of the neural network with the representation variable h used to characterize the occupancy state of the i+1th timing preset frequency band _i+1 for comparison; modify the initial parameters according to the comparison results to obtain prediction parameters; b _i is the signal strength of the i-th timing preset frequency band, i is a positive integer, and m is a positive integer smaller than i; the value of i determines the set The initial value, the value of m determines the length of the set; in this step, the signal strength of the continuous m+1 time series is used as the input and the occupancy state of the preset frequency band in the next time series of the m+1 time series is the target value to train the neural network, so that the predicted value of the output of the neural network and the target value meet a certain error; if the error between the two is less than the preset error, the initial parameter of the neural network is the final prediction parameter; if it is greater than the preset error If the error needs to adjust the corresponding parameters, it is necessary to re-input {bi ,bi _-1 ,…,bi _im } to calculate the result, and compare the calculation result with h _{i +1} again, and then correct the parameters according to the comparison result. Finally, the operation result is smaller than the preset error, so that the initial parameter is repeatedly corrected, so that the final predicted value and the target value are smaller than the preset error time parameter as the final predicted parameter; by this method, the final parameters of the neural network are determined, and the implementation is simple. And for the final prediction result, there is a purpose of accurate prediction result.

Step S3: Simultaneously input each element in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the neural network outputs a predictive characterization variable representing the occupancy state of the i+2th time-series preset frequency band H _i+2 , judging the occupancy state of the preset frequency band at the i+2 time sequence according to the predicted characteristic variable H _i+2 .

In the specific communication process, the wireless spectrum is divided into frequency bands as carriers. Some frequency bands are used for uplink communication, some are used for downlink communication, and some are used for specific regional communication. Therefore, in different wireless communications, each communication frequency band different spectrum ranges. Therefore, this embodiment sets preset frequency bands for different communication needs according to the needs of user communication types. Since the neural network has a good generalization ability, it can well realize the correction of prediction parameters for different frequency bands and different timings, so that it can be used for different frequency bands. Prediction of spectrum occupancy information of frequency bands and different time series.

The above,

First of all, this embodiment is based on the prediction method of the spectrum occupancy state of the neural network, and uses the neural network to simulate the mapping relationship between the signal strength of the preset frequency band in the historical time series representing the occupancy state and the occupancy state of the preset frequency band in the prediction time series, so it does not involve The distribution information of the preset frequency bands, so there is no problem that it is difficult for the secondary users to obtain the distribution information of the frequency bands;

Secondly, the distribution information of general frequency bands satisfies the unavoidable need for highly complex operations such as differentiation, integration, and convolution integration when calculating such as Gaussian distribution and Poisson distribution. However, the method described in this embodiment does not involve frequency band distribution information. Therefore, it does not involve high-complexity calculations such as differentiation and integration, and the neural network itself has the characteristics of simple calculation and fast calculation, so this method has low computational complexity compared with traditional methods, and the relative computational complexity Small pluses.

Again, the method described in this embodiment is to adjust the prediction parameters in real time for different frequency bands and different timings, and the adjustment of the prediction parameters is closely related to factors such as the current network environment, so the prediction parameters and prediction methods are accurate, so that the prediction results are also more accurate ;

In addition, the method described in this embodiment uses a neural network for prediction. Since the neural network itself has a strong generalization ability, it can be quickly and effectively used in two different frequency bands and different wireless network environments through the training in step S2. The prediction of the spectrum occupancy state is more practical and applicable than the traditional method.

Embodiment two:

In this embodiment, the prediction method of the spectrum occupancy state based on the neural network includes the following steps:

Step S0: Collect {b _i , b _i-1 , . . . , b _im } and h _i+1 . The values of i and m determine the set length. In a specific implementation process, the value of m is usually a positive integer of 4-6.

