CN114363218B - A communication reachable rate detection method based on end-to-end learning - Google Patents

Abstract

The invention discloses a communication reachable rate detection method based on end-to-end learning. The steps include: 1. Using neural network model training to calculate logarithmic likelihood ratio (LLR); 2. Accurately calculating the result through gradient descent algorithm 3. According to the required log-likelihood ratio, calculate the information attainable rate based on bit decoding—generalized mutual information (GMI). The invention can greatly improve the calculation efficiency while ensuring the accuracy of the calculation results, thereby improving the real-time calculation of the information attainable rate.

Description

A communication reachable rate detection method based on end-to-end learning

technical field

The invention relates to the field of communication technology, to end-to-end learning and machine learning, and in particular to a generalized mutual information calculation method for calculating logarithmic likelihood ratio LLR based on a neural network model.

Background technique

In recent years, machine learning has been widely used in the field of communication. From coding to channel models, to decoding, and modulation and demodulation, machine learning can be applied to a certain extent and can achieve good results. Using machine learning methods to train some parameters in the neural network can make the training structure tend to be optimal, so that the calculation results are more accurate.

Along with a series of technological revolutions, the development of communication technology has made an important contribution to the growth of Internet capacity. However, the capacity of the communication core network cannot meet the increasing traffic demand of people in the information age. In order to make the information transmission rate close to the Shannon limit, the information theory analysis of the communication system is particularly important.

Achievable information rate (Achievable information rates, AIR) is a performance evaluation index based on information theory information measurement, defined as the maximum amount of information that can be reliably transmitted in a given channel, it can be used to evaluate the maximum achievable transmission rate of the channel, generalized mutual Information, as an achievable information rate based on a bit decoder coded modulation system, has important practical significance in practical engineering applications. However, in the prior art, since the calculation of the generalized mutual information needs to consider many factors, such as the actual channel distribution, the channel dimension, and the mutual interference between bits, etc., the practical application of the generalized mutual information in the prior art is less, and in practical applications The calculation steps of the generalized mutual information in the generalized mutual information are relatively complicated, take a long time, and cannot be absolutely accurate, so there are many applications of mutual information, another achievable information rate that does not consider the relationship between bits, but compared In other words, because the generalized mutual information considers more factors, it has more practical significance in practical engineering applications. Therefore, measuring the generalized mutual information of the channel is of great significance for understanding the ability of the channel to transmit information and evaluating the performance of the channel.

Contents of the invention

In order to solve the shortcomings of the above-mentioned prior art, the present invention proposes a communication reachable rate detection method based on end-to-end learning, in order to ensure the accuracy of the calculation results and greatly improve the calculation efficiency, thereby improving the generalized mutual Real-time computing of information.

The present invention adopts following technical scheme in order to achieve the above-mentioned purpose of the invention:

The characteristics of a communication reachable rate detection method based on end-to-end learning of the present invention include:

Step 1. Define the sending signal sequence of the sending end as s={s ₁ , s ₂ ,...,s _i ,...,s _n }, s _i represents the i-th sending signal, and define the receiving signal sequence of the receiving end as Represents the i-th received signal, i∈[1,n], n represents the sequence length;

Define Receive Signal Sequence Each received signal in is mapped to a bit sequence with a length of m, and the sequence formed by defining the kth bit of the received signal sequence mapping is B _k ={B _k,1 ,B _k,2 ,…,B _k,i , ...,B _1,n }, where, B _k,i represents the received signal /> The kth bit of the mapping, k∈[1,m], B _k,i ∈{0,1};

Step 2, the received signal sequence Each received signal in is divided into two parts, the real part and the imaginary part, and it is recorded as the input signal sequence of the neural network as /> where, /> Indicates the 2nth input value of the input layer, that is, the 0th layer;

Let the node bias vector from the input layer to the hidden layer be denoted as where, /> Represents the bias of the 2nth node of the input layer;

Make the input layer of the neural network, the number of nodes of the hidden layer be 2n, make the number of nodes of the output layer of the neural network be m; The number of layers of the hidden layer is H;

Let the weight matrix of the 2n nodes in the input layer to the 2n nodes in the first hidden layer be denoted as where, /> Represent the weight of the 2nth node in the input layer to the 2nth node in the first hidden layer;

