A vehicle intelligent anti-rollover warning method, system, device and storage medium

Abstract

This invention provides a vehicle intelligent rollover prevention alarm method, system, device, and storage medium, belonging to the field of intelligent vehicle technology. The method includes: extracting the feature vector of the accident road segment when a rollover accident is detected in vehicle one; adjusting the LTR alarm threshold of the accident road segment according to the accident frequency of the accident road segment; uploading the accident road segment feature vector and the adjusted LTR alarm threshold to the cloud; extracting the driving road segment feature vector of vehicle two; matching an accident road segment feature vector from the cloud based on the driving road segment feature vector, and using the corresponding accident road segment's LTR alarm threshold as the LTR alarm threshold for vehicle two; and issuing a rollover prevention alarm for vehicle two based on the LTR alarm threshold of vehicle two. This invention achieves dynamic adjustment of the LTR alarm threshold, successfully transforming an isolated single-vehicle accident into safety knowledge that can protect all subsequent vehicles.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle technology, and in particular to a vehicle intelligent anti-rollover alarm method, system, device and storage medium. Background Technology

[0002] Current vehicle rollover warnings typically rely on a fixed lateral load transfer rate (LTR) threshold, which cannot dynamically adapt to changes in road conditions.

[0003] In addition, driverless cleaning fleets typically consist of multiple vehicles of the same model, performing daily cleaning tasks in fixed areas and along fixed routes. One safety issue with this operating model is that if one vehicle rolls over on a certain section of road, other vehicles in the same group still face the same risk of rollover when passing through the same area.

[0004] Traditional fleet safety management relies on the transmission of human experience, meaning that after an accident, managers verbally or in writing notify other drivers of the risks associated with a particular road section. This method suffers from significant delays (notification takes time) and incompleteness (the notification typically only states that a certain road section is dangerous, lacking quantifiable risk parameters). For autonomous vehicles, since there are no human drivers, this experience transmission mechanism completely fails. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a vehicle intelligent anti-rollover alarm method, system, device, and storage medium.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a vehicle intelligent anti-rollover warning method, the method comprising:

[0008] When a vehicle rollover accident is detected, the feature vector of the accident section is extracted.

[0009] Adjust the LTR alarm threshold of the accident section according to the frequency of accidents in the accident section;

[0010] Upload the feature vector of the accident section and the adjusted LTR alarm threshold to the cloud;

[0011] Extract the feature vector of the road segment in which vehicle 2 travels;

[0012] Based on the driving segment feature vector, an accident segment feature vector is matched from the cloud, and the LTR alarm threshold of the corresponding accident segment is used as the LTR alarm threshold of vehicle two.

[0013] Based on the LTR alarm threshold of vehicle two, a rollover prevention alarm is issued for vehicle two.

[0014] Furthermore, the step of extracting the feature vector of the accident section when a vehicle rollover accident is detected includes:

[0015] (a) The LTR of the current vehicle is greater than the preset LTR threshold and the duration is greater than the preset duration threshold.

[0016] (b) The vehicle roll angle is greater than the preset roll angle threshold;

[0017] (c) The vehicle body roll rate is greater than the preset angular velocity threshold;

[0018] The occurrence of a rollover accident is considered to be the fulfillment of any of the above conditions.

[0019] Furthermore, the method for extracting the feature vector of the accident section or the feature vector of the driving section includes:

[0020] Collect multi-dimensional vehicle data;

[0021] The multidimensional vehicle data is used to extract features using a preset feature extraction model to obtain feature vectors of accident sections or driving sections.

[0022] Furthermore, the multidimensional vehicle data includes: vehicle coordinates, vibration spectrum characteristics of the inertial measurement unit, peak lateral acceleration, maximum LTR value, vehicle speed, steering wheel angle, braking status, weather conditions, and road surface humidity.

[0023] Furthermore, the feature extraction model employs a long short-term memory neural network.

[0024] Furthermore, the long short-term memory neural network includes an input layer, a first LSTM hidden layer, a second LSTM hidden layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence.

[0025] Furthermore, the long short-term memory neural network is fine-tuned through an incremental learning mechanism after training.

[0026] Furthermore, adjusting the LTR alarm threshold of the accident section based on the accident frequency of the accident section includes:

[0027] For road sections where the frequency of rollover accidents is less than the preset frequency, the first LTR alarm threshold is used;

[0028] For road sections where the frequency of rollover accidents is greater than or equal to a preset frequency, a second LTR alarm threshold is used; the first LTR alarm threshold is greater than the second LTR alarm threshold.

[0029] Furthermore, the step of matching an accident road segment feature vector from the cloud based on the driving road segment feature vector includes:

[0030] For each accident road segment feature vector in the cloud, calculate their cosine similarity with the driving road segment feature vector;

[0031] The feature vector of the accident road segment with the largest cosine similarity greater than the preset similarity threshold is taken as the feature vector of the matched accident road segment.

[0032] Furthermore, before issuing a rollover prevention warning for vehicle two based on the LTR warning threshold of vehicle two, the method further includes: correcting the LTR warning threshold of vehicle two according to the weather conditions of the current driving section, and issuing a rollover prevention warning for vehicle two using the corrected LTR warning threshold.

[0033] Furthermore, the LTR alarm threshold of vehicle two is corrected based on the weather conditions of the current driving segment. The correction method includes:

[0034] Get the weather conditions for the current driving route;

[0035] The adhesion coefficient is read from the preset first weather-adhesion coefficient mapping table according to the current weather conditions. The adhesion coefficient is used to correct the LTR alarm threshold of vehicle two.

[0036] Furthermore, the correction method also includes:

[0037] Get the humidity of the current driving route;

[0038] Obtain the humidity range corresponding to the weather conditions of the current driving route;

[0039] Determine whether the humidity of the current driving section is within the humidity range corresponding to the weather conditions of the current driving section;

[0040] When the humidity of the current driving segment is within the humidity range corresponding to the weather conditions of the current driving segment, the preset first weather-adhesion coefficient mapping table is used.

[0041] When the humidity of the current driving segment is outside the humidity range corresponding to the weather conditions of the current driving segment, a preset second weather-adhesion coefficient mapping table is used; compared with the first weather-adhesion coefficient mapping table, the adhesion coefficient under each weather condition in the second weather-adhesion coefficient mapping table is reduced by a preset ratio.

[0042] Furthermore, the correction method also includes:

[0043] If weather conditions for the current driving route cannot be obtained, obtain the humidity of the current driving route.

[0044] The adhesion coefficient is read from a preset humidity-adhesion coefficient mapping table based on the current humidity. The adhesion coefficient is used to correct the LTR alarm threshold of vehicle two.

