A C-V2X communication auxiliary system based on a self-attention mechanism residual composite network

Abstract

This invention belongs to the field of vehicle-to-everything (V2X) communication assistance system, specifically relating to a C-V2X communication assistance system based on a self-attention mechanism residual composite network. Deployed in an onboard unit (OBU), the system includes a traffic condition acquisition module, a channel detection module, a congestion prediction module, a congestion determination module, and a data offloading module. The traffic condition acquisition module acquires vehicle condition data; the channel detection module acquires channel information data from ahead; the vehicle condition information and channel information data are input into the congestion prediction module, where a neural network calculates and outputs the calculated parameters to the congestion determination module. The congestion determination module determines whether to offload the data based on the parameter determination results. This system enables the prediction of communication quality for the road segment ahead in the user's lane and, based on the prediction results, the selection of whether to offload communication data, achieving the function of early prediction and real-time assistance for communication quality.

Description

Technical Field

[0001] This invention belongs to the field of vehicle networking, and specifically relates to a C-v2X communication assistance system. Background Technology

[0002] Before C-v2X technology is widely deployed at the vehicle and roadside levels, channel congestion often occurs in large-scale test scenarios, simulating congestion caused by factors such as vehicle traffic. When the channel capacity reaches a pre-set congestion window threshold, congestion control algorithms alleviate congestion to some extent. Commonly used congestion control algorithms include four processes: slow start, congestion avoidance, congestion occurrence (fast retransmission), and fast recovery. These methods focus on taking appropriate measures to alleviate congestion after it occurs.

[0003] Regarding the above, the shortcomings of the existing technology are as follows:

[0004] 1. Current measures to address channel congestion involve activating different congestion control algorithms to reduce the channel busy rate when it reaches a set threshold. However, these methods typically only work when the channel is already at a certain level of congestion. In congested scenarios such as intersections with heavy traffic, switching to a C-v2X communication network after a certain degree of congestion may not effectively improve communication quality.

[0005] 2. Currently, most C-v2X communication is based on cellular base stations. When traffic is congested, the number of communication messages may exceed the channel capacity of the on-board unit (OBU) or roadside unit (RSU), causing packet loss and other issues that may endanger the personal safety of vehicle users. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a C-V2X communication assistance system based on a self-attention mechanism residual composite network. This system predicts the communication quality of the road segment ahead of its lane and selects whether to divert communication data based on the prediction results, thereby achieving the function of predicting communication quality in advance and providing real-time assistance.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows.

[0008] A C-V2X communication assistance system based on a self-attention mechanism residual composite network, the system is deployed on an on-board OBU unit, characterized in that the system includes a traffic condition acquisition module, a channel detection module, a congestion prediction module, a congestion determination module, and a data splitting module;

[0009] The traffic condition acquisition module acquires vehicle condition data; the channel detection module acquires channel information data ahead; the vehicle condition information and channel information data are input into the congestion prediction module, the neural network inside the congestion prediction module calculates and outputs the calculated parameters to the congestion determination module, and the congestion determination module selects whether to divert the data based on the parameter determination results;

[0010] The prediction function in the congestion prediction module is implemented by a residual composite network consisting of an LSTM long short-term recurrent neural network and an encoder-decoder based on a self-attention mechanism.

[0011] The system can be configured by deploying it in the vehicle's OBU unit to predict the communication quality of the road segment ahead of the lane, and select whether to divert the communication data based on the prediction results, thereby achieving the function of predicting communication quality in advance and providing real-time assistance.

[0012] Furthermore, the system also includes a terminal display module, which displays the congestion determination results from the congestion determination module.

[0013] Furthermore, the road condition acquisition module obtains vehicle condition data as the average speed v of the road segment.

