Vehicle stability evaluation method, system, medium and vehicle based on GPS-INS device

Abstract

The application discloses a vehicle stability evaluation method and system based on a GPS / INS device, a medium and a vehicle. The method comprises the following steps: collecting multi-dimensional time sequence state parameters of the vehicle through the GPS / INS device, calculating a physical instability index, and merging the original parameters to form an input feature sequence; inputting the input feature sequence into a time sequence prediction model, and outputting the stability level of the vehicle at each time. The time sequence prediction model adopts a Transformer encoder to extract time sequence features, and uses a ordinal regression prediction head to model the ordered structure of the level. The model training adopts a hybrid supervised learning framework, uses the maximum level label of the segment level to supervise the prediction at each time, and uses a coarse-to-fine multi-stage progressive training strategy to gradually refine the level distinction. The application converts subjective feelings into objective quantitative indicators, can be applied to ESP calibration acceptance and automatic driving stability boundary monitoring, and realizes effective prediction from the rough label of the segment level to the fine level at each time.

Description

Technical Field

[0001] This application belongs to the field of vehicle dynamic performance quantitative evaluation technology, and in particular relates to a vehicle stability evaluation method, system, medium and vehicle based on GPS-INS equipment. Background Technology

[0002] In vehicle chassis development and intelligent driving systems, evaluating vehicle driving stability is a core aspect. Currently, the calibration and acceptance of ESP systems primarily rely on subjective evaluation. The specific process is as follows: after the calibration engineer completes the initial calibration of the ESP parameters, the acceptance engineer performs standard test conditions such as single lane change, double lane change (moose test), circular maneuver, and J-turn at the test track, and conducts a qualitative evaluation based on subjective feelings during the driving process. If the acceptance engineer believes that "the vehicle attitude is controllable," "there is no unacceptable fishtailing," or "the steering response is linear," it is judged as qualified; otherwise, it needs to be returned to the calibration engineer to readjust the parameters, entering a new cycle of "calibration-acceptance-recalibration." This model suffers from the following fundamental flaws: Evaluation criteria are subjective and vary from person to person. Different acceptance engineers have different thresholds for perceiving "stability." Some engineers prefer sensitive handling response and can accept a certain degree of drift, while others prefer an absolutely stable vehicle posture and are cautious about any lateral slippage. This difference in preference leads to drastically different conclusions for the same vehicle, making it impossible for calibration engineers to obtain definite optimization goals and often resulting in inefficiency. Furthermore, there is a lack of quantifiable acceptance criteria. Currently, there is no universally accepted, quantifiable standard for classifying stability levels, making it impossible to provide clear design indicators such as "under the 80km / h moose test, the maximum lateral stability level of the vehicle must not exceed level 5." This ambiguity in design goals and acceptance criteria hinders horizontal comparisons between different models and development stages.

[0003] In some engineering practices, fixed threshold methods based on physical parameters are used for objective evaluation. For example, hard thresholds are set such as "lateral acceleration must not exceed 0.3g" or "yaw rate must not exceed 0.2 rad / s". When the sensor reading exceeds the threshold, instability is determined. However, this method has significant technical limitations: fixed thresholds cannot adapt to complex and changing driving conditions. The adhesion limit of a vehicle on icy and snowy roads differs by several times from that on dry asphalt roads, and the response of a vehicle to the same angular input varies drastically at different speeds. Fixed thresholds also cannot adapt to the differences in dynamic characteristics of different vehicle models. Due to differences in parameters such as center of gravity height, wheelbase, and suspension stiffness, the stability boundaries of sports cars, SUVs, and commercial trucks vary greatly. In addition, fixed thresholds ignore the multi-parameter coupling effect. Vehicle instability is often the result of the combined effect of multiple parameters. Single-parameter threshold methods cannot capture this coupling relationship and can only output a binary judgment of "exceeding the limit / not exceeding the limit", which cannot reflect the gradual process of stability change.

[0004] In recent years, a few studies have attempted to use machine learning methods for vehicle stability assessment, such as using support vector machines or long short-term memory networks to classify vehicle states, thus shifting from manual rules to data-driven approaches. However, most existing methods are supervised learning, requiring precise time-by-time level labels for training. In practical engineering, accurately labeling the stability level at every moment during vehicle operation is extremely difficult—engineers can reliably determine that "the overall performance of this double lane change test is not good, and the most unstable point is around level 5," but they cannot precisely point out that "from 3.25 seconds to 3.78 seconds is level 5, and from 3.79 seconds to 4.12 seconds is level 6," leading to high data labeling costs and difficulty in guaranteeing labeling quality. Furthermore, most existing methods treat different stability levels as independent categories, using standard cross-entropy classification loss, ignoring the inherent ordinal relationships between levels—in reality, stability levels constitute a clear ordered progressive structure, with the similarity between adjacent levels being much greater than the similarity across levels. Ignoring this ordinal structure results in the loss of important prior information. Furthermore, existing methods fail to fully utilize prior physical knowledge of vehicle dynamics. Purely data-driven models require a large number of samples to learn physical laws, while data collection under extreme conditions is inherently risky and costly, resulting in a scarcity of extreme condition samples.

[0005] In the field of autonomous driving, existing technologies primarily employ vehicle dynamics models combined with state estimators (such as Kalman filters) to predict the future state of a vehicle and compare it with predefined safety boundaries. Typical safety boundaries include tire force boundaries based on the friction circle and stability boundaries based on the phase plane method using a centroid sideslip angle-yaw rate phase diagram. However, accurate vehicle dynamics models require a large number of physical parameters (such as tire sideslip stiffness, moment of inertia, and centroid position). Accurate acquisition of these parameters requires specialized bench testing, and these parameters drift in real time with tire wear and load changes. Furthermore, the simplified handling of unmodeled dynamics in the model results in limited boundary prediction accuracy, and the model parameters differ significantly between different vehicle models. Migrating from one model to another requires remodeling and recalibration, lacking universality.

