Importance-based data collection storage method and device, equipment and medium

Abstract

The application provides a data collection and storage method and device based on importance, equipment and medium, the data collection and storage method based on importance is applied to an edge computing system including an edge server, an edge device and a central processor; the method comprises the following steps: determining a loss function corresponding to a convolutional neural network model to be trained, updating the convolutional neural network model based on a small batch random gradient descent algorithm and the loss function; determining data upload information and cache queue information according to the loss function, attribute information of the edge server and attribute information of the edge device; determining a random optimization problem based on the data upload information and the cache queue information; determining a long-term optimization strategy and a short-term optimization strategy based on the random optimization problem, and performing data collection and storage based on the long-term optimization strategy and the short-term optimization strategy. The application realizes the improvement of algorithm performance and reduces network resource loss.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication technology and artificial intelligence technology, and in particular to a data acquisition and storage method, apparatus, device and medium based on importance. Background Technology

[0002] The convergence of AI (Artificial Intelligence) and IoT (Internet of Things) has facilitated the development of fields such as smart homes, wearable devices, and industrial automation. To adapt to dynamic environments, online learning continuously learns new knowledge from scenarios using real-time data streams. However, because online learning cannot obtain a feature-complete dataset, AI model training typically requires many more updates to reach convergence. Secondly, the massive demand for training samples from edge servers in real-world scenarios also puts enormous pressure on communication channels and cache queues.

[0003] From a model learning perspective, some data selection schemes have proven that prioritizing important data input to the model can effectively improve training performance. However, most existing schemes ignore the dynamic needs of the model for different types of data at different stages, resulting in no significant improvement in learning efficiency. Meanwhile, inspired by importance selection in machine learning, research on wireless resource allocation has also begun to focus on the importance of transmitted content, prioritizing the transmission of important data with limited resources. However, importance-based scheduling schemes do not consider the heavy burden of data acquisition on devices; an imbalance between acquisition and transmission speeds can lead to buffer overflows, causing system crashes. Summary of the Invention

[0004] This invention provides a data acquisition and storage method, apparatus, device, and medium based on importance, to improve the training efficiency of AI models and solve the technical problem of system crash caused by the imbalance between acquisition and transmission speeds.

[0005] This invention provides a data acquisition and storage method based on importance, applicable to an edge computing system including edge servers, edge devices, and a central processing unit; the data acquisition and storage method based on importance includes:

[0006] Determine the loss function corresponding to the convolutional neural network model to be trained, and update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function;

[0007] Based on the loss function, the attribute information of the edge server, and the attribute information of the edge device, the data upload information and cache queue information are determined.

[0008] Based on the data upload information and the cache queue information, a random optimization problem is determined;

[0009] Based on the stochastic optimization problem, a long-term optimization strategy and a short-term optimization strategy are determined, and data is collected and stored based on the long-term optimization strategy and the short-term optimization strategy.

[0010] According to the present invention, a data acquisition and storage method based on importance is provided, wherein determining data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device includes:

[0011] Based on the updated model convergence index and queue backlog information, the mini-batch value, data computation vector and central processing unit frequency corresponding to each iteration process are determined.

[0012] The computational delay for each iteration process is determined based on the mini-batch value, sample size, workload required for sample gradient calculation, and CPU frequency.

[0013] Based on the constants corresponding to the edge server and the frequency of the central processing unit, the power consumption corresponding to each of the iterative processes is determined;

[0014] Based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge server and the edge device, and the data calculation vector, the data scheduling vector and the transmission power are determined.

[0015] The data upload rate is determined based on the channel fading coefficient between the edge server and the edge device, the total bandwidth of the edge computing system, the channel noise, the transmit power, and the data scheduling vector.

[0016] According to the importance-based data acquisition and storage method provided by the present invention, the step of determining data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device further includes:

[0017] The data acquisition vector is determined based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge server and the edge device, and the data calculation vector.

[0018] The dynamic changes in the queue of the edge device are determined based on the data upload rate, the data acquisition vector, and the queue backlog of the edge device.

