Port mobile machine driving behavior ai monitoring and early warning and active intervention method based on multi-source information fusion

Abstract

The present application relates to the technical field of port engineering machinery safety control, specifically to a port mobile machinery driving behavior AI monitoring and early warning and active intervention method based on multi-source information fusion, comprising the following steps: data acquisition and preprocessing; multi-dimensional feature extraction and fusion; driving behavior risk reasoning grading; graded early warning and active intervention control. The present application generates a dynamic expected interval by operation micro-phase probability weighting, combines double-dimensional deviation and transitive correlation quantity recursion operation response inconsistency state, can distinguish normal operation fluctuation and continuous dangerous behavior, reduces false judgment and late judgment caused by fixed threshold, improves the accuracy and continuity of risk identification; at the same time, the vehicle correction ability boundary is corrected by the load stable occupation factor coupled with multiple factors, the intervention mode and intensity are constrained by the risk recovery margin of the asymmetric recursion, the intervention adaptation risk is reduced, and the safety of the graded early warning and active intervention in the port operation scene is improved.

Description

Technical Field

[0001] This invention relates to the field of port engineering machinery safety management and control technology, specifically to a method for AI monitoring, early warning and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion. Background Technology

[0002] As port operations move towards digitalization, mobile machinery such as forklifts, reach stackers, and tractor units are widely equipped with onboard controllers, positioning terminals, video sensing devices, and wireless communication modules. Driver behavior monitoring, a key component of intelligent port safety management, aims to identify real-time changes in personnel operating status, vehicle running status, and the working environment, providing a basis for proactive early warning and safety control.

[0003] Currently, port mobile machinery driving safety monitoring mainly adopts two types of solutions: one is a driver status monitoring solution based on vehicle-mounted video, which uses a cab camera and image processing unit to identify faces, eyes, head postures, and body movements to determine behaviors such as fatigue, distraction, smoking, and making or receiving phone calls. It has the characteristics of relatively independent deployment, intuitive behavioral characteristics, and convenient modification. The other is a vehicle operation monitoring solution based on vehicle networking terminals, which uses vehicle controllers, satellite positioning terminals, inertial sensors, vehicle bus interfaces, and back-end platforms to acquire information such as speed, acceleration, steering, braking, position, and electronic fences to identify behaviors such as speeding, sudden acceleration or deceleration, sharp turns, and area violations. It has the advantages of continuous data collection, wide management scope, and convenient centralized traceability.

[0004] The aforementioned solutions generally rely on single-type perceived information or preset threshold rules as the main basis for judgment. However, under conditions of frequent switching between port mobile machinery operation stages, continuous changes in load status, and mutual influence between personnel operation and vehicle response, two common technical problems remain: First, existing solutions lack dynamic expectation modeling of expected operations and vehicle responses at the current micro-stage of operation (such as approaching the stack, precise forklift picking, load transfer, etc.). This makes it difficult to effectively distinguish between short-term fluctuations in driver operation, normal operational adjustments that conform to the working conditions, and continuous dangerous operational behaviors. Consequently, there is a lack of continuous correlation analysis between abnormal driver operation and actual vehicle operational risks, easily leading to delayed risk identification or repeated alarms. Second, existing risk assessment and intervention strategies generally ignore the impact of multiple factors such as load height, lateral load offset, and vehicle steering motion on vehicle stability. They cannot accurately quantify the vehicle's actual remaining longitudinal deceleration, lateral steering, and mechanism correction capabilities available for safe correction under the current working conditions. Furthermore, it is difficult to reasonably select and constrain the methods and intensity of proactive intervention based on this, resulting in inconsistencies between warning results and actual operational risks, and insufficient scenario adaptability of intervention measures. The aforementioned issues indicate that there is still room for improvement in the refinement and proactive control of port mobile machinery driving safety monitoring. Therefore, it is necessary to enhance the capabilities of multi-source information collaborative analysis, dynamic risk assessment, and graded intervention to meet the safety control needs under complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based monitoring, early warning, and proactive intervention method for the driving behavior of port mobile machinery based on multi-source information fusion. This method addresses the lack of continuous dynamic correlation analysis between abnormal driver operation and operational risks in existing technologies, as well as the insufficient assessment of vehicle stability and corrective capabilities under load-coupled conditions by risk assessment and intervention strategies. Ultimately, this method enables refined and proactive safety monitoring and control under complex operating conditions.

[0006] To address the aforementioned technical problems, this invention provides an AI-based monitoring, early warning, and proactive intervention method for the driving behavior of port mobile machinery based on multi-source information fusion, comprising the following steps: S1. Data Acquisition and Preprocessing: Real-time acquisition of onboard operation data, driver status and behavior data, port operation environment and spatial positioning data, load and operation mechanism status data of port mobile machinery under different operating conditions, and time-series alignment and standardization preprocessing of multi-source data to obtain standardized multi-source monitoring datasets. S2. Multidimensional feature extraction and fusion: Based on the standardized multi-source monitoring dataset, multidimensional features reflecting the evolution of driving behavior risks are extracted through spatiotemporal dimension alignment, multidimensional feature mapping, and feature-level fusion processing to obtain a driving behavior risk fusion feature set; S3. Driving behavior risk reasoning and classification: Based on the driving behavior risk fusion feature set, the driving behavior risk is judged by a joint reasoning mechanism of work response inconsistency state recursion, driving risk recovery margin recursion, and work load stability occupancy factor correction, and the driving behavior risk classification result includes risk level, intervention urgency and allowable intervention method constraints. S4. Tiered early warning and active intervention control: Based on the risk grading results of the driving behavior, the tiered early warning strategy and the intervention instructions of the on-board equipment are dynamically matched and issued for execution, in order to reduce the driving safety risks during the operation of mobile machinery in the port.

[0007] As a further improvement to this technical solution, in step S1, time-series alignment and standardization preprocessing are performed on the multi-source data to obtain a standardized multi-source monitoring dataset, including the following steps: S1.1 Collect on-board operation data, driver status and behavior data, port operation environment and spatial positioning data, and load and operation mechanism status data of port mobile machinery under different operating conditions through corresponding sensing terminals according to data categories, and obtain raw multi-source monitoring data. S1.2. Use a unified clock reference to perform timestamp alignment and sampling granularity unification processing on the original multi-source monitoring data to generate time-synchronized multi-source monitoring data; S1.3. Perform outlier identification and jump data correction on the time-synchronized multi-source monitoring data, remove invalid data segments, and obtain cleaned multi-source monitoring data. S1.4 Perform dimensional unification and standardization mapping on the cleaned multi-source monitoring data to output a standardized multi-source monitoring dataset.

