An automobile driving behavior data collection method based on internet of things

Abstract

The application discloses a kind of based on Internet of Things's automobile driving behavior data acquisition method, belong to intelligent network connection automobile data acquisition technical field.This method obtains multiple source driving behavior data and carries out time benchmark alignment, to the aligned driving data stream is controlled right switch identification, to the switch mark data stream is differentiating resampling, to candidate driving behavior segment is standardized encapsulation, to standardization driving behavior data packet is fragment level annotation, to the annotated driving behavior data packet is collected and filed processing, obtains target driving behavior acquisition result.The present application is completed time alignment, right of control attribution determination, credibility evaluation and transmission disposal decision in acquisition stage, to a certain extent, the problem that the data dilution of key switching period caused by the uniform acquisition of fixed frequency in prior art and the problem of relying on cloud post-processing are solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicle data acquisition technology, specifically relating to a method for acquiring vehicle driving behavior data based on the Internet of Things. Background Technology

[0002] In recent years, with the development of intelligent connected vehicles towards mass production and large-scale deployment, mass-produced vehicles with combined driving assistance functions have become the main operating vehicles in complex road scenarios such as urban expressways, ramps, toll stations, construction zones, tunnel entrances, congested following zones, and underground parking garage entrances and exits. During actual driving, vehicles frequently switch between manual driving, assisted driving, and takeover transition states. Simultaneously, they need to handle issues such as different sampling frequencies, inconsistent time bases, short-term packet loss, and unclear control attribution among multiple data sources, including the CAN bus, inertial sensors, positioning modules, driver monitoring units, combined driving assistance state channels, and onboard communication modules. Under such conditions of frequent switching between real-world driving and assisted driving, the semantic integrity, time alignment accuracy, and segment-level reliability of driving behavior data exhibit significant time-varying characteristics. This leads to problems such as data dilution during critical takeover periods, ambiguity of control attribution, unevaluable data collection quality, and a lack of basis for data upload strategies, thereby affecting the reliability and consistency of subsequent driving behavior analysis, liability tracing, model training, and insurance pricing.

[0003] Currently, the industry typically uses in-vehicle terminals or T-Boxes to connect to the CAN bus and GPS, collecting vehicle speed, steering, braking, acceleration, and location information at fixed sampling intervals. Events such as rapid acceleration, deceleration, and sharp turns are reported, or data is uploaded to the cloud after each trip. The backend then identifies and labels driving behavior based on threshold rules or statistical models. For vehicles with combined driver assistance functions, some solutions additionally record system on / off status or driver status information. Such solutions have formed a relatively mature technical path and application system for collecting vehicle operation data and triggering event reporting under single manual driving conditions.

[0004] However, the above-mentioned solutions still have significant limitations in applications involving frequent switching between human and assisted driving: On the one hand, existing data collection links still organize driving behavior data using a fixed frequency and uniform collection method throughout the process, making it difficult to shift the sampling focus around the control switching point and the system capability boundary change point. This results in the inability to accurately align steering wheel torque, pedal changes, vehicle posture changes, system state changes, and driver operation intentions within critical seconds, and the critical switching period is diluted by low-frequency sampling. On the other hand, existing data collection links typically perform cleaning, data point supplementation, and label correction in the cloud, rather than establishing a unified time benchmark and observability evaluation for multi-source data during the vehicle-side data collection stage. This makes it impossible to synchronously output standardized driving behavior data packages, control ownership labels, and data collection credibility labels at the collection end, and it is also impossible to provide transmission processing instructions for immediate upload, re-upload, or cache retention based on the credibility results. Consequently, it is impossible to simultaneously meet the requirements of real-time performance, completeness, credibility, and traceability. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a method for collecting vehicle driving behavior data based on the Internet of Things. The purpose is to at least partially solve the following problems in existing collection schemes: the use of fixed-frequency uniform collection leads to the dilution of key data during the control switching period, and the inability to synchronously output one or more of the standardized data packets, control ownership labels, and collection credibility labels at the collection end due to the fact that the annotation and correction are performed in the cloud post-processing.

[0007] To address the above problems, the present invention provides the following technical solution: A method for collecting vehicle driving behavior data based on the Internet of Things, comprising: Acquire multi-source driving behavior data and align the multi-source driving behavior data with a time reference to obtain an aligned driving data stream; The control switchover is identified by performing control switchover identification on the aligned driving data stream to obtain a switchover marker data stream; Differential resampling is performed on the switching marker data stream to obtain candidate driving behavior segments; The candidate driving behavior segments are standardized and encapsulated to obtain standardized driving behavior data packets; The standardized driving behavior data packet is annotated at the fragment level to obtain an annotated driving behavior data packet; The labeled driving behavior data packets are collected and archived to obtain the target driving behavior collection results.

[0008] As a preferred embodiment, the acquisition of multi-source driving behavior data and the time-base alignment of the multi-source driving behavior data to obtain an aligned driving data stream include: Vehicle operating status data, driver behavior data, and driver assistance system status data are acquired separately to obtain multi-source driving behavior data; The local timestamps of each data source in the multi-source driving behavior data are mapped to a unified clock according to the same reference time axis to obtain unified time axis data; Short-term packet loss detection is performed on the unified timeline data to obtain packet loss marker data, and missing segments are registered in the packet loss marker data to obtain an aligned driving data stream.

[0009] As a preferred embodiment, the step of identifying the control handover in the aligned driving data stream to obtain a handover marker data stream includes: The sliding window state transition mutation detection algorithm is used to detect state jumps in the driving assistance system state signals in the aligned driving data stream to obtain system switching events; Abrupt changes are detected in the steering wheel torque signal and brake pedal signal in the aligned driving data stream to obtain manual intervention events; The system switching event and the manual intervention event are fused and determined to obtain a switching marker data stream.

[0010] As a preferred embodiment, the fusion determination of the system switching event and the manual intervention event includes: The torque threshold verification is performed on the steering wheel torque change in the aforementioned manual intervention event to obtain the torque verification result; The change in the brake pedal during the aforementioned manual intervention event is subjected to a pedal threshold verification to obtain the pedal verification result. The duration of the sustained event in the aforementioned human intervention event is verified using a duration threshold to obtain the duration verification result. The torque verification result, the pedal verification result, and the duration verification result are combined for a pass determination to obtain a valid intervention event; The effective intervention events and the system switching events are merged and arranged in chronological order, and duplicate markers are removed to complete the fusion determination.

[0011] In a preferred embodiment, the differential resampling of the switching marker data stream to obtain candidate driving behavior segments includes: Taking each switching marker in the switching marker data stream as the center, extract the preceding window data of a preset duration and the following window data of a preset duration to obtain the window data segment; The window data segment is resampled differentially using an information entropy adaptive hierarchical resampling algorithm to obtain candidate driving behavior segments.

[0012] As a preferred embodiment, the step of extracting the preceding window data for a preset duration and the following window data for a preset duration includes: For each switching marker, backtrack for a preset time period, extract high-frequency data for the corresponding time period from the circular buffer, and obtain the front window data; High-frequency data of each switching marker is collected in real time for a preset time backward to obtain the data of the rear window; The data in the preceding window and the data in the following window are concatenated in chronological order to extract the window data segment.

[0013] In a preferred embodiment, the standardization and encapsulation of the candidate driving behavior segments to obtain a standardized driving behavior data packet includes: The various signals in the candidate driving behavior segments are time-aligned at the segment level to obtain a time-aligned signal set, and the time-aligned signal set is resampled at a uniform frequency to obtain an aligned signal set. The aligned signal set is semantically encapsulated according to the segment start and end time, vehicle identification, control quantity sequence, system state sequence, driver operation sequence, positioning trajectory sequence, and missing field registration results to obtain a standardized driving behavior data packet.

[0014] In a preferred embodiment, the step of performing fragment-level annotation on the standardized driving behavior data packet to obtain annotated driving behavior data packet includes: The dominant control source is determined for the system state, driver operation, and vehicle response at each sampling time in the standardized driving behavior data packet, and the control ownership label is obtained. The data integrity and inter-source consistency of the standardized driving behavior data packet are evaluated to obtain a collection credibility label. The control ownership label and the collection credibility label are associated and assembled to obtain a labeled driving behavior data packet.

[0015] As a preferred embodiment, the reliability calculation of the data integrity and inter-source consistency of the standardized driving behavior data packet includes: The time base integrity, key field integrity, inter-source consistency, and packet loss rate of the standardized driving behavior data packets are calculated to obtain various reliability indicators. The credibility fusion calculation is performed on each of the credibility indicators using a mutual information decoupling weighted fusion algorithm to obtain the segment-level acquisition credibility value; The credibility values ​​of the segment-level acquisitions are categorized by threshold to complete the credibility calculation.

[0016] As a preferred embodiment, the step of collecting and archiving the labeled driving behavior data packet to obtain the target driving behavior collection result includes: The labeled driving behavior data packets are classified by confidence level to obtain classified driving behavior data packets, and the classified driving behavior data packets are sorted according to transmission priority to obtain sorted driving behavior data packets. The sorted driving behavior data packets are transmitted and processed to obtain processed driving behavior data packets, and the processed driving behavior data packets are archived and registered to obtain the target driving behavior collection results.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. By performing differentiated resampling around the control handover anchor point, compared with the fixed-frequency uniform acquisition method, higher density sampling data can be retained during the control handover period. This helps to reduce the dilution of multi-source signals by low-frequency sampling during critical handover, and can improve the temporal resolution and semantic integrity of driving behavior data during critical periods to a certain extent.

[0018] 2. By synchronously generating control ownership labels and collection credibility labels for standardized driving behavior data packets during the collection phase, compared to the method of labeling and correcting only in the cloud post-processing phase, it helps to ensure that each data packet carries ownership and availability information at the time of generation, thereby reducing the dependence on cloud backhaul cleaning and improving the real-time performance and autonomy of the collection link.