[0019] The number of time slots corresponding to each of the iterative processes is determined based on the computational delay.

[0020] Based on the data calculation vector, determine the number of samples extracted from the cache queue corresponding to the edge server;

[0021] Based on the number of samples extracted from the cache queue and the number of time slots corresponding to each iteration process, the average number of samples leaving the cache queue is determined.

[0022] According to the importance-based data acquisition and storage method provided by the present invention, the step of determining data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device further includes:

[0023] The dynamic changes in the edge server's queue are determined based on the sample size, the data upload rate, the average number of samples leaving the cache queue, and the queue backlog of the edge server.

[0024] According to the present invention, a data acquisition and storage method based on importance is provided, wherein determining the stochastic optimization problem based on the data upload information and the cache queue information includes:

[0025] Based on the model convergence metric, the data upload information, and the cache queue information, the target optimization problem is determined.

[0026] The objective optimization problem is transformed into a stochastic optimization problem through Lyapunov optimization.

[0027] According to the present invention, a data acquisition and storage method based on importance is provided, wherein determining a long-term optimization strategy and a short-term optimization strategy based on the stochastic optimization problem, and performing data acquisition and storage based on the long-term optimization strategy and the short-term optimization strategy, includes:

[0028] Based on the test set loss value of the current convolutional neural network model and the queue backlog corresponding to each iteration process, the model computation vector and the edge server computation frequency are determined.

[0029] Based on the observation state of the upper-layer agent corresponding to the PDQN algorithm and the model calculation vector, the convolutional neural network model is updated to determine the model learning performance reward corresponding to the long-term optimization strategy.

[0030] Based on the channel state vector, the queue backlog of the edge devices, and the queue backlog of the edge servers, the data scheduling vector, transmit power, and data acquisition vector of each edge device are determined.

[0031] Based on the observation state of the lower-level agent corresponding to the DQN algorithm, the data scheduling vector, the transmission power, the data acquisition vector, and the average number of samples leaving the buffer queue, the queue stability performance reward corresponding to the short-term optimization strategy is determined.

[0032] The present invention also provides a data acquisition and storage device based on importance, comprising:

[0033] The model update module is used to determine the loss function corresponding to the convolutional neural network model to be trained, and update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function.

[0034] The first determining module is used to determine data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device;

[0035] The second determining module is used to determine the random optimization problem based on the data upload information and the cache queue information;

[0036] The third determining module is used to determine long-term optimization strategies and short-term optimization strategies based on the stochastic optimization problem, and to perform importance-based data collection and storage based on the long-term optimization strategies and the short-term optimization strategies.

[0037] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the importance-based data acquisition and storage method as described above.

[0038] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the importance-based data acquisition and storage method as described above.

[0039] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the importance-based data acquisition and storage method as described above.

[0040] The importance-based data acquisition and storage method, apparatus, device, and medium provided by this invention determine the loss function corresponding to the convolutional neural network model to be trained, update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function, and then determine the data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device. Based on the data upload information and cache queue information, a stochastic optimization problem is determined, and finally, based on the stochastic optimization problem, a long-term optimization strategy and a short-term optimization strategy are determined. Based on the long-term optimization strategy and the short-term optimization strategy, importance-based data acquisition and storage is performed, which improves the algorithm performance, reduces network resource consumption, and formulates an optimization problem that maximizes the learning efficiency of joint data acquisition, scheduling, and computation. The main feature of the importance-based data acquisition and storage method provided by this invention is that it can determine the acquisition, transmission, storage, and computation strategies according to the data importance and channel and system cache constraints. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0042] Figure 1 This is one of the flowcharts illustrating the importance-based data acquisition and storage method provided by the present invention;

[0043] Figure 2 This is the second flowchart of the importance-based data acquisition and storage method provided by the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of the importance-based data acquisition and storage device provided by the present invention;

[0045] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0047] The following is combined with Figures 1-2The present invention describes a data acquisition and storage method based on importance.