Let the weight matrix of any 2n nodes in the hth hidden layer to the 2n nodes in the h+1th hidden layer be denoted as where, /> Indicates the weight of the 2nth node in the hth hidden layer to the 2nth node in the h+1th hidden layer;

Let the weight matrix from the 2n nodes in the hidden layer of the H layer to the m nodes in the output layer be denoted as where, /> Represent the weight of the 2nth node in the hidden layer of the H layer to the mth node in the output layer, h∈[1,H];

Let the node bias vector in any h-th hidden layer be denoted as where, /> Indicates the bias of the 2nth node in the hth hidden layer;

Let the calculation result sequence of any h-th hidden layer be recorded as where, /> Indicates the calculation result of the 2nth node in the hidden layer of the hth layer;

Let the linear equation of the h-th hidden layer be y ^(h) = ω ^(h) x ^T + b ^(h) , where T represents transposition;

Let the output sequence after activation of the hth hidden layer be where, /> Indicates the 2nth output after activation of the hth hidden layer, and x ^(h) = f(y ^(h) ), f( ) is the activation function;

The output sequence z ^(H) after activation of the hidden layer of the H layer is used as the input sequence of the output layer, and the linear equation y=ω ^(H) (x ^(H) ) ^T + b ^(H) in the output layer The output calculation result sequence is y={y1,y2,...,yn}, where yn represents the calculation result output by the nth node of the output layer; x ^(H) represents the output sequence after the H-th hidden layer is activated, b ^(H) represents the node bias vector in the hidden layer of the Hth layer;

Step 2.1, assuming channel distribution f _Y|X (y|x), using the maximum logarithm approximation method shown in formula (1) to estimate the logarithmic likelihood ratio sequence of the binary mapping bit sequence of n received signals is where, /> Represents the log-likelihood ratio sequence of the kth bit of n received signal mapping /> Represents the log-likelihood ratio of the k-th bit mapped to the n-th received signal, k∈[1,m];

In formula (1), Respectively represent the maximum probability that the kth bit of the n received signal mapping is judged as 1 or 0, f _Y|X (y|x) represents the channel transition probability of the transmitted signal as x and the received signal as y, and X and Y respectively represent Send and receive signal sequences;

Step 2.2, define the current number of iterations as I, the maximum number of iterations as I _max , and initialize I=1;

Randomly initialize all weight matrices and node bias vectors in the neural network when I=1, and record the initialized weight matrix and node bias vectors as the set of the I iteration θ ^I ={ω ⁽⁰⁾ ,ω ⁽¹⁾ ,…,ω ^(H) ; b ⁽¹⁾ ,b ⁽²⁾ ,…b ^(H) };

Step 2.3, using formula (2) to establish the loss function l(θ ^I ) of the first iteration of the neural network:

In formula (2), Indicates the calculation result sequence output by the kth node of the output layer after the I iteration, and where, /> Indicates the log-likelihood ratio of the i-th received signal output by the k-th node after the i-th iteration;

Step 2.4, using formula (3) to update the gradient θ ^I of the I iteration, that is, to update all weight matrices and bias vectors in the network, to obtain the gradient θ I+ ¹ of the I+1 iteration;

In formula (3), α is the learning rate in machine learning, and α>0;

Step 2.5, after assigning I+1 to I, judge whether I>I _max is true, if true, it means that the neural network training is completed, and the calculated network model after training is obtained, otherwise, return to step 2.3 and execute sequentially until the loss Until the function l(θ ^I ) is smaller than the set training standard ε, the training is stopped, and the trained computing network model is obtained, which is used to calculate the received signal sequence The optimal log-likelihood ratio sequence for each signal mapping bit sequence in in, Represents the optimal log-likelihood ratio sequence output by the kth node of the output layer /> Indicates the optimal logarithmic likelihood ratio of the i-th received signal output by the k-th node of the network output layer, and the neural network parameter after training is recorded as θ, and used as the network parameter in the calculation stage;

Step 2.6, calculate the generalized mutual information G through formula (4);

In formula (4), E represents the mathematical expectation, Indicates that the decision is /> when the reception is B _k The conditional probability of , /> means /> probability distribution.