Step S1: Construct a neural network and use a pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

Step S2: Input each element in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output result of the neural network with the representation variable h used to characterize the occupancy state of the i+1th timing preset frequency band _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th timing preset frequency band, i is a positive integer, and m is a positive integer less than i;

Step S3: Simultaneously input each element in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the neural network outputs a predictive characterization variable representing the occupancy state of the i+2th time-series preset frequency band H _i+2 , judging the occupancy state of the preset frequency band at the i+2 time sequence according to the predicted characteristic variable H _i+2 .

Due to the needs of step S2, the {b _i+1 , b _i , ..., b _i-m+1 } and h _i+1 can be input by peripherals or can be collected by themselves, but in this embodiment The method of self-collection, the specific collection method can be actual measurement or empirical data.

Embodiment three:

As shown in FIG. 1 , on the basis of the previous embodiment, this embodiment implements a method for predicting spectrum occupancy status based on a neural network, and further includes step S0 between the step S0 and the step S1 The step of normalizing the data collected in the process. Through normalization processing, the input data units are unified and within the processing range of the neural network, thereby improving the calculation speed of the neural network and shortening the calculation time.

In the specific implementation process, the initial parameters and the prediction parameters include the number of neurons, the activation function of each neuron, the learning rate, the maximum number of iterations, the preset error, and the weight coefficient parameters between each neuron layer .

As a further optimization of this embodiment, in the step S3, the initial parameters are corrected with a backpropagation algorithm according to the comparison result. The backpropagation algorithm is used to correct the initial parameters, which is simple and fast.

Embodiment four:

In this embodiment, the prediction method of the spectrum occupancy state based on the neural network includes the following steps:

Step S1: Construct a neural network and use a pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

Step S2: Input each element in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output result of the neural network with the representation variable h used to characterize the occupancy state of the i+1th timing preset frequency band _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th timing preset frequency band, i is a positive integer, and m is a positive integer less than i;

Step S3: Simultaneously input each element in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the neural network outputs a predictive characterization variable representing the occupancy state of the i+2th time-series preset frequency band H _i+2 , using a threshold determination method to determine the occupancy state of the time series of the preset frequency band i+2 corresponding to the predictive characterization variable H _i+2 .

In the specific implementation process, other judgment methods such as function judgment can be used, and the threshold judgment method has the advantages of simple implementation, so that the prediction method of the spectrum occupancy state based on the neural network described in this embodiment has the advantages of accurate prediction value and low calculation amount. Small in size and low in computational complexity, it also has the advantages of simple and clear judgment and easy implementation.

The neural network-based spectrum occupancy state prediction device of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

First embodiment:

The apparatus for predicting the spectrum occupancy state based on the neural network in this embodiment includes:

The prediction device of the spectrum occupancy state based on the neural network comprises:

The neural network building block is used to construct the neural network and use the pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

The neural network training module is used to input the elements in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output results of the neural network to represent the occupancy state of the i+1th timing preset frequency band The characteristic variable h _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th time series preset frequency band, i is a positive integer, and m is a positive integer smaller than i;

The spectrum occupancy state prediction module is used to input each element in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the output of the neural network represents the occupancy state of the i+2th timing preset frequency band The prediction characteristic variable H _{i +2 is used to determine the occupancy state of the i+2 time sequence preset frequency band according to the prediction characteristic variable H i+2} .

In this embodiment, the device for spectrum occupancy state prediction realizes spectrum prediction through neural network building blocks, neural network training modules, and spectrum occupancy state prediction modules. Compared with traditional devices, it has low computational complexity and low computational load The advantage of small, accurate predictions.

Second embodiment:

As shown in Figure 2, the prediction device of the spectrum occupancy state based on the neural network in this embodiment includes:

A data collection module, used to collect {b _i , b _i-1 ,..., b _im } and h _i+1 ;

The neural network building block is used to construct the neural network and use the pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

The neural network training module is used to input the elements in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output results of the neural network to represent the occupancy state of the i+1th timing preset frequency band The characteristic variable h _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th time series preset frequency band, i is a positive integer, and m is a positive integer smaller than i;

The spectrum occupancy state prediction module is used to simultaneously input the elements in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the output of the neural network represents the occupancy of the i+2th timing preset frequency band The predictive characterization variable H _i+2 of the state is used to determine the occupancy state of the preset frequency band at the i+2 time sequence according to the predictive characterization variable H _i+2 .