[0008] As a further improvement to this technical solution, in step S2, multi-dimensional features reflecting the evolution law of driving behavior risk are extracted based on the standardized multi-source monitoring dataset to obtain a driving behavior risk fusion feature set, including the following steps: S2.1 Using the standardized multi-source monitoring dataset as the processing object, unify the spatial coordinate system, complete the spatiotemporal benchmark calibration of the multi-source data, and generate spatiotemporal dimension aligned multi-source basic feature data; S2.2 Map the spatiotemporal aligned multi-source basic feature data to the operation requirement feature subset, driving response feature subset, vehicle execution feature subset, and load stability feature subset according to the representation attributes, and complete the dimensional feature classification; S2.3 Extract time-series fluctuation features and cross-dimensional correlation and interaction features for each feature subset to generate derived features that reflect the evolution of driving behavior risks; S2.4. Perform feature-level fusion and splicing on the subset of operation requirements features, subset of driving response features, subset of vehicle execution features, subset of load stability features, and derived features according to the time sequence dimension to output the fusion feature set of driving behavior risks.

[0009] As a further improvement to this technical solution, the driving behavior risk reasoning and classification in S3 includes the following steps: S3.1 Match the current job process with the corresponding job micro-stage and generate a scenario-based driving response expectation range and vehicle execution expectation range; S3.2 Calculate the degree of mismatch in the driving response dimension and the vehicle execution dimension respectively, and assess the degree of transmission of operational abnormalities to operational risks; S3.3. Generate inconsistent operation response states through time-series recursion to quantify the formation and persistence of risks; S3.4 Calculate the working load stability occupancy factor based on load and motion conditions, and correct the vehicle's safety correction capability boundary; S3.5. Generate driving risk recovery margin and combine it with the inconsistent state of operation response, driving risk recovery margin and load constraints to complete the joint classification and output the driving behavior risk classification result.

[0010] As a further improvement to this technical solution, step S3.1, matching the current job process's corresponding job micro-stage and generating a scenario-based expected range, includes the following steps: S3.11. Extract vehicle spatial location, driving direction, operating mechanism status, load status and time sequence change features from the driving behavior risk fusion feature set, and obtain the belonging probability of each operation micro-stage at the current time through time sequence feature matching and probabilistic inference, and generate operation micro-stage probability vector. S3.12. Using the probability of each stage in the operation micro-stage probability vector as the weight, the driving response expectation template and vehicle execution expectation template corresponding to each pre-configured operation micro-stage are weighted and fused to generate a driving response expectation interval and a vehicle execution expectation interval that match the current operation process.

[0011] As a further improvement to this technical solution, step S3.2, calculating the degree of mismatch in both dimensions and assessing the degree of risk transmission, includes the following steps: S3.21. The driver's state and behavior features in the driving behavior risk fusion feature set are compared with the expected range of driving response feature by feature. For features that exceed the expected range, the standardized deviation is calculated. The deviation is then weighted and summed with the feature weights corresponding to the current operation stage to obtain the comprehensive driving response deviation. S3.22. The on-board operation and work mechanism features in the driving behavior risk fusion feature set are compared with the vehicle execution expectation range feature by feature. For features that exceed the expectation range, the standardized deviation is calculated. The deviation is then weighted and summed with the feature weights corresponding to the current operation stage to obtain the comprehensive vehicle execution deviation. S3.23. Within the preset time window, calculate the temporal transmission matching degree between driving response deviation and vehicle execution deviation through the deviation correlation matrix corresponding to the current operation stage, and generate the transmission correlation quantity of driving response deviation to vehicle execution result.

[0012] As a further improvement to this technical solution, in S3.3, generating an inconsistent job response state specifically includes: The inconsistency state of the operation response at the previous moment is obtained. Combined with the risk memory coefficient corresponding to the current operation stage, the comprehensive driving response deviation, the comprehensive vehicle execution deviation, and the transmission correlation, the inconsistency state of the operation response at the current moment is calculated using a time-series recursive rule with memory decay. When the vehicle is in reversing positioning, load transfer, densely populated areas, or near the operation boundary, the value of the risk memory coefficient is increased accordingly to prolong the duration of the risk state's continuous memory, thereby quantifying the formation and persistence of the risk of mismatch between driving response and operation requirements.

[0013] As a further improvement to this technical solution, step S3.4, calculating the operational load stability occupancy factor and correcting the vehicle safety correction capability boundary, includes the following steps: S3.41 Extract four types of normalized features from the driving behavior risk fusion feature set: load amount, load height, load lateral offset, and vehicle steering motion intensity. Obtain the working load stability occupancy factor through multi-factor coupling calculation to quantify the degree of occupancy of the vehicle's stability margin by the current load and motion conditions. S3.42. Using the aforementioned work load stability occupancy factor as a correction coefficient, the basic correctable capabilities of the three correction channels—longitudinal deceleration, lateral steering, and work mechanism retraction—are attenuated and corrected to obtain the effective correctable capabilities of the vehicle corresponding to each channel under the current load condition.

[0014] As a further improvement to this technical solution, in S3.5, generating the driving risk recovery margin and completing the joint classification output includes the following steps: S3.51. Based on the comprehensive vehicle execution deviation, calculate the necessary correction amounts for the three channels of longitudinal deceleration, lateral steering, and working mechanism reversal required to restore to a safe state, and calculate the normalized recovery margin for each correction channel in combination with the vehicle's effective correctability. S3.52. Extract the response hysteresis, operation amplitude, and deviation descent rate features of the driver's corrective operation from the driving behavior risk fusion feature set. Use the minimum normalized recovery margin of each correction channel as the recursive base value. Adjust the decay step size of risk deterioration with response hysteresis and adjust the recovery coefficient of risk recovery with operation amplitude and deviation descent rate. Use the asymmetric recursive rule of rapid decline when risk deteriorates and slow rise when risk recovers to calculate the driving risk recovery margin at the current moment. S3.53. The abnormal threshold of the inconsistent operation response is used as the basis for determining the establishment of risk, and the driving risk recovery margin is lower than the safety threshold as the basis for determining the intervention trigger. Combined with the operation load stability occupancy factor to limit the range of permissible intervention methods, the risk level is jointly classified and the intervention urgency is calculated. The result of driving behavior risk classification, which includes risk level, intervention urgency and permissible intervention method constraints, is output.

