WO2022120757A1 - 一种星座符号检测方法及装置 - Google Patents

一种星座符号检测方法及装置 Download PDF

Info

Publication number
WO2022120757A1
WO2022120757A1 PCT/CN2020/135430 CN2020135430W WO2022120757A1 WO 2022120757 A1 WO2022120757 A1 WO 2022120757A1 CN 2020135430 W CN2020135430 W CN 2020135430W WO 2022120757 A1 WO2022120757 A1 WO 2022120757A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
weight
modulation
trained
weights
Prior art date
Application number
PCT/CN2020/135430
Other languages
English (en)
French (fr)
Inventor
杨志鸿
梁璟
曾兴
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to PCT/CN2020/135430 priority Critical patent/WO2022120757A1/zh
Publication of WO2022120757A1 publication Critical patent/WO2022120757A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00—Modulated-carrier systems

Definitions

  • the present application relates to the field of communications, and more particularly, to a method and apparatus for detecting constellation symbols.
  • the transmitter In a communication system, the transmitter generally performs symbol mapping on the coded bits, which is called symbol modulation, and then modulates and transmits the carrier.
  • symbol modulation At the receiving end, baseband processing needs to detect the transmitted constellation symbols.
  • a feasible method is to use the powerful nonlinear fitting ability of neural network technology to solve the problem of constellation symbol detection.
  • the core of neural network technology is to store the detected inference parameters in the network weights. Generally, since the number of discrete constellation symbol points under different modulations is different, and for different modulation methods, different network weights need to be trained separately. Therefore, when the number of modulation methods that need to be supported in the system increases, the storage of network weights The cost will also increase.
  • the present application provides a constellation symbol detection method and device, so as to reduce the storage overhead of network weights under different modulation modes.
  • a first aspect provides a method for constellation symbol detection, the method comprising: acquiring a first network weight of first modulation, where the first network weight is an M ⁇ N matrix, and M and N are both positive integers;
  • the first network weight determines the second network weight to be trained, the second network weight to be trained includes all parameters of the first network weight, and the second network weight to be trained is the second modulated network weight value, the modulation order of the second modulation is greater than the modulation order of the first modulation, and the weight of the second network to be trained is a (M+a) ⁇ (N+b) matrix, where a and b are both positive Integer; train the parameters other than the first network weight in the second network weight to be trained.
  • all parameters in the first network weight of the first modulation are inserted into the second network weight to be trained corresponding to the second modulation, and the second network weight to be trained is trained. parameters other than the first network weight, thereby reducing the storage overhead of network weights in different modulation modes.
  • the first modulation and the second modulation are adjacent-order modulations having the same modulation mode.
  • the first network weight includes a first weight and a first offset
  • the second network weight to be trained includes a second weight and a second offset
  • the The second weight includes all parameters of the first weight
  • the second bias includes all parameters of the first bias
  • the first weight includes at least two sub-weights, and the at least two sub-weights are respectively inserted into the second weight, and the insertion positions of the at least two sub-weights do not overlap.
  • the first offset includes at least two sub-offsets, the at least two sub-offsets are respectively inserted into the second offset, and the insertion of the at least two sub-offsets The positions do not coincide.
  • the first network weight includes at least two sub-network weights, the at least two sub-network weights include a weight and a bias, and the at least two sub-network weights are respectively Inserted into the second network weight to be trained, the insertion positions of at least two sub-network weights do not coincide.
  • the second network weight to be trained includes all parameters of the first network weight, including evenly and equally spaced M ⁇ N parameters of the first network weight Insert into the (M+a) ⁇ (N+b) matrix of the weights of the second network to be trained.
  • the modulation modes of the first modulation and the second modulation include quadrature amplitude modulation (QAM).
  • QAM quadrature amplitude modulation
  • the method further includes: acquiring a second network weight after training; and detecting a constellation symbol vector according to the second network weight after training.
  • a constellation symbol detection device in a second aspect, includes: an acquisition unit configured to acquire a first network weight of the first modulation, the first network weight is an M ⁇ N matrix, and both M and N are positive Integer; processing unit, the processing unit is used to determine the second network weight to be trained according to the first network weight, the second network weight to be trained includes all parameters of the first network weight, the second network weight to be trained is the network weight of the second modulation, the modulation order of the second modulation is greater than the modulation order of the first modulation, and the weight of the second network to be trained is a (M+a) ⁇ (N+b) matrix, where, Both a and b are positive integers; the processing unit is further configured to train parameters other than the first network weight in the second network weight to be trained.
  • all parameters in the first network weight of the first modulation are inserted into the second network weight to be trained corresponding to the second modulation, and the second network weight to be trained is trained. parameters other than the first network weight, thereby reducing the storage overhead of network weights in different modulation modes.
  • the first modulation and the second modulation are adjacent-order modulations having the same modulation mode.
  • the first network weight includes a first weight and a first offset
  • the second network weight to be trained includes a second weight and a second offset
  • the first network weight includes a first weight and a first offset.
  • the second weight includes all parameters of the first weight
  • the second bias includes all parameters of the first bias.
  • the first weight includes at least two sub-weights, and the at least two sub-weights are respectively inserted into the second weight, and the insertion positions of the at least two sub-weights do not overlap.
  • the first offset includes at least two sub-biases, the at least two sub-biases are respectively inserted into the second offset, and the at least two sub-biases are The insertion positions do not coincide.
  • the first network weight includes at least two sub-network weights, the at least two sub-network weights include a weight and a bias, and the at least two sub-network weights are respectively Inserted into the second network weight to be trained, the insertion positions of the at least two sub-network weights do not coincide.
  • the second network weight to be trained includes all parameters of the first network weight, including evenly and equally spaced M ⁇ N parameters of the first network weight Insert into the (M+a) ⁇ (N+b) matrix of the weights of the second network to be trained.
  • the modulation modes of the first modulation and the second modulation include quadrature amplitude modulation (QAM).
  • QAM quadrature amplitude modulation
  • the obtaining unit is further configured to obtain the trained second network weight; the processing unit is further configured to detect the constellation symbol vector according to the trained second network weight.
  • a constellation symbol detection device comprising: a memory for storing a program; a processor for executing the program stored in the memory, when the program stored in the memory is executed, the The processor is configured to execute the method in any one of the implementation manners of the above first aspect.
  • the present application provides a computer-readable storage medium, where computer instructions are stored in the computer-readable storage medium, and when the computer instructions are executed on a computer, the computer is made to execute any one of the implementations in the first aspect. method.
  • the present application provides a computer program product, the computer program product includes computer program code, when the computer program code is run on a computer, the computer is made to execute any one of the implementations of the first aspect. method.
  • a sixth aspect provides a chip, the chip includes a processor and a data interface, and the processor reads an instruction stored in a memory through the data interface to execute any one of the implementations in the first aspect. method.
  • the chip may further include a memory, where instructions are stored in the memory, and the processor is configured to execute the instructions stored in the memory, when the instructions When executed, the processor is configured to execute the method in any one of the implementations of the first aspect.
  • FIG. 1 is a schematic structural diagram of a fully connected neural network provided by an embodiment of the present application.
  • FIG. 2 is a schematic flowchart of a method for constellation symbol detection provided by an embodiment of the present application
  • FIG. 3 is a schematic flowchart of a method for nesting network weights provided by an embodiment of the present application
  • FIG. 4 is another schematic flowchart of a method for nesting network weights provided by an embodiment of the present application.
  • FIG. 5 is another schematic flowchart of a method for nesting network weights provided by an embodiment of the present application.
  • FIG. 6 is another schematic flowchart of a method for nesting network weights provided by an embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of a constellation symbol detection device provided by an embodiment of the present application.
  • FIG. 8 is another schematic structural diagram of a constellation symbol detection device provided by an embodiment of the present application.
  • FIG. 9 is another schematic structural diagram of a constellation symbol detection apparatus provided by an embodiment of the present application.
  • GSM global system of mobile communication
  • CDMA code division multiple access
  • WCDMA wideband code division multiple access
  • general packet radio service general packet radio service
  • GPRS general packet radio service
  • long term evolution long term evolution
  • LTE long term evolution
  • LTE frequency division duplex frequency division duplex
  • TDD time division duplex
  • UMTS universal mobile telecommunication system
  • WiMAX worldwide interoperability for microwave access
  • 5G 5th generation
  • NR new radio
  • a future evolved wireless communication system etc.
  • FIG. 1 is a schematic structural diagram of a fully connected neural network.
  • an MLP includes an input layer (far left), an output layer (far right), and multiple hidden layers (only two layers are shown in Fig. 1).
  • Each layer consists of several nodes, called neurons. Neurons in two adjacent layers are connected two by two, and neurons in the same layer are not connected. The number of nodes included in each layer is called the width of the layer.
  • the input layer represents a feature vector
  • each neuron of the input layer represents a feature value.
  • the output h of the neurons of the next layer is the functional relationship between all the neurons x of the previous layer connected to it, which can be expressed as formula (1).
  • w is the weight matrix
  • b is the bias vector
  • f is the activation function
  • the neural network can use the error back propagation (BP) algorithm to correct the size of the parameters in the initial neural network model during the training process, so that the reconstruction error loss of the neural network model becomes smaller and smaller.
  • BP error back propagation
  • the input signal is passed forward until the output will generate error loss, and the parameters in the initial neural network model are updated by back-propagating the error loss information, so that the error loss converges.
  • the back-propagation algorithm is a back-propagation movement dominated by error loss, aiming to obtain the parameters of the optimal neural network model, such as the weight matrix.
  • the maximum likelihood algorithm is the optimal detection method.
  • the maximum likelihood detection is to compare the received signal with all possible transmitted signals to generate a maximum likelihood estimate. Its principle is to find the transmitted signal (or signal vector space) with the smallest distance from the received signal after channel transformation in the transmitted signal (or signal vector space).
  • the transmitter In a communication system, the transmitter generally performs symbol mapping on the coded bits, and then modulates and transmits the carrier. Different modulation orders contain different numbers of constellation sets. Take quadrature amplitude modulation (QAM) as an example. For example, 16QAM contains 16 optional constellation sets, and 64QAM contains 64 optional constellation sets. gather. Considering that the real part and the imaginary part are independently modulated in the complex plane during QAM modulation (usually called I/Q path), then, when the modulation mode is 16QAM, each branch has 4 states, and when the modulation mode is 64QAM , each branch has 8 states.
  • QAM quadrature amplitude modulation
  • the number of states of the optional constellation set for QAM modulation increases exponentially, and the number of states for the optional constellation set for QAM modulation increases exponentially.
  • baseband processing requires detection of the transmitted constellation symbols.
  • OFDM orthogonal frequency division multiplexing
  • MIMO multi-input multi-output
  • y represents the vector of the received signal in the frequency domain
  • H represents the channel matrix of the MIMO system
  • s represents the constellation symbol vector
  • n represents the additive Gaussian noise vector. Therefore, in a MIMO system, y and H are generally obtained through pilot information, and s is recovered.
  • MLD detection or linear minimum mean square error (LMMSE) detection is used, and LMMSE detection models the transmitted constellation symbols as a random Gaussian distribution.
  • MLD detection fully considers the modulation information of the transmitted constellation symbols, optimizes the accurate posterior probability of the modulation constellation set, and has the best performance under the Gaussian noise model.
  • the neural network technology has a strong nonlinear fitting ability, which is used to solve the MLD detection problem and approach the optimal MLD performance with low complexity.
  • the core of neural network technology is to store the inference parameters detected by MLD in the network weights, which are used to approximate the posterior probability of the discrete constellation set. Because the number of discrete constellation points is different under different modulation modes, and for different modulation modes, it is generally necessary to train separately to obtain different network weights. Based on this principle, this application proposes a method for network weight nesting.
