Data classification method based on self-attention mechanism in iscc system

Abstract

The application provides a data classification method based on a self-attention mechanism in an ISCC system, belongs to the field of perception communication and computing integration, and comprises the following steps: introducing attention weights by using the self-attention mechanism to measure the information amount of each ISAC device, fusing the local features of each ISAC device into global queries, taking the local features of each ISAC device as a key function and a value function, and using dot product to calculate the similarity to obtain the attention weights of the ISAC device; and determining whether the data of the device is transmitted to an edge server to participate in training of a classification model by judging whether the attention weight value reaches a set threshold value or not; the method improves the subsequent classification accuracy, reduces the calculation amount, avoids unnecessary waste of communication resources caused by transmission of redundant features, and balances the contradiction between the subsequent classifier and the number of feature selection. The method not only reduces the communication and calculation resource consumption but also improves the accuracy of the classification reasoning model.

Description

Technical Field

[0001] This invention belongs to the field of integrated perception, communication and computing, and specifically relates to a data classification method based on self-attention mechanism in the ISCC system. Background Technology

[0002] With the rapid development of technologies such as the Internet of Things (IoT), Artificial Intelligence (AI), and Big Data, the demand for computing power is experiencing explosive growth, giving rise to computing power networks as an important way to address this challenge. Computing power networks rely on high-speed, mobile, secure, and ubiquitous network connections, integrating multi-layered computing resources from cloud, edge, and endpoint to provide integrated services such as data sensing, transmission, storage, and computation. Meanwhile, with the surge in the number of terminal devices and the increasing complexity of application requirements, Integrated Sensing, Computation, and Communication (ISCC) is gradually becoming an important development direction for intelligent edge systems. The traditional "sensing-communication-computing" separation architecture, where data transmission, reception, and processing are all staged, is no longer sufficient to meet the system requirements for low latency, high throughput, and strong robustness. ISCC achieves efficient reuse of hardware resources by partially sharing radio frequency transceivers among sensing, computing, and communication, improving spectrum and bandwidth utilization and enhancing system real-time performance. However, the computing power of terminal devices is insufficient to meet the demands for low latency and high reliability. Introducing a computing network into the ISCC system can effectively overcome the computing power bottleneck of terminal devices. By flexibly allocating computing tasks among network nodes, the local computing burden can be reduced, and the efficiency of integrated sensing, communication and computing can be improved.

[0003] Existing ISCC systems typically consist of multiple ISCC (Integrated Sensing and Communication) devices with sensing and communication capabilities and a server located at the network edge. Each ISCC device is responsible for collecting data from the target object. Considering the limited computing power of the ISCC devices, the collected data is uploaded to the edge server, where it performs large-scale data processing and classification inference calculations.

[0004] Self-attention (SA) is a mechanism used to capture the relationships between different positions in a sequence. By allowing positions in the same sequence to act as both queries and keys and values, it enables global interaction between elements within the sequence. It can effectively capture the dependencies between positions in the sequence and is suitable for processing long sequence data.

[0005] Based on existing literature and research, the research content of ISCC mainly focuses on the following aspects:

[0006] 1) ISCC Joint Waveform and Signal Design: For ISCC devices that simultaneously perform sensing and communication functions, this study investigates how to design joint signal waveforms for high-resolution sensing and high-spectral-efficiency communication, including multi-carrier, phased array radar waveforms, and adaptive orthogonal frequency division multiplexing, in order to achieve synergistic optimization of sensing and communication performance.

[0007] 2) Network resource allocation in the ISCC system: This mainly involves coordinating the needs of the sensing task for high resolution and low false detection rate with the communication task for high throughput and low latency, given limited spectrum, time slots, and power resources. Existing research typically constructs a joint optimization model that incorporates sensing accuracy, communication latency, and energy consumption targets into the same constraint, and dynamically allocates network resources through convex optimization algorithms or deep reinforcement learning.

[0008] 3) Distributed collaborative perception and edge inference: To address the collaboration requirements between distributed ISAC device nodes, we explore multi-node data fusion and distributed inference algorithms, including federated learning, graph neural networks and collaborative filtering, to improve the perception accuracy and robustness of the entire system while reducing dependence on the central server.