[0048] Please refer to Figure 1 This invention provides a data acquisition and storage method based on importance, comprising:

[0049] Step 100: Determine the loss function corresponding to the convolutional neural network model to be trained, and update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function;

[0050] Specifically, the importance-based data acquisition and storage method provided in this invention proposes an importance-based approach to data acquisition, data transmission, data storage, and data computation. This method can automatically schedule edge devices to collect important data and upload it to edge servers for model training, effectively improving the online learning efficiency of models in edge computing systems and ensuring system stability. The importance-based data acquisition and storage method provided in this invention can jointly schedule acquisition, transmission, storage, and computation based on channel status, queue status, and the importance of data to the algorithm. The scheduling decision is based on a newly designed data importance index, which can describe the relationship between data and algorithm performance. In edge computing systems, this method can significantly improve algorithm performance and reduce network resource consumption. Specifically, it includes:

[0051] Step 1: Before each model update, based on the learning performance and data queue status, extract key data that can accelerate learning from the edge server cache. Input this key data into the CPU (central processing unit) for training, and allocate computational frequency for model updates. Step 2: Based on the data queue status and channel conditions, allocate transmit power to some edge devices to schedule their data uploads, while simultaneously allocating acquisition frequency to all edge devices. Acquired data that has not yet been transmitted is temporarily stored in the edge device's cache. Repeat steps 1 and 2 until the AI ​​model training converges.

[0052] Suppose an edge computing system has one ES (Edge Server) and k ED (Edge Devices), each equipped with a buffer. The ES relies on data samples collected by the EDs to train a classifier model online. To accelerate model training, the ES selects important data to update the model. Furthermore, data collection and transmission are adaptively adjusted based on the model's training performance and buffer status. For example, to perform an image classification task, the edge computing system needs to train a CNN (Convolutional Neural Networks) model. By adjusting the parameters w, the CNN model can learn the mapping relationship between inputs and predicted values ​​from labeled samples (x, y), where x is the input sample and y is the label corresponding to the input sample. Given a loss function ψ(w; x, y), the goal of model training is to minimize... w Ψ(w), as shown in Formula 1.

[0053]

[0054]

[0055]

[0056] Where N is the total sample size, the mini-batch SGD (stochastic gradient descent) algorithm is used to solve the above loss function (Equation 1) minimization problem. In the i-th iteration, i∈I={1,2,…,I}, ES calculates the gradient of a mini-batch to update the convolutional neural network model, as shown in Equation 2. Where γ is the learning rate of the model training, m i This refers to the size of the mini-batch in the i-th iteration. Both of these parameters significantly impact model convergence. Considering the limited storage in Elasticsearch and the ability of the terminal to guarantee real-time uploads, samples are deleted after each model update.

[0057] Step 200: Determine data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device;

[0058] Specifically, after updating the convolutional neural network model, based on the attribute information of the edge server (e.g., the constants corresponding to the edge server and the channel fading coefficient between the edge server and the edge device), the attribute information of the edge device (e.g., the acquisition frequency of the edge device and the queue backlog of the edge device), and the frequency of the central processing unit, the data upload information (e.g., the data upload rate and the data upload volume) and the cache queue information (e.g., the dynamic changes of the queue and the average number of samples leaving the cache queue) are determined. The specific confirmation process of the data upload information and the cache queue information will be described below.

[0059] Step 300: Based on the data upload information and the cache queue information, determine the random optimization problem;

[0060] Specifically, after determining the data upload information and cache queue information, an MCM (Model Convergence Metric) is used to dynamically evaluate the importance of the data for model learning. Based on the MCM, the optimization problem P1, which jointly involves data acquisition, scheduling, and computation, can be formulated as follows:

[0061] Clearly, P1 is a highly complex stochastic optimization problem. This is manifested in several ways: 1. The gradient calculated using the current mini-batch not only determines the MCM of this iteration but also affects the long-term cumulative gradient. Therefore, data computation decisions are time-coupled. 2: Due to the limited buffer space of the device, data computation decisions affect data scheduling decisions, and similarly, data scheduling decisions also affect data acquisition decisions. Therefore, there is also coupling between multiple variables. 3: When the data computation strategy and channel conditions are unknown, the per-slot decision is difficult to satisfy the long-term constraint of queue stability. Based on the above analysis and Lyapunov optimization techniques, P1 can be transformed into a stochastic optimization problem P2 without long-term constraints, which will be described in detail below.

