Cloud-edge cooperative industrial wireless channel prediction method based on multi-time scale attention mechanism

Abstract

The invention discloses a cloud-edge cooperative industrial wireless channel prediction method based on a multi-time scale attention mechanism. The method comprises the following steps: establishing a cloud-edge cooperative industrial wireless channel prediction architecture; establishing a multi-target dynamic channel prediction task scheduling algorithm and deploying the multi-target dynamic channel prediction task scheduling algorithm in an edge server; establishing a cloud-edge cooperative industrial wireless channel prediction model comprising a cloud prediction model deployed on a cloud server and an edge prediction model deployed on an edge server; the industrial terminal equipment collects channel data in real time, marks the channel data as LoS or NLoS scene data based on channel characteristics, and uploads the marked channel data to the edge server; and after the edge server receives the marked channel data uploaded by the industrial terminal equipment, a channel prediction task is allocated to the edge server or the cloud server through a multi-target dynamic channel prediction task scheduling algorithm, the channel prediction task is completed, and a channel prediction result is obtained. The method is high in robustness, generalization and practicability, high in precision and low in delay.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication, and particularly to an industrial wireless channel prediction method. Background Art

[0003] Traditional deep learning methods (such as CNN, LSTM), for example, the invention with application number 202211281296.X discloses a wireless network channel state prediction method, device, equipment and storage medium. Although good results have been achieved in some applications, there are two problems in heterogeneous environments. First, existing models are difficult to effectively distinguish the commonalities and differences of channels in LoS and NLoS scenarios, resulting in limited generalization performance. Although multi-task learning (MTL) improves heterogeneous modeling ability by sharing features, its multi-objective optimization process is complex and difficult to deploy in resource-constrained environments. Second, traditional deep learning models lack a real-time adjustment mechanism, and their robustness and adaptability are insufficient in the face of drastic fluctuations in channel states.

[0004] With the development of cloud computing and edge computing, the cloud-edge collaborative architecture, as an effective solution, has gradually been applied to the prediction task of industrial wireless channels. The cloud server is mainly responsible for processing global feature modeling and complex data calculations, and analyzing the global patterns and long-term trends of channels; the edge computing node is responsible for near-real-time data processing and short-term dynamic prediction to reduce communication latency and computing load. However, the existing cloud-edge collaborative architecture still has some deficiencies, especially in the task scheduling strategy. Traditional static scheduling and heuristic methods are often difficult to adapt to the dynamic changes of channel states, resulting in difficulty in achieving an ideal balance between real-time performance and prediction accuracy.

[0005] To solve this problem, in recent years, the deep reinforcement learning (DRL) method has been widely used in multi-task optimization and resource scheduling. Through dynamic decision optimization, DRL can adapt to complex industrial environments. However, most of the existing DRL-based scheduling methods focus on the optimization of a single objective (such as minimizing latency), lacking a mechanism to effectively handle the conflicts between multiple objectives (such as latency and prediction accuracy), which limits its application in industrial environments.

[0006] To make up for the deficiencies in the accuracy of the above channel prediction and the latency of the prediction task, it is necessary to design an industrial wireless channel prediction method that is more suitable for actual industrial scenarios, has better generalization, high accuracy, and low latency. Summary of the Invention

[0007] Aiming at the technical problems of low accuracy and high latency in the prediction of industrial wireless channels by existing methods, the present invention proposes a cloud-edge collaborative industrial wireless channel prediction method based on a multi-time scale attention mechanism. A prediction model with a multi-time scale attention model as the core is deployed in the designed cloud-edge collaborative industrial wireless channel prediction architecture, and a multi-objective dynamic channel prediction task scheduling algorithm is used to intelligently allocate prediction tasks, realizing a dynamic balance between prediction accuracy and latency between the cloud and the edge.

[0008] To achieve the above object, the technical solution of the present invention is realized as follows:

[0009] A cloud-edge collaborative industrial wireless channel prediction method based on a multi-time scale attention mechanism, comprising the following steps:

[0010] S1. Establish a cloud-edge collaborative industrial wireless channel prediction architecture including one cloud server, at least one edge server communicating with the cloud server, and at least one industrial terminal device communicating with the edge server;

[0011] S2. Based on the double deep Q network and the time series weighted attention mechanism, establish a multi-objective dynamic channel prediction task scheduling algorithm and deploy it on the edge server; establish a cloud-edge collaborative industrial wireless channel prediction model based on the multi-time scale attention mechanism, including a cloud prediction model deployed on the cloud server and an edge prediction model deployed on the edge server;

[0012] S3. The industrial terminal device collects channel data in real time and marks the channel data as LoS or NLoS scenario data based on the channel characteristics, and uploads the marked channel data to the edge server;

[0013] S4. After the edge server receives the marked channel data uploaded by the industrial terminal device, it performs channel prediction task allocation through the multi-objective dynamic channel prediction task scheduling algorithm; if the channel prediction task is assigned to the edge server, the edge server inputs the marked channel data into the edge prediction model to obtain a channel state prediction result; if the channel prediction task is assigned to the cloud server, the edge server uploads the marked channel data to the cloud server, and the cloud server obtains a channel state prediction result through the cloud prediction model.