[0009] However, in actual ISAC device deployment scenarios, due to various factors such as environmental changes, device noise, and installation location, the data collected by ISAC devices in different locations often exhibits certain differences and redundancy. In traditional data processing workflows, data collected from all devices is often simply spliced ​​or averaged as input data for subsequent modeling or analysis. Furthermore, as the scale of ISAC device deployment continues to expand, massive amounts of data can easily lead to low training efficiency and poor generalization ability of edge server models, and even introduce noise due to the introduction of invalid features, affecting the final inference results. Noise or irrelevant data generated by environmental interference or aging of some ISAC devices exacerbates the communication burden and affects the inference accuracy of the model; therefore, it is crucial to effectively select the most important features for the classification inference model from multiple ISAC devices to improve model performance. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this application proposes a data classification method based on a self-attention mechanism in the ISCC system. An attention weight is introduced through the self-attention mechanism to measure the information content of each ISCC device, thereby selecting data to participate in the training of the classification inference model. Since a single ISCC device can only collect data within a certain radius or viewpoint, each ISCC extracts local feature information. To obtain global features that can be used as a global query to estimate the importance of each ISCC, this invention fuses the local features of each ISCC device into a global query. The local features of each ISCC device are used as key and value functions, and similarity is calculated using dot products to obtain the attention weight of each ISCC device. Whether the data of a device is transmitted to the server for classification model training is determined by whether the attention weight value reaches a set threshold. If the attention weight exceeds the set threshold, the feature data of that ISCC device participates in model training; if it is less than the weight threshold, the data of that ISCC device is discarded.

[0011] In a first aspect, the present invention provides a data classification method based on a self-attention mechanism in an ISCC system, applied to the device side, including:

[0012] K ISAC devices are deployed at different locations. The ISAC devices collect sensing data. The collected sensing data is sampled and divided into T fixed-length sequences with non-overlapping time intervals to obtain the local feature set of each ISAC device.

[0013] The ISAC device receives the query function sent by the edge server;

[0014] Each ISAC device calculates a key function based on its local feature set, and then calculates the attention weight of each ISAC device using an attention function based on the key function and the query function.

[0015] If the attention weight of the ISAC device is lower than the preset threshold, the local feature set of the ISAC device is discarded; if the attention weight of the ISAC device is higher than or equal to the preset threshold, the local feature set of the ISAC device is selected, and the value function of each ISAC device is calculated based on the selected local feature set of the ISAC device, and the weight matrix of the value function is randomly initialized.

[0016] Weighted features are calculated for each ISAC device based on its value function and attention weight.

[0017] The ISAC device sends the weighted features to a designated channel, performs over-the-air computation on the designated channel, fuses the weighted features of the selected ISAC device to obtain the fused weighted features, and sends the fused weighted features to the edge server through the designated channel.

[0018] The attention weight of each ISAC device is calculated using an attention function based on the key function and the query function, as shown in the following formula:

[0019] ;

[0020] in, Let be the attention function. As a global feature, For the first The first ISAC device Local feature data for each time interval, For the softmax function, For key functions, For the first Query functions for each ISAC device It is the dimension of the mapping space. For the first Attention weights for each ISAC device.

[0021] The key function is calculated as follows:

[0022] ;

[0023] in, For the first The first ISAC device Key functions for time intervals, It is the weight matrix of the keys. For the first The first ISAC device Local feature data for each time interval.

[0024] The value function for each ISAC device is calculated based on the local feature set of the selected ISAC devices, and the calculation formula is as follows:

[0025] ;

[0026] in, For the first The first ISAC device A function of values ​​for each time interval. It is a weight matrix of values. For the first The first ISAC device Local feature data for each time interval.

[0027] The weighted features are calculated in the ISAC device, and the calculation formula is as follows:

[0028] ;

[0029] in, For the first The first ISAC device Weighted features of time intervals, For the first Attention weights of each ISAC device For the first The first ISAC device A function that calculates the value of a time interval.

[0030] The fusion weighted feature is calculated as follows:

[0031] ;

[0032] in, For the first The fusion weighted features of each time interval, For the first The first ISAC device Weighted features of time intervals, For the selected set of ISAC devices.