[0045] Furthermore, the step of issuing a rollover prevention alarm for vehicle two based on the LTR alarm threshold of vehicle two includes:

[0046] Calculate the current LTR value of vehicle two;

[0047] If the current LTR value is greater than the LTR alarm threshold of vehicle two, then an anti-rollover alarm will be issued for vehicle two.

[0048] Furthermore, the cloud platform is used to manage the feature vectors of each accident section and their corresponding LTR alarm thresholds, the management including:

[0049] If the accident frequency of any accident section is 0 within a preset continuous time period, the LTR alarm threshold of the corresponding accident section will be increased by a preset increment value every preset time period.

[0050] The present invention also provides a vehicle intelligent rollover prevention warning system, using the above-described vehicle intelligent rollover prevention warning method, the system comprising:

[0051] The feature extraction module is used to extract feature vectors of accident sections and driving sections;

[0052] The threshold adjustment module is used to adjust the LTR alarm threshold of the accident section according to the frequency of accidents in the accident section;

[0053] The matching module is used to match an accident road segment feature vector from the cloud based on the driving road segment feature vector, and use the LTR alarm threshold of the corresponding accident road segment as the LTR alarm threshold of the vehicle.

[0054] The alarm module is used to issue rollover prevention alarms based on the vehicle's LTR alarm threshold.

[0055] The present invention also provides a computer device, characterized in that it includes: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the above-described vehicle intelligent rollover prevention alarm method.

[0056] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above-described vehicle intelligent anti-rollover alarm method.

[0057] Compared with the prior art, the present invention has the following advantages:

[0058] (1) The present invention introduces a threshold adjustment mechanism based on accident frequency, which configures different alarm thresholds for road sections with different risk levels. Unlike the fixed thresholds in traditional technologies, it achieves the optimal balance between safety and comfort.

[0059] (2) The feature vector of the accident section and the adjusted LTR alarm threshold are uploaded to the cloud. When other vehicles travel to similar sections, they upload their own travel section feature vectors. The cloud can then use a cosine similarity matching algorithm to accurately match the most similar accident case in history and issue the corresponding risk threshold. This transforms an isolated accident into public safety knowledge that can protect all subsequent vehicles. Even if other vehicles have never traveled on that section, they can be informed of its potential risks in advance, achieving a leap from passive response to proactive prevention. Attached Figure Description

[0060] The following is a brief explanation of the content depicted in each of the accompanying drawings in this specification:

[0061] Figure 1 A flowchart of a vehicle intelligent anti-rollover alarm method provided in an embodiment of the present invention;

[0062] Figure 2 This is a logic diagram for rollover accident detection provided in an embodiment of the present invention;

[0063] Figure 3 This is a flowchart of road segment feature vector extraction provided in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of a long short-term memory neural network structure provided in an embodiment of the present invention;

[0065] Figure 5 A flowchart of dynamic adjustment of LTR alarm threshold based on accident frequency provided in an embodiment of the present invention;

[0066] Figure 6 A flowchart for matching feature vectors of accident-prone road sections provided in an embodiment of the present invention;

[0067] Figure 7 A flowchart of LTR alarm threshold correction based on weather conditions provided in an embodiment of the present invention;

[0068] Figure 8 A flowchart of the humidity verification mechanism provided in an embodiment of the present invention;

[0069] Figure 9 A flowchart of the LTR alarm threshold downgrade correction strategy provided in this embodiment of the invention;

[0070] Figure 10 This is a schematic diagram of a vehicle intelligent anti-rollover warning system provided in an embodiment of the present invention. Detailed Implementation

[0071] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0072] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0073] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0074] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0075] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0076] In traditional technical solutions, vehicle rollover warnings typically rely on a lateral load transfer rate (LTR) threshold pre-calibrated and embedded in the vehicle controller by the vehicle manufacturer. This threshold is a static constant, often set based on idealized standard road conditions (such as dry, smooth asphalt) and standard vehicle loading conditions.

[0077] However, in real-world road transport environments, vehicles face extremely complex and varied operating conditions. Different road sections possess different geometric characteristics, such as curve radii, superelevation (cross slope) angles, and longitudinal slope gradients; different road surfaces exhibit different adhesion conditions, such as dry asphalt roads, slippery waterlogged roads, and icy or snowy roads with extremely low adhesion; furthermore, weather conditions (such as rain, snow, and fog) and road surface moisture are constantly changing. Under the combined effect of these complex factors, a fixed LTR (Low-Temperature Reduction) alarm threshold inevitably cannot achieve accurate warnings in all scenarios.

[0078] Specifically, on dry roads with good grip, a relatively high LTR threshold is safe because the vehicle tires still have sufficient grip reserves to resist skidding. If the threshold is set too low, it will lead to frequent false alarms, interfering with the driver's normal driving, causing alarm fatigue, and weakening the driver's trust in the warning system. Conversely, on wet or snow-covered roads, tire grip drops sharply, and the vehicle may skid or even roll over before the LTR value reaches the safe upper limit corresponding to dry roads. If a high fixed threshold is used in this case, the best time to avoid danger will be missed due to the alarm being too late, leading to serious traffic accidents.

[0079] Furthermore, current warning technologies rely entirely on real-time sensor data from the vehicle at the current moment, lacking effective utilization of historical accident risks ahead or on the target road segment. In other words, each vehicle is isolated, unable to learn from accidents previously involving other vehicles on that segment. If a road segment has a history of frequent rollover accidents, this strongly suggests the presence of inherent, potentially rollover-prone hazards, such as a poorly designed sharp bend or a section with a persistently slippery surface due to water seepage. Current technologies fail to translate this historical risk information into effective warnings for passing vehicles, thus missing the opportunity to learn from data and improve overall road safety.

[0080] To address the problems in the prior art, this embodiment provides a vehicle intelligent anti-rollover warning method, such as... Figure 1 As shown, the method includes:

[0081] When a vehicle rollover accident is detected, the feature vector of the accident section is extracted.

[0082] Adjust the LTR alarm threshold of the accident section according to the frequency of accidents in the accident section;

[0083] Upload the feature vector of the accident section and the adjusted LTR alarm threshold to the cloud;

[0084] Extract the feature vector of the road segment in which vehicle 2 travels;

[0085] Based on the driving segment feature vector, an accident segment feature vector is matched from the cloud, and the LTR alarm threshold of the corresponding accident segment is used as the LTR alarm threshold of vehicle two.

[0086] Based on the LTR alarm threshold of vehicle two, a rollover prevention alarm is issued for vehicle two.

[0087] The method of this embodiment will be described in detail below.