[0014] Further, before establishing the multi-objective dynamic channel prediction task scheduling algorithm, it is necessary to obtain a Markov decision model for multi-objective dynamic channel prediction task scheduling. The method is: model the state space and the action space of the edge server; establish a reward function that comprehensively considers channel prediction accuracy and channel prediction task latency;

[0015] The modeling of the state space and the action space of the edge server includes the following steps:

[0016] Within time slot t, the system state consists of the load of the edge server, network bandwidth, data volume, and transmission delay;

[0017] The action a selected by the agent t ={0, 1}, where the action a t =0 indicates that the channel prediction task is assigned to the edge server, and the action a t =1 indicates that the channel prediction task is assigned to the cloud server. When the load L of the edge server t is greater than or equal to the upper load limit L of the edge server max , the channel prediction task is only uploaded to the cloud server, and at this time the action a t =1, where L max represents the upper load limit of the edge server;

[0018] The method for establishing the reward function that comprehensively considers channel prediction accuracy and channel prediction task delay is as follows: establish the delay reward within time slot t; establish the error reward within time slot t; combine the delay reward and the error reward to obtain the reward function.

[0019] Furthermore, the expression of the system state is: s t ={L t , B t , C t , D t}, where L t is the load of the edge server, B t is the network bandwidth of the edge server, C t is the data volume of the edge server, and D t is the transmission delay of the edge server;

[0020] The expression of the delay reward is:

[0021]

[0022] where t tans represents the data transmission delay, and t infer represents the inference delay of the cloud-edge collaborative industrial wireless channel prediction model; the expression of the error reward is:

[0023]

[0024] where Error(s t , a t ) represents the specific measure of the prediction error;

[0025] The expression of the reward function is: where α1 and α2 are weight coefficients and are both greater than zero.

[0026] Furthermore, the method for establishing a multi-objective dynamic channel prediction task scheduling algorithm based on the double deep Q-network and the temporal weighted attention mechanism is as follows: Establish a double deep Q-network including an independent delay network and an error network, and introduce the designed temporal weighted attention mechanism before calculating the Q value to obtain a double deep Q-network based on temporal weighted attention;

[0027] After the edge server receives the marked channel data uploaded by the industrial terminal device, the method for allocating channel prediction tasks through the multi-objective dynamic channel prediction task scheduling algorithm is as follows: The edge server uses the double deep Q-network based on temporal weighted attention, with the current system state s t as the input, calculates the reward value through the reward function r t , selects the action a t based on the reward value through the ε-greedy strategy, and allocates channel prediction tasks according to the action a t .

[0028] Furthermore, the method for the double deep Q-network based on temporal weighted attention to output the Q value through the designed temporal weighted attention mechanism is as follows:

[0029] Estimate the delay Q value and the error Q value

[0030]

[0031] where θ l and θ e are the parameters of the delay network and the error network respectively;

[0032] Dynamically adjust the weights of the delay Q value and the error Q value based on the historical context vector and the average reward signal to generate dynamic weights where and are the average delay and average error respectively, and represent the weights of delay and error respectively;

[0033] Multiply the dynamic weights element-wise with the context vector v t of the current time slot to obtain the weighted context vector and add it to the delay Q value and the error Q value Combine the input multi-head attention mechanism to capture short-term and long-term dependence characteristics, then concatenate the outputs of the multi-head attention mechanism and generate the comprehensive Q value Q through a fully connected layer t 。

[0034] Furthermore, the cloud prediction model is composed of a multi-time scale attention layer, a task-specific branch layer, and an output layer connected in sequence; the edge prediction model is composed of a multi-time scale attention layer and an output layer connected in sequence; the multi-time scale attention layers in the cloud prediction model and the edge prediction model are the same;

[0035] The multi-time scale attention layer is composed of a variable selection network, an LSTM encoder, a Gate and an Add&Norm layer Ⅰ, a gated residual network Ⅰ, a multi-scale local self-attention network, a Gate and an Add&Norm layer Ⅱ, and a gated residual network Ⅱ connected in sequence.

[0036] Furthermore, both the cloud prediction model and the edge prediction model extract multi-time scale features from the labeled channel data through the multi-time scale attention layer, including the following steps:

[0037] First, organize the labeled channel data into a sliding window time series, and the sliding window time series is fed into the corresponding variable selection network. After being processed by the variable selection network, the weighted features of each feature vector in the sliding window time series are obtained, and then the weighted feature matrix of the sliding window time series is obtained;

[0038] The weighted feature matrix of the sliding window time series is input into the LSTM encoder to extract the time series dependence features, and a hidden state sequence is obtained;

[0039] Perform a dropout operation on the hidden state sequence, generate a sparsified state sequence and feed it into the Gate and Add&Norm layer Ⅰ to obtain a gated state sequence;

[0040] Further extract features from the gated state sequence through the gated residual network Ⅰ, generate the features input to the multi-scale local self-attention network, and after being processed by the multi-scale local self-attention network, obtain a multi-time scale feature matrix;

[0041] The multi-time scale feature matrix is processed by the Gate and Add&Norm layer Ⅱ and the gated residual network Ⅱ in sequence to obtain multi-time scale features.

