An obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method

Abstract

The present invention discloses an obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method, which relates to the field of wireless communication technology. Industrial terminal equipment collects channel state data in real time and uses a data preprocessing unit to preprocess the channel state data, and then transmits the preprocessed channel state data to an edge server. The edge server uses a long short-term memory network to detect and classify obstacles based on the preprocessed channel state data, generates obstacle category information, and uploads the obstacle category information and preprocessed channel state data to a cloud server. An obstacle-aware channel conversion architecture model is constructed, and the cloud server obtains channel state prediction results based on the received obstacle category information and preprocessed channel state data through the obstacle-aware channel conversion architecture model. The present invention has high accuracy, is more adaptable in complex environments, and has good reliability and scalability in practical applications.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to a high-precision channel state prediction method. Background Art

[0002] Industrial wireless networks (IWNs) are increasingly being used in automation, smart manufacturing, and real-time monitoring. These networks, with their flexibility and efficiency, have become the core infrastructure for connecting a wide range of industrial devices. However, the efficient operation of IWNs relies on accurate channel state prediction to cope with the frequently changing channel characteristics in dynamic environments. Therefore, accurately predicting channel state has become a pressing technical challenge.

[0003] In IWNs, channel states are affected by a variety of factors, including the rapid movement of equipment, the presence of obstacles in the environment, and multipath propagation. These factors make the dynamic changes in the channel complex and unpredictable, severely impacting the reliability and real-time performance of communications. Traditional channel prediction methods, such as deep learning-based convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), have demonstrated good performance in certain static or relatively simple environments. However, the effectiveness of these methods is often limited by the frequent changes in obstacles and dynamic movement of equipment in complex industrial environments, resulting in a significant decrease in prediction accuracy.

[0004] To address these issues, edge computing technology is gradually being introduced into IWNs. By distributing data processing tasks to edge servers close to the data source, edge computing aims to reduce data transmission latency and improve response speed. This architecture reduces the burden on servers, enabling more efficient real-time data processing and decision-making. However, effective implementation of edge computing relies on robust channel state prediction mechanisms to ensure real-time processing of large amounts of data from industrial equipment and rapid response. Furthermore, edge computing applications must consider how to effectively integrate data from diverse sources to improve prediction accuracy and reliability.

[0005] In order to overcome the problems of dynamic changes in wireless channels and task processing delays in existing industrial wireless networks, it is necessary to design a new channel prediction model that can effectively capture the dynamic changes in channel status and have real-time feedback and adaptive capabilities.

[0006] The invention with application number 202211281296.X discloses a method, apparatus, device and storage medium for predicting the channel state of a wireless network, including obtaining channel occupancy state sample data, constructing a network state time series prediction model based on a Transformer model, and training the network state time series prediction model using the channel occupancy state sample data, including: stacking multiple encoders and connecting them to a fully connected layer to construct a network state time series prediction model, calculating the error according to a loss function, training the network state time series prediction model using an optimization algorithm, iterating repeatedly until the loss function converges to obtain a trained network state time series prediction model, using the trained network state time series prediction model to predict the channel state of the wireless network to be detected, and obtaining a channel prediction result, and using the channel occupancy state sample data to train the network state time series prediction model to achieve channel state prediction in the wireless network. However, the above invention relies on the Transformer model for channel state prediction, lacks the ability to adapt to dynamic industrial environments, the centralized architecture leads to insufficient real-time performance, the feature extraction is single and does not combine multi-module optimization, and the model of the above invention lacks a dynamic update mechanism, making it difficult to cope with channel changes in complex scenarios. Summary of the Invention

[0007] In response to the technical problem that the dynamic changes of wireless channels in industrial wireless networks make it difficult for existing channel prediction methods to effectively capture the dynamic changes of channel states and the channel prediction accuracy is low, the present invention proposes an obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method, which realizes accurate prediction of channel states in complex and changeable industrial environments, greatly improves the communication stability and adaptability of industrial wireless networks, and provides a highly robust prediction solution for channel changes in dynamic industrial scenarios.

[0008] In order to achieve the above object, the technical solution of the present invention is achieved as follows:

[0009] An obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method. The cloud-edge-end collaborative architecture includes a cloud server, at least one edge server, and several industrial terminal devices. The edge server is connected to the cloud server via a high-speed network, and the industrial terminal devices are connected to the edge server via a wireless network.