[0014] Furthermore, the specific steps for obtaining the average speed v of a road segment are as follows: The road condition acquisition module reads the driving route from the vehicle navigation system, requests the vehicle condition information of the corresponding road segment from the roadside units (RSUs) ahead of the driving route, including the number of vehicles m and the average speed vi per vehicle; and calculates the corresponding average speed v of the road segment using m and vi, as follows:

[0015]

[0016] Where m is the number of vehicles on the corresponding road segment.

[0017] Furthermore, the channel detection module acquires channel information data from the front, including channel busy rate c, packet loss rate p, and air interface delay l; and detects the network quality between base stations and between base stations and satellite signals. The detection results are sorted in descending order of value to construct a corresponding ordered table of C-V2X network quality.

[0018] Furthermore, the congestion prediction module integrates and preprocesses the data output from the traffic condition acquisition module and the channel detection module, using it as input parameters for the neural network in the congestion prediction module, and outputs the corresponding results.

[0019] Furthermore, the congestion prediction module integrates and preprocesses the data output from the traffic condition acquisition module and the channel detection module, mainly including the following steps:

[0020] Step 1: Parameter initialization;

[0021] Step 2: Data partitioning;

[0022] Step 3: Congestion prediction.

[0023] Furthermore, in step 1, the average speed v of the road segment, the channel busy rate c, the packet loss rate p, and the air interface delay l are normalized using the following formula:

[0024]

[0025] Where, x norm The results are the normalized values ​​of the average speed v of the road segment, the channel busy rate c, the packet loss rate p, and the air interface delay l.

[0026] Furthermore, in step 2, the calculation results in step 1 are divided into different data truncations X of length L to adapt to the feature extraction mode of the hidden layer of the recurrent neural network.

[0027] Furthermore, in step 3, the segmented feature data X containing vehicle condition and channel information is... t (The data of X at a certain time t is X) t As the original input to the residual composite network, the trend vector is extracted through iterative processing by a separate Long Short-Term Recurrent Neural Network (LSTM).

[0028]

[0029] X' t =LSTM(X) t );

[0030]

[0031] Subtract the trend value from the original value to obtain the residual δ;

[0032] δ=X' t -X t ;

[0033] δ is used as the input to the encoder network and the background vector containing δ feature information is encoded and extracted. The encoder uses a randomized LSTM long short-term recurrent neural network.

[0034] C = [c t ,h t ] = Encoder(δ);

[0035] Among them, c t h is the cell state vector saved by the encoder during the iteration process. t The hidden state vector is saved by the encoder during the iteration process, and the two are concatenated to form the background vector C.

[0036] In the self-attention sublayer SA of the decoder, the background vector C is interleaved with a linear matrix. The dot product is mapped to the query matrix Q, the key matrix K, and the value matrix V. The self-attention relationship between Q, K, and V is then calculated.

[0037]

[0038] Wherein, s(k j ,q i The score for the correlation fraction equation is:

[0039]

[0040] The decoder-side LSTM unit is initialized with a random matrix zi containing self-attention relationships between road segment, lane, and channel information, and the residual δ is used as the decoder-side input. The input result δ' after constraining the differences in the residuals through the decoder-side network is:

[0041] δ'=Decoder(z i ,δ);

[0042] The constrained residual δ' is superimposed with the predicted trend σ to obtain the final prediction result σ':

[0043]

[0044] The beneficial effect of this invention is that by deploying it in the vehicle-mounted OBU unit, the communication quality of the road segment ahead of the lane can be predicted, and the communication data can be diverted according to the prediction results, thereby achieving the function of predicting the communication quality in advance and providing real-time assistance. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of a residual composite network.

[0046] Figure 2 This is a diagram of the system's prediction process. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0048] See Figure 1-2This embodiment provides a C-V2X communication assistance system based on a self-attention mechanism residual composite network. The system comprises six modules: a traffic condition acquisition module, a channel detection module, a congestion prediction module, a congestion determination module, a data splitting module, and a terminal display module. Vehicle condition and channel information data are acquired from the traffic condition acquisition and channel detection modules, respectively, and input into the congestion prediction module. The congestion prediction module calculates and outputs corresponding parameters based on its internal neural network, which are then sent to the congestion determination module. Based on the determination result, the module selects whether to split the data. The terminal display module provides visualization functionality for the above results.