  • This application takes an MLD detection algorithm designed by the receiving end in a MIMO system using neural network technology as an example.
  • the detection algorithm can also be an approximate MLD algorithm, such as sphere decoding (SD) algorithm, K-best algorithm tree search algorithm, etc.
  • the detection algorithm can also be various linear or linear enhancement algorithms, such as linear minimum mean square error (LMMSE) algorithm, sequential interference cancellation (successive interference cancellation, SIC) algorithm, expectation propagation (expectation propagation, EP) ) detection algorithm, etc., which are not limited here.
  • LMMSE linear minimum mean square error
  • SIC sequential interference cancellation
  • EP expectation propagation
  • the input vector of the MLD algorithm can be represented as X in
  • the neural network function is represented by f
  • the corresponding network weight set is represented by ⁇
  • the network output is represented by X out , that is, there is,
  • a constellation symbol detection network For a constellation symbol detection network, it generally has a fixed network structure, and has the same input for different modulation modes, that is, the same X in . That is to say, for a neural network algorithm that approximates MLD, for different modulation methods, the selected network input features (ie, input data) are the same as the neural network structure. In the embodiment of the present application, there is only one hidden layer as an example, that is, f is the same. Although the structure of the neural network is the same and the data input type is the same, the network weight ⁇ is often different, which is related to the number and value of the parameters in the network weight ⁇ .
  • the preset neural network structure is used, but only 8 nodes in the hidden layer are selected, and for the 16th order, 16 nodes in the hidden layer are selected.
  • the neural network structure includes a hidden layer as an example, and its MLP network structure f can be expressed as,
  • a linear rectified linear unit (ReLu) is selected as the activation function of the hidden layer.
  • the activation function of the hidden layer may also include a Sigmoid function, a Tanh function, and the like. It should be understood that the activation function introduces a nonlinear factor to the neuron, so that the neural network can approximate any nonlinear function arbitrarily, so that the neural network can be applied to many nonlinear models. If the activation function is not used, the output of each layer is a linear function of the input of the upper layer. No matter how many layers of the neural network function, the output is a linear combination of the input. This application does not make any limitation on the selection of the activation function.
  • the network weight ⁇ ⁇ W 1 , b 1 , W 2 , b 2 ⁇ .
  • W1 represents the weight from the input layer to the hidden layer
  • b1 represents the bias from the input layer to the hidden layer
  • W2 represents the weight from the hidden layer to the output layer
  • b2 represents the bias from the hidden layer to the output layer.
  • the network weight is generally a matrix or vector related to the modulation method, that is to say, the modulation method will affect the size of the matrix and the value of each element in the matrix, that is, different modulation methods have different dimensions of the network weights.
  • FIG. 2 is a schematic diagram of a network weight embedding method provided by an embodiment of the present application.
  • S220 Determine a second network weight to be trained according to the first network weight, where the second network weight to be trained includes all parameters of the first network weight, and the second network weight to be trained is a second modulation
  • the network weight of the second modulation is greater than the modulation order of the first modulation, and the weight of the second network to be trained is a (M+a) ⁇ (N+b) matrix, where a and b are all positive integers;
  • obtaining the first network weight of the first modulation can usually be obtained by training the input value. That is to say, when the modulation order of the first modulation is the lowest order in the communication system with modulation, the first network weight of the first modulation is generally obtained by training a predefined value, or when the first modulation is When the network weights of the previous order modulation are used as input values, they can be obtained by training.
  • the network weight in this embodiment of the present application is a multi-dimensional matrix
  • the second network weight may be a (M+a) ⁇ (N+b) matrix, where , where M, N, a, and b are all positive integers.
  • a second network weight to be trained is determined according to the first network weight, and the second network weight to be trained includes all parameters of the first network weight by inserting all parameters in M ⁇ N into (M+a) ⁇ (N+b) matrix.
  • the weight of the first network is a matrix of order 4 ⁇ 4, which contains 16 parameters
  • the weight of the second network is a matrix of order 8 ⁇ 8, which contains 64 parameters
  • 16 of the matrix of order 4 ⁇ 4 are The parameters are inserted into the 8 ⁇ 8-order matrix, which means that 16 of the 64 parameters of the 8 ⁇ 8-order matrix are from the 4 ⁇ 4-order matrix, and the remaining 48 parameters can be obtained by training.
  • parameters other than the first network weight among the second network weights to be trained are trained, that is, the 48 parameters are trained, that is, among the 64 parameters, 16 are from the 4 ⁇ 4 order matrix Parameters are not trained.
  • the first network weight may also be of order 3 ⁇ 2
  • the second network weight may be of order 5 ⁇ 6, that is to say, as long as the dimension of the second network weight is greater than the first network weight, the The dimensions of the first network weight and the second network weight will not be described in detail.
  • the first modulation may be 4QAM, 16QAM, or 64QAM modulation in QAM modulation, and the modulation mode and modulation order are as shown in Table 1 below, as long as it is the highest-order modulation.
  • the first modulation may also be other modulations, such as 2PSK, 4PSK, or 8PSK in multi-phase shift keying (MPSK) modulation, and the corresponding modulation modes and modulation orders are shown in Table 2 below.
  • MPSK multi-phase shift keying
  • the present application will hereinafter describe the first modulation as 4QAM in QAM modulation as an example, but the present application can also be used for different modulation orders of other modulation types, and the first modulation is not limited herein.
  • the communication system there are 4QAM, 16QAM and 64QAM in the communication system. That is to say, the highest-order modulation in this system is 64QAM, so in order to save overhead, a network weight of modulation lower than that of 64QAM modulation order is inserted into the network weight of 64QAM.
  • MPSK and QAM modulation methods can also exist in the communication system.
  • the network weights of low-order modulation such as BPSK or QPSK can also be included in 16QAM, 64QAM, and 256QAM. In the network weights of equal high-order modulation.
  • the modulation order of 4QAM is lower than that of 16QAM, and the modulation order of 16QAM is lower than that of 64QAM, so first insert the network weight ⁇ (4QAM) of 4QAM into the network weight of 16QAM, The network weights of 16QAM are trained, and then the network weights ⁇ (16QAM) in the trained 16QAM are inserted into the network weights of 64QAM, and the network weights of 64QAM are trained.
  • the network weight of 4QAM when the network weight of 4QAM is inserted into the network weight of 16QAM, 4QAM is the first modulation, the network weight ⁇ (4QAM) of 4QAM is the first network weight, and 16QAM is the second modulation, The network weight ⁇ (16QAM) of 16QAM is the second network weight; when the network weight ⁇ (16QAM) of 16QAM is inserted into the network weight of 64QAM, 16QAM is the first modulation, and the network weight of 16QAM is is the first network weight, 64QAM is the second modulation, and the network weight ⁇ (64QAM) of 64QAM is the second network weight.
  • the network weight ⁇ (4QAM) of 4QAM is the network weight that has been trained.
  • each parameter in ⁇ (4QAM) is first inserted into the waiting
  • the training updates the parameters except ⁇ (4QAM) to obtain the network weights ⁇ (16QAM) of 16QAM, and then inserts each parameter in ⁇ (16QAM) into the 64QAM to be trained.
  • the network weights of training and updating the parameters except ⁇ (16QAM) to obtain the network weights ⁇ (64QAM) of 64QAM, so that the network weights of low-order modulation can be inserted into the network weights of high-order modulation. middle.
  • the above embodiment is only a preferred implementation, which can greatly save the network weight storage overhead.
  • the network weight of 4QAM can also be inserted into the network weight of 64QAM, as long as the low-order network weight can be realized
  • the insertion of the modulated network weight into the high-order modulated network weight falls within the protection scope of this embodiment of the present application, and details are not described here.
  • the insertion method that saves the storage cost of network weights is to insert sequentially from low to high, That is, the network weight of 4 1 QAM Inserted into the network weights of 4 2 QAM, training divided parameters other than the Then the network weights of 4 2 QAM Inserted into the network weights of 4 3 QAM, training divided parameters other than the And so on, the network weight of 4 q-1 QAM Inserted into the network weights of 4q QAM, the training divides parameters other than the Then the network weights of 4 q QAM Inserted into the network weights of 4 q+1 QAM, training divided by parameters other than the The final implementation inserts the low-order modulated network weights into the high-order modulated network weights.
  • FIG. 4 is only a preferred embodiment.
  • the network weight of 4 1 QAM can also be inserted into any network weight higher than its order.
  • the ability to insert the low-order modulated network weight into the high-order network weight falls within the protection scope of the embodiments of the present application, and details are not described here.
  • Fig. 5 shows a nested form of network weights.
  • the low-order modulation is 4QAM as an example, and the high-order modulation is 16QAM and 64QAM respectively. It should be understood that this is only an exemplary description, the low-order modulation can be any modulation, the high-order modulation can also be any modulation, and the number of modulation modes is not limited.
  • (a), (b) and (c) in Figure 5 represent three second network weights to be trained according to the first network weights, wherein (a) in Figure 5 represents that all parameters in 4QAM are inserted into In the upper left corner of the 16QAM matrix, all parameters in the trained 16QAM are inserted into the upper left corner of the 64QAM matrix. It can be roughly represented by a matrix as:
  • the weight of the 4QAM network is a 2 ⁇ 2 order matrix:
  • the weight of the 16QAM network is a 4 ⁇ 4 order matrix:
  • the weight of the 64QAM network is an 8 ⁇ 8 order matrix:
  • the above-mentioned matrices corresponding to different modulation modes and various parameters in the matrices are only exemplary descriptions, and the specific matrix size and parameter values can be determined by themselves according to the actual situation, which will not be repeated here.
  • the higher the modulation order the larger the dimension of the corresponding network weight matrix.
  • the dimension of the network weight matrix is equal to its corresponding modulation order, such as the network weight matrix corresponding to each modulation order in the foregoing exemplary description.
  • the dimension of the network weight matrix does not have to be the same as its corresponding modulation order.
  • the network weight matrix of 16QAM is a 3 ⁇ 3 order matrix, as long as the network weights of the lower order modulation are smaller
  • the dimension of the matrix can be large.
  • the weight of the 4QAM network is a 2 ⁇ 2 order matrix:
  • the weight of the 16QAM network is a 4 ⁇ 4 order matrix:
  • the weight of the 16QAM network is an 8 ⁇ 8 order matrix:
  • (c) in Figure 5 indicates that the network weights in 4QAM include four sub-network weights. Insert these four sub-network weights into any position of the 16QAM matrix, and the four positions do not overlap with each other.
  • the trained 16QAM All parameters are divided into four sub-network weights. Similarly, these four sub-network weights are inserted into any position of the 64QAM matrix, and the four positions do not coincide with each other.
  • inserting them into the upper left corner. It can be roughly represented by a matrix as:
  • the weight of the 4QAM network is a 2 ⁇ 2 order matrix:
  • the weight of the 16QAM network is a 4 ⁇ 4 order matrix:
  • the weight of the 16QAM network is an 8 ⁇ 8 order matrix:
  • sub-network weights in the modulation mode are uniformly divided into blocks and inserted into the upper left corner, which may also be the middle position or the lower right corner. Therefore, in practical applications, the sub-network weights may be unevenly divided into blocks, and the insertion methods of the sub-network weights may also be different, which will not be limited here.
  • Fig. 6 shows a nested form of network weights. As shown in Fig. 6, the low-order modulation is 4QAM as an example, and the high-order modulation is 16QAM as an example for description.
  • first weights weights (first weights) and biases (first biases)
  • second weights biases
  • second bias bias
  • the weight W 1 of 4QAM is a matrix of order 4 ⁇ 4, the bias b 1 is a matrix of order 4 ⁇ 1, and the weight W 2 of 16QAM is a matrix of order 8 ⁇ 4, and the bias For example, b 2 is a matrix of order 8 ⁇ 1.
  • the weights in 4QAM are inserted into the weight matrix to be trained in 16QAM in row units evenly and at equal intervals.
  • the biases in 4QAM are inserted into the offset matrix to be trained in 16QAM in row units evenly and at equal intervals.
  • the weight matrix corresponding to 4QAM has 4 rows
  • the weight matrix corresponding to 16QAM has 8 rows
  • insert the first row in the 4QAM weight matrix into the first row in the 16QAM weight matrix and insert the second row in the 4QAM weight matrix into to the third row in the 16QAM weight matrix
  • insert the third row in the 4QAM weight matrix into the fifth row in the 16QAM weight matrix insert the fourth row in the 4QAM weight matrix into the seventh row in the 16QAM weight matrix, 4QAM
  • the insertion method of the bias matrix of which will not be described too much.
  • the weight W 1 of 4QAM is a matrix of order 4 ⁇ 4
  • the bias b 1 is a matrix of order 4 ⁇ 1
  • the weight W 2 of 16QAM is a matrix of order 8 ⁇ 8, and the bias
  • b 2 is a matrix of order 8 ⁇ 1.
  • the weights in 4QAM are inserted into the weight matrix to be trained in 16QAM with each parameter as the unit, and the bias in 4QAM is inserted into the weight matrix to be trained in 16QAM with each parameter as the unit. in the bias matrix. That is, the weight matrix corresponding to 4QAM has 16 parameters, and the weight matrix corresponding to 64QAM has 64 parameters.
  • the 16 parameters can be inserted into the 64 parameters evenly and at equal intervals.
  • the specific implementation method can be as follows.
  • the parameters of the first column of the row are inserted into the first row and the first column of the 64QAM weight matrix, and the parameters of the first row and the second column of the 16QAM weight matrix are inserted into the first row and the third column of the 64QAM weight matrix.