[0042] Furthermore, in the edge prediction model, the multi-time scale features are fed into the output layer, and in the output layer, the multi-time scale features are mapped to the channel state prediction results to obtain the channel state prediction sequence for the next T prediction time steps

[0043] Further, in the cloud prediction model, the multi-time scale features are passed to the task-specific branch layer for feature extraction for LoS and NLoS scenarios respectively;

[0044] For the LoS scenario, the multi-time scale features are sequentially subjected to feature extraction through LSTM layer Ⅰ, gated residual network Ⅲ, and LSTM layer Ⅱ to obtain long-term global features O′ is the multi-time scale feature, k is the window length, d h is the hidden layer dimension, GRN is the gated residual network operation, and LSTM is the long short-term memory network operation;

[0045] For the NLoS scenario, the multi-time scale features are sequentially subjected to feature extraction through one-dimensional convolutional network Ⅰ, gated residual network Ⅳ, and one-dimensional convolutional network Ⅱ to obtain local features Conv-1D is the one-dimensional convolution operation;

[0046] The long-term global features and local features are fused through a gating mechanism. The fusion process is as follows:

[0047] G LoS =σ(W LoS H LoS +b LoS )

[0048] H fusion =G LoS ⊙H LoS +(1 - G LoS )⊙H NLoS

[0049] where W LoS and b LoS are the weight and bias of the gating unit respectively, G Los is the gating value, σ(·) is the Sigmoid activation function, and H fusion is the fused feature matrix;

[0050] The fused feature matrix H fusion is passed to the output layer to obtain the channel state prediction sequence for the next T prediction time steps

[0051] Further, the edge prediction model shares the parameters of the multi-time scale attention layer with the cloud prediction model through a shallow parameter sharing mechanism, and the edge prediction model updates the parameters of the multi-time scale attention layer through a shallow parameter update mechanism:

[0052] (1) In the shallow parameter sharing mechanism, the parameters Θ sharedIt is unloaded to the edge server, replacing the parameter set of the multi-time scale attention layer in the edge prediction model. The parameters of the edge prediction model are represented as:

[0053] Θ Edge-PM ={Θ shared , Θ adjustable}

[0054] Among them, Θ adjustable represents the specific parameters of the edge prediction model. The parameters of the shared multi-time scale attention layer are frozen, and only the Θ adjustable part needs to be trained to adapt to the specific task requirements of the edge server;

[0055] (2) In the shallow parameter update mechanism, when the real-time data collected by the edge server reaches the set capacity, the data will be uploaded to the cloud server. The cloud server fine-tunes the parameters of the multi-time scale attention layer of the cloud prediction model and shares them with the edge server.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] 1. The present invention comprehensively considers the complexity and high dynamics of the industrial wireless channel environment, and based on data-driven deep learning and reinforcement learning, establishes a cloud-edge collaborative industrial wireless channel prediction architecture based on hierarchical intelligent scheduling, with high accuracy and low latency.

[0058] 2. The present invention has strong generalization and practicability. The cloud-edge collaborative industrial wireless channel prediction model based on the multi-time scale attention mechanism can solve problems such as the heterogeneity of channel data in complex dynamic scenarios, and makes up for the lack of accuracy of past algorithm channel predictions.

[0059] 3. The present invention has strong robustness. Through the designed multi-objective dynamic channel prediction task scheduling, it can adapt to industrial scenarios with harsh environments, and has strong real-time performance and high reliability in actual industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0061] Figure 1 It is a schematic structural diagram of the cloud-edge collaborative industrial wireless channel prediction architecture based on hierarchical intelligent scheduling of the present invention.

[0062] Figure 2 It is a flowchart of the present invention.

[0063] Figure 3 It is a schematic structural diagram of the multi-objective dynamic channel prediction task scheduling algorithm of the present invention.

[0064] Figure 4 It is an architecture diagram of the time-series weighted attention mechanism of the present invention.

[0065] Figure 5 It is a diagram of the cloud-edge collaborative industrial wireless channel prediction model based on the multi-time-scale attention mechanism of the present invention.

[0066] Figure 6 It is a multi-scale local self-attention network diagram of the present invention. Detailed implementation manners

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0068] As Figure 1 and Figure 2 shown, a cloud-edge collaborative industrial wireless channel prediction method based on the multi-time-scale attention mechanism includes the following steps:

[0069] S1. Considering the complex dynamic channel scenario in industry, a cloud-edge collaborative industrial wireless channel prediction architecture based on hierarchical intelligent scheduling is established. The cloud-edge collaborative industrial wireless channel prediction architecture includes: one cloud server, N (N≥1) edge servers communicating with the cloud server, and M (M≥1) industrial terminal devices communicating with the edge servers.

