Method for reducing burst error rate based on meta self-supervised learning in wireless communication

Abstract

The present application relates to the field of wireless communication and mobile communication, and discloses a method for reducing burst bit error rate based on meta self-supervised learning in wireless communication, comprising the following steps: first step: determining the path loss function between a small cell base station (SBS) and its service users and the received signal expression; second step: determining the content popularity distribution and the cache model; third step: proposing a burst bit error rate mechanism and modeling the optimization problem; fourth step: introducing burst bit error rate (BBER) as a measure of channel quality, and solving the joint decision cache placement problem by using a meta-learning-based self-supervised learning algorithm, which can realize online iteration through few-shot learning, make real-time decisions according to the channel state, and reduce transmission errors caused by burst errors in signal transmission.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication and mobile communication, specifically to a method for reducing burst error rate in wireless communication based on a meta-self-supervised learning algorithm. Background Technology

[0002] With the development and evolution of fifth-generation mobile communication networks, users' demand for mobile data traffic will grow rapidly. In order to alleviate the pressure on the backhaul link from the cloud to the communication base station during peak business periods and improve the user experience quality of content delivery, mobile edge caching has been proposed as a key technology. By pre-transmitting user-requested content to the base station for storage during off-peak business periods, it aims to distribute requested content to mobile users in a timely manner, reduce response latency, and improve content delivery quality.

[0003] In narrowband communication, flat fading channels can be blocked by obstacles, leading to deep fading. Burst errors typically occur in deep fading situations where the signal is affected by burst fading. Combined with additive white Gaussian noise (AWGN), this results in consecutive bit errors at the receiver. While such errors can be reduced through channel coding, this can increase the coding / decoding complexity of the system. Therefore, we propose a method for reducing burst error rate in wireless communication based on meta-self-supervised learning. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a method for reducing burst bit error rate in wireless communication based on a meta-self-supervised learning algorithm, thus solving the aforementioned problems.

[0006] (II) Technical Solution

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for reducing burst bit error rate in wireless communication based on a meta-self-supervised learning algorithm, comprising the following steps:

[0008] Step 1: Determine the path loss function and the received signal expression between each SBS and its served users;

[0009] Step 2: Determine the content popularity distribution and caching model;

[0010] Step 3: Propose a burst error rate mechanism and optimize problem modeling;

[0011] Step 4: Solve the joint cache placement problem online by using a meta-learning-based self-supervised learning algorithm.

[0012] The specific details of the first step are as follows:

[0013] The path loss function between each SBS and its served user is expressed as P r = gd κ P t , where g is the average channel gain, κ is the path loss exponent, and d is the distance between the SBS and the user;

[0014] The received signal consists of K paths, and the delay of each multipath is τ k , and τ k is much smaller than the reciprocal of the signal bandwidth, that is: τ k << B -1 , and the received signal is:

[0015]

[0016] where f c is the carrier frequency, a k (t) and φ k (t) are the amplitude and phase of the k-th multipath component.

[0017] The specific content of the second step is as follows:

[0018] There are F contents requested by each user. Among them, the intersection of the content spaces of different users is empty, and the content satisfies the Zipf distribution model. γ is the Zipf factor, that is, the popularity of the m-th content (1 ≤ f ≤ F) can be expressed as:

[0019]

[0020] The capacity of the base station is C, then C << NF, N is the wireless system of N users, and F is the number of contents requested by each user.

[0021] The specific content of the third step is as follows:

[0022] S1: AFD is defined as the average time when the received signal envelope is lower than the given signal recognition threshold Z. Z is defined as the signal power of the lowest recognizable signal-to-noise ratio. AFD is expressed as where ρ 2 = Z 2 / P r , P r is the average received power, and f d is the Doppler shift;

[0023] S2: Let L z be the expectation of the number of crossings per second at the threshold Z in the Rayleigh channel, then

[0024] S3: Introduce the definition of BBER and use BBER to evaluate the communication performance under burst fading.

