A deep learning-based vehicle-mounted communication signal control method

Abstract

The application discloses a kind of vehicle communication signal control methods based on deep learning, specifically relates to vehicle communication control technical field, historical vehicle sensor data, historical communication network data and historical traffic flow data are collected and are time-space alignment, historical multi-source data after time-space alignment are sequentially carried out data cleaning, feature extraction and feature fusion, construct deep learning model, training set, verification set and test set are divided according to time sequence and then carried out multi-objective optimization training, sequentially to the model trained laboratory simulation verification, closed site real vehicle verification and open road small-scale verification, after verification passes, model is deployed to carry out subsequent vehicle communication control, improve the accuracy of communication control scheme generation while being able to reduce the generation time of communication control scheme, and then reduce the regulation and control delay, greatly improve the vehicle driving safety and vehicle communication quality.

Description

Technical Field

[0001] This invention relates to the field of vehicle communication control technology, and more specifically, to a deep learning-based vehicle communication signal control method. Background Technology

[0002] With the rapid development of intelligent transportation systems, vehicle-to-everything (V2X) communication technology has become crucial for enabling information exchange between vehicles and between vehicles and infrastructure. However, in real-world traffic environments, due to high vehicle density, complex and variable road conditions, and the presence of various interference factors, traditional V2X communication signal control methods struggle to meet the requirements for efficient and stable communication, necessitating the search for a superior V2X communication signal control method.

[0003] An existing vehicle communication signal control method acquires vehicle status information through a CAN bus, acquires environmental perception data through a visual perception device, then performs feature extraction, analyzes the extracted features to determine whether the current vehicle status needs adjustment, if adjustment is required, generates an adjustment plan, and controls the vehicle according to the adjustment plan. To a certain extent, this method can improve communication quality, enhance vehicle safety, and optimize vehicle control.

[0004] However, existing methods still have some problems: generating vehicle control schemes based solely on vehicle status information and environmental perception data has limitations, as it does not involve the collection of vehicle communication information, and the final generated control parameters are difficult to guarantee the quality of vehicle communication; generating adjustment schemes based solely on simple data analysis is a rather mechanical process that takes a lot of time, which is not conducive to instantaneous vehicle communication control. Vehicle communication quality should be further improved and control latency reduced. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a vehicle communication signal control method based on deep learning to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based vehicle communication signal control method, comprising the following steps:

[0007] S1. Historical Multi-Source Data Acquisition: Acquire historical vehicle sensor data, historical communication network data, and historical traffic flow data, and perform spatiotemporal alignment.

[0008] S2. Historical Multi-Source Data Feature Fusion: The historical multi-source data after spatiotemporal alignment is sequentially cleaned, feature extracted, and feature fused.

[0009] S3. Deep learning model training: Construct a deep learning model, divide it into training set, validation set and test set according to time sequence and then perform multi-objective optimization training.

[0010] S4. Multi-dimensional evaluation of model training effect: The trained model is successively verified by laboratory simulation, closed-field real vehicle verification and open road small-scale verification. The key indicators of each verification stage are evaluated and different instructions are executed based on the evaluation results.

[0011] S5. Real-time communication control: Real-time communication control is achieved through the coordinated action of real-time data stream processing control, security control, and anomaly handling control.

[0012] S6. Communication Control Monitoring: Monitor the communication control process and record communication quality data, control performance data, and system security data.

[0013] S7. Communication Control Analysis: Analyze the recorded communication quality data, control performance data, and system security data. Based on the analysis results, determine whether to generate a model update instruction and whether to generate a communication control quality index update instruction.

[0014] S8, Instruction Execution: Update the model after generating the model update instruction, and calculate the communication control quality index after generating the communication control quality index update instruction;

[0015] S9. Data transmission: Transmit the calculated communication control quality index and the updated model to the vehicle communication control backend management terminal.

[0016] The technical effects and advantages of this invention are as follows:

[0017] This invention collects historical vehicle sensor data, historical communication network data, and historical traffic flow data and performs spatiotemporal alignment. The spatiotemporally aligned historical multi-source data is then sequentially cleaned, feature extracted, and fused to construct a deep learning model. After dividing the model into training, validation, and test sets according to time sequence, multi-objective optimization training is performed. The trained model is then sequentially validated through laboratory simulations, closed-field real-vehicle validation, and small-scale open-road validation. Once all validations are successful, the model is deployed for subsequent vehicle communication control. This improves the accuracy of communication control scheme generation while reducing the generation time, thereby reducing control latency and significantly improving vehicle driving safety and vehicle communication quality.

[0018] This invention monitors the communication control process, records communication quality data, control performance data, and system security data, analyzes the recorded communication quality data, control performance data, and system security data, and determines whether to generate a model update instruction based on the analysis results, and whether to generate a communication control quality index update instruction based on the analysis results. This can ensure the continuous stability of the model's performance, thereby ensuring the stability of communication control. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the method steps of the present invention.

[0020] Figure 2 This is a system structure block diagram of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 The embodiment shown provides a deep learning-based vehicle communication signal control method, including the following steps:

[0023] S1. Historical Multi-Source Data Acquisition: Acquire historical vehicle sensor data, historical communication network data, and historical traffic flow data, and perform spatiotemporal alignment.