Compared with prior art, the beneficial effect of the present invention is:

1. The present invention realizes the calculation and training of the logarithmic likelihood ratio LLR based on the neural network model, the network structure is simple, and the algorithm is simple to implement, which overcomes the inability of the prior art to take into account the measurement of the accuracy and real-time performance of the channel performance index. By improving the measurement The algorithm used by the device makes the measurement results relatively real-time, and at the same time makes the measurement results accurate through the gradient descent of the end-to-end neural network and minimizes the value of the loss function, so that the detection method is easy to design and has good practical value.

2. The design logic of the present invention is simple. The approximate logarithmic likelihood ratio LLR of the channel is estimated by the maximum likelihood method, and the training is stopped when the value of the loss function is less than a certain selected value ε, and the neural network under this parameter can be used The model is used as the calculation model of LLR, which greatly improves the calculation efficiency of LLR.

3. The present invention uses the logarithmic likelihood ratio LLR and the generalized mutual information GMI to measure channel performance, and there is a certain relationship between the two, so the GMI can be calculated by using the LLR, thereby simplifying the calculation of the GMI and improving the GMI Computational efficiency, and by adjusting the update time to make the resulting GMI real-time.

Description of drawings

Fig. 1 is a schematic structural diagram of the communication reachable rate detection method based on end-to-end learning of the present invention;

Fig. 2 is the flow chart of the detection method of communication reachable rate based on end-to-end learning of the present invention;

Fig. 3 is the neural network diagram of the parameter training of the communication reachable rate detection method based on end-to-end learning of the present invention;

FIG. 4 is a neural network calculation model diagram of the end-to-end learning-based communication reachable rate detection method of the present invention.

Detailed ways

In the actual communication system, the implementation flow chart of the detection method based on end-to-end learning is shown in Figure 2. By constructing the neural network model and initializing its parameters, training the neural network and calculating the value of the loss function , stop training when the number of training times exceeds the maximum number of iterations, otherwise calculate the loss function value, stop training when it is less than ε, take ε=2, use the trained model to calculate the log-likelihood ratio LLR of the signal, and finally use the LLR to calculate The generalized mutual information GMI of the signal is output and the calculation result is output, and the channel condition is detected at the same time. Set a fixed update time to retrain, update network parameters and recalculate LLR and GMI, so as to ensure that the calculation results are accurate and in line with the actual situation.

In this embodiment, a schematic structural diagram of an end-to-end learning-based communication reachable rate detection method is shown in Figure 1, which includes four modules, and the input parameters are the signal sequence s={s ₁ ,s ₂ sent by the sending end ,…, s _n } and the signal sequence received by the receiver Take n=32, the logarithmic likelihood ratio LLR value λ _m calculated by the neural network model and the LLR value calculated by the maximum logarithm approximation/> As the independent variable of the loss function, when the number of training does not reach the maximum number of iterations I _max and the value of the loss function does not meet the requirements, that is, l(θ)≥ε(ε=2), enter the training mode, and pass the parameter initialization module and the network training module. Network parameters are trained; when l(θ)<ε(ε=2), the trained network parameter θ is used to enter the calculation mode, and after the logarithmic likelihood ratio LLR and generalized mutual information GMI calculation stage, the GMI is realized. For real-time detection, the generalized mutual information of the channel can be measured by connecting the interfaces at both ends of the detection device to the two ends of the channel; Compared with the neural network model of LLR, this design method can realize real-time calculation of channel generalized mutual information GMI.

The method can be extended to multi-dimensional signals, and the detection device can be extended to a multi-dimensional signal generalized mutual information detector by adjusting the number of neurons in the input layer of the network. Specifically, the method is carried out as follows:

Step 1. Define the sending signal sequence of the sending end as s={s ₁ , s ₂ ,...,s _i ,...,s _n }, s _i represents the i-th sending signal, and define the receiving signal sequence of the receiving end as Represents the i-th received signal, i∈[1,n], n represents the sequence length.