[0049] The specific functions of each module are as follows:

[0050] 1. Traffic Condition Data Acquisition Module:

[0051] Based on the vehicle navigation system, the system reads the vehicle's route and requests traffic information for the corresponding road segment from the roadside units (RSUs) ahead of that route, including the number of vehicles (m) and the average speed per vehicle (vi). The average speed (v) for the corresponding road segment is then calculated using m and vi, as follows:

[0052]

[0053] Where m is the number of vehicles on the corresponding road segment.

[0054] 2. Channel detection module:

[0055] The system detects the channel busy rate (c), packet loss rate (p), and air interface latency (l) of the road segment ahead along the driving route by transmitting data between the vehicle-mounted OBU and the roadside RSU unit. It also detects the network quality between base stations and between base stations and satellite signals. The detection results are sorted in descending order of numerical value to construct a corresponding ordered C-V2X network quality table, which is constructed based on the descending order of the detection results.

[0056] The detection method involves calculating the Received Signal Strength Index (RSSI) between the onboard OBU unit and signals from different base stations and low-Earth orbit satellites.

[0057] RSSI=Tx_Power+Path_Loss+Rx_Gain+System_Gain

[0058] Where Tx_Power is the signal transmit power, Path_Loss is the path loss, Rx_Gain is the receive antenna gain, and System_Gain is the system gain.

[0059] 3. Congestion Prediction Module

[0060] The congestion prediction module integrates and preprocesses the data output from the traffic condition acquisition module and the channel detection module, using this data as input parameters for the neural network within the congestion prediction module, and outputs the corresponding results. Specifically, this neural network is based on the open-source tool PyTorch. Figure 1 The architecture can also be built using other open-source tools such as TensorFlow 2.0. The prediction function in the congestion prediction module is implemented by a residual composite network consisting of an LSTM long short-term recurrent neural network and an encoder-decoder based on a self-attention mechanism.

[0061] 3.1 Parameter Initialization

[0062] Because the average road speed v, channel congestion rate c, packet loss rate p, and air interface delay l have different dimensions and magnitudes (e.g., the average road speed v is 20 km / h, the channel congestion rate c is 50%, and the air interface delay l is 10 ms), the gradient propagation process of network parameters may cause the learning center of gravity of the residual composite network to shift, and slow network convergence may lead to invalid results. The following formula is used to normalize each parameter:

[0063]

[0064] Where, x norm The normalized results are the average speed v of the road segment, the channel busy rate c, the packet loss rate p, and the air interface delay l.

[0065] The specific normalization process is an existing technology and is the standardization process for most neural networks before parameter input. It can accelerate network convergence and avoid learning center shift. The normalization process maps the original values ​​from the original interval to the interval [0,1].

[0066] 3.2 Data Segmentation

[0067] The calculation results in section 3.1 are segmented into different data truncated segments X of length L using a sliding dynamic window to adapt to the feature extraction mode of the hidden layers of the recurrent neural network. For example, at time t=1, the dynamic window starts at the first data group and ends at the Lth data group:

[0068]

[0069] The specific v represents the defined average speed concept of the road segment. v1 to vL are the average speed values ​​of the road segment at a specific moment; the others C1-C L and C, P1-PL and P, l1 and lL and l are the same as above.

[0070] After the data within this window is processed, the window's start and end positions are shifted one position to the right. That is, at t=2, the data truncation X2 is:

[0071]

[0072] And then move forward sequentially until the end of the data.

[0073] Specifically, it is as follows:

[0074]

[0075] Where X is the segmented and recombined feature data, L∈N + v1,…,v L The average speed of the road segment at different times.