[0024] Furthermore, the historical vehicle sensor data includes vehicle status data and environmental perception data. Vehicle status data includes basic status data, powertrain data, location data, and dynamic behavior data; environmental perception data includes visual perception data, radar scan data, and weather data. The historical communication network data includes V2X communication records and network performance indicators. V2X communication records include V2V, V2I, and V2N communication data; network performance indicators include transmission quality data, spectrum characteristic data, and protocol parameter data. Historical traffic flow data includes segment-level data, intersection-level data, and dynamic prediction data. Segment-level data includes traffic flow characteristic data, speed distribution data, and congestion index; intersection-level data includes traffic light phase data, queue length, and conflict point analysis; and dynamic prediction data includes short-term traffic prediction and event impact. The spatiotemporal alignment refers to associating all collected data using a unified timestamp and spatial coordinates.

[0025] Specifically, in this embodiment, the acquisition frequency of vehicle sensor data can be greater than or equal to 10Hz, the acquisition rule for communication data is event-triggered recording, and the acquisition frequency for traffic flow data can be 1 minute granularity, synchronized with the traffic light cycle. Basic status data includes, but is not limited to, vehicle speed, acceleration, steering angle, braking status, and throttle opening; powertrain data includes, but is not limited to, engine speed, battery SOC, and motor torque; location information data includes, but is not limited to, GPS latitude and longitude coordinates, heading angle, and altitude; dynamic behavior data includes, but is not limited to, acceleration variance (frequency of rapid acceleration / deceleration) and lateral offset (lane-keeping stability); visual perception data includes, but is not limited to, lane line recognition results and distance to vehicles ahead collected by cameras; radar scan data includes, but is not limited to, millimeter-wave radar / LiDAR point clouds (obstacle relative speed, azimuth angle); weather data includes, but is not limited to, rainfall data and light intensity; V2V communication data, i.e., vehicle-to-vehicle communication data, includes, but is not limited to, nearby vehicle IDs, signal strength (RSSI), and communication latency; and V2I communication data, i.e., vehicle-to-infrastructure communication data, includes, but is not limited to, roadside units. (RSU) location, signal coverage, channel occupancy; V2N communication data, i.e., vehicle-to-network communication data, including but not limited to base station handover records, cellular network signal-to-noise ratio (SNR); transmission quality data, including but not limited to bit error rate (BER), packet loss rate, and throughput; spectrum characteristic data, including but not limited to channel bandwidth and Doppler shift; protocol parameter data, including but not limited to TCP / UDP retransmission count and MAC layer collision count; traffic flow characteristic data, including but not limited to the number of vehicles passing through per unit time and lane occupancy; speed distribution data, including but not limited to average road segment speed and speed standard deviation; congestion index, which is the congestion level based on historical data, such as a normalized score in the range of 0-1; traffic light phase data, including but not limited to traffic light cycle and phase switching time; queue length, which refers to the number of vehicles waiting in each direction; conflict point analysis, including but not limited to vehicle lane change frequency and intersection emergency braking event records.

[0026] In this embodiment, it should be specifically noted that the specific formula for calculating the Doppler frequency shift Δf is as follows: v a f0, θ a c g The parameters are, in order: relative speed between the vehicle and the signal source (base station / other vehicles), signal carrier frequency, angle between the vehicle's direction of motion and the signal propagation direction, and speed of light; congestion index C. final The specific calculation formula is as follows: a f C index , t represents the attenuation coefficient, the basic congestion index, and the basic congestion index C at time t, respectively. index The specific calculation formula is as follows: Ibase f behavior The components are, in order, the basic congestion index and the behavior correction factor, and the specific calculation formulas are as follows: , a1, a2, a3, v e v free o e o crit q e q max a0, v std The weighting factors are, in order: weighting factor, weighting factor, weighting factor, average speed of road segment, free-flow speed (design value), lane occupancy, critical occupancy, vehicle queue length, maximum queue length, frequency of emergency braking events, and speed standard deviation.

[0027] S2. Historical Multi-Source Data Feature Fusion: The historical multi-source data after spatiotemporal alignment is sequentially cleaned, feature extracted, and feature fused.

[0028] Furthermore, the specific steps for feature fusion of historical multi-source data are as follows:

[0029] S21. Data cleaning: Process outliers in historical vehicle sensor data, historical communication network data, and historical traffic flow data based on the 3σ criterion, and process missing values ​​in historical vehicle sensor data, historical communication network data, and historical traffic flow data.

[0030] S22. Feature Extraction: Extract vehicle dynamic features, communication environment features, and traffic flow features. The extracted vehicle dynamic feature categories include basic state, driving behavior, and energy state. The extracted communication environment feature categories include signal quality and network state. The extracted traffic flow feature categories include micro-behavior, macro-state, and event impact.

[0031] S23. Feature Fusion: Construct a spatiotemporal feature tensor to generate a feature matrix with timestamps.