[0015] As a further improvement to this technical solution, in step S4, the dynamic matching and execution of the graded early warning strategy and the intervention command from the vehicle-mounted equipment includes the following steps: S4.1 Analyze the driving behavior risk classification results and extract the corresponding risk level, intervention urgency and allowable intervention method constraints; S4.2. Based on the risk level and the urgency of intervention, match the corresponding level of early warning strategy and determine the early warning triggering method and output sequence; S4.3. Based on the constraints of permissible intervention methods, select on-board equipment intervention commands that are suitable for the current load conditions, and clarify the action channel and execution intensity of the intervention. S4.4. The graded early warning strategy and the intervention instructions of the vehicle equipment are sent to the corresponding execution terminal for execution, so as to realize the graded control of driving behavior risks.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention identifies the current micro-stage of port mobile machinery operation, generates a probability vector for that micro-stage, and then weights and fuses the expected templates of driver response and vehicle execution to form a dynamic expectation interval that matches the actual operation process. Furthermore, by combining the deviation of driver response, the deviation of vehicle execution, and the temporal correlation between the two, an inconsistent state of operation response is recursively generated. Thus, continuous and correlated analysis of driver operation and vehicle execution results can be performed at different operation stages, accurately distinguishing between short-term operational fluctuations, normal adjustments conforming to working conditions, and continuous dangerous operations. This reduces misjudgments, delayed judgments, and repeated alarms caused by fixed thresholds or single behavior identification, improving the accuracy, continuity, and timeliness of driver behavior risk identification. 2. This invention generates a stable occupancy factor for the operating load through multi-factor coupled calculations of load amount, load height, lateral load offset, and vehicle steering motion intensity. Based on this factor, the basic correctable capabilities of the three correction channels—longitudinal deceleration, lateral steering, and operating mechanism reversal—are attenuated and corrected. Furthermore, by combining the necessary correction amount, driver response lag, operational amplitude, and deviation descent rate, an asymmetric recursive rule is used to calculate the driving risk recovery margin, which decreases rapidly when risk worsens and increases slowly when risk recovers. This allows for a more objective reflection of the vehicle's actual remaining correction capability under different loads and motion conditions. It also limits the active intervention methods, action channels, and execution intensity, reducing the risk of incompatibility between intervention measures and vehicle stability, and improving the safety and applicability of graded early warning and active intervention in complex port operation scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall method steps of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment provides an AI-based monitoring, early warning, and proactive intervention method for the driving behavior of port mobile machinery based on multi-source information fusion. It is applicable to various scenarios involving the management of driving safety risks for mobile machinery such as forklifts, reach stackers, and tractors within ports, under different working conditions including loading / unloading, transfer, and yard operations. This embodiment can be implemented using an onboard sensing and computing platform. This platform integrates a driver status monitoring terminal, an onboard operation data acquisition module, positioning and operating environment sensing equipment, load and operating mechanism status sensors, and an onboard edge computing unit. Each sensing terminal achieves multi-source data aggregation and clock synchronization through an in-vehicle communication network.

[0020] In this embodiment, driven by the dynamic expectations of the micro-stage of the operation, a temporal recursive correlation is constructed between the deviation of driving response and vehicle execution in two dimensions. Combined with the safety boundary correction for stable load occupancy, this achieves dynamic quantitative assessment and graded intervention of driving behavior risks. After preprocessing and feature fusion, multi-source data sequentially enters the risk reasoning and grading stages, forming a closed-loop collaboration of "perception—feature—reasoning—control." The specific steps are as follows: S1. Data Acquisition and Preprocessing: Real-time acquisition of onboard operation data, driver status and behavior data, port operation environment and spatial positioning data, and load and operating mechanism status data of port mobile machinery under different operating conditions. Time-series alignment and standardization preprocessing are then performed on the multi-source data to obtain a standardized multi-source monitoring dataset; including the following steps: S1.1 Collect on-board operation data, driver status and behavior data, port operation environment and spatial positioning data, and load and operation mechanism status data of port mobile machinery under different operating conditions through corresponding sensing terminals according to data categories, and obtain raw multi-source monitoring data. Specifically, the original multi-source monitoring data involved in this embodiment comes from various types of sensing terminals configured on port mobile machinery. Each terminal achieves data aggregation through the in-vehicle communication network and the vehicle-mounted edge computing unit.

[0021] Furthermore, vehicle operating data is collected through the vehicle bus interface and inertial sensors, including vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, braking pressure, throttle opening, gear status, and steering angle. The sampling granularity is based on the vehicle bus refresh cycle, typically ranging from 10 Hz to 50 Hz.

[0022] Driver status and behavior data are collected by a camera in the cab and an image processing unit, including facial orientation, gaze direction, head posture angle, eyelid opening, upper limb operation type and operation frequency. After being extracted into structured behavioral features by the image processing unit, the data is output at a frame rate of no less than 5 Hz.

[0023] Port operation environment and spatial positioning data are acquired in collaboration with reference base stations in the operation area through satellite positioning terminals, lidar or visual positioning modules. This includes the vehicle's global coordinate position, heading angle, driving speed vector, minimum distance from the operation boundary and obstacles, and identification of densely populated areas. The positioning data update rate is usually 10 Hz to 20 Hz.

[0024] Load and operating mechanism status data are collected through load sensors, boom angle sensors, gantry height sensors, and hydraulic system pressure sensors. This includes load weight, load height, load lateral offset, lifting / tilting / lateral movement stroke of the operating mechanism, and corresponding action rate. The sampling granularity is consistent with the vehicle-mounted operating data or aligned after synchronization.

[0025] Finally, after each sensing terminal is powered on, it automatically completes time synchronization and communication link verification. During the acquisition process, it continuously outputs raw sensing data streams with their own local timestamps, which together constitute the original multi-source monitoring dataset.

[0026] S1.2. The original multi-source monitoring data is timestamped and sampled using a unified clock reference to generate time-synchronized multi-source monitoring data. Specifically, different sensing terminals have different sampling frequencies and timestamp sources. A unified clock reference is built into the vehicle-mounted edge computing unit. This unified clock reference uses the timing pulse from the satellite positioning terminal as a reference and combines it with the in-vehicle network clock synchronization protocol to uniformly align the local timestamps of various sensing terminals. For any sampling moment, using the timestamp generated by the unified clock reference as an index, the sampled value closest to that timestamp within a preset time tolerance window in each data category is taken as the valid data point for that moment.

[0027] Furthermore, to address the issue of inconsistent sampling granularity, data in the low-speed sampling category is filled with nearest neighbor hold between adjacent valid sampling points, while data in the high-frequency sampling category is downsampled or mean-aggregated at a uniform sampling granularity. This unifies the sampling granularity of all data categories to the same step size, with a typical uniform sampling step size set to 100ms.

[0028] Through the above processing, each data category is aligned point-by-point in the time dimension, forming multi-source monitoring data with time-series synchronization indexed by a unified time label.

[0029] S1.3. Perform outlier identification and jump data correction on the time-synchronized multi-source monitoring data, remove invalid data segments, and obtain cleaned multi-source monitoring data. Specifically, for continuous numerical features, a sliding window-based statistical test method is used to identify outliers. When the value of a sampling point deviates from the local median within its window by more than a preset multiple of the local absolute deviation, it is determined to be an outlier. Isolated outliers are corrected by linear interpolation of the preceding and following valid sampling points. For consecutive outlier segments or outlier segments whose duration exceeds a preset threshold, all data channels for the corresponding time period are directly removed.