[0062] Step 400: Based on the stochastic optimization problem, determine the long-term optimization strategy and the short-term optimization strategy, and collect and store data based on the long-term optimization strategy and the short-term optimization strategy.

[0063] Specifically, although the long-term constraint problem has been solved, solving P2 still faces the following challenges: 1. The objective contains finite-time expectations, making commonly used deterministic optimization schemes unsuitable for problems without closed-form solutions. 2. An inherent time difference exists between model updates and channel changes, requiring a dual-time-scale optimization scheme, including a long-term model computation strategy {D}. I ,f i} and short-term data scheduling and acquisition strategies {α t ,pt ,λ t}

[0064] Reinforcement learning is a common method for solving the aforementioned problems without closure solutions. To overcome the curse of dimensionality, hierarchical reinforcement learning involves upper-level and lower-level agents exploring the environment at two levels of temporal abstraction. Specifically, the upper-level agent selects a sub-objective at a high-level temporal abstraction, and then the lower-level agent selects actions based on the environment and the selected sub-objective at several consecutive lower-level temporal abstractions. The collaborative learning between the two agents at different levels of temporal abstraction in hierarchical reinforcement learning is suitable for the separation required in P2 multi-stage optimization. Therefore, the long-term model computation strategy and the short-term data scheduling and acquisition strategy correspond to the learning of the upper-level agent and the learning of the lower-level agent, respectively.

[0065] This embodiment determines the loss function corresponding to the convolutional neural network model to be trained, updates the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function, and then determines the data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device. Based on the data upload information and cache queue information, a stochastic optimization problem is determined. Finally, based on the stochastic optimization problem, a long-term optimization strategy and a short-term optimization strategy are determined. Based on the long-term optimization strategy and the short-term optimization strategy, importance-based data acquisition and storage are performed, which improves the algorithm performance, reduces network resource consumption, and formulates an optimization problem that maximizes learning efficiency through joint data acquisition, scheduling, and computation. The main feature of the importance-based data acquisition and storage method provided in this embodiment is that it can determine the acquisition, transmission, storage, and computation strategies based on data importance and channel and system cache constraints.

[0066] Please refer to Figure 2 In one embodiment, the importance-based data acquisition and storage method provided in this application may further include:

[0067] Step 201: Based on the updated model convergence index and queue backlog information, determine the mini-batch value and important sample sources for each iteration process;

[0068] Step 202: Determine the computation delay corresponding to each iteration process based on the mini-batch value, sample size, workload required for sample gradient calculation, and CPU frequency.

[0069] Step 203: Determine the power consumption corresponding to each iteration process based on the constants corresponding to the edge server and the frequency of the central processing unit;

[0070] Step 204: Based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge server and the edge device, and the data calculation vector, determine the data scheduling vector and the transmit power;

[0071] Step 205: Determine the data upload rate based on the channel fading coefficient between the edge server and the edge device, the total bandwidth of the edge computing system, the channel noise, the transmit power, and the data scheduling vector.

[0072] Specifically, in this embodiment, the key sample source is edge devices capable of collecting important samples. It is assumed that the CPU frequency used by the Elasticsearch engine for model updates is adjusted only before the start of each model iteration. The CPU frequency for the i-th iteration is denoted as f. i The size of each sample is denoted as δ (i.e., the sample size in this embodiment, in bits / sample), and the workload required to compute the gradient of each sample (i.e., the workload required to compute the gradient of a sample in this embodiment) is denoted as ν (cycles / bit). Therefore, given the training task, the computational delay τ required for the i-th iteration process is... i As shown in Equation 3, the power consumption p of ES for the i-th iteration. s,i It can be represented as Here, κ is a constant related to the edge server.