[0033] Secondly, this invention provides a data classification method based on a self-attention mechanism in the ISCC system, applied to an edge server, including:

[0034] The edge server receives the local feature set of each ISAC device, fuses the local feature sets of each ISAC device through 1×1 convolution, and constructs global features based on the convolutionally fused data;

[0035] The edge server calculates the query function based on global features and sends the query function to each ISAC device;

[0036] The edge server receives the fused weighted features. In the edge server, the fused weighted features are combined with the global query function through residual connections to obtain global fused features. The global fused features are input into a one-dimensional convolutional neural network classification model to classify the sensing data collected by the ISAC device and obtain the classification result. The one-dimensional convolutional neural network updates the network parameters using the gradient descent method. The classification result is used to determine the status of the current monitoring target.

[0037] The global feature is calculated as follows:

[0038] ;

[0039] in, As a global feature, For the first A 1×1 convolution kernel for each ISAC device, For the first The first ISAC device Local feature data for each time interval, K represents the length of the layout feature data, and K represents the total number of ISAC devices.

[0040] The query function is calculated as follows:

[0041] ;

[0042] in, For the first Query functions for each ISAC device For querying the matrix, This is a global feature.

[0043] The global fusion feature is calculated as follows:

[0044] ;

[0045] in, For global fusion features, This is a global query function for all ISAC devices. For the first The fusion weighted features of each time interval.

[0046] The one-dimensional convolutional neural network updates its parameters using gradient descent. The update process includes:

[0047] Step S100: Initialize network parameters, including: setting the maximum number of iterations. The current iteration number =1, learning rate convolution kernel The weight matrix of the query function The weight matrix of the key function and the weight matrix of the value function ;

[0048] Step S101: Calculate the attention weight for each ISAC device. : ,in, As a global feature, For the first The first ISAC device Local feature data for each time interval, For the softmax function, For key functions, For the first Query functions for each ISAC device It is the dimension of the mapping space. For the first Attention weights for each ISAC device;

[0049] Step S102: Calculate the loss function : ,in, For batch size, For the first The true value of each sample For the first Predicted values ​​for each sample;

[0050] Step S103: Update the convolution kernel: ,in, The updated convolutional kernel;

[0051] Step S104: Update the weight matrix of the query function, the weight matrix of the key function, and the weight matrix of the value function: , , ,in, The weight matrix for the updated query function. The weight matrix of the updated key function. The updated value function weight matrix;

[0052] Step S105: Increment the current iteration count by 1. Greater than or equal to the maximum number of iterations or loss function When the value converges, the iteration ends and the final convolution kernel is output. Weight matrix of query function The weight matrix of the key function The weight matrix of the value function and attention weight Otherwise, proceed to step S101 to calculate the attention weight of the next ISAC device and iterate again.

[0053] Beneficial effects:

[0054] This application proposes a data classification method based on self-attention mechanism in the ISCC system. Compared with the prior art, this application reduces the amount of data involved in model training by adding attention weights to filter ISCC devices containing important information. This not only reduces communication and computing resource overhead but also improves the accuracy of the classification inference model. Attached Figure Description

[0055] Figure 1 Flowchart of the data classification method based on self-attention mechanism in the ISCC system of this invention;

[0056] Figure 2 A schematic diagram of the data classification method based on self-attention mechanism in the ISCC system of this invention;

[0057] Figure 3 Comparison of classification model verification accuracy under different thresholds in embodiments of the present invention;

[0058] Figure 4 A comparison chart of the test accuracy of the classification model under different thresholds in this embodiment of the invention. Detailed Implementation

[0059] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0060] Example 1:

[0061] This embodiment provides a data classification method based on the self-attention mechanism in the ISCC system, such as... Figure 1 , Figure 2 As shown, it includes:

[0062] Step S1: Deploy K ISAC devices at different locations. The ISAC devices collect sensing data. The collected sensing data is sampled and divided into T fixed-length sequences with non-overlapping time intervals to obtain the local feature set of each ISAC device.

[0063] In this embodiment, K ISAC devices deployed at different locations perform sensing and communication functions through a time-division model. After collecting sensing data, the ISAC devices sample the data using a uniform sampling rate, and the sampled data segments are divided into... A fixed-length sequence of non-overlapping time intervals. For the first The first device Local feature data for each time interval, where For data length, Time interval ;

[0064] In this embodiment, the ISAC device samples rolling bearing data, the sampling state is 20Hz-0V, the number of ISAC devices K is 8, and the sample size is... The value is 2056.