Based on the neural network-based spectrum occupancy state prediction device in this embodiment, a data collection module is added on the basis of the previous embodiment to collect required information, so that data collection can be realized without the help of peripheral devices.

Third embodiment:

On the basis of the previous embodiment, this embodiment further specifies the data collection module, and the data collection module also includes a normalization sub-module for normalizing the collected data . The collected data can be normalized through the setting of the normalization sub-module, so that the prediction efficiency and prediction effect of the device are better.

Fourth embodiment:

The apparatus for predicting the spectrum occupancy state based on the neural network in this embodiment includes:

A data collection module, used to collect {b _i , b _i-1 ,..., b _im } and h _i+1 ;

The neural network building block is used to construct the neural network and use the pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron;

The neural network training module is used to input the elements in the set {b _i , b _i-1 ,..., b _im } to the neural network, and compare the output results of the neural network to represent the occupancy state of the i+1th timing preset frequency band The characteristic variable h _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th time series preset frequency band, i is a positive integer, and m is a positive integer smaller than i; the initial parameter And the prediction parameters at least include weight coefficients between two adjacent layers of the neural network; the neural network training module includes a parameter correction sub-module for correcting initial parameters with a backpropagation algorithm according to the comparison result.

The spectrum occupancy state prediction module is used to simultaneously input the elements in the set {b _i+1 , b _i ,..., b _i-m+1 } to the neural network, and the output of the neural network represents the occupancy of the i+2th timing preset frequency band The predicted characteristic variable H _i+2 of the state, according to the predicted characteristic variable H _{i+2 ,} determines the occupancy state of the i+2 time sequence preset frequency band; A submodule for judging the occupancy state of the preset frequency band corresponding to time i+2.

In summary, the neural network-based spectrum occupancy state prediction method and device of the present invention can directly predict the occupancy of the required frequency band without obtaining the current spectrum distribution state information when predicting spectrum occupancy information. state, and the neural network is used to predict the mapping relationship between the intensity signal of the spectrum corresponding to the frequency band and the occupancy state, which has the advantages of small amount of calculation, low computational complexity and accurate prediction results.

The above embodiments are only used to illustrate the present invention, but not to limit the present invention. Those of ordinary skill in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all Equivalent technical solutions also belong to the category of the present invention, and the scope of patent protection of the present invention should be defined by the claims.

Claims (10)