[0010] The channel prediction method comprises the following steps:

[0011] Step 1: The industrial terminal device collects channel status data in real time and uses the data preprocessing unit to preprocess the channel status data, and transmits the preprocessed channel status data to the edge server;

[0012] Step 2: The edge server uses a long short-term memory network to detect and classify obstacles based on the preprocessed channel state data, generates obstacle category information, and uploads the obstacle category information and preprocessed channel state data to the cloud server.

[0013] Step 3: Build an obstacle-aware channel conversion architecture model. Based on the received obstacle category information and preprocessed channel state data, the cloud server obtains the channel state prediction result through the obstacle-aware channel conversion architecture model. The model parameters of the obstacle-aware channel conversion architecture model are dynamically updated using the feedback update mechanism to adapt to environmental changes.

[0014] Preferably, the channel state data includes signal-to-noise ratio, received signal strength indicator, reference signal received power, transmission time difference, noise power and location information;

[0015] At time slot t, the channel status data x collected by industrial terminal device m m (t) is:

[0016] x m (t) = [SNR m (t),RSSI m (t),RSRP m (t),TimeDifference m (t),NoisePower m (t),Location m (t)]

[0017] Among them, SNR m (t) is the signal-to-noise ratio from industrial terminal equipment m to industrial terminal equipment n at time slot t, RSSI m (t) is the received signal strength indicator of industrial terminal device m, RSRP m (t) is the reference signal receiving power of industrial terminal equipment m, TimeDifference m (t) is the transmission time difference of industrial terminal equipment m, NoisePower m (t) is the noise power of industrial terminal equipment m, Location m (t) is the location information of the industrial terminal device m.

[0018] Preferably, the data preprocessing unit includes a denoising module, a normalization module, a sliding window statistical analysis module and a threshold detection module; the preprocessing includes sequentially using the denoising module to perform denoising processing, using the normalization module to perform normalization processing, and using the sliding window statistical analysis module to calculate the channel state data x within a predetermined window size. mThe mean μ(t) and standard deviation σ(t) of (t) are obtained by using the threshold detection module based on the analysis results of the sliding window statistical analysis module. The upper and lower thresholds μ(t)±kσ(t) are set. Data points exceeding [μ(t)-kσ(t),μ(t)+kσ(t)] will be marked as abnormal and removed, which is expressed as:

[0019] x m ′(t)∈[μ(t)-kσ(t),μ(t)+kσ(t)]

[0020] Among them, x m ′(t) is the preprocessed channel state data, and k is the adjustment parameter.

[0021] Preferably, the edge server uses a long short-term memory network to perform obstacle detection and classification based on the pre-processed channel state data. The method for generating obstacle category information is as follows: the edge server converts the pre-processed channel state data x m ′(t) is transformed into the input sequence X m (t), input sequence X m (t) is sent to the LSTM layer through the input layer, and the LSTM layer outputs the hidden state h t , hidden state h t Is input into the fully connected layer to generate the classification result of the obstacle The classification results of obstacles are calculated by the Softmax function in the Softmax layer The probability of each category in the output layer outputs the obstacle category information c m (t).

[0022] Preferably, the input sequence X m (t) = [x m ′(tT),x m ′(t)], T is the input sequence X m (t) step size;

[0023] The classification results of the obstacles Obstacle classification results Contains P+1 categories, P is the number of obstacle types, category 0 means no obstacle, W out and b out are the weight and bias of the fully connected layer respectively, and Softmax is the Softmax function.

[0024] Preferably, the classification result of the obstacle The predicted probability of class p in Among them, W p is the weight corresponding to category p, W jis the weight matrix corresponding to category j, and the predicted probability of category p Provides obstacle category information c m (t).

[0025] Preferably, the obstacle category information c m (t) = f clf (x m ′(t)), f clf (·) is the obstacle detection and classification function.

[0026] Preferably, the structure of the obstacle-aware channel conversion architecture model specifically includes: a multi-size convolution module, a feature map attention module, a Transformer timing modeling module and a fully connected layer connected in sequence.