[0088] When Vehicle 1 (i.e., the reference vehicle involved in the accident) rolls over, this embodiment immediately triggers the accident detection mechanism and extracts the feature vector of the accident section. This step is the data source of the entire cloud-based collaborative early warning mechanism, and its accuracy and reliability directly determine the effectiveness of all subsequent steps. Therefore, this embodiment sets a rigorous and multi-dimensional triggering logic for the detection and judgment of rollover accidents to avoid misjudging ordinary dangerous driving behaviors or occasional sensor noise as accidents, thereby polluting the cloud database. Specifically, this embodiment adopts a multi-condition parallel detection logic, such as... Figure 2 As shown, it includes:

[0089] (a) The LTR of the current vehicle is greater than the preset LTR threshold and the duration is greater than the preset duration threshold. In this embodiment, the corresponding setting is: LTR is greater than 1 and lasts for more than 200ms (indicating that the vehicle load transfer rate has exceeded 100% and the inner and outer wheels are about to leave the ground). This condition identifies the real rollover risk rather than instantaneous fluctuations by monitoring the continuity of lateral load transfer. LTR is the core indicator for measuring the degree of vehicle rollover danger. It is defined as the ratio of the difference between the vertical loads of the left and right wheels of the vehicle to the total vertical load. The value range is usually between -1 and 1. The closer its absolute value is to 1, the closer the vehicle is to the rollover critical point.

[0090] (b) The vehicle roll angle is greater than the preset roll angle threshold. In this embodiment, the corresponding setting is: the vehicle roll angle is greater than 45 degrees (the vehicle is severely tilted). This condition directly reflects the degree of vehicle tilt and is the most intuitive physical manifestation of rollover.

[0091] (c) The vehicle body roll rate is greater than a preset roll rate threshold, which in this embodiment is set to: a vehicle body roll rate greater than 30 degrees per second (the vehicle is rapidly rolling). The roll rate is the time derivative of the roll angle, which describes the speed of the vehicle roll. In some dynamic rollover scenarios (such as a sharp turn caused by emergency obstacle avoidance), the vehicle may lose control due to an excessive roll rate before the roll angle reaches its limit. An excessive roll rate means that the vehicle has undergone a significant change in attitude in a very short time. This dynamic instability process is often more dangerous than a slow roll and is more difficult to correct through driver operation. Therefore, using the roll rate as an independent trigger condition can effectively capture this transient instability condition.

[0092] If any of the above conditions are met, the vehicle is considered to have rolled over.

[0093] The reason for using parallel logic, where any condition being met is considered an accident, is that rollover accidents exhibit diverse forms: some accidents show a continuous exceeding of the LTR (Long-Term Ratio) limit but a gradual change in roll angle (e.g., slow rollover on a high-speed curve), while others show a sudden surge in roll angle but a lack of sustained LTR accumulation (e.g., rollover after tripping over a curb). Angular velocity conditions can capture sudden roll movements. Using serial logic (where all conditions are met simultaneously) would miss a large number of atypical rollovers; parallel logic ensures that accident confirmation is triggered as soon as any indicator of vehicle attitude, load, or kinematic parameters exceeds the safety boundary, thereby maximizing the accident detection rate and accumulating a rich sample of accidents for the cloud database.

[0094] like Figure 3 As shown, after confirming the accident, this embodiment immediately collects multi-dimensional vehicle data at the moment of the accident to record and encapsulate all relevant data within the time window from the start of the accident to its determination as an accident, including vehicle coordinates (longitude, latitude), vibration spectrum characteristics of the inertial measurement unit (IMU), peak lateral acceleration, maximum LTR value (maximum LTR value during vehicle travel), vehicle speed, steering wheel angle, braking status, weather conditions, and road surface humidity.

[0095] Among these features, precise vehicle coordinates can be obtained through the Global Navigation Satellite System (GNSS) and high-precision maps; three-axis acceleration, three-axis angular velocity, and the resulting vibration spectrum characteristics can be obtained through the Inertial Measurement Unit (IMU); peak lateral acceleration and real-time maximum LTR value can be calculated through wheel speed sensors and vehicle attitude sensors; vehicle speed can be obtained through vehicle speed sensors; steering wheel angle can be obtained through steering wheel angle sensors; braking status can be obtained through the braking system control unit; weather conditions can be obtained through vehicle networking or onboard meteorological APIs; and road surface humidity can be obtained through road surface humidity sensors.

[0096] These data were selected because they collectively constitute a complete picture of a rollover accident: vehicle coordinates are used to locate the geospatial features of the accident section; IMU vibration spectrum characteristics reflect road surface unevenness and vehicle suspension status; peak lateral acceleration and maximum LTR value are direct quantitative indicators of rollover dynamic intensity; vehicle speed and steering wheel angle reveal the driver's operating behavior and vehicle motion status; braking status reflects the impact of braking force distribution on rollover; and weather conditions and road surface humidity characterize the environmental adhesion coefficient.

[0097] Subsequently, this embodiment utilizes a pre-defined feature extraction model to extract features from the aforementioned multidimensional data, obtaining a feature vector for the accident section. Multidimensional vehicle data is often susceptible to noise interference, making it difficult to extract the most distinctive core features. However, the feature extraction model in this embodiment, trained offline on a large amount of data, can automatically learn and extract the feature vector that best characterizes the rollover risk of the road section. This feature vector is not a simple latitude and longitude coordinate, but a multidimensional data fingerprint that integrates vehicle dynamic response, road environment information, and driver operation characteristics. It can uniquely and accurately depict the dangerous state of the road section at the moment of the accident.

[0098] In this embodiment, the feature extraction model uses a long short-term memory neural network (LS segment and key feature information of vehicles on the segment TM).

[0099] LSTM was chosen because a vehicle's driving process on a road segment is essentially a time series, and its rollover risk is not only related to its state at a single instant but also closely related to the trend of its state evolution over a period of time. For example, whether a vehicle is steadily cornering or is already on the verge of losing control and sliding needs to be determined by analyzing the data change patterns within a continuous time window. LSTM networks, with their unique gating structure (input gate, forget gate, output gate), can effectively learn and remember key information from long-term time series, while overcoming the gradient vanishing or gradient exploding problems that traditional recurrent neural networks (RNNs) are prone to during training. Therefore, the LSTM-based feature extraction model can extract dynamic features that reflect the evolution trend of vehicle instability from the original time series data, making the final generated road segment feature vector more representative and discriminative, providing a reliable data foundation for subsequent high-precision similarity matching in the cloud.

[0100] like Figure 4 As shown, the Long Short-Term Memory Neural Network includes an input layer (10-dimensional time-series data, sequence length 4000 points, i.e., 40 seconds × 100Hz), a first LSTM hidden layer (containing 64 LSTM units, dropout rate 0.2), a second LSTM hidden layer (containing 64 LSTM units, dropout rate 0.2), a first fully connected layer (containing 32 neurons, ReLU activation), a second fully connected layer (containing 32 neurons, ReLU activation), and an output layer (128-dimensional, tanh activation, output range [-1, 1] for each dimension).