[0019] 3. By classifying, sorting, and transmitting data packets labeled with driving behavior based on the collection credibility tags, compared to the method of uploading them all at once after a single trip, it helps to complete the priority decision of data packets at the vehicle end to a certain extent, so that high-credibility data is transmitted first and low-credibility data is retained locally for supplementary collection, thereby improving the overall credibility level of uploaded data under limited transmission bandwidth conditions. Attached Figure Description

[0020] Figure 1 An exemplary flowchart of an Internet of Things-based vehicle driving behavior data collection method provided in an embodiment of the present invention.

[0021] Figure 2 A comparison diagram showing the effects of the IoT-based vehicle driving behavior data collection method provided in this embodiment of the invention with existing technologies. Detailed Implementation

[0022] To make the technical means, creative features, and achieved objectives and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention. Unless otherwise specified, the experimental methods in the following embodiments are conventional methods, and the materials and reagents used in the following embodiments are commercially available unless otherwise specified.

[0023] Example 1 combined Figure 1 The following is an exemplary flowchart of an IoT-based vehicle driving behavior data collection method, and the specific implementation steps are as follows: Acquire multi-source driving behavior data and align the multi-source driving behavior data with a time reference to obtain an aligned driving data stream; The multi-source driving behavior data is a set of raw driving process data synchronously collected from different types of sensing devices and status channels through the vehicle Internet of Things acquisition link, which is used to provide a complete data input basis for subsequent control switch identification and differentiated resampling.

[0024] Specifically, the vehicle speed, steering wheel angle, steering wheel torque, brake pedal opening and accelerator pedal opening are read through the CAN bus; longitudinal acceleration, lateral acceleration and yaw rate are read through inertial sensors; and vehicle longitude, vehicle latitude and vehicle heading angle are read through the positioning module. The above signals are combined to obtain vehicle operating status data.

[0025] The driver's gaze direction and head posture angle are read by the gaze tracking module in the driver monitoring unit, the eyelid opening and closing degree is read by the eyelid detection module, and the hand position status is read by the hand detection module. The above signals are combined to obtain driver behavior data.

[0026] By combining the driver assistance system status channels, the start / stop status, current operating mode, takeover prompt status, and system availability level of the driver assistance system are read and aggregated to obtain driver assistance system status data. The vehicle operating status data, driver behavior data, and driver assistance system status data are then aggregated to obtain multi-source driving behavior data. This multi-source driving behavior data refers to the raw data set characterizing the vehicle's driving process and changes in control state, acquired through the vehicle IoT acquisition link from the vehicle operating status channel, driver behavior channel, and driver assistance system status channel.

[0027] The nominal sampling periods of the above data sources are different: 10ms for CAN bus data, 5ms for inertial sensor data, 100ms for positioning module data, 33ms for driver monitoring unit data, and 20ms for combined driving assistance status channel data. In addition, the clock sources of each data source are independent of each other, and there are start offsets and clock drifts.

[0028] The time reference alignment is a process of uniformly mapping the local timestamps carried by each data source to the same reference time axis, and performing packet loss detection and missing data registration on the mapped data. This is used to eliminate time axis differences between different data sources and to ensure that the sampling points of each channel corresponding to the same physical moment can be accurately matched.

[0029] Specifically, the time base alignment goes through three stages in sequence: unified clock mapping, short-term packet loss detection, and missing segment registration.

[0030] During the unified clock mapping phase, the satellite time synchronization timestamp output by the positioning module is selected as the reference time axis. When data acquisition for each channel begins, the local timestamp of the first sampling point of that channel and the satellite time synchronization timestamp at the same moment on the reference time axis are read. The initial offset of that channel is obtained by subtracting the satellite time synchronization timestamp from the local timestamp.

[0031] For each sampling point in the channel, the local timestamp of the sampling point is subtracted from the initial offset, and the result of the subtraction is used as the mapped timestamp of the sampling point on the reference time axis. After completing the process point by point, the time starting point of all sampling points in the channel is aligned with the reference time axis, thus completing the translation correction.

[0032] After translation correction, due to slight differences in the clock crystal frequencies of each channel, the mapped timestamps of each channel will gradually deviate from the reference time axis as the acquisition time increases. To eliminate this cumulative drift, the acquisition time is divided into several verification segments according to a preset time interval. At the end of each verification segment, the difference between the mapped timestamp of that channel and the satellite time synchronization timestamp at the same moment on the reference time axis is calculated to obtain the residual deviation at the end of that segment. Using the residual deviations at the ends of two adjacent verification segments as endpoint values, the mapped timestamps of all sampling points between the two verification segments are proportionally interpolated and corrected according to their time positions. That is, the sampling point closer to the end of the previous verification segment has a greater proportion of correction based on the previous residual deviation value, and the sampling point closer to the end of the next verification segment has a greater proportion of correction based on the next residual deviation value. The correction amount is superimposed on the mapped timestamps of each sampling point. After completing segment by segment, the cumulative drift of each channel is eliminated. After translation correction and linear drift compensation, the sampling points corresponding to the same physical moment in the data of each channel are aligned to the same time coordinate, resulting in unified time axis data.

[0033] During the short-term packet loss detection phase, the unified timeline data is traversed channel by channel. For the current channel, the nominal sampling period of the channel is read. Starting from the first sampling point of the channel, the mapped timestamp of the current sampling point and the mapped timestamp of the next sampling point are taken. The mapped timestamp of the current sampling point is subtracted from the mapped timestamp of the next sampling point to obtain the actual time interval between adjacent sampling points. The actual time interval is divided by the nominal sampling period of the channel to obtain the interval deviation ratio.

[0034] The interval deviation ratio is compared with a preset deviation threshold. When the interval deviation ratio is less than or equal to the preset deviation threshold, the sampling at that position is determined to be normal. Then, the sampling point is moved backward and the above difference and ratio calculation is repeated.

[0035] When the interval deviation ratio exceeds a preset deviation threshold, a short-term packet loss is determined to exist at that location. The mapping timestamp of the current sampling point is recorded as the packet loss start time. The interval deviation ratio is then calculated point by point until it falls back below the preset deviation threshold. The mapping timestamp of the sampling point corresponding to the fallback time is recorded as the packet loss end time. The packet loss duration is obtained by subtracting the packet loss start time from the packet loss end time, and the channel identifier where the short-term packet loss occurred is recorded. The packet loss start time, the packet loss end time, the packet loss duration, and the channel identifier are encapsulated into a packet loss record. After traversing the remaining sampling points of the channel in the above manner, the process is repeated for the next channel. After all channels have been traversed, all packet loss records generated by each channel are aggregated to obtain packet loss marker data.

[0036] During the missing segment registration phase, all packet loss records in the packet loss marker data are grouped by channel identifier, ensuring that packet loss records from the same channel are grouped together. Within each channel group, the packet loss records are then sorted from earliest to latest according to their packet loss start time.

[0037] After sorting, starting from the first packet loss record in the channel group, take the packet loss end time of the current packet loss record and the packet loss start time of the next packet loss record, and subtract the packet loss end time of the current packet loss record from the packet loss start time of the next packet loss record to obtain the interval between adjacent packet loss records.

[0038] The interval duration is compared with a preset merging threshold. When the interval duration is less than the preset merging threshold, the current packet loss record and the next packet loss record are merged into a missing segment record. The packet loss start time of the current packet loss record is taken as the missing start time of the missing segment record, and the packet loss end time of the next packet loss record is taken as the missing end time of the missing segment record. The missing duration is obtained by subtracting the missing start time from the missing end time. After merging, the interval duration of the missing segment record is compared with the next packet loss record, and the above merging determination is repeated until the interval duration is greater than or equal to the preset merging threshold. At this point, merging stops, the current missing segment record is finalized, and the next packet loss record is used as the new starting record to continue the determination process.

[0039] When the interval duration is greater than or equal to a preset merging threshold, the current packet loss record is directly retained as an independent missing segment record and is not merged with the next packet loss record. After all packet loss records in this channel group are determined, the process is repeated for the next channel group. After all channel groups are processed, all missing segment records generated by each channel are aggregated to generate a missing segment registration table. The unified timeline data is associated and encapsulated with the missing segment registration table to obtain the aligned driving data stream; the aligned driving data stream refers to a time-series data stream formed by associating vehicle operating status data, driver behavior data, driving assistance system status data, and missing segment registration information under a unified reference timeline, which is used as input for subsequent control handover identification.

[0040] The control switchover is identified by performing control switchover identification on the aligned driving data stream to obtain a switchover marker data stream; The control switch identification process involves extracting and fusing events from the driver assistance system status signals and manual operation signals in the aligned driving data stream, and writing the determination results to the corresponding time position. This process is used to determine the specific time and type of control switch during driving. The switch marker data stream is time-series data formed by writing the switch time, switch type, event source, and marker sequence number onto the aligned driving data stream.

[0041] Specifically, the control handover identification includes three steps: using a sliding window state transition mutation detection algorithm to detect state jumps in the driving assistance system status signals to obtain system handover events; performing mutation detection on the steering wheel torque signal and brake pedal signal to obtain manual intervention events; and fusing the system handover events and manual intervention events to obtain a handover marker data stream.

[0042] The sliding window state transition abrupt change detection algorithm is a processing algorithm that jointly quantizes discrete state changes and continuous level changes in the state signal of the driving assistance system to identify state jumps.

[0043] In the state transition detection phase, the start / stop status, current operating mode, takeover prompt status, and system availability level of the driver assistance system are first read from the aligned driving data stream to form a driver assistance system state signal sequence. Using the start time on a unified reference time axis as the starting point of the first window, all driver assistance system state sample values ​​within a continuous 300ms range from this starting point are extracted as the first state sliding window data segment. Subsequently, the window starting point is moved forward 50ms along the unified reference time axis to extract the next state sliding window data segment. This process continues until the window endpoint reaches the end of the driver assistance system state signal sequence, resulting in a set of state sliding window data segments.

