Mobile medical internet of things spectrum intelligent prediction algorithm based on DB-LSTM neural network

Abstract

The application discloses a mobile medical Internet of Things spectrum intelligent prediction algorithm based on a DB-LSTM neural network, and comprises the following steps: firstly, sampling local mobile medical Internet of Things signal data, performing spectrum sensing operation by using a ResNet-CNN model, and obtaining a main user historical use spectrum time sequence binary data set; then, constructing a spectrum prediction data set with time sequence characteristics by using a sliding window, training a DB-LSTM neural network model by using the data set, and realizing spectrum prediction by using the trained DB-LSTM neural network model. The method can effectively improve spectrum prediction accuracy and reduce spectrum prediction mean square error.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a mobile medical Internet of Things spectrum intelligent prediction algorithm based on a DB-LSTM neural network. BACKGROUND

[0002] According to the remote medical multi-parameter health monitoring system, various wireless biological sensor devices continuously monitor and collect various index information of the patient's body, so that medical staff can timely master the information without contacting the patient, better manage and treat, and lay a foundation for medical information and resource sharing. The medical Internet of Things has been integrated with life, and medical consultation, real-time monitoring, remote medical treatment, real-time information interaction between doctors and patients, expert remote consultation to guide patient treatment schemes, etc.

[0003] A. U. Haq et al. proposed an early breast cancer diagnosis method based on a convolutional neural network (CNN). The cross-validation method combines the best model selection and super parameter adjustment, and good results are obtained. Y.-S. Su et al. observed the relative relationship between blood reflux and skin temperature by using the data obtained by pulse medical Internet of Things sensor equipment, realized a screening system for membrane flap heart disease, and facilitated self-examination of patients and timely discovery of problems. L. Jia et al. proposed an intelligent electronic stethoscope based on a piezoelectric micromechanical ultrasonic transducer (PMUT) array for the Internet of Medical Things (IoMT). Remote medical treatment and digital intelligent medical treatment are realized. Feasibility tests are carried out on patients with existing diseases. S. Y. Siddiqui et al. carried out intelligent breast cancer staging prediction based on medical Internet of Things through medical data images, further researched the development of diseases, and obtained relatively accurate results. H. Wang et al. proposed a safety improvement framework for medical Internet of Things in the field of minimally invasive surgery, which is used to assist surgeons in implementing minimally invasive surgery and guide the next stage of the surgery process, reduce peripheral complications caused by the non-deterministic nature of the surgery workflow, and improve the safety of the surgery. Scholars play an important role in the medical Internet of Things from medical consultation, health monitoring, disease diagnosis, prediction, and corresponding disease surgery treatment, and operate in all directions. However, with the rapid development of the Internet of Things technology in the medical field, while providing convenient services for doctors and medical users, more and more medical devices are continuously connected, the data volume transmitted by the medical Internet of Things is increasing, and the frequency of using wireless networks is increasing, so that the contradiction between supply and demand of spectrum resources becomes increasingly serious, and the medical devices cannot be better served.

[0004] The traditional static spectrum allocation strategy only allows specific users to use, which cannot meet the growing demand for spectrum. How to efficiently use spectrum resources is an important problem. Cognitive radio network (CRN) gives secondary users (SU) the right to dynamically use PU idle time slot spectrum resources for communication without interfering with primary users (PU), which improves spectrum resource utilization.

[0005] Z. Chen et al. proposed a STFT-CNN spectrum sensing algorithm using time-frequency domain information of the signal, without any prior information of the primary user. S. Lin et al. used intelligent reflecting surface to enhance PU signal technology, and proposed a weighted energy detection method. S. Surekha et al. proposed an ALOHA spectrum sensing method to improve the throughput of medical sensor network, which reduces the interference to the primary user. A. Moawad et al. proposed a cepstrum covariance detector, which effectively overcomes the situation of poor detection under low signal-to-noise ratio.

