Cellular-free large-scale MIMO-oriented quantitative federated learning method and system

Abstract

The invention discloses a quantized federated learning method and a quantized federated learning system for non-cellular large-scale MIMO (Multiple Input Multiple Output). According to the method, the distributed characteristic and excellent network performance of the non-cellular large-scale MIMO network are utilized, and a distributed machine learning framework, namely federal learning, is deployed on the non-cellular large-scale MIMO network. In order to relieve the overhead problem caused by frequent exchange of a large number of model parameters in federated learning, the model parameter updating value is quantized before a user uploads the model parameter updating value, so that the transmission bit number is reduced, and the communication overhead of the system is reduced. Considering that in actual communication, users are frequently failed to access due to interference or resource limitation, a random user selection mechanism is further introduced to adapt to different user participation rates, the model training performance is optimized, and the communication overhead is reduced. The method has been proved to have convergence in consideration of channel estimation errors, quantization errors and partial user participation. In addition, the transmission bit number can be reduced while the learning performance is ensured, and the communication overhead is reduced.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a quantitative federated learning method and system for non-cellular large-scale MIMO. Background Art

[0002] In recent years, cell-free massive MIMO technology has been considered as an effective solution to improve the performance of 6G networks. Compared with traditional cellular networks, it can not only improve network capacity and reduce interference, but also provide more uniform coverage. At present, user-centric cell-free massive MIMO network architecture has become a practical way to deploy such communication systems. By building service clusters for users, this architecture makes it possible to estimate fronthaul transmission and channel state information in the case of a surge in mobile traffic. At the same time, as an emerging distributed machine learning framework, federated learning only requires exchanging model parameters, ensuring local processing of raw data. Compared with traditional distributed learning, federated learning enhances data privacy and reduces communication overhead. However, the sharp increase in mobile devices and the amount of collected data has significantly increased the communication cost of federated learning. Therefore, developing an efficient wireless federated learning system is crucial to cope with the challenge of huge communication overhead caused by frequent data exchange of a large number of mobile terminals. Recently, inspired by the need for federated learning to improve communication efficiency and the dual advantages of cell-free massive MIMO in terms of distributed characteristics and excellent network performance, some studies have begun to focus on deploying federated learning on cell-free massive MIMO networks. Nevertheless, these studies mainly focus on indicators such as training time, energy consumption and federated learning performance, ignoring the huge communication overhead caused by a large number of devices and data transmission. Therefore, it is particularly necessary to conduct research on efficient communication federated learning methods for non-cellular massive MIMO networks. Summary of the invention

[0003] Purpose of the invention: In view of the shortcomings of the prior art, the purpose of the present invention is to provide a quantitative federated learning method and system for non-cellular massive MIMO, which quantizes the parameter update value of the local model, can significantly reduce the number of transmission bits, and at the same time can ensure the performance of federated learning. At the same time, further research on the model training performance and communication overhead under different user participation rates can also achieve the goal of reducing overhead while ensuring learning performance.

[0004] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a quantitative federated learning method for non-cellular massive MIMO, comprising the following steps:

[0006] Deploy a quantized federated learning framework on a user - centered cell - free massive MIMO network; in the uplink transmission phase, the locally updated model values at the user terminals are quantized and then encoded into OFDM symbols to reduce the number of transmitted bits and communication overhead. The central processing unit receives the estimated signals at the access points and obtains the aggregated updated model values.

[0007] Select wireless network parameters and federated learning training parameters, including the scaling factor related to transmission power and the quantization order, to achieve a trade - off between the global model convergence speed and the convergence error.

[0008] Run the quantized federated learning under the above - set wireless network and federated learning training parameters to train the global model parameters and obtain the optimal model parameters of the training network.

[0009] Furthermore, in each round of global communication and training, the central processing unit randomly selects a fixed number of user terminals to participate in the training; the parameters selected based on the convergence analysis include the scaling factor related to transmission power, the quantization order, and the number of user terminals participating in the training.