[0006] In summary, the existing technology mainly faces the following technical challenges: (1) There is a lack of objective and quantitative evaluation standards, and subjective evaluations vary from person to person and are not reproducible; (2) It lacks the ability to make detailed evaluations at each time step. The fixed threshold method can only provide coarse-grained binarization judgments. (3) There is a bottleneck in data annotation. Accurate time-by-time annotation is extremely costly and difficult to guarantee quality. Traditional supervised learning methods are difficult to apply directly. (4) The model has insufficient ability to distinguish between different levels. Under the weak supervision setting with only fragment-level coarse labels, the existing methods are difficult to extract fine level distinction information from the coarse-grained supervision signal. Summary of the Invention

[0007] To address the aforementioned technical challenges, this invention provides a vehicle stability assessment method based on GPS-INS devices. By organically combining physical prior feature construction, ordinal regression prediction head, weakly supervised multi-instance learning, and a coarse-to-fine progressive training strategy, it achieves effective learning from fragment-level labels to time-by-time stability levels, providing an objective, precise, and reproducible quantitative solution for vehicle stability assessment.

[0008] This application achieves the above objective through the following technical solution: a vehicle stability assessment method based on GPS-INS equipment, comprising the following steps: Data acquisition steps: Collect multi-dimensional time-series state parameters of the vehicle during driving using a GPS-INS device at a predetermined sampling frequency; the multi-dimensional time-series state parameters include at least three-axis acceleration, three-axis angular velocity, sideslip angle, velocity components, and attitude angle; Feature construction steps: Based on the multidimensional time-series state parameters, calculate the physical instability index that characterizes the current instability of the vehicle, and merge the physical instability index with the multidimensional time-series state parameters to form an input feature sequence; Model prediction steps: The input feature sequence is input into a trained time-series prediction model, which outputs the stability level of the vehicle at each time step. The stability level is used to quantitatively evaluate the dynamic stability of the vehicle.

[0009] Furthermore, the multidimensional time-series state parameters include: longitudinal acceleration, lateral acceleration, vertical acceleration, roll rate, pitch rate, yaw rate, sideslip angle, northward velocity, eastward velocity, roll angle, pitch angle, yaw angle, and physical instability index.

[0010] Furthermore, the formula for calculating the physical instability index is as follows: ; ; in, , , and These are the values ​​of the absolute value of lateral acceleration, yaw rate, sideslip angle, and roll rate after global z-score normalization. The mean lateral acceleration. The standard deviation of lateral acceleration; , , , These are the weighting coefficients.

[0011] Furthermore, the time-series prediction model includes: A temporal encoder is used to extract features from the input feature sequence, capture the long-distance dependency between any two time points in the sequence, and output the hidden state vector at each time point. An ordinal regression prediction head is used to predict the probability distribution of stability level at each time step based on the hidden state vector at that time step. The ordinal regression prediction head uses a cumulative linking framework to model the ordered relationship between the levels.

[0012] Furthermore, the time-series encoder employs a Transformer encoder, which uses a self-attention mechanism to capture long-distance dependencies in the time-series data.

[0013] Furthermore, the ordinal regression prediction head calculates the rank probability at each time step in the following manner: For each time t, predict K-1 cumulative probabilities: ; in, For the sigmoid function, Let be the hidden state vector at time t, and w be the projection vector. For ordered bias thresholds; The probability of each level k The cumulative probability is obtained by performing a difference operation. ; Expectation level at this moment We obtain the result by weighted summation of the probabilities at each level: .

[0014] Furthermore, the time-series prediction model is trained using a weakly supervised learning strategy, which utilizes fragment-level label supervision signals to learn time-by-time stability level predictions.

[0015] Furthermore, the stability grade is used for at least one of the following purposes: Output objective quantitative evaluation indicators to the vehicle electronic stability program calibration system as the criteria for calibration acceptance; Provide stability boundary constraints to the autonomous driving decision-making system, and when the predicted stability level exceeds a predetermined safety threshold, suppress dangerous driving operations or perform deceleration and path correction.

[0016] Another objective of this application is to provide a vehicle stability assessment system, comprising: The data acquisition module is used to collect multi-dimensional time-series status parameters of the vehicle via GPS-INS devices; The feature processing module is used to calculate the physical instability index based on the multidimensional temporal state parameters, and merge the physical instability index with the multidimensional temporal state parameters to construct the input feature sequence; Storage module for storing trained time-series prediction models; The evaluation output module is used to input the input feature sequence into the time series prediction model to obtain and output the stability level of the vehicle at each time step.

[0017] Another object of this application is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0018] The final objective of this application is to provide a vehicle comprising: GPS-INS devices are used to collect multi-dimensional time-series status parameters of vehicles; An electronic control unit, communicatively connected to the GPS-INS device, is configured to perform the method described above.

[0019] Compared with existing technologies, the advantages of this application's vehicle stability assessment method, system, medium, and vehicle based on GPS-INS equipment are as follows: (1) This application uses the physical instability index as the model input feature and uses the prior knowledge of vehicle dynamics to guide the model learning. Compared with the method of using physical constraints as loss terms, it does not introduce additional hyperparameters and gradient competition. Moreover, the model can adaptively learn the mapping relationship between the feature and the level, which is suitable for the actual situation that the physical index and the real level are not strictly monotonic in the actual data.

[0020] (2) This application uses the Transformer encoder as the backbone network for temporal feature extraction. Its self-attention mechanism can effectively capture the long-distance dependency between any two moments in the sequence. It is particularly suitable for modeling the rapid evolution of vehicle state under extreme conditions, and improves the model's ability to express complex temporal patterns.

[0021] (3) This application adopts the ordinal regression prediction head of the cumulative linking framework to explicitly model the ordered structure of stability level. The ordered constraint of the bias threshold is used to ensure the monotonicity of the cumulative probability with respect to the level. Compared with the traditional multi-classification method that treats the level as an unordered category, it makes fuller use of the prior information of the ordinal structure of the level, thus improving the accuracy and rationality of the prediction.