[0032] In this embodiment, the data cleaning process is as follows: Input spatiotemporally aligned historical vehicle sensor data, historical communication network data, and historical traffic flow data; calculate the mean and standard deviation for each data type according to a time window; if a data point exceeds the mean ± 3σ range within the window, it is marked as an anomaly; replace historical vehicle sensor data anomalies with linear interpolation within the window; mark communication data anomalies when the number of consecutive anomalies reaches a preset value; and fill traffic flow data anomalies with concurrent data from adjacent road segments. The specific steps for linear interpolation to replace historical vehicle sensor data anomalies are as follows: Input the original time series vehicle sensor data x = {x i1 ,x i2 ,...,x ij, where i refers to the i-th type of data, and j refers to the data with time series j in the i-th type of data; define the sliding window length W a and the step size A a = 1; calculate the window mean μ iw , and the specific formula is: , calculate the window standard deviation σ iw , and the specific formula is: ; if |x ij - μ iw | > 3σ iw , then mark x ij as an outlier, traverse the data points, and mark the index set A = {a1, a2,..., a m} of the outlier data points; determine the interpolation interval: if a i is an isolated point, then the left boundary L a = a i - 1, the right boundary R a = a i + 1, if a j , a j + 1,..., a j + k are consecutive outliers, then the left boundary L = the position of the last normal point with an order less than a j , the right boundary R = the position of the first normal point with an order greater than a j + k , if the outlier is at the start of the sequence, use the two nearest normal points on the right for extrapolation, if the outlier is at the end of the sequence, use the two nearest normal points on the left for extrapolation; linear interpolation calculation: for the interpolation point x it in the interval [L, R], (L < t < R), there is , x iL , x iR are the normal values of the left boundary and the right boundary in turn. When dealing with consecutive outliers, all points in the interpolation interval are generated according to this formula; replace the outlier values according to the linear interpolation calculation results.

[0033] Specifically in this embodiment, after filtering redundant features through random forest feature importance evaluation, the basic state features in vehicle dynamic features include but are not limited to the rate of change of speed and the variance of acceleration, the driving behavior features include but are not limited to the frequency of rapid acceleration, the frequency of rapid deceleration, and the lane keeping stability, and the energy state includes but is not limited to the rate of change of battery SOC; the signal quality features in communication environment features include but are not limited to the multipath fading coefficient and the Doppler frequency shift estimation, and the network state features include but are not limited to the channel conflict probability and the base station handover frequency; the microscopic behavior features in traffic flow features include but are not limited to the coding of following behavior patterns, the macroscopic state features include but are not limited to the road section congestion index, and the event impact features include but are not limited to the abnormal event propagation radius.

[0034] In this embodiment, the specific steps of feature fusion are as follows: The spatiotemporal grid is divided to determine the time axis and spatial axis; vehicle sensor features are mapped to their respective grids, communication features (such as RSSI) are mapped to multiple adjacent grids covered by the signal, and traffic flow features are mapped to corresponding grids according to road segment-lane; the tensor structure is determined, with dimensions consisting of time step × spatial grid × feature channel; the weights of different modal features are calculated using a multi-head self-attention mechanism; the tensor is divided according to a fixed time window (such as 10 seconds), with each slice containing features from all spatiotemporal grids within the window; the [H×W×C] dimension of each time slice is flattened into a 1-dimensional vector, such as 50x50 grid × 10 features → 25000-dimensional vector; the feature matrix with timestamps is stored using a time series database (such as InfluxDB) indexed by timestamps.

[0035] S3. Deep learning model training: Construct a deep learning model, divide it into training set, validation set and test set according to time sequence and then perform multi-objective optimization training.

[0036] Furthermore, the specific steps for training a deep learning model are as follows:

[0037] S31. Construct a TCN-GRU hybrid network: Temporal convolutional layers (TCN) are used to extract local temporal patterns, and gated recurrent units (GRU) are used to capture long-term dependencies.

[0038] S32. Data partitioning: Divide the training set, validation set, and test set in chronological order, and use a sliding window to generate the input sequence;

[0039] S33. Multi-objective optimization training: After determining the main objective, secondary objective, and constraints, the total loss function is determined. Then, the deep learning model is trained according to the set optimization strategy. The main objective is to minimize communication latency, the secondary objective is to minimize signal volatility, and the constraint is that the energy consumption growth rate is less than a preset value.

[0040] S4. Multi-dimensional evaluation of model training effect: The trained model is successively verified by laboratory simulation, closed-field real vehicle verification and open road small-scale verification. The key indicators of each verification stage are evaluated and different instructions are executed based on the evaluation results.

[0041] Furthermore, the specific steps for multi-dimensional evaluation of model training performance are as follows:

[0042] S41. Construct a digital twin platform, inject historical fault data into the platform, and perform simulation verification by randomly generating a preset number of extreme scenarios based on the Monte Carlo method. Record the number m of scenario groups in which the model output conforms to the expected control logic. ac Number of preset extreme scenarios (m) az Simulation environment output value ysim , actual historical data value y real Model prediction delay t pred i Actual latency t gt i Total number of inferences in the model m tz and the total inference time t of the model az ;

[0043] S42. Calculate the model fidelity coefficient, which is expressed as the difference between the numerical value 1 and the ratio of the root mean square error of the simulation environment output value and the real historical data value to the average value of the real historical data value. The specific formula is as follows: N a y represents the number of true historical data values. reale The latency stability coefficient is calculated using the average of real historical data. It is expressed as the average of the sum of the absolute values ​​of the differences between the model-predicted latency and the corresponding actual latency, and the ratio of these sums to the model-predicted latency. The specific formula is as follows: N t t represents the total number of predictions made by the model. pred i, t gt i represents the prediction delay and the actual delay of the i-th model inference, respectively, and σ represents the prediction delay and the actual delay, respectively. t i Let the standard deviation of the latency of model inference for i times be the extreme scenario coverage coefficient, which is expressed as the ratio of the number of scenario groups in which the model output conforms to the expected control logic to the number of extreme scenarios in the preset groups. The specific formula is as follows: The inference efficiency coefficient is calculated as the ratio of the total number of inferences to the total inference time. The specific formula is as follows: The calculated indicators are compared with the corresponding set expected indicators. If the calculated value is greater than or equal to the set expected value, the key indicator is judged to meet the standard. Otherwise, the key indicator is judged to fail to meet the standard. If all key indicators meet the standard, the model will be tested and verified in a closed field. Otherwise, the model will be reconstructed for training.