Define Receive Signal Sequence Each received signal in is mapped to a bit sequence with a length of m, and the sequence formed by defining the kth bit of the received signal sequence mapping is B _k ={B _k,1 ,B _k,2 ,…,B _k,i , ...,B _1,n }, where, B _k,i represents the received signal /> The kth bit of the mapping, k∈[1,m], B _k,i ∈{0,1}; the sending and receiving process is shown in Figure 1, taking the bit sequence length of each signal mapping as 8 as an example, That is, m=8.

Step 2, will receive the signal sequence Each received signal in is divided into two parts, the real part and the imaginary part, and it is recorded as the input signal sequence of the neural network as /> where, /> Indicates the 2nth input value of the input layer;

Let the node bias vector from the input layer to the hidden layer be denoted as where, /> Represents the bias of the 2nth node of the input layer;

The number of nodes of the input layer and the hidden layer of the neural network is 2n, the number of nodes of the output layer of the neural network is m; the number of layers of the hidden layer is H;

Let the weight matrix from 2n nodes in the input layer to 2n nodes in the first hidden layer be denoted as where, /> Indicates the weight of the 2nth node in the input layer to the 2nth node in the hidden layer of the first layer;

Let the weight matrix of any 2n nodes in the hth hidden layer to the 2n nodes in the h+1th hidden layer be denoted as where, /> Represents the weight of the 2nth node in the hth hidden layer to the 2nth node in the h+1th hidden layer;

Let the weight matrix from the 2n nodes in the hidden layer of the H layer to the m nodes in the output layer be denoted as where, /> Indicates the weight of the 2nth node in the hidden layer of the H layer to the mth node in the output layer, h∈[1,H];

Let the node bias vector in any h-th hidden layer be denoted as where, /> Indicates the bias of the 2nth node in the hth hidden layer;

Let the calculation result sequence of any h-th hidden layer be recorded as where, /> Indicates the calculation result of the 2nth node in the hidden layer of the hth layer;

Let the linear equation of the h-th hidden layer be y ^(h) = ω ^(h) x ^T + b ^(h) , where T represents transposition;

Let the output sequence x ^(h) after activation of the hth hidden layer ={x ₁ ^(h) ,x ₂ ^(h) ,…,x _2n ^(h) }, where, Indicates the 2nth output after activation of the hth hidden layer, and x ^(h) = f(y ^(h) ), f( ) is the activation function;

The output sequence z ^(H) after activation of the H-th hidden layer is used as the input sequence of the output layer, and is output by the linear equation y=ω ^(H) (x ^(H) ) ^T +b ^(H) in the output layer The calculation result sequence is y={y1,y2,...,yn}, where yn represents the calculation result output by the nth node of the output layer; x ^(H) represents the output sequence after the H-th hidden layer is activated, and b ^{( H)} represents the node bias vector in the hidden layer of the H layer; the neural network example diagram is as shown in Figure 3, Re represents the real part of the signal in the input node, and Im represents the imaginary part of the signal, taking n=32 as an example, input nodes The number is 2n=64, taking one layer of hidden layer as an example, that is, H=1, the number of nodes in the hidden layer is 64, the number of nodes in the output layer is m=8, and the output result of the network is m logarithmic likelihood than sequence. The calculation process of the first node in the hidden layer is shown in Figure 4, where Represents all the weights of all nodes in the input layer to the first node in the hidden layer, and the calculation result /> The activation function takes the ReLU function as an example, that is, f(x)=max(0,x), and the activation function is applied to /> after getting /> as input to the next layer of nodes.

Step 2.1, assuming channel distribution f _Y|X (y|x), using the maximum logarithm approximation method shown in formula (1) to estimate the logarithmic likelihood ratio sequence of the binary mapping bit sequence of n received signals is where, /> Represents the log-likelihood ratio sequence of the kth bit of n received signal mapping /> Indicates the log-likelihood ratio of the k-th bit representing the n-th received signal mapping, k∈[1,m];

In formula (1), Respectively represent the maximum probability that the kth bit of the n received signal mapping is judged as 1 or 0, f _Y|X (y|x) represents the channel transition probability of the transmitted signal as x and the received signal as y, and X and Y respectively represent Send and receive signal sequences;

Step 2.2, define the current number of iterations as I, the maximum number of iterations as I _max , and initialize I=1;