[0101] The input layer is used to receive preprocessed and standardized multidimensional time series data, and its input dimension matches the feature dimension of the aforementioned multidimensional vehicle data.

[0102] The first LSTM hidden layer contains several LSTM units, 64 of which are set in this embodiment. It is responsible for performing preliminary feature abstraction and time series pattern capture on the input time series data.

[0103] The second LSTM hidden layer, stacked on top of the first LSTM hidden layer, is used to perform deeper abstraction and fusion of the high-level features output by the first layer, extracting more complex temporal dependencies. By stacking two LSTM layers, the network can learn higher-level nonlinear features and temporal dynamics in the data.

[0104] After two LSTM layers, the network is connected to a first fully connected layer, which flattens and linearly transforms the high-dimensional temporal features output by the LSTM layers, mapping them to an intermediate feature space.

[0105] This is followed by a second fully connected layer, which further performs non-linear transformations and dimensionality compression on the features, ultimately mapping the features to a vector space of a preset dimension.

[0106] Finally, there is the output layer, which outputs the final feature vector, namely the feature vector of the accident section or the feature vector of the driving section.

[0107] The feature extraction model in this embodiment, namely the Long Short-Term Memory Neural Network, needs to be trained before practical application. The goal of training is to enable the model to accurately map the input multidimensional raw data to the feature space, so that data with similar rollover risk features are close together in the feature space, while data with large differences in risk features are far apart.

[0108] The training dataset contains a large amount of historically collected multidimensional vehicle data. Data preprocessing includes normalization (mean 0, standard deviation 1) and sliding window segmentation (window size 10 seconds, step size 5 seconds). The optimizer uses Adam, with the loss function being mean squared error (MSE), a batch size of 32, 100 training epochs, a validation set ratio of 20%, and 10 epochs of early stop patience settings. After training, the Long Short-Term Memory (LSTM) neural network can extract a 128-dimensional feature representation from the input time-series data. This feature vector condenses the key feature information of road segments and vehicles on those segments.

[0109] Each training sample in the training dataset is pre-assigned a 128-dimensional label vector. This label vector serves as the supervision signal and comparison standard during training. It is not generated by the Long Short-Term Memory (LSTM) neural network itself, but is pre-designed based on expert experience, domain knowledge, historical rollover accident data, real-vehicle test data, simulation calibration data, and / or industry standards. The construction of the label vector exhibits the property that data with similar rollover risk features are close in distance in the feature space, while data with significantly different risk features are far apart. Therefore, MSE training allows the output features to inherit this property.

[0110] To further enhance the model's adaptability to constantly changing road environments and vehicle conditions, the Long Short-Term Memory (LSTM) neural network in this embodiment also supports online incremental learning. That is, after the LSM neural network is trained, it will continue to collect new multi-dimensional vehicle data and fine-tune the trained LSM neural network through the incremental learning mechanism, without having to retrain the entire network from scratch.

[0111] Specifically, when new training data arrives, the network is fine-tuned based on the existing model. The new training data comes from: automatically collecting multi-dimensional vehicle data at the time of each actual rollover accident after confirmation, or collecting corresponding multi-dimensional vehicle data by driving test vehicles under standardized operating conditions on specific road sections during accident-free periods.

[0112] The fine-tuning strategy involves freezing the first three layers (to retain general feature extraction capabilities), namely the input layer, the first LSTM hidden layer, and the second LSTM hidden layer, and training only the last two fully connected layers (to adapt to new road segment features), namely the first and second fully connected layers. The fine-tuning process is repeated 10-20 times, with a learning rate of 0.0001 (an order of magnitude lower than the initial training rate). This incremental learning mechanism allows the model to continuously adapt to new road segment features while avoiding excessive forgetting of existing knowledge.

[0113] While extracting the feature vector of the accident section, this embodiment also dynamically adjusts the LTR alarm threshold according to the accident frequency of the accident section.

[0114] The LTR (Local Time Tolerance) warning threshold should not be a static constant, but rather a dynamic variable that reflects the inherent hazard level of a road segment. The underlying logic is that the accident frequency of a road segment is a comprehensive statistical result of the long-term effects of various potential hazard factors (such as road geometry design defects, pavement material aging, poor visibility, and complex traffic environments). If a road segment experiences frequent rollover accidents, this is not accidental, but strongly suggests that the segment has an extremely high risk of rollover under specific conditions (such as rainy weather). For such road segments, if the same LTR warning threshold is used as for ordinary, good road segments, then when a vehicle reaches that segment and begins to show signs of skidding, its LTR value may not have even reached the threshold before the vehicle loses control and rolls over due to insufficient road surface adhesion. Therefore, to achieve focused attention on high-risk road segments, this embodiment introduces a threshold adjustment mechanism based on accident frequency.

[0115] Specifically, such as Figure 5As shown, the cloud-based or vehicle-mounted system will count the total number of rollover accidents occurring on a certain accident section within a preset historical time window (e.g., the past year) and calculate the accident frequency accordingly. Then, this frequency is compared with a preset frequency threshold. Specifically, for road sections with a rollover accident frequency lower than the preset frequency, a first LTR alarm threshold is used; for road sections with a rollover accident frequency greater than or equal to the preset frequency, a second LTR alarm threshold is used, and the first LTR alarm threshold is greater than the second LTR alarm threshold.

[0116] For example, for road sections where only one rollover accident has occurred within 30 days, the LTR alarm threshold is set to 0.85; for road sections where the number of rollover accidents is greater than or equal to 2 within 30 days, the LTR alarm threshold is set to 0.80.

[0117] The reason for this design is that the frequency of accidents is essentially a statistical representation of the degree of danger of a road segment. For high-risk road segments with frequent accidents (such as sharp bends, steep slopes, or perpetually slippery surfaces), there are inherent defects in their road geometry or environmental conditions, resulting in a higher risk of rollover for vehicles traveling on such segments. If a high first LTR (Least Rapid Turnover) warning threshold is used, the warning will be triggered too late, and drivers will lose valuable time to avoid danger. Therefore, a lower second LTR warning threshold is used for such road segments to achieve early warning and intervention, compensating for the high-risk nature of the road segment by lowering the trigger threshold.

[0118] Conversely, for low-accident road sections with relatively good road conditions, rollovers are mostly caused by extreme driver maneuvers. In these cases, using a higher first LTR (Low-Temperature Reduction) warning threshold can avoid excessive warnings for normal driving behaviors (such as high-speed lane changes), reduce false alarms that interfere with the driver, and lower the probability of overlooking genuine risks due to warning fatigue. This frequency-based hierarchical threshold strategy achieves adaptive matching between warning sensitivity and road section risk level, optimizing warning accuracy while ensuring safety.

