Intelligent driving behavior analysis method based on internet of vehicles big data

Abstract

The application discloses an intelligent driving behavior analysis method based on Internet of Vehicles big data, which comprises the following steps: collecting the multi-source driving behavior data flow and external environment perception data of a target vehicle in real time; dynamically intercepting the transition period driving behavior segment corresponding to the starting process based on the vehicle motion state; performing feature coding on the environment data to generate an environment context state vector; comparing the transition period segment with historical benchmarks to calculate an identity consistency likelihood index representing the degree of coincidence; based on the environment context state vector, using causal analysis to determine the degree to which behavior changes can be explained by the environment, and generating an environment attribution explanation degree index; and comprehensively considering the identity consistency and the environment attribution index to generate a driver switching determination result and execute a safety response. The application effectively distinguishes the behavior change induced by environmental factors and the actual driver switching by introducing an open set recognition and causal inference mechanism, and solves the misjudgment problem caused by the misrecognition of unknown drivers and environmental interference.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of Internet of Vehicles and intelligent transportation, and particularly relates to an intelligent driving behavior analysis method based on Internet of Vehicles big data. BACKGROUND

[0002] With the deep integration of Internet of Vehicles and intelligent transportation system, real-time confirmation of the driver's identity has key technical significance for ensuring the safety of vehicle control rights and implementing personalized auxiliary driving strategies. Non-contact identity authentication by analyzing micro-control features (such as pedal control and steering habits) in sensor data has become a core technical path to improve the active safety protection capability of intelligent vehicles.

[0003] The prior art mainly uses closed-set classification algorithms (such as random forests and convolutional neural networks) based on supervised learning to identify driving behavior. The system usually collects vehicle CAN bus data, extracts time-domain or frequency-domain features, and trains a classifier to distinguish between registered drivers. In the detection process, real-time data is intercepted by a sliding window, and its statistical correlation or distance measure (such as Euclidean distance) with historical reference features is calculated. When the feature deviates beyond a pre-set fixed threshold, the system outputs an identity abnormality alarm or switching decision result.

[0004] However, the above scheme has technical problems of failure to identify unknown targets, weak environmental anti-interference capability, and insufficient decision timeliness in open road environments and complex interaction scenarios. First, the classification model based on the closed-set assumption forces all observed samples to be classified as registered users, which makes it easy to misidentify unauthorized strangers taking over the vehicle as legitimate users, and there is a risk of false negatives. Second, the existing method mainly relies on statistical correlation rather than causal logic, making it difficult to distinguish whether the change in driving behavior is caused by driver switching or environmental factors such as rain and snow weather, traffic congestion, etc., resulting in a high false alarm rate in environmental mutations. In addition, the decision mechanism of fixed time window or fixed confidence threshold lacks flexibility, making it difficult to quickly respond to switching events while ensuring low false alarm rate. SUMMARY

[0005] The application aims to provide an intelligent driving behavior analysis method based on Internet of Vehicles big data to solve the above problems in the prior art.

[0006] The technical scheme is an intelligent driving behavior analysis method based on Internet of Vehicles big data, which comprises:

[0007] real-time collection of multi-source driving behavior data streams and external environment perception data of a target vehicle;

[0008] based on vehicle motion state parameters in the multi-source driving behavior data streams, dynamic interception of a transition period driving behavior segment corresponding to the vehicle starting process;

[0009] encoding the external environment perception data, extracting and generating an environmental context state vector representing the current driving scene;

[0010] comparing the transitional driving behavior segment with the pre-stored candidate driver historical behavior benchmark, calculating an identity consistency likelihood indicator representing the degree of consistency between the current driving behavior and the historical benchmark;

[0011] based on the environmental context state vector, analyzing the correlation between the behavior feature changes in the transitional driving behavior segment and environmental factors, generating an environmental attribution explanation degree indicator representing the degree to which the behavior changes can be explained by environmental factors;

[0012] integrating the identity consistency likelihood indicator and the environmental attribution explanation degree indicator to generate a driver switching decision result; based on which, performing vehicle control strategy adjustment or safety warning operation.

[0013] According to one aspect of the present application, calculating the identity consistency likelihood indicator comprises:

[0014] extracting multi-dimensional features of the transitional driving behavior segment to generate a transitional behavior feature vector;

[0015] obtaining historical behavior probability density distribution models corresponding to each candidate driver pre-constructed based on historical samples in the candidate driver historical behavior benchmark;

[0016] inputting the transitional behavior feature vector into the historical behavior probability density distribution model of each candidate driver, calculating the posterior probability value or likelihood value of the transitional behavior feature vector belonging to the distribution space of each candidate driver, and taking it as the identity consistency likelihood indicator.

[0017] According to one aspect of the present application, the historical behavior probability density distribution model is a Gaussian mixture model;

[0018] calculating the posterior probability value or likelihood value comprises:

[0019] calculating the minimum Mahalanobis distance of the transitional behavior feature vector to the center of each Gaussian component in the Gaussian mixture model;

[0020] fitting the tail of the minimum Mahalanobis distance distribution to construct a distance extreme value distribution model;

[0021] based on the distance extreme value distribution model, calculating the unknown class probability of the transitional behavior feature vector not belonging to any pre-stored candidate driver;

[0022] combining the unknown class probability and the posterior probability of the Gaussian mixture model to generate an open set probability distribution containing the unknown class dimension, and determining it as the identity consistency likelihood indicator.

[0023] According to an aspect of the present application, the historical behavior probability density distribution model is a single Gaussian distribution model comprising a mean vector and a covariance matrix;

[0024] The posterior probability value or the likelihood value is calculated, comprising:

[0025] The instantaneous likelihood of the transition period behavior feature vector at the current time belonging to each candidate driver is calculated based on the single Gaussian distribution model;

[0026] The cumulative posterior probability at the previous time is obtained, the instantaneous likelihood is multiplied by the cumulative posterior probability at the previous time, and the product is normalized to obtain the cumulative posterior probability at the current time;

[0027] The cumulative posterior probability at the current time is determined as the identity consistency likelihood index.

[0028] According to an aspect of the present application, the environmental attribution explanation degree index is generated, comprising:

[0029] The actual behavior change feature of the transition period driving behavior segment relative to the historical benchmark is extracted;

[0030] Based on the environmental context state vector, the expected behavior influence amount of the current environmental factor on the driving behavior feature is determined by using a preset environmental impact evaluation model;

[0031] The attribution deviation between the actual behavior change feature and the expected behavior influence amount is calculated;

[0032] Based on the significance of the attribution deviation, the environmental attribution explanation degree index is quantitatively generated.

[0033] According to an aspect of the present application, the environmental impact evaluation model is a causal forest model, the expected behavior influence amount is an counterfactual expected behavior change, and the attribution deviation is a causal residual;

[0034] The environmental attribution explanation degree index is quantitatively generated, comprising:

[0035] The environmental context state vector is input into a pre-trained causal forest model to estimate the conditional average treatment effect of the environmental factor on each dimension of the driving behavior feature. Accordingly, the counterfactual expected behavior change that should occur under the current environmental condition is calculated;

[0036] The counterfactual expected behavior change is subtracted from the actual behavior change feature to obtain a causal residual after removing the environmental impact;

[0037] The causal residual is subjected to a statistical significance test, and when the causal residual is not statistically significant, an environmental attribution explanation degree index indicating high explanation degree is generated.

[0038] According to one aspect of the present application, the environmental impact evaluation model is an environment-behavior response knowledge base, and the expected behavior impact quantity is an expected change range;

[0039] The quantitative generation of the environmental attribution explanation index includes:

[0040] According to the environmental context state vector, a matched environment-behavior response rule is retrieved from a pre-stored environment-behavior response knowledge base;

[0041] The environment-behavior response rule is parsed to determine the expected change range of each dimension of driving behavior characteristics under the current environment;

[0042] It is determined whether the actual behavior change characteristics fall within the expected change range, and the proportion of the characteristics dimensions falling within the expected change range is counted;

[0043] The proportion is determined as the environmental attribution explanation index.

[0044] According to one aspect of the present application, the generation of the driver switching decision result includes:

[0045] The identity consistency likelihood index and the environmental attribution explanation index at the current time are weighted and fused to generate an instantaneous switching evidence strength representing the possibility of driver switching at the current time;

[0046] Based on the instantaneous switching evidence strength, the pre-stored decision state at the previous time is time-series updated to generate a cumulative decision statistic reflecting the overall evidence level from the transition period start time to the current time;

[0047] The cumulative decision statistic is compared with a preset decision threshold interval, and a driver switching decision result is generated according to the comparison result.

[0048] According to one aspect of the present application, the cumulative decision statistic is a cumulative log-likelihood ratio sequence;

[0049] The comparison of the cumulative decision statistic with the preset decision threshold interval includes performing a sequential probability ratio test:

[0050] Based on the preset expected false alarm rate and expected missed alarm rate, the upper decision boundary and the lower decision boundary of the sequential test are calculated;

[0051] At each sampling time, it is determined whether the current value of the cumulative log-likelihood ratio sequence is greater than the upper decision boundary and the lower decision boundary;

[0052] If the current value is greater than the upper decision boundary, a decision result of confirming switching is generated; if the current value is less than the lower decision boundary, a decision result of maintaining the original identity is generated; if the current value is between the lower decision boundary and the upper decision boundary, a decision result of continuing observation is generated and the next time data is collected.

[0053] According to one aspect of the present application, the environmental context state vector is extracted and generated, comprising:

[0054] The external environment perception data is parsed into discrete environmental elements and continuous environmental elements;

[0055] The one-hot encoding processing is performed on the discrete environmental elements to generate a sparse feature vector;

[0056] Based on the preset quantile or distribution statistical characteristics, the adaptive interval binning processing is performed on the continuous environmental elements to map the continuous numerical values to discrete interval encoding;

[0057] The sparse feature vector and the discrete interval encoding are spliced and normalized to generate the environmental context state vector.

[0058] Beneficial effects, the present application introduces an open set recognition and causal inference mechanism, effectively distinguishes the behavior change induced by environmental factors and the actual driver switching, solves the misjudgment problem caused by unknown driver misrecognition and environmental interference. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A step flowchart of the intelligent driving behavior analysis method based on vehicle networking big data provided by the embodiment of the present application.