[0025] The probability of an event with i consecutive bits of error is p. i The probability of no bit error is p0, and the expected length of burst errors is... The average BBER can then be calculated as follows:

[0026]

[0027] Define the BBER between the user and the cloud as b0. The requested content is cached in the Service Provider Shortcut (SBS) providing the service, and the SBS can immediately transmit it to the requesting user. In this case, the BBER of the transmission link is much smaller than b0, and the total BBER in time slot t is:

[0028]

[0029] This indicates whether the f-th content in the u-th user's file space was requested at time t. A value of 1 indicates it was requested, and 0 indicates it wasn't. t This represents the caching strategy at time t. and Let f represent whether the m-th base station has cached content f at time t, and let f represent the average BBER of the channel link at that time.

[0030] The optimization problem in the third step:

[0031]

[0032]

[0033] Preferably, the specific algorithm flow in the fourth step is as follows:

[0034] S1: Input: p(τ): Joint distribution of tasks, α, β: Learning rates of the inner and outer loops;

[0035] S2: Randomly initialize θ;

[0036] S3: while not done do;

[0037] S4: Sample training samples T i ~p data ;

[0038] S5: From T i Extracting support set S i ;

[0039] S6: Calculate the self-supervised objective matrix and predict the network output.

[0040] S7: Update the network parameters of the inner loop using the binary cross-entropy loss function.

[0041] S8: From T i Extracting query set Q i ;

[0042] S9: Calculate the query set in the network parameters using the binary cross-entropy loss function. The cumulative gradient descent direction is determined and updated on the initial network;

[0043] S10: end while.

[0044] (III) Beneficial Effects

[0045] Compared with existing technologies, this invention provides a method for reducing burst bit error rate in wireless communication based on a meta-self-supervised learning algorithm, which has the following beneficial effects:

[0046] 1. A method for reducing burst bit error rate based on meta-self-supervised learning algorithm is proposed. Burst bit error rate (BBER) is introduced as a measure of channel quality to reduce transmission errors caused by burst errors in signal transmission.

[0047] 2. A method for reducing burst error rate based on meta-self-supervised learning algorithm. The proposed algorithm can quickly adapt to a new environment through learning from a small number of samples, and therefore can better adapt to dynamically changing channels.

[0048] 3. A novel meta-learning-based self-supervised learning algorithm is proposed to reduce burst error rate and address the cache placement problem. Simulation results show that this method has a fast convergence speed and high accuracy when the training samples are few. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the system model;

[0050] Figure 2 A diagram illustrating the algorithm parameter settings;

[0051] Figure 3 This diagram illustrates the performance comparison of the algorithm under different iteration numbers.

[0052] Figure 4 This is a diagram showing the performance comparison of various algorithms under different amounts of content.

[0053] Figure 5 This is a diagram showing the performance comparison of various algorithms under different signal-to-noise ratios. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] System Description

[0056] Consider a wireless system with M small cell base stations (SBSs) and N users. The M SBSs are located in fixed positions, while the users are located in arbitrary positions. The N users move at random speeds and in random directions within the coverage area of ​​the SBSs. Figure 1 As shown. The user's velocity V and direction θ vary randomly with time, but remain constant within each coherent time slot. The transmit power of each SBS is p. t .

[0057] Using a simple free space model, the path loss function between each SBS and its served users is expressed as P. r =gd κ P t Where g is the average channel gain, κ is the path loss exponent, and d is the distance between the SBS and the user. The SBS provides service to each user within a dedicated resource block (defined by orthogonal carrier frequencies or time slots) to avoid transmission interference. Assume the receiver signal consists of K paths, and the delay of each multipath path is τ. k , and τ k Much smaller than the reciprocal of the signal bandwidth, i.e.: τ k < -1 Therefore, the received signal can be written as:

[0058]

[0059] Among them, f c It is the carrier frequency, a k (t) and φ k (t) represents the amplitude and phase of the k-th multipath component.

[0060] Content distribution and caching model.