[0022] (4) This application constructs a hybrid supervised learning framework that combines fully supervised data and weakly supervised data. It only needs to label the data segments with the highest level to achieve time-by-time level prediction, which greatly reduces the difficulty and cost of data labeling and makes model training based on a large amount of real vehicle data feasible.

[0023] (5) By adopting the Top-k soft aggregation module and the upper bound constraint of the level, this application enables the weak supervision signal to be effectively applied to the identification at key moments, while preventing the model prediction from violating the known physical constraints, thus improving the effect and stability of weak supervision learning.

[0024] (6) This application designs a joint loss function consisting of four terms: instance loss, fragment loss, sparse regularization loss and smoothing constraint loss. The sparse regularization effectively prevents the model from predicting all time points as the maximum level of the fragment level. The smoothing constraint ensures the physical continuity of the level output of adjacent time points. The synergistic effect of the various losses enables effective learning from coarse-grained fragment labels to fine-grained time-by-time levels.

[0025] (7) This application adopts a multi-stage progressive training strategy from coarse to fine, which decomposes the difficult fine-level prediction task into three sub-tasks from easy to difficult and completes them in stages. The temporal feature representation established in the previous stage provides a good initialization for the fine training in the next stage, which significantly reduces the learning difficulty of fine-level prediction under weak supervision and improves the prediction accuracy of the final model. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the vehicle stability assessment method in this application. Detailed Implementation

[0027] I. Overview of Technical Issues and Solutions This application addresses the problems of inconsistent subjective evaluation standards and lack of time-by-time quantitative indicators in vehicle stability assessment by providing a vehicle stability assessment method and system based on GPS-INS equipment. This application uses a GPS-INS device to collect multi-dimensional temporal state parameters during vehicle operation at a sampling frequency of 100Hz, constructs an input feature sequence including a physical instability index, and uses a time-series prediction model (based on a Transformer encoder and ordinal regression predictor) trained with weakly supervised learning to output the vehicle's stability level (0-7) at each time moment, thereby transforming subjective perception into objective quantitative indicators. The technical solution of this application is described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Please refer to Figure 1 This embodiment provides a vehicle stability assessment method based on GPS-INS equipment, comprising the following steps: Data acquisition steps: Collect multi-dimensional time-series state parameters of the vehicle during driving using a GPS-INS device at a predetermined sampling frequency; the multi-dimensional time-series state parameters include at least three-axis acceleration, three-axis angular velocity, sideslip angle, velocity components, and attitude angle; Feature construction steps: Based on the multidimensional time-series state parameters, calculate the physical instability index that characterizes the current instability of the vehicle, and merge the physical instability index with the multidimensional time-series state parameters to form an input feature sequence; Model prediction steps: The input feature sequence is input into a trained time-series prediction model, which outputs the stability level of the vehicle at each time step. The stability level is used to quantitatively evaluate the dynamic stability of the vehicle.

[0029] II. Specific Implementation Methods for Data Acquisition and Preprocessing 2.1 Data Acquisition Equipment and Installation In this embodiment, data acquisition utilizes a MEMS (Micro-Electro-Mechanical Systems) based integrated navigation system, preferably the NovAtel SPAN-IGM-S1 high-precision integrated navigation system, which integrates a MEMS inertial navigation module and a dual-antenna GPS / GLONASS receiver. The equipment installation requirements are as follows: the inertial navigation module is rigidly fixed near the vehicle's center of gravity, with its sensitive axis aligned with the vehicle's coordinate system, and the installation deviation not exceeding ±2°; the dual antennas are installed along the vehicle's longitudinal centerline, with a baseline length of not less than 1.5m, to provide high-precision heading angle information. After powering on, the equipment must be left to align for at least 5 minutes to ensure initial alignment accuracy.

[0030] 2.2 Vehicle Coordinate System Definition This application adopts a right-handed Cartesian vehicle coordinate system conforming to ISO 8855: the x-axis points in the direction of vehicle movement, the y-axis points to the driver's left, and the z-axis points upwards (from the roof to the sky). Rotation about the x-axis is the roll angle φ, rotation about the y-axis is the pitch angle θ, and rotation about the z-axis is the yaw angle. angular velocities , , The positive direction is determined by the right-hand rule.

[0031] 2.3 Multidimensional Time-Series State Parameter Acquisition The GPS-INS device continuously collects multi-dimensional temporal state parameters during vehicle operation at a sampling frequency of 100Hz. Let a complete test data segment be defined as a time series. Where T is the sequence length, typically between 1000 and 3000 (corresponding to a continuous driving segment of 10 to 30 seconds), and D=13 is the feature dimension. At each time t, the following 12 signal dimensions are collected and recorded:

[0032] Of the above signals, the northbound velocity and eastward speed The signal is provided by GPS, with the remaining signals provided by the MEMS inertial navigation module. Side slip angle. The attitude is calculated in real time by the integrated navigation system based on the difference between the velocity direction and the heading angle. It should be noted that the attitude calculation of MEMS inertial navigation under high dynamic conditions (such as sharp turns and emergency lane changes) has certain drift and errors. This error is an inherent characteristic of MEMS devices. The standardization process in the physical instability index constructed later in this application can partially eliminate the dimensional differences and offset effects of different vehicles and different test conditions. However, the model learns an end-to-end mapping from the observed signal to the stability level, which does not depend on the absolute accuracy of the inertial navigation data.

[0033] 2.4 Calculation of Physical Instability Index After data cleaning, the raw 12-dimensional signals acquired by GPS-INS are further used to calculate the 13th feature dimension—the physical instability index. The index is designed based on prior knowledge of vehicle dynamics, aiming to capture multiple dimensions of vehicle instability from a physical perspective.

[0034] First, for each signal variable, its global mean and standard deviation are calculated on the full training dataset. This is expressed in terms of the absolute value of lateral acceleration. For example, its global mean and standard deviation The calculation method is as follows: ; ; in This represents the total number of time steps in the full training dataset. After global z-score normalization, The expression is: .

[0035] Similarly, calculate the yaw rate separately. Side slip angle and roll rate The global mean and standard deviation are calculated, and then z-score standardization is performed to obtain... , and .