[0065] Step S2: The edge server receives the local feature set of each ISAC device, fuses the local feature sets of each ISAC device through 1×1 convolution, and constructs global features based on the convolutionally fused data; the edge server calculates the query function based on the global features and sends the query function to each ISAC device; the ISAC device receives the query function sent by the edge server.

[0066] In this embodiment, based on each ISAC device Local feature set at time step ,pass Convolutional processing fuses local information from different ISAC devices, which is then transmitted to an edge server to construct global features. ,in, As a global feature, For the first A 1×1 convolution kernel for each ISAC device, For the first The first ISAC device Local feature data for each time interval, Where K is the length of the layout feature data, and K is the total number of ISAC devices. Based on the obtained global features... Execute the query function on it The transformed data is then broadcast to each ISAC device.

[0067] The query function is calculated as follows:

[0068] ;

[0069] in, For the first Query functions for each ISAC device For querying the matrix, This is a global feature.

[0070] Step S3: Each ISAC device calculates the key function based on its local feature set, and calculates the attention weight of each ISAC device using the attention function based on the key function and the query function;

[0071] The key function is calculated as follows:

[0072] ;

[0073] in, For the first The first ISAC device Key functions for time intervals, It is the weight matrix of the keys. For the first The first ISAC device Local feature data for each time interval.

[0074] In this embodiment, the local features of each device are processed using key functions. The transformation is then performed, and the attention weights for each device are calculated using the attention function. :

[0075] ;

[0076] in, Let be the attention function. As a global feature, For the first The first ISAC device Local feature data for each time interval, For the softmax function, For key functions, For the first Query functions for each ISAC device It is the dimension of the mapping space, and its value is fixed at 64. For the first Attention weights for each ISAC device.

[0077] Step S4: If the attention weight of the ISAC device is lower than the preset threshold, discard the local feature set of the ISAC device; if the attention weight of the ISAC device is higher than or equal to the preset threshold, select the local feature set of the ISAC device, calculate the value function of each ISAC device based on the selected local feature set of the ISAC device, and randomly initialize the weight matrix of the value function.

[0078] In this embodiment, a weighted threshold is introduced. Select the ISAC device and discard the weights. Below The data collected by the devices, the selected set of ISAC devices is ,in The local features of the selected ISAC device. Value function transformation is performed to obtain ,in , This is the weight matrix of the value function.

[0079] The value function for each ISAC device is calculated based on the local feature set of the selected ISAC devices, and the calculation formula is as follows:

[0080] ;

[0081] in, For the first The first ISAC device A function of values ​​for each time interval. It is a weight matrix of values. For the first The first ISAC device Local feature data for each time interval.

[0082] Step S5: Calculate the weighted features in each ISAC device based on the value function and attention weight of each ISAC device;

[0083] Step S6: The ISAC device sends the weighted features to the designated channel, performs over-the-air calculations in the designated channel, fuses the weighted features of the selected ISAC device to obtain the fused weighted features, and sends the fused weighted features to the edge server through the designated channel.

[0084] Step S7: The edge server receives the fused weighted features. In the edge server, the fused weighted features are combined with the global query function through residual connections to obtain global fused features. The global fused features are input into a one-dimensional convolutional neural network classification model to classify the sensing data collected by the ISAC device and obtain the classification result. The one-dimensional convolutional neural network updates the network parameters using the gradient descent method. The classification result is used to determine the status of the current monitoring target.

[0085] In this embodiment, the local features of different ISAC devices are used. With the corresponding attention weights Multiplication yields the weighted features:

[0086] ;

[0087] in, For the first The first ISAC device Weighted features of time intervals, For the first Attention weights of each ISAC device For the first The first ISAC device A function that calculates the value of a time interval.

[0088] Then set The weighted features of the ISAC devices in the dataset are summed to obtain the fused weighted features:

[0089] ;

[0090] in, For the first The fusion weighted features of each time interval, For the first The first ISAC device Weighted features of time intervals, For the selected set of ISAC devices.