[0006] But spectrum sensing needs to be perceived in each time slot, susceptible to noise interference, consume a lot of resources, cause certain sensing delay, not conducive to data transmission. Spectrum prediction technology in cognitive radio uses channel state information fed back to secondary users by spectrum sensing to predict the rules of PU communication by mining the internal correlation of historical channel state data. Combined with spectrum sensing, it reduces the energy and time loss of spectrum sensing and reduces the error of received signal. Therefore, it is particularly important to explore an accurate and effective spectrum prediction algorithm. In order to quickly and efficiently obtain the channel state of PU, ensure efficient and accurate collection and transmission of medical data, spectrum prediction algorithm has been widely studied by scholars. R. Ahmed et al. proposed a new sensing and communication integrated framework, called CR assisted IoV network (CRAV Net), which effectively predicts the spectrum occupancy in the main spectrum without prior knowledge of the radio environment. Z. Cong et al. designed a switching attention module (SAM) that adaptively weights the received signal, solving the problem of spectrum sensing with random arrival and departure of main signals. Compared with the detection performance of existing detectors, the detection performance is greatly improved. S. Luo et al. proposed a prediction method based on HT-HMM for unmanned aerial vehicle communication, which achieved good prediction effect. In the face of complex signals in wireless channels affected by external factors, the lack of data, time series characteristics and the increase in scale easily cause the limitation of accuracy and effectiveness of traditional machine learning methods and the increase of convergence time. X. Li et al. proposed a spatiotemporal spectrum prediction network based on three time characteristics of spectrum data, which significantly improved the prediction performance. J. Sun et al. proposed a prediction idea based on multiple "images" composed of historical spectrum data, realizing a long-term spectrum prediction scheme based on tensor completion (LSP-TC). In order to reduce the complexity of the spectrum prediction network model, N. Radhakrishnan et al. proposed a "Kandeepan Niranjana (KN) initialization method, which significantly improved the training convergence of the spectrum prediction model based on LSTM. R. Ding et al. proposed a radio frequency machine learning driven spectrum prediction scheme, which uses a unit with only two parameters in each layer of the AR network to speed up network training.

[0007] The spectrum sensing method in the above research is susceptible to noise interference and the feature of the perceived signal is fuzzy. The traditional prediction model of spectrum prediction technology is rough and the reverse characteristics of time series data are less considered, which fails to fully mine the internal correlation. SUMMARY

[0008] The technical problem to be solved by the present application is how to provide a method capable of effectively improving the spectrum prediction accuracy and reducing the mean square error of spectrum prediction.

[0009] To solve the above technical problems, the technical scheme adopted by the present application is: a mobile medical Internet of Things spectrum intelligent prediction algorithm based on a DB-LSTM neural network, characterized by comprising the following steps:

[0010] First, the local mobile medical Internet of Things signal data is sampled, and a ResNet-CNN model is used for spectrum sensing operation to obtain a main user historical use spectrum time series binary data set;

[0011] Then, a sliding window is used to construct a spectrum prediction data set with time sequence characteristics, and the DB-LSTM neural network model is trained using this data set, and the trained DB-LSTM neural network model is used to realize spectrum prediction.

[0012] The beneficial effects produced by the above technical scheme are that the method is aimed at the problems of complex noise in the mobile medical Internet of Things communication environment, wireless signal susceptible to noise interference, and low signal feature quality. A neural network structure based on ResNet-CNN combination is proposed. Two layers of residual shortcut connection module structure are used to alleviate the gradient problem to some extent in the network training process, effectively reduce the loss of data information transmission, and ensure the integrity of the data information. After convolution operation, batch normalization and ReLU function activation are used to enhance the ability to extract wireless signal data features, making the network structure more robust. CNN uses three layers of convolution operation without pooling layer to extract data features while reducing the number of neural network parameters, reducing the loss of important features of wireless signal data, and ensuring the original signal feature information as much as possible.

[0013] The sliding window method is used to construct the spectrum prediction data set, ensuring the time sequence characteristics of the data while covering more data points, and fully utilizing the data. A DB-LSTM neural network is proposed. The LSTM branch structure contains 128 memory units and is activated by a tanh function. The forward internal correlation of the time series data is updated and extracted at each time step, and the Bi-LSTM branch is equivalent to two layers of LSTM containing 64 memory units. The extracted features at each time step are activated by a tanh function, and the forward and backward directions are performed simultaneously. The reverse characteristics of the time series data are extracted, and the ability to extract data features in the forward direction is enhanced. DB-LSTM fully excavates the internal correlation characteristics of the forward and reverse time series of the time series data, and predicts the internal regularity of the authorized user's use of the spectrum.