[0010] Furthermore, the user - centered cell - free massive MIMO network includes multiple user terminals equipped with a single antenna and multiple access points equipped with multiple antennas; the access points are connected to the central processing unit through a fronthaul link and act as relays to support wireless communication between the central processing unit and the user terminals; by adopting a dynamic cooperation cluster scheme, an association matrix is defined so that each user is served by only one wireless access point cluster in the network.

[0011] Furthermore, in the uplink transmission phase, the n - th OFDM symbol on the s - th sub - carrier actually received by the wireless access point l is

[0012]

[0013] where is the n - th symbol on the s - th sub - carrier at the user terminal k, α t is the scaling factor related to transmission power, h k,l is the channel gain vector, is the additive noise; use a distributed maximum ratio combining vector to estimate the received signal, where is the estimated value of the channel gain vector h k,l (t), D k,l is the association matrix between the access point l and the user terminal k, and then obtain the partial information of the n - th OFDM symbol of the user terminal k at the access point l on the s - th sub - carrier The data is transmitted to the central processing unit via the fronthaul link, where it is executed

[0014]

[0015] final, The estimate is

[0016]

[0017] where tr(·) represents the trace of the matrix, D k =diag(D k,1 ,…,D k,L ) and R k =diag(R k,1 ,…,R k,L ), diag(·) represents a diagonal matrix.

[0018] Furthermore, in the uplink transmission phase, using S subcarriers and N transmissions, the nth symbol on the sth subcarrier at the user terminal k after quantization is Among them, Q b (·) represents a quantization operation, Represents the local model update value at user terminal k The qth element of and represents the upward rounding function, d is the model parameter; the central processing unit is based on the estimation get And take the real part and the imaginary part respectively to get the corresponding elements in the corresponding aggregated model update value.

[0019] Furthermore, when the local model update value at the user terminal is quantized, a parameterized quantization method is considered, in which the interval [0,1] is divided into b quantization intervals according to the quantization order b, and the absolute value of the vector element to be quantized is normalized by the vector binary norm and then uniformly quantized according to the quantization interval to which it belongs. The quantization value is the endpoint of the interval, and the final quantized vector element is the product of its sign value, the vector binary norm and its uniform quantization value.

[0020] Furthermore, according to the deviation term in the convergence analysis Select the quantization order b and the number of user terminals participating in the training E; when the model parameter scale d of the training network increases, according to the deviation term in the convergence analysis Choose the scaling factor α t , where η t is the learning rate, τ is the number of local training iterations, H is a constant that limits the expected value of the stochastic gradient 2 norm, ρ is a parameter related to the channel condition, is the noise variance.

[0021] Furthermore, the federated learning process includes a series of global communication rounds, in which the participating user terminals and the central processing unit jointly train a global model parameter θ by periodically exchanging model parameters; in the t-th round of the federated learning process, the central processing unit first broadcasts the parameter θ(t) to the participating user terminals through the downlink of the cell-free massive MIMO network, and the user terminal k will receive the parameter θ k (t) = θ(t), and through several iterative trainings of the local model parameters by using the local dataset, and obtain the updated value of the local model parameters; all participating user terminals upload the local model update values through the uplink of the cell-free massive MIMO network, and the central processing unit receives and executes the aggregation of the model update values, thereby updating the global model parameters.

[0022] In a second aspect, the present invention provides a quantization federated learning system for cell-free massive MIMO, including:

[0023] A quantization federated learning deployment module, configured to deploy a quantization federated learning framework on a user-centric cell-free massive MIMO network; wherein in the uplink transmission phase, the local model update values at the user terminals are quantized and encoded into OFDM symbols to reduce the number of transmitted bits and reduce the communication overhead, and the central processing unit receives the estimated signals at the access points to obtain the aggregated model update values;

[0024] A parameter selection and training module, configured to select wireless network parameters and federated learning training parameters, including a scaling factor and a quantization order related to the transmission power, to achieve a trade-off between the global model convergence speed and the convergence error; and run the quantization federated learning under the above-mentioned wireless network and federated learning training parameter settings to train the global model parameters and obtain the optimal model parameters of the training network.