[0044] For each state sliding window data segment, the start / stop status value and current operating mode value of the driver assistance system corresponding to its start and end positions are read respectively. The start / stop status value at the end position is compared with the start / stop status value at the start position; if they are different, it is recorded as a start / stop state transition; if they are the same, it is recorded as a start / stop state hold. The operating mode value at the end position is compared with the operating mode value at the start position; if they are different, it is recorded as an operating mode transition; if they are the same, it is recorded as an operating mode hold. Simultaneously, the takeover prompt status value and system availability level value corresponding to the start and end positions are read. The change in prompt value is obtained by subtracting the takeover prompt status value at the start position from the takeover prompt status value at the end position, and the change in availability is obtained by subtracting the system availability level value at the start position from the system availability level value at the end position.

[0045] The above start / stop state transition results, operating mode transition results, prompt changes, and availability changes are input into the sliding window state transition abrupt change detection algorithm to calculate the state jump variable corresponding to each state sliding window data segment. Since the start / stop state and current operating mode of the driver assistance system are discrete state quantities, their changes are manifested in whether the state at the beginning and end of the window switches. Therefore, the discrete state changes are first quantified to obtain the discrete state transition quantity of the i-th state sliding window data segment: , in, This represents the discrete state transition amount of the i-th state sliding window data segment; This represents the discrete state transition weight coefficient corresponding to the start / stop state of the driver assistance system. This represents the discrete state transition weight coefficient corresponding to the current operating mode. This indicates an indicator function that returns 1 if the condition within the parentheses is true, and 0 if the condition within the parentheses is false. This represents the start / stop status value of the driver assistance system corresponding to the starting position of the i-th state sliding window data segment. This represents the start / stop status value of the driver assistance system corresponding to the end position of the i-th state sliding window data segment. This represents the current operating mode value corresponding to the starting position of the i-th state sliding window data segment. This represents the current operating mode value corresponding to the end position of the i-th state sliding window data segment, s represents the value corresponding to the start position, e represents the value corresponding to the end position, and i is the index of the state sliding window data segment.

[0046] Furthermore, the takeover notification status and system availability level are continuously changing quantities, and their changes are manifested as differences in values ​​at the beginning and end of the window. Therefore, the continuous state changes are quantified to obtain the continuous state change quantity of the i-th state sliding window data segment: , in, This represents the continuous state change amount of the i-th state sliding window data segment. This represents the weighting coefficient of continuous state changes corresponding to the takeover alert status. This represents the weighting coefficient for continuous state changes corresponding to the system availability level. This represents the takeover prompt status value corresponding to the starting position of the i-th state sliding window data segment. This represents the takeover prompt status value corresponding to the end position of the i-th state sliding window data segment. This represents the system availability level value corresponding to the starting position of the i-th state sliding window data segment. This represents the system availability level value corresponding to the end position of the i-th state sliding window data segment.

[0047] After calculating the discrete and continuous state changes, the two are weighted and combined to obtain the state jump variable for the i-th state sliding window data segment: , in, This represents the state jump variable for the i-th state sliding window data segment; Based on the above definition, the sliding window state transition mutation detection algorithm first passes through Characterize the intensity of discrete state switching, and then through... Characterize the intensity of continuous state changes, and based on this, through A unified representation is provided for the overall state transition degree of the current state sliding window data segment.

[0048] Subsequently, the state jump variable With preset state transition threshold Compare; when When a state transition occurs in the i-th state sliding window data segment, the window end time of that state sliding window data segment is determined as the system switching time, and a system switching event is generated; when If no state transition occurs in the i-th state sliding window data segment, no system switching event is generated, and the process moves to the next state sliding window data segment for further detection. After all state sliding window data segments have been detected, a sequence of system switching events arranged in chronological order is obtained.

[0049] In this embodiment, Take 0.35, Take 0.25, Take 0.20, Take 0.20, The value is set to 0.5. The calibration process for the aforementioned weighting coefficients and state transition thresholds is as follows: During the trial operation phase, at least 200 manually labeled positive and negative samples are collected. Positive samples are state window data segments where a control switch has been manually confirmed, and negative samples are state window data segments where a control switch has not been manually confirmed. The mean values ​​of the state transition variables for the positive and negative sample sets are calculated separately. The objective is to maximize the difference between the mean values ​​of the positive and negative sample state transition variables. , , , Perform grid search adjustments; after determining the weight coefficients, take the lower 5th quantile of the positive sample state jump variable distribution as... Recommended value.

[0050] The mutation detection is a process of scanning the change amplitude of the steering wheel torque signal and the brake pedal signal window by window within a short time window, and extracting change events that exceed a preset change threshold, which is used to identify the moment when the driver actively applies operational intervention in the assisted driving state.

[0051] In the mutation detection stage, steering wheel torque signal and brake pedal signal are read from the aligned driving data stream to form a manual operation signal sequence. Taking the first time position of this sequence on a unified reference time axis as the starting point of the window, all steering wheel torque sample values ​​and brake pedal sample values ​​within a continuous 200ms range from this starting point are extracted as the first manual intervention sliding window data segment. Thereafter, the starting point of the window is moved back 20ms to extract the next data segment, and this process continues until the end, resulting in a set of manual intervention sliding window data segments.

[0052] For each manually intervened sliding window data segment, adjacent steering wheel torque sample values ​​are read sequentially according to the sampling order. The difference between the previous and subsequent values ​​is subtracted point by point to form a steering wheel torque change sequence. The change with the largest absolute value is extracted from this sequence to obtain the steering wheel torque change amount. Similarly, the difference between adjacent brake pedal sample values ​​is calculated point by point to form a brake pedal change sequence, and the change with the largest absolute value is extracted to obtain the brake pedal change amount. After obtaining the steering wheel torque change amount and the brake pedal change amount, the duration for which the steering wheel torque change amount continuously exceeds the steering wheel torque threshold and the duration for which the brake pedal change amount continuously exceeds the brake pedal change threshold are statistically analyzed. The shorter of the two is taken as the duration of sustained change.

[0053] In this embodiment, the steering wheel torque threshold is taken as... The brake pedal change threshold is set at 15% of the opening change, and the duration threshold is set at 300ms. The calibration method for the above three thresholds is as follows: during the trial operation phase, the distribution of steering wheel torque change, brake pedal change, and duration of continuous hold during the actual driver takeover and non-takeover processes are collected, and the lower 10th percentile of the distribution of each indicator in the actual takeover sample is taken as the recommended value of the corresponding threshold.

[0054] The system then proceeds to torque threshold verification, pedal threshold verification, and duration threshold verification. The change in steering wheel torque is compared to a steering wheel torque threshold; if the threshold is reached, the torque verification passes; otherwise, it fails. Similarly, the change in brake pedal position is compared to a brake pedal change threshold; if the threshold is reached, the pedal verification passes; otherwise, it fails. Finally, the duration of brake pedal engagement is compared to a duration threshold; if the threshold is reached, the duration verification passes; otherwise, it fails.

[0055] The joint pass determination is a process of simultaneously judging the pass conditions of the torque verification result, the pedal verification result, and the duration verification result. Only when all three verification results are pass results is the current manual intervention window data segment determined to constitute a valid intervention event, which is used to exclude non-substantive takeover behaviors such as the driver slightly touching the steering wheel or unintentionally stepping on the brake pedal.

[0056] After completing the three verifications, the torque verification result, pedal verification result, and duration verification result are jointly judged for pass. When all three results are pass, the window end time of the corresponding manual intervention sliding window data segment is determined as the manual intervention time. This manual intervention time, steering wheel torque change, brake pedal change, and duration of hold are written into the event record to generate a valid intervention event. When at least one of the three results fails, no valid intervention event is generated, the data segment is registered as an invalid intervention window, and the process moves to the next data segment for further testing. After all manual intervention sliding window data segments have been tested, a sequence of valid intervention events arranged in chronological order is obtained. This sequence is used as the manual intervention events in subsequent fusion judgments.

[0057] In the fusion determination phase, the event time, event type, and event source of each event in the system switching event sequence and the effective intervention event sequence are read separately and merged to form a set of events to be fused. This set is then sorted by event time from earliest to latest to obtain a time-sorted event set. Starting from the first event in the time-sorted event set, the event time of the current event is read, followed by the event time of the next event. The time interval between adjacent events is obtained by subtracting the current event time from the next event time. When this interval is less than 500ms, the two events are grouped into the same switching event group, and the current switching event group is compared with subsequent adjacent events using the same time interval. When the interval is greater than or equal to 500ms, the current event or the current switching event group is retained as an independent switching event group, and the next event is used as the new starting event to continue the determination process. After all events are grouped, a set of switching event groups is obtained.

[0058] Source analysis is performed on each switching event group in the switching event group set: when both system switching events and effective intervention events exist within the same switching event group, it is determined as a control switching event; when only system switching events exist, it is determined as a system-side switching event; and when only effective intervention events exist, it is determined as a manual-side switching event. The earliest event time in each switching event group is determined as the switching time, the determination result is determined as the switching type, and the event source combination participating in the grouping is used as the event source, generating a marker sequence number in chronological order. The effective intervention events and the system switching events are merged and arranged in chronological order, and duplicate markers appearing at the same time are removed. The switching time, switching type, event source, and marker sequence number corresponding to each removed switching event are written into the record position corresponding to the switching time in the aligned driving data stream. After all switching event groups are written, the switching marker data stream is obtained.

[0059] The effective intervention event is a manual intervention event that has passed the torque threshold verification, pedal threshold verification, and duration threshold verification, and is used to participate in the subsequent fusion judgment with system switching events.

[0060] The system switching event is a system-side control state change event extracted from the state transition results in the driving assistance system state signal. It includes at least the system switching time, the corresponding state sliding window number, and the driving assistance system start / stop status, current operating mode, takeover prompt status, and system availability level corresponding to the switching time.

[0061] The manual intervention event is a manual operation intervention event extracted based on the abrupt change detection results of the steering wheel torque signal and the brake pedal signal, and includes at least the time of manual intervention, the amount of change in steering wheel torque, the amount of change in brake pedal, and the duration of the intervention.