[0014] Through simulation experiments, compared with ELM, BP and CNN neural network algorithms, the spectrum prediction accuracy of the algorithm is increased by 2%, 3% and 8% respectively, and the mean square error is reduced by 9%, 44% and 41% respectively, indicating that the algorithm has obvious advantages in both spectrum prediction accuracy and mean square error reduction. BRIEF DESCRIPTION OF DRAWINGS

[0015] The application will be described in further detail below with reference to the drawings and specific embodiments.

[0016] Figure 1 is a flow chart of a spectrum prediction algorithm in an embodiment of the application;

[0017] Figure 2 is a model of a spectrum prediction cognitive communication system in an embodiment of the application;

[0018] Figure 3 is a result diagram of a ResNet-CNN network in an embodiment of the application;

[0019] Figure 4 is a specific structure diagram of a CNN network in an embodiment of the application;

[0020] Figure 5 is a module structure diagram of a ResNet network in an embodiment of the application;

[0021] Figure 6 is a sliding window diagram in an embodiment of the application;

[0022] Figure 7 is a data set diagram in an embodiment of the application;

[0023] Figure 8 is a structure diagram of an LSTM unit in an embodiment of the application;

[0024] Figure 9 is a network structure diagram of a Bi-LSTM in an embodiment of the application;

[0025] Figure 10 is a DB-LSTM neural network diagram in an embodiment of the application;

[0026] Figure 11 is a spectrum sensing accuracy comparison diagram under different signal-to-noise ratios in an embodiment of the application;

[0027] Figure 12 is a spectrum sensing false alarm probability diagram under different signal-to-noise ratios in an embodiment of the application;

[0028] Figure 13 is an ELM and DB-LSTM test set error diagram in an embodiment of the application;

[0029] Figure 14 is an ELM and DB-LSTM set data mean square error diagram in an embodiment of the application;

[0030] Figure 15 is a BP and DB-LSTM test set error diagram in an embodiment of the application;

[0031] Figure 16is a BP and DB-LSTM ensemble data mean square error graph in the embodiment of the application;

[0032] Figure 17 is a CNN and DB-LSTM test set error graph in the embodiment of the application;

[0033] Figure 18 is a CNN and DB-LSTM ensemble data mean square error graph in the embodiment of the application;

[0034] Figure 19 is a prediction accuracy graph under different algorithms in the embodiment of the application;

[0035] Figure 20 is a mean square error graph under different communication intensities in the embodiment of the application; DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the application.

[0037] In the following description, many specific details are set forth in order to provide a thorough understanding of the application. However, the application can be practiced without the specific details, which are not described in the present application, and it can be apparent to those skilled in the art that the application can be implemented in other different ways, and therefore the application is not limited to the specific embodiments disclosed below.

[0038] Generally, as shown in Figure 1 The embodiment of the application discloses a mobile medical Internet of Things spectrum intelligent prediction algorithm based on a DB-LSTM neural network, which comprises the following steps:

[0039] First, the local mobile medical Internet of Things signal data is sampled, and a ResNet-CNN model is used for spectrum sensing operation to obtain a primary user historical use spectrum time sequence binary data set;

[0040] Then, a sliding window is used to construct a spectrum prediction data set with time sequence characteristics, a DB-LSTM neural network model is trained by using the data set, and the trained DB-LSTM neural network model is used to realize spectrum prediction.

[0041] The above steps will be described in detail below in combination with specific contents

[0042] System model

[0043] The spectrum prediction cognitive system model constructed in the application is as shown in Figure 2As shown, the primary user can unconditionally use the spectrum resource for communication, and in the primary user network, the secondary user obtains the history timing binary data of the primary user using the spectrum through spectrum sensing, and the specific process is shown as follows:

[0044] H0:x(t)=n(t) (1)

[0045] H1:x(t)=h(t)*s(t)+n(t) (2)

[0046] In the formula, H0 and H1 represent available spectrum resources and no available spectrum resources respectively, x(t) is a set of discrete signal sequences received by the receiving end, n(t) is Gaussian white noise under the condition of standard normal distribution, the channel gain generated in the N-Nakagami channel is h(t), and the discrete signal data set transmitted by the primary user transmitter is represented by s(t). The operator * represents convolution operation. The secondary user judges whether there is available idle spectrum resource through spectrum sensing.