[0070] Furthermore, the industrial terminal devices include intelligent devices such as automatically guided vehicles (AGVs) deployed distributively. The industrial terminal device m∈M is responsible for collecting channel data in real time, preliminarily differentiating the collected channel data based on channel characteristics (such as statistical characteristics such as signal path loss, reflection degree, and signal-to-noise ratio) in each time slot t, and marking it as LoS or NLoS scenario data. The marked channel data is uploaded to the edge server;

[0071] Furthermore, the edge server is used for dynamic scheduling and local execution of channel prediction tasks. After receiving the marked channel data uploaded by the industrial terminal device, the edge server makes a dynamic trade-off based on load, bandwidth, data volume, and transmission delay, and decides whether to execute the channel prediction task locally to reduce latency, or upload the marked channel data to the cloud server for the cloud server to execute the channel prediction task to obtain a higher-precision prediction result.

[0072] Furthermore, the cloud server is used for global modeling of complex channel characteristics and performing high-precision channel prediction tasks, and sharing shallow parameters with the edge server to dynamically update the performance of the edge prediction model in the edge server.

[0073] S2. Considering the problem of imbalance between prediction accuracy and real-time performance caused by complex dynamic channel environments, a multi-objective dynamic channel prediction task scheduling algorithm is established based on the double deep Q-network and the temporal weighted attention mechanism and deployed on the edge server to jointly optimize the channel prediction accuracy and the processing delay of the channel prediction task. A cloud-edge collaborative industrial wireless channel prediction model based on a multi-time scale attention mechanism is established, including a cloud prediction model deployed on the cloud server and an edge prediction model deployed on the edge server.

[0074] Furthermore, before establishing the multi-objective dynamic channel prediction task scheduling algorithm, a Markov decision model based on the multi-objective dynamic channel prediction task scheduling needs to be obtained. The method is as follows: modeling the state space and the action space of the edge server; establishing a reward function that comprehensively considers the channel prediction accuracy and the channel prediction task delay.

[0075] Specifically, modeling the state space and the action space of the edge server includes the following steps:

[0076] In time slot t, the system state consists of the load L t of the edge server, the network bandwidth B t , the data volume C t and the transmission delay D t . The expression of the system state is:

[0077] s t = {L t , B t , C t , D t}

[0078] The action selected by the agent is defined as a t = {0, 1}, where the action a t = 0 means that the channel prediction task is assigned to the edge server, and the action a t = 1 means that the channel prediction task is assigned to the cloud server. When the load of the edge server is greater than or equal to the load upper limit of the edge server, that is, L t ≥ L max , the channel prediction task is only uploaded to the cloud server. At this time, the action a t = 1, where L max represents the load upper limit of the edge server.

[0079] Establish a reward function that comprehensively considers the channel prediction accuracy and the latency of the channel prediction task, including the following steps:

[0080] To minimize the processing latency of the channel prediction task, within time slot t, define the latency reward

[0081] where t tans represents the data transmission latency, and t infer represents the inference latency of the cloud-edge collaborative industrial wireless channel prediction model.

[0082] Within time slot t, measure the accuracy of channel prediction through the error reward, and the error reward is defined as:

[0083]

[0084] where Error(s t , a t ) represents the specific measure of the prediction error (such as the mean squared error MSE).

[0085] In summary, at time slot t, the composite reward function r of the edge server t is expressed as:

[0086]

[0087] where α1 and α2 are weight coefficients and are both greater than zero.

[0088] Furthermore, the method for establishing a multi-objective dynamic channel prediction task scheduling algorithm based on the double deep Q-network and the temporal weighted attention mechanism is as follows: establish a double deep Q-network including an independent latency network and an error network, and introduce the designed temporal weighted attention mechanism before calculating the Q value to obtain a double deep Q-network based on temporal weighted attention.

[0089] Furthermore, as Figure 3 shown, before the multi-objective dynamic channel prediction task scheduling algorithm is deployed on the edge server, it needs to be trained. In the training stage, it includes two parts: action training and network training. The specific steps are as follows:

[0090] First, in the action training part, use the independent latency network and error network to estimate the latency Q value and the error Q value

[0091]

[0092] where θ l and θ e are the parameters of the latency network and the error network respectively.

[0093] As Figure 4 shown, the delayed Q-value is dynamically adjusted through the designed temporal weighted attention mechanism and the error Q-value weights, based on the historical context vector and the average reward signal where and are the average delay and average error respectively, and through the GRU (Gated Recurrent Unit) module, a dynamic weight is generated where and represent the weights of delay and error respectively.

[0094] The dynamic weight is element-wise multiplied by the context vector v of the current time slot t to obtain the weighted context vector and combined with the delayed Q-value and the error Q-value to input into the multi-head attention mechanism to capture short-term and long-term dependence characteristics, and then the outputs of the multi-head attention mechanism are concatenated and passed through a fully connected layer to generate the comprehensive Q-value Q t .

