Cellular weak signal feature extraction and application method and system under vehicle driving scene
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AUTOMOTIVE INTELLIGENT TECHNOLOGY (TIANJIN) CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-08-07
AI Technical Summary
[0004]传统方法通常仅基于单一维度的信号参数（如参考信号接收功率）进行特征描述，无法区分“隧道遮挡导致的信号衰减”与“城市峡谷多径效应导致的信号波动”等不同弱信号本质，导致特征刻画精度低、场景适配性差，难以满足智能网联车辆的动态应用需求
1.场景关联性强：本申请结合车辆运动轨迹与环境属性构建场景分类体系，突破传统方法“脱离行驶场景”的局限，使特征提取更贴合车辆动态行驶实际情况；
Smart Images
Figure CN122534404A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of vehicle communication technology, and more specifically, to a method and system for extracting and applying weak cellular signal features in vehicle driving scenarios. Background Technology
[0002] With the rapid development of intelligent connected vehicle technology, cellular communication (4G / 5G) has become a core supporting technology for vehicle-to-everything (V2X) communication. The stability of its signal transmission directly determines the accuracy of environmental perception, the timeliness of decision-making and response in intelligent driving, and the user experience of navigation, entertainment, and remote control functions in the intelligent cockpit. During dynamic driving, vehicles frequently traverse complex scenarios such as tunnels, urban canyons, tree-lined roads, and suburbs. These scenarios are prone to weak cellular communication signals due to geographical obstruction, electromagnetic interference, multipath propagation effects, or insufficient base station coverage. This manifests as a sudden drop in signal strength, increased transmission latency, and a higher bit error rate.
[0003] Weak signal issues have become a key bottleneck restricting the development of vehicle-to-everything (V2X) communication. In intelligent driving scenarios, weak signals can lead to interruptions in environmental data transmission for automated driving assistance systems, causing vehicle decision-making errors. In smart cockpit scenarios, weak signals can cause navigation and positioning drift, audio-visual playback stuttering, and remote control failures. Therefore, accurately extracting the characteristic patterns of weak cellular communication signals under different driving scenarios is a prerequisite for achieving weak signal early warning and adaptive adjustment of communication links, and is of great significance for improving the reliability of V2X communication.
[0004] Traditional methods typically describe features based on a single-dimensional signal parameter (such as the received power of the reference signal), which cannot distinguish between different weak signal characteristics such as "signal attenuation caused by tunnel obstruction" and "signal fluctuation caused by multipath effect in urban canyons". This results in low feature characterization accuracy and poor scene adaptability, making it difficult to meet the dynamic application needs of intelligent connected vehicles. Summary of the Invention
[0005] The purpose of this application is to provide a method and system for extracting and applying weak cellular signal features in vehicle driving scenarios. By constructing a scenario classification system, mining multi-dimensional signal features, and establishing a feature-scenario mapping model, the method can accurately depict the essential laws of weak signals in different driving scenarios and apply them in practice.
[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application provides a method for extracting and applying weak cellular signal features in a vehicle driving scenario, including: During vehicle operation, multimodal data is collected, including driving data and cellular signal data. Feature extraction is performed on the multimodal data to obtain core features; The core features are input into a weak signal feature-scene mapping model based on random forest, and the current driving scene type and the contribution weight of the core features to the scene type are output. The vehicle's communication link and / or communication algorithm are adaptively adjusted based on the current driving scenario type and contribution weight.
[0007] Secondly, this application provides a cellular weak signal feature extraction and application system for vehicle driving scenarios, including: Onboard sensors are used to collect multimodal data during vehicle operation, including driving data and cellular signal data. The vehicle communication controller is used to extract features from the multimodal data to obtain core features; input the core features into a weak signal feature-scene mapping model based on random forest, output the current driving scene type and the contribution weight of the core features to the scene type; and adaptively adjust the vehicle's communication link and / or communication algorithm according to the current driving scene type and contribution weight.
[0008] Compared with the prior art, this application has the following beneficial effects: 1. Strong scene relevance: This application combines vehicle motion trajectory and environmental attributes to construct a scene classification system, breaking through the limitation of traditional methods that are "detached from the driving scene", making feature extraction more in line with the actual dynamic driving situation of vehicles; 2. Precise feature characterization: This application extracts features from multiple dimensions such as signal amplitude, time, and space. Compared with single-parameter analysis, it can more comprehensively capture the essential laws of weak signals in different scenarios. 3. High practicality: The feature-scene mapping model constructed in this application can directly provide input for the weak signal early warning mechanism, and quickly match scene types through core features, providing a decision basis for adaptive adjustment of communication links. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0010] Figure 1 This is a flowchart illustrating a method for extracting and applying weak cellular signal features in a vehicle driving scenario, as provided in an embodiment of this application. Figure 2 This is a schematic diagram of the vehicle provided in the embodiments of this application. Detailed Implementation
[0011] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.