[0047] Through the spectrum sensing operation based on the ResNet-CNN spectrum intelligent sensing model, spectrum state binary data with timing characteristics can be obtained, reflecting the overall timing change process of the primary user using the spectrum resource, providing clearer and more accurate spectrum sensing data for spectrum prediction, and effectively improving the prediction effect of spectrum prediction.

[0048] The spectrum prediction data set is constructed by using the sliding window, and the historical data is trained and analyzed by the DB-LSTM neural network spectrum prediction model to realize the prediction of the spectrum state. According to the combination of sensing and prediction information, the secondary user can use the idle spectrum resource for communication without affecting the communication of the primary user.

[0049] Spectrum sensing based on ResNet-CNN model

[0050] According to the characteristics of small signal data dimension, the pooling layer in the neural network is discarded, so that the transmission of feature information is more complete. After the receiving end obtains the discrete signal data set, the data is first normalized and dimensionally transformed. The ResNet residual block and CNN network structure are used to further extract data features, so that the features input into the full connection network are more sufficient and accurate, and the final output spectrum state classification result is more accurate.

[0051] Data preprocessing

[0052] After obtaining the discrete signal data set, the normalization operation is performed to eliminate the influence of bad data on the sensing result. The removal of data unit limitation facilitates the comparison and weighting of data with different units or orders of magnitude, and is also conducive to the training of neural network. The normalization formula is shown as follows:

[0053]

[0054] Where X is the original data, X is the normalized data, Max and Min are the maximum and minimum values of the original data, respectively.

[0055] The neural network in this paper uses the ReLU activation function:

[0056]

[0057] Where c represents the feature map output after the convolution operation is completed.

[0058] The fully connected layer uses the Softmax classifier:

[0059]

[0060] In formula (5), z i is the output value of the i-th node, J is the number of output nodes, the output value of multi-classification is converted into a probability distribution with a range of [0, 1] and a sum of 1 through the Softmax function, and the class of the maximum value is classified.

[0061] ResNet-CNN network module

[0062] ResNet-CNN combined neural network

[0063] The overall structure of the ResNet-CNN spectrum sensing model includes three layers of convolutional neural networks, two layers of residual modules, and three layers of fully connected neural networks. The specific network structure is shown in Figure 3 .

[0064] Improved CNN network structure

[0065] The convolutional neural network can have a small amount of calculation and stable learning features due to the sharing of convolution kernel parameters in the hidden layer and the sparsity of interlayer connections. Since the number of extracted signal features is small and is a one-dimensional vector, it is not suitable for traditional convolutional neural networks. According to the characteristics of the small input value dimension, the pooling layer used to reduce the network parameter dimension is removed, avoiding the loss of important features of the signal data. The feature map of the previous convolution layer is convolved with the convolution kernel of the current layer to produce the output of the current layer feature map through the corresponding activation function. The convolution operation process is shown in formula (6):

[0066]

[0067] In the formula, x (L) is the output feature map of the current layer convolution, x (L-1) is the output feature map of the upper layer convolution, W (L) , and b (L)are the convolution kernel and bias value of the current layer respectively, and f(·) is the ReLU activation function. The output formula of a two-dimensional convolution of a certain hidden layer is shown in equation (7):

[0068]

[0069] In equation (7) represents the jth feature map of the output of the ith hidden layer after convolution, is the kth channel feature map of the (i-1)th layer, is the kth weight of the jth feature map of the ith layer, is the bias term corresponding to the ith layer.

[0070] ResNet network module structure

[0071] The skip shortcut branch structure in the residual neural network module can effectively alleviate the gradient problem of the neural network in the training process. When the residual value of the residual module is 0, it can be understood that the current layer of neurons is a linear stack of the output features of the previous layer, and the overall performance of the network will not decrease. In actual training, the residual value cannot always be 0, so the depth of the neural network can be further increased. At the same time, it can reduce the loss of network information transmission to a certain extent, ensure the integrity of the data information, and enhance the ability to extract data features. Only the difference between the input and the output needs to be learned, which simplifies the learning goal and difficulty. The residual learning module calculation formula is shown in equation (8):

[0072] x l+1 =x l +F(x l ,w l ) (8)

[0073] In equation (8), x l+1 is the output of the current residual block, x l is the output of the previous layer, F(x l ,w l ) = x l+1 -x l is the residual function, and w l is the corresponding weight. If F(x l ,w l )→0, it is equivalent to an identity mapping. L layers can be represented by any I layer with fewer layers than L and the residual part between them, as shown in equation (9):