  • the parameters of the first row and third column are inserted into the first row and fifth column of the 64QAM weight matrix, and the parameters of the first row and fourth column of the 16QAM weight matrix are inserted into the first row and seventh column of the 64QAM weight matrix.
  • 16QAM weight The parameters of the second row and first column of the matrix are inserted into the second row and first column of the 64QAM weight matrix, and the parameters of the second row and second column of the 16QAM weight matrix are inserted into the second row and third column of the 64QAM weight matrix, etc. and so on, until the 16 parameters are inserted evenly and at equal intervals, the same is true for the insertion method of the 4QAM bias matrix, which will not be described too much.
  • the weight W 1 of 4QAM is a matrix of order 4 ⁇ 4
  • the bias b 1 is a matrix of order 4 ⁇ 1
  • the weight W 2 of 16QAM is a matrix of order 8 ⁇ 4
  • the bias For example, b 2 is a matrix of order 8 ⁇ 1.
  • the weights in 4QAM are inserted into the weight matrix to be trained in 16QAM in units of the entire weight matrix, and similarly, the offsets in 4QAM are inserted into the offset matrix to be trained in 16QAM in units of the entire weight matrix.
  • the weighted matrix corresponding to 4QAM is a 4 ⁇ 4-order matrix, with 4 rows and 4 columns, while the weighted matrix corresponding to 16QAM is of order 8 ⁇ 4, with 8 rows and 4 columns, so insert the parameters of 4 rows in 4QAM into 16QAM
  • the first 4 lines of the same is true for the insertion of the bias matrix of 4QAM, which will not be described too much.
  • the weight W 1 of 4QAM is a matrix of order 4 ⁇ 4
  • the bias b 1 is a matrix of order 4 ⁇ 1
  • the weight W 2 of 16QAM is a matrix of order 8 ⁇ 8, and the bias
  • b 2 is a matrix of order 8 ⁇ 1.
  • the weights in 4QAM are inserted into the upper left corner of the weight matrix to be trained in 16QAM in units of the entire weight matrix.
  • the biases in 4QAM are inserted into the offset matrix to be trained in 16QAM in units of the entire weight matrix.
  • the weight matrix corresponding to 4QAM has 4 rows and 4 columns
  • the weight matrix corresponding to 16QAM has 8 rows and 8 columns, so all parameters in 4 rows and 8 columns of 4QAM can be inserted into the first 4 rows and the first 4 columns corresponding to 16QAM.
  • the network weight corresponding to 4QAM is a matrix of 2 ⁇ 2
  • the network weight corresponding to 16QAM is a matrix of 4 ⁇ 4
  • the network weight corresponding to 64QAM is a matrix of 8 ⁇ 8
  • the network weight corresponding to 256QAM is 16 ⁇ 16 matrix
  • the storage overhead of the network weights is saved.
  • constellation symbol detection it is only necessary to obtain the network weight corresponding to the modulation mode by inference according to the corresponding modulation mode, so that the constellation symbol detection can be performed. That is to say, the method provided in the embodiment of the present application obtains the network weight corresponding to the highest-order modulation mode, and when performing constellation symbol detection, the corresponding network weight is obtained according to the corresponding modulation order for detection.
  • the system may save the network weight matrix corresponding to the high-order modulation scheme obtained according to the foregoing embodiment, for constellation symbol detection.
  • the network weight of the modulation scheme corresponding to the constellation symbol can be obtained from the network weight matrix of the high-order modulation scheme according to the modulation scheme corresponding to the constellation symbol, and the detection is performed.
  • the modulation order of the high-order modulation scheme is greater than or equal to the modulation scheme corresponding to the constellation symbols, and the network weight matrix of the high-order modulation scheme includes the network weights of the modulation scheme corresponding to the constellation symbols.
  • how to nest the network weights of the low-order modulation method in the network weights of the high-order modulation method, and how to obtain the network weights of the low-order modulation method from the network weights of the high-order modulation method can refer to the foregoing embodiments. No longer.
  • FIG. 7 is a schematic structural diagram of a constellation symbol detection apparatus provided by an embodiment of the present application.
  • the apparatus 700 shown in FIG. 7 includes an acquisition unit 701 and a processing unit 702 .
  • an obtaining unit 701 configured to obtain a first network weight of the first modulation, where the first network weight is an M ⁇ N matrix, and M and N are both positive integers;
  • the processing unit 702 is configured to determine the second network weight to be trained according to the first network weight, the second network weight to be trained includes all parameters of the first network weight, the second network weight to be trained is the network weight of the second modulation, the modulation order of the second modulation is greater than the modulation order of the first modulation, and the weight of the second network to be trained is a (M+a) ⁇ (N+b) matrix, where , a and b are both positive integers;
  • the processing unit 702 is further configured to train parameters other than the first network weight in the second network weight to be trained.
  • the first modulation and the second modulation are adjacent-order modulations having the same modulation mode.
  • the adjacent orders may be adjacent orders such as 1, 2, 3, ... q, or may be adjacent orders such as 1, 2, 4, ... 2 q , or may be 4, 16, 64 ; Adjacent-order modulation can be considered, and there is no need to limit too much here.
  • the first network weight includes a first weight and a first offset
  • the second network weight to be trained includes a second weight and a second offset
  • the second weight includes all parameters of the first weight
  • the first weight The second bias includes all parameters of the first bias
  • the first weight includes at least two sub-weights, the at least two sub-weights are respectively inserted into the second weight, and the insertion positions of the at least two sub-weights do not overlap.
  • the first offset includes at least two sub-offsets, the at least two sub-offsets are respectively inserted into the second offset, and the insertion positions of the at least two sub-offsets do not overlap.
  • the first network weight includes at least two sub-network weights, the at least two sub-network weights include a weight and a bias, and the at least two sub-network weights are respectively inserted into the second network weight to be trained, at least The insertion positions of the weights of the two sub-networks do not coincide.
  • the second network weight to be trained includes all parameters of the first network weight, including inserting M ⁇ N parameters in the first network weight into the second network weight to be trained at equal intervals. (M+a) ⁇ (N+b) matrix.
  • the modulation modes of the first modulation and the second modulation include quadrature amplitude modulation (QAM).
  • QAM quadrature amplitude modulation
  • the obtaining unit 701 is further configured to obtain the second network weight after training, and the processing unit 702 is further configured to detect the constellation symbol vector according to the second network weight after training.
  • FIG. 8 shows a schematic structural diagram of another apparatus for constellation symbol detection provided by an embodiment of the present application.
  • the apparatus 800 shown in FIG. 8 includes a storage unit 801 , an acquisition unit 802 and a processing unit 803 .
  • the storage unit 801 is configured to store the network weight matrix corresponding to the high-order modulation mode.
  • the obtaining unit 802 is configured to obtain the network weight of the modulation scheme corresponding to the constellation symbol from the network weight matrix of the higher-order modulation scheme according to the modulation scheme corresponding to the constellation symbol, so that the processing unit 803 can perform constellation symbol detection.
  • the modulation order of the high-order modulation scheme is greater than or equal to the modulation scheme corresponding to the constellation symbol, and the network weight matrix of the high-order modulation scheme includes the network weight of the modulation scheme corresponding to the constellation symbol.
  • the network weight matrix corresponding to the high-order modulation scheme may be obtained by the apparatus 800 trained according to any of the methods in FIG. 2 to FIG.
  • the network weight matrix corresponding to the high-order modulation scheme may be obtained through training by other devices, such as the device 700, and then the device 800 obtains the network weight matrix from other devices and stores it in the storage In the unit 801, the device 800 and other devices (eg, the device 700) are completely different devices at this time, which will not be repeated here.
  • FIG. 9 is a schematic diagram of a hardware structure of a constellation symbol detection apparatus according to an embodiment of the present application.
  • the apparatus 900 shown in FIG. 9 (the apparatus 900 may specifically be a computer device) includes a memory 901 , a processor 902 , a communication interface 903 and a bus 901 .
  • the memory 901 , the processor 902 , and the communication interface 903 are connected to each other through the bus 904 for communication.
  • the memory 901 may be a read only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM).
  • the memory 901 may store a program, and when the program stored in the memory 901 is executed by the processor 902, the processor 902 is configured to execute each step of the constellation symbol detection method of the embodiment of the present application.
  • the processor 902 can be a general-purpose central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), a graphics processor (graphics processing unit, GPU), or one or more
  • the integrated circuit is used to execute a related program to implement the constellation symbol detection method of the embodiment of the present application.
  • the processor 902 may also be an integrated circuit chip with signal processing capability.
  • each step of the constellation symbol detection method of the present application may be completed by a hardware integrated logic circuit in the processor 902 or an instruction in the form of software.
  • the above-mentioned processor 902 may also be a general-purpose processor, a digital signal processor (digital signal processing, DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, Discrete gate or transistor logic devices, discrete hardware components.
  • DSP digital signal processing
  • ASIC application-specific integrated circuit
  • FPGA field programmable gate array
  • a general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the steps of the methods disclosed in conjunction with the embodiments of the present application may be directly embodied as executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor.
  • the software module may be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other storage media mature in the art.
  • the storage medium is located in the memory 901, and the processor 902 reads the information in the memory 901, and combines its hardware to complete the functions required by the units included in the constellation symbol detection apparatus, or execute the constellation symbol detection method of the method embodiment of the present application.
  • the communication interface 903 uses a transceiver device such as, but not limited to, a transceiver to implement communication between the device 900 and other devices or a communication network.
  • a transceiver device such as, but not limited to, a transceiver to implement communication between the device 900 and other devices or a communication network.
  • the first modulated first network weight can be acquired through the communication interface 903 .
  • Bus 904 may include a pathway for communicating information between various components of device 900 (eg, memory 901, processor 902, communication interface 903).
  • the apparatus 900 may also include other devices necessary for normal operation. Meanwhile, according to specific needs, those skilled in the art should understand that the apparatus 900 may further include hardware devices that implement other additional functions. In addition, those skilled in the art should understand that the apparatus 900 may only include the necessary devices for implementing the embodiments of the present application, and does not necessarily include all the devices shown in FIG. 9 .
  • An embodiment of the present application further provides a constellation symbol detection device, including: at least one processor and a communication interface, where the communication interface is used for the constellation symbol detection device to perform information interaction with other communication devices, when the program instruction is in the at least one communication interface.
  • the constellation symbol detection apparatus is made to execute the above method.
  • Embodiments of the present application further provide a computer program storage medium, characterized in that, the computer program storage medium has program instructions, and when the program instructions are directly or indirectly executed, the foregoing method can be implemented.
  • An embodiment of the present application further provides a chip system, characterized in that, the chip system includes at least one processor, and when a program instruction is executed in the at least one processor, the foregoing method can be implemented.
  • a component may be, but is not limited to, a process running on a processor, a processor, an object, an executable, a thread of execution, a program, and/or a computer.
  • an application running on a computing device and the computing device may be components.
  • One or more components may reside in a process and/or thread of execution, and a component may be localized on one computer and/or distributed between 2 or more computers.
  • these components can execute from various computer readable media having various data structures stored thereon.
  • a component may, for example, be based on a signal having one or more data packets (eg, data from two components interacting with another component between a local system, a distributed system, and/or a network, such as the Internet interacting with other systems via signals) Communicate through local and/or remote processes.
  • data packets eg, data from two components interacting with another component between a local system, a distributed system, and/or a network, such as the Internet interacting with other systems via signals
  • the disclosed system, apparatus and method may be implemented in other manners.
  • the apparatus embodiments described above are only illustrative.
  • the division of the units is only a logical function division. In actual implementation, there may be other division methods.
  • multiple units or components may be combined or Can be integrated into another system, or some features can be ignored, or not implemented.
  • the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.
  • the functions, if implemented in the form of software functional units and sold or used as independent products, may be stored in a computer-readable storage medium.
  • the technical solution of the present application can be embodied in the form of a software product in essence, or the part that contributes to the prior art or the part of the technical solution, and the computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and other media that can store program codes .