[0012] Figure 1 This is a flowchart of a method for extracting and applying weak cellular signal features in a vehicle driving scenario, provided in this embodiment. This embodiment can accurately extract the amplitude attenuation characteristics, temporal correlation, and spatial distribution patterns of weak signals in different scenarios, providing reliable feature support for optimizing vehicle network communication performance.
[0013] The purpose of this embodiment is to provide a method for extracting and applying weak cellular signal features in typical vehicle driving scenarios, which is used to extract weak communication signals and identify typical weak cellular network scenarios in vehicles.
[0014] See Figure 1 The method provided in this embodiment includes: S110. During vehicle operation, multimodal data is collected, including driving data and cellular signal data.
[0015] The main body executing this step is the vehicle-mounted sensor. The driving data includes: vehicle speed and driving direction; the cellular signal data includes: reference signal received power (RSRP), reference signal received quality (RSRQ), signal-to-noise ratio (SINR), signal fluctuation frequency, transmission delay, and bit error rate.
[0016] For example, vehicle speed is collected through wheel speed sensors, driving direction is collected through steering wheel angle sensors, and cellular signal data is collected through cellular communication modules.
[0017] S120. Extract features from multimodal data to obtain core features.
[0018] The collected multimodal data is denoised. For example, the Kalman filter algorithm is used to eliminate random noise in signal parameters such as RSRP and RSRQ. The missing values in the data acquisition gaps are filled by interpolation. The driving data and cellular signal data are accurately aligned based on the timestamp.
[0019] The core features are those that can distinguish (vehicle driving) scenario types, including: RSRP attenuation magnitude, SINR fluctuation frequency, signal attenuation duration, and the gradient of signal quality changes with vehicle speed. These core features need to be pre-selected, which will be described in subsequent embodiments.
[0020] The process of extracting core features from multimodal data includes: RSRP attenuation is the difference between the highest and lowest points of signal strength during a significant decrease. In continuously acquired RSRP data, first identify segments where the signal continuously decreases (e.g., RSRP decreases at several consecutive sampling points). For each such segment, subtract the lowest value at the end of the segment from the initial signal strength value; the difference is the attenuation magnitude.
[0021] SINR fluctuation frequency is the rate at which the SINR value fluctuates up and down per unit time, measured in Hertz (Hz). First, remove the slowly changing portions of the SINR data (e.g., subtract the average value over several seconds), retaining only the fast-moving segments. Then, count the number of times the SINR value changes from positive to negative or vice versa (crossing the average value) within these fast-moving segments. Divide this number by 2 to obtain the number of fluctuation cycles. Dividing the number of fluctuation cycles by the time duration gives the fluctuation frequency.
[0022] The signal attenuation duration is the time it takes for RSRP to rise from its initial decline to its lowest point. Specifically, first, detect events where RSRP continuously declines. Record the start time of the event (the point where the signal begins its continuous decline) and the end time of the event (after reaching the lowest point, it stops declining, begins to rise again, or remains flat). Subtract the start time of the event from the end time; the resulting time difference is the duration of this attenuation.
[0023] The gradient of signal quality as a function of vehicle speed is the change in signal quality (RSRP or SINR) per 1 km / h change in vehicle speed, expressed in dB. Specifically, a scatter plot is created with vehicle speed on the x-axis and signal quality on the y-axis. A straight line is fitted to this scatter plot, and the slope of this line is the gradient.
[0024] S130. Input the core features into the weak signal feature-scene mapping model based on random forest, and output the current driving scene type and the contribution weight of the core features to the scene type.
[0025] The structure and function of the weak signal feature-scene mapping model based on random forest include: 1) Input layer, used to input multi-dimensional core features. The core features need to be converted into fixed-length numerical vectors, and each core feature is normalized or standardized.
[0026] 2) The intermediate layer consists of N independent decision trees. Each tree outputs a predicted scene category. The construction process for each tree is as follows: Bootstrap sampling is performed from the original dataset. At each node split, m features are randomly selected, and the best splitting feature and threshold are chosen from these. Gini impurity or information gain is used to evaluate the splitting quality. The splitting stops when the minimum number of samples per leaf node or the maximum tree depth is reached.