[0073]

[0074]

[0075] The uplink sample transmission is controlled by orthogonal frequency division multiple access technology. Each time slot ES will divide the total bandwidth B of the edge computing system into multiple subbands b, and schedule ED to transmit important samples. Let α represent the data scheduling vector for the t-th time slot. k,t =1 indicates that the k-th ED uploads a sample in the t-th time slot. h k,t The channel fading coefficient between the ED and ES is represented by the data upload rate of the k-th ED in the t-th time slot, as shown in Formula 4, where p k,t For the transmission power, σ 2 Let be the channel noise. If the duration of each time slot is Δ, then the data upload amount of the k-th ED in the t-th time slot is equal to the data upload rate.

[0076] This embodiment determines the data upload rate by using the attribute information of the edge server, the attribute information of the edge device, and the frequency of the central processing unit.

[0077] In one embodiment, the importance-based data acquisition and storage method provided in this application may further include:

[0078] Step 211: Determine the data acquisition vector based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge device, and the data calculation vector;

[0079] Step 212: Determine the dynamic changes in the queue of the edge device based on the data upload rate, the data acquisition vector, and the queue backlog of the edge device;

[0080] Step 213: Determine the number of time slots corresponding to each iteration process based on the computational delay;

[0081] Step 214: Calculate the vector based on the data to determine the number of samples extracted from the cache queue corresponding to the edge server;

[0082] Step 215: Based on the number of samples extracted from the cache queue and the number of time slots corresponding to each iteration process, determine the average number of samples leaving the cache queue.

[0083] Specifically, after collecting real-time data from the environment, the ED (Edge Analyzer) labels the collected data to create samples, and then temporarily stores the samples in a local cache awaiting uploading. Considering that whether a sample leaves the cache queue depends on the selective scheduling of the ES (Elasticsearch Engine), the ED's collection rate is adjustable to prevent cache queue overflow. Let λ represent the data acquisition vector for the t-th time slot. k,t This represents the sampling frequency of EDk in the t-th time slot. Let represent the queue backlog of EDk at the beginning of the t-th time slot. The dynamic changes of the queue are shown in Formula 5.

[0084]

[0085]

[0086] At the same time, ES will also reserve a corresponding cache queue to store samples from ED. In order to improve the efficiency of online learning, only important samples will be selected as input for model training. Let D represent the data computation vector for the i-th iteration, where D k,i This represents the number of samples extracted from the k-th cache queue in each iteration. To facilitate data scheduling, the number of samples required for each iteration is mapped to each time slot. This represents the number of time slots during one iteration of the model; therefore, the average number of samples leaving the ES queue per time slot, D. k,t As shown in Formula 6.

[0087] This embodiment determines the average number of samples leaving the cache queue by using the attribute information of the edge server, the attribute information of the edge device, and the frequency of the central processing unit.

[0088] In one embodiment, the importance-based data acquisition and storage method provided in this application may further include:

[0089] Step 221: Determine the dynamic changes of the edge server's queue based on the sample size, the data upload rate, the average number of samples leaving the cache queue, and the queue backlog of the edge server.

[0090] Specifically, similarly, Let t represent the queue backlog of ES at the beginning of the t-th time slot. Then, the dynamic changes of the queue of the edge server are shown in Equation 7.

[0091]

[0092] This embodiment determines the dynamic changes in the edge server's queue by considering factors such as sample size, data upload rate, average number of samples leaving the cache queue, and queue backlog on the edge server.

[0093] In one embodiment, the importance-based data acquisition and storage method provided in this application may further include:

[0094] Step 310: Determine the target optimization problem based on the model convergence metric, the data upload information, and the cache queue information;

[0095] Step 320: The objective optimization problem is transformed into a stochastic optimization problem through Lyapunov optimization.

[0096] Specifically, after determining the data upload information and cache queue information, an MCM (Model Convergence Metric) is used to dynamically evaluate the importance of the data for model learning. The model convergence metric represents the importance of the samples collected from the sample source, and this, combined with factors such as channel buffering, determines the acquisition, transmission, storage, and computation strategy. As shown in Equation 8, where c1 and c2 are scaling factors related to the initial settings, the optimization problem P1 based on the MCM, which jointly involves data acquisition, scheduling, and computation, can be expressed as: C1, C3, and C5 show the feasible domain of data acquisition, scheduling, and computation variables. C2 and C4 ensure that the transmit power and acquisition frequency of each ED do not exceed the limits. C6 and C7 guarantee queue stability, i.e., that overflow will not occur for a long time.