[0044] S43. Construct a high-density traffic flow scenario in a closed test track and create communication interference. Test vehicles travel in formation at a constant speed, and record the number of times m the relative error between the command and the execution result is less than a preset value. ba Total number of instructions m bz Number of accidental emergency braking incidents (m) ca Test duration t b Test strategy energy consumption E test Baseline strategy energy consumption E base Vehicle i receives timestamp t i rx Vehicle j sends timestamp t j sx ;

[0045] S44. Calculate the success rate of control command execution, expressed as the ratio of the number of times the relative error between the command and the execution result is less than a preset value to the total number of commands. The specific formula is as follows: The emergency braking false trigger rate is calculated as the ratio of the number of false emergency braking triggers to the test duration. The specific formula is as follows: The energy consumption growth rate is calculated as the ratio of the difference between the energy consumption of the test strategy and the energy consumption of the benchmark strategy to the energy consumption of the benchmark strategy. The specific formula is as follows: The multi-vehicle coordination error coefficient is calculated, which is expressed as the maximum clock deviation of all vehicle pairs. The specific formula is as follows: The calculated indicators are compared with the corresponding set expected indicators. If the calculated value is greater than or equal to the set expected value, the key indicator is judged to meet the standard. Otherwise, the key indicator is judged to fail to meet the standard. If all key indicators meet the standard, the open road small-scale verification will be carried out. Otherwise, the model parameters will be adjusted and the three-stage test will be carried out again.

[0046] S45. Select a typical open road scenario, collect data on unexpected events in natural traffic flow, and record the total normal operating time t of the system. yz Number of failures m da The weight b of the i-th scene ri The compliance coefficient b for the i-th scenario zi Number of manual interventions (m) da Test mileage L s Total number of communication signal transmissions m ez The number of times m is the actual transmitted power of the communication signal less than or equal to the standard maximum power. ea ;

[0047] S46. Calculate the operational reliability coefficient, which is expressed as the ratio of total normal operating time to the number of failures. The specific formula is as follows: The environmental adaptability coefficient is calculated using the following formula: When the i-th scenario meets the standard, b zi =1, otherwise b zi =0, calculate the manual intervention rate coefficient, which is the ratio of the number of manual interventions to the test mileage. The specific formula is: The communication compliance coefficient is calculated as the ratio of the number of times the actual transmission power of the communication signal is less than or equal to the standard maximum power to the total number of transmissions of the communication signal parameters. The specific formula is as follows: The calculated indicators are compared with the corresponding set expected indicators. If the calculated value is greater than or equal to the set expected value, the key indicator is determined to meet the standard. Otherwise, the key indicator is determined to fail. If all key indicators meet the standard, the model is deployed. If non-safety indicators fail to meet the standard, the model parameters are tuned and the three-stage test is repeated. If safety indicators fail to meet the standard, the model is reconstructed and trained.

[0048] S5. Real-time communication control: Real-time communication control is achieved through the coordinated action of real-time data stream processing control, security control, and anomaly handling control.

[0049] Furthermore, the real-time data stream processing control flow is as follows:

[0050] S511, Data Acquisition: Real-time synchronous acquisition of vehicle sensor data, communication data, and environmental data via CAN bus, V2X communication module, and visual perception device;

[0051] S512, Dual-buffered preprocessing: BufferA receives the real-time data collected, and BufferB preloads the new data of the next frame. It is physically isolated from BufferA. Buffer swapping is performed once every preset interval. The swapping process is performed atomically. The collected data is filtered for outliers and filled with missing values ​​in sequence.

[0052] S513, Spatiotemporal Alignment: Perform time and spatial alignment on the preprocessed acquired data, where spatial alignment includes coordinate system transformation and sensor extrinsic parameter calibration;

[0053] S514, Feature Drift Detection: Calculate the mean μ and standard deviation σ of each feature within the sliding window for the spatiotemporally aligned data. Perform Box-Cox transformation on non-Gaussian distributed features. If the current feature value exceeds the range of μ±3σ, trigger a feature drift alarm. At the same time, apply KL divergence detection. When the KL divergence is greater than the preset value, it is determined to be a significant drift.

[0054] S515, Model Inference: Input the feature vector after feature drift detection into the deep learning model for inference;

[0055] S516, Model Output: Obtain the original control parameter vector of the output, and calculate the variance matrix of the model output based on the Monte Carlo Dropout method.

[0056] Furthermore, the security control process is as follows:

[0057] S521, Confidence Detection: Perform a comprehensive confidence calculation on the parameter vector and variance matrix output by the model, and output the calculated confidence score and anomaly type label;

[0058] In this embodiment, it should be specifically noted that the confidence score is calculated using the following formula: b u1 Var ypred Var max b u2 Sim(F) current ,F history ), F current Fhistory The weights, in order, are: the weight coefficient for prediction uncertainty, the variance of the model's predicted output, the maximum permissible variance (normalized baseline), the weight coefficient for feature consistency, the similarity between the current feature and historical features, the feature vector at the current time step, and the historical feature reference baseline. The weight coefficient for prediction uncertainty can be optimized through grid search, typically with a value range of [0.6, 0.8]. The variance of the model's predicted output is actually the variance of the prediction results from multiple inferences (MonteCarlo Dropout) on the same input. The maximum permissible variance (normalized baseline) is determined based on the 99th quantile of the predicted variance distribution in the training set. The weight coefficient for feature consistency is b. u2 =1-b u1 Sim(F) current ,F history It can be calculated using the cosine similarity formula, specifically: F current F history Since all features are feature vectors, the dot product method is used. The feature vector at the current time step is the multidimensional vector of real-time data after feature engineering, and the historical feature reference is the exponentially weighted average of the feature vectors in the sliding window. F current t F current t-1 b s F history t The reference values ​​are, in order: historical feature reference benchmark at time t, historical feature reference benchmark at time t-1, historical memory intensity decay factor, and feature vector at time t.