[0060] Figure 2 A step flowchart of calculating the identity consistency likelihood index provided by the embodiment of the present application.

[0061] Figure 3 A step flowchart of calculating the posterior probability value or likelihood value provided by the embodiment of the present application.

[0062] Figure 4 A step flowchart of generating the environmental attribution explanation index provided by the embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the protection scope of the present application.

[0064] It is to be understood that the terms include and comprising and their conjugations, as used herein, are intended to be open-ended, that is, to mean including, without limitation, two or more of the listed steps or elements, components, or options, as appropriate.

[0065] As shown in Figure 1 The intelligent driving behavior analysis method based on Internet of Vehicles big data includes the following steps:

[0066] Real-time collection of multi-source driving behavior data streams and external environment perception data of the target vehicle.

[0067] In other words, real-time collection of multi-source driving behavior data streams and external environment perception data of the target vehicle, and time sequence synchronization alignment, to obtain multi-source driving behavior data streams and external environment perception data after time sequence synchronization alignment.

[0068] In this embodiment, data collection is the basis of the entire analysis process. Multi-source driving behavior data streams refer to time series data that can reflect the micro-control state and macro-motion state of the vehicle. Specifically, multi-source driving behavior data streams can include vehicle chassis signals obtained through the vehicle controller area network (CAN) bus network, such as throttle pedal opening, brake pedal pressure, steering wheel angle, engine speed, and instantaneous vehicle speed, etc.; and also include high-frequency attitude data such as vehicle three-axis acceleration, three-axis angular velocity, and Euler angle collected through the inertial measurement unit (IMU). External environment perception data refers to information describing the characteristics of the driving environment in which the vehicle is located, which can include image data captured by the vehicle-mounted camera, target distance and relative speed data detected by the millimeter wave radar, and road traffic state information such as the front congestion condition or signal phase received through the vehicle-to-road (V2X) communication module. In order to ensure the consistency of multi-source data in the time dimension, the system uses a unified system clock as the reference to synchronize and resample the timestamps of the above types of data, eliminating the time offset error caused by the inconsistency of sensor sampling frequencies.

[0069] Based on the vehicle motion state parameters in the multi-source driving behavior data streams, the transition period driving behavior segment corresponding to the vehicle starting process is dynamically intercepted.

[0070] Specifically, the transition period refers to the initial stage when the driver takes over the vehicle and starts to perform driving operations, and the behavior characteristics of this stage best reflect the individual physiological and psychological characteristics of the driver. The vehicle motion state parameters mainly include vehicle speed and acceleration signals. The system continuously monitors the changes of these parameters, and when it detects that the vehicle has transitioned from a stationary state to a motion state and meets the preset start determination logic, the data interception mechanism is triggered. The interception operation does not use a fixed time length, but adjusts the time window length adaptively according to the current start working condition to obtain a transition period driving behavior segment that contains complete start characteristics. The dynamic interception mechanism can effectively deal with the start differences under different driving habits and different road conditions, so that the intercepted data segment neither loses information by being too short nor introduces steady-state noise by being too long.

[0071] The external environment perception data is feature encoded, and an environment context state vector representing the current driving scene is extracted and generated.

[0072] In this embodiment, unstructured or semi-structured environment perception raw data is converted into structured numerical vectors to facilitate subsequent mathematical model processing. The feature encoding process includes numerical mapping of discrete variables (such as weather type, road grade), and binning or normalization processing of continuous variables (such as traffic density, road curvature). The generated environment context state vector is a high-dimensional feature vector that comprehensively depicts the external environmental factors that have a significant impact on driving behavior at the current time. The environment context state vector not only serves as an input condition for subsequent attribution analysis, but also provides a scene benchmark for cross-situation comparison of driver behavior.

[0073] The transition period driving behavior segment is compared with the pre-stored candidate driver historical behavior benchmark to calculate an identity consistency likelihood indicator representing the degree of agreement between the current driving behavior and the historical benchmark.

[0074] Specifically, the system attempts to answer the question of who the current driving behavior resembles. The candidate driver historical behavior benchmark is a mathematical model library that the system pre-establishes and contains the behavior characteristic distributions of multiple registered drivers. The comparison process involves extracting the current intercepted transition period driving behavior segment as a feature vector and projecting it into the feature space of each candidate driver to calculate the probability density value or distance metric of the observed data under each model. The output identity consistency likelihood indicator quantifies the likelihood that the current driver is a known registered user or the degree of abnormality if he / she does not belong to any known user.

[0075] Based on the environment context state vector, the correlation between the behavior characteristic changes in the transition period driving behavior segment and the environmental factors is analyzed to generate an environmental attribution explanation degree indicator representing the degree to which the behavior changes can be explained by the environmental factors.

[0076] In this embodiment, the idea of causal inference or rule association is introduced, which can solve the core problem of whether the behavior change is caused by the change of the person or the environment. Specifically, the system uses a pre-set model to evaluate what degree of driving behavior change the current environmental context state vector should lead to, i.e. the expected behavior impact, and compares it with the actual observed behavior change. If the actual change is highly consistent with the expected change, it means that the behavior difference is mainly caused by the environment, and the attribution explanation degree is high; on the contrary, if the environment cannot explain the current behavior mutation, the attribution explanation degree is low, which implies that the driver switching may have occurred.

[0077] The driver switching determination result is generated by integrating the identity consistency likelihood indicator and the environmental attribution explanation degree indicator.

[0078] Specifically, the system integrates the evidence from the identity matching dimension, i.e. the identity consistency likelihood indicator, and the evidence from the environmental explanation dimension, i.e. the environmental attribution explanation degree indicator. The integration can use weighted summation, logical rule judgment or more complex sequential probability test methods. By comprehensively considering the information from the two aspects, the system can effectively reduce the misjudgment rate caused by environmental interference, and output the driver switching determination result such as confirming the switching, maintaining the original identity or pending observation.

[0079] According to the driver switching determination result, the vehicle control strategy adjustment or safety warning operation is performed.

[0080] In this embodiment, after obtaining the determined determination result, the system performs the corresponding closed-loop response. If it is determined that the driver switching is authorized, the system can automatically load the personalized settings of the driver, such as seat position, driving mode preference; if it is determined that the switching is unauthorized or abnormal takeover, the system triggers the safety warning mechanism, such as sending alarm information to the Internet of Vehicles platform, limiting the maximum driving speed of the vehicle or requesting remote human intervention, to ensure the safety of the vehicle and information.

[0081] In one possible implementation, based on the vehicle motion state parameters in the multi-source driving behavior data stream, the transition period driving behavior segment corresponding to the vehicle starting process is dynamically intercepted, including:

[0082] The vehicle speed signal and acceleration signal are monitored, and when it is detected that the vehicle speed increases from the static state and exceeds the pre-set starting threshold, the transition period starting time is determined.

[0083] In this embodiment, the accurate identification of the start of the starting motion is the prerequisite for subsequent analysis. The system reads the pre-processed vehicle speed signal v(t) and longitudinal acceleration signal a(t) in real time. In order to exclude the slight vibration or sensor drift interference when the vehicle is stationary, the preset starting threshold is usually set to a small positive value, for example, 3 km / h. At the same time, the system also verifies the persistence of the acceleration signal, that is, it requires the acceleration to remain positive for a continuous period of time (for example, 0.5 seconds). When the above conditions are met at the same time, the time is marked as the transition period start time t start .

[0084] The average acceleration of the vehicle in the short time window after the transition period start time is calculated, and the dynamic window length adjustment factor is calculated based on the average acceleration.

[0085] Specifically, the system evaluates the severity of the start. In the short time window after t start , for example, 2 seconds, the average longitudinal acceleration of the vehicle is calculated, denoted as a*. In order to quantify the difference between the current starting intensity and the standard working condition, the standard starting acceleration parameter a ref is introduced, which can be set to a typical value of 1.5 m / s2. The dynamic window length adjustment factor k adj is defined as the ratio of the standard starting acceleration to the current average acceleration. The specific calculation formula is:

[0086] k adj =a ref / max(a*,a min );

[0087] Where a min is a preset minimum effective acceleration threshold, for example, 0.3 m / s2, to prevent overflow caused by too small denominator. The calculation formula of the dynamic window length adjustment factor reflects that the greater the acceleration, the shorter the time required to reach a stable driving state, so the window should be shortened; on the contrary, the smaller the acceleration, the longer the starting process, and the window should be extended.

[0088] The preset reference window length is modified using the dynamic window length adjustment factor to obtain the current transition period window length.

[0089] In this embodiment, the influence of the dynamic characteristics and the environmental conditions on the starting process is considered comprehensively. Specifically, the base window length is calculated based on the adjustment factor:

[0090] T base_adj = T base × k adj ;

[0091] Where T base_adj is the base window length corrected by acceleration; Tbase is a preset reference window length, for example, 10 seconds.

[0092] Further, the preset reference window length is modified by a dynamic window length adjustment factor to obtain a current transition period window length, and further comprising executing a multi-scenario compensation mechanism:

[0093] Obtain road slope data, engine coolant temperature data and load estimation data of the vehicle; when the road slope data indicates that the vehicle is in a hill start state, increase the hill compensation length based on the dynamic window length.

[0094] Exemplarily, the compensation amount ΔT slope may be calculated using a linear model, for example:

[0095] ΔT slope = |θ slope | × C slope ;

[0096] where θ slope is the slope angle, and C slope is the slope coefficient, for example, 0.1 seconds per degree.

[0097] When the engine coolant temperature data is lower than a preset threshold, increase the cold start compensation length based on the dynamic window length.

[0098] Specifically, because the engine and transmission respond slowly at low temperatures, the driver's operation feedback chain is lengthened. The compensation amount ΔT cold may be calculated as:

[0099] ΔT cold = (T thresh - T coolant ) × C temp ;

[0100] where T thresh is the temperature threshold, for example, 40 degrees Celsius; T coolant is the current engine coolant temperature, and C temp is the temperature coefficient, for example, 0.05 seconds per degree Celsius.

[0101] When the load estimation data exceeds a preset load standard, increase the load compensation length ΔT load based on the dynamic window length.

[0102] For example, increase by 0.8 seconds fixedly. The final current transition period window length T window is the sum of the above:

[0103] T window = T base_adj + ΔTslope + ΔT cold + ΔT load .