[0095] After the masking operation filters out infeasible actions, a set of feasible actions is generated where m t is the mask vector, and the ε-greedy strategy is used to select the action a from the set of feasible actions t , finally, the interaction experience obtained after executing the action a t is stored in the prioritized experience replay pool. Where, s t+1 is the system state within time slot t+1.

[0096] After the edge server receives the marked channel data uploaded by the industrial terminal device, the method for channel prediction task allocation through the multi-objective dynamic channel prediction task scheduling algorithm is as follows: The edge server uses the dual deep Q-network based on temporal weighted attention, with the current system state s t as the input, calculates the reward value through the reward function r t , and selects the action a based on the reward value through the ε-greedy strategy t : Locally execute the channel prediction task (a t =0) to reduce the delay, or upload the marked channel data to the cloud server (a t =1) to improve the prediction accuracy.

[0097] To increase the training efficiency, the network training part is trained and optimized by a dual deep Q-network based on temporal weighted attention, including the following steps:

[0098] During the training process of the entire multi-objective dynamic channel prediction task scheduling algorithm, training samples are sampled according to the priority of TD error, and the TD errors of the delay network and the error network are calculated:

[0099]

[0100] where δ l is the TD error of the delay network, δ e is the TD error of the error network, γ ∈ (0, 1) is the discount factor, and are the target Q-values of the delay network and the error network respectively, generated by the target networks corresponding to the delay network and the error network respectively, a t ′ is the action selected by s t+1 , is the parameter of the target network corresponding to the delay network, is the parameter of the target network corresponding to the error network, and is periodically synchronized and updated from the policy networks of the delay network and the error network.

[0101] The loss functions of the delay network and the error network are expressed as:

[0102]

[0103] where w n is the weight based on the priority of the nth sample, and the parameters of the delay network and the error network are updated separately by gradient descent, δ l,n and δ e,n are the TD errors of the nth empirical sample calculated by the delay network and the error network respectively;

[0104] The parameters of the delay target network and the error target network are updated by minimizing the loss function, expressed as:

[0105]

[0106] where α is the learning rate, is the gradient of the loss function L l (θ l ) with respect to the parameter θ l , is the gradient of the loss function L e (θ e ) with respect to the parameter θ e ; To ensure the training stability, the target network parameters and are synchronized with the policy network after a fixed step size C.

[0107] S3. Industrial terminal devices collect channel data in real time, mark the channel data as LoS or NLoS scenario data based on channel characteristics, and upload the marked channel data to the edge server.

[0108] S4. After receiving the marked channel data uploaded by industrial terminal devices, the edge server performs channel prediction task allocation through a multi-objective dynamic channel prediction task scheduling algorithm; if the channel prediction task is assigned to the edge server, the edge server inputs the marked channel data into the edge prediction model to obtain the channel state prediction result; if the channel prediction task is assigned to the cloud server, the edge server uploads the marked channel data to the cloud server, and the cloud server obtains the channel state prediction result through the cloud prediction model.

[0109] Furthermore, the cloud prediction model synchronously optimizes feature extraction through multi-task learning, dynamically integrates multi-time scale features, the edge prediction model adopts shallow parameter sharing to achieve lightweight deployment, and combines a dynamic update mechanism to maintain prediction performance in a complex dynamic industrial environment. The overall architecture of the cloud-edge collaborative industrial wireless channel prediction model is as Figure 5 shown.

[0110] Specifically, the cloud prediction model is composed of a multi-time scale attention layer, a task-specific branch layer, and an output layer connected in sequence; the edge prediction model is composed of a multi-time scale attention layer and an output layer connected in sequence. The multi-time scale attention layers in the cloud prediction model and the edge prediction model are the same. The multi-time scale attention layer is composed of a variable selection network (VSN), an LSTM encoder, a Gate (gating mechanism layer), and an Add&Norm (residual connection normalization) layer Ⅰ, a gated residual network Ⅰ (GRN), a multi-scale local self-attention network, a Gate (gating mechanism layer), and an Add&Norm (residual connection normalization) layer Ⅱ, and a gated residual network Ⅱ (GRN) connected in sequence.

[0111] Furthermore, both the cloud prediction model and the edge prediction model extract multi-time scale features from the marked channel data through the multi-time scale attention layer, including the following steps:

[0112] At each time slot t, the marked channel data is x t ;

[0113] Convert the original channel data into a format suitable for time series processing to capture historical information, and organize the input marked channel data x t into a sliding window time series X = [x t-k+1 , x t-k+2 ,..., x t , where Let \(k\) be the window length and \(D\) be the feature dimension of each time slot. Among them, the labeled channel data \(x\) in the sliding window time series \(X\) t is represented in the form of a feature vector. At this time, where is the \(j\)-th channel feature component.