[0062] In one embodiment, the sliding window state transition abrupt change detection algorithm further includes: A window status record is generated for each state sliding window data segment in the driving assistance system status signal sequence. Each window status record includes at least the window start time, window end time, driving assistance system start / stop status, current operating mode, takeover warning status, and system availability level corresponding to that state sliding window data segment. After arranging the window status records in order of window number, the current window status record and the previous window status record are read sequentially, and the driving assistance system start / stop status, current operating mode, takeover warning status, and system availability level in the two are compared to determine whether there are any abrupt changes in the current window status record compared to the previous window status record.

[0063] Specifically, when the driving assistance system's start / stop status changes from on to off in the current window status record, or the current operating mode changes from assisted driving mode to takeover transition mode, or the takeover prompt status changes from no prompt to prompt, or the system availability level changes from available level to degraded level, the corresponding change is registered as a state transition candidate. All state transition candidates in the same state sliding window data segment are summarized to form the state transition candidate record corresponding to that state sliding window data segment. Subsequently, the continuity of the state transition candidate records is verified. When a state transition candidate of the same type appears consecutively in two adjacent state sliding window data segments, the state transition candidate is determined as a valid state transition item. When a state transition candidate of the same type appears only in a single state sliding window data segment and does not appear again in any of the preceding and following state sliding window data segments, the state transition candidate is determined as a transient disturbance item and removed from subsequent system switching event extraction. For valid state transition items after continuity verification, time positioning is performed. The end time of the state sliding window data segment where the first valid state transition item appears is determined as the system switching time, and the window state record corresponding to the system switching time is written into the system switching event sequence.

[0064] In a specific scenario, when a vehicle enters a tunnel on an urban expressway, the driver assistance system's status signal first shows a decrease in system availability, followed by an activation of the takeover prompt, and then the operating mode switches from assisted driving mode to takeover transition mode. The sliding window state transition abrupt change detection algorithm detects each window of continuous state sliding window data within this time period, identifying the aforementioned continuous state change chain as a valid state transition item and generating a system switching event. This ensures that subsequent differentiated resampling can focus the sampling on the time period when control actually changes.

[0065] Furthermore, to avoid interference from single-point jitter in the driver assistance system status signal on the extraction of system switching events, the sliding window state transition abrupt change detection algorithm also includes suppression processing of state transition candidate records. When a certain state field changes in the current state sliding window data segment but returns to its original state in the next state sliding window data segment, and the takeover prompt status and system availability level do not change synchronously, the change in the state field is judged as an instantaneous jitter and is not written into the system switching event sequence as a system switching event. By suppressing the instantaneous jitter, a stable system switching event sequence is obtained, providing input for subsequent fusion judgment of the system switching event and the manual intervention event.

[0066] Differential resampling is performed on the switching marker data stream to obtain candidate driving behavior segments; The differentiated resampling process involves extracting high-frequency driving behavior data before and after a switch, centered on the switch time corresponding to each switch marker in the switch marker data stream, and then performing adaptive hierarchical resampling of the extracted window data segments based on information entropy. The candidate driving behavior segment is a local time-series data segment formed around a single switch marker and obtained after differentiated resampling; the local time-series data segment covers the pre-switch stage, the switch occurrence stage, and the post-switch stage.

[0067] Specifically, the differentiated resampling includes two steps: window data segment extraction and information entropy adaptive hierarchical resampling.

[0068] In the window data segment extraction stage, each switching marker is first read from the switching marker data stream in chronological order, and the switching time, switching type, event source, and marker sequence number corresponding to each switching marker are extracted to form a switching marker sequence. For any switching marker in the switching marker sequence, using the switching time corresponding to that switching marker as the time center, a pre-window data of a preset duration is extracted forward, and a post-window data of a preset duration is extracted backward to obtain the window data segment. Specifically, for each switching marker, a preset duration is traced back, and high-frequency data of the corresponding time period is extracted from the circular buffer to obtain the pre-window data; The circular cache is a short-term data cache area in the vehicle edge gateway that is written to in a cyclic manner according to a fixed storage capacity. When the storage capacity is full, it automatically overwrites the earliest written data. It is used to retain high-frequency original records for a period of time before the control switch event occurs for backtracking retrieval.

[0069] The high-frequency data refers to the raw sampling records that are continuously output by each data source according to its own nominal sampling period without down-frequency processing.

[0070] The circular cache continuously stores high-frequency original records of the aligned driving data stream within the most recent preset cache duration, and cyclically overwrites them in chronological order. In this embodiment, the preset cache duration is 10 seconds, and the preset backtracking duration is 3 seconds. When the backtracking start time corresponding to a certain switching marker is within the effective time range of the circular cache, all high-frequency records with timestamps within that backtracking period are extracted from the circular cache one by one to obtain the front window data; when the backtracking start time corresponding to a certain switching marker is earlier than the earliest storage time of the circular cache, the earliest high-frequency record that can be extracted from the circular cache is used as the starting record of the front window data, and missing time periods are registered.

[0071] After extracting the preceding window data, the subsequent window data for a preset duration is extracted from each switching marker. This includes: using the switching time corresponding to the current switching marker as the starting time, high-frequency data for each switching marker is collected in real time for a preset duration to obtain the subsequent window data. In this embodiment, the preset duration for extraction is 5 seconds. High-frequency records continuously written within the subsequent window time range are cached according to timestamps from early to late. When the subsequent window time range ends, real-time acquisition corresponding to the current switching marker is stopped, and all cached high-frequency records are determined as the subsequent window data. When a sampling interruption occurs within the subsequent window time range, the interruption start time, interruption end time, and interruption duration are registered, and the corresponding missing information is written into the subsequent window data.

[0072] After extracting the data from the preceding and following windows, the preceding and following window data are concatenated in chronological order to extract the window data segment. Specifically, the preceding window data is arranged from earliest to latest timestamp as the first half of the window data segment, and the following window data is arranged from earliest to latest timestamp as the second half of the window data segment. The position of the switching time in the concatenated sequence is then used as the switching center index and written into the window data segment header information to obtain the window data segment corresponding to the switching marker. This process is repeated for all switching markers to obtain a set of window data segments.

[0073] The preceding window data consists of high-frequency data for the corresponding time period extracted from the circular cache by tracing back a preset time around the switching time corresponding to the switching marker.

[0074] The rear window data is high-frequency data collected in real time for a preset duration around the switching time corresponding to the switching marker.

[0075] The window data segment is a local time-series data segment formed by splicing the preceding window data and the following window data in chronological order, surrounding a single switching marker.

[0076] In the information entropy adaptive hierarchical resampling stage, each window data segment in the set of window data segments is first divided into several local statistical units according to a unified statistical interval. In this embodiment, the unified statistical interval is 100ms. For any local statistical unit, the steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, lateral acceleration signal, driver assistance system start / stop status, current operating mode, takeover prompt status, and system availability level are read to form a local feature set. When discretizing the local feature set, continuous signals are binned with equal width according to a preset amplitude range, wherein the steering wheel torque signal is divided according to... The brake pedal signal is divided into compartments based on a 5% opening angle based on the compartment width, and the vehicle speed signal is divided according to... The box is divided into boxes according to its width, and the longitudinal acceleration signal and the lateral acceleration signal are divided according to... Binning is performed based on the bin width; discrete state signals are directly classified according to state value categories. Specifically, the start / stop status of the driver assistance system is classified into two categories: on and off; the current operating mode is classified into three categories: assisted driving mode, takeover transition mode, and exit mode; the takeover prompt status is classified into two categories: inactive and active; and the system availability level is classified into three categories: high availability, medium availability, and low availability. After binning and classification, the frequency of occurrence of each feature category within the current local statistical unit is counted to obtain the frequency distribution.

[0077] After obtaining the frequency distribution of each feature category, an adaptive hierarchical resampling algorithm based on information entropy is applied to the window data segment to calculate different sampling densities based on the differences in the degree of change in driving behavior during different time periods before and after the switch. This algorithm determines the sampling level based on the complexity of changes in local statistical units within the window data segment and performs differentiated resampling. Specifically, each window data segment is first divided into several local statistical units according to a uniform statistical interval. Then, the steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, lateral acceleration signal, driver assistance system start / stop status, current operating mode, takeover prompt status, and system availability level in each local statistical unit are discretized to obtain the frequency distribution of each feature category within each local statistical unit. Based on this, the sample proportion of each feature category in the corresponding local statistical unit is calculated as the input for subsequent information entropy calculation. The sample proportion of the m-th feature in the j-th local statistical unit is defined as: , in, This represents the proportion of samples with the m-th feature in the j-th local statistical unit. This represents the frequency of the m-th feature in the j-th local statistical unit. Let represent the total number of samples in the j-th local statistical unit, where j represents the number of the local statistical unit, and m represents the feature category number obtained after discretization.

[0078] After obtaining the sample proportion of each feature category, the uncertainty of all feature categories within the j-th local statistical unit is further quantified to obtain the information entropy value corresponding to that local statistical unit. The algorithm formula for the information entropy value is as follows: , in, This represents the information entropy value of the j-th local statistical unit. This represents the total number of feature categories obtained after discretization in the j-th local statistical unit. It is used to characterize the distribution ratio of the m-th type of feature within the j-th local statistical unit in all samples.

[0079] Therefore, the information entropy value The larger the value, the more complex the representation of changes in driving behavior within the j-th local statistical unit; the information entropy value The smaller the value, the more stable the changes in driving behavior within the j-th local statistical unit.

[0080] Based on the aforementioned information entropy value, sampling levels are divided for each local statistical unit, and the corresponding resampling interval is further determined. Specifically, the resampling interval corresponding to the j-th local statistical unit is defined as: , in, This represents the resampling interval corresponding to the j-th local statistical unit. This indicates the high-frequency sampling interval corresponding to a high level of change. This indicates the intermediate frequency sampling interval corresponding to the level of variation. This indicates the low-frequency sampling interval corresponding to the low-level change. This represents the first threshold level corresponding to a high level of change. This represents the second grading threshold between the medium and low change levels.