Randomly initialize all weight matrices and node bias vectors in the neural network when I=1, and record the initialized weight matrix and node bias vectors as the set of the I iteration θ ^I ={ω ⁽⁰⁾ ,ω ⁽¹⁾ ,…,ω ^(H) ; b ⁽¹⁾ ,b ⁽²⁾ ,…b ^(H) };

Step 2.3, using formula (2) to establish the loss function l(θ ^I ) of the first iteration of the neural network:

In formula (2), Indicates the calculation result sequence output by the kth node of the output layer after the I iteration, denoted as where, /> Indicates the log-likelihood ratio of the i-th received signal output by the k-th node after the i-th iteration;

Step 2.4, using formula (3) to update the gradient θ ^I of the I iteration, that is, to update all weight matrices and bias vectors in the network, to obtain the gradient θ I+ ¹ of the I+1 iteration;

In formula (3), α is the learning rate in machine learning, and α>0; taking α=0.1 as an example, the training speed and calculation speed of the network are improved while ensuring the calculation accuracy;

Step 2.5. After assigning I+1 to I, judge whether I>I _max is true. If it is true, it means that the neural network training is completed and the trained computing network model is obtained. Otherwise, return to step 2.3 and execute sequentially until the loss function l (θ ^I ) is less than the set training standard ε, stop the training, and get the trained computing network model, which is used to calculate the received signal sequence The optimal log-likelihood ratio sequence of each signal mapping bit sequence in , and the calculation result is recorded as where, /> Represents the optimal log-likelihood ratio sequence output by the kth node of the output layer , where /> Indicates the calculation of the optimal logarithmic likelihood ratio of the i-th received signal output by the k-th node of the network output layer, and the neural network parameter after training is recorded as θ, which is used as the network parameter in the calculation stage; as shown in Figure 1, in In the training mode, the trained network parameter θ is transferred to the network in the calculation mode, and the log likelihood ratio of the signal can be quickly calculated by using the network.

Step 2.6, calculate the generalized mutual information G through formula (4);

In formula (4), E represents seeking mathematical expectation, m and B _k are defined in the above-mentioned step 1, As defined in step 2.5 above, /> Indicates that the decision is /> when the reception is B _k The conditional probability of , /> means /> probability distribution. As shown in Figure 1 and Figure 2, the detector includes a parameter initialization module, a network training module, an LLR calculation module and a generalized mutual information GMI calculation module to realize real-time detection of GMI. By adjusting the number of network input nodes, multi-dimensional For the AIR detection of the signal, the modification of the training end flag can adjust the calculation time and calculation accuracy, and the determination of the appropriate end flag can improve the real-time detection and ensure the accuracy of the calculation.

Claims (1)