[0104] The system also imposes upper and lower bound constraints on the final result, such as 3 seconds to 15 seconds, to ensure the real-time performance of the system.

[0105] From the multi-source driving behavior data stream, a data sequence starting from the beginning of the transition period and lasting for the length of the current transition period window is intercepted to obtain the transition period driving behavior segment.

[0106] In this embodiment, based on the accurate time boundary [t start , t start + T window ] calculated above, the system extracts all dimensions of sensor data in this time period from the data cache area, including CAN signals and IMU signals. The data sequence is the transition period driving behavior segment after adaptive cutting, which eliminates the static redundant data before starting and completely covers the whole process from starting to tending to steady state, providing high-quality input for subsequent feature extraction.

[0107] As shown in Figure 2 , according to one aspect of the present application, the identity consistency likelihood index is calculated, including:

[0108] Multi-dimensional feature extraction is performed on the transition period driving behavior segment to generate a transition period behavior feature vector.

[0109] In this embodiment, considering that the driver's starting process has a clear time sequence evolution rule, the system does not generally calculate the statistical value of the whole window, but adopts a phased strategy.

[0110] In a preferred implementation, multi-dimensional feature extraction is performed on the transition period driving behavior segment to generate a transition period behavior feature vector, including performing phased feature extraction:

[0111] The transition period driving behavior segment is divided into a start-up initial stage, an acceleration transition stage, and a steady convergence stage in chronological order.

[0112] Specifically, the division of the three stages can be based on vehicle speed normalization thresholds or fixed time ratios. For example, the start-up initial stage corresponds to the process of increasing the vehicle speed from 0 to 10% of the target cruising speed; the acceleration transition stage corresponds to the process of increasing the vehicle speed from 10% to 80%; and the steady convergence stage corresponds to the process of the vehicle speed tending to be stable from 80%.

[0113] In the start-up initial stage, the response delay time and initial stepping rate of the accelerator pedal are extracted.

[0114] Specifically, at the initial stage of starting, the driver's unconscious habits are most prominent. The system extracts the response delay time of the accelerator pedal, i.e. the time difference from t start to the first time the accelerator pedal opening exceeds the idle threshold; and the initial depression rate, i.e. the mean first derivative of the accelerator pedal opening in the first 0.5 seconds. Reflects the driver's starting reaction speed and character.

[0115] In the acceleration transition sub-stage, the pedal opening rate and steering wheel fine adjustment frequency are extracted.

[0116] That is, during acceleration, the system pays attention to the driver's control style. The pedal opening rate, i.e. the time series standard deviation of the accelerator pedal position, is extracted to represent the smoothness of the acceleration process; at the same time, the steering wheel fine adjustment frequency, i.e. the frequency of the steering wheel angle fluctuating in a small range (such as ±2 degrees), reflects the driver's habit of correcting the direction.

[0117] In the steady convergence sub-stage, the convergence speed of the control signal fluctuation amplitude is extracted.

[0118] Specifically, when the vehicle approaches steady state, the system analyzes how the driver ends the acceleration action. The convergence speed of the control signal fluctuation amplitude can be achieved by calculating the decay rate of the envelope line of the accelerator pedal signal. Reflects the convergence characteristics of whether the driver is used to quickly or slowly releasing the accelerator.

[0119] The feature parameters extracted in each sub-stage are spliced to generate a transition period behavior feature vector.

[0120] Optionally, the multi-dimensional feature extraction of the transition period driving behavior segment also includes performing transient feature enhancement processing, specifically:

[0121] First-order difference operation is performed on the continuous signal sequence in the transition period driving behavior segment to obtain a difference fluctuation sequence.

[0122] In this embodiment, in order to capture the high-frequency operation details hidden in the smooth signal, the system performs enhancement processing. First-order difference operation is performed on the continuous signal sequence x(t) such as the accelerator pedal, brake pedal and steering wheel angle to obtain the difference fluctuation sequence Δx(t) = x(t) - x(t-1), which eliminates the trend of the signal and highlights the instantaneous operation jitter.

[0123] The peak distribution density and zero-crossing rate of the difference fluctuation sequence are calculated.

[0124] Specifically, based on the differential wave sequence, the system calculates the peak distribution density, i.e. the number of local maxima of Δx(t) exceeding a preset noise threshold per unit time, and the zero-crossing rate, i.e. the frequency of sign flipping of Δx(t). These two indicators effectively characterize the subtlety and fine-tuning frequency of the driver's operation, for example, a driver with more frequent fine actions will have a higher zero-crossing rate.

[0125] The switching time interval and amplitude connection characteristics of the accelerator pedal signal and the brake pedal signal are counted.

[0126] In this embodiment, for the pedal switching action, the system identifies the time interval from the driver releasing the accelerator to stepping on the brake (or vice versa). The mean and variance of this interval are counted, as well as the connection relationship of the opening values of the two pedals at the switching moment, such as whether there is an overlapping stepping area. This constitutes a unique foot action fingerprint of the driver.

[0127] The feature parameters extracted in each sub-stage are spliced with the peak distribution density, zero-crossing rate and switching characteristics to generate a transition period behavior feature vector.

[0128] In other words, the peak distribution density, zero-crossing rate and switching characteristics are added as enhancement components to the transition period behavior feature vector.

[0129] Specifically, the system concatenates all the scalar features mentioned above in a predefined order to form a high-dimensional numerical vector, i.e. the transition period behavior feature vector. It combines macro features in time sequence and micro transient features in differential enhancement, providing a high-discriminability input representation for the subsequent identity recognition model.

[0130] Obtain the historical behavior probability density distribution model corresponding to each candidate driver, which is pre-constructed based on historical samples in the historical behavior benchmark of the candidate driver.

[0131] In other words, obtain the historical behavior probability density distribution model corresponding to each candidate driver, which is pre-constructed based on historical samples in the historical behavior benchmark of the candidate driver.

[0132] Input the transition period behavior feature vector into the historical behavior probability density distribution model of each candidate driver, calculate the posterior probability value or likelihood value of the transition period behavior feature vector belonging to the distribution space of each candidate driver, and use it as an identity consistency likelihood indicator.

[0133] In a preferred embodiment, the historical behavior probability density distribution model is a Gaussian mixture model.

[0134] In this embodiment, the system abandons the simple single Gaussian assumption and adopts a Gaussian Mixture Model (GMM) to characterize the complex behavioral patterns of drivers. This is because the same driver may exhibit multimodal distribution characteristics under different moods or physiological states (such as aggressive mode, fatigue mode, and normal mode). For the k-th candidate driver, its Gaussian mixture model consists of M Gaussian components, each containing a weighting coefficient π. m Mean vector μ m and covariance matrix Σ m These parameters were obtained by training the driver's historical data offline using the Expectation-Maximization (EM) algorithm. For example, the weight coefficients satisfy a normalization condition, meaning the sum of the weights of all components is 1. The mean vector μ... m Located at the cluster center of this driving behavior pattern in the feature space, and the covariance matrix Σ m This describes the degree of dispersion and correlation of behavioral characteristics under this model.

[0135] like Figure 3 As shown, the posterior probability or likelihood value of the transitional behavioral feature vector belonging to the distribution space of each candidate driver is calculated, including:

[0136] Calculate the minimum Mahalanobis distance from the transitional behavior feature vector to the center of each Gaussian component in the Gaussian mixture model.

[0137] In this embodiment, when the system obtains the current transitional behavior feature vector x, it calculates the distance between the transitional behavior feature vector and each Gaussian component in the driver's GMM model. To eliminate the influence of different feature dimension dimensions and correlations, Mahalanobis distance is preferred over Euclidean distance. Specifically, the distance between the feature vector x and the center μ of the m-th Gaussian component is calculated. m The squared Mahalanobis distance. The system iterates through all M components and selects the minimum value as the representative distance d of the feature vector relative to the driver model. min (x). The minimum Mahalanobis distance reflects the degree of deviation between the current behavior and the historical pattern most similar to that of the driver. If the minimum Mahalanobis distance is too large, it suggests that the current behavior may not belong to that driver.

[0138] By fitting the tail of the minimum Mahalanobis distance distribution, a distance extreme value distribution model is constructed.

[0139] In this embodiment, in order to quantify the probability of too large distance, the Extreme Value Theory (EVT) is introduced to fit the tail of the distribution of the minimum Mahalanobis distance, that is, the system focuses on the tail features of the distance distribution, that is, the distribution law of the abnormal samples far away from the cluster center. Specifically, the maximum value of the pre-stored historical matching distance is selected, for example, the first 10% of the data, and the parameters of the Weibull distribution are fitted by using the maximum likelihood estimation method. The constructed distance extreme value distribution model includes the shape parameter κ and the scale parameter λ. The Weibull distribution can well describe the statistical characteristics of the extreme value, and provides a mathematical basis for subsequent calculation of the probability that the sample belongs to an unknown space.

[0140] Based on the distance extreme value distribution model, the probability of the transition period behavior feature vector belonging to an unknown category of any pre-stored candidate driver is calculated.

[0141] Specifically, based on the fitted Weibull distribution, the system calculates the cumulative distribution function (CDF) value of the current observed minimum Mahalanobis distance d min (x) reflecting the normality of the minimum Mahalanobis distance. In order to obtain the unknown probability, the system defines the unknown category probability P unknown as the complementary probability of the cumulative distribution function value, or is mapped through a special open set identification function (such as the OpenMax activation function). The specific calculation formula can be expressed as:

[0142] P unknown (x) = 1 - exp( - (d min (x) / λ) κ );

[0143] When the observed behavior feature distance is very far from the center of all known modes, P unknown will quickly tend to 1, indicating that the current driver is most likely an unregistered stranger.

[0144] The open set probability distribution containing the unknown category dimension is generated by combining the unknown category probability with the posterior probability of the Gaussian mixture model, and the open set probability distribution is determined as the identity consistency likelihood indicator.

[0145] That is, based on the Gaussian mixture model, the closed set posterior probability of the transition period behavior feature vector under each candidate driver model is calculated; the open set probability distribution containing the unknown category dimension is generated by combining the unknown category probability with the closed set posterior probability, and the open set probability distribution is determined as the identity consistency likelihood indicator.