[0119] By setting this differentiated threshold, this embodiment achieves an optimal balance between safety and comfort. More importantly, this adjusted threshold does not only serve the vehicle involved in the accident; it serves as a digital label for the hazard level of that road segment, and is uploaded to the cloud along with the feature vector of the accident segment, becoming a warning basis for all subsequent vehicles passing through that segment (such as vehicle two). This is equivalent to transforming an accident into a reusable, quantifiable safety parameter, realizing the accumulation and transmission of safety knowledge.

[0120] In this embodiment, after extracting the feature vector of the accident section and adjusting the LTR alarm threshold, the feature vector of the accident section and the adjusted LTR alarm threshold are further uploaded to the cloud. The cloud, as a central data management platform, is responsible for storing the feature vectors of accident sections from numerous vehicles and their corresponding LTR alarm thresholds for those accident sections. If a newly uploaded LTR alarm threshold for a certain section already exists in the cloud, the existing LTR alarm threshold in the cloud will be overwritten and updated by the newly uploaded LTR alarm threshold.

[0121] To achieve efficient and reliable data transmission, in this embodiment, the vehicle's onboard terminal establishes a secure data connection with the cloud server via an onboard communication module (such as 4G / 5G cellular network, C-V2X vehicle-to-everything communication, etc.). The uploaded feature vector can be shared with all vehicles communicating with the cloud. In this embodiment, the feature vector is stored in the cloud platform in the following format: Vector ID (road segment GPS coordinate hash value, 8 bytes), 128-dimensional floating-point array (512 bytes), generation timestamp (8 bytes), confidence level (4 bytes, 0-1), data source vehicle ID (4 bytes), and validity period (4 bytes, default 90 days), with each record approximately 540 bytes.

[0122] In this embodiment, the cloud platform is also used for continuous data management of the feature vectors of each accident section and their corresponding LTR alarm thresholds. Specifically, the management includes:

[0123] If the frequency of accidents on any accident road segment is 0 within a preset continuous period of time (e.g., 90 consecutive days), the threshold increment mechanism is activated: the LTR alarm threshold of the corresponding accident road segment is incremented by a preset increment value (2%) every preset period of time (e.g., every week).

[0124] The technical logic behind this mechanism is that a zero accident frequency indicates that the hazards on that road section may have been eliminated (e.g., completed curve modifications, road surface repairs, improved drainage systems, etc.), or that previous warnings have effectively prompted drivers to change their driving behavior. Maintaining the original low threshold at this point would lead to unnecessary frequent warnings. By progressively increasing the threshold, this embodiment can tentatively reduce the warning sensitivity until it returns to the standard level or is readjusted when an accident occurs again. This forgetting mechanism avoids the threshold rigidity caused by the accumulation of historical data in the cloud database, giving this embodiment dynamic evolution capabilities, allowing it to adapt to improvements in road infrastructure and changes in the traffic environment.

[0125] This embodiment breaks down the information silos of traditional in-vehicle systems, enabling the safety experience of each vehicle to be fed into a big data pool in the cloud in real time. The richer the data in the cloud, the more comprehensive the understanding of road risks, and the more accurate the warning services provided to other vehicles. This constitutes a virtuous cycle of data ecosystem, allowing the intelligence level of the method in this embodiment to continuously evolve with the accumulation of data.

[0126] For vehicle two, which is currently in motion, this embodiment extracts the feature vector of its current travel segment in real time. Vehicle two is connected to the cloud and can exchange data with it. Typically, vehicle two and vehicle one belong to the same fleet and are connected to the same cloud. This step is a prerequisite for vehicle two to initiate a query request to the cloud. Its purpose is to transform the travel segment where vehicle two is currently in an unknown risk level into a standardized format that can be compared with the known risk features in the cloud.

[0127] The method for extracting feature vectors for driving sections is consistent with that for accident sections. Specifically, it involves collecting multi-dimensional vehicle data of the same dimensions (vehicle coordinates, IMU vibration spectrum features, peak lateral acceleration, maximum LTR value, vehicle speed, steering wheel angle, braking status, weather conditions, and road surface humidity), and compressing these features using the same feature extraction model. Maintaining consistency in feature extraction ensures that the feature vectors for driving sections and accident sections reside in the same vector space, thereby guaranteeing the effectiveness of subsequent similarity calculations.

[0128] Subsequently, this embodiment matches the most similar accident segment feature vector from the cloud based on the feature vector of the driving segment. Specifically, as follows: Figure 6 As shown, for each accident road segment feature vector in the cloud, its cosine similarity with the driving road segment feature vector is calculated. Among all the calculated cosine similarities, candidate vectors with a similarity greater than the preset similarity threshold are selected, and the accident road segment feature vector with the largest cosine similarity value is selected as the final matching result. At the same time, the LTR alarm threshold corresponding to the accident road segment is used as the current LTR alarm threshold of vehicle two.

[0129] The reason for using cosine similarity is that it focuses on the angle between vector directions rather than the absolute value, effectively eliminating scale differences caused by data with different dimensions (such as coordinate values ​​and angle values), and concentrating on the similarity of road features and operating conditions. The dual mechanism of setting a similarity threshold and taking the maximum value ensures both accuracy and conservatism in matching: threshold filtering excludes historical accident samples that differ too much from the current road segment (such as distinguishing between straight road segments and winding mountain roads), avoiding false alarms caused by incorrectly applying high-risk thresholds; taking the maximum value among qualified candidates means selecting the accident case from historical data that best matches the current road conditions and has the most reference value, ensuring that the selected threshold comes from the most similar historical scenario and maximizing the rationality of threshold transplantation.

[0130] Through this matching mechanism, Vehicle 2 can inherit historical experience and lessons learned from similar road sections without accumulating its own accident data, thus achieving cross-vehicle and cross-time-space safety knowledge sharing.

[0131] It should be noted that if a matching accident segment feature vector cannot be found in the cloud, the default LTR alarm threshold will be used as the current LTR alarm threshold for vehicle two.

[0132] By using this threshold acquisition method based on similarity matching, even if vehicle 2 is driving on a new road segment where no accidents have ever occurred, as long as the characteristics of the road segment are highly similar to a historical accident road segment, it can learn from the risk level of that historical road segment and obtain a scientific and reasonable warning threshold, thus truly realizing the generalization capability of risk warning.

[0133] After obtaining the LTR alarm threshold matched in the cloud, this embodiment does not directly use the threshold to issue an alarm. Instead, it further corrects the threshold according to the weather conditions of the current driving section and uses the corrected LTR alarm threshold to issue an anti-rollover alarm for vehicle two.

[0134] Specifically, such as Figure 7 As shown, in this embodiment, the meteorological API interface first obtains the weather conditions (such as sunny, rainy, snowy, foggy, etc.) of the current driving segment via HTTP / HTTPS protocol, updating every 15 minutes. Then, based on the weather conditions, the corresponding adhesion coefficient is read from a preset first weather-adhesion coefficient mapping table, and this adhesion coefficient is used to correct the LTR alarm threshold of vehicle two. This mapping table is obtained in advance through extensive experiments or simulations, and it quantifies the adhesion capability of typical road surfaces under different weather conditions.