[0027] Preferably, the cloud server outputs the channel state prediction value of the future time slot based on the received obstacle category information and the pre-processed channel state data using the obstacle-aware channel conversion architecture model. The cloud server converts the received obstacle category information c m (t) and the preprocessed channel state data x m ′(t) is constructed as a global dataset And input to the obstacle-aware channel conversion architecture model, the global dataset Enter the multi-scale convolution module of the obstacle-aware channel conversion architecture model;

[0028] Global Dataset The multi-scale convolution module of the obstacle-aware channel conversion architecture model uses convolution kernels of various sizes to compare the global dataset. Perform one-dimensional convolution to obtain feature maps of different scales, then stack all the feature maps to output a multi-scale feature map F multi-scale ;

[0029] Multi-scale feature map F multi-scale is passed to the feature map attention module, which uses the attention mechanism to provide multi-scale feature maps F multi-scale The feature maps of different scales are weighted and the weighted multi-scale feature map F is output. weighted ;

[0030] Weighted multi-scale feature map F weighted It is input into the Transformer temporal modeling module, and the global temporal dependency is established through the multi-head self-attention mechanism of the Transformer temporal modeling module to generate a temporal feature map;

[0031] Before being passed to the fully connected layer, the time series feature map is flattened into a one-dimensional vector and finally passed to the fully connected layer, which maps these features into predicted values. Output the predicted channel state.

[0032] Preferably, the method for dynamically updating the model parameters of the obstacle-aware channel switching architecture model using a feedback update mechanism is:

[0033] ①Evaluate model performance: Collect actual channel state data over a period of time and the prediction results of the obstacle-aware channel conversion architecture model. For sample s, calculate the actual channel state y s (t) and the predicted channel state The error between s (t), based on the error e s (t) Calculate the loss function Loss Function Using mean square error:

[0034]

[0035]

[0036] Where N is the number of samples;

[0037] ②Update model parameters: Use Adam optimization algorithm according to loss function Update the model parameters θ, the update formula is:

[0038]

[0039] Where η is the learning rate, m(t) and υ(t) are the first-order moment and second-order moment estimates of the gradient of the loss function to the model parameters, and ∈ is a constant used to prevent division by zero errors;

[0040] After the obstacle-aware channel conversion architecture model is deployed, the cloud server periodically executes the above steps ① and ② to update the model parameters of the obstacle-aware channel conversion architecture model.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention introduces a cloud-edge-end collaborative architecture into the channel prediction system, uses industrial terminal devices to collect channel status data in real time, performs obstacle detection and classification in the edge server, and the cloud server uses the Obstacle-Aware Channel Transition Architecture (OCTA) model to accurately predict the channel status. By partially transferring the computing load to the edge server, the computing pressure of the cloud server is significantly reduced, and it is more adaptable in complex environments.

[0043] The Obstacle-Aware Channel Transformation Architecture (OCTA) model proposed in this paper uses multi-module feature fusion technology to comprehensively capture the multi-dimensional features in industrial wireless networks. First, a multi-scale convolution module is used to extract the global data set. Spatial features at different time scales are used to generate multi-scale feature maps. The feature map attention module then weights these multi-scale feature maps to adaptively retain the most critical information. Finally, the weighted multi-scale feature maps are fed into the Transformer temporal modeling module, where its multi-head self-attention mechanism establishes global temporal dependencies. Through this process, the Obstacle-Aware Channel Transformation Architecture (OCTA) model fully captures the multi-scale local features in channel data and the global temporal features over long time spans, significantly improving the real-time, accuracy, and robustness of channel prediction.

[0044] The present invention introduces a feedback update mechanism in the cloud server. In dynamic environments such as device movement and obstacle occlusion, the present invention demonstrates excellent stability and generalization capabilities, and has good reliability and scalability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 Schematic diagram of the cloud-edge-end collaborative industrial wireless network channel prediction architecture of the present invention.

[0047] Figure 2 Flowchart of the present invention.