[0146] In this embodiment, the system will calculate the Bayesian posterior probability P knownP unknown The probabilities are normalized to ensure completeness of the probability space. The resulting open-set probability distribution is [P1, P2,..., P K , P unknown ]. The open-set probability distribution is the final output of the identity consistency likelihood indicator. For example, if the result is [0.1, 0.1, 0.05, 0.75], the system can determine that the current driver has a 75% probability of being an unknown person, avoiding the error of the traditional closed-set classification which forces it to be the first driver with a slightly higher probability (0.1).

[0147] In a detailed numerical case, assume that there are two drivers registered in the system, driver A and driver B, and the current transitional behavior feature vector is x. The minimum Mahalanobis distance of x to each driver GMM model is calculated. Assume that the minimum Mahalanobis distance of x to driver A model is d A =2.5, and the minimum Mahalanobis distance of x to driver B model is d B =4.2. Based on the pre-fitted Weibull distribution parameters, assume that κ=1.8 and λ=3.0, the unknown class probability is calculated. Replace d min =min(2.5, 4.2)=2.5 into the formula, and P unknown =1-exp(-(2.5 / 3.0) 1.8 )≈0.42 is obtained. The closed-set posterior probability of each known driver is calculated. Assume that the Bayesian calculation based on GMM obtains P(A|x)=0.65 and P(B|x)=0.35. Among them, P(A|x) is the closed-set posterior probability of the behavior feature belonging to driver A under the condition that the current transitional behavior feature vector x is observed; that is, the probability of judging that the current driver is the registered driver A according to the extracted behavior feature x. The unknown class probability is combined for normalization adjustment. After introducing the unknown class dimension, the open-set probability distribution is: P'(A)=0.65×(1-0.42) / Z≈0.30, P'(B)=0.35×(1-0.42) / Z≈0.16, and P'(unknown)=0.42 / Z≈0.54; wherein Z is the normalization coefficient. The result shows that the current driver has a 54% probability of being an unregistered stranger, and the system should pay attention to it.

[0148] In another preferred embodiment, the historical behavior probability density distribution model is a single Gaussian distribution model including a mean vector and a covariance matrix.

[0149] In the embodiment, to reduce the computational complexity, it is assumed that the behavior characteristics of each driver obey a multivariate single Gaussian model (SGM). For the kth candidate driver, the model parameters only include a mean vector μ k and a covariance matrix Σ k . To prevent the covariance matrix from being non-invertible due to insufficient samples or feature collinearity, the system uses a regularization technique when building the model, i.e., a small positive number (for example, 0.01 times the ratio of the matrix trace to the dimension) is added to the diagonal of the covariance matrix to ensure numerical stability of the matrix operation.

[0150] The posterior probability value or the likelihood value of the transition period behavior characteristic vector belonging to the distribution space of each candidate driver includes:

[0151] The instantaneous likelihood of the transition period behavior characteristic vector belonging to each candidate driver at the current time is calculated based on the single Gaussian distribution model.

[0152] In the embodiment, the recursive Bayesian update mechanism is preferably used. As the transition period data continuously flows in, the system extracts the feature vector x t at each time step t. The probability density value of the feature vector belonging to the kth driver, i.e., the instantaneous likelihood L k (t), is calculated using the probability density function of the multivariate Gaussian distribution. The specific calculation formula is as follows:

[0153] L k (t) = (1 / sqrt((2π) d × |Σ k |)) × exp( -0.5 × (x t - μ k ) T × Σ k -1 ×(x t - μ k ));

[0154] where L k (t) is the instantaneous likelihood of the kth driver at time t; π is the circular constant; d is the dimension of the feature vector; |Σ k | is the determinant of the covariance matrix of the kth driver; x t is the transition period behavior characteristic vector at time t; μ k is the mean vector of the kth driver; T denotes vector transposition; Σ k -1is the inverse of the covariance matrix of the kth driver. Intuitively, it reflects how similar the current behavior (e.g. a sudden acceleration) is to the driver's historical habits.

[0155] The cumulative posterior probability at the previous time is obtained, the instantaneous likelihood is multiplied by the cumulative posterior probability at the previous time, and the product is normalized to obtain the cumulative posterior probability at the current time.

[0156] Specifically, the system sets a state register to store the cumulative posterior probability P k (t-1) at the previous time t-1. At the initial time t=0, the prior probability of all drivers is initialized to an equal value (1 / K) or is set according to the historical usage frequency. At time t, the system fuses the current instantaneous likelihood L k (t) into the historical belief using the Bayesian formula. The update formula is as follows:

[0157] P k (t) = (L k (t) × P k (t-1)) / Σ j=1 K (L j (t) × P j (t-1));

[0158] where P k (t) is the cumulative posterior probability of the kth driver at time t; K is the number of all candidate drivers; L j (t) is the instantaneous likelihood of the jth candidate driver at time t; P j (t-1) is the cumulative posterior probability of the jth candidate driver at time t-1. The denominator part is the sum of the likelihood and the prior of all candidate drivers, which is used for normalization to ensure that the sum of all P k (t) is 1.

[0159] The cumulative posterior probability at the current time is determined as the identity consistency likelihood index.

[0160] In this embodiment, the set of P k (t) after normalization constitutes the current identity consistency likelihood index. The recursive update mechanism has strong anti-interference ability: occasional abnormal operation resulting in low instantaneous likelihood will be smoothed by the high probability accumulated in history, and only when the driver's behavior continuously deviates from a certain pattern, the cumulative posterior probability will significantly decrease. This makes the system gradually converge in confidence during data accumulation, and realizes the recognition effect of more opening and more accuracy.

[0161] In an exemplary embodiment, the external environment perception data is feature encoded, and an environmental context state vector representing the current driving scene is extracted and generated, including:

[0162] The external environment perception data is parsed into discrete environmental elements and continuous environmental elements; the discrete environmental elements are subjected to one-hot encoding processing to generate a sparse feature vector.

[0163] In the present embodiment, the environmental data is described in a fine-grained digital manner. The external environment perception data is parsed into discrete elements (such as weather type, road grade) and continuous elements (such as traffic density, road curvature). For discrete elements, one-hot encoding is used. For example, if the weather type is divided into four categories of sunny, rainy, snowy, and foggy, and it is currently raining, the encoding is [0, 1, 0, 0]. The numerical size between categories is eliminated, and an orthogonal sparse feature vector is constructed.

[0164] Based on the preset quantile or distribution statistical characteristics, the continuous environmental elements are subjected to adaptive interval binning processing to map the continuous numerical values to discrete interval encoding.

[0165] Specifically, for continuous elements, in order to capture nonlinear environmental effects, the system uses an adaptive binning strategy. Based on the historical statistical distribution of continuous elements, quantiles (such as 20%, 40%, 60%, and 80% quantile points) are used to determine the bin boundaries, ensuring that the number of samples in each bin is relatively balanced. For example, traffic density is divided into five levels of very low, low, medium, high, and congestion, and the current numerical value is mapped to the corresponding discrete interval encoding. Complex continuous variables can be converted into categorical features that are easy to process in causal models.

[0166] The sparse feature vector and the discrete interval encoding are spliced and normalized to generate an environmental context state vector.

[0167] In the present embodiment, the system splices the above two parts of encoding and may add some auxiliary global variables, such as time period identifier, to form a high-dimensional numerical vector, for example, a 55-dimensional vector. This vector comprehensively and structurally represents the current driving situation, i.e., the environmental context state vector E context , which can be used as a conditional variable to query the expected influence of the environment on behavior.

[0168] As shown in Figure 4 , an environmental attribution explanation index is generated, including:

[0169] Extracting the actual behavior change characteristics of the transition period driving behavior segment relative to the historical baseline;

[0170] Based on the environmental context state vector, the expected behavior influence amount of the current environmental factor on the driving behavior feature is determined by using a preset environmental impact evaluation model;

[0171] The attribution deviation between the actual behavior change feature and the expected behavior influence amount is calculated;

[0172] Based on the significance of the attribution deviation, the environmental attribution explanation degree index is quantitatively generated.

[0173] In a preferred implementation, the environmental impact evaluation model is a causal forest model, the expected behavior influence amount is counterfactual expected behavior change, and the attribution deviation is causal residual;

[0174] The environmental attribution explanation degree index is quantitatively generated, including:

[0175] The environmental context state vector is input into the pre-trained causal forest model, and the conditional average treatment effect of the environmental factor on each dimension of the driving behavior feature is estimated.

[0176] In the online running stage of the embodiment, the system loads the causal forest model trained based on the causal diagram. The current environmental context state vector E context is input into the model, and the model outputs the conditional average treatment effect (CATE) of the environmental factor on the specific behavior feature (such as average speed, frequency of sudden braking). The CATE function τ (x) quantifies how much change in the behavior feature is expected to occur due to the change in the environmental factor under the context x. For example, the model may output: under the current heavy rain and highway conditions, the expected average speed should be reduced by 20% than the baseline.

[0177] Based on the conditional average treatment effect, the counterfactual expected behavior change that should occur under the current environmental condition is calculated.

[0178] Specifically, based on the CATE, the system calculates the counterfactual expected behavior change. Counterfactual refers to the behavior that the original driver should have shown if he was driving under the current environment. The specific calculation formula is:

[0179] ΔB expected = Σ i (τ ij (E context ) ×ΔE i );

[0180] Where ΔB expected is the expected behavior change amount (counterfactual prediction) under the current environmental change; Σ i represents the summation of all related environmental factors i; τ ij (E context) is the conditional average treatment effect function output by the causal forest, representing the expected behavior change on the current environmental context E context is the influence coefficient of the environmental factor i on the behavior j; ΔE i is the actual change amount of the environmental factor i. The behavior change amount represents the reasonable behavior drift amount caused purely by the environmental factor.

[0181] Subtract the counterfactual expected behavior change from the actual behavior change characteristic to obtain the causal residual after removing the environmental impact.

[0182] In this embodiment, the system calculates the actual observed behavior change ΔB observed between the above-mentioned counterfactual expectation ΔB expected , that is, the causal residual ε:

[0183] ε= ΔB observed - ΔB expected ;

[0184] The causal residual represents the part of the behavior change that cannot be explained by the environmental change.

[0185] Statistical significance test is performed on the causal residual, and when the causal residual is not statistically significant, an environmental attribution explanation degree indicator indicating high explanation degree is generated.