[0097]

[0098]

[0099] Clearly, P1 is a highly complex stochastic optimization problem. This is manifested in several ways: 1. The gradient calculated using the current mini-batch not only determines the MCM of this iteration but also affects the long-term cumulative gradient. Therefore, data computation decisions are time-coupled. 2: Due to the limited buffer space of the device, data computation decisions affect data scheduling decisions, and similarly, data scheduling decisions also affect data acquisition decisions. Therefore, there is also coupling between multiple variables. 3: When the data computation strategy and channel conditions are unknown, the decision for each time slot is difficult to satisfy the long-term constraint of queue stability. Based on the above analysis and Lyapunov optimization techniques, P1 can be transformed into a stochastic optimization problem P2, as shown in Equation 9.

[0100] In this embodiment, the objective optimization problem is transformed into a stochastic optimization problem through Lyapunov optimization.

[0101] In one embodiment, the importance-based data acquisition and storage method provided in this application may further include:

[0102] Step 410: Based on the test set loss value of the current convolutional neural network model and the queue backlog corresponding to each iteration process, determine the model computation vector and the edge server computation frequency.

[0103] Step 420: Update the convolutional neural network model based on the observation state of the upper-layer agent corresponding to the PDQN algorithm and the model calculation vector, and determine the model learning performance reward corresponding to the long-term optimization strategy.

[0104] Step 430: Based on the channel state vector, the queue backlog of the edge devices, and the queue backlog of the edge servers, determine the data scheduling vector, transmit power, and data acquisition vector for each edge device;

[0105] Step 440: Determine the queue stability performance reward corresponding to the short-term optimization strategy based on the lower-level agent observation state corresponding to the DQN algorithm, the data scheduling vector, the transmission power, the data acquisition vector, and the average number of samples leaving the buffer queue.

[0106] Specifically, the long-term optimization strategy: Considering that the long-term optimization strategy contains both continuous and discrete variables, the PDQN (Parametrized Deep Q-Network) algorithm is used to solve it.

[0107] State Space: Although the purpose of data computation is to improve the learning efficiency of the model, it is also limited by cache resources. Therefore, the loss value of the current model on the test set, the queue backlog of ES and ED are selected as the states that the upper-layer agent needs to observe. in, and These are the backlog vectors of the edge device and edge server cache queues at the start of the i-th iteration, respectively.

[0108] Action Space: At the beginning of the i-th iteration, the upper-layer agent selects important samples to input into the learning model and allocates computational frequencies to the ES.

[0109] Reward function: The upper-level reward consists of two parts: one is the cumulative reward from the lower layers over multiple time slots, and the other is a reward related to the model's learning performance.

[0110] Short-term optimization strategy: Considering that the short-term optimization strategy only contains discrete variables, the DQN (Deep Q-Network) algorithm is used to solve it.

[0111] State space: The lower-level agent needs to observe two parts of the state: one part is the channel state and environment state obtained from the environment, and the other part is the average departure rate of the ES queue data determined by the sub-target. Among them, h t D represents the channel state vector. t This represents the average departure rate vector of the ES queue.

[0112] Action Space: The lower-level agent needs to learn the data acquisition and transmission strategy of the ED in each time slot.

[0113] Reward Function: Given a long-term model computation strategy, the goal of the lower-level learning is to ensure queue stability. Based on the stability component of the P2 objective, the internal reward is defined as follows:

[0114] In this embodiment, the long-term model computation strategy and the short-term data scheduling and acquisition strategy correspond to the upper-layer agent learning and the lower-layer agent learning, respectively.

[0115] The importance-based data acquisition and storage device provided by the present invention will be described below. The importance-based data acquisition and storage device described below and the importance-based data acquisition and storage method described above can be referred to in correspondence.