[0030] Furthermore, for categorical features, when values ​​exceeding the preset valid category set or state transitions that do not meet physical continuity constraints occur, they are also treated as anomalous data and removed or marked. In addition, for persistent data loss caused by sensor communication interruptions, power supply anomalies, or sensing terminal malfunctions, corresponding invalid data segment identifiers are inserted into the dataset for complete removal in subsequent steps.

[0031] After the above processing, cleaned multi-source monitoring data are obtained.

[0032] S1.4. Perform dimensional unification and standardization mapping on the cleaned multi-source monitoring data to output a standardized multi-source monitoring dataset. Specifically, since there are significant differences in the dimensions and numerical ranges of various features in the cleaned multi-source monitoring data, this step is used to perform unified standardization processing.

[0033] Specifically, continuous physical quantities such as velocity, acceleration, angle, and distance are first uniformly converted to the International System of Units (SI) or a preset standard unit system based on the factory calibration parameters or pre-calibrated conversion relationships of the corresponding sensing terminals, thereby eliminating dimensional differences between different terminal output protocols.

[0034] Furthermore, each continuous feature is standardized using Z-score with preset normalization parameters, mapping it to dimensionless standardized values. The reference mean and reference standard deviation are derived from the statistical values ​​of each feature under normal operating conditions, obtained through offline statistical analysis or historical operational data calibration. For categorical features and status indicators, a preset coding mapping table is used to convert them into standardized coded values.

[0035] After the above processing, each data channel forms a standardized multi-source monitoring dataset with consistent numerical values ​​and dimensions under the same time index. This dataset serves as the input for step S2.

[0036] S2. Multidimensional Feature Extraction and Fusion: Based on a standardized multi-source monitoring dataset, multidimensional features reflecting the evolution of driving behavior risks are extracted through spatiotemporal alignment, multidimensional feature mapping, and feature-level fusion processing, resulting in a driving behavior risk fusion feature set; including the following steps: S2.1. Using standardized multi-source monitoring datasets as the processing object, unify the spatial coordinate system, complete the spatiotemporal benchmark calibration of multi-source data, and generate spatiotemporal dimension-aligned multi-source basic feature data; Specifically, the vehicle's global coordinate position and heading angle are derived from the satellite positioning terminal, with their original coordinate values ​​defined in a geocentric coordinate system or a local northeast-sky coordinate system of the port. The distances to the work boundary and obstacles are derived from the vehicle-mounted lidar or visual positioning module, with their original measurements using the sensor's installation position as the origin and the sensor's detection direction as the reference direction. This step uniformly transforms the above spatial data into a vehicle body coordinate system with the vehicle's geometric center as the origin, the current longitudinal axis as the positive X-axis, and the current transverse axis as the positive Y-axis. The rotation and translation parameters required for coordinate transformation are jointly determined by the calibration values ​​of each sensor's installation position on the vehicle body and the vehicle's current heading angle.

[0037] After the spatial coordinates are unified, the time labels of each feature channel at the same sampling time are verified a second time, based on the unified sampling step size determined in S1.2. If a feature channel has a time deviation at that time that exceeds the preset tolerance range, it is corrected according to the sampling filling rules in S1.2. Abnormal jumps are handled according to the interpolation correction rules in S1.3. After the spatial coordinate system is unified and the time labels are verified, each feature is aligned in both spatial and temporal dimensions, generating multi-source basic feature data with spatiotemporal alignment.

[0038] S2.2 The multi-source basic feature data aligned in spatiotemporal dimensions are mapped to the operation requirement feature subset, driving response feature subset, vehicle execution feature subset, and load stability feature subset according to their representational attributes, completing the dimensional feature classification; among which: The task requirement feature subset includes features that describe the objective requirements of the current task on vehicle movement and driver operation, including the vehicle's global coordinate position, heading angle, driving speed vector, shortest distance to the task boundary, shortest distance to obstacles, and densely populated area identification. The above spatial position and motion state features also serve as the input basis for subsequent task micro-stage identification. The driving response feature subset contains features that describe the driver’s operational behavior after perceiving the task requirements, including facial orientation angle, gaze direction vector, head posture angle, eyelid opening percentage, upper limb operation type encoding, operation frequency, throttle opening change rate, and gear shift sequence. The vehicle execution feature subset includes features that describe the motion response of the vehicle body and working mechanism to the driver's operation, including vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, braking pressure, steering angle, lifting stroke, tilting stroke, lateral movement stroke of the working mechanism, and the action rate corresponding to the above strokes; The load stability feature subset contains features that describe the current load state and its impact on vehicle stability, including load weight, load height, load lateral offset, and the rates of change of load weight, load height, and load lateral offset.

[0039] Furthermore, each of the four feature subsets is organized with an independent data structure. Within each subset, all feature dimensions share the same time label sequence, and the subsets establish a correspondence between sampling points through time labels. The specific feature dimensions included in each subset can be adjusted according to the type and number of sensing terminals actually configured on the port's mobile machinery, but the affiliation boundaries between the four subsets remain fixed, and the same basic feature belongs to only one subset.

[0040] S2.3 Extract time-series fluctuation features and cross-dimensional correlation and interaction features for each feature subset to generate derived features that reflect the evolution of driving behavior risks; Specifically, for the temporal fluctuation characteristics within each feature subset, the following statistics are calculated for each continuous feature within a preset sliding time window: the mean of the first difference within the window, reflecting the average trend of the feature within the window; the standard deviation of the first difference within the window, reflecting the amplitude of the feature's fluctuation within the window; the difference between the maximum and minimum values ​​within the window, reflecting the extreme value deviation range of the feature within the window; and the number of times the feature value crosses a preset reference value within the window, reflecting the oscillation frequency of the feature within the window. These statistics are updated point-by-point as the window slides, constituting the temporal fluctuation characteristics of each feature subset.

[0041] Cross-dimensional correlation and interaction features are used to quantify the coupling influence between different feature subsets. Between the driving response feature subset and the vehicle execution feature subset, the time-delay cross-correlation function of the throttle opening rate of change sequence and the longitudinal acceleration sequence within a sliding window is calculated. The peak value of the cross-correlation is taken as the correlation strength of the driving acceleration response transmitted to the vehicle's longitudinal execution, and the corresponding time delay is taken as the response delay of the operation transmission. Simultaneously, the time-delay cross-correlation function of the steering angle rate of change sequence and the yaw rate of change sequence within a sliding window is calculated. The peak value of the cross-correlation and the corresponding time delay are taken as the correlation strength and response delay of the steering operation transmitted to the lateral execution. Between the vehicle execution feature subset and the load stability feature subset, the Pearson correlation coefficient of the lateral acceleration sequence and the load lateral offset rate of change sequence within the same sliding window is calculated as the disturbance strength of lateral motion on the load's lateral stability; the Pearson correlation coefficient of the longitudinal acceleration sequence and the load height rate of change sequence within the same sliding window is calculated as the disturbance strength of longitudinal motion on the load's vertical stability.