[0186] Specifically, the system can use a double machine learning (DML) framework or a chi-square test to perform significance analysis on the residual ε. The standardized statistic of the residual is calculated, and it is judged whether it falls within the acceptance domain of the null hypothesis that no driver switching has occurred. If the residual is small and statistically insignificant, for example, the p-value is greater than 0.05, it means that the actual behavior change is consistent with the expected change of the environment, and the behavior difference is mainly caused by the environment, so a high explanation degree indicator, for example, 0.9, is generated, which tends to determine that no switching has occurred. On the contrary, if the residual is significant, for example, the expected speed is reduced by 20%, but the actual speed is accelerated by 10%, it means that there is an anomaly that cannot be explained by the environment, and a low explanation degree indicator is generated, which tends to determine switching.

[0187] In further embodiments, the causal forest model is constructed based on an environment-behavior causal graph structure, which is pre-learned in the following way:

[0188] Obtain historical driving big data containing a large number of environment states and driving behavior records;

[0189] Use PC algorithm or FCI algorithm to perform conditional independence test on the environmental variables and behavior variables in the historical driving big data;

[0190] Identify the causal connection direction between variables, construct a directed acyclic graph describing the causal dependence relationship between environmental factors and driving behavior as an environment-behavior causal graph structure.

[0191] In this embodiment, before online analysis, the system uses massive historical data to construct a causal model in the offline stage. The historical driving big data containing a large number of environmental state and driving behavior records are used to perform conditional independence test between variables by using Peter-Clark algorithm (PC) or fast causal inference (FCI) algorithm. For example, test whether the rain and the vehicle speed are independent under the given congestion degree. Through a series of tests, the system identifies the causal connection direction between variables, constructs a directed acyclic graph (DAG) describing the causal dependence relationship between environmental factors and driving behavior, and indicates which environmental factors directly lead to the change of which behavior characteristics, and eliminates the pseudo-correlation relationship.

[0192] In another preferred implementation, the environmental impact assessment model is an environment-behavior response knowledge base, and the expected behavior impact quantity is an expected change range.

[0193] In this embodiment, the system predefines a structured environment-behavior response knowledge base. The knowledge base is not a black box model, but is composed of a series of explicit IF-THEN rules. Each rule defines a specific environmental condition (IF part) and an expected driving behavior change range (THEN part). These rules are derived from statistical laws of traffic engineering, regulatory standards, or summaries of experienced driving experts. For example, a typical rule can describe that when the road curvature exceeds 0.05 and the road surface is wet, the expected average vehicle speed decreases by 15% to 30%, and the steering wheel fine-tuning frequency increases by 20% to 50%. The expected change range is the closed interval of the upper and lower limits of the change of the behavior indicators defined in the above rule.

[0194] Quantitative generation of environment attribution explanation index, including:

[0195] According to the environmental context state vector, the matching environment-behavior response rule is retrieved from the pre-stored environment-behavior response knowledge base.

[0196] Specifically, after the system generates the current environmental context state vector, it is used as a query key for matching retrieval in the knowledge base. The matching process can adopt exact matching (for discrete elements) or fuzzy matching (for continuous elements). For example, if the current state vector indicates heavy rain, highway, and light congestion, the system will traverse the knowledge base to filter out a subset of rules whose premise conditions contain the above elements. If there are multiple rules that hit at the same time, for example, one about heavy rain and one about congestion, the system can use a weighted fusion strategy or a priority strategy (such as priority of severe weather rules) to determine the final applicable combined rule.

[0197] Parse the environment-behavior response rules to determine the expected variation range of each dimension of driving behavior characteristics under the current environment.

[0198] In this embodiment, the system parses the retrieved rules and converts the qualitative description into quantitative numerical intervals. For each dimension j of driving behavior characteristics (such as vehicle speed, following distance), the system determines its expected variation range [L j , U j ] under the current environment. For example, if the reference vehicle speed is 100 km / h and the rule states that the speed should be reduced by 20%, the expected speed range may be set to [75, 85] km / h. A safety envelope of environmental influence is constructed, and behaviors falling within this envelope are considered as normal reactions caused by the environment.

[0199] Determine whether the actual behavior variation characteristics fall within the expected variation range, and count the proportion of characteristics dimensions that fall within the expected variation range.

[0200] Specifically, the system extracts the actual behavior variation characteristics during the current transition period and checks whether it is within the corresponding expected variation range dimension by dimension. To enhance the robustness of the judgment, a counterfactual reasoning verification mechanism can also be introduced, which checks the time sequence relationship between behavior variation and environmental change: only when the environmental change occurs before the behavior change (or simultaneously), and the delay time is within the reasonable physiological reaction range, such as 0.5 to 2 seconds, is the matching of this dimension considered valid. This effectively eliminates coincidences that may be logically explained by the environment, although the values meet the rules.

[0201] Determine the proportion as the environmental attribution explanation degree index.

[0202] In this embodiment, the system counts how many dimensions k successfully fall within the expected range and pass the timing check among the total N key behavior dimensions. Calculate the explanation proportion:

[0203] R explain =k / N;

[0204] Where R explain is the environmental attribution explanation degree index (proportion value); k is the number of behavior characteristic dimensions whose actual variation falls within the expected rule range; N is the total number of key behavior characteristic dimensions participating in the evaluation. This proportion is directly used as the environmental attribution explanation degree index. For example, if 8 out of 10 characteristics can be explained by the environment rules, the index is 0.8, indicating that the current behavior deviation is most likely caused by environmental factors rather than driver switching. The quantitative method based on rules is intuitive and easy to understand, making it convenient for users to explain the basis of the system's judgment.

[0205] The embodiment is applicable to a scenario that lacks large-scale historical data for causal graph learning but has rich expert experience or industry standard specifications, and has the characteristics of strong interpretability and low consumption of computing resources.

[0206] In one possible embodiment, a driver switching decision result is generated, including:

[0207] The identity consistency likelihood indicator and the environmental attribution explanation degree indicator at the current moment are weighted and fused to generate an instantaneous switching evidence strength representing the possibility of driver switching at the current moment.

[0208] In the embodiment, the system normalizes and fuses the evidence from different dimensions. It is assumed that the identity consistency likelihood indicator (such as the posterior probability of belonging to the original owner) at time t is P owner (t), and the environmental attribution explanation degree indicator is R env (t). The system calculates the instantaneous switching evidence strength E switch (t). One preferred fusion method is to emphasize the unexplained deviation, that is, when the identity matching degree is low and the environment cannot be explained, the switching evidence is the strongest. The calculation formula can be:

[0209] E switch (t) = (1 - P owner (t)) × (1 - λ × R env (t));

[0210] where λ is the adjustment coefficient of the environmental weight, ranging from 0 to 1. If the environmental explanation degree is very high (R env tends to 1), even if the identity matching degree is slightly low, the switching evidence will be weakened; on the contrary, if neither of them supports the original owner, E switch will tend to 1.

[0211] As an alternative embodiment, weighted fusion can also be implemented using a Sigmoid nonlinear function. Specifically, a decision index is constructed:

[0212] I(t) = sigmoid(α × S evidence - β × S env + γ × D deviation );

[0213] where I(t) is the normalized switching confidence; sigmoid( ) is a Sigmoid activation function; α, β, γ are the corresponding fusion weight coefficients; S evidence is the identity difference score; S env is the environmental attribution score; and D deviation is the behavior deviation degree indicator.

[0214] Based on the instantaneous switching evidence strength, the pre-stored decision state of the last time is updated in time sequence to generate a cumulative decision statistic reflecting the overall evidence level from the starting time of the transition period to the current time;

[0215] The cumulative decision statistic is compared with a preset decision threshold interval, and a driver switching decision result is generated according to the comparison result.

[0216] In a preferred implementation, the cumulative decision statistic is a cumulative log-likelihood ratio sequence.

[0217] In this embodiment, in order to realize sequential decision, the system converts the instantaneous evidence into a log-likelihood ratio (LLR). It is assumed that H1 is switching, and H0 is no switching. The log-likelihood ratio increment λ t The instantaneous switching evidence strength E switch (t) is calculated as follows:

[0218] λ t =ln[(P(E switch (t)│H1)) / (P(E switch (t)│H0))].

[0219] In a simplified implementation, the instantaneous probability can be directly approximated as: λ t = ln(E switch (t) / (1 - E switch (t))) The system maintains a cumulative log-likelihood ratio sequence Λ t , and the update rule is: Λ t =Λ t-1 +λ t . The cumulative mechanism allows weak evidence to be superimposed over time, and extracts a robust trend in noise.

[0220] The cumulative decision statistic is compared with a preset decision threshold interval, including performing a sequential probability ratio test, specifically:

[0221] Based on the preset expected false alarm rate and expected missed alarm rate, the upper decision boundary and the lower decision boundary of the sequential test are calculated.

[0222] In this embodiment, the decision boundary of the sequential probability ratio test (SPRT) is strictly derived by the system's tolerance to the error rate. The system allows the false alarm rate (Type I Error, i.e. misjudging switching) to be α, for example 0.05, and the allowed missed alarm rate (Type II Error, i.e. missing switching) is β, for example 0.1. According to the Wald theory, the upper decision boundary A and the lower decision boundary B are calculated as follows:

[0223] A = ln((1 - β) / α);

[0224] B = ln(β / (1 - α));

[0225] These two boundaries remain constant during the detection process, providing an objective mathematical criterion for decision making.

[0226] At each sampling time, the system determines the size relationship between the current value of the cumulative log-likelihood ratio sequence and the upper and lower decision boundaries.

[0227] Specifically, the system compares Λt with A and B immediately after updating Λt at each time step t. This process is carried out in real time and continuously, without waiting for the fixed observation window to end.

[0228] If the current value is greater than the upper decision boundary, a decision result of confirmed switching is generated; if the current value is less than the lower decision boundary, a decision result of maintaining the original identity is generated; if the current value is between the lower and upper decision boundaries, a decision result of continuing observation is generated and the next time data is collected.