[0059] S522, Confidence Level Response: The received confidence score is compared with the high confidence threshold, medium confidence threshold, low confidence threshold, and dangerous confidence threshold. If the confidence score is greater than or equal to the high confidence threshold, it is marked as high confidence and the model output is directly transmitted to the actuator. If the confidence score is less than the high confidence threshold but greater than or equal to the medium confidence threshold, it is marked as medium confidence. An exponential moving average is applied for smoothing filtering, and the adjustment range of the delay parameter is limited. If the confidence score is less than the medium confidence threshold but greater than or equal to the dangerous confidence threshold, it is marked as low confidence. The parameter change rate hard constraint is enabled, and the online feature verification module is triggered to re-verify the validity of the input data. If the confidence score is less than the low confidence threshold, it is marked as dangerous confidence, and the system immediately switches to the preset safety mode.

[0060] S523, Command Transmission: The final control parameters after security processing are encoded into commands, and then a transmission protocol is selected. The transmission guarantee mechanism is applied to transmit the control commands to the actuator.

[0061] Furthermore, the exception handling control process is as follows:

[0062] S531, Anomaly Detection: Compare the real-time sensor data, communication data, model output confidence score, and system health status with the corresponding preset normal standard values. If the data does not meet the preset normal standard values, mark the corresponding data as abnormal and output the anomaly event marking result.

[0063] S532, Anomaly Type Identification: Based on duration, scope of impact, and physical source, anomaly types are classified into transient anomalies, persistent anomalies, local anomalies, global anomalies, hardware anomalies, and software or algorithm anomalies. Anomaly identification is performed based on the feature identifiers corresponding to different anomaly types.

[0064] In this embodiment, it should be specifically noted that the characteristics of transient anomalies are single-time triggering and automatic recovery, the characteristics of continuous anomalies are consecutive triggering ≥3 times, the characteristics of local anomalies are affecting only a single subsystem, the characteristics of global anomalies are causing control commands to fail, the characteristics of hardware anomalies are accompanied by hardware error codes, and the characteristics of software or algorithm anomalies are low model confidence and no hardware error.

[0065] S533, Anomaly Recovery Strategy Selection: Automatically match the corresponding anomaly recovery strategy based on the identified anomaly type;

[0066] Specifically, in this embodiment, when a transient sensor malfunctions, data interpolation compensation can be selected; when communication is continuously interrupted, a redundant communication link can be switched to; when the model confidence remains low, a rollback to the previous stable model version can be selected; and when a hardware failure occurs, the faulty unit can be isolated and backup hardware can be enabled.

[0067] S534, Abnormal Recovery Strategy Execution: After locking relevant resources and backing up the current state, perform abnormal recovery according to the selected abnormal recovery strategy. Check whether the indicators after recovery reach the corresponding preset normal standard values. If they do, the abnormal recovery strategy is executed successfully; otherwise, the abnormal recovery strategy fails and enters deep recovery mode.

[0068] S6. Communication Control Monitoring: Monitor the communication control process and record communication quality data, control performance data, and system security data.

[0069] Furthermore, the communication quality data includes the amount of successfully transmitted data, m. sa Total transmission time t s The prediction delay t of the i-th model inference pred i Actual latency t gt i Number of smooth switching times m qaand total number of switch attempts m qz The control performance data includes the expected response time t for the i-th command. ui The actual response time t for the i-th instruction wi Effective communication volume m sr and total communication energy consumption E tz The system security data includes m instances where the actual transmitted power of the communication signal exceeds the standard maximum power. fa Safe mode duration t ha and total runtime t hz .

[0070] S7. Communication Control Analysis: Analyze the recorded communication quality data, control performance data, and system security data. Based on the analysis results, determine whether to generate a model update instruction and whether to generate a communication control quality index update instruction.

[0071] Furthermore, the specific steps of communication control analysis are as follows:

[0072] S71. Calculate the effective throughput coefficient X ap The specific formula is as follows: Calculate the time delay stability coefficient A tp The specific formula is as follows: N t σ represents the total number of predictions made by the model. t i Calculate the frequency band handover success rate coefficient X, which is the standard deviation of the delay for i-th model inference iterations. bp The specific formula is as follows: ;

[0073] S72. Calculate the command response timeliness coefficient X cp The specific formula is as follows: N b Calculate the energy efficiency ratio X for each instruction response. dp The specific formula is as follows: ;

[0074] S73. Calculate the frequency coefficient of violation operation X fp The specific formula is as follows: Calculate the safe mode trigger coefficient X ep The specific formula is as follows: ;