[0042] The aforementioned temporal fluctuation features and cross-dimensional correlation and interaction features together constitute derived features, which, together with the basic features of the four feature subsets in S2.2, enter the subsequent fusion steps.

[0043] S2.4. Perform feature-level fusion and splicing on the subset of operation requirements features, subset of driving response features, subset of vehicle execution features, subset of load stability features, and derived features according to the time sequence dimension to output the fusion feature set of driving behavior risks.

[0044] Specifically, using a unified time label as an index, for each sampling moment, all basic feature values ​​contained in the subsets of job requirement features, driving response features, vehicle execution features, and load stability features, as well as all time-series fluctuation feature values ​​and cross-dimensional correlation interaction feature values ​​corresponding to that moment, are arranged and concatenated into a fixed-length feature vector in a preset order. This vector serves as the driving behavior risk fusion feature vector for that sampling moment. The concatenation order is determined during system initialization and remains unchanged during operation. A typical order is: job requirement features, driving response features, vehicle execution features, load stability features, time-series fluctuation features, and cross-dimensional correlation interaction features.

[0045] Furthermore, the fused feature vectors from each sampling moment are arranged sequentially along the time dimension, forming a two-dimensional data table with time as the row index and feature dimension as the column index, which is the driving behavior risk fusion feature set. Each row corresponds to a complete risk feature description for a sampling moment, and each column corresponds to a specific risk-related feature dimension. The total number of columns for feature dimensions is the sum of the number of basic feature dimensions of the four feature subsets and the total number of derived feature dimensions.

[0046] After the above processing, the driving behavior risk fusion feature set serves as the input for step S3, providing unified and structured feature data for task micro-stage identification, dual-dimensional mismatch calculation of driving response and vehicle execution, recursive deduction of inconsistent task response states, and calculation of driving risk recovery margin.

[0047] S3. Driving Behavior Risk Reasoning and Grading: Based on the fusion feature set of driving behavior risks, a joint reasoning mechanism is used to assess driving behavior risks through inconsistent state deduction of operational response, inconsistent driving risk recovery margin deduction, and correction of operational load stability occupancy factors. The output is a driving behavior risk grading result that includes risk level, intervention urgency, and permissible intervention method constraints. Specifically, this step uses the fusion feature set of driving behavior risks output from S2 as input. Through a joint reasoning mechanism involving dynamic expectation matching of operational micro-stages, dual-dimensional mismatch calculation of driving response and vehicle execution, inconsistent state deduction of operational response, correction of operational load stability occupancy factors, and assessment of driving risk recovery margin, the current driving behavior risks of port mobile machinery are quantitatively assessed. The output is a driving behavior risk grading result that includes risk level, intervention urgency, and permissible intervention method constraints, providing a decision-making basis for the graded early warning and proactive intervention control in S4. This includes the following steps: S3.1 Match the current job process with the corresponding job micro-stage to generate the scenario-based driving response expectation range and vehicle execution expectation range; the following will explain each of the sub-steps in detail: S3.11 Extract vehicle spatial location, driving direction, operating mechanism status, load status and time sequence change features from the driving behavior risk fusion feature set. Through time sequence feature matching and probabilistic inference, obtain the belonging probability of each operation micro-stage at the current moment and generate the operation micro-stage probability vector. In this step, the extracted features include: vehicle global coordinate position, heading angle, travel speed vector, shortest distance to the work boundary, shortest distance to the obstacle, lifting stroke of the working mechanism and its rate of change, load weight and its rate of change, and the changing trends of the above features within a preset time window. These features are arranged in a preset order to form the feature vector for the current moment. Each micro-stage of the operation corresponds to a preset reference feature vector, which is composed of the statistical mean of each feature under typical working conditions at that stage, and its dimension is consistent with the extracted feature vector.

[0048] The Euclidean distance between the current feature vector and the preset reference feature vectors for each micro-stage is calculated to obtain the distance value between the current feature vector and the reference vectors for each stage. Further, the distance values ​​corresponding to each stage are converted into assignment probabilities using a softmax mapping function. The calculation method is as follows: after taking the negative value of the distance value, softmax normalization is applied, so that stages with smaller distances receive higher assignment probabilities, and the sum of the assignment probabilities of all stages is 1. The micro-stages are arranged in a preset order, and their assignment probabilities constitute a micro-stage probability vector. The vector dimension is equal to the preset total number of micro-stage types, and the value range of each element is [0,1]. This probability vector represents the degree to which the current vehicle operation status is assigned to each micro-stage.

[0049] S3.12. Using the probability of each stage in the operation micro-stage probability vector as the weight, the pre-configured driving response expectation template and vehicle execution expectation template corresponding to each operation micro-stage are weighted and fused to generate a driving response expectation interval and a vehicle execution expectation interval that match the current operation process.

[0050] In this step, the driving response expectation template for each operational micro-stage defines the expected lower and upper limits of various driving response characteristics for that stage, including the expected ranges for throttle opening change rate, steering angle, operating frequency, and line-of-sight angle. The vehicle execution expectation template defines the expected lower and upper limits of various vehicle execution characteristics for that stage, including the expected ranges for vehicle speed, longitudinal acceleration, lateral acceleration, yaw rate, and operating mechanism movement rate. These expectation templates are obtained based on historical operational data statistics or operational procedure calibration of port mobile machinery under normal operating conditions and are pre-stored in the storage medium of the onboard computing environment.

[0051] The weighted fusion calculation method is as follows: for each feature in the driving response expectation template, its expected lower limit value at each operational micro-stage is multiplied by the corresponding stage's attribution probability and then summed to obtain the dynamic expected lower limit value of that feature at the current moment; the dynamic expected upper limit value is calculated in the same way. For the th Each driving response feature has a dynamic expectation range as follows: ; in, This indicates the total number of micro-stages in the task. Indicates the current time No. The probability of belonging to each micro-stage of the task. and They represent the first The first micro-stage of the task The expected lower limit and expected upper limit of each driving response characteristic. and They represent the weighted fusion results of the first... The lower limit and upper limit of the dynamic expectation of each driving response feature.