[0091] Finally, the fused weighted features are combined with the global query through residual connections. Combined to obtain global fusion features:

[0092] ;

[0093] in, For global fusion features, This is a global query function for all ISAC devices. For the first The fusion weighted features of each time interval.

[0094] Based on the obtained global fusion features A one-dimensional convolutional neural network classification model is then built on the edge server. The model consists of two 3×3 convolutional layers, two 2×2 max pooling layers, one fully connected layer, and one fully connected output layer, with cross-entropy as the loss function. ,in It refers to the batch size. For the first The true value of each sample For the first The predicted value for each sample. The model uses gradient descent to update the parameters.

[0095] The one-dimensional convolutional neural network updates its parameters using gradient descent. The update process includes:

[0096] Step S100: Initialize network parameters, including: setting the maximum number of iterations. The current iteration number =1, learning rate convolution kernel The weight matrix of the query function The weight matrix of the key function and the weight matrix of the value function ;

[0097] Step S101: Calculate the attention weight for each ISAC device. : ,in, As a global feature, For the first The first ISAC device Local feature data for each time interval, For the softmax function, For key functions, For the first Query functions for each ISAC device It is the dimension of the mapping space. For the first Attention weights for each ISAC device;

[0098] Step S102: Calculate the loss function : ,in, For batch size, For the first The true value of each sample For the first Predicted values ​​for each sample;

[0099] Step S103: Update the convolution kernel: ,in, The updated convolutional kernel;

[0100] Step S104: Update the weight matrix of the query function, the weight matrix of the key function, and the weight matrix of the value function: , , ,in, The weight matrix for the updated query function. The weight matrix of the updated key function. The updated value function weight matrix;

[0101] Step S105: Increment the current iteration count by 1. Greater than or equal to the maximum number of iterations or loss function When the value converges, the iteration ends and the final convolution kernel is output. Weight matrix of query function The weight matrix of the key function The weight matrix of the value function and attention weight Otherwise, proceed to step S101 to calculate the attention weight of the next ISAC device and iterate again.

[0102] The classification model categorizes bearing faults based on data collected by ISAC devices to determine the current status of the monitored target. Through result analysis and extensive experimental comparisons, model convergence can be further accelerated and classification accuracy improved. This reduces the number of ISAC devices involved in model training, thereby lowering communication overhead and increasing the convergence speed and accuracy of the classification model. This enables timely monitoring of bearing status and prompt identification of fault types.

[0103] like Figure 3 , Figure 4 As shown, simulation experiments were conducted using real data to demonstrate the proposed method. Based on the ISCC system model, a bearing dataset collected by a university at eight locations on a gearbox was simulated as data collected by multiple ISCC devices at different locations. The edge server used a one-dimensional convolutional neural network model for inference computation. For ease of analysis, the system did not consider the effects of channel fading and noise. The current method filters ISCC devices by configuring different weight thresholds and transmits data from devices meeting the weight requirements to the server for model training and inference. This aims to illustrate the impact of the number of participating ISCC devices on the model validation accuracy and test accuracy. Setting a smaller weight threshold indicates that more ISCC device data can be selected to participate in the server-side inference computation. Analysis of the experimental results shows that the optimal performance of the inference model does not require all ISCC device data to participate in model training.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