[0229] In this embodiment, according to the comparison result, the system performs a three-branch logic: when Λ t ≥ A, it means that the evidence supporting switching is already strong enough, the system immediately terminates the detection and outputs the confirmed switching result, ensuring that the fastest response can be made when the evidence is sufficient. When Λ t ≤ B, it means that the evidence supporting the original owner is dominant, the system immediately terminates the detection and outputs the result of maintaining the original identity, avoiding unnecessary waste of computing resources. When B < Λ t < A, it means that the current evidence is not yet sufficient to make a reliable judgment, it is in the gray area or the uncertain area, at this time the system outputs the instruction of continuing observation, keeps the detection state open, and waits for the arrival of the next frame of data (t+1). Naturally, the adaptive adjustment of the observation length is realized: for obvious switching, the observation time is short; for ambiguous cases, the observation time is automatically extended until the evidence is solid.

[0230] This embodiment not only solves the trade-off problem between response speed and accuracy rate of the traditional fixed threshold method, but also provides another implementation of double evidence fusion.

[0231] According to an aspect of the present application, the candidate driver historical behavior benchmark is a situation-conditioned benchmark constructed by grouping according to environmental situation categories; the construction process includes:

[0232] The historical sample data of each candidate driver is classified according to the corresponding environmental scene label and divided into multiple situation subsets.

[0233] In this embodiment, in order to solve the problem of large difference in behavior of the same driver in different situations, the system adopts a divide-and-conquer strategy in the offline modeling stage. Specifically, long-term historical driving data of each registered driver is collected. Using environment coding technology, the historical driving data is labeled with situation tags, such as high-speed-sunny, urban-congestion, mountain-rainy. The data belonging to the same driver is divided into K situation subsets according to the situation tags. The data distribution within each subset is more compact and single, conforming to the assumption of Gaussian distribution.

[0234] For each situation subset, the mean, variance and correlation between features of the candidate driver's behavior in that situation are respectively calculated.

[0235] That is, for each situation subset, the system independently calculates its statistical features. Specifically, the feature mean vector μ kc and the covariance matrix Σ kc of the driver in that specific situation are calculated. The fine-grained statistical work reveals the situation adaptability fingerprint of the driver. For example, a driver may perform stably on the highway, i.e. small variance, but may perform impatiently on the congested road, i.e. large variance, and strong negative correlation between throttle and brake.

[0236] The situation-specific probability density parameter set corresponding to each situation category is generated as the candidate driver's historical behavior benchmark, so that in the identity matching analysis, the corresponding parameter set is selected according to the current environmental context state vector for comparison.

[0237] In this embodiment, the system packages the parameters of all subsets to build an indexed benchmark library. When running online, the system identifies the current environment E context , and like looking up a dictionary, retrieves the set of parameter models that best match the current environment from the benchmark library for identity comparison. The situation-conditioning mechanism improves the accuracy of the matching benchmark and avoids the error of using the average face to match a specific expression.

[0238] In an embodiment of the present application, when the driver switching determination result indicates that the driver has switched and the new driver identity is an unknown driver, the method further comprises performing an incremental registration process, specifically:

[0239] Establishing a temporary driver profile for the current driver and storing the current transitional driving behavior segment in the temporary driver profile.

[0240] In this embodiment, when the open set recognition module determines that the current driver is unknown and the decision module confirms this determination, the system enters the incremental registration mode. At this time, the system does not directly block the alarm, but intelligently starts the learning and interaction process. The system creates a new storage space in the memory and creates a temporary file with an ID of Temp_User_X. The transition period driving behavior segment that triggered the determination just now, that is, the data that caused the system to think he was a stranger, is stored as the first batch of samples in the file. At this time, the system has actually begun to profile the new driver.

[0241] Send an identity confirmation request to the preset authorized terminal, and if an authorization confirmation instruction for the temporary driver file is received, convert the temporary driver file to a formal file.

[0242] Specifically, in order to ensure safety, the system sends a pop-up window to the vehicle owner (administrator) through the vehicle machine screen or the bound mobile phone APP: a new driving style is detected, is it an authorized driver? If the vehicle owner confirms, for example, by entering a password or fingerprint, it means that this is a legal but previously unregistered new user, such as a family member. The system then assigns a formal ID to the temporary file.

[0243] Update the pre-stored candidate driver historical behavior benchmark with the data in the formal file, and include the new driver in the subsequent identity matching analysis range.

[0244] In this embodiment, after obtaining authorization, the system trains the GMM or SGM model of the new user using the data accumulated in the temporary file and the data collected subsequently, and adds it to the candidate driver library. At this point, the new user has completed the conversion from unknown to known. Next time the system will be able to directly identify him, and no longer determine him as a stranger, realizing the self-evolution and expansion of the system.

[0245] This embodiment perfects the whole life cycle management of the system, enabling the model to adapt to diverse environments and unknown users.

[0246] According to an aspect of the present application, in real-time collection of multi-source driving behavior data stream of target vehicle and external environment perception data and environment perception process, the system may encounter inconsistent descriptions of the same environmental elements by different sensors. For example, there is a deviation between the vehicle distance measured by millimeter wave radar and the vehicle distance measured by visual sensor. The deviation of each sensor measurement value can be calculated, and when the deviation exceeds a preset consistency threshold, for example, 10% of the measurement range, it is determined that there is a data conflict. The system queries the historical reliability score of each sensor and the current working state, such as whether the camera is in backlight state or the radar is in blind area. The resolution strategy adopts a weighted fusion method, that is, the reliability score of the sensor is used as the weight to perform weighted averaging on the conflicting data. If a sensor is in a non-optimal working interval, its weight is reduced, and verified high-confidence environmental data is generated.

[0247] According to another aspect of the present application, for data frame loss caused by communication interruption or sensor occasional failure, the system adopts a multi-level filling strategy. The first level is time series prediction filling, that is, based on the historical change trend of the signal dimension, the linear extrapolation or exponential smoothing method is used to predict the value at the current time, which is suitable for short missing, for example, less than 3 seconds. The second level is spatial correlation filling, which uses other signals that are physically related to make an estimate, for example, according to the light intensity sensor data and time information to estimate the missing weather state data. The third level is default safety value filling, when the above two levels are not available, the parameter is set to a conservative default value, for example, the front vehicle distance is set to the minimum safety distance, so that the subsequent calculation process does not interrupt.

[0248] According to another aspect of the present application, in the evidence fusion process, a fixed weight may not be suitable for all scenarios. An online learning mechanism can be introduced, and the accuracy rates of identity evidence and environmental evidence are counted according to the historical post-determination verification results, such as the post-feedback of the vehicle owner or manual review. The system calculates the weight adjustment increment Aw, the formula is as follows:

[0249] Aw = η × (Acc behav - Acc context );

[0250] Where η is the learning rate, for example, 0.05; Acc behav is the accuracy rate of identity behavior evidence; and Acc context is the accuracy rate of environmental attribution evidence. The system adds the weight adjustment increment to the weight of the behavior evidence and performs normalization processing. This allows the system to automatically tilt towards more reliable evidence sources according to actual operation results.

[0251] In one possible implementation, the setting of the decision boundary or decision threshold for the sequential probability ratio test is dynamically calibrated by the system according to the safety requirement level and the false alarm rate statistics. When the vehicle is in a high-risk state (such as high-speed driving), the system automatically reduces the false alarm rate tolerance α and increases the value of the upper decision boundary A, requiring more convincing evidence to confirm the switch. At the same time, the system monitors the recent false alarm rate (the ratio of false alarms that do not switch), and when the false alarm rate exceeds the tolerance limit, the decision threshold is automatically adjusted upward; when the missed detection rate is too high, the decision threshold is automatically adjusted downward. This allows the system to maintain an optimal balance between sensitivity and specificity at different safety levels.

[0252] In another embodiment of the present application, the intelligent driving behavior analysis method based on Internet of Vehicles big data can also be: through the vehicle-mounted gateway, the original CAN signal frame transmitted by the CAN bus is read in real time, and driving control signals such as steering wheel angle, accelerator pedal opening, brake pedal pressure, vehicle speed, and engine speed are obtained by analysis; the original IMU data output by the inertial measurement unit are synchronously collected, containing three-axis acceleration and three-axis angular velocity information, forming a multi-source original data set containing time stamps. Based on the system unified clock, the linear interpolation method is used to resample signals of different sampling rates to a unified frequency, eliminate the time offset between channels, and generate a synchronous driving behavior data sequence with strictly aligned time sequence. The sliding window detection is performed on each channel signal in the synchronous driving behavior data sequence, the abnormal jump points generated due to sensor failure or communication interference are identified and removed, the low-pass filter is used to eliminate high-frequency noise interference, and the preprocessed driving behavior data sequence with reliable quality is output.

[0253] The vehicle speed signal in the preprocessed driving behavior data sequence is monitored, and when it is detected that the vehicle changes from a stationary state to a driving state, the time point is marked as the transition period starting point; the data segment of T trans seconds after the starting point is intercepted, where T transAccording to the dynamic adjustment of the vehicle speed growth rate, the window is shortened when the vehicle speed rises quickly, and the window is lengthened when the vehicle speed rises slowly, to ensure that the original data window of the transition period contains enough effective driving operation information. In view of the characteristics of limited data volume and unstable behavior pattern in the transition period, a multi-scale transient feature extraction strategy is constructed to mine high-discrimination identity feature information from the original data window of the transition period. Specifically, the transition period window is divided into three sub-stages in time sequence, namely the start-up initial stage, the acceleration transition period and the steady-state period, which correspond to the driver's start-up habit, acceleration style and state convergence characteristics respectively. In the start-up initial stage, the response delay time of the first pedal, the initial pedal depth and the pedal rate are extracted to capture the driver's start-up response characteristics; in the acceleration transition period, the slope sequence of the accelerator pedal opening degree curve and the steering wheel fine adjustment frequency are calculated to represent the driver's acceleration control style; in the steady-state period, the convergence speed of the fluctuation amplitude of each control signal is calculated to reflect the transition characteristics of the driver reaching steady-state driving. In view of the problem of insufficient stability of traditional statistical features under short time sequence conditions, a transient feature enhancement mechanism is introduced. The first-order difference sequence of each control signal between adjacent sampling points is calculated, and the peak value distribution density and zero-crossing rate are extracted from the difference sequence to describe the rhythm characteristics of the driver's control action; the joint change mode of the accelerator and brake pedal signals is constructed, and the time interval distribution and amplitude connection characteristics of the switch between the two are extracted to represent the driver's pedal coordination habit. The above sub-stage features and transient enhanced features are spliced according to the weight to form the transition period behavior feature vector F trans , where the start-up initial stage features are given a higher weight to highlight the start-up stage information that is most significant to individual driver differences.