[0116] Please refer to Figure 3The present invention also provides a data acquisition and storage device based on importance, comprising:

[0117] The model update module 301 is used to determine the loss function corresponding to the convolutional neural network model to be trained, and update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function.

[0118] The first determining module 302 is used to determine data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device.

[0119] The second determining module 303 is used to determine the random optimization problem based on the data upload information and the cache queue information;

[0120] The third determining module 304 is used to determine a long-term optimization strategy and a short-term optimization strategy based on the stochastic optimization problem, and to collect and store data based on the long-term optimization strategy and the short-term optimization strategy.

[0121] Optionally, the first determining module includes:

[0122] The mini-batch value determination unit is used to determine the mini-batch value and important sample sources for each iteration process based on the updated model convergence index and queue backlog information.

[0123] The computation delay determination unit is used to determine the computation delay corresponding to each of the iteration processes based on the mini-batch value, sample size, workload required for sample gradient calculation, and central processing unit frequency.

[0124] The power consumption determination unit is used to determine the power consumption corresponding to each of the iterative processes based on the constant corresponding to the edge server and the frequency of the central processing unit.

[0125] The data scheduling vector determination unit is used to determine the data scheduling vector and transmit power based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge server and the edge device, and the data calculation vector;

[0126] The data upload rate determination unit is used to determine the data upload rate based on the channel fading coefficient between the edge server and the edge device, the total bandwidth of the edge computing system, the channel noise, the transmit power, and the data scheduling vector.

[0127] Optionally, the first determining module further includes:

[0128] The queue dynamic change determination unit is used to determine the data acquisition vector based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge device, and the data calculation vector, and to determine the queue dynamic change of the edge device according to the data upload rate, the data acquisition vector, and the queue backlog of the edge device.

[0129] A time slot number determination unit is used to determine the number of time slots corresponding to each iteration process based on the computation delay;

[0130] The sample quantity determination unit is used to calculate the vector based on the data and determine the number of samples extracted from the cache queue corresponding to the edge server.

[0131] The average number of samples leaving the cache queue is determined based on the number of samples extracted from the cache queue and the number of time slots corresponding to each iteration process.

[0132] Optionally, the first determining module further includes:

[0133] The queue dynamic change determination unit is used to determine the queue dynamic change of the edge server based on the sample size, the data upload rate, the average number of samples leaving the cache queue, and the queue backlog of the edge server.

[0134] Optionally, the second determining module includes:

[0135] The target optimization problem determination unit is used to determine the target optimization problem based on the model convergence index, the data upload information, and the cache queue information.

[0136] The problem transformation unit is used to transform the objective optimization problem into a stochastic optimization problem through Lyapunov optimization.

[0137] Optionally, the third determining module includes:

[0138] The first determining unit is used to determine the model computation vector and the edge server computation frequency based on the test set loss value of the current convolutional neural network model and the queue backlog corresponding to each iteration process.

[0139] The second determining unit is used to update the convolutional neural network model based on the observation state of the upper-layer agent corresponding to the PDQN algorithm and the model calculation vector, and to determine the model learning performance reward corresponding to the long-term optimization strategy.

[0140] The third determining unit is used to determine the data scheduling vector, transmit power, and data acquisition vector of each edge device based on the channel state vector, the queue backlog of the edge device, and the queue backlog of the edge server.

[0141] The fourth determining unit is used to determine the queue stability performance reward corresponding to the short-term optimization strategy based on the observation state of the lower-level agent corresponding to the DQN algorithm, the data scheduling vector, the transmission power, the data acquisition vector, and the average number of samples leaving the buffer queue.

[0142] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440. The processor 410, communications interface 420, and memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions from the memory 430 to execute a data acquisition and storage method based on importance.

[0143] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0144] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the importance-based data acquisition and storage methods provided by the above methods.

[0145] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the importance-based data acquisition and storage methods provided by the above methods.