[0061] Suppose that each user requests a total of F pieces of content, where the intersection of the content spaces of different users is empty. These contents follow the Zipf distribution model, where γ is the Zipf factor, i.e., the popularity of the m-th content (1≤f≤F) can be expressed as:

[0062]

[0063] ​In the actual application environment, the cache capacity of the base station is much smaller than the total number of contents. Suppose there are M base stations, and the capacity of each base station is C, then C << NF. Therefore, in order to achieve the best network performance, the proposed method needs to allocate the limited cache space to the more popular requested contents and optimize the system performance through an effective caching mechanism. Imagine that if each base station stores the most popular contents simultaneously, when users request these contents, all base stations can provide services, so the system diversity is the highest. However, using the same capacity space to store the same content, this caching allocation method will lead to a decrease in the total number of cached contents. When users request those uncached contents, the SBS cannot directly provide services and needs to obtain the requested content from the central server through the wireless backhaul before forwarding the content to the requesting customers. This transmission link has a higher bit error rate and consumes more power, thus reducing the overall transmission performance. Therefore, simply improving the diversity by caching popular contents is not necessarily the optimal solution. On the contrary, if each base station stores different contents according to a certain strategy, the utilization rate of the cached contents will be higher in this case.

[0064] In the caching model, we need to consider two constraints: the base station capacity and the cooperation relationship. It is necessary to ensure that each base station can only store C contents, and at the same time, the cached contents among different base stations are different.

[0065] Modeling of the burst bit error rate mechanism and optimization problem.

[0066] An optimization problem is formulated to deploy a joint edge caching placement scheme with the goal of minimizing the burst error. Based on the burst fading of the Rayleigh channel under narrowband communication, the parameter AFD (average burst fading duration) belonging to the physical layer is introduced to evaluate the channel quality between the user and the SBS. AFD is defined as the average time when the signal envelope is lower than the given signal recognizable threshold Z, and Z is defined as the signal power of the lowest recognizable signal-to-noise ratio (that is, the signal-to-noise ratio lower than Z does not meet the signal decoding, and the signal-to-noise ratio during burst errors is lower than Z). Thus, AFD can be expressed as where ρ 2 = Z 2 / P r , P r is the average received power, and f d is the Doppler shift. In addition, let L z be the expectation of the number of crossings per second at the threshold Z under the Rayleigh channel, then For the convenience of analysis and without loss of generality, we introduce the definition of the burst bit error rate (BBER) and use the BBER to evaluate the communication performance under burst fading. Specifically, assume that the probability of consecutive errors of i bits in a burst event is p i , and the probability of no bit error is p0, and the expectation of the length of the burst bit error is The average BBER can then be calculated as follows:

[0067]

[0068] Meanwhile, we define the BBER between the user and the cloud as b0. If the requested content is cached in the SBS providing the service, the SBS can be immediately transmitted to the requesting user. In this case, the BBER of the transmission link is much smaller than b0, therefore, the total BBER of time slot t is:

[0069]

[0070] This indicates whether the f-th content in the u-th user's file space was requested at time t. A value of 1 indicates it was requested, and a value of 0 indicates it wasn't. A t This represents the caching strategy at time t. and Let f represent whether the m-th base station has cached content f at time t, and let f represent the average BBER of the channel link at that time.

[0071] Our goal is to minimize equation (5). The optimization problem is thus presented below:

[0072]

[0073]

[0074] Because all SBSs within the service area cooperate with each other, they transmit the collected channel information to the cloud. The cloud server determines the cooperative caching strategy, effectively avoiding cache conflicts caused by multiple base stations caching the same content. Next, we carefully designed a self-supervised learning method based on meta-learning to solve the proposed optimization problem, as detailed below:

[0075] Meta-learning, initially applied to image recognition and also known as learning how to learn, aims to rapidly adapt to new environments through training on a small number of samples. Model-agnostic meta-learning (MAML) is a promising meta-learning algorithm that trains a meta-model on a small number of prepared training tasks, enabling personalized models to perform better on new tasks. For problem (P1), we propose a self-supervised meta-learning-based algorithm to centrally determine the caching strategy for all SBSs. Our goal is to train a multilayer perceptron (MLP) to predict the caching deployment strategy for each base station by observing the BBER of different channels at a given time, under conditions where the content popularity is unknown. The details of the entire algorithm will be elaborated below.