[0076] 3.3 Congestion Prediction

[0077] The segmented feature data X, which includes lane and channel information of the road segment ahead, is used to... t As a residual composite network (specifically, the residual composite network is...) Figure 1 The network structure name refers to a custom-designed network structure implemented using existing technology. Because the computation process requires calculating the residual between the trend vector and the original value, and the encoder / decoder below the network structure performs input-calculation-output operations based on this residual, the network structure is named a residual composite network based on the characteristics of the computation process. The original value is input through a separate Long Short-Term Recurrent Neural Network (LSTM) that iterates repeatedly (specifically, this iteration is implemented using existing technology. The essence of a recurrent neural network is to use the unit state ct and hidden state ht, etc., output from the previous stage as part of the input for the next stage, iterating in this way until the end of the loop) to extract the trend vector.

[0078] X' t =LSTM(X) t )

[0079]

[0080] Where X' t Let t represent the trend of average speed of road segment, channel busy rate, packet loss rate, and air interface delay output by the LSTM network at time t.

[0081] The residual δ is obtained by subtracting the trend value from the original value. This residual represents the magnitude of the difference between the predicted trend and the actual value in the congestion prediction module.

[0082] δ=X' t -X t

[0083] Using δ as Figure 1 The encoder receives the input from the encoder network and encodes the background vector containing delta feature information. The encoder uses a separate randomly initialized LSTM long short-term recurrent neural network.

[0084] C = [c t ,h t ] = Encoder(δ)

[0085] Among them, c t h is the cell state vector saved by the encoder during the iteration process. t The hidden state vector is saved by the encoder during the iteration process, and the two are concatenated to form the background vector C.

[0086] In the self-attention sublayer SA of the decoder, the background vector C is interleaved with a linear matrix. The dot product is mapped to the query matrix Q, the key matrix K, and the value matrix V. The self-attention relationship between Q, K, and V is then calculated.

[0087]

[0088] Wherein, s(k j ,q i The score for the correlation fraction equation is:

[0089]

[0090] The decoder-side LSTM unit is initialized with a random matrix zi containing self-attention relationships between road segment, lane, and channel information, and the residual δ is used as the decoder-side input. The input result δ' after constraining the differences in the residuals through the decoder-side network is:

[0091] δ'=Decoder(z i ,δ)

[0092] The constrained residual δ' is superimposed with the predicted trend σ to obtain the final prediction result σ':

[0093]

[0094] 4. Congestion Determination Module

[0095] Based on the prediction results from the congestion prediction module, it is determined whether the network channel on the upcoming road segment is congested if driving continues. The determination rule is as follows:

[0096] Channel busy rate > 50%;

[0097] Packet loss rate > 20%;

[0098] Transmission latency > 100ms

[0099] Specifically, if any two of the above three conditions are met, the network channel accessing the preceding road segment is determined to be congested; otherwise, it is determined to be uncongested. The relationship between these data and the final prediction result is a threshold relationship. If the final prediction result is greater than this value, and the determination rules are met, a channel congestion determination result can be output.

[0100] For example,

[0101] 5. Data splitting module

[0102] A C-V2X network quality ordered table is established in the channel acquisition module. Based on the results of the congestion determination module, if congestion occurs, the system selects to access either a low-Earth orbit satellite network or a cellular network according to the ordered table results, and provides automatic or manual access functions.

[0103] 6. Terminal display module

[0104] Provides data visualization capabilities. Based on the first item in the network quality ordered table (i.e., the currently accessed network), it displays the current C-V2X network communication status and the current network selection status in real time. It also shows whether the communication assistance system is enabled and the percentage improvement in performance of the selected network compared to other networks in the ordered table.