[0081] After determining the resampling interval, differential resampling is performed on each local statistical unit. For those that meet the requirements... The local statistical units retain the original high-frequency sampling points according to the high-frequency sampling interval in order to preserve the rapid change process near the switch; For satisfying The local statistical units are sampled according to the intermediate frequency sampling interval to characterize the changing trend of the transition phase; For satisfying The local statistical unit extracts sampling points according to the low-frequency sampling interval, and merges and retains continuous repetitive state values ​​to compress data redundancy in the stable phase. In this embodiment, η1 is 0.85 and η2 is 0.45. Take 20ms. Take 50ms. Take 100ms.

[0082] Furthermore, to ensure the temporal resolution of key changes near the switching moment, a switching core segment is constructed within the window data segment, centered on the switching moment. For local statistical units within the switching core segment, instead of individually reducing the sampling frequency based on the information entropy value, a uniform high-frequency sampling interval is used for resampling. In this embodiment, the switching core segment takes a time range of 500ms before and after the switching moment. After resampling each local statistical unit, the obtained sampled data is reassembled in its original chronological order, retaining the switching moment location index, sampling level marker, and missing registration information to obtain candidate driving behavior segments.

[0083] In one embodiment, the information entropy adaptive hierarchical resampling algorithm further includes: The window data segment is partially segmented at a uniform statistical interval to obtain multiple local statistical units. The steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, lateral acceleration signal, driver assistance system start / stop status, current operating mode, takeover prompt status, and system availability level in each local statistical unit are jointly read. The frequency, amplitude, and number of state transitions of each signal within the current local statistical unit are statistically analyzed to obtain local change characterization results. The change level of each local statistical unit is determined based on the local change characterization results, and the corresponding sampling interval is determined according to the change level. Local statistical units with high change levels retain denser sampling points, local statistical units with medium change levels retain medium-density sampling points, and local statistical units with low change levels retain sparser sampling points. The data from each local statistical unit after resampling are then reassembled in the original time sequence to obtain candidate driving behavior segments.

[0084] Specifically, when a local statistical unit simultaneously exhibits rapid changes in steering wheel torque, a sudden increase in brake pedal opening, a short-term decrease in vehicle speed, and activation of the takeover warning status, that local statistical unit is identified as a high-change segment, and high-frequency resampling is applied to this high-change segment. When a local statistical unit only exhibits gradual changes in vehicle speed and longitudinal acceleration, while the driver assistance system's start / stop status, current operating mode, and takeover warning status remain unchanged, that local statistical unit is identified as a low-change segment, and low-frequency resampling is applied to this low-change segment. When a local statistical unit contains a single signal change that does not meet the criteria for a high-change segment, that local statistical unit is identified as a medium-change segment, and medium-frequency resampling is applied to this medium-change segment.

[0085] In a specific scenario, when a vehicle enters a tunnel on an urban expressway, changes in lighting conditions, changes in takeover warning status, a decrease in system availability, and an increase in driver steering torque occur consecutively before and after the switchover time corresponding to the switchover marker. After partially segmenting the window data segment corresponding to this switchover marker, the changes in multi-source signals within several local statistical units before and after the switchover time are significantly concentrated. The information entropy adaptive hierarchical resampling algorithm identifies these local statistical units as high-change segments and retains higher-density sampling points. In contrast, in the stable cruise segment before the switchover and the stable takeover segment after the switchover, the changes in multi-source signals are less. The information entropy adaptive hierarchical resampling algorithm identifies the corresponding local statistical units as medium-change or low-change segments and reduces the sampling density. Through the above processing, the obtained candidate driving behavior segments retain a more complete state evolution process near the switchover time and reduce redundant data in stable segments, thereby enabling the subsequent standardization and encapsulation to simultaneously obtain complete key behavioral information and continuous temporal correlations before and after the switchover.

[0086] Furthermore, to ensure data continuity at the switching center location, local statistical units within a preset range before and after the switching time are uniformly processed as high-change segments, ensuring continuous high-frequency recording of candidate driving behavior segments near the switching time. For local statistical units outside the preset range, corresponding resampling is performed based on their change level. By applying different sampling densities to different segments within the window data segment, the information entropy adaptive hierarchical resampling algorithm completes differentiated resampling of the window data segment and obtains candidate driving behavior segments.

[0087] The candidate driving behavior segments are standardized and encapsulated to obtain standardized driving behavior data packets.

[0088] The standardized encapsulation is a process of performing segment-level time alignment and unified frequency resampling on various signals from different data sources and different sampling frequencies in the candidate driving behavior segments, and encapsulating them according to a preset semantic field structure. This process is used to transform the non-uniform time-series data after differential resampling into standardized data units with unified structure and complete fields.

[0089] Specifically, the start and end times of the candidate driving behavior segments are first read, and a local time axis for each segment is constructed using the segment start and end times as endpoints. Segment-level time alignment is then performed on various signals within the candidate driving behavior segments, mapping the sampling times of each signal to their corresponding time positions on the segment's local time axis. Continuous signals are linearly interpolated at missing positions using the preceding and following valid sample values, while discrete state signals are continued at missing positions using the previous valid state value, resulting in a time-aligned signal set. This time-aligned signal set is the set of signals formed after performing segment-level time alignment on various signals within the candidate driving behavior segments.

[0090] A standard time sampling sequence is then generated on the local time axis of the segment at fixed time intervals of 50ms. The time-aligned signal set is then resampled at a uniform frequency according to this standard time sampling sequence to obtain the aligned signal set. The aligned signal set is a set of uniform time interval signals formed after performing uniform frequency resampling on the time-aligned signal set.

[0091] Subsequently, the aligned signal set is semantically encapsulated according to the segment start and end times, vehicle identification, control quantity sequence, system state sequence, driver operation sequence, positioning trajectory sequence, and missing field registration results. Specifically, steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, and lateral acceleration signal are written into the control quantity sequence; driver assistance system start / stop status, current operating mode, takeover prompt status, and system availability level are written into the system state sequence; steering wheel torque change record, brake pedal change record, and manual intervention time marker are written into the driver operation sequence; longitude value, latitude value, and trajectory position index are written into the positioning trajectory sequence. Sampling positions that still cannot be completed are registered as missing field registration results, resulting in a standardized driving behavior data package. The header field of the standardized driving behavior data package records the segment number, switching type, and segment start and end times, while the body field stores the above sequence data in chronological order according to the standard time sampling sequence.

[0092] The standardized driving behavior data packet is annotated at the fragment level to obtain an annotated driving behavior data packet.

[0093] The segment-level labeling is a process of evaluating the control source attribution and overall segment acquisition quality at each sampling moment in the standardized driving behavior data packet and writing the evaluation results into the data packet. This is used to ensure that each standardized driving behavior data packet carries a control source attribution label and an acquisition credibility label during the generation stage.

[0094] Specifically, the system status, driver operation, and vehicle response are read sequentially at each sampling moment in the standard time sampling sequence. The dominant control source is determined for these three records: a system control label is generated when the driver assistance system is in an "on" state and the driver's operation does not meet the intervention criteria; a manual control label is generated when the driver's operation meets the intervention criteria and the vehicle response is consistent with the driver's operation direction; and a transitional control label is generated when the takeover prompt is activated and the system status and driver operation alternate in adjacent sampling moments. The labels for all sampling moments are then summarized in chronological order to form a control ownership label. Next, four indicators are calculated for the standardized driving behavior data packet: time base integrity, key field integrity, inter-source consistency, and packet loss rate. Time base integrity is the ratio of valid time-aligned sampling points to total sampling points; key field integrity is the ratio of valid fields to required fields in each sequence; inter-source consistency is the ratio of sampling moments where the logical correspondence between the system status, driver operation, and vehicle response records is established to total sampling moments; and packet loss rate is the ratio of missing sampling points to total sampling points. Four indicators are input into a mutual information decoupling weighted fusion algorithm to calculate the credibility fusion value, resulting in a segment-level acquisition credibility value. This segment-level acquisition credibility value is then categorized according to a first credibility threshold and a second credibility threshold to obtain acquisition credibility labels. Finally, the control ownership label and the acquisition credibility label are associated and assembled to obtain a labeled driving behavior data package.

[0095] The mutual information decoupling weighted fusion algorithm is a processing algorithm that determines the decoupling weights based on the mutual information between each credibility index and performs fusion calculations on each credibility index.

[0096] The control ownership label is a set of time-series labels formed by marking the dominant control source corresponding to each sampling moment in the standard time sampling sequence. It is used to characterize whether the control at the corresponding sampling moment is dominated by the system, manually, or in a transitional state.

[0097] The labeled driving behavior data packets are collected and archived to obtain the target driving behavior collection results.

[0098] The data collection and archiving process is a process of classifying, sorting, transmitting, and archiving data packets according to the data collection credibility tags in the labeled driving behavior data packets. This is used to complete the credibility classification and transmission decision of data packets during the vehicle-side data collection stage.

[0099] The archiving registration is a process of writing the fragment identification information, annotation results, and handling methods of the data packets for handling driving behavior into the archiving index and storing them in the archiving storage area.

[0100] The transmission processing involves determining the corresponding data packet based on its classification and then performing an immediate upload, retransmission request, or local retention.

[0101] The labeled driving behavior data package is a structured fragment data package formed by writing control ownership tags and collection credibility tags into the standardized driving behavior data package, and is used as input for subsequent collection and archiving processing.

[0102] Specifically, the labeled driving behavior data packets are first divided into three levels—high confidence, medium confidence, and low confidence—based on the collection confidence labels to obtain graded driving behavior data packets. Then, the graded driving behavior data packets are sorted according to the rules of high confidence priority, control transition segment priority, and earlier segment end time priority to obtain sorted driving behavior data packets. Subsequently, the sorted driving behavior data packets are processed according to the graded results, including immediate upload, retransmission request, or local retention, to obtain processed driving behavior data packets. Finally, archive records are generated for the processed driving behavior data packets and written to archive storage to obtain the target driving behavior collection results.

[0103] The transmission priority is the order in which data packets are transmitted, determined by the classification category, control ownership label, and time sequence of the segments.