[0135] The underlying technical logic is that weather conditions directly affect the coefficient of friction between the tire and the road surface, thus influencing the vehicle's rollover stability boundary. In low-friction conditions such as rain or snow, tire grip decreases, making vehicles more prone to skidding and rollover. In these situations, the LTR (Label Traction Rate) warning threshold needs to be lowered for earlier warnings. Conversely, in sunny, high-friction conditions, the vehicle has a larger stability margin, allowing for a more appropriate increase in the threshold to reduce interference. By introducing a weather-friction coefficient mapping table, this embodiment transforms macroscopic meteorological information into microscopic dynamic parameter corrections, achieving a quantitative bridge from environmental perception to safety decision-making.

[0136] For example, the specific values ​​of the first weather-adhesion coefficient mapping table are:

[0137] Sunny day k_weather=1.0 (baseline value, LTR alarm threshold is not adjusted);

[0138] Cloudy k_weather=0.95 (LTR alarm threshold reduced by 5%);

[0139] Xiaoyu k_weather=0.85 (LTR alarm threshold reduced by 15%);

[0140] Moderate rain k_weather=0.78 (LTR alarm threshold reduced by 22%);

[0141] Heavy rain k_weather=0.73 (LTR alarm threshold reduced by 27%);

[0142] Xiaoxue k_weather=0.75 (LTR alarm threshold reduced by 25%);

[0143] Heavy snow k_weather=0.65 (LTR alarm threshold reduced by 35%);

[0144] Freeze k_weather=0.58 (LTR alarm threshold reduced by 42%);

[0145] Heavy fog k_weather=0.80 (LTR alarm threshold reduced by 20%).

[0146] The corrected formula is:

[0147] LTR_th_weather = k_weather × LTR_th_base.

[0148] LTR_th_weather represents the corrected LTR alarm threshold; k_weather represents the adhesion coefficient; LTR_th_base represents the original LTR alarm threshold.

[0149] Furthermore, to avoid correction deviations caused by weather judgment errors in a single mapping table, this invention also introduces a humidity verification mechanism. Specifically, such as... Figure 8 As shown, it includes:

[0150] Get the humidity of the current driving route;

[0151] Obtain the humidity range corresponding to the weather conditions of the current driving route;

[0152] Determine whether the humidity of the current driving section is within the humidity range corresponding to the weather conditions of the current driving section;

[0153] When the humidity of the current driving segment is within the humidity range corresponding to the weather conditions of the current driving segment, the preset first weather-adhesion coefficient mapping table is used.

[0154] When the humidity of the current driving segment is outside the humidity range corresponding to the weather conditions of the current driving segment, a preset second weather-adhesion coefficient mapping table is used; compared with the first weather-adhesion coefficient mapping table, the adhesion coefficient under each weather condition in the second weather-adhesion coefficient mapping table is reduced by a preset ratio.

[0155] For example, when a vehicle enters a certain road section, this embodiment queries the meteorological API to obtain the current weather conditions and obtains the actual humidity through the road surface humidity sensor. If the two are consistent (e.g., the API shows light rain and humidity > 60%), the preset first weather-adhesion coefficient mapping table is directly used. If the two are inconsistent (e.g., the API shows sunny but humidity > 40%, and humidity on sunny days is usually less than 40%), the preset second weather-adhesion coefficient mapping table is used.

[0156] In the second weather-adhesion coefficient mapping table, compared to the first weather-adhesion coefficient mapping table, the adhesion coefficients under each weather condition are reduced by a preset ratio. This means that when weather and humidity data contradict each other, the method in this embodiment tends to conservatively estimate a lower adhesion coefficient, thereby further reducing the LTR alarm threshold and addressing data uncertainty with a "higher rather than lower" safety principle. This dual verification mechanism effectively solves the problem of threshold correction distortion caused by meteorological sensor errors or local microclimate anomalies, avoids the risk of missed reports that may be caused by relying on a single information source, and improves the robustness of this embodiment in complex environments.

[0157] Furthermore, in extreme situations where weather conditions for the current driving route cannot be obtained (e.g., the vehicle is driving in a remote area without network coverage, or the onboard weather sensor malfunctions), this embodiment activates a degradation strategy, such as... Figure 9 As shown, it includes:

[0158] Directly obtain humidity data for the current driving route;

[0159] The adhesion coefficient is read from a preset humidity-adhesion coefficient mapping table based on the current humidity. The adhesion coefficient is used to correct the LTR alarm threshold of vehicle two.

[0160] Since the road surface humidity sensor is a local on-board sensor and does not rely on external communication, it can still perform basic threshold correction even in offline mode. Humidity was chosen as the degradation criterion because humidity and road surface adhesion coefficient are strongly correlated (higher humidity means a wetter and slipperier road surface, and a lower adhesion coefficient), and humidity data is relatively easy to obtain reliably through on-board sensors (such as windshield humidity sensors or wheel speed difference estimation). This degradation mechanism ensures that this embodiment still possesses basic environmental adaptability even when the weather perception function fails, avoiding alarm function degradation due to the failure of a single data source.

[0161] The humidity-adhesion coefficient mapping table establishes a direct correlation between road surface humidity and the coefficient of adhesion. The underlying physics is that road surface humidity is one of the most direct factors affecting tire-road adhesion. As humidity increases, a water film gradually forms and thickens on the road surface, reducing the contact area between the tire and the road, and significantly decreasing friction. Therefore, a relatively reasonable coefficient of adhesion can be directly estimated using the humidity value. For example, the specific values ​​in the humidity-adhesion coefficient mapping table are:

[0162] When the humidity is less than 30%, k_weather=1.0;

[0163] When the humidity is greater than or equal to 30% and less than or equal to 60%, k_weather=0.85;

[0164] When the humidity is greater than 60%, k_weather=0.78.

[0165] After completing the above threshold determination and correction, the method of this embodiment enters the real-time alarm stage.

[0166] Specifically, the on-board computing unit calculates the current LTR (Lateral Load Transfer Rate) of vehicle 2 in real time. This value is usually estimated by the vehicle's lateral acceleration, suspension displacement or wheel load, and is used to quantify the degree of vertical load transfer on the left and right wheels of the vehicle.

[0167] If the calculated current LTR value is greater than the current LTR alarm threshold of vehicle 2 (i.e., the dynamic threshold after cloud matching and weather correction), it indicates that the vehicle has entered a dangerous state and immediately triggers the rollover prevention alarm.