[0075] S74. Retrieve the preset standard values ​​for the effective throughput coefficient, latency stability coefficient, frequency band switching success rate coefficient, command response timeliness coefficient, energy efficiency ratio, violation operation frequency coefficient, and safety mode trigger coefficient, and denote them as X. aq A tq Xbq X cq X dq X fq X eq ;

[0076] Specifically, this embodiment requires clarification on the following: Based on the IEEE 802.11p protocol, the theoretical maximum single-channel speed is 27Mbps. Considering the actual environmental attenuation and retaining an 18.5% margin, to meet the concurrent transmission requirements of high-definition map updates (2Mbps) + collaborative sensing data (3Mbps), and considering the minimum throughput of 5.2Mbps measured in 90% urban scenarios, the preset standard value for the effective throughput coefficient can be 5Mbps. The measured latency stability coefficient shows an 8-fold increase in communication interruption probability when it is below 0.7; therefore, the preset standard value for the latency stability coefficient can be 0.7. Referring to the industry benchmark for 4G / 5G network handover success rate (99.3%–99.7%), and retaining 0.2% redundancy for V2X critical messages, the preset standard value for the frequency band handover success rate coefficient can be 99.5%. Considering that the actual command response time should be less than the expected command response time, and the smaller the time, the more timely the response, the preset standard value for the command response timeliness coefficient can be... Assuming a power consumption limit of 15W (36kJ / h) for the vehicle communication module, and requiring the transmission of 75GB of data per hour (all sensor data for autonomous driving), 75GB / 36kJ = 2.08GB / kJ. With a margin, the preset standard value for the energy efficiency ratio can be 2.1GB / kJ. Based on FCC Part 15's tolerance for ISM band transmit power violations, the requirement of MTBF (Mean Time Between Failures) ≥ 1000 hours, and the temperature rise rate ΔT < 3℃ / s of power devices under over-limit operation, the preset standard value for the violation frequency coefficient can be 1 time / hour. Based on the corresponding system availability level 3 standard of 99.9%, the economic balance point of requiring an additional $50,000 in hardware costs for every 0.01% improvement, and the acceptance criterion of RPN (Risk Priority Number) < 80 in FMEA analysis, the preset standard value for the safety mode trigger coefficient can be 0.1%.

[0077] S75. Compare the calculated values ​​of the effective throughput coefficient, delay stability coefficient, frequency band switching success rate coefficient, command response timeliness coefficient, energy efficiency ratio, violation operation frequency coefficient, and safety mode triggering coefficient with their corresponding standard values. If the calculated value of any of the following is less than the preset standard value, or if the calculated value of any of the following is greater than the preset standard value, a model update command is generated; otherwise, a communication control quality index update command is generated.

[0078] S8, Instruction Execution: Update the model after generating the model update instruction, and calculate the communication control quality index after generating the communication control quality index update instruction;

[0079] Furthermore, the specific steps for calculating the communication control quality index are as follows:

[0080] S81. Calculate the communication quality coefficient Y A The specific formula is as follows: ;

[0081] S82, Calculate the control effectiveness coefficient Y B The specific formula is as follows: ;

[0082] S83, Calculate the system security factor Y C The specific formula is as follows: ;

[0083] S84, Calculate the Communication Control Quality Index Z Y The specific formula is as follows: .

[0084] S9. Data transmission: Transmit the calculated communication control quality index and the updated model to the vehicle communication control backend management terminal.

[0085] In this embodiment, it should be noted that the preset values, set values, and expected values ​​used are all selected based on actual needs, and no specific value limit is imposed here. In order to avoid the denominator of the formula being 0, the denominator in some formulas is increased by 1 for adjustment.

[0086] like Figure 2 This embodiment provides a deep learning-based vehicle communication signal control system, including a historical multi-source data acquisition module, a historical multi-source data feature fusion module, a deep learning model training module, a multi-dimensional evaluation module for model training effect, a real-time communication control module, a communication control monitoring module, a communication control analysis module, a model update module, a communication control quality index update module, and a database. The historical multi-source data acquisition module, historical multi-source data feature fusion module, deep learning model training module, multi-dimensional evaluation module for model training effect, real-time communication control module, communication control monitoring module, and communication control analysis module are connected sequentially. The communication control analysis module is connected to the model update module and the communication control quality index update module. All modules in the system are connected to the database.

[0087] The historical multi-source data acquisition module is used to collect historical vehicle sensor data, historical communication network data, and historical traffic flow data and perform spatiotemporal alignment.

[0088] The historical multi-source data feature fusion module is used to perform data cleaning, feature extraction and feature fusion sequentially on the spatiotemporally aligned historical multi-source data.

[0089] The deep learning model training module is used to construct a deep learning model, and performs multi-objective optimization training after dividing the training set, validation set and test set in chronological order.

[0090] The multi-dimensional evaluation module for model training effect sequentially performs laboratory simulation verification, closed-site real vehicle verification, and open road small-scale verification on the trained model, evaluates the key indicators of each verification stage, and executes different instructions based on the evaluation results.

[0091] The real-time communication control module performs real-time communication control through the coordinated action of real-time data stream processing control, security control, and anomaly handling control.

[0092] The communication control and monitoring module is used to monitor the communication control process and record communication quality data, control performance data, and system security data.

[0093] The communication control analysis module is used to analyze the recorded communication quality data, control performance data, and system security data, and to determine whether to generate a model update instruction based on the analysis results, and whether to generate a communication control quality index update instruction based on the analysis results.

[0094] The model update module is used to update the model and transmit the updated model to the vehicle communication control backend management terminal after the model update is completed.