[0104] The driving behavior data packet being processed is the driving behavior data packet formed after completing the corresponding transmission processing of the immediate upload, retransmission request, or local retention.

[0105] The sorted driving behavior data packets are driving behavior data packets that have been sorted according to transmission priority and written with transmission sequence numbers.

[0106] The graded driving behavior data package is a driving behavior data package formed after grading based on the collected credibility tags.

[0107] Combination Figure 2 The above-described comparison chart of the effects of the IoT-based vehicle driving behavior data collection method shows that the black bars represent the technical effects of the present invention, while the gray bars represent existing technologies. Figure 2 It can be seen that the present invention is superior to the prior art to a certain extent.

[0108] Example 2, a method for collecting vehicle driving behavior data based on Example 1, is as follows: In one embodiment, the standardization and encapsulation of the candidate driving behavior segments to obtain a standardized driving behavior data packet includes: The standardized encapsulation is a process of performing segment-level time alignment and unified frequency resampling on various signals from different data sources and different sampling frequencies in the candidate driving behavior segments, and encapsulating them according to a preset semantic field structure. This process is used to transform the non-uniform time-series data after differential resampling into standardized data units with unified structure and complete fields.

[0109] After obtaining the candidate driving behavior segments, the system first reads the steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, lateral acceleration signal, driver assistance system start / stop status, current operating mode, takeover prompt status, system availability level, and positioning trajectory signal, arranged in chronological order within the candidate driving behavior segments. Then, it categorizes and groups these signals according to their source to obtain a multi-source signal set within each segment. Subsequently, it reads the switching time location index corresponding to each candidate driving behavior segment and constructs a segment local time axis based on the segment's start and end times. This segment local time axis covers the continuous time range between the segment's start and end times, and is used to perform segment-level processing on all types of signals under a unified time reference.

[0110] The segment-level time alignment is the process of uniformly mapping the sampling times of various signals from different data sources and different sampling frequencies in the candidate driving behavior segment to the same time coordinate on the local time axis of the segment, in order to eliminate the residual time deviation between signals within the same segment.

[0111] After constructing the local time axis of the segment, segment-level time alignment is performed on various signals in the candidate driving behavior segment to obtain a time-aligned signal set. Segment-level time alignment is the process of uniformly mapping the sampling times of various signals from different data sources and different sampling frequencies in the candidate driving behavior segment to the same time coordinate on the segment's local time axis, used to eliminate residual time deviations between signals within the same segment. The signals in the candidate driving behavior segment are classified into three categories according to their properties: continuous signals, discrete state signals, and positioning trajectory signals. The time alignment methods for these three types of signals differ depending on their signal properties.

[0112] For continuous signals such as steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, and lateral acceleration signal, first read the sampling time and sample value, and map each sampling time to the corresponding time position on the local time axis of the segment. When multiple sample values ​​of the same type exist at the same time position, retain the sample value closest to the center time of that time position; when no corresponding sample value exists at a certain time position, read the two most recent valid sample values ​​before and after that time position, perform linear interpolation based on the time distance ratio between the sampling time of the valid sample values ​​before and after and the current time position, and use the interpolation result as the padding value for that time position.

[0113] Discrete state signals are processed differently from continuous signals. Discrete quantities such as the start / stop status of the driver assistance system, the current operating mode, the takeover prompt status, and the system availability level are not suitable for linear interpolation. When a state value is missing at a certain time position, the most recently recorded state value before that time position is written forward. When there are multiple state values ​​at the same time position, the state value with the later time sequence is taken as the current state value.

[0114] The processing method for positioning trajectory signals falls between that of continuous and discrete signals. Longitude and latitude values ​​are continuous quantities; when missing, position interpolation is performed based on the time ratio between consecutive valid trajectory points. After processing the three types of signals as described above, a time-aligned signal set is obtained.

[0115] The unified frequency resampling is a process of generating a standard time sampling sequence on the local time axis of a segment at fixed time intervals and remapping various signals to the standard time sampling sequence. It is used to convert various signals that retain their original sampling frequencies after segment-level time alignment into uniformly spaced sampling data with a unified frequency.

[0116] After obtaining the time-aligned signal set, the time-aligned signal set is resampled at a unified frequency to obtain the aligned signal set. Specifically, a unified resampling frequency is first determined according to the encapsulation frequency requirements of the standardized driving behavior data packet, and a standard time sampling sequence is generated on the local time axis of the segment using the unified resampling frequency; wherein, each sampling moment in the standard time sampling sequence is arranged sequentially according to a fixed time interval. In this embodiment, the fixed time interval is 50ms. Subsequently, various signals in the time-aligned signal set are remapped to the standard time sampling sequence; for continuous signals, two adjacent valid sample values ​​that have completed time alignment are read, and the resampled value of the corresponding sampling moment is determined according to the time ratio between the corresponding moments of the two valid sample values ​​in the standard time sampling sequence; for discrete state signals, the state value corresponding to each standard sampling moment is determined as the most recent valid state value before the standard sampling moment; for positioning trajectory signals, the position value corresponding to each standard sampling moment is determined as the time ratio position value between the trajectory points before and after the standard sampling moment. For sampling positions that cannot be filled in the standard time sampling sequence, these positions are registered as missing positions, and the corresponding signal name, missing start time, missing end time, and missing duration are written into the missing field registration table. After uniform frequency resampling of the time-aligned signal set, the aligned signal set is obtained.

[0117] The semantic field encapsulation is a process of writing various signals in the alignment signal set into the corresponding semantic fields according to a preset field structure, which is used to organize the time-series sampling data into structured, addressable, and standardized data units.

[0118] After obtaining the alignment signal set, the alignment signal set is semantically encapsulated according to the segment start and end time, vehicle identification, control quantity sequence, system state sequence, driver operation sequence, positioning trajectory sequence and missing field registration results to obtain a standardized driving behavior data packet. Specifically, firstly, the header and body fields of the standardized driving behavior data package are constructed; then, the segment start and end times, vehicle identification, switching time location index, switching type, and segment number corresponding to the candidate driving behavior segments are written into the header field to obtain the segment identification field group; next, the steering wheel torque signal, brake pedal signal, vehicle speed signal, longitudinal acceleration signal, and lateral acceleration signal in the aligned signal set are written into the control quantity sequence according to the time order of the standard time sampling sequence; the start / stop status, current operating mode, takeover prompt status, and system availability level of the driving assistance system are written into the system status sequence according to the time order of the standard time sampling sequence; the steering wheel torque change record, brake pedal change record, manual intervention time marker, and status record related to driver operation are written into the driver operation sequence according to the time order of the standard time sampling sequence; the longitude value, latitude value, and trajectory location index after unified frequency resampling are written into the positioning trajectory sequence according to the time order of the standard time sampling sequence; and the missing field registration table is organized into missing field registration results and written into the corresponding fields. After completing the writing of the segment identifier field group, control quantity sequence, system status sequence, driver operation sequence, positioning trajectory sequence and missing field registration results, semantic field encapsulation is performed according to the preset field order to obtain a standardized driving behavior data packet.

[0119] The standardized driving behavior data packet is annotated at the fragment level to obtain an annotated driving behavior data packet, including: After obtaining the standardized driving behavior data packet, the standard time sampling sequence, system state sequence, driver operation sequence, vehicle response sequence, and missing field registration results in the standardized driving behavior data packet are read first. Then, each sampling time is used as an index to establish a sampling time association record. Each sampling time association record is written with the system state, driver operation, and vehicle response corresponding to that sampling time.

[0120] The determination of the dominant control source involves logically comparing the records of system status, driver operation, and vehicle response at the same sampling time to determine whether the driving control at that time is dominated by the system, manually, or in a transitional co-driving state.

[0121] Subsequently, the dominant control source is determined for the records associated with each sampling time to obtain the control ownership label; then, the data integrity and inter-source consistency of the standardized driving behavior data package are calculated to obtain the collection credibility label, whereby the data integrity includes the time base integrity and the key field integrity; finally, the control ownership label and the collection credibility label are associated and assembled to obtain the labeled driving behavior data package.

[0122] The collection confidence label is a fragment-level quality label obtained by thresholding the fragment-level collection confidence value, and is used to characterize the collection quality level of the corresponding standardized driving behavior data packet.

[0123] During the determination of the dominant control source, the system status record is obtained by first reading the start / stop status, current operating mode, takeover prompt status, and system availability level of the driver assistance system in the system status sequence at the sampling time; then, the system status record is obtained by reading the steering wheel torque change record, brake pedal change record, and manual intervention time marker in the driver operation sequence at the same sampling time; and the system status record is obtained by reading the vehicle speed value, longitudinal acceleration value, and lateral acceleration value in the vehicle response sequence at the same sampling time. Subsequently, the system status records, driver operation records, and vehicle response records at the same sampling time are compared and contrasted. When the driver assistance system is in the "on" state, the current operating mode is assisted driving mode, the takeover prompt is inactive, and neither the steering wheel torque change record nor the brake pedal change record meets the manual intervention judgment conditions, this sampling time is determined as the system-dominant control time, and a system control label is generated. When the steering wheel torque change record or brake pedal change record meets the manual intervention judgment conditions, and the direction of change of the vehicle speed value, longitudinal acceleration value, or lateral acceleration value is consistent with the driver's operation direction, this sampling time is determined as the manual-dominant control time, and a manual control label is generated. When the takeover prompt is activated, and the system status record and driver operation record alternate continuously in adjacent sampling times, this sampling time is determined as the control transition time, and a transition control label is generated. After completing the above judgment for all sampling times, the system control label, manual control label, and transition control label are written into the control source labeling sequence in chronological order to obtain the control ownership label.

[0124] The data completeness is a set of indicators used to quantitatively evaluate the effective coverage of time bases and key fields in the standardized driving behavior data package, including two sub-indicators: time base completeness and key field completeness.

[0125] The inter-source consistency is an indicator that checks and quantifies whether the logical correspondence between system states, driver operations, and vehicle responses from different data sources holds true at the same sampling time.