[0036] Physical instability index The calculation formula is as follows: ; ; in , , , These are the weighting coefficients.

[0037] In one specific embodiment, the weighting coefficients are determined by minimizing the linear regression error between the physical instability index and the manually labeled maximum fragment-level rank. Specifically, a validation dataset containing N samples is collected, where each sample i has a maximum fragment-level rank. Calculate the maximum value of the physical instability index at all times within this segment. Solve the following optimization problem using the least squares method: , And satisfy , .

[0038] In one example, the typical weight coefficients obtained from the optimization solution are: =0.35 (lateral acceleration has the highest weight because it is the most direct physical quantity for measuring the lateral stability of a vehicle). =0.25 (Yaw rate characterizes the severity of the vehicle's yaw response). =0.20 (Slip angle reflects tire adhesion). =0.20 (roll angular velocity reflects vehicle roll dynamics). It should be noted that the above weighting coefficients are only exemplary values ​​and can be adjusted according to the differences in vehicle type (sedan, SUV, commercial vehicle, etc.) and test conditions (moose test, sine sweep, J-Turn, etc.) in actual applications.

[0039] After completing the above calculations, As the 13th feature dimension, it is combined with the aforementioned 12 signals to form a complete input feature sequence. .

[0040] Using this feature directly as input to the model, rather than as a separate ranking loss term, offers several advantages: First, it avoids introducing additional loss hyperparameters and gradient competition, simplifying the training process. Second, the model can adaptively learn the mapping relationship between the feature and the ranking, rather than being forced into a monotonic mapping, thus adapting to estimation biases in real-world MEMS inertial navigation systems under high-dynamic conditions. The actual situation is not strictly monotonous and does not correspond to the true level.

[0041] 2.5 Data Segmentation and Labeling The vehicle stability level set is G = {0, 1, 2, 3, 4, 4.5, 5, 5.5, 6, 7}, a total of 10 levels (0 represents stationary, 7 represents the most unstable). The model's task is to learn a mapping f: R T×D -> {0, ..., K-1} T Such that for any input time series X, the output is... To approximate the true time-by-time hierarchical sequence as closely as possible.

[0042] When testing on a vehicle, it's impossible to record the stability level at every moment. A feasible approach is to record the most unstable level of that data segment. This results in the collected data only having fragment-level labels, not labels for each moment. To address the problem of lacking moment-by-moment labels for unstable data, this application constructs a hybrid supervised learning framework.

[0043] The continuously collected long sequence data is divided into independent data segments according to a set duration (e.g., 10-30 seconds). When segmenting, avoid segmenting across different operating conditions (i.e., ensure that the vehicle's driving state is relatively continuous and complete within each segment), and prioritize segmentation points when the vehicle's state is relatively stable (e.g., during a straight-line constant speed segment).

[0044] Each data segment is labeled with a segment-level tag. This represents the maximum true rating across all time points in the data segment. Labeling is performed collaboratively by at least three professional test engineers who have received standardized training, with the final label determined through majority voting. The labeling is based on the engineers' overall perception of the "least unstable" level of the entire segment, rather than precise time-by-time determination. Specifically: Let the segment-level label of the i-th data segment X be . K represents the number of levels in the current training phase, which is the maximum value of the true level index across all time points in this segment. .

[0045] Based on the data characteristics, the tags present a two-layer structure: (1) Fully supervised layer: for fragment-level labels The fragments, whose true level at each moment is known. = 0 corresponds to the vehicle being stationary at all times. ; = 1 corresponds to straight-line driving or small lateral dynamic conditions, at all times. Where the stationary moment (e.g., velocity v is less than 1 m / s and Less than 0.1 m / s 2Level 0 is marked as 0, and all other times are marked as 1. Since level 0 represents a stationary vehicle, identifying level 0 allows us to determine the times for level 1. These two types of segments provide precise time-by-time level labels.

[0046] (2) Weak supervision layer: for fragment-level labels The fragment, only known The segment contains a mixture of multiple levels (for example, a level 7 segment must contain times of levels 1 to 6), and the specific level of each time is unobservable and cannot be identified manually.

[0047] The above The fragments constitute a fully supervised dataset .for The fragments constitute a weakly supervised dataset. .

[0048] In one embodiment, the dataset size is as follows: the fully supervised dataset contains approximately 2,000 segments (approximately 500 segments for Y=0 and approximately 1,500 segments for Y=1), and the weakly supervised dataset contains approximately 15,000 segments, with a roughly balanced distribution across different levels (approximately 1,800 to 2,000 segments each for Y=2, 3, 4, 4.5, 5, 5.5, 6, and 7).

[0049] III. Detailed Implementation of the Model Architecture The time series prediction model used in this application consists of three cascaded modules: a time series encoder, an ordinal regression prediction head, and a bag-level aggregation module. The specific implementation methods of each module are described below.

[0050] 3.1 Timing Encoder The temporal encoder uses the Transformer encoder as the backbone network. The core advantage of the Transformer encoder is that its self-attention mechanism can capture the long-distance dependency between any two moments in the sequence, which is especially suitable for modeling the rapid evolution of the vehicle's state under extreme conditions (such as the entire process of the vehicle going from steady state to transient state and then back to steady state in a single lane change).

[0051] The model input is a complete time series. First, position encoding is performed on the input: ; Location coding The position information is encoded using sine / cosine functions: ; ; in The index of the position in the sequence at time ( ), where i is the dimension index ( ), Hide dimensions for the model.

[0052] In one specific embodiment, the configuration parameters of the Transformer encoder are as follows: number of encoder layers L=4, number of multi-head self-attention heads h=8, and model hidden dimension. =256, feedforward network dimensions =1024, Dropout rate set to 0.1. Calculated layer by layer by the encoder: ; Each TransformerEncoderLayer contains a multi-head self-attention sublayer and a feedforward sublayer, each with residual connections and layer normalization. The multi-head self-attention is calculated as follows: ; ; ; in , , , All are learnable projection matrices. .

[0053] Final encoder output The hidden state vector at each time t is .