1. a prediction method based on the spectrum occupancy state of neural network, it is characterized in that, the prediction method of the spectrum occupancy state based on neural network comprises the following steps successively: Step S1: Construct a neural network and use a pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron; Step S2: Input each element in the set {b _i , b _i-1 ,...,b _im } to the neural network, and compare the output result of the neural network with the representation variable h used to characterize the occupancy state of the i+1th timing preset frequency band _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th timing preset frequency band, i is a positive integer, and m is a positive integer less than i; Step S3: Simultaneously input each element in the set {b _i+1 , b _i ,...,b _i-m+1 } to the neural network, and the neural network outputs a predictive characterization variable representing the occupancy state of the i+2th timing preset frequency band H _i+2 , judging the occupancy state of the i+2 timing preset frequency band according to the predicted characteristic variable H _i+2 ; Wherein, said modifying the initial parameter according to the comparison result to obtain the preset parameter includes: if the error between the output result of the neural network and the characteristic variable h _i+1 representing the occupancy state of the i+1th time sequence preset frequency band is smaller than the preset error, then The initial parameter is the final prediction parameter; If the error between the output of the neural network and the characteristic variable h _i+1 representing the occupancy state of the preset frequency band at the i+1th time sequence is greater than the preset error, then input the set {b _i ,b _i-1 ,…,b _im }, and compare the output result of the neural network with the characterization variable h _i+1 representing the occupancy state of the preset frequency band in the i+1th time sequence, and then modify the initial parameters according to the comparison results to obtain the prediction parameters, and finally make the neural network The error between the output result of and the characteristic variable h _i+1 representing the occupancy state of the preset frequency band at the i+1th time sequence is smaller than the preset error. 2. The prediction method of the spectrum occupancy state based on neural network according to claim 1, is characterized in that, also comprises step S0 before described step S1; The step S0: collect {bi, bi _-1 , ..., bi _im } and h _{i +1} . 3. The prediction method of the spectrum occupancy state based on neural network according to claim 2, characterized in that, between said step S0 and said step S1, also includes normalizing the data collected in said step S0 processing steps. 4. The prediction method based on the spectrum occupancy state of neural network according to claim 1, 2 or 3, characterized in that, the initial parameters and the predicted parameters at least comprise the weight between the adjacent two layers of the neural network coefficient; In the step S2, the initial parameters are corrected with a backpropagation algorithm according to the comparison result. 5. according to the prediction method of the spectrum occupancy state based on neural network described in claim 1,2 or 3, it is characterized in that, in the described step S3, determine the corresponding i+2 of described H _i+2 by the threshold judgment method Occupancy status of the preset frequency band at all times. 6. A neural network-based prediction device for a spectrum occupancy state, characterized in that the neural network-based prediction device for a spectrum occupancy state comprises: The neural network building block is used to construct the neural network and use the pruning algorithm to determine the initial parameters of each layer of the neural network and each neuron; The neural network training module is used to input each element in the set {b _i , b _i-1 ,...,b _im } to the neural network, and compare the output result of the neural network to represent the occupancy state of the i+1th timing preset frequency band The characteristic variable h _i+1 is compared; according to the comparison result, the initial parameter is corrected to obtain the predicted parameter; b _i is the signal strength of the i-th time series preset frequency band, i is a positive integer, and m is a positive integer smaller than i; The spectrum occupancy state prediction module is used to input each element in the set {b _i+1 , b _i ,...,b _i-m+1 } to the neural network at the same time, and the output of the neural network represents the occupancy of the i+2th timing preset frequency band The predictive characterization variable H _i+2 of the state, according to the predictive characterization variable H _i+2 , determines the occupancy state of the i+2 time sequence preset frequency band; Wherein, in the neural network training module, the correction of the initial parameters according to the comparison result to obtain the preset parameters includes: if the output result of the neural network and the characteristic variable h _i+1 representing the occupancy state of the preset frequency band at the i+1th time sequence If the error is less than the preset error, the initial parameter is the final prediction parameter; If the error between the output of the neural network and the characteristic variable h _i+1 representing the occupancy state of the preset frequency band at the i+1th time sequence is greater than the preset error, then input the set {b _i ,b _i-1 ,…,b _im }, and compare the output result of the neural network with the characterization variable h _i+1 representing the occupancy state of the preset frequency band in the i+1th time sequence, and then modify the initial parameters according to the comparison results to obtain the prediction parameters, and finally make the neural network The error between the output result of and the characteristic variable h _i+1 representing the occupancy state of the preset frequency band at the i+1th time sequence is smaller than the preset error. 7. The prediction device of the spectrum occupancy state based on the neural network according to claim 6, characterized in that, the prediction device of the spectrum occupancy state based on the neural network also includes collecting { _bi , bi _-1 ,..., b _im } and h _i+1 data acquisition module. 8 . The neural network-based prediction device for spectrum occupancy status according to claim 7 , wherein the data acquisition module further comprises a normalization sub-module for normalizing the collected data. 9 . 9. The prediction device based on the spectrum occupancy state of the neural network according to any one of claims 6-7, wherein the initial parameters and the prediction parameters at least include the weight between two adjacent layers of the neural network coefficient; The neural network training module includes a parameter modification sub-module for modifying initial parameters with a backpropagation algorithm according to the comparison result. 10. The prediction device based on the spectrum occupancy state of neural network according to any one of claims 6-7, characterized in that, the spectrum occupancy state prediction module includes a method for determining the corresponding H _i+2 by a threshold determination method A submodule for judging the occupancy state of the preset frequency band at time i+2.