[0095] The communication control quality index update module is used to calculate the communication control quality index and transmit the calculated communication control quality index to the vehicle communication control backend management terminal after the calculation is completed.

[0096] The database is used to store data information for all modules in the system.

[0097] In conclusion, 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 within the protection scope of the present invention.

Claims

1. A deep learning-based vehicle communication signal control method, characterized in that: Includes the following steps: S1. Historical Multi-Source Data Acquisition: Acquire historical vehicle sensor data, historical communication network data, and historical traffic flow data, and perform spatiotemporal alignment. S2. Historical Multi-Source Data Feature Fusion: The historical multi-source data after spatiotemporal alignment is sequentially cleaned, feature extracted, and feature fused. S3. Deep learning model training: Construct a deep learning model, divide it into training set, validation set and test set according to time sequence and then perform multi-objective optimization training. S4. Multi-dimensional evaluation of model training effect: The trained model is successively verified by laboratory simulation, closed-field real vehicle verification and open road small-scale verification. The key indicators of each verification stage are evaluated and different instructions are executed based on the evaluation results. The specific steps for multi-dimensional evaluation of the model training effect are as follows: S41. Construct a digital twin platform, inject historical fault data into the digital twin platform, and randomly generate a preset number of extreme scenarios based on the Monte Carlo method for simulation verification. Record the number of scenario groups in which the model output meets the expected control logic, the number of preset extreme scenarios, the simulation environment output value, the real historical data value, the model prediction delay, the real delay, the total number of model inferences, and the total model inference time. S42. Calculate the model fidelity coefficient, latency stability coefficient, extreme scenario coverage coefficient, and inference efficiency coefficient. Compare the calculated indicators with the corresponding set expected indicators. If the calculated value is greater than or equal to the set expected value, the key indicator is judged to meet the standard. Otherwise, the key indicator is judged to fail to meet the standard. If all key indicators meet the standard, the model will be tested and verified in a closed field. Otherwise, the model will be reconstructed for training. S43. Construct a high-density traffic flow scenario in a closed test field and create communication interference. Test vehicles drive in formation at a constant speed. Record the number of times the relative error between the instruction and the execution result is less than the preset value, the total number of instructions, the number of times the emergency braking is accidentally triggered, the test duration, the energy consumption of the test strategy, the energy consumption of the baseline strategy, the timestamp of vehicle i receiving, and the timestamp of vehicle j sending. S44. Calculate the success rate of control command execution, the accidental emergency braking rate, the energy consumption growth rate, and the multi-vehicle coordination error coefficient. Compare the calculated indicators with the corresponding set expected indicators. If the calculated value is greater than or equal to the set expected value, the key indicator is judged to meet the standard. Otherwise, the key indicator is judged to fail to meet the standard. If all key indicators meet the standard, the test will proceed to small-scale verification on open roads. Otherwise, the model parameters will be adjusted and the three-stage test will be carried out again. S45. Select typical open road scenarios, collect sudden events in natural traffic flow, and record the total normal operating time of the system, the number of failures, the weight of the i-th scenario, the compliance coefficient of the i-th scenario, the number of manual interventions, the test mileage, the total number of communication signal transmissions, and the number of times the actual transmission power of the communication signal is less than or equal to the standard maximum power. S46. Calculate the operational reliability coefficient, manual takeover rate coefficient, and communication compliance coefficient. Compare the calculated indicators with the corresponding set expected indicators. If the calculated value is greater than or equal to the set expected value, the key indicator is determined to meet the standard. Otherwise, the key indicator is determined to fail to meet the standard. If all key indicators meet the standard, the model is deployed. If non-safety indicators fail to meet the standard, the model parameters are adjusted and the three-stage test is carried out again. If safety indicators fail to meet the standard, the model is reconstructed and trained. S5. Real-time communication control: Real-time communication control is achieved through the coordinated action of real-time data stream processing control, security control, and anomaly handling control. S6. Communication Control Monitoring: Monitor the communication control process and record communication quality data, control performance data, and system security data. S7. Communication Control Analysis: Analyze the recorded communication quality data, control performance data, and system security data. Based on the analysis results, determine whether to generate a model update instruction and whether to generate a communication control quality index update instruction. S8, Instruction Execution: Update the model after generating the model update instruction, and calculate the communication control quality index after generating the communication control quality index update instruction; S9. Data transmission: Transmit the calculated communication control quality index and the updated model to the vehicle communication control backend management terminal.

2. The vehicle communication signal control method based on deep learning according to claim 1, characterized in that: The historical vehicle sensor data includes vehicle status data and environmental perception data. Vehicle status data includes basic status data, powertrain data, location data, and dynamic behavior data. Environmental perception data includes visual perception data, radar scan data, and weather data. The historical communication network data includes V2X communication records and network performance indicators. V2X communication records include V2V, V2I, and V2N communication data. Network performance indicators include transmission quality data, spectrum characteristic data, and protocol parameter data. Historical traffic flow data includes segment-level data, intersection-level data, and dynamic prediction data. Segment-level data includes traffic flow characteristic data, speed distribution data, and congestion index. Intersection-level data includes traffic light phase data, queue length, and conflict point analysis. Dynamic prediction data includes short-term traffic prediction and event impact. Spatiotemporal alignment refers to associating all collected data using a unified timestamp and spatial coordinates.