[0025] In a third aspect, the present invention provides a computer program product, including computer programs / instructions, which, when executed by a processor, implement the processing steps of the user terminal and / or the central processing unit in a quantization federated learning method for cell-free massive MIMO as described above.

[0026] Beneficial effects: Before the user terminal uploads the model parameter update value on the network uplink, the present invention performs a quantization operation on it to reduce the required number of transmission bits, thereby reducing the communication overhead of the system. And considering that in actual communication, the network often faces the situation that some users cannot access or are disconnected, a random user selection method applicable to general scenarios is further introduced to consider the trade-off between the model training performance and the communication overhead under different user participation rates, making it more in line with the actual application scenario. The present invention has proven its convergence under different channel estimation errors, quantization precisions, and random user selection parameters at the theoretical level. In addition, the present invention can significantly reduce the number of transmission bits on the premise of ensuring the performance of federated learning, thereby reducing the communication overhead. Description of the Drawings

[0027] Figure 1 It is a schematic diagram of the overall method flow of the embodiment of the present invention.

[0028] Figure 2 It is a schematic diagram of the quantization federated learning system for cell-free massive MIMO according to the embodiment of the present invention. Detailed Embodiments

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] As Figure 1As shown in the figure, an embodiment of the present invention discloses a quantization federated learning method for a cell-free massive MIMO network. The method first deploys a quantization federated learning framework on a user-centric cell-free massive MIMO network. In the uplink transmission phase of this framework, the local model update values at the user terminals are quantized and encoded into OFDM symbols to reduce the number of transmitted bits and communication overhead. The central processing unit receives the estimated signals at the access points and obtains the aggregated model update values. Then, based on the convergence analysis, wireless network parameters and federated learning training parameters are selected, including scaling factors related to transmission power, quantization order, etc., to achieve a trade-off between the global model convergence speed and the convergence error. Finally, the quantization federated learning process under the above wireless network and federated learning training parameter configurations is run to train the global model parameters and obtain the optimal model parameters. In addition, further considering that in actual communication, users often fail to access due to interference or resource limitations, more preferably, a random user selection mechanism is introduced to adapt to different user participation rates, optimize the model training performance, and reduce communication overhead. That is, in each round of global communication and training, the central processing unit randomly selects a fixed number of user terminals to participate in the training; the parameters selected based on the convergence analysis include the scaling factor related to transmission power, quantization order, and the number of user terminals participating in the training.

[0031] As Figure 2 shown, it is a schematic diagram of quantization federated learning for a cell-free massive MIMO network. Based on the constructed quantization federated learning system for a cell-free massive MIMO network, the global communication process of each round of federated learning is as follows: First, the central processing unit broadcasts the global model parameters to the user terminals participating in this round of training; Second, each user terminal uses local data for model training and quantizes the updated values of the trained model parameters; Then, the quantized updated values of the model parameters are uploaded to the central processing unit for aggregation; Finally, the central processing unit updates the global model parameters according to the aggregation result.

[0032] The method of this embodiment will be described in more detail below in combination with a specific scenario.

[0033] First, the user-centric cell-free massive MIMO network will be described. Specifically, in this embodiment, a user-centric cell-free massive MIMO network based on OFDM is considered, and this wireless network operates in the TDD mode. The network consists of K user terminals equipped with a single antenna and L access points equipped with M antennas. The access points establish a connection with the central processing unit through a fronthaul link and act as relays to support wireless communication between the central processing unit and the user terminals. It is assumed that the wireless channel remains unchanged during a round of global communication in federated learning. Considering that there are a total of τ b complex-valued samples available in each coherence block. These samples are allocated as τp uplink pilot transmission samples, τ u uplink data transmission samples, and τ d downlink data transmission samples, and satisfying τ b = τ p + τ d + τ u In this embodiment, it is assumed that the wireless channel remains unchanged during one round of global communication in federated learning.