[0074]

[0075] If it is derived from the 0th layer, the Lth layer can represent the accumulation of the features extracted by the residual blocks of each layer, as shown in equation (10):

[0076]

[0077] The optimization is performed using a cross-entropy loss function, as shown in equation (11):

[0078]

[0079] y is a label value, is a classification value. The loss function C for a certain x l The gradient of the loss function C with respect to x is shown in equation (12):

[0080]

[0081] Because It is not always 0, so it can solve the problem of gradient disappearance to some extent. The specific structure of the ResNet network module is shown in Figure 5

[0082] Spectrum prediction algorithm based on DB-LSTM neural network

[0083] The DB-LSTM neural network spectrum prediction algorithm proposed in the present application includes two main parts: first, sampling the local signal data, using the ResNet-CNN model for spectrum sensing operation, and obtaining the main user historical usage spectrum time series binary data set. Then, through the sliding window, a spectrum prediction data set with time series characteristics is constructed, and the DB-LSTM network model is trained using this data set to realize fast and accurate spectrum prediction.

[0084] Acquisition of signal data

[0085] The M / Geo / 1 queuing system model is used to simulate the internal law of the actual authorized user using the channel on the MATLAB simulation platform to model the wireless channel, and generate the wireless signal data set. In the M / Geo / 1 model, M represents the entire input process of the arrival of authorized users subject to a Poisson process with parameter λ, Geo represents that the authorized user uses the channel each time according to a geometric distribution, and the average channel usage time of the authorized user is set to μ; 1 represents a single channel.

[0086] The mathematical expression of the Poisson distribution is shown in equation (13):

[0087]

[0088] In the formula, represents the probability distribution of the number of authorized users arriving at any time t.

[0089] The mathematical expression of the geometric distribution is shown in equation (14):

[0090] P(x=k)=(1-μ) k-1 μ,k=1,2,3...N (14) ​

[0091] where T represents the time duration that the licensed user occupies the channel each time.

[0092] In order to obtain the state of the licensed user using the channel in different communication scenarios, the application changes the communication intensity p and the average arrival time t inter of the licensed user to achieve the relationship expression as shown in (15):

[0093]

[0094] where t serv is the average time of the licensed user occupying the channel for communication.

[0095] Sliding window construction dataset

[0096] According to the time sequence characteristics of the time sequence signal data, the order cannot be adjusted, disordered, and randomly disturbed. A sliding window method is used to construct a spectrum prediction dataset. Different sizes of sliding windows are set to meet the requirements of historical spectrum sensing data input in different time slots. The time sequence adjacent and accurate label dataset can be obtained by sequentially sliding forward. That is, the time sequence characteristics of the time sequence data are ensured, and more data points are covered. The specific content is shown in Figure 6 .

[0097] Figure 6 In the formula, it is assumed that the size of each sliding window is 5, and the corresponding modification is made according to the needs. S j represents the jth sliding, j takes 1, 2, …, N, and each value in the window represents the channel spectrum state obtained by spectrum sensing in a time slot. L j represents the label value corresponding to the S j th sliding window data. t represents time, that is, how many time slots. The constructed spectrum prediction dataset is shown in Figure 7 .

[0098] Figure 7 In the formula, the dataset constructed by the application is shown in the figure, each row is data plus label, and constitutes a group of training samples. Before inputting into the neural network, it is converted into the required dimension of the neural network model input through dimension conversion.

[0099] DB-LSTM neural network model spectrum prediction

[0100] 1) LSTM neural network

[0101] The LSTM neural network effectively solves the gradient problem in the recurrent neural network (RNN). A component called a memory unit is introduced in the network structure, and the memory module is expanded to replace the hidden nodes in the network, which has a significant advantage in processing problems with time characteristics, long intervals, and long time delays. The binary spectrum state data obtained by spectrum sensing can be regarded as time sequence data, and it is reasonable and feasible to use the LSTM neural network for spectrum prediction. The specific network unit structure used in the present application is as shown in Figure 8