[0048] Figure 3 Schematic diagram of the Obstacle-Aware Channel Transfer Architecture (OCTA) model of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0050] An obstacle-aware cloud-edge-device collaborative industrial wireless network channel prediction method, such as Figure 1As shown, the cloud-edge-end collaborative architecture includes a cloud server, at least one edge server, and several industrial terminal devices. The edge server is connected to the cloud server via a high-speed network, and the industrial terminal devices are connected to the edge server via a wireless network. The industrial terminal devices are used to collect channel status data in real time and upload it to the edge server; the edge server is used to process the channel status data and perform obstacle classification tasks; and the cloud server is used to receive data sent by the edge server to perform channel prediction tasks.

[0051] Specifically, the industrial terminal equipment includes distributedly deployed intelligent devices such as automatic guided vehicles (AGVs) and mobile robots.

[0052] like Figure 2 As shown, the channel prediction method includes the following steps:

[0053] Step 1: The industrial terminal equipment collects channel status data in real time and uses the data preprocessing unit to preprocess the channel status data, and transmits the preprocessed channel status data to the edge server.

[0054] Furthermore, the channel state data includes signal-to-noise ratio, received signal strength indicator, reference signal received power, transmission time difference, noise power and location information.

[0055] Specifically, at time slot t, the channel status data x collected by industrial terminal device m m (t) is:

[0056] x m (t) = [SNR m (t),RSSI m (t),RSRP m (t),TimeDifference m (t),NoisePower m (t),Location m (t)]

[0057] Among them, SNR m (t) is the signal-to-noise ratio from industrial terminal equipment m to industrial terminal equipment n at time slot t, RSSI m (t) is the received signal strength indicator of industrial terminal device m, RSRP m (t) is the reference signal receiving power of industrial terminal equipment m, TimeDifference m (t) is the transmission time difference of industrial terminal equipment m, NoisePower m (t) is the noise power of industrial terminal equipment m, Location m (t) is the location information of the industrial terminal device m.

[0058] The data preprocessing unit includes a denoising module, a normalization module, a sliding window statistical analysis module and a threshold detection module; the preprocessing includes sequentially using the denoising module to perform denoising processing, using the normalization module to perform normalization processing, and using the sliding window statistical analysis module to calculate the channel state data x within a predetermined window size. m The mean μ(t) and standard deviation σ(t) of the channel state data x m (t) Conduct analysis.

[0059]

[0060] Where W is the window size;

[0061] The threshold detection module is used based on the channel state data x m (t) is the mean μ(t) and standard deviation σ(t), and upper and lower thresholds μ(t)±kσ(t) are set. Data points exceeding [μ(t)-kσ(t), μ(t)+kσ(t)] will be marked as abnormal and removed, which can be expressed as:

[0062] x m ′(t)∈[μ(t)-kσ(t),μ(t)+kσ(t)]

[0063] Among them, x m ′(t) is the preprocessed channel state data, and k is the adjustment parameter based on the normal distribution.

[0064] Step 2: Based on the preprocessed channel state data, the edge server uses a long short-term memory network (LSTM) to perform obstacle detection and classification, generates obstacle category information, and uploads the obstacle category information and preprocessed channel state data to the cloud server.

[0065] Furthermore, the edge server detects and classifies obstacles in the industrial environment, and the pre-processed channel state data x m ′(t) is transmitted to the edge server at each time slot t, and the edge server is based on the pre-processed channel state data x m ′(t), the long short-term memory network (LSTM) is used to detect and classify obstacles. First, it is determined whether there are obstacles in the current communication environment. Then, the type of obstacles is further determined based on the detection results to obtain the obstacle category information c m (t).

[0066] Furthermore, the edge server sends the pre-processed channel state data x m ′(t) is transformed into the input sequence X m (t) = [x m ′(tT),xm ′(t)] is input to the long short-term memory network (LSTM), where T is the input sequence X m The step size of (t) is used to define the time span of the input channel state data; the input sequence X m (t) is sent to the LSTM layer through the input layer. After processing for multiple time steps, the LSTM layer outputs the hidden state h t , hidden state h t It is input into the fully connected layer (Dense layer) to generate the classification results of obstacles The classification results of obstacles are calculated by the Softmax function in the Softmax layer The probability of each category in the output layer outputs the obstacle category information c m (t).