[0168] Warning methods may include flashing dashboard warning lights, buzzer alarms, seat vibration, or automatic deceleration suggestions. These multimodal warning methods are designed to attract the driver's attention in the most direct and effective way, prompting the driver to take the correct evasive action immediately, such as releasing the accelerator pedal, lightly applying the brakes, or correcting the steering direction.

[0169] This embodiment compares the real-time LTR value with a dynamic threshold that has been optimized through both historical experience and real-time correction. This embodiment ensures that the alarm trigger point is neither too early (avoiding false alarms) nor too late (avoiding missed alarms). It can provide the driver with valuable warning time before the critical point of vehicle loss of control, thereby minimizing the probability of rollover accidents and protecting the lives and property of passengers.

[0170] The reason for choosing LTR as the final warning criterion is that LTR is a direct physical quantity of rollover risk. A value of 0.5 typically corresponds to one wheel starting to lift off the ground, and a value close to 1.0 indicates the vehicle is about to completely roll over. By comparing the dynamic threshold with the real-time LTR value, this embodiment can intervene in time at the critical moment before a rollover occurs, providing the driver with sufficient reaction time, thereby effectively avoiding rollover accidents or mitigating their consequences.

[0171] In summary, this embodiment has the following technical effects:

[0172] (1) This embodiment realizes the sharing of knowledge about rollover risks across vehicle road sections, breaking down the data silos of single-vehicle perception:

[0173] By extracting road segment feature vectors from accident vehicles and uploading them to the cloud, subsequent vehicles passing through similar road segments can reuse the LTR alarm thresholds of historical accident road segments based on vector similarity matching. High-risk road segments can be identified without a single vehicle being involved in an accident. A single rollover accident is transformed into reusable road segment risk quantification parameters, forming a closed loop of continuous accumulation and iteration of safety data within the fleet, thereby improving the fleet's ability to generalize warnings of unknown similar dangerous road segments.

[0174] (2) This embodiment adopts a multi-condition parallel rollover accident judgment logic, which takes into account both the accident detection rate and the ability to prevent false judgment:

[0175] The system employs a parallel mechanism that determines an accident based on the fulfillment of any one of the following conditions: LTR (Long-Term Reversal) exceeding limits, roll angle exceeding limits, or roll rate exceeding limits. This mechanism covers various rollover accident scenarios, including gradual rollovers on gentle curves, rollovers caused by tripping over curbs, and transient instability during emergency turns. It avoids missing atypical rollovers due to serial judgment logic. At the same time, it sets constraints such as duration thresholds to filter out instantaneous noise from sensors and normal aggressive driving behaviors, preventing the accidental recording of non-accident data into the cloud database and ensuring the reliability of cloud samples.

[0176] (3) This embodiment uses an LSTM time series model for feature extraction, which improves the accuracy of road segment similarity matching:

[0177] This embodiment utilizes an LSTM temporal model to mine the temporal dynamic features of vehicle driving, capturing the dynamic trend of rollover risk evolving over time and overcoming the gradient vanishing problem of traditional RNNs. The model outputs multi-dimensional feature vectors, rather than simple GPS coordinates, transforming road segment risks into a unified vector space. Combined with cosine similarity matching, it accurately identifies high-risk road segments with similar features but different geographical locations. The model supports an incremental learning mechanism, freezing the underlying general temporal feature layer and only fine-tuning the upper fully connected layers. New training data can update the model without requiring complete retraining, reducing model iteration costs and continuously adapting to the ever-changing road environment.

[0178] (4) This embodiment sets the LTR basic threshold based on the historical accident frequency of the road segment, balancing alarm sensitivity and false alarm rate, so that the alarm threshold can dynamically evolve with the risk level of the road segment:

[0179] Lower LTR (Low Trial Rate) alarm thresholds are used for high-frequency accident sections to provide early warnings and compensate for the high rollover risk caused by inherent road defects. Higher thresholds are used for low-frequency accident sections to reduce invalid alarms in normal aggressive driving scenarios and prevent driver fatigue from alarms. A cloud-based threshold forgetting mechanism is also included: when a road section has no accidents for a long period of time, the alarm threshold is gradually increased to adapt to risk attenuation scenarios such as road reconstruction and pavement repair, avoiding the permanent solidification of thresholds based on historical high-risk data and solving the problem that static thresholds cannot keep up with the evolution of road risks.

[0180] (5) This embodiment adopts a cloud feature matching mechanism based on cosine similarity to realize the transfer of historical accident knowledge across vehicles, enabling new vehicles or vehicles on unfamiliar road sections to inherit safety experience in similar scenarios:

[0181] Cosine similarity can effectively eliminate scale differences caused by data with different dimensions (such as coordinate values ​​and angle values), focusing on the similarity of road features and operating conditions.

[0182] (6) This embodiment modifies the LTR alarm threshold based on weather conditions and road surface humidity, ensuring that the accuracy and robustness of the alarm remain even under complex meteorological conditions and sensor failures:

[0183] By leveraging meteorological APIs to obtain weather data, environmental corrections are applied to the base thresholds using a weather-adhesion coefficient mapping table. Alarm thresholds are automatically lowered in low-adhesion scenarios such as rain, snow, and freezing. A road surface humidity sensor is added for cross-validation. When weather forecasts and measured humidity differ, the system automatically switches to a more conservative second mapping table, mitigating threshold distortion caused by errors from a single meteorological data source and local microclimate conditions. An offline degradation strategy is designed to directly estimate the adhesion coefficient and correct thresholds in scenarios without network access or where meteorological sensors fail, ensuring the alarm function continues to operate normally even in communication failure environments.

[0184] (7) In this embodiment, the real-time LTR is compared with the dynamic threshold after cloud matching and environmental correction. Multimodal alarm is triggered before the vehicle reaches the critical state of rollover, and the time for the driver to take evasive action is reserved. Compared with fixed threshold alarm, the alarm threshold can be dynamically adjusted according to the inherent risk of the road section and the real-time road surface adhesion conditions, reducing missed alarms, reducing the probability of rollover accidents, and mitigating accident damage.

[0185] This embodiment also provides a vehicle intelligent rollover prevention alarm system, using the above-described vehicle intelligent rollover prevention alarm method, such as... Figure 10 As shown, the system includes:

[0186] The feature extraction module is used to extract feature vectors of accident sections and driving sections;

[0187] The threshold adjustment module is used to adjust the LTR alarm threshold of the accident section according to the frequency of accidents in the accident section;

[0188] The matching module is used to match an accident road segment feature vector from the cloud based on the driving road segment feature vector, and use the LTR alarm threshold of the corresponding accident road segment as the LTR alarm threshold of the vehicle.

[0189] The alarm module is used to issue rollover prevention alarms based on the vehicle's LTR alarm threshold.

[0190] This embodiment also provides a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions that the processor can execute. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the vehicle intelligent anti-rollover alarm method described above.