[0126] During the credibility calculation process, indicators such as time base integrity, key field integrity, inter-source consistency, and packet loss rate of the standardized driving behavior data packets are calculated. The calculation approach for all four indicators is to statistically analyze the percentage of valid items, but their respective inspection targets and valid judgment conditions differ.

[0127] When calculating the integrity of the time base, check each sampling moment in the standard time sampling sequence to see if there is a valid time alignment record. Count the number of sampling moments with valid time alignment records and divide the number by the total number of sampling moments in the standard time sampling sequence to obtain the integrity of the time base.

[0128] The calculation method for critical field completeness is similar to that for time base completeness. The difference is that the object of inspection is not the sampling time itself, but whether each field that should be filled in the control quantity sequence, system state sequence, driver operation sequence, and positioning trajectory sequence has a valid value. The number of fields with valid values ​​is divided by the total number of fields that should be filled to obtain the critical field completeness.

[0129] The logic for determining inter-source consistency differs from the previous two. It's not a simple matter of checking for presence or absence, but rather an examination of the logical correspondence between the system state, driver operation, and vehicle response records at the same sampling time. When the sampling time is determined to be a system-dominant control time, the vehicle response should be consistent with the system state change, and the driver operation record should not show any operation changes that meet the conditions for manual intervention. When the sampling time is determined to be a manually-dominant control time, the vehicle response should be consistent with the driver operation change. When the sampling time is determined to be a control transition time, the takeover prompt status should appear before or simultaneously with the driver operation change. After checking all sampling times, the number of sampling times that satisfy the correspondence is divided by the total number of sampling times involved in the check to obtain the inter-source consistency.

[0130] The packet loss rate can be obtained by directly reading the number of missing sampling points in the missing field registration result and dividing it by the total number of sampling points in the standard time sampling sequence.

[0131] After obtaining the time base integrity, key field integrity, inter-source consistency, and packet loss rate, these four metrics are used as input indicators for the mutual information decoupling weighted fusion algorithm. The algorithm performs integrity fusion calculations on each integrity indicator to obtain the fragment-level collection integrity value. Specifically, during the trial operation calibration phase, no fewer than 500 standardized driving behavior data packet samples are collected. For each sample, the four indicator values ​​of time base integrity, key field integrity, inter-source consistency, and packet loss rate are calculated. Each indicator is divided into 20 value intervals with a width of 0.05. The joint occurrence frequency of any two integrity indicators within each value interval combination is statistically analyzed, forming an indicator joint distribution table. The joint probability distribution and marginal probability distribution between any two integrity indicators are then determined. Based on this, the mutual information between any two integrity indicators is calculated to obtain the relevant information content between the indicators.

[0132] In this embodiment, the mutual information between any two credibility indices a and b is defined as: , in, This represents the mutual information between credibility index a and credibility index b. This represents the set of value ranges for the credibility index 'a'. This represents the set of value ranges for the credibility index b. This indicates that the credibility index 'a' takes values ​​within the range... Furthermore, the credibility index b takes values ​​within the range The joint probability at time, This represents the marginal probability when the confidence index 'a' takes a value in the interval x. This represents the marginal probability when the credibility index b takes a value in the interval y, where a and b represent any one of the following indices: time base integrity, key field integrity, inter-source consistency, and packet loss rate.

[0133] After obtaining the mutual information among the various credibility indicators, the redundancy of each credibility indicator is determined based on the total mutual information between each indicator and the other credibility indicators. Then, the credibility indicators are decoupled using the basic weight coefficients to obtain the fusion weight. The fusion weight of any credibility indicator 'a' is defined as follows: , in, This represents the fusion weight of the credibility index 'a'. q represents the basic weight coefficient of credibility index a, and r represents the credibility index number participating in the fusion calculation.

[0134] After determining the fusion weights, a weighted fusion is performed on the time base integrity, key field integrity, inter-source consistency, and packet loss rate to obtain a segment-level acquisition reliability value. The segment-level acquisition reliability value is defined as follows: , in, This represents the confidence value of fragment-level data collection. This represents the fusion weight corresponding to the completeness of the time reference. This indicates the fusion weight corresponding to the completeness of the key fields. This represents the fusion weight corresponding to the consistency between sources. This represents the fusion weight corresponding to the packet loss rate. Indicates the completeness of the time base. Indicates the completeness of key fields. This indicates consistency between sources. This indicates the packet loss rate.

[0135] In this embodiment, the basic weight coefficient for time reference integrity is 0.30, the basic weight coefficient for key field integrity is 0.25, the basic weight coefficient for inter-source consistency is 0.30, and the basic weight coefficient for packet loss rate is 0.15. The above basic weight coefficients are determined as follows: experts assign values ​​based on the degree of influence of each indicator on subsequent driving behavior analysis tasks. Time reference integrity and inter-source consistency have the greatest impact on the alignment quality of multi-source signals, so they are assigned higher weights. Packet loss rate has a relatively small impact on overall reliability when short-term packet loss has been registered in previous steps, so it is assigned a lower weight.

[0136] The threshold grading is a process of dividing continuous fragment-level acquisition confidence values ​​into discrete level labels according to a preset confidence threshold. This is used to convert quantitative confidence calculation results into hierarchical labels that can directly drive subsequent transmission and processing decisions.

[0137] After obtaining the fragment-level acquisition confidence value, the fragment-level acquisition confidence value is threshold-classified to complete the confidence calculation and obtain the acquisition confidence label. Specifically, the fragment-level acquisition confidence value is compared with a first confidence threshold and a second confidence threshold; when the fragment-level acquisition confidence value is greater than or equal to the first confidence threshold, a high confidence label is generated; when the fragment-level acquisition confidence value is less than the first confidence threshold but greater than or equal to the second confidence threshold, a medium confidence label is generated; when the fragment-level acquisition confidence value is less than the second confidence threshold, a low confidence label is generated. After threshold classification, the acquisition confidence label is obtained.

[0138] In this embodiment, the first confidence threshold is set to 0.80, and the second confidence threshold is set to 0.60. The two thresholds are determined as follows: during the trial operation calibration phase, the standardized driving behavior data packets collected are manually sampled and inspected. The distribution of the segment-level collection confidence values ​​corresponding to the samples that are manually determined to be directly usable is statistically analyzed, and the lower 10% quantile of this distribution is taken as the first confidence threshold. The distribution of the segment-level collection confidence values ​​corresponding to the samples that are manually determined to require supplementary collection or review is statistically analyzed, and the lower 10% quantile of this distribution is taken as the second confidence threshold.

[0139] After obtaining the control ownership label and the data collection confidence label, the control ownership label and the data collection confidence label are associated and assembled to obtain an annotated driving behavior data package. Specifically, firstly, the control ownership label is read time-by-time according to the standard time sampling sequence, and the control ownership label corresponding to each sampling time is written into the time sequence annotation field of the standardized driving behavior data package; then, the data collection confidence label is read and written into the segment-level annotation field of the standardized driving behavior data package; after completing the writing of the time sequence annotation field and the segment-level annotation field, all the written annotation fields are re-encapsulated with the original fields in the standardized driving behavior data package according to the preset field order to obtain the annotated driving behavior data package.

[0140] In one embodiment, the mutual information decoupling weighted fusion algorithm further includes: After a vehicle enters a tunnel on an urban expressway, the standardized driving behavior data package continuously records events such as takeover alert activation, decreased system availability level, increased steering wheel torque, and vehicle speed changes. When performing segment-level annotation on this standardized driving behavior data package, the dominant control source is first determined moment-by-moment based on system status, driver operation, and vehicle response, resulting in a control ownership label. Then, the data integrity and inter-source consistency of the standardized driving behavior data package are calculated for reliability, yielding four reliability indicators: time base integrity, key field integrity, inter-source consistency, and packet loss rate. A mutual information decoupling weighted fusion algorithm is used to complete the reliability fusion calculation, resulting in a segment-level acquisition reliability value. Subsequently, the segment-level acquisition reliability value is thresholded to obtain an acquisition reliability label. Finally, the control ownership label and the acquisition reliability label are associated and assembled to obtain an annotated driving behavior data package. This annotated driving behavior data package serves as input for subsequent acquisition and archiving processing steps.

[0141] In one embodiment, the process of collecting and archiving the labeled driving behavior data packets to obtain the target driving behavior collection results includes: After obtaining the labeled driving behavior data packet, the following are first read: the segment identifier field group, control ownership label, acquisition confidence label, control quantity sequence, system status sequence, driver operation sequence, positioning trajectory sequence, and missing field registration results. A record of segments to be processed is then created according to the segment number, switching type, and segment start and end time. Each record of segments to be processed corresponds to a labeled driving behavior data packet, and the corresponding segment number, switching type, segment start and end time, vehicle identifier, control ownership label, and acquisition confidence label are written into the record. Subsequently, the labeled driving behavior data packets are categorized according to the acquisition confidence label to obtain categorized driving behavior data packets. After categorization, the categorized driving behavior data packets are sorted according to transmission priority to obtain sorted driving behavior data packets. These sorted driving behavior data packets are then transmitted and processed to obtain processed driving behavior data packets. Finally, the processed driving behavior data packets are archived and registered to obtain the target driving behavior acquisition result.

[0142] During the grading process, the collection confidence tags in each record to be processed are read one by one, and these tags are matched with the grading conditions in the grading rule table. The grading rule table at least registers the grading categories corresponding to high-confidence tags, medium-confidence tags, and low-confidence tags. When the collection confidence tag in a record to be processed is a high-confidence tag, the corresponding labeled driving behavior data packet is written to the high-confidence fragment set; when the collection confidence tag is a medium-confidence tag, the corresponding labeled driving behavior data packet is written to the medium-confidence fragment set; and when the collection confidence tag is a low-confidence tag, the corresponding labeled driving behavior data packet is written to the low-confidence fragment set. After matching all records to be processed, the high-confidence fragment set, medium-confidence fragment set, and low-confidence fragment set are written as grading results into the grading field, resulting in grading driving behavior data packets.

[0143] The transmission priority is determined by comprehensively considering the classification, control ownership label, and time sequence of each driving behavior data packet, thus establishing the order of transmission.