[0076] Classical meta-learning methods employ two main processes to update network parameters: an inner loop and an outer loop. In the i-th iteration, a small amount of training data is split into two training sets: the support set S. i and query set Q i The former is used in the inner loop to train and update the initial network parameters into new temporary network parameters using the support set. The latter is used in the outer loop to train the network parameters obtained in the inner loop using the query set, calculate the cumulative gradient of the temporary network, and then update the initial network with the new gradient.

[0077] As is well known, self-supervised learning (SSL) is a type of unsupervised learning used to learn meaningful information as labels from unlabeled data. These labels can be used as targets for fine-tuning training or as fixed feature extractors for models such as logistic regression. Let J(θ) be the cost function, and θ be the parameters of the MLP network. The cost function is defined as:

[0078]

[0079] Where x∈R MN×1 The training samples are the inverse of the BBER of the (M×N) user-SBS links associated with AFD, p is the empirical distribution of the training data, and L is the loss function, which is binary cross-entropy (BCE). This is the prediction result of the network model; it's the network output given the input x and network parameters θ. Π(x) is the self-supervised learning label, constructed by multiplying the inverse of the BBERs observed in each time slot by the request probability of the content at a given time. Through this process, the network parameters are updated as follows:

[0080]

[0081] α is the learning rate of the network parameters. Once all tasks in the support set have been trained, the update of the meta-learning inner loop is complete. The updated temporary network parameters are obtained upon completion of the inner loop. Next, the outer loop is needed to update the cumulative gradient descent direction obtained from training on the temporary network to the initial network θ.

[0082] The formula for calculating the loss function is as follows:

[0083] L(φ(x;θ),Π(x))=Π(x)logφ(x;θ)+(1-Π(x))log(1-φ(x;θ)) (9)

[0084] The gradient of the network parameters is calculated using the loss function described above, but it is not used in... Instead of performing gradient descent updates based on the initial network θ, the network parameters in the outer loop are updated within the initial network θ. The update process for the network parameters in the outer loop is as follows:

[0085]

[0086] The specific algorithm flow is as follows:

[0087] Input: p(τ): Joint distribution of tasks;

[0088] α, β: Learning rates for the inner and outer loops;

[0089] 1: Randomly initialize θ;

[0090] 2: while not done do;

[0091] 3: Sampling training samples T i ~p data ;

[0092] 4: From T i Extracting support set S i ;

[0093] 5: Calculate Π(S) (i) And predict the network output.

[0094] 6: Update the network parameters of the inner loop using formula (8)

[0095] 7: From T i Extracting query set Q i ;

[0096] 8: Use formula (10) to find the gradient descent direction of the query set samples on the network and update it on the initial network;

[0097] 9:end while.

[0098] The performance of the proposed method and existing comparable methods is compared through experimental simulations. First, the system model parameters are set as shown in the table below. Consider a rectangular area with sides of 100 meters. Four SBSs are located at the vertices of the rectangle, and eight users are randomly distributed within the rectangular area, moving at random speeds and directions. Each user requests a different set of content. Pre-training and DQN, two traditional deep learning and deep reinforcement learning methods, are used as comparative experiments.

[0099] See Figure 3The accuracy of the cache placement schemes achieved by the proposed method and the comparison methods under different iteration counts is compared. Gaussian smoothed curves of each method are used to more clearly reflect their performance. It can be observed that for a random environment where the user's position, movement speed, and direction change uniformly and randomly, the proposed algorithm has higher learning accuracy and faster convergence speed. Therefore, the proposed caching strategy is superior to the compared algorithms. It is worth noting that even with only a small number of samples for training, this method can achieve better performance than DQN (e.g., ...). Figure 2 As shown, the incremental gradient of our proposed method is steeper than other methods. This characteristic facilitates online execution of the algorithm, allowing it to be trained with only a small number of samples and quickly adapt to different environments. From the definition of learning accuracy, this result demonstrates that the cache using this method has a better BBER gain compared to other methods, thereby reducing burst errors during transmission.

[0100] See Figure 4 The proposed method and the comparison methods were compared under the condition that the amount of content per user was fixed at 200 and 1000 respectively. The average BBER performance of the caching schemes calculated by the algorithm was compared under different iteration numbers. As the content space increases, the request probability of each piece of content is diluted, so overall the performance at F=200 is better than that at F=1000. It can also be seen that the convergence speed and performance of the proposed algorithm are indeed better than other algorithms.