1. A communication reachable rate detection method based on end-to-end learning, its characteristics comprising: Step 1. Define the sending signal sequence of the sending end as s={s ₁ , s ₂ ,...,s _i ,...,s _n }, s _i represents the i-th sending signal, and define the receiving signal sequence of the receiving end as Represents the i-th received signal, i∈[1,n], n represents the sequence length; Define Receive Signal Sequence Each received signal in is mapped to a bit sequence with a length of m, and the sequence formed by defining the kth bit of the received signal sequence mapping is B _k ={B _k,1 ,B _k,2 ,…,B _k,i , ...,B _1,n }, where, B _k,i represents the received signal /> The kth bit of the mapping, k∈[1,m], B _k,i ∈{0,1}; Step 2, the received signal sequence Each received signal in is divided into two parts, the real part and the imaginary part, and it is recorded as the input signal sequence of the neural network as /> where, /> Indicates the 2nth input value of the input layer, that is, the 0th layer; Let the node bias vector from the input layer to the hidden layer be denoted as where, /> Represents the bias of the 2nth node of the input layer; Make the input layer of the neural network, the number of nodes of the hidden layer be 2n, make the number of nodes of the output layer of the neural network be m; The number of layers of the hidden layer is H; Let the weight matrix of the 2n nodes in the input layer to the 2n nodes in the first hidden layer be denoted as where, /> Represent the weight of the 2nth node in the input layer to the 2nth node in the first hidden layer; Let the weight matrix of any 2n nodes in the hth hidden layer to the 2n nodes in the h+1th hidden layer be denoted as where, /> Indicates the weight of the 2nth node in the hth hidden layer to the 2nth node in the h+1th hidden layer; Let the weight matrix from the 2n nodes in the hidden layer of the H layer to the m nodes in the output layer be denoted as where, /> Represent the weight of the 2nth node in the hidden layer of the H layer to the mth node in the output layer, h∈[1,H]; Let the node bias vector in any h-th hidden layer be denoted as where, /> Indicates the bias of the 2nth node in the hth hidden layer; Let the calculation result sequence of any h-th hidden layer be recorded as where, /> Indicates the calculation result of the 2nth node in the hidden layer of the hth layer; Let the linear equation of the h-th hidden layer be y ^(h) = ω ^(h) x ^T + b ^(h) , where T represents transposition; Let the output sequence x ^(h) after activation of the hth hidden layer ={x ₁ ^(h) ,x ₂ ^(h) ,…,x _2n ^(h) }, where, Indicates the 2nth output after activation of the hth hidden layer, and x ^(h) = f(y ^(h) ), f( ) is the activation function; The output sequence z ^(H) after activation of the hidden layer of the H layer is used as the input sequence of the output layer, and the linear equation y=ω ^(H) (x ^(H) ) ^T + b ^(H) in the output layer The output calculation result sequence is y={y ₁ ,y ₂ ,…,y _n }, where y _n represents the calculation result output by the nth node of the output layer; x ^(H) represents the activation of the hidden layer of the Hth layer The output sequence of , b ^(H) represents the node bias vector in the hidden layer of the Hth layer; Step 2.1, assuming channel distribution f _Y|X (y|x), using the maximum logarithm approximation method shown in formula (1) to estimate the logarithmic likelihood ratio sequence of the binary mapping bit sequence of n received signals is where, /> Represents the log-likelihood ratio sequence of the kth bit of n received signal mapping /> Represents the log-likelihood ratio of the k-th bit mapped to the n-th received signal, k∈[1,m]; In formula (1), Respectively represent the maximum probability that the kth bit of the n received signal mapping is judged as 1 or 0, f _Y|X (y|x) represents the channel transition probability of the transmitted signal as x and the received signal as y, and X and Y respectively represent Send and receive signal sequences; Step 2.2, define the current number of iterations as I, the maximum number of iterations as I _max , and initialize I=1; Randomly initialize all weight matrices and node bias vectors in the neural network when I=1, and record the initialized weight matrix and node bias vectors as the set of the I iteration θ ^I ={ω ⁽⁰⁾ ,ω ⁽¹⁾ ,…,ω ^(H) ; b ⁽¹⁾ ,b ⁽²⁾ ,…b ^(H) }; Step 2.3, using formula (2) to establish the loss function l(θ ^I ) of the first iteration of the neural network: In formula (2), Indicates the calculation result sequence output by the kth node of the output layer after the I iteration, and where, /> Indicates the log-likelihood ratio of the i-th received signal output by the k-th node after the i-th iteration; Step 2.4, using formula (3) to update the gradient θ ^I of the I iteration, that is, to update all weight matrices and bias vectors in the network, to obtain the gradient θ I+ ¹ of the I+1 iteration; In formula (3), α is the learning rate in machine learning, and α>0; Step 2.5, after assigning I+1 to I, judge whether I>I _max is true, if true, it means that the neural network training is completed, and the calculated network model after training is obtained, otherwise, return to step 2.3 and execute sequentially until the loss Until the function l(θ ^I ) is smaller than the set training standard ε, the training is stopped, and the trained computing network model is obtained, which is used to calculate the received signal sequence The optimal log-likelihood ratio sequence for each signal mapping bit sequence in where, /> Represents the optimal log-likelihood ratio sequence output by the kth node of the output layer /> Indicates the optimal logarithmic likelihood ratio of the i-th received signal output by the k-th node of the network output layer, and the neural network parameter after training is recorded as θ, and used as the network parameter in the calculation stage; Step 2.6, calculate the generalized mutual information G through formula (4); In formula (4), E represents the mathematical expectation, Indicates that the decision is /> when the reception is B _k The conditional probability of , /> means /> probability distribution.