[0146] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0147] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data acquisition and storage method based on importance, characterized in that, The importance-based data acquisition and storage method is applied to edge computing systems that include edge servers, edge devices, and central processing units. Importance-based data acquisition and storage methods include: Determine the loss function corresponding to the convolutional neural network model to be trained, and update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function; Based on the loss function, the attribute information of the edge server, and the attribute information of the edge device, the data upload information and cache queue information are determined. Based on the model convergence metric, the data upload information, and the cache queue information, the objective optimization problem is determined; the objective optimization problem is then transformed into a stochastic optimization problem through Lyapunov optimization. Based on the stochastic optimization problem, a long-term optimization strategy and a short-term optimization strategy are determined, and data is collected and stored based on the long-term optimization strategy and the short-term optimization strategy.

2. The importance-based data acquisition and storage method according to claim 1, characterized in that, The step of determining the data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device includes: Based on the updated model convergence metric and queue backlog information, the mini-batch value and important sample sources corresponding to each iteration process are determined. The computational delay for each iteration process is determined based on the mini-batch value, sample size, workload required for sample gradient calculation, and CPU frequency. Based on the constants corresponding to the edge server and the frequency of the central processing unit, the power consumption corresponding to each of the iterative processes is determined; Based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge server and the edge device, and the data calculation vector, the data scheduling vector and the transmit power are determined. The data upload rate is determined based on the channel fading coefficient between the edge server and the edge device, the total bandwidth of the edge computing system, the channel noise, the transmit power, and the data scheduling vector.

3. The importance-based data acquisition and storage method according to claim 2, characterized in that, The step of determining the data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device further includes: The data acquisition vector is determined based on the channel fading coefficient between the edge server and the edge device, the queue backlog of the edge device, and the data calculation vector. The dynamic changes in the queue of the edge device are determined based on the data upload rate, the data acquisition vector, and the queue backlog of the edge device. The number of time slots corresponding to each of the iterative processes is determined based on the computational delay. Based on the data calculation vector, determine the number of samples extracted from the cache queue corresponding to the edge server; Based on the number of samples extracted from the cache queue and the number of time slots corresponding to each iteration process, the average number of samples leaving the cache queue is determined.

4. The importance-based data acquisition and storage method according to claim 3, characterized in that, The step of determining the data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device further includes: The dynamic changes in the edge server's queue are determined based on the sample size, the data upload rate, the average number of samples leaving the cache queue, and the queue backlog of the edge server.

5. The importance-based data acquisition and storage method according to claim 1, characterized in that, The process of determining long-term and short-term optimization strategies based on the stochastic optimization problem, and then performing importance-based data collection and storage based on the long-term and short-term optimization strategies, includes: Based on the test set loss value of the current convolutional neural network model and the queue backlog corresponding to each iteration process, determine the model computation vector and the edge server computation frequency; Based on the observation state of the upper-layer agent corresponding to the PDQN algorithm and the model calculation vector, the convolutional neural network model is updated to determine the model learning performance reward corresponding to the long-term optimization strategy. Based on the channel state vector, the queue backlog of the edge devices, and the queue backlog of the edge servers, the data scheduling vector, transmit power, and data acquisition vector of each edge device are determined. Based on the observation state of the lower-level agent corresponding to the DQN algorithm, the data scheduling vector, the transmission power, the data acquisition vector, and the average number of samples leaving the buffer queue, the queue stability performance reward corresponding to the short-term optimization strategy is determined.

6. A data acquisition and storage device based on importance, characterized in that, include: The model update module is used to determine the loss function corresponding to the convolutional neural network model to be trained, and update the convolutional neural network model based on the mini-batch stochastic gradient descent algorithm and the loss function. The first determining module is used to determine data upload information and cache queue information based on the loss function, the attribute information of the edge server, and the attribute information of the edge device; The second determining module is used to determine the target optimization problem based on the model convergence index, the data upload information, and the cache queue information. The objective optimization problem is transformed into a stochastic optimization problem through Lyapunov optimization. The third determining module is used to determine the long-term optimization strategy and the short-term optimization strategy based on the stochastic optimization problem, and to collect and store data based on the long-term optimization strategy and the short-term optimization strategy.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the importance-based data acquisition and storage method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the importance-based data acquisition and storage method as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the importance-based data acquisition and storage method as described in any one of claims 1 to 5.

Images