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Digital Transmission Methods That Use Modulated Carrier Waves (AREA)

Abstract

本申请提供了一种星座符号检测方法及装置，该方法包括：获取第一调制的第一网络权值，其根据第一网络权值确定待训练的第二网络权值，待训练的第二网络权值包括第一网络权值的所有参数，待训练的第二网络权值为第二调制的网络权值，第二调制的调制阶数大于第一调制的调制阶数，然后训练待训练的第二网络权值中除第一网络权值之外的参数。该方法通过将低阶调制的网络权值插入到高阶调制的网络权值中，从而降低了网络权值的存储开销。

Description

一种星座符号检测方法及装置 技术领域
本申请涉及通信领域，并且更具体地，涉及一种星座符号检测方法及装置。
背景技术
在通信系统中，发射端一般会对编码比特进行符号映射，称为符号调制，然后进行载波调制并发送。在接收端，基带处理需要对发送的星座符号进行检测，一种可行的方法是利用神经网络技术的强大的非线性拟合能力来解决星座符号检测的问题。神经网络技术的核心在于将检测的推理参数存储在网络权值中。一般地，由于不同调制下的离散星座符号点数不同，并且对于不同的调制方式，需要分别训练得到不同的网络权值，因此，当系统中需要支持的调制方式增多时，对网络权值的存储开销也会变大。
发明内容
本申请提供一种星座符号检测方法及装置，以便降低不同调制方式下的网络权值的存储开销。
第一方面，提供了一种星座符号检测的方法，该方法包括：获取第一调制的第一网络权值，该第一网络权值为M×N矩阵，M和N均为正整数；根据第一网络权值确定待训练的第二网络权值，该待训练的第二网络权值包括第一网络权值的所有参数，该待训练的第二网络权值为第二调制的网络权值，该第二调制的调制阶数大于第一调制的调制阶数，该待训练的第二网络权值为(M+a)×(N+b)矩阵，其中，a和b均为正整数；训练该待训练的第二网络权值中除第一网络权值之外的参数。
在本申请实施例中，通过将第一调制的第一网络权值中的全部参数插入到第二调制对应的待训练的第二网络权值中，并训练该待训练的第二网络权值中除第一网络权值之外的参数，从而降低不同调制方式下的网络权值的存储开销。
结合第一方面，在第一方面的某些实现方式中，第一调制与第二调制为具有相同调制方式的相邻阶调制。
结合第一方面，在第一方面的某些实现方式中，第一网络权值包括第一权重和第一偏置，待训练的第二网络权值包括第二权重和第二偏置，该第二权重包括第一权重的所有参数，且该第二偏置包括第一偏置的所有参数。
结合第一方面，在第一方面的某些实现方式中，第一权重包括至少两个子权重，将该至少两个子权重分别插入在第二权重中，该至少两个子权重的插入位置不重合。
结合第一方面，在第一方面的某些实现方式中，第一偏置包括至少两个子偏置，将该至少两个子偏置分别插入在第二偏置中，至少两个子偏置的插入位置不重合。
结合第一方面，在第一方面的某些实现方式中，第一网络权值包括至少两个子网络权值，至少两个子网络权值包括权重和偏置，将该至少两个子网络权值分别插入在待训练的 第二网络权值中，至少两个子网络权值的插入位置不重合。
结合第一方面，在第一方面的某些实现方式中，待训练的第二网络权值包括第一网络权值的所有参数，包括将第一网络权值的M×N个参数均匀等间隔插入到待训练的第二网络权值的(M+a)×(N+b)矩阵中。
结合第一方面，在第一方面的某些实现方式中，第一调制和第二调制的调制方式包括正交振幅调制QAM。
结合第一方面，在第一方面的某些实现方式中，该方法还包括：获取训练后的第二网络权值；根据训练后的第二网络权值检测星座符号向量。
第二方面，提供了一种星座符号检测装置，该装置包括：获取单元，用于获取第一调制的第一网络权值，第一网络权值为M×N矩阵，M和N均为正整数；处理单元，处理单元用于根据第一网络权值确定待训练的第二网络权值，待训练的第二网络权值包括第一网络权值的所有参数，待训练的第二网络权值为第二调制的网络权值，第二调制的调制阶数大于第一调制的调制阶数，待训练的第二网络权值为(M+a)×(N+b)矩阵，其中，a和b均为正整数；所述处理单元还用于训练待训练的第二网络权值中除第一网络权值之外的参数。
在本申请实施例中，通过将第一调制的第一网络权值中的全部参数插入到第二调制对应的待训练的第二网络权值中，并训练该待训练的第二网络权值中除第一网络权值之外的参数，从而降低不同调制方式下的网络权值的存储开销。
结合第二方面，在第二方面的某些实现方式中，第一调制与第二调制为具有相同调制方式的相邻阶调制。
结合第二方面，在第二方面的某些实现方式中，第一网络权值包括第一权重和第一偏置，待训练的第二网络权值包括第二权重和第二偏置，第二权重包括第一权重的所有参数，且第二偏置包括第一偏置的所有参数。
结合第二方面，在第二方面的某些实现方式中，第一权重包括至少两个子权重，将该至少两个子权重分别插入在第二权重中，该至少两个子权重的插入位置不重合。
结合第二方面，在第二方面的某些实现方式中，第一偏置包括至少两个子偏置，将该至少两个子偏置分别插入在第二偏置中，该至少两个子偏置的插入位置不重合。
结合第二方面，在第二方面的某些实现方式中，第一网络权值包括至少两个子网络权值，至少两个子网络权值包括权重和偏置，将该至少两个子网络权值分别插入在待训练的第二网络权值中，该至少两个子网络权值的插入位置不重合。
结合第二方面，在第二方面的某些实现方式中，待训练的第二网络权值包括第一网络权值的所有参数，包括将第一网络权值的M×N个参数均匀等间隔插入到待训练的第二网络权值的(M+a)×(N+b)矩阵中。
结合第二方面，在第二方面的某些实现方式中，第一调制和第二调制的调制方式包括正交振幅调制QAM。
结合第二方面，在第二方面的某些实现方式中，获取单元还用于获取训练后的第二网络权值；处理单元还用于根据训练后的第二网络权值检测星座符号向量。
第三方面，提供了一种星座符号检测装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器 用于执行上述第一方面中的任意一种实现方式中的方法。
第四方面，本申请提供一种计算机可读存储介质，计算机可读存储介质中存储有计算机指令，当计算机指令在计算机上运行时，使得计算机执行第一方面中的任意一种实现方式中的方法。
第五方面，本申请提供一种计算机程序产品，所述计算机程序产品包括计算机程序代码，当所述计算机程序代码在计算机上运行时，使得计算机执行第一方面中的任意一种实现方式中的方法。
第六方面，提供一种芯片，所述芯片包括处理器与数据接口，所述处理器通过所述数据接口读取存储器上存储的指令，以执行第一方面中的任意一种实现方式中的方法。
结合第六方面，在第六方面的某些实现方式中，所述芯片还可以包括存储器，所述存储器中存储有指令，所述处理器用于执行所述存储器上存储的指令，当所述指令被执行时，所述处理器用于执行第一方面中的任意一种实现方式中的方法。
附图说明
图1是本申请实施例提供的一种全连接神经网络的结构示意图；
图2是本申请实施例提供的一种星座符号检测的方法的示意性流程图；
图3是本申请实施例提供的一种网络权值嵌套方法的示意性流程图；
图4是本申请实施例提供的一种网络权值嵌套方法再一示意性流程图；
图5是本申请实施例提供的一种网络权值嵌套方法再一示意性流程图；
图6是本申请实施例提供的一种网络权值嵌套方法再一示意性流程图；
图7是本申请实施例提供的一种星座符号检测装置的结构示意图；
图8是本申请实施例提供的一种星座符号检测装置的再一结构示意图；
图9是本申请实施例提供的一种星座符号检测装置的再一结构示意图。
具体实施方式
下面将结合附图，对本申请中的技术方案进行描述。
本申请实施例的技术方案可以应用于各种通信系统，例如：全球移动通讯(global system of mobile communication，GSM)系统、码分多址(code division multiple access，CDMA)系统、宽带码分多址(wideband code division multiple access，WCDMA)系统、通用分组无线业务(general packet radio service，GPRS)、长期演进(long term evolution，LTE)系统、LTE频分双工(frequency division duplex，FDD)系统、LTE时分双工(time division duplex，TDD)、通用移动通信系统(universal mobile telecommunication system，UMTS)、全球互联微波接入(worldwide interoperability for microwave access，WiMAX)通信系统、第五代(5th generation，5G)系统或新无线(new radio，NR)、或者未来演进的无线通信系统等。
首先对本申请实施例中涉及的相关技术或概念作简单介绍。
1、全连接神经网络
全连接神经网络又叫做多层感知机(multi-layer perceptron,MLP)。参见图1，图1是全连接神经网络的结构示意图。如图1所示，一个MLP包括一个输入层(最左侧)、 一个输出层(最右侧)和多个隐藏层(图1中仅示出两层)。每个层包括若干个节点，称为神经元。相邻两层的神经元两两相连，同层的神经元之间不连接。其中，各层所包括的节点的数量称为该层的宽度。一般地，输入层代表特征向量，输入层的每一个神经元代表一个特征值。
以相邻两层的神经元来看，下一层的神经元的输出h为与所有与之相连的上一层的神经元x之间的函数关系可以表示为式(1)。
h＝f(wx+b)                             (1)
其中，w为权重矩阵，b为偏置向量，f为激活函数。
因此，MLP的输出可以递归表达为式(2)。
y＝f n(w nf n-1(...)+b n)                         (2)
2、损失函数
在训练神经网络的过程中，因为希望神经网络的输出尽可能的接近真正想要预测的值，所以可以通过比较当前网络的预测值和真正想要的目标值，再根据两者之间的差异情况来更新每一层神经网络的权重向量(当然，在第一次更新之前通常会有初始化的过程，即为深度神经网络中的各层预先配置参数)，比如，如果网络的预测值高了，就调整权重向量让它预测低一些，不断地调整，直到神经网络能够预测出真正想要的目标值或与真正想要的目标值非常接近的值。因此，就需要预先定义“如何比较预测值和目标值之间的差异”，这便是损失函数(loss function)或目标函数(objective function)，它们是用于衡量预测值和目标值的差异的重要方程。其中，以损失函数举例，损失函数的输出值(loss)越高表示差异越大，那么神经网络的训练就变成了尽可能缩小这个loss的过程。
3、反向传播算法
神经网络可以采用误差反向传播(back propagation，BP)算法在训练过程中修正初始的神经网络模型中参数的大小，使得神经网络模型的重建误差损失越来越小。具体地，前向传递输入信号直至输出会产生误差损失，通过反向传播误差损失信息来更新初始的神经网络模型中参数，从而使误差损失收敛。反向传播算法是以误差损失为主导的反向传播运动，旨在得到最优的神经网络模型的参数，例如权重矩阵。
4、最大似然检测方法(maximal likelihood detection，MLD)
理论上，最大似然算法是最优的检测方法，最大似然检测是让接收信号和所有可能的发送信号进行比较，产生最大似然估计值。它的原理是在发射信号(或信号向量空间)中寻找其经过信道变换以后到接收信号距离最小的那个发射信号(或信号向量空间)。
假设调制信号星座图空间大小为Q，对于发送流数为M的MIMO系统，需要的复杂度为
Figure PCTCN2020135430-appb-000001
随着流数和符号调制阶数呈指数增长。
在通信系统中，发射端一般会对编码比特进行符号映射，然后进行载波调制并发送。不同的调制阶数包含了不同的数目的星座集合，以正交振幅调制(quadrature amplitude modulation，QAM)为例，例如16QAM包含了16个可选的星座集合，64QAM包含了64个可选的星座集合。考虑到QAM调制时在复平面分实部和虚部独立调制(通常称为I/Q路)，那么，当调制方式为16QAM时，每条支路有4种状态，当调制方式为64QAM时，每条支路有8种状态。总之，随着调制方式的升高，QAM调制可选星座集合的状态数成倍增加，QAM调制可选星座集合的状态数成倍增加。在接收端，基带处理需要对发送星 座符号进行检测。比如，对于多天线技术(multi-input multi-output，MIMO)中的正交频分复用技术(orthogonal frequency division multiplexing，OFDM)而言，接收信号模型可以表示为：
y＝Hs+n                               (3)
其中，y表示频域接收信号的向量，H表示MIMO系统的信道矩阵，s表示星座符号向量，n表示加性高斯噪声向量。所以，在MIMO系统中，一般通过导频信息得到y和H，恢复s。通常用MLD检测或者线性最小均方差(linear minimum mean square error，LMMSE)检测，LMMSE检测将发送星座符号建模为随机高斯分布。MLD检测充分考虑发送星座符号的调制信息，最优化调制星座集合的准确后验概率，在高斯噪声模型下具有最佳性能。而神经网络技术具有强大的非线性拟合能力，用于解决MLD检测问题，以低复杂度逼近最优MLD性能。