[0102] Figure 8 In the LSTM neural network unit structure, three threshold structures are introduced: input gate, forget gate, and output gate; the input gate determines which information in the current input stream is important and needs to be added to the memory unit. The forget gate determines which related information in the previous memory unit needs to be retained to further reduce the load of the network. The output gate determines the value of the hidden state at the next moment. The cell state C t representing long-term memory is introduced; the candidate state σ represents the Sigmoid function, represents element multiplication operation, represents element addition operation, and tanh represents the tanh function; the three thresholds are functions of the input features x t at the current moment and the short-term memory h t-1 at the previous moment, and the formulas are respectively:

[0103] Input gate:

[0104]

[0105] i t =σ(W i ·[h t-1 ,x t ]+b i ) (17)

[0106]

[0107] Forget gate:

[0108] f t =σ(W f ·[h t-1 ,x t ]+b f ) (19)

[0109] Output gate:

[0110] O t =σ(W o ·[h​t-1 ,x t ]+b o (20)

[0111]

[0112] 2) Bi-LSTM Neural Network

[0113] Bi-LSTM can simultaneously utilize the inherent correlations contained in past and future time slots of a time series. A single-layer Bi-LSTM neural network can be viewed as a combination of two LSTM neural networks: one layer processes the time series forward, and the other layer processes it backward. After all time steps have been processed, the forward and backward data state features extracted by the two LSTM layers are concatenated to form the final data output structure features. The network structure diagram used in this application is shown below. Figure 9 As shown:

[0114] The hidden state h of Bi-LSTM at time t t Including the hidden state of the forward LSTM Hidden state of inverse LSTM

[0115]

[0116]

[0117]

[0118] In the formula, T represents the length of the time sequence.

[0119] 3) DB-LSTM Neural Network

[0120] After constructing the spectrum prediction dataset from the time-series binary result sequence obtained by spectrum intelligent sensing, the dataset is first transformed in terms of feature dimension to meet the input dimension requirements of a recurrent neural network. In the input network structure, the network structure and parameters are adjusted to find the optimal configuration. Based on the characteristics of the time-series data of signal sensing results, a DB-LSTM neural network is proposed to address the features of both forward and backward data. This fully extracts historical and future information from the time-series data, improving the accuracy of the network model's predictions and the efficiency of network training. The spectrum prediction network model structure in this application is as follows: Figure 10 As shown;

[0121] Depend on Figure 10It can be seen that after the DB-LSTM spectrum prediction network model is trained to obtain the optimal network structure, there is a certain error between the predicted value g of the given target input neural network and the actual expected value label, and the output value cannot be immediately determined as the channel state, so a binary decision threshold is added after the final output, and the threshold value is set as λ=0.5.

[0122]

[0123] Experimental analysis

[0124] Based on the M / Geo / 1 queuing model under the N-Nakagami channel, the application simulates to generate a signal data set with a signal-to-noise ratio of-15dB-0dB, an interval of 3dB and added Gaussian white noise under the 2FSK modulation mode. 11220 groups of signal data are generated under each signal-to-noise ratio. The first 1-2000 groups of continuous data are taken as the training set, the 2000-3000 groups are taken as the test set, and the subsequent 8220 groups of continuous data are used for comparative experiments. The sensing accuracy and false alarm rate of the neural network spectrum sensing model are compared. The simulation parameters are shown in Table 1.

[0125] Table 1-Simulation parameters

[0126]

[0127] Spectrum sensing performance comparison:

[0128] Figure 11 、 12 As can be seen from the above table, the spectrum sensing algorithm proposed in the application has better overall performance in the test set compared to the LVQ, Elman and CNN algorithms, and both the spectrum sensing accuracy and the false alarm probability of spectrum sensing are better. The spectrum sensing accuracy is improved by 3.8%, 30.9% and 1% respectively at-9dB. It shows that the ResNet-CNN neural network can accurately and fully extract signal features from noise interference, and obtain accurate and efficient spectrum sensing results, and has good spectrum sensing ability.

[0129] Spectrum prediction performance comparison

[0130] The application selects 8000 groups of continuous data from the channel historical state data obtained by spectrum sensing, and sets the sliding window size to 4, wherein each group of data contains 4 historical spectrum states as target input and 1 label value as prediction model output value. The first 1-1800 groups of data are taken as training data, the 1900-2000 groups of data are used for prediction experiment, and the 2000-8000 groups of data are set to 100 sets, each set containing 60 groups of data for set prediction experiment. The error and mean square error of the prediction results of the neural network prediction model and the actual value are compared, as shown in Table 2. Figures 13-18 ​

[0131] Figures 13-18 As can be seen in the test set and the set of test set, the performance of the algorithm proposed in the present application is compared with ELM, BP and CNN algorithm, the mean square error is reduced by 9%, 44% and 41% respectively. It shows that the DB-LSTM neural network model makes full use of the historical time series data, discovers the internal regularity of the authorized user using the spectrum, is more suitable for complex and changeable wireless communication environment, and shows good prediction ability.