[0054] 3.2 Ordinal Regression Prediction Head Traditional multi-class cross-entropy loss treats rank as an unordered category, ignoring the natural ordering relationships between ranks. This application employs a cumulative link framework to explicitly model the ordinal structure of ranks.

[0055] For each time t, the hidden state vector Mapped to scalar scores through a linear projection layer. ,in This is a learnable projection vector. It is then passed through K-1 ordered bias thresholds. Calculate the cumulative probability: ; in For the sigmoid activation function: Through constraints This can guarantee the cumulative probability with respect to level. Monotonically decreasing, that is This aligns with the physical meaning of ordinal structures—the probability of exceeding a higher level must be less than or equal to the probability of exceeding a lower level.

[0056] In practical implementation, the bias threshold The following methods are used to ensure orderliness: Let , ,in These are unconstrained, learnable parameters. Ensure that the difference between adjacent biases is always positive.

[0057] The probabilities of each level are obtained by differencing the cumulative probabilities: ; Expected level at each time point We obtain the result by weighted summation of the probabilities at each level: .

[0058] During the inference phase, the level corresponding to the highest probability can also be directly taken as the discrete prediction result: .

[0059] 3.3 Bag-level aggregation module For weakly supervised datasets The samples in the dataset only have fragment-level labels Y and no time-by-time level labels. To address this, this application designs a bag-level aggregation module that uses a gated attention mechanism to automatically identify the most unstable critical moments within a segment and aggregates the time-by-time prediction results into fragment-level predictions.

[0060] Specifically, for the hidden state vector at each time t The attention weights are calculated using a gating attention mechanism. : ; ; in This is a learnable attention weight vector. The weight matrix is ​​a learnable matrix. This is a learnable bias vector.

[0061] The intention behind the attention mechanism is that, in weakly supervised data, the segment-level label (maximum value) is determined by the few most unstable moments within the segment. The attention mechanism can automatically learn to assign higher weights to these critical moments and lower weights to stable moments, thereby achieving effective supervision from segment labels to critical moments.

[0062] After the attention weights are calculated, the first m time steps with the largest weights are selected to form a set T. k : ; in This is for rounding down. For example, with T=2000, This means selecting the 100 moments with the highest attention weight. The m value is designed to be 5% of the sequence length, which can both cover a sufficient number of key moments within the segment to resist single-point noise interference and focus on the extreme moments that truly determine the maximum level of the segment.

[0063] A weighted average of the selected m time points yields a segment-level prediction probability vector: ; in Let be the rank probability vector at time t.

[0064] During the training phase, a bounded order constraint needs to be imposed on the weakly supervised samples. Since the maximum order of this data segment is known to be Y, the predicted order at any given time should theoretically not exceed Y. This constraint is enforced through an output layer mask: ; ; in This is a normalization factor that ensures the sum of probabilities for all levels after masking is 1. This masking operation applies to all samples (including fully supervised and weakly supervised samples) during the training phase. This mask is not applied during the inference phase to output natural predictions.

[0065] IV. Specific Implementation Methods of the Loss Function The loss function for model training consists of four parts, which work together to achieve the goal of effectively learning time-by-time levels from fragment-level labels.

[0066] 4.1 Instance Loss

[0067] Instance loss applied to fully supervised datasets For each time step t, ordinal regression is trained using its time-by-time rank label. Ordinal regression treats K-1 cumulative probabilities as independent binary classification problems: ; in This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0068] Instance loss The average is taken over all time points of all fully supervised samples: .

[0069] 4.2 Fragment Loss

[0070] Fragment loss applied to weakly supervised datasets Apply a negative log-likelihood loss to the fragment-level prediction probability vector output by the bag-level aggregation module: ; in This represents the probability value corresponding to the true level Y in the fragment-level predicted probability vector.

[0071] 4.3 Sparse Regularization Loss

[0072] In multi-instance learning frameworks, a typical degradation pattern exists: the model predicts all time points within a segment as segment-level labels Y. Since max(Y, Y, ..., Y) = Y, this prediction results in... It is a valid solution for minimizing loss, but it cannot distinguish the differences in level between different times within the segment.

[0073] Sparse regularization breaks this degenerate equilibrium by constraining the mean of the predicted levels within segments. Specifically, it restricts the mean of the expected levels within segments to be significantly smaller than the segment-level label Y, forcing the model to concentrate high-level predictions on a few key moments: ; in , This represents the ratio of the expected mean to the expected maximum. The physical meaning is that within a segment, the average stability level should be significantly lower than the maximum level, because unstable moments usually only account for a small portion of the entire segment.

[0074] 4.4 Smoothing Constraint Loss

[0075] Vehicle stability is physically continuous. Since the sampling frequency in this application is 100 Hz, which is sufficiently large, abrupt changes in the stability levels between adjacent time steps are impossible. The smoothing constraint loss enforces temporal continuity by penalizing the difference in expected stability levels between adjacent time steps. ; in The entire training set is used. The loss constraint ensures that the expected level difference between adjacent time steps does not exceed 1. A linear penalty is applied when the difference is greater than 1, and the loss is 0 when the difference is less than or equal to 1. The meaning of a threshold of 1 is that in actual vehicle dynamics, the jump in stability level usually does not exceed one full level. For example, it is physically unreasonable to jump directly from "slightly unstable (level 3)" to "severely unstable (level 6)".

[0076] 4.5 Joint Losses

[0077] The total loss is the weighted sum of the above four losses. : ; In the formula, The loss is calculated based on fully supervised data. The loss is calculated for weakly supervised data. To reduce the sparsity of the regularized loss, prevent training degradation, and penalize the lazy behavior of guessing the highest level, As a temporal continuity constraint loss, the stability of a vehicle is a continuously changing physical quantity, and the level between adjacent time steps cannot jump too much (for example, jumping directly from level 1 to level 6). Its purpose is to make the level curve output by the model smooth and conform to physical laws. and These are the weighting coefficients.