[0191] This embodiment also provides a computer-readable storage medium storing a computer program. When the computer program is run by a processor, it performs the steps of the vehicle intelligent anti-rollover alarm method described above.

[0192] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A vehicle intelligent anti-rollover warning method, characterized in that, The method includes: When a vehicle rollover accident is detected, the feature vector of the accident section is extracted. Adjust the LTR alarm threshold of the accident section according to the frequency of accidents in the accident section; Upload the feature vector of the accident section and the adjusted LTR alarm threshold to the cloud; Extract the feature vector of the road segment in which vehicle 2 travels; Based on the driving segment feature vector, an accident segment feature vector is matched from the cloud, and the LTR alarm threshold of the corresponding accident segment is used as the LTR alarm threshold of vehicle two. Based on the LTR alarm threshold of vehicle two, a rollover prevention alarm is issued for vehicle two.

2. The intelligent anti-rollover warning method for vehicles according to claim 1, characterized in that, When a vehicle rollover accident is detected, the feature vector of the accident section is extracted, including: (a) The LTR of the current vehicle is greater than the preset LTR threshold and the duration is greater than the preset duration threshold. (b) The vehicle roll angle is greater than the preset roll angle threshold; (c) The vehicle body roll rate is greater than the preset angular velocity threshold; The occurrence of a rollover accident is considered to be the fulfillment of any of the above conditions.

3. The intelligent vehicle rollover prevention alarm method according to claim 1, characterized in that, The methods for extracting the feature vectors of accident sections or driving sections include: Collect multi-dimensional vehicle data; The multidimensional vehicle data is used to extract features using a preset feature extraction model to obtain feature vectors of accident sections or driving sections.

4. The intelligent vehicle rollover prevention alarm method according to claim 3, characterized in that, The multidimensional vehicle data includes: vehicle coordinates, vibration spectrum characteristics of the inertial measurement unit, peak lateral acceleration, maximum LTR value, vehicle speed, steering wheel angle, braking status, weather conditions, and road surface humidity.

5. A vehicle intelligent anti-rollover warning method according to claim 3 or 4, characterized in that, The feature extraction model employs a long short-term memory neural network.

6. The intelligent anti-rollover warning method for vehicles according to claim 5, characterized in that, The long short-term memory neural network includes an input layer, a first LSTM hidden layer, a second LSTM hidden layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence.

7. A vehicle intelligent anti-rollover warning method according to claim 6, characterized in that, After training, the long short-term memory neural network is fine-tuned through an incremental learning mechanism.

8. The intelligent vehicle rollover prevention alarm method according to claim 1, characterized in that, The adjustment of the LTR alarm threshold for accident-prone road sections based on the accident frequency includes: For road sections where the frequency of rollover accidents is less than the preset frequency, the first LTR alarm threshold is used; For road sections where the frequency of rollover accidents is greater than or equal to a preset frequency, a second LTR alarm threshold is used; the first LTR alarm threshold is greater than the second LTR alarm threshold.

9. A vehicle intelligent anti-rollover alarm method according to claim 1, characterized in that, The step of matching an accident road segment feature vector from the cloud based on the driving road segment feature vector includes: For each accident road segment feature vector in the cloud, calculate their cosine similarity with the driving road segment feature vector; The feature vector of the accident road segment with the largest cosine similarity greater than the preset similarity threshold is taken as the feature vector of the matched accident road segment.

10. A vehicle intelligent anti-rollover warning method according to claim 1, characterized in that, Before issuing a rollover prevention warning for vehicle two based on the LTR warning threshold of vehicle two, the method further includes: correcting the LTR warning threshold of vehicle two according to the weather conditions of the current road segment, and issuing a rollover prevention warning for vehicle two using the corrected LTR warning threshold.

11. A vehicle intelligent anti-rollover warning method according to claim 10, characterized in that, The method for correcting the LTR alarm threshold of vehicle two based on the weather conditions of the current road segment includes: Get the weather conditions for the current driving route; The adhesion coefficient is read from the preset first weather-adhesion coefficient mapping table according to the current weather conditions. The adhesion coefficient is used to correct the LTR alarm threshold of vehicle two.

12. A vehicle intelligent anti-rollover alarm method according to claim 11, characterized in that, The correction method further includes: Get the humidity of the current driving route; Obtain the humidity range corresponding to the weather conditions of the current driving route; Determine whether the humidity of the current driving section is within the humidity range corresponding to the weather conditions of the current driving section; When the humidity of the current driving segment is within the humidity range corresponding to the weather conditions of the current driving segment, the preset first weather-adhesion coefficient mapping table is used. When the humidity of the current driving segment is outside the humidity range corresponding to the weather conditions of the current driving segment, a preset second weather-adhesion coefficient mapping table is used; compared with the first weather-adhesion coefficient mapping table, the adhesion coefficient under each weather condition in the second weather-adhesion coefficient mapping table is reduced by a preset ratio.

13. A vehicle intelligent anti-rollover warning method according to claim 11 or 12, characterized in that, The correction method further includes: If weather conditions for the current driving route cannot be obtained, obtain the humidity of the current driving route. The adhesion coefficient is read from a preset humidity-adhesion coefficient mapping table based on the current humidity. The adhesion coefficient is used to correct the LTR alarm threshold of vehicle two.

14. The intelligent anti-rollover warning method for vehicles according to claim 1, characterized in that, The step of issuing a rollover prevention alarm for vehicle two based on the LTR alarm threshold of vehicle two includes: Calculate the current LTR value of vehicle two; If the current LTR value is greater than the LTR alarm threshold of vehicle two, then an anti-rollover alarm will be issued for vehicle two.

15. A vehicle intelligent anti-rollover warning method according to claim 1, characterized in that, The cloud-based system is used to manage the feature vectors of each accident-prone road segment and their corresponding LTR alarm thresholds. This management includes: If the accident frequency of any accident section is 0 within a preset continuous time period, the LTR alarm threshold of the corresponding accident section will be increased by a preset increment value every preset time period.

16. A vehicle intelligent rollover prevention warning system, using the vehicle intelligent rollover prevention warning method according to any one of claims 1-15, characterized in that, The system includes: The feature extraction module is used to extract feature vectors of accident sections and driving sections; The threshold adjustment module is used to adjust the LTR alarm threshold of the accident section according to the frequency of accidents in the accident section; The matching module is used to match an accident road segment feature vector from the cloud based on the driving road segment feature vector, and use the LTR alarm threshold of the corresponding accident road segment as the LTR alarm threshold of the vehicle. The alarm module is used to issue rollover prevention alarms based on the vehicle's LTR alarm threshold.

17. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the vehicle intelligent rollover prevention alarm method as described in any one of claims 1 to 15.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the vehicle intelligent rollover prevention alarm method as described in any one of claims 1 to 15.

Images