[0027] 3) Output layer: The prediction results of all decision trees are voted on by majority vote. The scenario category with the most votes is used as the final output of the model, which is the current driving scenario type.
[0028] Furthermore, based on the weak signal feature-scene mapping model of random forest, the contribution weight of core features to scene type, i.e., feature importance score, can also be obtained. This contribution weight is calculated by the trained random forest weak signal feature-scene mapping model on all training samples, reflecting the distinguishing ability of features in scene classification.
[0029] S140. Adaptively adjust the vehicle's communication link and / or communication algorithm based on the current driving scenario type and contribution weight.
[0030] In tunnel scenarios, if the contribution weight of RSRP attenuation exceeds a set value (e.g., 0.7), RSRP will usually experience severe attenuation at the tunnel entrance. This feature contributes the most to the determination of tunnel scenarios. Therefore, the algorithm weight of GNSS is actively reduced, offline maps are cached in advance, and inertial navigation is enabled to calculate the path. The relative displacement is obtained by IMU integration and is periodically corrected by vehicle speed sensor to ensure the continuity and short-term accuracy of positioning.
[0031] In urban canyon scenarios, if the contribution weights of SINR fluctuation frequency and signal attenuation duration both exceed set thresholds (e.g., 0.6 and 0.5 respectively), it indicates frequent multipath and signal handover. In this case, wide-beam communication is extended, and resource reservations are initiated in advance with the target base station to improve signal coverage stability and reduce handover interruptions. For example, the antenna array is controlled to switch from a narrow beam (high gain) to a wide beam (low gain but large coverage angle) to enhance the reception capability of signals from multiple reflection paths and mitigate fast fading caused by multipath. A Handover Preparation Request is sent from the current serving base station to adjacent, predicted candidate target base stations, carrying QoS requirements (such as guaranteed bit rate and latency tolerance); the number of Hybrid Automatic Repeat Request (HARQ) retransmissions at the Media Access Control (MAC) layer is increased to reduce the bit error rate.
[0032] Compared with the prior art, this application has the following beneficial effects: 1. Strong scene relevance: This application combines vehicle motion trajectory and environmental attributes to construct a scene classification system, breaking through the limitation of traditional methods that are "detached from the driving scene", making feature extraction more in line with the actual dynamic driving situation of vehicles; 2. Precise feature characterization: This application extracts features from multiple dimensions such as signal amplitude, time, and space. Compared with single-parameter analysis, it can more comprehensively capture the essential laws of weak signals in different scenarios. 3. High practicality: The feature-scene mapping model constructed in this application can directly provide input for the weak signal early warning mechanism, and quickly match scene types through core features, providing a decision basis for adaptive adjustment of communication links.
[0033] Optionally, before extracting features from the multimodal data to obtain core features, the method further includes: determining core features; and before determining core features, the method further includes: constructing a classification system for typical weak signal scenarios, including: multiple basic scenarios and sub-scenarios under each basic scenario.
[0034] This embodiment constructs a two-level weak signal scene classification system comprising "basic scenes - subdivided scenes". Weak signal scenes are divided into four basic scenes: urban dense area scenes, enclosed space scenes, open road scenes, and complex terrain scenes. Finally, the basic scenes are further subdivided: urban dense area scenes include urban canyons and tree-lined roads; enclosed space scenes include tunnels and underground parking garages; open road scenes include elevated roads and suburban roads; and complex terrain scenes include mountain roads and bridges, forming a classification system covering the main weak signal scenes.
[0035] Identify core features, including: Step 1: In the pre-training phase, collect multimodal data samples and scene type labels during vehicle operation.
[0036] During the pre-training phase, test routes covering all sub-scenarios are planned for vehicle driving, and multimodal data samples and scenario type labels are collected. Figure 2 This is a schematic diagram of a vehicle provided in an embodiment of this application. The vehicle is equipped with a cellular antenna, a positioning antenna, a host computer, a frequency sweeper, a GPS module, and an inertial measurement unit. For example, a smart connected vehicle equipped with a 5G communication module is selected, and a signal acquisition unit is integrated into the frequency sweeper. A high-precision GPS module (positioning accuracy ±1m) and an inertial measurement unit (sampling frequency 100Hz) are installed simultaneously. The planned test route includes typical scenarios such as densely populated urban areas, tunnels, elevated roads, and suburban roads, with a total length of 100km, of which sections prone to weak signals account for no less than 60%.