[0101] See Figure 5 The average BBER of the proposed method and the comparative methods were calculated under different signal-to-noise ratios (SNRs) for various cache placement schemes, with a fixed number of 500 iterations. The content size F = 200 was determined, and the performance obtained after 200 training iterations under different SNRs was investigated. It can be seen that the average BBER of all methods decreases with increasing SNR. However, our proposed method consistently outperforms the two comparative methods, especially in the high SNR region.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for reducing burst bit error rate in wireless communication based on meta-self-supervised learning. Its features include the following steps: Step 1: Determine the path loss function and received signal expression between each SBS and its served users; Step 2: Determine the content popularity distribution and caching model; Step 3: Propose the burst bit error rate mechanism and optimize the problem modeling; Step 4: Online solve the joint caching placement problem by using the meta-learning based self-supervised learning algorithm; The specific content of the third step is as follows: S1: AFD is defined as the average time the signal envelope is below a given signal identifiable threshold Z, where Z is defined as the signal power with the lowest identifiable signal-to-noise ratio. AFD is expressed as... Where ρ 2 =Z 2 / P r ,P r f is the average received power. d For Doppler frequency shift; S2: Let L z Let Z be the expected number of crossings per second at the threshold Z under a Rayleigh channel. S3: Introduce the definition of BBER and use BBER to evaluate the communication performance under burst fading.

2. The method for reducing burst bit error rate in wireless communication based on meta-self-supervised learning according to claim 1, characterized in that: The specific content of the first step is as follows: The path loss function between each SBS and its served users is represented by P. r =gd κ P t , where g is the average channel gain, κ is the path loss exponent, and d is the distance between the SBS and the user; The receiver signal consists of K paths, and the delay of each multipath path is τ. k , and τ k Much smaller than the reciprocal of the signal bandwidth, i.e.: τ k < -1 The received signal is:​ Among them, f c It is the carrier frequency, a k (t) and φ k (t) represents the amplitude and phase of the k-th multipath component.

3. The method for reducing burst bit error rate in wireless communication based on meta-self-supervised learning according to claim 1, characterized in that: The specific content of the second step is as follows: There are F contents requested by each user. Among them, the intersections of the content spaces of different users are all empty. The content popularity satisfies the Zipf distribution model. γ is the Zipf factor. That is, the popularity of the m-th content (1 ≤ f ≤ F) can be expressed as: The capacity of the base station is C, where C << NF, N is the wireless system of N users, and F is the number of contents requested by each user.

4. The method for reducing burst bit error rate in wireless communication based on meta-self-supervised learning according to claim 1, characterized in that: The probability of an event with i consecutive bits of error is p. i The probability of no bit error is p0, and the expected length of burst errors is... The average BBER can then be calculated as follows: Define the BBER between the user and the cloud as b0. The requested content is cached in the serving SBS, and the SBS can immediately transmit it to the requesting user. In this case, the BBER of the transmission link is much smaller than b0. The total BBER at time slot t is: This indicates whether the f-th content in the u-th user's content space was requested at time t. A value of 1 indicates it was requested, and 0 indicates it wasn't. t This represents the caching strategy at time t. and Let f represent whether the m-th base station has cached content f at time t, and let f represent the average BBER of the channel link at that time.

5. The method for reducing burst bit error rate in wireless communication based on meta-self-supervised learning according to claim 1, characterized in that: The optimization problem in the third step:

6. The method for reducing burst bit error rate in wireless communication based on meta-self-supervised learning according to claim 1, characterized in that: The specific algorithm process in the fourth step is: S1: Input: p(τ): the joint distribution of tasks, α, β: the learning rates of the inner and outer loops; S2: Randomly initialize θ; S3: while not done do; S4: Sample training samples T i ~p data ; S5: From T i Extracting support set S i ; S6: Calculate the self-supervised objective matrix and predict the network output. S7: Update the network parameters of the inner loop using the binary cross-entropy loss function. S8: From T i Extracting query set Q i ; S9: Calculate the query set in the network parameters using the binary cross-entropy loss function. Accumulate gradients on the network and update them on the initial network; S10: end while.

Images