[0034] To solve the scalability problem caused by each user connecting to all wireless access points in a traditional cell-free massive MIMO network, by adopting a dynamic cooperation cluster scheme, an association matrix is defined, such that each user is served by a cluster of wireless access points in the network instead of all wireless access points in the network, thereby solving the resource allocation and management complexity problem caused by each user connecting to all wireless access points. The association matrix defined in this embodiment is as follows

[0035]

[0036] where represents a set of specific user terminals that are within the service coverage of wireless access point l. It is assumed that access point l can serve at most one user terminal on each pilot sequence and uses all antennas, that is In the uplink transmission stage, the nth symbol on the sth subcarrier at user terminal k can be expressed as After transmission through the link, the nth OFDM symbol on the sth subcarrier actually received by wireless access point l is

[0037]

[0038] where α t is a scaling factor related to the transmission power, and the channel gain vector between user terminal k and access point l follows Rayleigh fading, where represents the additive noise at access point l, where is the noise variance, and I M is the M-order identity matrix.

[0039] By using a distributed maximum ratio combining vector to estimate this received signal, where is the estimated value of the channel gain vector h k,l (t). Thus, partial information about the nth OFDM symbol of the kth user terminal at wireless access point l on the sth subcarrier is obtained

[0040]

[0041] Then, The data is transmitted to the central processing unit via the fronthaul link.

[0042]

[0043] Obtain comprehensive information of the nth OFDM symbol of user terminal k on subcarrier s from all served access points. For user terminal k, the nth OFDM symbol on subcarrier s can be estimated as

[0044]

[0045] where tr(·) represents the trace of the matrix, D k =diag(D k,1 ,…,D k,L ) and R k =diag(R k,1 ,…,R k,L ), diag(·) represents a diagonal matrix. In fact, is an unbiased estimate.

[0046] In this embodiment, quantitative federated learning is deployed in the above-mentioned wireless network. Specifically, consider a federated learning system, which includes K clients and a central server. In fact, the user terminal and central processing unit in the wireless network are equivalent to the client and central server in the federated learning. According to the learning task, the corresponding training network can be deployed at the user terminal and the central processing unit. Exemplarily, for the classification task of the MNIST data set, a convolutional neural network can be used, and its architecture design is as follows: the network includes two convolutional layers, the convolution kernel size of each convolutional layer is 5×5, the ReLU activation function is used, and the same padding is used. The first convolutional layer contains 32 filters, and the second convolutional layer contains 64 filters. Each convolutional layer is connected to a 2×2 maximum pooling layer for downsampling to reduce the spatial dimension of the feature map. After the convolutional layer, the network is connected to a fully connected layer containing 128 units, and the ReLU activation function is also used. Finally, the network output layer is a softmax layer containing 10 units, which is used to map features to 10 categories of the MNIST data set. The network model parameters of the central processing unit can be expressed as θ.

[0047] In the uplink transmission phase, we first update the local model at user terminal k Encoded into OFDM symbols, using S subcarriers and N transmissions, the OFDM symbol transmitted on the sth subcarrier for the nth time is given by

[0048]

[0049] where represents the q-th element of Δθ k (t), and represents the ceiling function.