[0078] In one specific embodiment, the weighting coefficient is taken as: (The weight of the segment loss should not be too large, so as to avoid weak supervision signals dominating the training.) (Sparse regularization helps prevent degradation, but should not be too strong to avoid distorting normal predictions.) (The smoothing constraint gradually takes effect in later training stages). It should be noted that the values ​​of these weight coefficients need to be optimized based on the performance on the validation set. The above values ​​are preferred examples and do not constitute a limitation on the scope of protection of this application.

[0079] V. Specific Implementation Methods of Coarse-to-Fine Multi-Stage Training Training a weakly supervised model directly on the original 10 levels presents significant challenges: the data differences between adjacent levels (such as level 4 and level 4.5, level 5 and level 5.5) are extremely small. Under weak supervision, the model needs to infer whether a value belongs to level 4 or level 4.5 at each time step based solely on fragment-level labels. However, the subjective annotations by engineers themselves contain significant noise between these fine-grained levels, making learning extremely difficult. This application employs a coarse-to-fine multi-stage training strategy, decomposing the difficult task into three sub-tasks of increasing difficulty.

[0080] 5.1 Level Merging Mapping Scheme This application defines three granularity levels, with level 0 and level 1 remaining independent at all stages to ensure... Stable time-by-time monitoring signals can be provided in all stages. The hierarchy mapping of the three stages is shown in Table 1 below: Table 1

[0081] Stage 1 (K=3): The model only needs to distinguish three coarse-grained levels - 0 (stationary), 1 (stable driving), and 2 (unstable). The task is the simplest, and the model can quickly establish a basic temporal feature representation.

[0082] Stage 2 (K=5): Further refine the coarse-grained model to 0, 1, 2 (slightly unstable), 3 (moderately unstable), and 4 (severely unstable). The model further learns to distinguish the degree of instability based on the learned feature representations.

[0083] Stage 3 (K=10): Restore all 10 original levels and perform the final refined prediction.

[0084] 5.2 Internal Training Process for Each Stage Each training phase contains two sub-phases: Sub-phase 1: Supervised pre-training. Using only the fully supervised dataset. (Levels 0 and 1), the loss function is Weakly supervised data is not used in this stage, with the aim of allowing the model's temporal encoder to establish good initial feature representations under reliable supervision signals.

[0085] Sub-phase 2: Hybrid supervised training. Using all data. The loss function is a complete four-term joint loss. At this stage, weakly supervised data begins to play a role, and the model uses fragment-level labels to learn the internal structure that distinguishes levels.

[0086] Specifically, in the initial stage of sub-stage 2, the sparse regularization loss is adjusted. Weighting coefficients Apply linear warmup: within the first N epochs, The value increases linearly from 0 to the target value (e.g., 0.1). After the warm-up period, The target value is fixed. The purpose of warming up is to prevent sparse regularization from imposing too strong a constraint in the early stages of training when the model has not yet learned any meaningful predictions, which could lead to training instability.

[0087] 5.3 Stage 1: Coarse-grained training (K=3) Sub-phase 1 (Supervised Pre-training): Starting from random initialization, using... The model is trained, and the loss function is... The optimizer used is AdamW, with an initial learning rate of 1×10⁻⁶. -3 The batch size is 64, and the number of training epochs is 80. The learning rate uses cosine annealing decay, eventually decaying to 1×10⁻⁶.-5 .

[0088] Sub-phase 2 (hybrid supervised training): using all data The loss function is a four-term joint loss. At this point... All segments (original levels 2-7) are mapped to coarse level 2, and the segment-level labels are uniformly set to 2. The learning rate is reduced to 3×10. -4 (1 / 3 of sub-stage 1), with 40 training epochs. Linearly preheat from 0 to 0.1 within the first 10 epochs.

[0089] 5.4 Stage 2: Medium-grained training (K=5) Weights are transferred from the final model of Stage 1: all parameters of the temporal encoder and all parameters of the attention module are directly inherited from the values ​​after Stage 1 training and used as the initialization for Stage 2. The number of primal rank predictions for ordinal regression is changed from 3 to 5 (the number of bias thresholds is changed from K-1=2 to 4), and their parameters are re-randomized.

[0090] Sub-phase 1 (Supervised Pre-training): Using only The loss is Since the encoder was already well initialized, the number of training epochs was reduced to 20, and the learning rate was 5×10⁻⁶. -4 .

[0091] Sub-phase 2 (hybrid supervised training): Using all data, the loss is a four-term joint loss. At this point... It covers coarse levels 2, 3, and 4 (corresponding to original levels 2-7), and fragment-level labels provide discriminative signals for three-class classification (providing richer information compared to the binary classification of Stage 1). The learning rate is reduced to 1.5 × 10⁻⁶. -4 The number of training epochs is 30. The temperature was linearly preheated from 0 to 0.1 over the first 8 epochs.

[0092] 5.5 Stage 3: Fine-grained training (K=10) Weights are transferred from the final Stage2 model: all parameters of the temporal encoder and attention module are directly inherited from the values ​​after Stage2 training. The ordinal regression prediction head is reinitialized to a configuration of K=10 (9 bias thresholds).

[0093] Sub-phase 1 (Supervised Pre-training): Using only The loss is The training epochs were 15, and the learning rate was 2×10⁻⁶. -4 .

[0094] Sub-phase 2 (hybrid supervised training): Using all data, the loss is a four-term joint loss. At this point... Covering all levels from the original level 2 onwards (2, 3, 4, 4.5, 5, 5.5, 6, 7), fragment-level labels provide rich discriminative signals across 8 levels. The learning rate is reduced to 5×102. -5 The number of training epochs is 30. The model is linearly warmed up from 0 to 0.1 over the first 6 epochs. After training, the final stability level prediction model is obtained.

[0095] 5.6 Key Technical Details of the Training Process In sub-stage 2 of Stage 2 and Stage 3, when the sample size is insufficient in the early stages of training, data augmentation strategies can be used to expand the weak supervision data for some weak supervision levels. Specifically, the following augmentation operations are performed on the original sequences: (1) Time axis scaling - randomly stretching or compressing the sequence (scaling factor 0.9~1.1). (2) Additive noise – Gaussian noise is added to the input features (the standard deviation is 1% to 5% of the standard deviation of the original signal); (3) Time Segment Discarding – Randomly discard no more than 10% of consecutive time segments. The above enhancements are only used on weakly supervised data and not on fully supervised data, so as not to destroy the accurate time-by-time level labels.