[0067] The LSTM layer outputs the hidden state h t The process is as follows:

[0068] f t =σ(W f [h t-1 ,x t ]+b f ),i t =σ(W i [h t-1 ,x t ]+b i )

[0069] Among them, f t Represents the state of the forget gate, i t represents the state of the input gate, σ(·) is the Sigmoid activation function, W f is the weight matrix of the forget gate, W i is the weight matrix of the input gate, b f is the bias term of the forget gate, b i is the bias term of the input gate, h t-1 is the hidden state at the previous moment, x t is the current input data;

[0070] Memory unit C t The updated expression is:

[0071] C t =f t ⊙C t-1 +i t ⊙tanh(W C [h t-1 ,x t ]+b C )

[0072] Where ⊙ represents the element-wise multiplication operation (Hadamard product), tanh(·) is the hyperbolic tangent activation function, and W C is the weight matrix of the memory unit, b C is the bias term, C t is the memory unit state at the current moment;

[0073] Output gate o t Control hidden state h t The output expression is:

[0074] h t =o t ⊙tanh(C t )

[0075] Furthermore, the hidden state h t It is input into a fully connected layer (Dense layer) to generate the classification results of obstacles. The Softmax function is used to calculate the probability of each category in the Softmax layer, and the output is expressed as:

[0076]

[0077] Among them, the classification results of obstacles Contains P+1 categories, P is the number of obstacle types, category 0 means no obstacle, W out and b out are the weights and biases of the fully connected layer respectively, and Softmax is the Softmax function. The Softmax function is used to convert the output into a probability distribution, which is expressed as:

[0078]

[0079] in, is the predicted probability of category p, and the weight W corresponding to category p p and the model characteristics of the long short-term memory network, W j is the weight matrix corresponding to category j. The predicted probability of category p Provides obstacle category information c m (t), through the maximum probability principle, the category with the highest predicted probability is selected as the final classification result of time slot t. m (t)∈{0,1,2,...,P} is the obstacle category information of time slot t, 0 indicates no obstacle, and other categories p∈{1,2,...,P} indicate different types of obstacles.

[0080] Furthermore, the obstacle category information c m (t) is expressed as:

[0081] cm (t) = f clf (x m ′(t))

[0082] Among them, f clf (·) is the obstacle detection and classification function, which outputs the obstacle category information at time slot t. The obstacle category information and the preprocessed channel state data are then passed to the cloud server, which performs the channel prediction task.

[0083] Step 3: Build an Obstacle-Aware Channel Transfer Architecture (OCTA) model. Based on the received obstacle category information and pre-processed channel state data, the cloud server obtains the channel state prediction result through the OCTA model. The feedback update mechanism is used to dynamically update the model parameters of the OCTA model to adapt to environmental changes.

[0084] Furthermore, if Figure 3 As shown in FIG, the structure of the obstacle-aware channel conversion architecture model specifically includes: a multi-size convolution module, a feature map attention module, a Transformer timing modeling module and a fully connected layer connected in sequence.

[0085] Furthermore, the cloud server uses the obstacle-aware channel conversion architecture model to output the channel state prediction value of the future time slot as follows:

[0086] The cloud server receives the obstacle category information c m (t) and the preprocessed channel state data x m ′(t) is constructed as a global dataset And input into the obstacle-aware channel conversion architecture model.

[0087] Global Dataset The multi-scale convolution module of the obstacle-aware channel conversion architecture model uses convolution kernels of various sizes to compare the global dataset. Perform one-dimensional convolution to capture features at different time scales. The convolution operation of the convolution kernel with a size of n is represented as Conv1D n , then the output of the convolution kernel of size n is expressed as:

[0088]

[0089] where f n (t) is the feature extracted by the convolution kernel of size n, b n It is the bias of the convolution operation of the convolution kernel of size n. Through the convolution operation of convolution kernels of various sizes, feature maps of different scales are obtained. By stacking all the feature maps, a multi-scale feature map F is generated. multi-scale, ReLU(·) represents the activation function.

[0090] Multi-scale feature map F multi-scale is passed to the feature map attention module, which uses the attention mechanism to provide multi-scale feature maps F multi-scale The feature maps of different scales are given weights, and the feature map attention module assigns a weight α to each scale feature map n , and dynamically adjust the weight of each scale feature map according to the characteristics of the input data, expressed as:

[0091] F weighted =∑ n α n ·f n (t)

[0092] where α n is the weight of the feature map of size n, generated by the attention mechanism. Weighted multi-scale feature map F weighted The most critical information is retained, enabling the obstacle-aware channel conversion architecture model to more accurately model the dynamic changes of the channel.