[0132] By Figure 19 As can be seen, the DB-LSTM neural network model is compared with ELM, BP and CNN algorithm, the mobile spectrum prediction accuracy is increased by 2%, 3% and 8% respectively, which shows that the DB-LSTM neural network model makes full use of the historical spectrum sensing data information, and further improves the spectrum prediction accuracy.

[0133] Figure 20 The mean square error of the three algorithms under different communication intensity ρ is shown in the table, it can be seen that the algorithm proposed in the present application is better than the other three algorithms, when ρ is 0.8, compared with ELM, BP and CNN algorithm, the mean square error is reduced by 10%, 66% and 45% respectively.

[0134] ResNet-CNN spectrum sensing model can effectively overcome noise interference, extract clear sensing signal features, improve the accuracy of spectrum sensing results, and further improve the accuracy of spectrum prediction. The DB-LSTM neural network structure not only can fully extract the reverse internal correlation characteristics of time series data, but also further enhances the forward internal correlation of the extracted main user spectrum use rule, which can effectively reduce the mean square error of spectrum prediction and improve the accuracy of spectrum prediction. The simulation results show that the spectrum prediction performance of the algorithm in the present application is better than ELM, BP and CNN algorithm.

Claims

1. A mobile medical IoT spectrum intelligent prediction algorithm based on DB-LSTM neural network, characterized in that... Includes the following steps: First, local mobile medical IoT signal data is sampled, and spectrum sensing is performed using the ResNet-CNN model to obtain a binary dataset of the main user's historical usage spectrum time series. Then, a spectrum prediction dataset with time-series characteristics is constructed by using a sliding window. This dataset is used to train the DB-LSTM neural network model, and the trained DB-LSTM neural network model is used to achieve spectrum prediction. The DB-LSTM neural network consists of two branch structures, LSTM and Bi-LSTM. The feature information extracted from the Bi-LSTM branch units and the LSTM branch units of different dimensions is concatenated to generate a feature map. Finally, the feature map is fed into a fully connected layer to map the high-dimensional features to the classification space and output the final spectrum prediction result. By setting sliding windows of different sizes, the input requirements for historical spectrum sensing data in different time slots can be met; by sliding forward sequentially, a dataset with adjacent time sequences and accurate labels can be obtained. The overall structure of the ResNet-CNN spectrum sensing model consists of three convolutional neural networks, two residual modules, and three fully connected neural networks. In the CNN network structure: The convolution operation process is shown in formula (4): （4） In the formula This is the output feature map of this convolutional layer. The upper convolutional layer outputs a feature map. , These represent the kernel weights and biases of this layer, respectively. for ReLU Activation function; the output formula of a two-dimensional convolution in a hidden layer is shown in equation (5): （5） In formula (5) Indicates the first i The output of the nth hidden layer after convolution is the first... j Each feature map For the first i-1 The first layer k Each channel feature map It is the first i layer j The first feature map k Each weight, For the first i The bias term corresponding to the layer; In the ResNet network module structure: The calculation formula for the residual learning module is shown in (6): （6） In formula (6), This is the output of the residual block in this layer. This is the output of the previous layer. For the residual function, For the corresponding weights; if This is equivalent to an identity mapping; L A layer can be divided by any layer that is shallower than it. I The residuals between the layers are expressed as in formula (7): （7） If we start the derivation from layer 0, the... L The layer represents the sum of the features extracted from the residual blocks of each layer, as shown in formula (8): （8） Optimization is performed using the cross-entropy loss function, as shown in Equation (9): （9） y For label values, Classification value; loss function C For a certain x l The gradient is shown in equation (10): （10） because Since it cannot always be 0, it can solve the problem of gradient vanishing to some extent.

2. The intelligent spectrum prediction algorithm for mobile medical IoT based on DB-LSTM neural network as described in claim 1, characterized in that, Spectrum sensing operations using the ResNet-CNN model include the following steps: After the receiver acquires the discrete signal dataset, it first performs normalization and dimensionality transformation on the data, and then uses the ResNet residual blocks and CNN network structure in the ResNet-CNN model to further extract data features.