[0096] The hardware environment for model training was an NVIDIA A100 GPU (40GB VRAM), and the software framework was PyTorch 1.12. The total training time was approximately 48 hours (three phases combined). In one embodiment, an early stopping strategy was employed: training for the current phase was stopped when the fragment-level classification accuracy on the validation set no longer improved for 10 consecutive epochs, in order to avoid overfitting.

[0097] The hyperparameters during training are summarized in Table 2 below: Table 2

[0098] VI. Specific Implementation Methods of the Reasoning Stage After model training is complete, it is deployed to the vehicle stability assessment system for inference. The data acquisition method during the inference phase is exactly the same as that during the training phase: the GPS-INS device acquires 12-dimensional signals in real time at a sampling frequency of 100Hz, and calculates the physical instability index in real time. The features are then merged into a 13-dimensional feature vector and fed into the time series prediction model.

[0099] The main difference between the inference phase and the training phase is that: (1) no upper bound mask is applied ( (1) Without any truncation or renormalization, output a natural level probability distribution; (2) The output frequency is 100Hz, that is, a stability level is output at each time, realizing a true time-by-time evaluation.

[0100] During inference, the model outputs the rank probability vector at each time step. The system can choose to output one of the following formats based on application requirements: (1) Discrete level: It is suitable for intuitive rating output in calibration and acceptance scenarios; (2) Expected level: It is suitable for autonomous driving decision-making modules that require continuous numerical feedback, and the output value is a continuous real number from 0 to 7.

[0101] VII. Specific Implementation Methods for Application Scenarios 7.1 Applied to ESP calibration and acceptance In the ESP (Electronic Stability Program) calibration and acceptance scenario, the test vehicle performs moose test (ISO3888-2), sine sweep, J-turn, and circular maneuvering according to standard test specifications at the test track. The system in this application records and outputs the stability level curves at each moment throughout the entire process in real time.

[0102] Acceptance engineers can use pre-defined threshold standards from the vehicle manufacturer to make quantitative judgments. For example, "under the 80km / h moose test condition, the maximum stability level must not exceed level 5 throughout the test," and "the continuous duration of levels exceeding level 4 must not exceed 0.5 seconds." If the test results exceed these standards, the test is deemed unqualified and returned to the calibration engineer for parameter adjustment. These quantitative standards can be kept consistent among different acceptance engineers, eliminating individual differences in subjective evaluation.

[0103] The system's time-by-time stability rating curves can also be used to accurately pinpoint the onset and recovery points of vehicle instability, providing calibration engineers with optimization directions. For example, if the rating peak occurs at the start point of the second lane change in a dual lane change scenario, engineers can specifically adjust the ESP yaw moment control parameters in that area.

[0104] 7.2 Application to Decision Constraints in Autonomous Driving In Level 3 and above autonomous driving systems, the model in this application is deployed in the vehicle decision control unit. Before the vehicle is about to perform operations such as automatic lane changing or emergency obstacle avoidance, the decision module calls the model in this application to predict the stability level for future moments (such as the next 2-3 seconds) based on the current moment and historical moment (such as the past 5 seconds) state data.

[0105] If the predicted stability level exceeds the safety threshold (e.g., level ≥ 5), the decision module automatically executes one of the following strategies: (1) Suppress the operation, issue a warning to the driver and request takeover; (2) Reassess after reducing the intensity of operation (such as reducing the steering wheel angle and reducing the lane change speed); (3) Perform path correction (such as increasing the radius of curvature of the lane change trajectory) to reduce stability risks.

[0106] In practical implementation, the model in this application can achieve an inference speed of less than 5ms per frame (when batch size=1) on in-vehicle computing platforms such as NVIDIA Orin or Horizon Robotics Journey 5, which meets the real-time requirements.

[0107] VIII. Experimental Results and Effect Verification 8.1 Experimental Setup To verify the validity of this application, a test dataset containing approximately 17,000 segments was collected at a closed test track. The test conditions covered typical test conditions such as the moose test (vehicle speed 30~90km / h), dual lane change (vehicle speed 30~120km / h), sinusoidal sweep (frequency 0.5~3Hz, amplitude gradually increasing to the limit), J-Turn (vehicle speed 50~120km / h), circular maneuver (radius 30~100m, vehicle speed 30~90km / h), and single lane change (vehicle speed 30~120km / h).

[0108] The evaluation metrics used are: (1) Max-Level Accuracy; (2) Root Mean Square Error (RMSE) of each time step, where the true level of each time step is obtained by taking the majority vote after three experts have labeled it back to back, and is only used for evaluation of the test set; (3) Smoothness Violation Rate, which is used to evaluate the smoothness of the output curve.

[0109] 8.2 Comparison Methods This application is compared with the following methods: Baseline-1 (Fixed Threshold Method): Levels are set based on the magnitude of the absolute value of lateral acceleration. <0.2g is level 0~2, 0.2~0.4g is level 3~4, >0.4g is level 5~7, without considering the timing context.

[0110] Baseline-2 (LSTM+Direct Multi-Class): Uses a 2-layer LSTM (256 hidden dimensions) instead of the Transformer encoder, uses standard cross-entropy multi-class loss, does not use ordinal regression and weak supervision mechanisms, and is trained only on fully supervised data.

[0111] Baseline-3 (Transformer + Multi-class, Unsupervised): Uses the same Transformer encoder as in this application, cross-entropy multi-class loss, and is trained only on fully supervised data.

[0112] Baseline-4 (This application removes progressive training and trains directly at K=10): It retains all the mechanisms of this application (Transformer + ordinal regression + weak supervision + four loss terms), but removes coarse-to-fine progressive training and trains directly from scratch at K=10.