[0037] The vehicle is started and driven along the planned route, while the following data is collected simultaneously: cellular signal data (RSRP, RSRQ, SINR sampling frequency 10Hz; signal fluctuation frequency, transmission delay, and bit error rate sampling frequency 1Hz), driving data (vehicle speed and driving direction sampling frequency 5Hz), and GPS location data. The GPS location data is then mapped onto a high-precision map to obtain scene type labels, and three rounds of test data are continuously collected to form a sample set.
[0038] For example, based on GPS location data and high-precision maps, six sub-categories of scenarios were obtained: urban canyons (densely populated areas with high-rise buildings in the city center), tree-lined roads (width of tree cover on both sides > 5m), tunnels (length > 800m), elevated roads (height above ground > 10m), suburban roads (open areas without obstruction), and mountain roads (slope > 15°). Each data point was labeled with a scenario type.
[0039] Step 2: The host computer preprocesses the multimodal data samples.
[0040] For example, the host computer uses the Kalman filter algorithm to denoise the RSRP and RSRQ data, with the filter coefficient set to 0.8; it uses linear interpolation to fill in the missing 5% of signal data, and aligns the cellular signal data with the vehicle driving data based on the timestamp to form a standardized dataset.
[0041] Step 3: The host computer extracts feature sets from the three dimensions of amplitude characteristics, temporal characteristics, and spatial characteristics for the preprocessed multimodal data samples; the Relief-F feature selection algorithm is used to calculate the correlation coefficient between each feature in the feature set and the scene type label; based on the correlation coefficient, core features with high scene discrimination are selected.
[0042] Specifically, feature sets are calculated for each scenario. For example, in the tunnel scenario, the average RSRP attenuation amplitude is 28 dBm, and the average attenuation duration is 45 seconds; in the urban canyon scenario, the SINR fluctuation range is 5-18 dB, and the average fluctuation frequency is 6.2 Hz; in the suburban road scenario, the average RSRQ is -9 dB, and the signal quality gradient coefficient with vehicle speed is 0.03 dB / (km / h). Then, the Relief-F algorithm is used to calculate feature correlation, and four core features are selected: RSRP attenuation amplitude, SINR fluctuation frequency, signal attenuation duration, and the gradient of signal quality with vehicle speed. The correlation coefficients between these features and the scenario labels are all >0.75. For example, the core features of the tunnel scenario include RSRP attenuation amplitude (>20 dBm) and attenuation duration being positively correlated with tunnel length; the core features of the urban canyon scenario include high-frequency SINR fluctuation (fluctuation frequency >5 Hz, and significant signal quality changes with driving direction).
[0043] Relief-F is a filtering feature selection algorithm used to evaluate the relevance of each feature to the classification problem and the class label. The algorithm outputs a correlation coefficient of 0 to 1 or -1 to 1. If "all correlation coefficients > 0.75", it means that the weights of these four core features exceed the high threshold of 0.75, indicating that the four core features are very effective in distinguishing different scenarios (such as multipath, switching, open areas, etc.).
[0044] Optionally, before inputting the core features into the weak signal feature-scene mapping model based on random forest and outputting the current driving scene type and the contribution weight of the core features to the scene type, the model can also be trained. In the pre-training phase, multimodal data samples are preprocessed and core features are extracted to obtain input samples. The preprocessing and core feature extraction processes are described in the above embodiments and will not be repeated here. The input samples are then fed into a weak signal feature-scene mapping model based on random forest to obtain the predicted scene type. By minimizing the difference between the predicted scene type and the scene type label, the number and depth of the decision trees in the model are optimized to obtain the trained weak signal feature-scene mapping model based on random forest.
[0045] Optionally, the trained weak signal feature-scene mapping model based on random forest is deployed to the vehicle communication controller to receive multimodal data during vehicle operation in real time. The model quickly matches the scene type and outputs feature contribution weights. Under different scenarios, the vehicle communication controller can adaptively adjust the vehicle's communication link and / or communication algorithm according to the current driving scene type and contribution weights. Simultaneously, new measured data is collected quarterly to incrementally train the model, ensuring its adaptability to new scenarios (such as new elevated bridges).
[0046] This application also provides a cellular weak signal feature extraction and application system for vehicle driving scenarios, including: Onboard sensors are used to collect multimodal data during vehicle operation, including driving data and cellular signal data. The vehicle communication controller is used to extract features from the multimodal data to obtain core features; input the core features into a weak signal feature-scene mapping model based on random forest, output the current driving scene type and the contribution weight of the core features to the scene type; and adaptively adjust the vehicle's communication link and / or communication algorithm according to the current driving scene type and contribution weight.