[0093] Weighted multi-scale feature map F weighted It is input into the Transformer time series modeling module, and the global time dependency is established through the multi-head self-attention mechanism of the Transformer time series modeling module to generate a time series feature map. Before being passed to the fully connected layer, the time series feature map is flattened into a one-dimensional vector and finally passed to the fully connected layer, which maps these features into predicted values. That is, the channel state of the future time slot.

[0094] The method for dynamically updating the model parameters of the obstacle-aware channel conversion architecture model using the feedback update mechanism is:

[0095] ①Evaluate model performance: Collect actual channel state data and obstacle-aware channel conversion architecture model prediction results over a period of time. For sample s, calculate the actual channel state y s (t) and the predicted channel state The error between s (t), based on the error e s (t) Calculate the loss function Loss Function The mean square error (MSE) is used.

[0096]

[0097]

[0098] Where n is the number of samples;

[0099] ②Update model parameters: Use Adam optimization algorithm according to loss function Update the model parameters θ, the update formula is:

[0100]

[0101] Where η is the learning rate, m(t) and v(t) are the first-order moment and second-order moment estimates of the gradient of the loss function to the model parameters, respectively, and ∈ is a constant used to prevent division by zero errors.

[0102] After the obstacle-aware channel conversion architecture model is deployed, the cloud server periodically executes steps ① and ② above to fine-tune the model parameters of the obstacle-aware channel conversion architecture model to adapt to the dynamic changes in channel conditions in the industrial environment, thereby ensuring that the obstacle-aware channel conversion architecture model continues to maintain high prediction accuracy and adaptability.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method, characterized in that: The cloud-edge-end collaborative architecture includes a cloud server, at least one edge server, and several industrial terminal devices. The edge server is connected to the cloud server via a high-speed network, and the industrial terminal devices are connected to the edge server via a wireless network. The channel prediction method comprises the following steps: Step 1: The industrial terminal device collects channel status data in real time and uses the data preprocessing unit to preprocess the channel status data, and transmits the preprocessed channel status data to the edge server; Step 2: The edge server uses a long short-term memory network to detect and classify obstacles based on the preprocessed channel state data, generates obstacle category information, and uploads the obstacle category information and preprocessed channel state data to the cloud server. Step 3: Build an obstacle-aware channel conversion architecture model. Based on the received obstacle category information and preprocessed channel state data, the cloud server obtains the channel state prediction result through the obstacle-aware channel conversion architecture model. The model parameters of the obstacle-aware channel conversion architecture model are dynamically updated using the feedback update mechanism to adapt to environmental changes.

2. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 1 is characterized in that: The channel state data includes signal-to-noise ratio, received signal strength indicator, reference signal received power, transmission time difference, noise power and location information; At time slot t, the channel status data x collected by industrial terminal device m m (t) is: x m (t)=[SNR m (t),RSSI m (t),RSRP m (t),TimeDifference m (t),NoisePower m (t),Location m (t)] Among them, SNR m (t) is the signal-to-noise ratio from industrial terminal equipment m to industrial terminal equipment n at time slot t, RSSI m (t) is the received signal strength indicator of industrial terminal device m, RSRP m (t) is the reference signal receiving power of industrial terminal equipment m, TimeDifference m (t) is the transmission time difference of industrial terminal equipment m, NoisePower m (t) is the noise power of industrial terminal equipment m, Location m (t) is the location information of the industrial terminal device m.

3. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 1 or 2, characterized in that: The data preprocessing unit includes a denoising module, a normalization module, a sliding window statistical analysis module and a threshold detection module; the preprocessing includes sequentially using the denoising module to perform denoising processing, using the normalization module to perform normalization processing, and using the sliding window statistical analysis module to calculate the channel state data x within a predetermined window size. m The mean μ(t) and standard deviation σ(t) of (t) are obtained by using the threshold detection module based on the analysis results of the sliding window statistical analysis module. The upper and lower thresholds μ(t)±kσ(t) are set. Data points exceeding [μ(t)-kσ(t),μ(t)+kσ(t)] will be marked as abnormal and removed, which is expressed as: x m ′(t)∈[μ(t)-kσ(t),μ(t)+kσ(t)] Among them, x m ′(t) is the preprocessed channel state data, and k is the adjustment parameter.

4. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 3 is characterized in that: The edge server uses a long short-term memory network to perform obstacle detection and classification based on the pre-processed channel state data. The method for generating obstacle category information is as follows: the edge server converts the pre-processed channel state data x m ′(t) is transformed into the input sequence X m (t), input sequence X m (t) is sent to the LSTM layer through the input layer, and the LSTM layer outputs the hidden state h t , hidden state h t Is input into the fully connected layer to generate the classification result of the obstacle The classification results of obstacles are calculated by the Softmax function in the Softmax layer The probability of each category in the output layer outputs the obstacle category information c m (t).

5. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 4 is characterized in that: The input sequence X m (t) = [x m ′(tT),x m ′(t)], T is the input sequence X m (t) step size; The classification results of the obstacles Obstacle classification results Contains P+1 categories, P is the number of obstacle types, category 0 means no obstacle, W out and b out are the weight and bias of the fully connected layer respectively, and Softmax is the Softmax function.

6. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 5 is characterized in that: The classification results of the obstacles The predicted probability of class p in Among them, W p is the weight matrix corresponding to category p, W j is the weight matrix corresponding to category j, and the predicted probability of category p Provides obstacle category information c m (t).

7. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 6 is characterized in that: The obstacle category information c m (t) = f clf (x m ′(t)), f clf (·) is the obstacle detection and classification function.

8. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to any one of claims 1, 2, 4-7, characterized in that: The structure of the obstacle-aware channel conversion architecture model specifically includes: a multi-size convolution module, a feature map attention module, a Transformer timing modeling module and a fully connected layer connected in sequence.

9. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 8 is characterized in that: The cloud server outputs the channel state prediction value of the future time slot based on the received obstacle category information and the pre-processed channel state data using the obstacle-aware channel conversion architecture model. The cloud server converts the received obstacle category information c m (t) and the preprocessed channel state data x m ′(t) is constructed as a global dataset And input to the obstacle-aware channel conversion architecture model, the global dataset Enter the multi-scale convolution module of the obstacle-aware channel conversion architecture model; Global Dataset The multi-scale convolution module of the obstacle-aware channel conversion architecture model uses convolution kernels of various sizes to compare the global dataset. Perform one-dimensional convolution to obtain feature maps of different scales, then stack all the feature maps to output a multi-scale feature map F multi-scale ; Multi-scale feature map F multi-scale is passed to the feature map attention module, which uses the attention mechanism to provide multi-scale feature maps F multi-scale The feature maps of different scales are weighted and the weighted multi-scale feature map F is output. weighted ; Weighted multi-scale feature map F weighted It is input into the Transformer temporal modeling module, and the global temporal dependency is established through the multi-head self-attention mechanism of the Transformer temporal modeling module to generate a temporal feature map; Before being passed to the fully connected layer, the time series feature map is flattened into a one-dimensional vector and finally passed to the fully connected layer, which maps these features into predicted values. Output the predicted channel state.

10. The obstacle-aware cloud-edge-end collaborative industrial wireless network channel prediction method according to claim 1 or 9, characterized in that: The method for dynamically updating the model parameters of the obstacle-aware channel conversion architecture model using the feedback update mechanism is: ①Evaluate model performance: Collect actual channel state data over a period of time and the prediction results of the obstacle-aware channel conversion architecture model. For sample s, calculate the actual channel state y s (t) and the predicted channel state The error between s (t), based on the error e s (t) Calculate the loss function Loss Function Using mean square error: Where N is the number of samples; ②Update model parameters: Use Adam optimization algorithm according to loss function Update the model parameters θ, the update formula is: Where η is the learning rate, m(t) and v(t) are the first-order moment and second-order moment estimates of the loss function on the gradient of the model parameters, respectively. is a constant used to prevent division by zero errors; After the obstacle-aware channel conversion architecture model is deployed, the cloud server periodically executes the above steps ① and ② to update the model parameters of the obstacle-aware channel conversion architecture model.

Images