[0052] The weighted fusion method for the vehicle's expected range is the same as described above. For the first... The dynamic expectation range of a vehicle's execution characteristics is: ; in, and They represent the first The first micro-stage of the task The expected lower limit and expected upper limit of the performance characteristics of a vehicle. and They represent the weighted fusion results of the first... The lower limit and upper limit of the dynamic expectation of each vehicle's execution characteristics.

[0053] Because the probability vector dynamically changes with the feature extraction results at the current moment, the generated expectation interval can smoothly follow the progress and stage switching of the actual operation, avoiding the expectation deviation caused by fixed thresholds or single stage templates during stage transitions. The generated driving response expectation interval and vehicle execution expectation interval serve as reference benchmarks for deviation calculation in S3.2.

[0054] S3.2 Calculate the degree of mismatch in the driving response dimension and the vehicle execution dimension respectively, and assess the degree of transmission of operational abnormalities to operational risks. Specifically, driver operational deviations are not directly equivalent to vehicle operational risks; the correlation between the two depends on the type and magnitude of the operational deviation and the vehicle's sensitivity to the operational deviation under specific operating conditions. This step calculates the deviations in the driving response dimension and the vehicle execution dimension respectively, and quantifies the temporal transmission relationship between the two through a deviation correlation matrix, using the transmission correlation as a quantitative indicator of the degree of transmission. The following sections will explain each sub-step in detail: S3.21. Compare the driver's state and behavior characteristics in the driving behavior risk fusion feature set with the expected range of driving response feature by feature. Calculate the standardized deviation for features that exceed the expected range. Combine the weighted sum of the features corresponding to the current operation stage to obtain the comprehensive driving response deviation. In this step, for the first Each driving response feature, its actual value at the current moment. The dynamic expectation interval generated by S3.12 Perform a comparison. If... If the value lies within the expected range, then the standardized deviation of the feature is zero; if If the value exceeds the upper limit of the expected range, the standardized deviation is the difference between the actual value and the upper limit divided by half the width of the expected range; if If the value is below the lower limit of the expected range, the standardized deviation is the difference between the lower limit and the actual value divided by half the width of the expected range. ; in, Indicates the first Individual driving response characteristics The standardized deviation at time is dimensionless and takes the value of a non-negative real number, where zero indicates no deviation and a larger value indicates a more severe deviation. Indicates the first Individual driving response characteristics The actual value at that moment.

[0055] The overall driving response deviation is obtained by weighted summation of the standardized deviations of each driving response feature: ; in, This represents the total number of driving response characteristics. Indicates the first The weighting coefficients of each driving response characteristic in the comprehensive deviation evaluation satisfy the following: The weights are obtained by looking up the maximum belonging stage index of the probability vector of the micro-stage of the task from the preset weight configuration table, which simplifies the implementation of the project; alternatively, the weighted fusion of the probability vectors of the micro-stage of the task can be used to calculate the weights of the features corresponding to each micro-stage of the task, so as to obtain the dynamic weight coefficients and achieve smooth switching during stage transition. express The overall driving response deviation at any given moment is dimensionless.

[0056] S3.22. The on-board operation and work mechanism features of the driving behavior risk fusion feature set are compared with the expected range of vehicle execution feature by feature. For features that exceed the expected range, the standardized deviation is calculated. The deviation is then weighted and summed with the feature weights corresponding to the current operation stage to obtain the comprehensive vehicle execution deviation. In this step, the calculation method is the same as in S3.21, for the first... Actual value of each vehicle's performance characteristics Its dynamic expected range Perform a comparison and calculate the standardized deviation. , definition form and Consistent. The total deviation of the vehicle's performance is: ; in, This represents the total number of vehicle execution features. Indicates the first The weighting coefficients of each vehicle's execution characteristics satisfy... The weights are obtained by looking up the maximum belonging stage index of the probability vector of the micro-stage of the task from the preset weight configuration table, which simplifies the implementation of the project; alternatively, the weighted fusion of the probability vectors of the micro-stage of the task can be used to calculate the weights of the features corresponding to each micro-stage of the task, so as to obtain the dynamic weight coefficients and achieve smooth switching during stage transition. express The overall vehicle deviation at any given time is dimensionless. Indicates the first Individual vehicle execution characteristics Standardized deviation at time, Indicates the first Individual vehicle execution characteristics The actual value at that moment.

[0057] S3.23. Within the preset time window, calculate the temporal transmission matching degree between driving response deviation and vehicle execution deviation through the deviation correlation matrix corresponding to the current operation stage, and generate the transmission correlation quantity of driving response deviation to vehicle execution result.

[0058] In this step, the deviation correlation matrix is ​​used to characterize the transmission influence of each driving response characteristic deviation on each vehicle execution characteristic deviation under a specific operational stage. For the first... Each micro-stage of the operation deviates from the correlation matrix. The dimension is Its elements Indicates the first The driving response characteristic deviates from the first The transmission coefficients are the deviations in vehicle performance characteristics. The values ​​of these transmission coefficients are obtained through statistical analysis of historical data from the operational process or through expert calibration. A typical calibration method involves collecting sample data showing simultaneous deviations in driving response characteristics and vehicle performance characteristics at each micro-stage of the operation, and determining the magnitude of the transmission coefficients through correlation or regression analysis. The sum of the elements in each row of the matrix does not need to equal 1; each transmission coefficient independently reflects a specific cross-dimensional transmission relationship.

[0059] Considering that the current task is not in a single stage, this step uses the probability vector of the task's micro-stage as the weight to perform a weighted fusion of the deviation correlation matrices of each stage, thus obtaining the dynamic deviation correlation matrix at the current moment: ; in, express The dynamic deviation correlation matrix at time t.

[0060] Furthermore, within a preset time window, the standardized deviation vector of the driving response at each sampling time within the window is obtained. and vehicle standardized deviation vector Calculate the transitive correlation: ; in, This indicates the number of sampling points contained in the timing window. This represents the standardized deviation of each driving response characteristic. 3D column vector, This represents the standardized deviation of the performance characteristics of each vehicle. 3D column vector, superscript Indicates transpose. express The transmission correlation quantity at any given moment has a dimension of one. The larger the value, the stronger the transmission correlation between the driving response deviation and the vehicle execution deviation within the current time window. In other words, the abnormal driver operation has been clearly transmitted into an abnormal response of the vehicle's operating state.

[0061] Comprehensive driving response deviation Comprehensive vehicle deviation and the amount of correlation Together they serve as inputs for the recursive derivation of inconsistent job response states in S3.3.

[0062] S3.3. Generate inconsistent operation response states through time-series recursion to quantify the formation and persistence of risks. Specifically: Obtain the inconsistent operation response state at the previous moment, and combine it with the risk memory coefficient, comprehensive driving response deviation, comprehensive vehicle execution deviation, and transmission correlation corresponding to the current operation stage. Use a time-series recursion rule with memory decay to calculate the inconsistent operation response state at the current moment. When the vehicle is in reversing positioning, load transfer, densely populated areas, or near the operation boundary, the value of the risk memory coefficient is increased accordingly to prolong the duration of the risk state's continuous memory, thus quantifying the formation and persistence of the risk of mismatch between driving response and operation requirements.