[0144] After obtaining the segmented driving behavior data packets, the data packets are sorted according to transmission priority. Specifically, the segmentation field, switching type, control ownership label, and segment start and end time of each segmented driving behavior data packet are first read; then, the data packets are sorted according to a preset priority rule. The preset priority rule includes at least the following: high-confidence segments have higher priority than medium-confidence segments, and medium-confidence segments have higher priority than low-confidence segments; under the same segmentation category, segments corresponding to control ownership transition have higher priority than segments corresponding to system control and manual control; under the same segmentation category and the same control ownership label, driving behavior data packets with earlier segment end times have higher priority than those with later segment end times. After completing the above comparisons, transmission sequence numbers are generated from high to low priority, and these transmission sequence numbers are written into the transmission sorting field corresponding to each segmented driving behavior data packet to obtain the sorted driving behavior data packets.

[0145] The transmission processing is a process of determining, based on the classification of each sorted driving behavior data packet, whether to perform an immediate upload, a retransmission request, or local retention transmission action.

[0146] After obtaining the ordered driving behavior data packets, the ordered driving behavior data packets are processed for transmission. Specifically, each ordered driving behavior data packet is read sequentially according to its transmission sequence number, then the classification field corresponding to each ordered driving behavior data packet is read, and the corresponding transmission processing method is determined according to the processing rule table; wherein, the processing rule table registers at least three types of processing methods: immediate upload, retransmission request, and local retention. When the classification field corresponding to a certain ordered driving behavior data packet is a high-confidence classification, the ordered driving behavior data packet is written to the immediate upload queue and an immediate upload processing mark is generated; when the classification field corresponding to a certain ordered driving behavior data packet is a medium-confidence classification, the ordered driving behavior data packet is written to the retransmission processing queue and a retransmission request processing mark is generated; when the classification field corresponding to a certain ordered driving behavior data packet is a low-confidence classification, the ordered driving behavior data packet is written to the local retention queue and a local retention processing mark is generated. After matching the handling methods for all sorted driving behavior data packets, the corresponding transmission execution processes are invoked: For sorted driving behavior data packets written to the immediate upload queue, upload message encapsulation is performed, and the encapsulated upload message is sent to the remote receiving end; for sorted driving behavior data packets written to the retransmission handling queue, a retransmission request record is generated, and the retransmission request record is sent to the acquisition end's retransmission task queue; for sorted driving behavior data packets written to the local retention queue, a local retention record is generated, and the corresponding data packet is written to the local retention storage area. After completing all transmission execution processes, the sorted driving behavior data packets with the immediate upload handling flag, retransmission request handling flag, or local retention handling flag are identified as the handled driving behavior data packets.

[0147] The archiving registration is the process of aggregating the fragment identification information, annotation results, and handling methods of each data packet related to driving behavior into an archive record and writing it into the archive storage.

[0148] After obtaining the driving behavior data packets, the data packets are archived and registered. The segment number, vehicle identifier, switching type, segment start and end time, control ownership tag, acquisition reliability tag, classification category, transmission sequence number, and handling method corresponding to each driving behavior data packet are aggregated to generate an archive record. This archive record is written into an archive index table, and the corresponding driving behavior data packets are written to the archive storage area. After all driving behavior data packets have been archived, the target driving behavior acquisition result is obtained.

[0149] The target driving behavior collection results include at least: a set of archived driving behavior data packets, a set of corresponding archived index records, and a record of the processing method corresponding to each driving behavior data packet; wherein, the processing method record is used to indicate which type of processing method has been executed for the corresponding labeled driving behavior data packet: immediate upload, re-upload request, or local retention.

[0150] In a specific scenario, after a vehicle enters a tunnel on an urban expressway, it generates multiple labeled driving behavior data packets. One portion of these packets corresponds to the control transition process and has a high-confidence tag; the other portion corresponds to the stable manual driving process after the switch and has a medium- or low-confidence tag. When performing collection and archiving processing on this batch of labeled driving behavior data packets, the data packets are first categorized according to their collection confidence tags, resulting in categorized driving behavior data packets. Then, the transmission priority is sorted according to the rule of high-confidence categorization first, medium-confidence categorization second, and low-confidence categorization third, resulting in sorted driving behavior data packets. Subsequently, the data packets corresponding to the high-confidence categorization are immediately uploaded, the data packets corresponding to the medium-confidence categorization are retransmitted, and the data packets corresponding to the low-confidence categorization are locally retained, resulting in processed driving behavior data packets. Finally, the archive number, categorization type, processing method, and storage location of each processed driving behavior data packet are written into the archive index table, and the corresponding data packets are written to the archive storage area, obtaining the target driving behavior collection result.

[0151] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An Internet of Things-based automobile driving behavior data collection method, characterized in that, include: Acquire multi-source driving behavior data and align the multi-source driving behavior data with a time reference to obtain an aligned driving data stream; The control switchover is identified by performing control switchover identification on the aligned driving data stream to obtain a switchover marker data stream; Differential resampling is performed on the switching marker data stream to obtain candidate driving behavior segments; The candidate driving behavior segments are standardized and encapsulated to obtain standardized driving behavior data packets; The standardized driving behavior data packet is annotated at the fragment level to obtain an annotated driving behavior data packet; The labeled driving behavior data packets are collected and archived to obtain the target driving behavior collection results.

2. The automobile driving behavior data collection method according to claim 1, characterized in that, Acquire multi-source driving behavior data and align the multi-source driving behavior data with a time reference to obtain an aligned driving data stream, including: Vehicle operating status data, driver behavior data, and driver assistance system status data are acquired separately to obtain multi-source driving behavior data; The local timestamps of each data source in the multi-source driving behavior data are mapped to a unified clock according to the same reference time axis to obtain unified time axis data; Short-term packet loss detection is performed on the unified timeline data to obtain packet loss marker data, and missing segments are registered in the packet loss marker data to obtain an aligned driving data stream.

3. The automobile driving behavior data collection method according to claim 1, wherein The aligned driving data stream is subjected to control handover identification to obtain a handover marker data stream, including: The sliding window state transition mutation detection algorithm is used to detect state jumps in the driving assistance system state signals in the aligned driving data stream to obtain system switching events; Abrupt changes are detected in the steering wheel torque signal and brake pedal signal in the aligned driving data stream to obtain manual intervention events. The system switching event and the manual intervention event are fused and determined to obtain a switching marker data stream.

4. The automobile driving behavior data collection method according to claim 3, characterized in that, The system switching event and the manual intervention event are fused and determined, including: The torque threshold verification is performed on the steering wheel torque change in the aforementioned manual intervention event to obtain the torque verification result; The change in the brake pedal during the aforementioned manual intervention event is subjected to a pedal threshold verification to obtain the pedal verification result. The duration of the manual intervention event is validated using a duration threshold to obtain the duration validation result. The torque verification result, the pedal verification result, and the duration verification result are combined for a pass determination to obtain a valid intervention event; The effective intervention events and the system switching events are merged and arranged in chronological order, and duplicate markers are removed to complete the fusion determination.

5. The method of claim 1, wherein Differential resampling is performed on the switching marker data stream to obtain candidate driving behavior segments, including: Taking each switching marker in the switching marker data stream as the center, extract the preceding window data of a preset duration and the following window data of a preset duration to obtain the window data segment; The window data segment is resampled differentially using an information entropy adaptive hierarchical resampling algorithm to obtain candidate driving behavior segments.

6. The automobile driving behavior data collection method according to claim 5, wherein Extracting data from the preceding window for a preset duration and extracting data from the following window for a preset duration, including: For each switching marker, backtrack for a preset time period, extract high-frequency data for the corresponding time period from the circular buffer, and obtain the front window data; High-frequency data of each switching marker is collected in real time for a preset time backward to obtain the data of the rear window; The data in the preceding window and the data in the following window are concatenated in chronological order to extract the window data segment.

7. The method of claim 1, wherein The candidate driving behavior segments are standardized and encapsulated to obtain a standardized driving behavior data packet, including: The various signals in the candidate driving behavior segments are time-aligned at the segment level to obtain a time-aligned signal set, and the time-aligned signal set is resampled at a uniform frequency to obtain an aligned signal set. The aligned signal set is semantically encapsulated according to the segment start and end time, vehicle identification, control quantity sequence, system state sequence, driver operation sequence, positioning trajectory sequence, and missing field registration results to obtain a standardized driving behavior data packet.

8. The method for collecting vehicle driving behavior data according to claim 1, characterized in that, The standardized driving behavior data packet is annotated at the fragment level to obtain an annotated driving behavior data packet, including: The dominant control source is determined for the system state, driver operation, and vehicle response at each sampling time in the standardized driving behavior data packet, and the control ownership label is obtained. The data integrity and inter-source consistency of the standardized driving behavior data packet are evaluated to obtain a collection credibility label. The control ownership label and the collection credibility label are associated and assembled to obtain a labeled driving behavior data packet.

9. The method for collecting vehicle driving behavior data according to claim 8, characterized in that, The reliability calculation of the data integrity and inter-source consistency of the standardized driving behavior data packet includes: The time base integrity, key field integrity, inter-source consistency, and packet loss rate of the standardized driving behavior data packets are calculated to obtain various reliability indicators. The credibility fusion calculation is performed on each of the credibility indicators using a mutual information decoupling weighted fusion algorithm to obtain the segment-level acquisition credibility value; The credibility values ​​of the segment-level acquisitions are categorized by threshold to complete the credibility calculation.

10. The method for collecting vehicle driving behavior data according to claim 1, characterized in that, The labeled driving behavior data packets are collected and archived to obtain the target driving behavior collection results, including: The labeled driving behavior data packets are classified by confidence level to obtain classified driving behavior data packets, and the classified driving behavior data packets are sorted according to transmission priority to obtain sorted driving behavior data packets. The sorted driving behavior data packets are transmitted and processed to obtain processed driving behavior data packets, and the processed driving behavior data packets are archived and registered to obtain the target driving behavior collection results.

Images