[0105] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0106] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of equivalent technology of this disclosure, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A data classification method based on self-attention mechanism in the ISCC system, applied to the device side, characterized in that, include: K ISAC devices are deployed at different locations. The ISAC devices collect sensing data. The collected sensing data is sampled and divided into T fixed-length sequences with non-overlapping time intervals to obtain the local feature set of each ISAC device. The ISAC device receives the query function sent by the edge server; Each ISAC device calculates a key function based on its local feature set, and then calculates the attention weight of each ISAC device using an attention function based on the key function and the query function. If the attention weight of the ISAC device is lower than the preset threshold, the local feature set of the ISAC device is discarded; if the attention weight of the ISAC device is higher than or equal to the preset threshold, the local feature set of the ISAC device is selected, and the value function of each ISAC device is calculated based on the selected local feature set of the ISAC device, and the weight matrix of the value function is randomly initialized. Weighted features are calculated for each ISAC device based on its value function and attention weight. The ISAC device sends the weighted features to a designated channel, performs over-the-air computation in the designated channel, fuses the weighted features of the selected ISAC device to obtain the fused weighted features, and sends the fused weighted features to the edge server through the designated channel. The attention weight of each ISAC device is calculated using an attention function based on the key function and the query function, as shown in the following formula: ； in, Let be the attention function. As a global feature, For the first The first ISAC device Local feature data for each time interval, For the softmax function, For key functions, For the first Query functions for each ISAC device It is the dimension of the mapping space. For the first Attention weights for each ISAC device; The key function is calculated as follows: ； in, For the first The first ISAC device Key functions for time intervals, It is the weight matrix of the keys. For the first The first ISAC device Local feature data for each time interval; The value function for each ISAC device is calculated based on the local feature set of the selected ISAC devices, and the calculation formula is as follows: ； in, For the first The first ISAC device A function of values ​​for each time interval. It is a weight matrix of values. For the first The first ISAC device Local feature data for each time interval.

2. The data classification method based on self-attention mechanism in the ISCC system according to claim 1, characterized in that, The weighted features are calculated in the ISAC device, and the calculation formula is as follows: ； in, For the first The first ISAC device Weighted features of time intervals, For the first Attention weights of each ISAC device For the first The first ISAC device A function that calculates the value of a time interval.

3. The data classification method based on self-attention mechanism in the ISCC system according to claim 1, characterized in that, The fusion weighted feature is calculated as follows: ； in, For the first The fusion weighted features of each time interval, For the first The first ISAC device Weighted features of time intervals, For the selected set of ISAC devices.

4. A data classification method based on self-attention mechanism in the ISCC system, applied to the edge server, is characterized by: include: The edge server receives the local feature set of each ISAC device, fuses the local feature sets of each ISAC device through 1×1 convolution, and constructs global features based on the convolutionally fused data; The edge server calculates the query function based on global features and sends the query function to each ISAC device; The edge server receives the fused weighted features. In the edge server, the fused weighted features are combined with the global query function through residual connections to obtain global fused features. The global fused features are input into a one-dimensional convolutional neural network classification model to classify the sensing data collected by the ISAC device and obtain the classification result. The one-dimensional convolutional neural network updates the network parameters using the gradient descent method. The classification result is used to determine the status of the current monitoring target. The one-dimensional convolutional neural network updates its parameters using gradient descent. The update process includes: Step S100: Initialize network parameters, including: setting the maximum number of iterations. The current iteration number =1, learning rate convolution kernel The weight matrix of the query function The weight matrix of the key function and the weight matrix of the value function ; Step S101: Calculate the attention weight for each ISAC device. : ,in, As a global feature, For the first The first ISAC device Local feature data for each time interval, For the softmax function, For key functions, For the first Query functions for each ISAC device It is the dimension of the mapping space. For the first Attention weights for each ISAC device; Step S102: Calculate the loss function : ,in, For batch size, For the first The true value of each sample For the first Predicted values ​​for each sample; Step S103: Update the convolution kernel: ,in, The updated convolutional kernel; Step S104: Update the weight matrix of the query function, the weight matrix of the key function, and the weight matrix of the value function: , , ,in, The weight matrix for the updated query function. The weight matrix of the updated key function. The updated value function weight matrix; Step S105: Increment the current iteration count by 1. Greater than or equal to the maximum number of iterations or loss function When the value converges, the iteration ends and the final convolution kernel is output. Weight matrix of query function The weight matrix of the key function The weight matrix of the value function and attention weight Otherwise, proceed to step S101 to calculate the attention weight of the next ISAC device and iterate again.

5. The data classification method based on self-attention mechanism in the ISCC system according to claim 4, characterized in that, The global feature is calculated as follows: ； in, As a global feature, For the first A 1×1 convolution kernel for each ISAC device, For the first The first ISAC device Local feature data for each time interval, K represents the length of the layout feature data, and K represents the total number of ISAC devices.

6. The data classification method based on self-attention mechanism in the ISCC system according to claim 4, characterized in that, The global fusion feature is calculated as follows: ； in, For global fusion features, This is a global query function for all ISAC devices. For the first The fusion weighted features of each time interval.

Images