[0050] To reduce the large number of transmitted bits and lower the communication overhead, a random quantization scheme is introduced. Consider a parameterized quantization method (with quantization order b). The principle is that according to the quantization order, the interval [0, 1] can be divided into b quantization intervals. After normalizing the absolute value of the vector element to be quantized by the vector two-norm, it is uniformly quantized according to the quantization interval it belongs to (the quantization value is the interval endpoint). Finally, the quantized vector element obtained is the product of its sign value, the vector two-norm, and its uniform quantization value. Specifically, the quantization model of user terminal k is updated to

[0051]

[0052] where sgn(·) is the sign function. For any i = 1, …, d, Φ i (Δθ k (t), b) are independent and identically distributed. Let 0 ≤ j < b and If then [j / b, (j + 1) / b] is the quantization interval of

[0053]

[0054] where p(a, b) = ab - j. In particular, if Δθ k (t) = 0, then Q b (Δθ k (t)) = 0. Therefore, the uplink transmission signal is given by

[0055]

[0056] Then, the central processing unit calculates its weighted sum based on the received estimated signal as

[0057]

[0058] and obtains an estimate of the aggregated model update value the q-th element of

[0059]

[0060]

[0061] Therefore, the global model parameters can be updated. Repeat the above process until the model parameters converge. In the above process, since all user terminals have different computing and communication capabilities, there are some users who cannot participate in the training due to problems such as ultra-high latency and congestion. In view of this, further consider the situation where only some users participate in each round of global communication and training, that is, adopt a random user selection strategy to select a fixed number of participating user terminals. Specifically, the selected user set is defined as The selection probability of the user terminal is For the convenience of derivation, it can be set that That is, in each global communication round, the central processing unit selects a proportion of user terminals to participate in the training.

[0062] In this embodiment, appropriate wireless network parameters and federated learning parameters can be selected according to the convergence analysis to achieve a trade-off between the convergence speed and the convergence error. Specifically, first consider three common assumptions about the local loss functions F 1 , …, F K for the convergence analysis of federated learning:

[0063] 1) Assume that F 1 , …, F K are μ-strongly convex. Specifically, for

[0064]

[0065] 2) Assume that F 1 , …, F K are L-smooth, that is, for all

[0066]

[0067] 3) For all the stochastic gradient satisfies where is the gradient operator, <·, ·> is the inner product operator, is the sampled data set of the i-th local training in the t-th global training of user terminal k, H is a constant used to limit the expectation of the two-norm of the stochastic gradient, and θ ★ represents the optimal model parameters for minimizing F(θ), and then define to measure the data heterogeneity. For simplicity, let F ＊ = F(θ ＊ ). If the value of ψ is larger, it indicates a higher degree of data heterogeneity among user terminals. Under the above assumptions and the learning rate satisfies Under the condition of, the convergence result is

[0068]

[0069] where A(t) = 1 - μη t (τ - η t τ - η t ) and

[0070]

[0071] Specifically, It should be noted that A(t) and B(t) jointly affect the deviation of the convergence result. The quantization process introduces additional deviation The fact that a lower quantization order leads to an increase in the optimality gap of the global model parameters is a natural and reasonable result. In addition, a large number of fixed users E will reduce the deviation of the convergence result. However, a higher quantization order and a larger E will increase the required transmitted data, resulting in an increase in communication overhead and a decrease in communication efficiency. Therefore, there is a trade-off between cost and performance when choosing b and E. In addition, when d is very large, the term will significantly increase the convergence error. Therefore, the scaling factor α should be carefully selected t .

[0072] Specifically, after completing the convergence analysis, appropriate parameters are selected, and quantized federated learning is run to train the global model parameters to obtain the optimal model parameters. The main goal of the federated learning process is to find the optimal model parameters θ that minimize the global loss function F(θ) ＊ , that is

[0073]

[0074] where F(·) is expressed as

[0075]

[0076] where P k represents the size of the local dataset collected from user terminal k, and is the total size of all local datasets. F k (·) represents the local loss function of user terminal k. The federated learning process consists of a series of rounds in which user terminals and the central processing unit jointly train a global model parameter θ by periodically exchanging model parameters. In the t-th round of the federated learning process, for user terminal k, the central processing unit first broadcasts the parameter θ(t) to the participating user terminals. If wireless channel fading and distortion are ignored during the transmission process, user terminal k will receive the parameter θ kθ(t) = θ(t). Then, the user terminal k performs several iterative trainings on the local model parameters by using the local data set and obtains the model update.