神经网络技术的核心在于将MLD检测的推理参数存储在网络权值中，用于近似计算离散星座集合的后验概率。由于不同调制方式下离散星座点数不同，并且，对于不同的调制方式，一般需要分别训练得到不同的网络权值。基于此原理，本申请提出一种网络权值嵌套的方法。
本申请以MIMO系统中的接收端利用神经网络技术设计的一种MLD检测算法为例，应理解，检测算法还可以为近似MLD算法，例如球形译码(sphere decoding，SD)算法、K-best树搜算法等。检测算法还可以为各种线性或线性增强算法，例如线性最小均方误差(linear minimum mean square error，LMMSE)算法、串行干扰抵消(successive interference cancellation，SIC)算法、期望传播(expectation propagation，EP)检测算法等，在此不做过多限定。
示例性的，MLD算法的输入向量可以表示为X in，神经网络函数用f表示，相应的网络权值集合用θ表示，网络输出用X out表示，即有，
X out＝f(X in；θ)                             (4)
对于一种星座符号检测网络而言，一般具有固定的网络结构，对于不同的调制方式具有相同的输入，即X in相同。也就是说，对于某种近似MLD的神经网络算法而言，对于不同的调制方式，其选择的网络输入特征(即输入数据)和神经网络结构相同。在本申请实施例中以只有一个隐藏层为例，即f相同。虽然神经网络的结构相同，数据输入类型也相同，但是网络权值θ往往不同，这与网络权值θ中参数的个数和取值有关。例如，对于4阶调制而言，采用预置的神经网络结构，但是只选取隐藏层中的8个节点，对于16阶而言，选取隐藏层的16个节点。以接收信号模型(3)为例，定义输入向量为X in＝vector{y,H,H H,Y}，神经网络结构以包含一个隐藏层为例，其MLP网络结构f可以表示为，
X out＝W 2[ReLu(W 1X in+b 1)]+b 2                     (5)
在公式(5)中，选用了线性整流单元(rectified linear unit，ReLu)作为隐藏层的激活函数。
可选的，隐藏层的激活函数还可以包括Sigmoid函数、Tanh函数等。应理解，激活函数是给神经元引入了非线性因素，使得神经网络可以任意逼近任何非线性函数，这样神经网络就可以应用到众多的非线性模型中。如果不用激活函数，每一层的输出都是上层的输 入的线性函数，无论神经网络函数有多少层，输出都是输入的线性组合。本申请对激活函数的选择不做任何限定。
在该神经网络结构中，网络权值θ＝{W 1,b 1,W 2,b 2}。其中，W 1表示输入层到隐藏层的权重，b 1表示输入层到隐藏层的偏置，W 2表示隐藏层到输出层的权重，b 2表示隐藏层到输出层的偏置。
应理解，网络权值一般是与调制方式有关的矩阵或向量，也就是说调制方式会影响矩阵的大小和矩阵中各元素的取值，即不同的调制方式，网络权值的维度也不同，调制方式越高，网络权值维度越高。因此，本申请可以实现网络权值嵌套，从而减少网络权值存储。
图2是本申请实施例提供的一种网络权值内嵌方法的示意图。
S210，获取第一调制的第一网络权值，该第一网络权值为M×N矩阵，M和N均为正整数；
S220，根据第一网络权值确定待训练的第二网络权值，该待训练的第二网络权值包括第一网络权值的所有参数，该待训练的第二网络权值为第二调制的网络权值，该第二调制的调制阶数大于第一调制的调制阶数，该待训练的第二网络权值为(M+a)×(N+b)矩阵，其中，a和b均为正整数；
S230，训练该待训练的第二网络权值中除第一网络权值之外的参数。
具体而言，获得第一调制的第一网络权值，通常可以通过训练输入的值来得到。也就是说，当第一调制的调制阶数是带有调制的通信系统中的最低阶时，第一调制的第一网络权值一般会通过训练预定义的值得到，或者当第一调制的上一阶调制的网络权值作为输入值时，可以训练得到。
应理解，本申请实施例中的网络权值为多维矩阵，其中第一网络权值为M×N矩阵时，第二网络权值可以为(M+a)×(N+b)矩阵，其中，M、N、a和b均是正整数。在步骤S220中，根据第一网络权值确定待训练的第二网络权值，该待训练的第二网络权值包括第一网络权值的所有参数为将M×N中所有的参数插入到(M+a)×(N+b)矩阵中。
示例性的，第一网络权值为4×4阶矩阵，其中含有16个参数，第二网络权值为8×8阶矩阵，其中含有64个参数，将4×4阶矩阵中的16个参数插入到8×8阶矩阵中，也就是说8×8阶矩阵的64个参数有16个参数是来自4×4阶矩阵的，而剩下的48个参数可以是训练得到的。在步骤S230中，训练待训练的第二网络权值中除第一网络权值之外的参数，即训练这48个参数，也就是在64个参数中，来自4×4阶矩阵的16个参数是不被训练的。
应理解上述4×4阶矩阵和8×8阶矩阵只是示例性说明。第一网络权值还可以是3×2阶，第二网络权值是5×6阶，也就是说只要第二网络权值的维度大于第一网络权值均可，本申请实施例对第一网络权值、第二网络权值的维度不做过多赘述。
示例性的，第一调制可以是QAM调制中的4QAM或16QAM或64QAM调制等，其调制方式和调制阶数对应如下表1所示，只要是不是最高阶的调制均可。第一调制还可以是其他调制，例如多进制相位(multiple phase shift keying，MPSK)调制中的2PSK、4PSK或8PSK等，其调制方式和调制阶数对应如下表2所示。本申请以下将以第一调制为QAM调制中的4QAM为例进行说明，但本申请还可以用于其他调制类型的不同调制阶数，在此不对第一调制做任何限定。
表1 QAM调制方式和调制阶数
调制方式 调制阶数Qm
4QAM 2
16QAM 4
64QAM 8
256QAM 16
表2 MPSK调制方式和调制阶数
调制方式 调制阶数Qm
2PSK(BPSK) 1
4PSK(QPSK) 2
8PSK 4
作为一个可选的示例，如图3所示，通信系统中存在4QAM、16QAM和64QAM。也就是说在该系统中最高阶调制为64QAM，所以为了节省开销，将比64QAM调制阶数低的调制的网络权值插入到64QAM的网络权值中。当然，通信系统中也可以同时存在MPSK和QAM调制方式，例如，BPSK，QPSK，16QAM，64QAM以及256QAM时，也可以将BPSK或QPSK等低阶调制的网络权值包含在16QAM，64QAM，以及256QAM等高阶调制的网络权值中。在图3中，4QAM的调制阶数比16QAM的调制阶数低，16QAM调制阶数比64QAM调制阶数低，所以先把4QAM的网络权值θ (4QAM)插入到16QAM的网络权值中，训练16QAM的网络权值，然后把训练好的16QAM中的网络权值θ (16QAM)插入到64QAM的网络权值中，训练64QAM的网络权值。应理解，当把4QAM的网络权值插入到16QAM的网络权值中时，此时4QAM为第一调制，4QAM的网络权值θ (4QAM)为第一网络权值，16QAM为第二调制，16QAM的网络权值θ (16QAM)为第二网络权值；当把16QAM的网络权值θ (16QAM)插入到64QAM的网络权值中时，此时16QAM为第一调制，16QAM的网络权值为第一网络权值，64QAM为第二调制，64QAM的网络权值θ (64QAM)为第二网络权值。
如图3所示，4QAM的网络权值θ (4QAM)为已经训练好的网络权值，在对16QAM的网络权值训练过程中，首先把θ (4QAM)中的每个参数全部插入到待训练的16QAM的网络权值中，训练更新除θ (4QAM)之外的参数，得到16QAM的网络权值θ (16QAM)，然后把θ (16QAM)中的每个参数全部插入到待训练的64QAM的网络权值中，训练更新除θ (16QAM)之外的参数，得到64QAM的网络权值θ (64QAM)，这样就可以实现将低阶调制的网络权值插入到高阶调制的网络权值中。
应理解，上述实施例仅作为一种较为优选的实施方式，可以极大地节省网络权值存储开销，当然其还可以是4QAM的网络权值插入到64QAM的网络权值中，只要能实现低阶调制的网络权值插入到高阶调制的网络权值均属于本申请实施例保护的范围，对此不做过多赘述。
同样地，如图4所示，当系统中存在从4 1QAM到最高阶为4 q+1QAM调制时，此时其最节省网络权值存储开销的插入方式是从低到高依次插入，即，将4 1QAM的网络权值
Figure PCTCN2020135430-appb-000002
插入到4 2QAM的网络权值中，训练除
Figure PCTCN2020135430-appb-000003
之外的参数得到
Figure PCTCN2020135430-appb-000004
然后将4 2QAM的网络权值
Figure PCTCN2020135430-appb-000005
插入到4 3QAM的网络权值中，训练除
Figure PCTCN2020135430-appb-000006
之外的参数得 到
Figure PCTCN2020135430-appb-000007
并以此类推，将4 q-1QAM的网络权值
Figure PCTCN2020135430-appb-000008
插入到4 qQAM的网络权值中，训练除
Figure PCTCN2020135430-appb-000009
之外的参数得到
Figure PCTCN2020135430-appb-000010
然后将4 q QAM的网络权值
Figure PCTCN2020135430-appb-000011
插入到4 q+1QAM的网络权值中，训练除
Figure PCTCN2020135430-appb-000012
之外的参数得到
Figure PCTCN2020135430-appb-000013
最终实现将低阶调制的网络权值插入到高阶调制的网络权值中。
应理解，图4示出的实施例也仅作为一种较为优选的实施例，当然还可以将4 1QAM的网络权值插入到任一比其阶数高的网络权值中，对此只要能实现将低阶调制的网络权值插入到高阶网络权值均属于本申请实施例保护的范围，在此不做过多赘述。
应理解，图4中，仅以4 q阶调制为例进行说明，其还可以是2 q阶、3 q阶或者q阶，q为正整数，也就是说，通信系统中，只要存在低阶调制和高阶调制均可采用本申请实施例，对此不做过多赘述。
图5示出了一种网络权值的嵌套形式，如图5所示，同样以低阶调制为4QAM为例，其中高阶调制分别为16QAM和64QAM。应理解，在此仅仅是示例性说明，低阶调制可以为任一调制，高阶调制也为任一调制，且调制方式的数量不限。
图5中的(a)、(b)和(c)表示三种根据第一网络权值确定待训练的第二网络权值，其中图5中的(a)表示4QAM中的所有参数插入到16QAM矩阵的左上角，训练好的16QAM中的所有参数插入到64QAM矩阵的左上角。其用矩阵大致可以表示为：
4QAM网络权值为2×2阶矩阵：
Figure PCTCN2020135430-appb-000014
16QAM网络权值为4×4阶矩阵：
Figure PCTCN2020135430-appb-000015
64QAM网络权值为8×8阶矩阵：
Figure PCTCN2020135430-appb-000016
应理解，上述不同调制方式对应的矩阵，及矩阵中的各个参数只是示例性说明，其具体的矩阵大小和参数值可以根据实际情况，自行确定，在此不做过多赘述。通常调制阶数越高，其对应的网络权值矩阵的维度越大。一种可能的实现方式中，网络权值矩阵的维度和其对应的调制阶数相等，如上述示例性说明中各调制阶数对应的网络权值矩阵。另一种可能的实现方式中，网络权值矩阵的维度也可以不必和其对应的调制阶数相同，例如16QAM的网络权值矩阵为3×3阶矩阵，只要比低阶调制的网络权值矩阵的维度大即可。
图5中的(b)表示4QAM中的所有参数插入到16QAM矩阵的中间部分，训练好的16QAM中的所有参数插入到64QAM矩阵的中间部分。其用矩阵大致可以表示为：
4QAM网络权值为2×2阶矩阵：
Figure PCTCN2020135430-appb-000017
16QAM网络权值为4×4阶矩阵：
Figure PCTCN2020135430-appb-000018
16QAM网络权值为8×8阶矩阵：
Figure PCTCN2020135430-appb-000019
应理解，上述不同调制方式对应的矩阵，及矩阵中的各个参数只是示例性说明，其具体的矩阵大小和参数值可以根据实际情况，自行确定，在此不做过多赘述。
图5中的(c)表示4QAM中网络权值包括四个子网络权值，将这四个子网络权值插入到16QAM矩阵的任一位置，且四个位置互不重合，训练好的16QAM中的所有参数分为四个子网络权值，同样的把这四个子网络权值插入到64QAM矩阵的任一位置，且四个位置互不重合，这里我们采用将其插入到左上角位置为例进行说明。其用矩阵大致可以表示为：
4QAM网络权值为2×2阶矩阵：
Figure PCTCN2020135430-appb-000020
16QAM网络权值为4×4阶矩阵：
Figure PCTCN2020135430-appb-000021
16QAM网络权值为8×8阶矩阵：
Figure PCTCN2020135430-appb-000022
应理解，上述不同调制方式对应的矩阵，及矩阵中的各个参数只是示例性说明，其具体的矩阵大小和参数值可以根据实际情况，自行确定，在此不做过多赘述。
应理解，上述将调制方式中的子网络权值是均匀分块且都插入到左上角，其还可以是中间位置，也可以是右下角。因此在实际的应用中，其子网络权值分块可以不均匀，且各 个子网络权值的插入方式也可以不同，在此不做过多的限定。
图6示出了一种网络权值的嵌套形式，如图6所示，同样以低阶调制为4QAM为例，其中高阶调制以16QAM为例进行说明。