3. The vehicle communication signal control method based on deep learning according to claim 1, characterized in that: The specific steps for fusing historical multi-source data features are as follows: S21. Data cleaning: Process outliers in historical vehicle sensor data, historical communication network data, and historical traffic flow data based on the 3σ criterion, and process missing values ​​in historical vehicle sensor data, historical communication network data, and historical traffic flow data. S22. Feature Extraction: Extract vehicle dynamic features, communication environment features, and traffic flow features. The extracted vehicle dynamic feature categories include basic state, driving behavior, and energy state. The extracted communication environment feature categories include signal quality and network state. The extracted traffic flow feature categories include micro-behavior, macro-state, and event impact. S23. Feature Fusion: Construct a spatiotemporal feature tensor to generate a feature matrix with timestamps.

4. The vehicle communication signal control method based on deep learning according to claim 1, characterized in that: The real-time data stream processing control flow is as follows: S511, Data Acquisition: Real-time synchronous acquisition of vehicle sensor data, communication data, and environmental data via CAN bus, V2X communication module, and visual perception device; S512, Dual-buffered preprocessing: BufferA receives the real-time data collected, and BufferB preloads the new data of the next frame. It is physically isolated from BufferA. Buffer swapping is performed once every preset interval. The swapping process is performed atomically. The collected data is filtered for outliers and filled with missing values ​​in sequence. S513, Spatiotemporal Alignment: Perform time and spatial alignment on the preprocessed acquired data, where spatial alignment includes coordinate system transformation and sensor extrinsic parameter calibration; S514, Feature Drift Detection: Calculate the mean μ and standard deviation σ of each feature within the sliding window for the spatiotemporally aligned data. Perform Box-Cox transformation on non-Gaussian distributed features. If the current feature value exceeds the range of μ±3σ, trigger a feature drift alarm. At the same time, apply KL divergence detection. When the KL divergence is greater than the preset value, it is determined to be a significant drift. S515, Model Inference: Input the feature vector after feature drift detection into the deep learning model for inference; S516, Model Output: Obtain the original control parameter vector of the output, and calculate the variance matrix of the model output based on the Monte Carlo Dropout method.

5. The vehicle communication signal control method based on deep learning according to claim 1, characterized in that: The security control process is as follows: S521, Confidence Detection: Perform a comprehensive confidence calculation on the parameter vector and variance matrix output by the model, and output the calculated confidence score and anomaly type label; S522, Confidence Level Response: The received confidence score is compared with the high confidence threshold, medium confidence threshold, low confidence threshold, and dangerous confidence threshold. If the confidence score is greater than or equal to the high confidence threshold, it is marked as high confidence and the model output is directly transmitted to the actuator. If the confidence score is less than the high confidence threshold but greater than or equal to the medium confidence threshold, it is marked as medium confidence. An exponential moving average is applied for smoothing filtering, and the adjustment range of the delay parameter is limited. If the confidence score is less than the medium confidence threshold but greater than or equal to the dangerous confidence threshold, it is marked as low confidence. The parameter change rate hard constraint is enabled, and the online feature verification module is triggered to re-verify the validity of the input data. If the confidence score is less than the low confidence threshold, it is marked as dangerous confidence, and the system immediately switches to the preset safety mode. S523, Command Transmission: The final control parameters after security processing are encoded into commands, and then a transmission protocol is selected. The transmission guarantee mechanism is applied to transmit the control commands to the actuator.

6. The vehicle communication signal control method based on deep learning according to claim 1, characterized in that: The communication quality data is used for communication control analysis, specifically including the amount of successfully transmitted data m. sa Total transmission time t s The prediction delay t of the i-th model inference pred i Actual latency t gt i Number of smooth switching times m qa and total number of switch attempts m qz The control performance data includes the expected response time t for the i-th command. ui The actual response time t for the i-th instruction wi Effective communication volume m sr and total communication energy consumption E tz ; The system security data includes the number of times m the actual transmission power of the communication signal exceeds the standard maximum power. fa Safe mode duration t ha and total runtime t hz .

7. The vehicle communication signal control method based on deep learning according to claim 6, characterized in that: The specific steps of the communication control analysis are as follows: S71. Calculate the effective throughput coefficient X ap The specific formula is as follows: Calculate the time delay stability coefficient A tp The specific formula is as follows: N t σ represents the total number of predictions made by the model. t i Calculate the frequency band handover success rate coefficient X, which is the standard deviation of the delay for i-th model inference iterations. bp The specific formula is as follows: ; S72. Calculate the command response timeliness coefficient X cp The specific formula is as follows: N b Calculate the energy efficiency ratio X for each instruction response. dp The specific formula is as follows: ; S73. Calculate the frequency coefficient of violation operation X fp The specific formula is as follows: Calculate the safe mode trigger coefficient X ep The specific formula is as follows: ; S74. Retrieve the preset standard values ​​for the effective throughput coefficient, latency stability coefficient, frequency band switching success rate coefficient, command response timeliness coefficient, energy efficiency ratio, violation operation frequency coefficient, and safety mode trigger coefficient, and denote them as X. aq A tq X bq X cq X dq X fq X eq ; S75. Compare the calculated values ​​of the effective throughput coefficient, delay stability coefficient, frequency band switching success rate coefficient, command response timeliness coefficient, energy efficiency ratio, violation operation frequency coefficient, and safety mode triggering coefficient with their corresponding standard values. If the calculated value of any of the following is less than the preset standard value, or if the calculated value of any of the following is greater than the preset standard value, a model update command is generated; otherwise, a communication control quality index update command is generated.

Images