[0063] In this step, the inconsistency in operational response is a continuous and dynamic measure of the risk of mismatch between driver operation and operational requirements. Unlike instantaneous deviation judgment based on a single sampling point, this step uses a time-series recursive rule with memory decay, giving the risk status cumulative and decaying characteristics, thereby distinguishing between short-term operational fluctuations and persistent hazardous behaviors.

[0064] Specifically, the current moment Inconsistent job response status From the state of the previous moment The remaining amount after memory decay consists of two parts: the incremental risk at the current moment and the additional risk. The recursive formula is as follows: ; in, express The job response at any given time is inconsistent, dimensionless, and has an initial value. Take zero, This indicates an inconsistent state in the job response from the previous moment. express The memory retention coefficient at time step [0,1] represents the proportion of the risk state from the previous time step that is retained at the current time step. This represents the increase in risk at the current moment. It is calculated recursively. If the preset limit is exceeded Then take , From normal operating conditions The high quantile values ​​of the statistical distribution are used to ensure that the risk status has clear physical boundaries and classification benchmarks.

[0065] Furthermore, the above The calculation is based on the three deviation indicators output from S3.2: ; in, , , The preset fusion coefficient satisfies These are used to adjust the contribution ratios of driving response deviation, vehicle execution deviation, and transmission correlation to the risk increment, respectively. It is a non-negative value; the larger the value, the stronger the risk of new mismatch at the current moment.

[0066] Furthermore, the memory retention coefficient Risk memory coefficient The relationship is: ; in, This represents the basic memory retention coefficient, with a value ranging from (0,1). It is preset by the system, with typical values ​​between 0.85 and 0.95, corresponding to the natural decay rate of a risk state under normal operating conditions when no new risks are introduced. This represents the risk memory coefficient, used to adaptively adjust the memory retention strength based on the risk level of the current work scenario. It is calculated as follows. If the value is greater than 1, then take the value 1, ensuring that its value always falls within the range of [0,1].

[0067] When the vehicle is in reversing positioning mode (determined by driving direction and speed characteristics), load transfer mode (determined by load weight and operating mechanism status characteristics), densely populated area (determined by densely populated area signage characteristics), or near the work boundary (determined by comparing the shortest distance to the work boundary with a preset boundary distance threshold), The memory retention coefficient is increased to a preset high-risk scenario memory coefficient value. This value can be calibrated using real vehicle test data or determined with reference to port safety operating procedures, with a typical range of 1.05 to 1.20, thus improving the memory retention coefficient. It is improved from the base value, getting closer to 1 to slow down the risk decay rate, thereby extending the duration of the risk state's memory. When the vehicle is not in the above-mentioned high-risk scenarios, The risk status is restored to 1.0, and the risk status decays normally at the base rate. Through this adaptive adjustment, the inconsistent operation response status can maintain a longer risk memory of intermittent deviations in high-risk scenarios, reducing missed judgments caused by drivers' brief convergence operations in high-risk scenarios.

[0068] S3.4 Calculate the operational load stability occupancy factor based on load and motion conditions to correct the vehicle's safety correction capability boundary. Specifically, in port mobile machinery's load transfer and stacking operations, the load's weight, height, lateral offset, and the vehicle's steering motion all consume the vehicle's stability margin, directly affecting the vehicle's longitudinal deceleration capability, lateral steering capability, and operating mechanism return capability available for safety correction. This step generates the operational load stability occupancy factor through multi-factor coupling calculation and performs differentiated attenuation correction on the basic correctable capability of the three correction channels to obtain the effective correctable capability of each channel under the current load and motion conditions. The following sections will explain each sub-step in detail: S3.41 Extract four types of normalized features from the driving behavior risk fusion feature set: load amount, load height, load lateral offset, and vehicle steering motion intensity. Obtain the working load stability occupancy factor through multi-factor coupling calculation to quantify the degree of occupancy of the vehicle's stability margin by the current load and motion conditions. In this step, the load normalization feature is... Defined as the ratio of the current load weight to the vehicle's rated load. Load height normalization characteristic. Defined as the ratio of the current load height to the vehicle's maximum lifting height. Load lateral offset normalization feature. Defined as the ratio of the current load lateral offset to the maximum permissible lateral offset. The load lateral offset is obtained by a load sensor or gantry displacement sensor, and the maximum permissible lateral offset is pre-calibrated based on the vehicle model and load height. Vehicle steering motion intensity normalization characteristic. It is defined as the ratio of the absolute value of the current yaw rate to the maximum allowable yaw rate at the current vehicle speed, where the maximum allowable yaw rate is obtained by looking up a table based on the speed-yaw relationship curve preset by the vehicle stability dynamics boundary.

[0069] The value range of the above four types of normalized features is [0,1]. The larger the value, the more serious the occupation of the stability margin by that dimension.

[0070] Operational load stability occupancy factor The following results were obtained from four types of normalized features through multi-factor coupling operations: ; in, Indicates the first The coupling weight coefficients of the normalized features, with values ​​ranging from (0,1], are used to characterize the relative importance of each dimension's influence on the overall vehicle stability margin. They are determined by vehicle dynamics calibration tests or simulation analysis. This represents the operating load stability occupancy factor, with values ​​between [0,1]. A larger value indicates a more severe occupancy of the vehicle's stability margin by the current load and operating conditions, resulting in a smaller remaining stability margin available for safety correction. When the value is close to 1, the vehicle has approached or reached the stability boundary, and the intensity and method of correction operations should be strictly limited.

[0071] S3.42. Using the working load stability occupancy factor as the correction coefficient, the basic correctable capabilities of the three correction channels—longitudinal deceleration, lateral steering, and working mechanism retraction—are attenuated and corrected to obtain the effective correctable capabilities of the vehicle corresponding to each channel under the current load condition.

[0072] In this step, the basic correctability of the three correction channels is defined as follows: The basic correctability of the longitudinal deceleration path It represents the absolute value of the maximum safe braking deceleration of the vehicle under no-load and normal driving conditions. Basic correctability of lateral steering channels This is the maximum safe steering rate of the vehicle under no-load and normal driving conditions. Basic correctability of the operating mechanism's callback channel This is the maximum safe return rate of the operating mechanism under no-load conditions.

[0073] The aforementioned basic values ​​are determined by vehicle technical parameters or calibration data and are pre-stored in the storage medium of the on-board computing environment.