在低阶调制对应的网络权值中，存在权重(第一权重)和偏置(第一偏置)，在待训练的高阶调制对应的网络权值中，同样也存在权重(第二权重)和偏置(第二偏置)，其中可以将低阶调制的第一权重中的参数全部插入到第二权重中，且将低阶调制的第一偏置中的参数全部插入到第二偏置中。
在图6的(a)中，以4QAM的权重W 1为4×4阶的矩阵，偏置b 1为4×1阶的矩阵，16QAM的权重W 2为8×4阶的矩阵，偏置b 2为8×1阶的矩阵为例。4QAM中的权重以行为单位均匀等间隔的插入到16QAM中待训练的权重矩阵中，同样地，4QAM中的偏置以行为单位均匀等间隔的插入到16QAM中待训练的偏置矩阵中。即4QAM对应的权重矩阵有4行，16QAM对应的权重矩阵有8行，则将4QAM权重矩阵中的第一行插入到16QAM权重矩阵中的第一行，将4QAM权重矩阵中的第二行插入到16QAM权重矩阵中的第三行，将4QAM权重矩阵中的第三行插入到16QAM权重矩阵中的第五行，将4QAM权重矩阵中的第四行插入到16QAM权重矩阵中的第七行，4QAM的偏置矩阵插入方式也是如此，对此不做过多的描述。
在图6的(b)中，以4QAM的权重W 1为4×4阶的矩阵，偏置b 1为4×1阶的矩阵，16QAM的权重W 2为8×8阶的矩阵，偏置b 2为8×1阶的矩阵为例。4QAM中的权重以每个参数为单位均匀等间隔的插入到16QAM中待训练的权重矩阵中，同样地，4QAM中的偏置以每个参数为单位均匀等间隔的插入到16QAM中待训练的偏置矩阵中。即4QAM对应的权重矩阵有16个参数，64QAM对应的权重矩阵有64个参数，可以将16个参数均匀等间隔的插入到64个参数中，其具体实现方式可以为，16QAM权重矩阵中第一行第一列的参数插入到64QAM权重矩阵中的第一行第一列，16QAM权重矩阵中第一行第二列的参数插入到64QAM权重矩阵中的第一行第三列，16QAM权重矩阵中第一行第三列的参数插入到64QAM权重矩阵中的第一行第五列，16QAM权重矩阵中第一行第四列的参数插入到64QAM权重矩阵中的第一行第七列，16QAM权重矩阵中第二行第一列的参数插入到64QAM权重矩阵中的第二行第一列，16QAM权重矩阵中第二行第二列的参数插入到64QAM权重矩阵中的第二行第三列等等，直至将16个参数均匀等间隔的插入，4QAM的偏置矩阵插入方式也是如此，对此不做过多的描述。
在图6的(c)中，以4QAM的权重W 1为4×4阶的矩阵，偏置b 1为4×1阶的矩阵，16QAM的权重W 2为8×4阶的矩阵，偏置b 2为8×1阶的矩阵为例。4QAM中的权重以整个权重矩阵为单位插入到16QAM中待训练的权重矩阵中，同样地，4QAM中的偏置以整个权重矩阵为单位插入到16QAM中待训练的偏置矩阵中。具体而言，4QAM对应的偏重矩阵为4×4阶矩阵，共有4行4列，而16QAM对应的偏重矩阵为8×4阶，共有8行4列，则将4QAM中4行参数插入到16QAM的前4行，4QAM的偏置矩阵插入方式也是如此，对此不做过多的描述。
在图6的(d)中，以4QAM的权重W 1为4×4阶的矩阵，偏置b 1为4×1阶的矩阵，16QAM的权重W 2为8×8阶的矩阵，偏置b 2为8×1阶的矩阵为例。4QAM中的权重以整个权重矩阵为单位插入到16QAM中待训练的权重矩阵的左上角中，同样地，4QAM中的 偏置以整个权重矩阵为单位插入到16QAM中待训练的偏置矩阵中。即4QAM对应的权重矩阵有4行4列，16QAM对应的权重矩阵有8行8列，则可以将4QAM的4行8列中的所有参数插入到16QAM对应的前4行且前4列中。
应理解，上述的图6中的(a)至(d)中的所有插入方式仅仅为示例性说明，其也可以组合，比如，低阶调制的权重矩阵中的参数均匀等间隔插入高阶调制的权重矩阵中，低阶调制的偏置矩阵的中参数以整个矩阵为单位插入到高阶调制的权阵中等，也就是说，只要能实现将低阶调制中的参数插入到高阶调制的矩阵中，且每个参数插入的位置不重合，均属于本申请实施例要求保护的范围，在此不做过多赘述。
本申请实施例中，把低阶调制的网络权值中的全部参数插入到高阶调制的网络权值，节省了网络权值的存储开销。例如，4QAM对应的网络权值为2×2的矩阵，16QAM对应的网络权值为4×4的矩阵，64QAM对应的网络权值为8×8的矩阵，256QAM对应的网络权值为16×16的矩阵，如果未采用将低阶调制的网络权值中的全部参数插入到高阶调制的网络权值时，其网络权值的存储开销为2×2+4×4+8×8+16×16＝340，但采用本申请实施例后，其可能仅需要16×16＝256个存储参数开销，此时可节省约25％的网络权值存储开销。
通过上述将低阶调制的网络权值插入到高阶的网络权值中，并训练高阶调制的网络权值，从而节省了网络权值的存储开销。当进行星座符号检测时，只需要按照对应的调制方式，通过推理得到该调制方式对应的网络权值，从而可以进行星座符号的检测。也就是说，本申请实施例提供的方法得到的是最高阶调制方式对应的网络权值，在进行星座符号检测时按照其对应的的调制阶数获取相应的网络权值以进行检测。
在本申请又一实施例中，系统可以保存根据前述实施例得到的高阶调制方式对应的网络权值矩阵，用于星座符号检测。可以根据星座符号对应的调制方式从高阶调制方式的网络权值矩阵中获取星座符号对应的调制方式的网络权值，并进行检测。其中，高阶调制方式的调制阶数大于或等于星座符号对应的调制方式，高阶调制方式的网络权值矩阵包括了星座符号对应的调制方式的网络权值。其中，高阶调制方式的网络权值如何嵌套低阶调制方式的网络权值，如何从高阶调制方式的网络权值得到低阶调制方式的网络权值可以参考前述各实施例，此处不再赘述。
图7是本申请实施例提供的一种星座符号检测装置的结构示意图。图7所示的装置700包括获取单元701和处理单元702。
获取单元701，用于获取第一调制的第一网络权值，该第一网络权值为M×N矩阵，M和N均为正整数；
处理单元702，用于根据第一网络权值确定待训练的第二网络权值，该待训练的第二网络权值包括第一网络权值的所有参数，该待训练的第二网络权值为第二调制的网络权值，该第二调制的调制阶数大于第一调制的调制阶数，该待训练的第二网络权值为(M+a)×(N+b)矩阵，其中，a和b均为正整数；
处理单元702还用于训练该待训练的第二网络权值中除第一网络权值之外的参数。
可选地，第一调制与第二调制为具有相同调制方式的相邻阶调制。
示例性地，以5G系统为例，在5G系统中存在4QAM，16QAM，64QAM，则4QAM和16QAM为相邻阶调制，16QAM和64QAM为相邻阶调制。换言之，相邻阶可以是1， 2，3，……q这样相邻的阶数，或者可以是1，2，4，……2 q这样相邻的阶数，或者可以是4，16，64……4 q这样相邻的阶数，其中，q为正整数，只要是任意两个调制方式之间不存在比其中一个调制阶数大，比另一个调制阶数小的调制方式，均可认为相邻阶调制，在此不做过多限定。
可选地，第一网络权值包括第一权重和第一偏置，待训练的第二网络权值包括第二权重和第二偏置，第二权重包括第一权重的所有参数，且第二偏置包括第一偏置的所有参数。
可选地，第一权重包括至少两个子权重，将至少两个子权重分别插入在第二权重中，至少两个子权重的插入位置不重合。
可选地，第一偏置包括至少两个子偏置，将至少两个子偏置分别插入在第二偏置中，至少两个子偏置的插入位置不重合。
可选地，第一网络权值包括至少两个子网络权值，至少两个子网络权值包括权重和偏置，将至少两个子网络权值分别插入在待训练的第二网络权值中，至少两个子网络权值的插入位置不重合。
可选地，待训练的第二网络权值中包括第一网络权值的所有参数，包括将第一网络权值中M×N个参数均匀等间隔插入到待训练的第二网络权值的(M+a)×(N+b)矩阵中。
可选地，第一调制和第二调制的调制方式包括正交振幅调制QAM。
可选地，获取单元701还用于获取训练后的第二网络权值，处理单元702还用于根据训练后的第二网络权值检测星座符号向量。
图8示出了本申请实施例提供的另一种星座符号检测装置的结构示意图。图8所示的装置800包括存储单元801、获取单元802和处理单元803。
示例性地，该存储单元801用于存储高阶调制方式对应的网络权值矩阵。
获取单元802用于根据星座符号对应的调制方式从高阶调制方式的网络权值矩阵中获取该星座符号对应的调制方式的网络权值，以便处理单元803进行星座符号检测。其中，高阶调制方式的调制阶数大于或等于星座符号对应的调制方式，高阶调制方式的网络权值矩阵包括了星座符号对应的调制方式的网络权值。
一种可能实现的方式中，高阶调制方式对应的网络权值矩阵可以是装置800根据图2至图6任一方法训练得到的，其具体的实现流程/步骤可以参照装置700。另一种可能实现的实现方式中，高阶调制方式对应的网络权值矩阵可以是通过其他装置训练得到的，例如装置700，然后装置800从其他装置中获取该网络权值矩阵并存储在存储单元801中，此时装置800和其他装置(例如装置700)是完全不同的装置，在此不做过多的赘述。
图9是本申请实施例的星座符号检测装置的硬件结构示意图。图9所示的装置900(该装置900具体可以是一种计算机设备)包括存储器901、处理器902、通信接口903以及总线901。其中，存储器901、处理器902、通信接口903通过总线904实现彼此之间的通信连接。
存储器901可以是只读存储器(read only memory，ROM)，静态存储设备，动态存储设备或者随机存取存储器(random access memory，RAM)。存储器901可以存储程序，当存储器901中存储的程序被处理器902执行时，处理器902用于执行本申请实施例的星座符号检测方法的各个步骤。
处理器902可以采用通用的中央处理器(central processing unit，CPU)，微处理器， 应用专用集成电路(application specific integrated circuit，ASIC)，图形处理器(graphics processing unit，GPU)或者一个或多个集成电路，用于执行相关程序，以实现本申请实施例的星座符号检测方法。
处理器902还可以是一种集成电路芯片，具有信号的处理能力。在实现过程中，本申请的星座符号检测方法的各个步骤可以通过处理器902中的硬件的集成逻辑电路或者软件形式的指令完成。
上述处理器902还可以是通用处理器、数字信号处理器(digital signal processing，DSP)、专用集成电路(ASIC)、现成可编程门阵列(field programmable gate array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。结合本申请实施例所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器901，处理器902读取存储器901中的信息，结合其硬件完成本星座符号检测装置中包括的单元所需执行的功能，或者执行本申请方法实施例的星座符号检测方法。
通信接口903使用例如但不限于收发器一类的收发装置，来实现装置900与其他设备或通信网络之间的通信。例如，可以通过通信接口903获取第一调制的第一网络权值。
总线904可包括在装置900各个部件(例如，存储器901、处理器902、通信接口903)之间传送信息的通路。
应注意，尽管上述装置900仅仅示出了存储器、处理器、通信接口，但是在具体实现过程中，本领域的技术人员应当理解，装置900还可以包括实现正常运行所必须的其他器件。同时，根据具体需要，本领域的技术人员应当理解，装置900还可包括实现其他附加功能的硬件器件。此外，本领域的技术人员应当理解，装置900也可仅仅包括实现本申请实施例所必须的器件，而不必包括图9中所示的全部器件。
本申请实施例还提供一种星座符号检测装置，包括：至少一个处理器和通信接口，所述通信接口用于所述星座符号检测装置与其他通信装置进行信息交互，当程序指令在所述至少一个处理器中执行时，使得所述星座符号检测装置执行上文中的方法。
本申请实施例还提供一种计算机程序存储介质，其特征在于，所述计算机程序存储介质具有程序指令，当所述程序指令被直接或者间接执行时，使得前文中的方法得以实现。
本申请实施例还提供一种芯片系统，其特征在于，所述芯片系统包括至少一个处理器，当程序指令在所述至少一个处理器中执行时，使得前文中的方法得以实现。
在本申请中，如果没有特别说明，“第一”、“第二”等词语仅用于区分不同的个体，例如，例如本申请实施例涉及的“第一网络权值”与“第二网络权值”为两个不同的网络权值，除此之外不存在其它限定。
在本说明书中使用的术语“部件”、“模块”、“系统”等用于表示计算机相关的实体、硬件、固件、硬件和软件的组合、软件、或执行中的软件。例如，部件可以是但不限于，在处理器上运行的进程、处理器、对象、可执行文件、执行线程、程序和/或计算机。通过图示，在计算设备上运行的应用和计算设备都可以是部件。一个或多个部件可驻留在 进程和/或执行线程中，部件可位于一个计算机上和/或分布在2个或更多个计算机之间。此外，这些部件可从在上面存储有各种数据结构的各种计算机可读介质执行。部件可例如根据具有一个或多个数据分组(例如来自与本地系统、分布式系统和/或网络间的另一部件交互的二个部件的数据，例如通过信号与其它系统交互的互联网)的信号通过本地和/或远程进程来通信。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(read-only memory，ROM)、随机存取存储器(random access memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (21)