3. The intelligent spectrum prediction algorithm for mobile medical IoT based on DB-LSTM neural network as described in claim 2, characterized in that, Normalizing and transforming the dimensions of data involves the following steps: The normalization formula is shown below: （1） in X The original data, It is normalized data. Max and Min These represent the maximum and minimum values ​​of the original data, respectively. Neural network use ReLU Activation function: （2） in c This represents the feature map output after the convolution operation. Fully connected layer Softmax Classifier: （3） In formula (3) z i For the first i The output value of each node, J To output the number of nodes, use Softmax The function converts the output values ​​of multi-class classification into a probability distribution in the range [0,1] with a sum of 1, and outputs the category of the maximum value.

4. The intelligent spectrum prediction algorithm for mobile medical IoT based on DB-LSTM neural network as described in claim 1, characterized in that, The acquisition of signal data includes the following steps: Using the MATLAB simulation platform M / Geo / 1. The queuing system model simulates the inherent laws of actual authorized users using the channel to model the wireless channel and generate a wireless signal dataset; M / Geo / In model 1, M The entire input process representing the authorized user follows the parameter: λ The Poisson process, Geo The channel usage time of authorized users follows a geometric distribution, and the average channel usage time of authorized users is set as... μ ; 1 indicates single channel; The mathematical expression for the Poisson distribution is shown in equation (11): （11） The formula represents any time. t In the middle, the number of authorized users reached x The probability distribution; The mathematical expression for the geometric distribution is shown in equation (12): （12） In the formula, the duration for which an authorized user occupies the channel each time is represented; To obtain the channel usage status of authorized users under different communication scenarios, the communication strength is adjusted. ρ and average arrival time of authorized users t inter To achieve this, the relational expression is shown in (13): （13） in, t serv The average time for authorized users to occupy the channel for communication.

5. The intelligent spectrum prediction algorithm for mobile medical IoT based on DB-LSTM neural network as described in claim 1, characterized in that: The LSTM neural network unit structure introduces three gate structures: input gate, forget gate, and output gate. The input gate determines which information in the current input stream is important and needs to be added to the memory unit. The forget gate determines which relevant information in the previous memory unit needs to be retained, further reducing the network load. The output gate determines the value of the hidden state at the next moment. Cellular states are introduced to represent long-term memory. C t ; Candidate states awaiting storage in long-term memory ; express Sigmoid function, ⨁ represents element-wise multiplication, and ⨁ represents element-wise addition. tanh represent tanh The function; all three thresholds are input features at the current time step. x t Short-term memory of the previous moment h t-1 The functions are expressed by the following formulas: Input Gate: （14） （15） （16） Forgotten Gate: （17） Output gate: （18） （19）。 6. The intelligent spectrum prediction algorithm for mobile medical IoT based on DB-LSTM neural network as described in claim 5, characterized in that: Bi-LSTM neural networks can simultaneously utilize the inherent correlations contained in past and future time slots of a time series. A single-layer Bi-LSTM neural network can be regarded as a combination of two-layer LSTM neural networks. One layer processes the time series in the forward direction, and the other layer processes the time series in the reverse direction. After all time steps have been processed, the forward and reverse data state features extracted by the two LSTM layers are spliced ​​together to form the final data output structure features. Bi-LSTM in t Hidden state of time Including the hidden state of the forward LSTM Hidden state of inverse LSTM : （20） （21） （22） In the formula, T Indicates the length of the time sequence.

7. The intelligent spectrum prediction algorithm for mobile medical IoT based on DB-LSTM neural network as described in claim 6, characterized in that: After constructing the spectrum prediction dataset from the time-series binary result sequence obtained by spectrum intelligent sensing, the dataset is first transformed in terms of feature dimension to meet the input dimension requirements of the recurrent neural network; in the input network structure, the network structure and parameters are adjusted to seek the optimal. After training, the DB-LSTM spectrum prediction network model obtains its optimal network structure. For a given target input, the neural network predicts the following value: g Compared to actual expected value label Since there is a certain error, the output value cannot immediately determine the channel state. Therefore, a binary decision threshold is added after the final output, and the threshold value is set to... λ= 0.5: （23）。

Images