[0114] Input the sliding window time series \(X = [x t-k+1 , x t-k+2 ,..., x t into the corresponding variable selection network (VSN). First, perform gated residual network (GRN) transformation on each feature component to obtain the enhanced feature component Then, the variable selection network (VSN) generates a dynamic weight vector \(\lambda\) t based on the labeled channel data \(x\) in the form of a feature vector t , where where is the weight of the \(d\)-th feature component . Different weights are assigned to each feature component to suppress irrelevant features and improve the efficiency and generalization ability of the model. Then, perform element-wise weighting on the enhanced feature component to obtain the weighted feature and construct the weighted feature matrix of the sliding window time series

[0115] Input the weighted feature matrix of the sliding window time series into the LSTM encoder to extract temporal dependence features and obtain the hidden state sequence where \(d\) h is the hidden layer dimension.

[0116] To prevent overfitting, the hidden state sequence \(H\) undergoes a dropout operation to generate a sparsified state sequence Subsequently, it is passed to the Gate (gating mechanism layer) and Add&Norm (residual connection normalization) layer Ⅰ to obtain the gated state sequence Then, through the gated residual network Ⅰ (GRN), the gated state sequence is further refined to make its features more representative, and finally, the features input to the multi-scale local self-attention network are generated

[0117] As Figure 6 shown, in the multi-scale local self-attention network, a query matrix \(Q\) i , a key matrix \(K\) i and a value matrix \(V\) i are generated, and multi-scale convolution processing is applied to the key and value matrices to obtain:

[0118] Among them, V i and K i are the key and value matrices after being processed by the multi-scale convolutional network respectively, MSC is the scale convolution operation, is the output channel number of the multi-scale one-dimensional convolution for processing the key matrix in the i-th branch, is the output channel number of the multi-scale one-dimensional convolution for processing the value matrix in the i-th branch.

[0119] On this basis, the time dynamic relationship is calculated through self-attention to capture the long-range dependence relationship:

[0120]

[0121] Among them, represents the dot product result of the query and the key, is the output matrix of the i-th attention head, d k is the dimension of the key.

[0122] Next, the output matrices of multiple attention heads are concatenated and linearly transformed to generate the final multi-time-scale feature matrix to improve the expression ability of the model, which is expressed as:

[0123] O = Concat(O1, O2, …, O h ,)W E

[0124] Among them, h is the number of attention heads, Concat is the concatenation operation, represents the output feature matrix of the multi-scale local self-attention network, d o is the output feature dimension, is the linear transformation matrix. The multi-time-scale feature matrix O is processed successively by the Gate (gating mechanism layer) and Add&Norm (residual connection normalization) layer II, and the gated residual network II (GRN) to obtain the multi-time-scale feature

[0125] In the edge prediction model, the multi-time-scale feature O' is input into the output layer, and the multi-time-scale feature is mapped to the channel state prediction result in the output layer, and finally the channel state prediction sequence for the future T prediction time steps is obtained

[0126] In the cloud prediction model, the multi-time-scale feature O' is passed to the task-specific branch layer to extract features for the LoS and NLoS scenarios respectively.

[0127] For the LoS scenario, the multi-time scale feature O′ is successively subjected to feature extraction by the LSTM layer Ⅰ, the gated residual network Ⅲ (GRN), and the LSTM layer Ⅱ to obtain the long-term global feature Among them, GRN is the gated residual network operation, and LSTM is the long short-term memory network operation.

[0128] For the NLoS scenario, the multi-time scale feature O′ is successively subjected to feature extraction by the one-dimensional convolutional network Ⅰ, the gated residual network Ⅳ (GRN), and the one-dimensional convolutional network Ⅱ. The one-dimensional convolutional network is used to extract local features, and GRN is used to further refine the feature expression to obtain local features Among them, Conv-1D is the one-dimensional convolutional operation.

[0129] To achieve the efficient fusion of the features of the two heterogeneous branches, a gated mechanism fusion method is introduced, and the fusion process is as follows:

[0130] G LoS =σ(W LoS H LoS +b LoS ),

[0131] H fusion =G LoS ⊙H LoS +(1 - G LoS )⊙H NLoS ,

[0132] Among them, W LoS and b LoS are the weight and bias of the gated unit respectively, G Los is the gated value, σ(·) is the Sigmoid activation function, and the fused feature matrix retains the complementary information of the channel features of the LoS and NLoS scenarios.

[0133] The fused feature matrix H fusion is passed to the output layer to obtain the channel state prediction sequence for the next T prediction time steps

[0134] Furthermore, the cloud prediction model is optimized using a multi-task loss function, and the multi-task loss function is:

[0135] L = αL LoS +β LNLoS

[0136] Among them, L LoS and L NLoS measure the mean square error of the channel prediction for the LoS and NLoS scenarios respectively, α and β are weight coefficients greater than zero and α + β = 1, and L is the value of the multi-task loss function.