  1. 一种星座符号检测方法，其特征在于，包括：
    获取第一调制的第一网络权值，所述第一网络权值为M×N矩阵，M和N均为正整数；
    根据所述第一网络权值确定待训练的第二网络权值，所述待训练的第二网络权值包括所述第一网络权值的所有参数，所述待训练的第二网络权值为第二调制的网络权值，所述第二调制的调制阶数大于所述第一调制的调制阶数，所述待训练的第二网络权值为(M+a)×(N+b)矩阵，其中，a和b均为正整数；
    训练所述待训练的第二网络权值中除所述第一网络权值之外的参数。
  2. 根据权利要求1所述的方法，其特征在于，所述第一调制与所述第二调制为具有相同调制方式的相邻阶调制。
  3. 根据权利要求1或2所述的方法，其特征在于，所述第一网络权值包括第一权重和第一偏置，所述待训练的第二网络权值包括第二权重和第二偏置，所述第二权重包括所述第一权重的所有参数，且所述第二偏置包括所述第一偏置的所有参数。
  4. 根据权利要求3所述的方法，其特征在于，所述第一权重包括至少两个子权重，将所述至少两个子权重分别插入在所述第二权重中，所述至少两个子权重的插入位置不重合。
  5. 根据权利要求3或4所述的方法，其特征在于，所述第一偏置包括至少两个子偏置，将所述至少两个子偏置分别插入在所述第二偏置中，所述至少两个子偏置的插入位置不重合。
  6. 根据权利要求1或2所述的方法，其特征在于，所述第一网络权值包括至少两个子网络权值，所述至少两个子网络权值包括权重和偏置，将所述至少两个子网络权值分别插入在所述待训练的第二网络权值中，所述至少两个子网络权值的插入位置不重合。
  7. 根据权利要求1至6中任一项所述的方法，其特征在于，所述待训练的第二网络权值包括所述第一网络权值的所有参数，包括将所述第一网络权值的M×N个参数均匀等间隔插入到所述待训练的第二网络权值的(M+a)×(N+b)矩阵中。
  8. 根据权利要求1至7中任一项所述的方法，其特征在于，所述第一调制和所述第二调制的调制方式包括正交振幅调制QAM。
  9. 根据权利要求1至8中任一项所述的方法，其特征在于，所述方法还包括：
    获取训练后的第二网络权值；
    根据所述训练后的第二网络权值检测星座符号向量。
  10. 一种星座符号检测装置，其特征在于，包括：
    获取单元，用于获取第一调制的第一网络权值，所述第一网络权值为M×N矩阵，M和N均为正整数；
    处理单元，所述处理单元用于：
    根据所述第一网络权值确定待训练的第二网络权值，所述待训练的第二网络权值包括所述第一网络权值的所有参数，所述待训练的第二网络权值为第二调制的网络权值，所述第二调制的调制阶数大于所述第一调制的调制阶数，所述待训练的第二网络权值为 (M+a)×(N+b)矩阵，其中，a和b均为正整数；
    训练所述待训练的第二网络权值中除所述第一网络权值之外的参数。
  11. 根据权利要求10所述的装置，其特征在于，所述第一调制与所述第二调制为具有相同调制方式的相邻阶调制。
  12. 根据权利要求10或11所述的装置，其特征在于，所述第一网络权值包括第一权重和第一偏置，所述待训练的第二网络权值包括第二权重和第二偏置，所述第二权重包括所述第一权重的所有参数，且所述第二偏置包括所述第一偏置的所有参数。
  13. 根据权利要求12所述的装置，其特征在于，所述第一权重包括至少两个子权重，将所述至少两个子权重分别插入在所述第二权重中，所述至少两个子权重的插入位置不重合。
  14. 根据权利要求12或13所述的装置，其特征在于，所述第一偏置包括至少两个子偏置，将所述至少两个子偏置分别插入在所述第二偏置中，所述至少两个子偏置的插入位置不重合。
  15. 根据权利要求10或11所述的装置，其特征在于，所述第一网络权值包括至少两个子网络权值，所述至少两个子网络权值包括权重和偏置，将所述至少两个子网络权值分别插入在所述待训练的第二网络权值中，所述至少两个子网络权值的插入位置不重合。
  16. 根据权利要求10至15中任一项所述的装置，其特征在于，所述待训练的第二网络权值包括所述第一网络权值的所有参数，包括将所述第一网络权值的M×N个参数均匀等间隔插入到所述待训练的第二网络权值的(M+a)×(N+b)矩阵中。
  17. 根据权利要求10至16中任一项所述的装置，其特征在于，所述第一调制和所述第二调制的调制方式包括正交振幅调制QAM。
  18. 根据权利要求10至17中任一项所述的装置，其特征在于，所述获取单元还用于获取训练后的第二网络权值；
    所述处理单元还用于根据所述训练后的第二网络权值检测星座符号向量。
  19. 一种星座符号检测装置，其特征在于，所述装置包括：
    存储器，用于存储程序；
    处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行如权利要求1至9中任一项所述的方法。
  20. 一种计算机可读存储介质，其特征在于，所述计算机可读介质存储用于设备执行的程序代码，所述程序代码包括用于执行如权利要求1至9中任一项所述的方法的指令。
  21. 一种芯片，其特征在于，所述芯片包括处理器与数据接口，所述处理器通过所述数据接口读取存储器上存储的指令，以执行如权利要求1至9中任一项所述的方法。
PCT/CN2020/135430 2020-12-10 2020-12-10 一种星座符号检测方法及装置 WO2022120757A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/135430 WO2022120757A1 (zh) 2020-12-10 2020-12-10 一种星座符号检测方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/135430 WO2022120757A1 (zh) 2020-12-10 2020-12-10 一种星座符号检测方法及装置