[0077]

[0078] where η t is the learning rate, is the sampled data set for the i-th local training iteration of the user terminal, τ is the number of local training iterations, and θ k,i (t) are the local model parameters after the i-th training iteration. Subsequently, the client transmits the model update, and the central processing unit performs the aggregation of the model update. Therefore, the global model parameters updated at the central processing unit are as follows:

[0079]

[0080] The above process is iterated until a predefined termination condition is reached.

[0081] Details not described in the present invention are all well-known techniques in the art.

[0082] Based on the same inventive concept, an embodiment of the present invention discloses a quantization federated learning system for a cell-free massive MIMO network, including: a quantization federated learning deployment module for deploying a quantization federated learning framework on a user-centric cell-free massive MIMO network; wherein in the uplink transmission phase, the local model update value at the user terminal is quantized and encoded into an OFDM symbol to reduce the number of transmitted bits and reduce the communication overhead, and the central processing unit receives the estimated signal at the access point to obtain the aggregated model update value; a parameter selection and training module for selecting wireless network parameters and federated learning training parameters to achieve a trade-off between the global model convergence speed and the convergence error; and running the quantization federated learning under the above-mentioned wireless network and federated learning training parameter settings to train the global model parameters and obtain the optimal model parameters of the training network.

[0083] Further, in the quantization federated learning deployment module, in each round of global communication and training, the central processing unit randomly selects a fixed number of user terminals to participate in the training; in the parameter selection and training module, the parameters selected based on the convergence analysis include a scaling factor related to the transmission power, the quantization order, and the number of user terminals participating in the training.

[0084] Based on the same inventive concept, an embodiment of the present invention discloses a computer program product, including a computer program / instructions, which when executed by a processor implement the processing steps of the user terminal and / or the central processing unit in the quantization federated learning method for a cell-free massive MIMO network described above.

[0085] The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations based on the concept of the present invention without creative efforts. Therefore, all technical solutions that can be obtained by those skilled in the art in the technical field according to the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art should fall within the protection scope determined by the claims.

Claims

1. A quantitative federated learning method for non-cellular massive MIMO, characterized in that: The following steps are involved: Deploy a quantized federated learning framework on user-centric cell-free massive MIMO networks; In the uplink transmission phase, the local model update value at the user terminal is quantized and encoded into OFDM symbols to reduce the number of transmission bits and reduce communication overhead. The central processing unit receives the estimated signal at the access point and obtains the aggregated model update value. Select wireless network parameters and federated learning training parameters, including scaling factors and quantization orders related to transmission power, to achieve a trade-off between global model convergence speed and convergence error; Run the quantized federated learning under the wireless network and federated learning training parameter settings to train the global model parameters and obtain the optimal model parameters of the training network.

2. The quantitative federated learning method for non-cellular massive MIMO according to claim 1, characterized in that: In each round of global communication and training, the central processing unit randomly selects a fixed number of user terminals to participate in the training; the parameters selected based on the convergence analysis include the scaling factor related to the transmission power, the quantization order and the number of user terminals participating in the training.

3. The quantitative federated learning method for non-cellular massive MIMO according to claim 1, characterized in that: The user-centric cellular-free massive MIMO network includes multiple user terminals equipped with a single antenna and multiple access points equipped with multiple antennas; the access points establish a connection with a central processing unit through a fronthaul link and serve as a relay to support wireless communication between the central processing unit and the user terminals; by adopting a dynamic collaborative clustering scheme, an association matrix is ​​defined so that each user is served by only one wireless access point cluster in the network.