[0047] The cellular weak signal feature extraction and application system for vehicle driving scenarios provided in this embodiment can execute the above-mentioned cellular weak signal feature extraction and application method for vehicle driving scenarios and has the corresponding technical effects, which will not be elaborated here.
[0048] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.
[0049] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A method for extracting and applying weak cellular signal features in a vehicle driving scenario, characterized in that, include: During vehicle operation, multimodal data is collected, including driving data and cellular signal data. Feature extraction is performed on the multimodal data to obtain core features; The core features are input into a weak signal feature-scene mapping model based on random forest, and the current driving scene type and the contribution weight of the core features to the scene type are output. The vehicle's communication link and / or communication algorithm are adaptively adjusted based on the current driving scenario type and contribution weight.
2. The method according to claim 1, characterized in that, Before extracting features from the multimodal data to obtain core features, the process also includes: Identify core features; The determination of core features includes: During the pre-training phase, multimodal data samples and scene type labels are collected while the vehicle is in motion; The multimodal data samples are preprocessed; For the preprocessed multimodal data samples, feature sets are extracted from three dimensions: amplitude characteristics, temporal characteristics, and spatial characteristics. The Relief-F feature selection algorithm is used to calculate the correlation coefficient between each feature in the feature set and the scene type label; Based on the correlation coefficient, core features with high scene discrimination are selected.
3. The method according to claim 2, characterized in that, Before determining the core features, the following also includes: A classification system for typical weak signal scenarios is constructed, including: multiple basic scenarios and sub-scenarios under each basic scenario.
4. The method according to claim 3, characterized in that, During the pre-training phase, multimodal data samples and scene type labels are collected while the vehicle is in motion, including: During the pre-training phase, test routes covering all sub-scenarios are planned for vehicle driving, and multimodal data samples and scenario type labels are collected.
5. The method according to claim 4, characterized in that, Before inputting the core features into a weak signal feature-scene mapping model based on random forest, and outputting the current driving scene type and the contribution weight of the core features to the scene type, the following steps are taken: During the pre-training phase, multimodal data samples are preprocessed and core features are extracted to obtain input samples; The input samples are fed into a weak signal feature-scene mapping model based on random forest to obtain the predicted scene type; By minimizing the difference between the predicted scene type and the scene type label, the number and depth of the decision trees in the model are optimized, resulting in a well-trained weak signal feature-scene mapping model based on random forest.
6. The method according to claim 5, characterized in that, Based on the current driving scenario type and contribution weight, the vehicle's communication link and / or communication algorithm are adaptively adjusted, including: In tunnel scenarios, if the contribution weight of RSRP attenuation exceeds a set value, the algorithm weight for GNSS is actively reduced, offline maps are cached in advance, and inertial navigation path calculation is enabled. In urban canyon scenarios, if the contribution weights of SINR fluctuation frequency and signal attenuation duration both exceed the set thresholds, it indicates multipath and frequent signal handover, expands to wide-beam communication, and initiates resource reservation to the target base station in advance to improve signal coverage stability and reduce handover interruptions.
7. The method according to claim 6, characterized in that, Driving data includes: vehicle speed and direction of travel; Cellular signal data includes: reference signal received power, reference signal received quality, signal-to-noise ratio, signal fluctuation frequency, transmission delay, and bit error rate.
8. The method according to claim 7, characterized in that, Key features include: RSRP attenuation magnitude, SINR fluctuation frequency, signal attenuation duration, and the gradient of signal quality as a function of vehicle speed.
9. The method according to claim 8, characterized in that, During the pre-training phase, scene type labels are collected while the vehicle is in motion, including: During the pre-training phase, GPS location data is collected while the vehicle is in motion; GPS location data is mapped onto a high-precision map to obtain scene type labels.
10. A system for extracting and applying weak cellular signal features in a vehicle driving scenario, characterized in that, include: Onboard sensors are used to collect multimodal data during vehicle operation, including driving data and cellular signal data. The vehicle communication controller is used to extract features from the multimodal data to obtain core features; input the core features into a weak signal feature-scene mapping model based on random forest, output the current driving scene type and the contribution weight of the core features to the scene type; and adaptively adjust the vehicle's communication link and / or communication algorithm according to the current driving scene type and contribution weight.