Publications (1)

Publication Number Publication Date
WO2022120757A1 true WO2022120757A1 (zh) 2022-06-16

Family

ID=81973008

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/135430 WO2022120757A1 (zh) 2020-12-10 2020-12-10 一种星座符号检测方法及装置

Country Status (1)

Country Link
WO (1) WO2022120757A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101453444A (zh) * 2007-11-30 2009-06-10 富士通株式会社 星座图简化装置和方法、接收机
CN101965720A (zh) * 2007-11-06 2011-02-02 诺基亚西门子通信公司 移动通信系统设备、系统和方法中的数据传输
CN107342962A (zh) * 2017-07-03 2017-11-10 北京邮电大学 基于卷积神经网络的深度学习智能星座图分析方法
CN107623565A (zh) * 2016-07-15 2018-01-23 华为技术有限公司 基于多信道传输信号的方法和装置
WO2020128690A1 (en) * 2018-12-19 2020-06-25 International Business Machines Corporation Demodulating modulated signals with artificial neural networks
CN111404856A (zh) * 2020-03-09 2020-07-10 西安电子科技大学 基于深度学习网络的高阶调制信号解调方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101965720A (zh) * 2007-11-06 2011-02-02 诺基亚西门子通信公司 移动通信系统设备、系统和方法中的数据传输
CN101453444A (zh) * 2007-11-30 2009-06-10 富士通株式会社 星座图简化装置和方法、接收机
CN107623565A (zh) * 2016-07-15 2018-01-23 华为技术有限公司 基于多信道传输信号的方法和装置
CN107342962A (zh) * 2017-07-03 2017-11-10 北京邮电大学 基于卷积神经网络的深度学习智能星座图分析方法
WO2020128690A1 (en) * 2018-12-19 2020-06-25 International Business Machines Corporation Demodulating modulated signals with artificial neural networks
CN111404856A (zh) * 2020-03-09 2020-07-10 西安电子科技大学 基于深度学习网络的高阶调制信号解调方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
GUO YUNXIN ,ZHANG YUE ,MA HONG: "Research on Digital Signal Modulation Recognition Method Based on DBN", MODERN ELECTRONICS TECHNIQUE, vol. 43, no. 5, 1 March 2020 (2020-03-01), pages 22 - 24+29, XP055940749, ISSN: 1004-373x, DOI: 10.16652/j.issn.1004⁃373x.2020.05.006 *
LIU MINGQIAN, ZHENG SHIFEI , LI BINGBING: "Modulation Recognition of MPSK Signals Based on Deep Learning", JOURNAL OF NATIONAL UNIVERSITY OF DEFENSE TECHNOLOGY, vol. 41, no. 5, 11 October 2019 (2019-10-11), pages 153 - 158, XP055940744, ISSN: 1001-2486, DOI: 10.11887/j.cn.201905022 *

Similar Documents

Publication Publication Date Title
US10263736B2 (en) Method and device for detecting data transmission and received signals
KR101154745B1 (ko) Ｍｉｍｏ 등화를 위한 고속 제곱근 알고리듬
US11005697B2 (en) Orthogonal frequency-division multiplexing equalization using deep neural network
US10756790B2 (en) Distributed radio system
US20160254937A1 (en) Low complexity scma/lds detection systems and methods
Sujitha et al. Genetic grey wolf optimizer based channel estimation in wireless communication system
JP6272574B2 (ja) 通信チャネルを介して受信されたデータブロックを復号するための方法および受信機
CN115943395A (zh) 用于无导频ofdm通信的无线电接收器、发射器和系统
CN114826832B (zh) 信道估计方法、神经网络的训练方法及装置、设备
US11722240B2 (en) Rate adaptation
CN109391315B (zh) 一种数据模型双驱动的mimo接收机
KR20140096559A (ko) 무선 통신 시스템에서 주파수-직각 진폭 변조된 신호의 복호 매트릭 생성 방법 및 장치
WO2022120757A1 (zh) 一种星座符号检测方法及装置
US8385489B2 (en) Methods and apparatus for wireless channel estimation using interpolation elimination in the Eigen domain
WO2018086405A1 (zh) 一种调制方式检测方法和装置
CN107395547B (zh) 一种iq支路不平衡条件下的补偿方法及系统
CN115037340B (zh) 信号检测方法、装置、电子设备及存储介质
CN115412416A (zh) 一种面向高速移动场景的低复杂度otfs信号检测方法
Xue et al. Progressive equalizer matrix calculation using QR decomposition in MIMO-OFDM systems
Turhan et al. Deep learning aided generalized frequency division multiplexing
KR102001107B1 (ko) 안테나 단마다 독립적인 발진기를 갖는 다중입출력 시스템 및 그 위상잡음 제거 방법
Rao et al. A low complex cuckoo search optimizer based OFDM carrier frequency offset estimation for 5G wireless technology
EP3836499B1 (en) Device and method for training a model
CN112751792B (zh) 一种信道估计方法及装置
Gizzini et al. Deep Neural Network Augmented Wireless Channel Estimation for Preamble-based OFDM PHY on Zynq System on Chip

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20964695

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20964695

Country of ref document: EP

Kind code of ref document: A1