[0113] This application (complete solution): Transformer + ordinal regression + weak supervision + four loss terms + three-stage progressive training.

[0114] 8.3 Main Results The performance comparison of each method on the test set is shown in Table 3 below: Table 3

[0115] Experimental results show that, compared with the fixed threshold method, this application improves the segment-level accuracy by 33.3 percentage points (from 38.2% to 71.5%), reduces the time-by-time RMSE by 47.6% (from 1.87 to 0.98), and reduces the smoothing violation rate by 74.2% (from 12.4% to 3.2%), verifying the significant technical effect brought about by the synergistic effect among the various technical features of this application.

[0116] 8.4 Ablation Experiment To further verify the effectiveness of each module, ablation experiments were conducted by removing each component one by one based on the complete scheme. The results are shown in Table 4. Table 4

[0117] Ablation experiments show that each technical module contributes positively to the final performance. Among them, sparse regularization and upper bound constraints are the most critical to prevent degradation. The progressive training strategy contributes significantly to improving the segment-level accuracy (about 7.7 percentage points), which verifies the effectiveness of decomposing difficult tasks into multi-stage learning from easy to difficult.

[0118] This application can be directly embedded into the existing ESP calibration process, replacing the subjective evaluation stage and providing objective and reproducible quantitative rating results. OEMs can integrate the evaluation algorithm of this application by modifying software modules, without requiring modifications to existing hardware; only GPS-INS devices need to be installed on the test vehicles, resulting in low deployment costs.

[0119] In the field of autonomous driving, this application can be integrated into the onboard decision control unit (DCU) or central domain controller as part of a safety monitoring module. Its time-by-time output can provide real-time feedback on stability boundaries to the planning and control module, improving the safety and comfort of the autonomous driving system.

[0120] Furthermore, the weakly supervised learning framework and coarse-to-fine training strategy proposed in this application are not only applicable to vehicle stability assessment, but can also be extended to other time-series data rating scenarios (such as driver behavior risk assessment, industrial equipment health status monitoring, etc.), and have broad application prospects.

[0121] For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the scope of protection of this application.

Claims

1. A vehicle stability assessment method based on GPS-INS equipment, characterized in that: Includes the following steps: Data acquisition steps: Collect multi-dimensional time-series state parameters of the vehicle during driving using a GPS-INS device at a predetermined sampling frequency; The multidimensional time-series state parameters include at least triaxial acceleration, triaxial angular velocity, sideslip angle, velocity components, and attitude angle; Feature construction steps: Based on the multidimensional time-series state parameters, calculate the physical instability index characterizing the current instability level of the vehicle; combine the physical instability index with the multidimensional time-series state parameters to form the input feature sequence; the calculation formula for the physical instability index is: ； ； in, , , and These are the values ​​of the absolute value of lateral acceleration, yaw rate, sideslip angle, and roll rate after global z-score normalization. The mean lateral acceleration. The standard deviation of lateral acceleration; , , , These are the weighting coefficients; Model prediction steps: The input feature sequence is input into a trained time-series prediction model, which outputs the stability level of the vehicle at each time step. The stability level is used to quantitatively evaluate the dynamic stability of the vehicle.

2. The method according to claim 1, characterized in that, The multidimensional time-series state parameters include: longitudinal acceleration, lateral acceleration, vertical acceleration, roll rate, pitch rate, yaw rate, sideslip angle, north velocity, east velocity, roll angle, pitch angle, yaw angle, and physical instability index.

3. The method according to claim 1, characterized in that, The time-series prediction model includes: A temporal encoder is used to extract features from the input feature sequence, capture the long-distance dependency between any two time points in the sequence, and output the hidden state vector at each time point. An ordinal regression prediction head is used to predict the probability distribution of stability level at each time step based on the hidden state vector at that time step. The ordinal regression prediction head uses a cumulative linking framework to model the ordered relationship between the levels.

4. The method according to claim 3, characterized in that, The time-series encoder uses a Transformer encoder, which captures long-distance dependencies in time-series data through a self-attention mechanism.

5. The method according to claim 3, characterized in that, The ordinal regression prediction head calculates the rank probability at each time step in the following manner: For each moment t Predict K-1 cumulative probabilities: ； in, For the sigmoid function, For a moment t The hidden state vector. w For projection vectors, b 0 < b 1 <···< b K−2 For ordered bias thresholds; All levels k probability The cumulative probability is obtained by performing a difference operation. ； Expectation level at this moment We obtain the result by weighted summation of the probabilities at each level: 。 6. The method according to claim 1, characterized in that, The time-series prediction model is trained using a weakly supervised learning strategy, which utilizes fragment-level label supervision signals to learn the stability level prediction at each time step.

7. The method according to claim 1, characterized in that, The stability rating is used for at least one of the following purposes: Output objective quantitative evaluation indicators to the vehicle electronic stability program calibration system as the criteria for calibration acceptance; Provide stability boundary constraints to the autonomous driving decision-making system, and when the predicted stability level exceeds a predetermined safety threshold, suppress dangerous driving operations or perform deceleration and path correction.

8. A vehicle stability assessment system, characterized in that, include: The data acquisition module is used to collect multi-dimensional time-series status parameters of the vehicle via GPS-INS devices; The feature processing module is used to calculate the physical instability index based on the multidimensional temporal state parameters, and merge the physical instability index with the multidimensional temporal state parameters to construct the input feature sequence; Storage module for storing trained time-series prediction models; The evaluation output module is used to input the input feature sequence into the time series prediction model, obtain and output the stability level of the vehicle at each time step; The formula for calculating the physical instability index is as follows: ； ； in, , , and These are the values ​​of the absolute value of lateral acceleration, yaw rate, sideslip angle, and roll rate after global z-score normalization. The mean lateral acceleration. The standard deviation of lateral acceleration; , , , These are the weighting coefficients.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

10. A vehicle, characterized in that, include: GPS-INS devices are used to collect multi-dimensional time-series status parameters of vehicles; An electronic control unit, communicatively connected to the GPS-INS device, is configured to perform the method as described in any one of claims 1 to 7.

Images