[0254] In view of the need to quickly give identity judgment under limited data conditions in the transition period, a gradual Bayesian update mechanism is constructed to realize the dynamic matching process of gradually improving the judgment confidence with data accumulation. The system pre-stores the historical behavior baseline feature set of each registered driver, and each driver corresponds to a set of baseline feature vectors B k and feature distribution parameters, including the mean vector μ k and the covariance matrix Σ k of each feature dimension. The matching process uses incremental calculation: when the first batch of data is collected in the transition period window, the initial posterior probability of the current observation belonging to each driver is calculated based on the extracted partial features; as the window data continues to accumulate, the Bayesian update is triggered once a new feature component is obtained, the posterior probability at the previous time is taken as the prior probability at the current time, and the updated posterior probability distribution is calculated by combining the likelihood of the new feature component. The specific calculation process is as follows: let the accumulated feature vector at the current time be F trans (t), for the kth registered driver, the likelihood L k (t) = exp(-0.5×(F trans (t) - μk ) T ×Σ k -1 ×(F trans (t) -μ k )), where μ k Let Σ be the feature mean vector of the k-th driver. k Let covariance matrix be the variance matrix. T This represents the transpose of a vector. -1 This represents finding the inverse of a matrix. It is combined with the prior probability P. k (t-1), update the posterior probability P according to Bayes' theorem. k (t) =L k (t)×P k (t-1) / Σ j (L j (t)×P j (t-1)), where the denominator is the sum of the products of the likelihoods of all registered drivers and their priors, used for normalization.

[0255] When the maximum posterior probability exceeds the fast decision threshold θ fast If the matching result is not triggered before the end of the full transition window, the matching result is output in advance without waiting for the transition window to close. If the fast determination is not triggered by the end of the window, the final posterior probability distribution is used as the matching result. The driver's identity corresponding to the highest posterior probability is output as the initial identity matching degree M. identity Simultaneously calculate the posterior probability distribution P of the candidate drivers. posterior Based on the transition period behavior feature vector F trans Baseline features of historical behavior of matched candidate drivers B matched This quantifies the deviation of current driving behavior from historical baselines from multiple dimensions. The Euclidean distance D in the feature space is calculated. euclidean = sqrt(Σ i (F trans_i - B matched_i ) 2 ), where the subscript i represents the i-th component of the eigenvector, sqrt represents the square root operation, and Σ i This represents the summation over all components, measuring the overall feature offset; it also calculates the standardized deviation Z for each feature dimension. i = (F trans_i -μ matched_i ) / σ matched_i , where μ matched_i and σ matched_i The historical mean and standard deviation of the matching driver's i-th feature are used to identify anomalous feature dimensions; the number of feature dimensions whose absolute standardized deviation exceeds a threshold is counted as the anomalous dimension count N. abnormal The above indicators are weighted and combined to generate the behavioral deviation index D.deviation = w1×D euclidean + w2×max(|Z i |) + w3×N abnormal , where w1, w2, w3 are preset weight coefficients, and max(|Z i |) represents the maximum value of the absolute values of all dimension standardized deviation values. The larger the index value is, the more significant the difference between the current behavior and the historical baseline is, and the higher the possibility of driver switching is.

[0256] The original environment perception data output by the vehicle-mounted environment perception module is read, including information such as front road curvature, lane width, surrounding vehicle distance and relative speed; the road static attribute data issued by the vehicle networking platform is synchronously received, covering the current road section type, speed limit value, traffic control state and weather condition, etc., to form an environment perception original data set. The context element analysis is performed on the environment perception original data set, and the key environmental factors affecting the driving behavior are extracted: the bend density and slope change rate of the current road section are calculated from the road geometry data; the surrounding vehicle density and the front vehicle following state are counted from the traffic flow data; the visibility level and road adhesion condition are determined from the meteorological data. The continuous elements in the driving context element set are discretized in intervals, and the category type elements are one-hot encoded to form a unified dimension numerical representation; the element importance weighting strategy is adopted, and higher weight is given to the situation that is easy to cause behavior change, such as high traffic complexity and bad weather condition, to finally splice and generate the environment context state vector E context as the context reference benchmark for subsequent behavior attribution analysis.

[0257] The behavior comparison information before and after the transition period is extracted from the pre-processed driving behavior data sequence to construct the time sequence pattern feature describing the dynamic process of behavior change. Specifically, the historical driving behavior data of the last driving segment before the transition period is read as a benchmark, compared with the current transition period behavior feature vector F trans , and the change amount ΔF i = F trans_i - F history_i is calculated for each feature dimension, where F history_i is the i-th component of the historical segment corresponding feature; the evolution trend of the change amount in the transition period is further analyzed, and if the change amount of each dimension monotonically converges over time, it is marked as a gradual change; if the change amount presents a stepwise mutation at the beginning of the transition period and remains stable thereafter, it is marked as a mutation change. The behavior change time sequence feature T change containing the change amount amplitude, change type label and dimension change direction is generated.

[0258] For the case that behavior change is induced by environmental factors, an environmental relevance verification mechanism based on causal inference framework is constructed to evaluate whether the current behavior deviation can be reasonably explained by environmental changes. Specifically, from the historical driving data accumulated by the system, the typical association patterns between various environmental context changes and driving behavior responses are statistically summarized. For each type of environmental factor change, the expected impact direction and impact amplitude range of each behavior feature dimension are established, forming an environmental-behavior response knowledge base R env . This knowledge base records causal response rules such as the expected increase of steering wheel operation frequency by 15% to 40% when entering a curved road section, the expected decrease of accelerator pedal average opening by 20% to 35% when encountering congestion, etc. The current environmental context state vector E context is read to identify the environmental factors that have undergone significant changes; for the change amount of each dimension in the behavior change time series feature T change , the environmental-behavior response knowledge base R env is retrieved one by one to determine whether the change falls within the expected response range of the corresponding environmental factor. The environmental explainable proportion R explain is defined as N matched / N total , where N matched is the number of behavior change dimensions that can be explained by environmental factors, and N total is the total number of behavior dimensions that have undergone significant changes. For the case that the environmental explainable proportion is high, the temporal causal relationship between behavior change and environmental change is further verified: if the behavior change occurs after the environmental change and there is a reasonable response delay, the confidence of environmental attribution is enhanced; if the behavior change occurs before the environmental change or the temporal relationship does not conform to the normal driver response characteristics, the credibility of environmental attribution is weakened. The causal verified environmental relevance score S env and the list of environmentally explainable behavior changes are output.

[0259] For the case that behavior change is caused by driver switching, a cross-situation feature consistency verification mechanism is constructed to verify whether the current behavior pattern conforms to the expected performance of a certain driver in the current situation. The historical driving data of each registered driver in the system is grouped by environmental situation, and for each driver, a behavior feature distribution model is established under different environmental categories. Specifically, the environmental situation is divided into typical categories such as urban congestion, urban free flow, highway, and mountainous curved road, and the feature mean, feature variance, and feature correlation of each driver in each category are statistically summarized to form a situation-conditioned driver feature library P conditional . According to the current environmental context state vector E context , the environmental situation category is determined, and the expected feature distribution of each driver in this situation is extracted from the situation-conditioned driver feature library P conditional . The current transitional period behavior feature vector Ftrans Match the situation-adaptive feature distribution of each driver, calculate the conditional Mahalanobis distance D mahal_k = sqrt((F trans - μ k_c ) T ×Σ k_c -1 ×(F trans - μ k_c )) where μ k_c and Σ k_c are the feature mean vector and covariance matrix of the kth driver under the current situation category c. If the current behavior highly matches the situation-adaptive features of a non-recent driver, further verify whether the match has cross-situation stability: retrieve the historical performance of the candidate driver under other similar situations, and check whether the deviation of the currently observed behavior conforms to the personal feature drift rule of the driver. If the consistency check passes, the confidence of the candidate driver as the switched identity is significantly improved. Output the situation-adaptive match score S match_k and the cross-situation consistency score S consist_k of each candidate driver.

[0260] Integrate the environmental correlation score S env , the list of environment-explainable behavior changes, the situation-adaptive match score S match_k of each driver, the cross-situation consistency score S consist_k , and the behavior deviation indicator D deviation , and output the final switching decision result through a dual-evidence fusion framework. Specifically, establish a decision-making matrix, with the horizontal axis representing the environmental attribution intensity (quantified by S env ) and the vertical axis representing the driver switching evidence intensity (quantified by S match_k and S consist_k jointly). When the environmental attribution intensity is high and the driver switching evidence is weak, the behavior change is determined to be environment-induced, and the original driver identity label is maintained; when the environmental attribution intensity is low and the driver switching evidence is strong, it is determined as a driver switching event, and the identity label is updated to the candidate driver with the highest match score; when the evidence intensity of both types is at a moderate level, enter the uncertain region, and start the extended observation mechanism to continuously collect subsequent driving data until the evidence is sufficient. Calculate the driver switching confidence C switch = sigmoid(α × S evidence - β × S env + γ × D deviation ), where S evidence = max(S match_k × S consist_k ) represents the strongest switching evidence score, α, β, γ are fusion weight parameters, and the sigmoid function normalizes the output to the interval 0 to 1. When Cswitch exceeds the decision threshold θ switch , output the switching decision result as confirmed switching, otherwise as maintaining original identity or pending observation.

[0261] The driver switching confidence C switch and the switching decision result are classified by confidence level and encapsulated. According to the confidence value, the decision result is classified into three levels of high confidence, medium confidence and low confidence, which correspond to the subsequent disposal strategies of direct execution, auxiliary verification and manual review respectively. The decision result, confidence level, and supporting evidence summary (including main deviation feature dimension, environmental attribution analysis conclusion, and candidate driver matching situation) are encapsulated as a structured switching detection intermediate result and transmitted to the result output module.

[0262] The switching detection intermediate result is read, and the switching decision conclusion, confidence level, switching time point, new driver identity matching result and key evidence information are integrated. A structured final switching detection report is generated according to the pre-defined template, which includes detailed analysis process for manual review and standardized decision label for automatic processing of the system. The decision result in the final switching detection report is compared with the authorized driver list of the system: if the driver after switching belongs to the authorized user, update the current driver identity label and synchronize the personalized driving assistance parameters; if the driver after switching is not in the authorized list or the identity cannot be confirmed, immediately send a safety warning instruction to the Internet of Vehicles platform to trigger remote monitoring intervention or vehicle function restriction and other safety response measures.