[0137] Furthermore, the edge prediction model shares the parameters of the multi-time scale attention layer with the cloud prediction model through a shallow parameter sharing mechanism, and the edge prediction model updates the parameters of the multi-time scale attention layer through a shallow parameter update mechanism:

[0138] (1) In the shallow parameter sharing mechanism, the parameters Θ of the multi-time scale attention layer shared from the cloud prediction model shared are offloaded to the edge server, replacing the parameter set of the multi-time scale attention layer in the edge prediction model. The parameters of the edge prediction model are represented as

[0139] Θ Edge-PM ={Θ shared , Θ adjustable}

[0140] where Θ adjustable represents the specific parameters of the edge prediction model. The shared parameters of the multi-time scale attention layer are frozen, and only the specific parameter Θ adjustable part of the edge prediction model needs to be trained to adapt to the specific task requirements of the edge server.

[0141] (2) In the shallow parameter update mechanism, when the real-time data collected by the edge server reaches the set capacity, the data will be uploaded to the cloud server. The cloud server fine-tunes the parameters of the multi-time scale attention layer of the cloud prediction model and shares them with the edge server.

[0142] For heterogeneous modeling of LoS / NLoS scenarios, the cloud prediction model synchronously optimizes feature extraction through multi-task learning, dynamically integrates multi-time scale features, and the edge prediction model achieves lightweight deployment through shallow parameter sharing, combined with a dynamic update mechanism to maintain prediction performance in a complex dynamic industrial environment. The entire cloud-edge collaborative industrial wireless channel prediction model has significant advantages in terms of real-time performance and prediction accuracy.

[0143] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism, characterized in that: The following steps are involved: S1. Establish a cloud-edge collaborative industrial wireless channel prediction architecture including a cloud server, at least one edge server communicating with the cloud server, and at least one industrial terminal device communicating with the edge server; S2. Based on dual deep Q network and time-weighted attention mechanism, a multi-target dynamic channel prediction task scheduling algorithm is established and deployed on the edge server; a cloud-edge collaborative industrial wireless channel prediction model based on multi-time scale attention mechanism is established, including a cloud prediction model deployed on the cloud server and an edge prediction model deployed on the edge server; S3, the industrial terminal equipment collects channel data in real time and marks the channel data as LoS or NLoS scenario data based on the channel characteristics, and uploads the marked channel data to the edge server; S4. After the edge server receives the marked channel data uploaded by the industrial terminal device, it allocates the channel prediction task through the multi-objective dynamic channel prediction task scheduling algorithm; if the channel prediction task is assigned to the edge server, the edge server inputs the marked channel data into the edge prediction model to obtain the channel state prediction result; if the channel prediction task is assigned to the cloud server, the edge server uploads the marked channel data to the cloud server, and the cloud server obtains the channel state prediction result through the cloud prediction model.

2. According to claim 1, the cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism is characterized in that: Before establishing a multi-objective dynamic channel prediction task scheduling algorithm, it is necessary to obtain a Markov decision model based on multi-objective dynamic channel prediction task scheduling. The method is as follows: modeling the state space and the action space of the edge server; establishing a reward function that comprehensively considers the channel prediction accuracy and the channel prediction task delay; The modeling of the state space and the action space of the edge server comprises the following steps: In time slot t, the system state consists of the edge server load, network bandwidth, data volume, and transmission delay; The action a chosen by the agent t ={0,1}, where action a t =0 means that the channel prediction task is assigned to the edge server, action a t =1 means that the channel prediction task is assigned to the cloud server. When the load of the edge server is L t Greater than or equal to the upper limit of the edge server load L mmm When the channel prediction task is only uploaded to the cloud server, action a t =1, where L mmm Indicates the upper limit of the edge server load; The method for establishing a reward function that comprehensively considers the channel prediction accuracy and the channel prediction task delay is: establishing a delay reward within time slot t; establishing an error reward within time slot t; and combining the delay reward and the error reward to obtain a reward function.

3. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 2 is characterized in that: The expression of the system state is: t ={L t ,B t ,C t ,D t }, where L t is the load of the edge server, B t is the network bandwidth of the edge server, C t is the data volume of the edge server, D t is the transmission delay of the edge server; The expression of the delayed reward is: Among them, t tmtt represents the data transmission delay, t itiii Represents the inference latency of the cloud-edge collaborative industrial wireless channel prediction model; The expression of the error reward is: Among them, Error(s t ,a t ) represents a specific measure of the prediction error; The expression of the reward function is: Among them, α1 and α2 are weight coefficients and are both greater than zero.

4. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 1 or 3 is characterized in that: The method for establishing a multi-objective dynamic channel prediction task scheduling algorithm based on a dual deep Q network and a temporal weighted attention mechanism is as follows: establishing a dual deep Q network including an independent delay network and an error network, and introducing a designed temporal weighted attention mechanism before Q value calculation to obtain a dual deep Q network based on temporal weighted attention; After the edge server receives the marked channel data uploaded by the industrial terminal device, the method for allocating channel prediction tasks through the multi-objective dynamic channel prediction task scheduling algorithm is as follows: the edge server uses a dual-depth Q network based on time-series weighted attention to allocate channel prediction tasks based on the current system state s t As input, through the reward function r t Calculate the reward value and select action a based on the reward value using the ε-greedy strategy t , according to action a t Assign channel prediction tasks.

5. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 4 is characterized in that: Through the designed temporal weighted attention mechanism, the method of outputting Q value of the dual deep Q network based on temporal weighted attention is: Estimate the delayed Q value through independent delay network and error network and error Q value Among them, θ l and θ i are the parameters of the delay network and the error network respectively; Dynamically adjust the delayed Q value through the designed temporal weighted attention mechanism and error Q value The weight of the historical context vector and the average reward signal Generate dynamic weights through the GRU module in, and are the average delay and average error respectively, and denote the weights of delay and error respectively; Dynamic weight and the context vector v of the current time slot t Multiply element by element to get the weighted context vector and with the delayed Q value and error Q value Combine the input multi-head attention mechanism to capture short-term and long-term dependency characteristics, then concatenate the output of the multi-head attention mechanism and generate a comprehensive Q value Q through a fully connected layer t .

6. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 1 or 5 is characterized in that: The cloud prediction model is composed of a multi-time scale attention layer, a task-specific branch layer and an output layer connected in sequence; the edge prediction model is composed of a multi-time scale attention layer and an output layer connected in sequence; the multi-time scale attention layer in the cloud prediction model and the edge prediction model is the same; The multi-time-scale attention layer is composed of a variable selection network, an LSTM encoder, a Gate and Add&Norm layer I, a gated residual network I, a multi-scale local self-attention network, a Gate and Add&Norm layer II, and a gated residual network II connected in sequence.

7. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 6 is characterized in that: The cloud prediction model and the edge prediction model both extract multi-time scale features from the labeled channel data through a multi-time scale attention layer, including the following steps: First, the labeled channel data is organized into a sliding window time series, which is then fed into the corresponding variable selection network. After being processed by the variable selection network, the weighted features of each feature vector in the sliding window time series are obtained, and then the weighted feature matrix of the sliding window time series is obtained. The weighted feature matrix of the sliding window time series is input into the LSTM encoder to extract the time-dependent features and obtain the hidden state sequence; Perform random inactivation on the hidden state sequence to generate a sparse state sequence and send it to the Gate and Add&Norm layer I to obtain a gated state sequence; The gated state sequence is further feature extracted through the gated residual network I to generate features that are input to the multi-scale local self-attention network. After being processed by the multi-scale local self-attention network, a multi-time scale feature matrix is ​​obtained. The multi-time-scale feature matrix is ​​processed by Gate and Add&Norm layer II and gated residual network II in sequence to obtain multi-time-scale features.

8. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 7 is characterized in that: In the edge prediction model, multi-time scale features are passed to the output layer, where they are mapped to channel state prediction results to obtain a channel state prediction sequence for the next T prediction time steps.

9. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 8 is characterized in that: In the cloud prediction model, multi-time scale features are passed to the task-specific branch layer to extract features for LoS and NLoS scenarios respectively; For the LoS scenario, the multi-time scale features are extracted in sequence through LSTM layer I, gated residual network III, and LSTM layer II to obtain long-term global features. O′ is a multi-time scale feature, k is the window length, d h is the hidden layer dimension, GRN is the gated residual network operation, and LSTM is the long short-term memory network operation; For the NLoS scenario, the multi-time scale features are extracted in turn through one-dimensional convolutional network I, gated residual network IV, and one-dimensional convolutional network II to obtain local features. Conv-1D is a one-dimensional convolution operation; The long-term global features are fused with the local features through the gating mechanism. The fusion process is as follows: G LLL =σ(W LLL H LLL +b LLL ) H iftiLt =G LLL ⊙H LLL +(1-G LLL )⊙H NLLL Among them, W LLL and b LLL are the weight and bias of the gate unit, G Los is the gate value, σ(·) is the Sigmoid activation function, H iftiLt is the fused feature matrix; The fused feature matrix H iftiLt is passed to the output layer to obtain the channel state prediction sequence of the next T prediction time steps 10. The cloud-edge collaborative industrial wireless channel prediction method based on multi-time scale attention mechanism according to claim 9 is characterized in that: The edge prediction model shares the parameters of the multi-time scale attention layer with the cloud prediction model through a shallow parameter sharing mechanism, and the edge prediction model updates the parameters of the multi-time scale attention layer through a shallow parameter update mechanism: (1) In the shallow parameter sharing mechanism, the multi-time scale attention layer parameters Θ shared by the cloud prediction model shared is offloaded to the edge server and replaces the parameter set of the multi-time scale attention layer in the edge prediction model. The parameters of the edge prediction model are expressed as: I EdEi-PP ={Θ thmiid ,I mdafttmali } Among them, Θ mdafttmali represents the specific parameters of the edge prediction model, and the shared multi-time scale attention layer parameters are frozen, requiring only training Θ mdafttmali Some are adapted to the specific task requirements of edge servers; (2) In the shallow parameter update mechanism, when the real-time data collected by the edge server reaches the set capacity, the data will be uploaded to the cloud server. The cloud server will fine-tune the multi-time scale attention layer parameters of the cloud prediction model and share them with the edge server.

Images