4. The quantitative federated learning method for non-cellular massive MIMO according to claim 3, characterized in that: In the uplink transmission phase, the nth OFDM symbol on the sth subcarrier actually received by the wireless access point l is in is the nth symbol on the sth subcarrier at user terminal k, α t is a scaling factor related to the transmission power, h k,l is the channel gain vector, is additive noise; using a distributed maximum ratio combining vector The received signal is estimated, where is the channel gain vector h k,l The estimated value of (t), D k,l is the association matrix between access point l and user terminal k, and then obtains the partial information of the nth OFDM symbol of user terminal k at access point l on the sth subcarrier The data is transmitted to the central processing unit via the fronthaul link, where it is executed final, The estimate is where tr(·) represents the trace of the matrix, D k =diag(D k,1 ,…,D k,L ) and R k =diag(R k,1 ,…,R k,L ), diag(·) represents a diagonal matrix.

5. The quantitative federated learning method for non-cellular massive MIMO according to claim 1, characterized in that: In the uplink transmission phase, using S subcarriers and N transmissions, the nth symbol on the sth subcarrier at user terminal k is quantized. Among them, Q b (·) represents a quantization operation, Represents the local model update value at user terminal k The qth element of and represents the upward rounding function, d is the model parameter; the central processing unit is based on the estimation get And take the real part and the imaginary part respectively to get the corresponding elements in the corresponding aggregated model update value.

6. The quantitative federated learning method for non-cellular massive MIMO according to claim 1, characterized in that: When quantizing the local model update value at the user terminal, a parameterized quantization method is considered, in which the interval [0,1] is divided into b quantization intervals according to the quantization order b, and the absolute value of the vector element to be quantized is normalized by the vector two norm and then uniformly quantized according to the quantization interval to which it belongs. The quantization value is the endpoint of the interval, and the final quantized vector element is the product of its sign value, the vector two norm and its uniform quantization value.

7. The quantitative federated learning method for non-cellular massive MIMO according to claim 1, characterized in that: According to the deviation term in the convergence analysis Select the quantization order b and the number of user terminals participating in the training E; when the model parameter scale d of the training network increases, according to the deviation term in the convergence analysis Choose the scaling factor α t , where η t is the learning rate, τ is the number of local training iterations, H is a constant that limits the expected value of the stochastic gradient 2 norm, ρ is a parameter related to the channel condition, is the noise variance.

8. The quantitative federated learning method for non-cellular massive MIMO according to claim 1, characterized in that: The federated learning process includes a series of global communication rounds, in which the user terminals participating in the training and the central processing unit jointly train a global model parameter θ by regularly exchanging model parameters; in the tth round of federated learning, the central processing unit first broadcasts the parameter θ(t) to the participating user terminals through the downlink of the cellular-free massive MIMO network, and the user terminal k will receive the parameter θ k (t) = θ(t), the local model parameters are iteratively trained several times using the local data set, and the updated values ​​of the local model parameters are obtained; All participating user terminals upload local model update values ​​via the uplink of the cell-free massive MIMO network, and the central processing unit receives and performs aggregation of the model update values ​​to update the global model parameters.

9. A quantitative federated learning system for non-cellular massive MIMO, characterized in that: include: A quantitative federated learning deployment module for deploying a quantitative federated learning framework on user-centric cell-free massive MIMO networks; In the uplink transmission phase, the local model update value at the user terminal is quantized and encoded into OFDM symbols to reduce the number of transmission bits and reduce communication overhead. The central processing unit receives the estimated signal at the access point and obtains the aggregated model update value. Parameter selection and training module, used to select wireless network parameters and federated learning training parameters, including scaling factors and quantization orders related to transmission power, to achieve a trade-off between global model convergence speed and convergence error; And running the quantized federated learning under the above wireless network and federated learning training parameter settings to train the global model parameters to obtain the optimal model parameters of the training network.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the processing steps of the user terminal and / or the central processing unit in the quantitative federated learning method for non-cellular massive MIMO according to any one of claims 1-8 are implemented.

Images