[0263] In a detailed embodiment, a certain commercial vehicle is equipped with the intelligent driving behavior analysis system of the present application. The driving behavior benchmark of the vehicle owner Zhang is pre-registered in the system. At 9:00 am on a certain day, the vehicle starts in the company parking lot. The system collects CAN bus data and environmental perception data in real time. At 9:00:15, it is detected that the vehicle speed increases from 0 to 5 km / h, which exceeds the start threshold of 3 km / h, and the transition period start time t start =9:00:15 is determined. The average acceleration a*=1.2 m / s 2 within 2 seconds after starting is calculated. Based on the standard acceleration a ref =1.5 m / s 2 , the dynamic adjustment factor k adj =1.5 / 1.2=1.25 is calculated. The current is flat ground (slope <5°), normal temperature (water temperature 65°C), and empty state, and no additional compensation is needed. The final window length T window =10×1.25=12.5 seconds. The current environment is sunny, urban road, and light congestion, and the coding generates an environment context state vector E context of 55 dimensions. The system extracts the transition period behavior feature vector x within 12.5 seconds, and calculates its minimum Mahalanobis distance dmin = 1.8. Based on the EVT, the P unknown = 0.12, and the closed-set posterior probability P(zhang | x) = 0.88. The open-set probability distribution is [0.78, 0.22] (zhang, unknown). The system detects that the current driving behavior is more conservative than Zhang's historical benchmark, with a 10% lower speed and a 15% larger following distance. The E context The input causal forest model estimates that the expected behavior change under light congestion conditions is a speed reduction of 8%-12% and an increase in following distance of 10%-20%. The causal residual is calculated as: speed residual = -2%, following distance residual = 0%, both within the statistical significance threshold (±5%). The environmental attribution explanation degree index R env = 0.92. The system calculates the instantaneous switching evidence strength E switch = (1-0.78) x (1-0.8 x 0.92) = 0.22 x 0.26 ≈ 0.06. After the cumulative log-likelihood ratio is updated, Λ = -1.2, which is far below the upper boundary A = 2.94 and close to the lower boundary B = -2.20. After 5 consecutive sampling periods, Λ converges to -2.35 < B, and the system outputs the original identity determination. The system confirms that the current driver is the vehicle owner Zhang, and automatically loads his personalized driving mode settings without triggering a safety warning. This embodiment demonstrates how the system correctly identifies the legal driver and excludes environmental factors (congestion) that interfere with behavior changes through causal attribution mechanisms, avoiding false positives.

[0264] The present application introduces an open-set identification mechanism based on Gaussian mixture models and extreme value theory. By fitting the tail characteristics of the sample distance distribution, the system can calculate the probability that the observed sample belongs to an unknown category, intercept illegal takeovers by unregistered drivers, and improve the safety and anti-theft capabilities of the system. A causal attribution analysis framework based on causal forest and double machine learning is constructed to quantify the conditional average treatment effect of environmental factors on driving behavior. By calculating the causal residual between actual behavior and counterfactual expectations, the non-essential influence of environmental factors (such as rain and snow, congestion) on behavior is mathematically removed, effectively avoiding false positives due to environmental changes. Using sequential probability ratio test decision mechanism, by setting dynamic decision boundaries based on false positive rate and false negative rate, and accumulating log-likelihood ratio evidence over time, the system can respond within milliseconds when the evidence is strong, and automatically extend the observation when the evidence is ambiguous, achieving the optimal response speed while ensuring the accuracy of the determination.

[0265] The above detailed the preferred embodiments of the present application, but the present application is not limited to the specific details in the above embodiments, and various equivalent transformations of the technical solutions of the present application can be made within the technical concept of the present application, which all belong to the protection scope of the present application.

Claims

1. A method for intelligent driving behavior analysis based on vehicle-to-everything (V2X) big data, characterized in that: include: Real-time acquisition of multi-source driving behavior data streams and external environment perception data of target vehicles; Based on the vehicle motion state parameters in the multi-source driving behavior data stream, dynamic segments of driving behavior during the transition period corresponding to the vehicle start-up process are extracted. Feature encoding is performed on the external environment perception data to extract and generate an environmental context state vector representing the current driving scenario; The transition period driving behavior segments are compared with the pre-stored candidate driver historical behavior benchmarks to calculate the identity consistency likelihood index, which represents the degree of consistency between the current driving behavior and the historical benchmarks. Based on the environmental context state vector, the correlation between behavioral characteristic changes and environmental factors in driving behavior segments during the transition period is analyzed, and an environmental attribution explanatory power index is generated to represent the degree to which behavioral changes can be explained by environmental factors. The driver switching determination result is generated by combining the identity consistency likelihood index and the environmental attribution explanatory power index. Based on this, vehicle control strategy adjustments or safety warning operations will be performed; Calculating the identity consistency likelihood metric includes: Multi-dimensional feature extraction is performed on driving behavior segments during the transition period to generate transition period behavior feature vectors; Obtain the historical behavior probability density distribution model of each candidate driver pre-constructed based on historical samples in the historical behavior benchmark of candidate drivers; Input the transitional behavior feature vectors into the historical behavior probability density distribution model of each candidate driver, calculate the posterior probability value or likelihood value of the transitional behavior feature vectors belonging to the distribution space of each candidate driver, and use it as the identity consistency likelihood index. Generative environment attribution explanatory power metrics include: Extract the actual behavioral change characteristics of driving behavior segments during the transition period relative to historical baselines; Based on the environmental context state vector, and using a pre-defined environmental impact assessment model, the expected behavioral impact of current environmental factors on driving behavior characteristics is determined. Calculate the attribution bias between the actual characteristics of behavioral changes and the expected impact of behavioral changes; Based on the significance of attribution bias, an index for the explanatory power of environmental attribution is quantified. Generate driver switchover determination results, including: We weighted and fused the identity consistency likelihood index and the environmental attribution explanatory power index at the current moment to generate the instantaneous switching evidence strength that represents the possibility of the driver switching at the current moment. Based on the instantaneous switching of evidence strength, the decision state of the previous moment is updated in time sequence to generate a cumulative decision statistic that reflects the overall evidence level from the start of the transition period to the current moment. The cumulative decision statistics are compared with the preset judgment threshold range, and the driver switching judgment result is generated based on the comparison result.

2. The method according to claim 1, characterized in that, The probability density distribution model of historical behavior is a Gaussian mixture model; Calculate the posterior probability value or likelihood value, including: Calculate the minimum Mahalanobis distance from the transitional behavior feature vector to the center of each Gaussian component in the Gaussian mixture model; By fitting the tail of the minimum Mahalanobis distance distribution, a distance extreme value distribution model is constructed; Based on the distance extreme value distribution model, the probability that the transition period behavior feature vector does not belong to any pre-stored candidate driver unknown category is calculated; By combining the unknown category probability with the posterior probability of the Gaussian mixture model, an open set probability distribution containing the unknown category dimension is generated and determined as the identity consistency likelihood index.

3. The method according to claim 1, characterized in that, The historical behavior probability density distribution model is a single Gaussian distribution model that includes a mean vector and a covariance matrix; Calculate the posterior probability value or likelihood value, including: The instantaneous likelihood of the current transition period behavior feature vector belonging to each candidate driver is calculated based on a single Gaussian distribution model. Obtain the cumulative posterior probability of the previous time step, multiply the instantaneous likelihood by the cumulative posterior probability of the previous time step, and normalize the product to obtain the cumulative posterior probability of the current time step. The cumulative posterior probability at the current moment is determined as the identity consistency likelihood index.

4. The method according to claim 1, characterized in that, The environmental impact assessment model is a causal forest model, the expected behavioral impact is the counterfactual expected behavioral change, and the attribution bias is the causal residual; Quantitative indicators of the explanatory power of environmental attribution include: The environmental context state vector is input into a pre-trained causal forest model to estimate the conditional average treatment effect of environmental factors on driving behavior characteristics of each dimension; based on this, the counterfactual expected behavior changes that should occur under the current environmental conditions are calculated. Subtracting the counterfactual expected behavioral changes from the actual behavioral changes yields the causal residuals after removing environmental influences; A statistical significance test is performed on the causal residuals. When the causal residuals are not statistically significant, an environmental attribution explanatory power index is generated to indicate high explanatory power.

5. The method according to claim 1, characterized in that, The environmental impact assessment model is an environmental-behavioral response knowledge base, and the expected behavioral impact is the expected range of change. Quantitative indicators of the explanatory power of environmental attribution include: Based on the environmental context state vector, retrieve matching environmental-behavior response rules from a pre-stored environmental-behavior response knowledge base; Analyze the environment-behavior response rules to determine the expected range of change of each dimension of driving behavior characteristics in the current environment; Determine whether the actual behavioral changes fall within the expected range of change, and statistically analyze the proportion of feature dimensions that fall within the expected range of change. The proportion was determined as an indicator of the explanatory power of environmental attribution.

6. The method according to claim 1, characterized in that, The cumulative decision statistic is a series of cumulative log-likelihood ratios; The cumulative decision statistic is compared with a preset decision threshold range, including performing a sequential probability ratio test: Based on the preset expected false positive rate and expected false negative rate, calculate the upper and lower decision boundaries of the sequential test; At each sampling time, determine the relationship between the current value of the cumulative log-likelihood ratio sequence and the upper and lower decision boundaries; If the current value is greater than the upper decision boundary, a decision result confirming the switch is generated; If the current value is less than the lower decision boundary, a decision result is generated to maintain the original identity; If the current value is between the lower decision boundary and the upper decision boundary, a decision to continue observation is generated and data for the next time step is collected.

7. The method according to claim 1, characterized in that, Extract and generate the environment context state vector, including: External environment perception data is analyzed into discrete environmental elements and continuous environmental elements; One-hot encoding is performed on discrete environmental elements to generate sparse feature vectors; Based on preset quantile or distribution statistical characteristics, adaptive interval binning is performed on continuous environmental elements to map continuous values ​​into discrete interval codes. The sparse feature vector is concatenated and normalized with the discrete interval code to generate the environment context state vector.

Images