[0105] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A C-V2X communication assistance system based on a self-attention mechanism residual composite network, the system being deployed in an on-board OBU unit, characterized in that, The system includes a traffic condition acquisition module, a channel detection module, a congestion prediction module, a congestion determination module, and a data diversion module; The traffic condition acquisition module acquires vehicle condition data; the channel detection module acquires channel information data ahead; the vehicle condition information and channel information data are input into the congestion prediction module, the neural network inside the congestion prediction module calculates and outputs the calculated parameters to the congestion determination module, and the congestion determination module selects whether to divert the data based on the parameter determination results; The prediction function in the congestion prediction module is implemented by a residual composite network consisting of an LSTM long short-term recurrent neural network and an encoder-decoder based on a self-attention mechanism. The congestion prediction module integrates and preprocesses the data output from the traffic condition acquisition module and the channel detection module, mainly including the following steps: Step 1: Parameter initialization; Step 2: Data partitioning; Step 3: Congestion prediction; In step 3, the segmented feature data X containing vehicle condition and channel information is... t As the original input to the residual composite network, the trend vector is extracted through iterative processing by a separate Long Short-Term Recurrent Neural Network (LSTM). Subtract the trend value from the original value to obtain the residual δ; δ=X' t -X t ; δ is used as the input to the encoder network and the background vector containing δ feature information is encoded and extracted. The encoder uses a randomized LSTM long short-term recurrent neural network. C=[c t ,h t ]=Encoder(δ); Among them, c t h is the cell state vector saved by the encoder during the iteration process. t The hidden state vectors saved by the encoder during the iteration process are concatenated to form the background vector C; In the self-attention sublayer SA of the decoder, the background vector C is interleaved with a linear matrix. The dot product is mapped to the query matrix Q, the key matrix K, and the value matrix V. The self-attention relationship between Q, K, and V is then calculated. Wherein, s(k j ,q i The score for the correlation fraction equation is: The decoder-side LSTM unit is initialized with a random matrix zi containing the self-attention relationship of road segment, lane, and channel information, and the residual δ is used as the decoder-side input; the input result δ' after constraining the difference of the residual through the decoder-side network is then: δ'=Decoder(z i ,d); The constrained residual δ' is superimposed with the predicted trend σ to obtain the final prediction result σ':

2. The C-V2X communication assistance system based on a self-attention mechanism residual composite network according to claim 1, characterized in that, The system also includes a terminal display module, which displays the congestion determination results from the congestion determination module.

3. The C-V2X communication assistance system based on a self-attention mechanism residual composite network according to claim 1, characterized in that, The road condition acquisition module obtains vehicle condition data as the average speed v of the road segment; The specific steps for obtaining the average speed v of a road segment are as follows: The road condition acquisition module reads the driving route from the vehicle navigation system, requests the corresponding road condition information from the roadside units (RSUs) ahead of the driving route, including the number of vehicles m and the average speed of a single vehicle vi; and calculates the corresponding average speed v of the road segment using m and vi. The calculation method is as follows: Where m is the number of vehicles on the corresponding road segment.

4. The C-V2X communication assistance system based on a self-attention mechanism residual composite network according to claim 3, characterized in that, The channel detection module acquires channel information data from the front, including channel busy rate c, packet loss rate p, and air interface delay l; and detects the network quality between base stations and between base stations and satellite signals. The detection results are sorted in descending order of value to construct a corresponding ordered table of C-V2X network quality.

5. A C-V2X communication assistance system based on a self-attention mechanism residual composite network according to claim 4, characterized in that, The congestion prediction module integrates and preprocesses the data output from the traffic condition acquisition module and the channel detection module, using it as input parameters for the neural network in the congestion prediction module, and outputs the corresponding results.

6. A C-V2X communication assistance system based on a self-attention mechanism residual composite network according to claim 1, characterized in that, In step 1, the average speed v of the road segment, the channel busy rate c, the packet loss rate p, and the air interface delay l are normalized.

7. A C-V2X communication assistance system based on a self-attention mechanism residual composite network according to claim 6, characterized in that, In step 2, the calculation results in step 1 are divided into different data segments X of length L to adapt to the feature extraction mode of the hidden layer of the recurrent neural network.

Images