[0074] Based on the stable occupancy factor of the operating load As a basis for attenuation correction, the effective correctability of each channel is as follows: ; ; ; in, , , They represent Effective corrective capability of the longitudinal deceleration channel, lateral steering channel, and working mechanism return channel at all times. , , These represent the load sensitivity coefficients for the longitudinal deceleration, lateral steering, and work mechanism return channels, respectively, reflecting the differences in sensitivity of each correction channel to load stability. The work mechanism return channel is typically the most sensitive to load conditions, with a relatively large sensitivity coefficient; the longitudinal deceleration channel is next; and the sensitivity coefficient for the lateral steering channel is determined comprehensively based on vehicle wheelbase, track width, and center of gravity height parameters. All three sensitivity coefficients are pre-stored in the onboard computing environment's storage medium after calibration. The effective correctability of each channel varies with... The increase and decrease reflect the narrowing of the actual safety correction space of the vehicle under load coupling conditions.

[0075] S3.5 generates the driving risk recovery margin and, combined with the inconsistent operation response state, driving risk recovery margin, and load constraints, completes a joint classification, outputting the driving behavior risk classification result. Specifically, the inconsistent operation response state output in S3.3... The cumulative risk level is quantified, and the effective correctable capability output by S3.4 defines the safe correction boundary under the current operating condition. This step further evaluates the matching relationship between the amount of correction required to restore the current vehicle deviation to a safe state and the actual available correction capability, and generates the driving risk recovery margin through asymmetric recursion. Finally, the three factors are combined to complete the risk classification. The following sections explain each sub-step in detail: S3.51. Based on the comprehensive vehicle execution deviation, calculate the necessary correction amounts for the three channels of longitudinal deceleration, lateral steering, and working mechanism reversal required to restore to a safe state, and calculate the normalized recovery margin for each correction channel in combination with the vehicle's effective correctability. In this step, the overall vehicle deviation is considered. As a comprehensive index with a single dimension, it needs to be mapped to the correction amounts required by three specific correction channels. The mapping relationship is achieved through a pre-defined channel allocation vector: ; ; ; in, , , They represent The necessary correction amounts for the longitudinal deceleration channel, the lateral steering channel, and the working mechanism return channel at all times. , , The channel allocation coefficient reflects the proportion of vehicle deviation distributed across each correction channel, satisfying... The allocation coefficients for each channel are obtained by looking up the maximum belonging stage of the current operation micro-stage probability vector from a preset allocation coefficient table, which simplifies the implementation of the project. Alternatively, a weighted fusion method of operation micro-stage probability vectors can be used to calculate the allocation coefficients corresponding to each operation micro-stage by weighting them to obtain dynamic allocation coefficients and achieve smooth switching during stage transitions.

[0076] Furthermore, the normalized recovery margin of each correction channel is defined as the ratio of the channel's current effective correctability minus the necessary correction amount to its effective correctability: ; ; ; in, , , These represent the normalized recovery margins for longitudinal deceleration, lateral steering, and working mechanism reversal channels, respectively. A positive normalized recovery margin indicates that the channel still has remaining corrective capacity to cover the current required correction. A normalized recovery margin approaching zero or being negative indicates that the channel's corrective capacity is insufficient or exhausted. Each channel's normalized recovery margin is a real number, with an upper limit of 1 and a lower limit that can be truncated to a preset negative boundary in actual calculations. When all corrective channels have negative normalized recovery margins, the vehicle's corrective capacity is deemed exhausted, and control is implemented at the highest risk level. Simultaneously, the intensity of intervention is strictly limited to not exceed the current effective corrective capacity to avoid secondary safety accidents.

[0077] S3.52. Extract the response hysteresis, operation amplitude, and deviation descent rate features of the driver's corrective operation from the driving behavior risk fusion feature set. Use the minimum normalized recovery margin of each correction channel as the recursive base value. Adjust the decay step size of risk deterioration with response hysteresis and adjust the recovery coefficient of risk recovery with operation amplitude and deviation descent rate. Use the asymmetric recursive rule of rapid decline when risk deteriorates and slow rise when risk recovers to calculate the driving risk recovery margin at the current moment. In this step, the driver's corrective action features are extracted from the driving behavior risk fusion feature set.

[0078] Response lag Defined as the average time delay from the occurrence of a vehicle deviation to the driver initiating an effective corrective action. The moment when the vehicle deviation occurs is... The first time the deviation exceeds the preset threshold The sampling time is used as the standard. The start time of an effective corrective operation is defined as the moment when the driver's operating characteristics undergo a significant change opposite to the direction of deviation. This significant change is determined by the amount of change in brake pedal travel, throttle opening, or steering angle exceeding a preset operating response threshold. The average of the above time differences is calculated within the most recent preset statistical window, with a window length of 5 to 10 seconds. The unit is seconds. If the driver continues to fail to perform effective corrective actions, then... Set a preset maximum hysteresis time limit to avoid the value from increasing unbounded.

[0079] Operating range Defined as the normalized value of the amplitude of the driver's current corrective action. When there is a single corrective action, the ratio of the actual travel or angle of the action to its maximum physical travel or angle is used as the normalized value; when there are multiple corrective actions, the maximum value among the normalized values ​​of each action is taken as the current action amplitude. The value range is [0,1].

[0080] Deviation from descent rate Defined as the rate of decrease of the overall vehicle deviation within the most recent preset time window, calculated in the form of a first-order difference in the discrete domain: ; in, This indicates the number of sampling points included in the deviation from the descent rate calculation window. Indicates a uniform sampling step size. A positive value indicates that the deviation is decreasing, while a zero or negative value indicates that the deviation is not decreasing or is still increasing.

[0081] Take the minimum value of the normalized recovery margin of the three correction channels as the recursive base value: ; in, express The minimum value of the normalized recovery margin of the three correction channels at any given time.

[0082] Furthermore, driving risk recovery margin Calculated using an asymmetric recursive rule, when the recovery margin deteriorates... Less than the value of the previous time step hour, The previous time value Subtract the attenuation amount from the base; when the recovery margin remains unchanged or improves, that is... When the value is greater than or equal to the value at the previous time step, Add the rebound amount to the previous time value: when This means that the recovery margin has worsened compared to the previous moment. Subtract the decay amount from the previous time value: ; when That is, to restore the remaining balance to maintain or improve the condition. Add the rebound amount to the previous time value: ; in, express The driving risk recovery margin at any given time, dimensionless, with a value range of [0,1], and an initial value. Take 1.0, This indicates the margin of recovery from driving risk at the previous moment. This indicates the decay step size when the risk worsens. This represents the recovery coefficient when the risk recovers. This represents the minimum normalized recovery margin of the three correction channels at the previous moment. This indicates a uniform sampling step size.

[0083] Furthermore, the decay step size Based on the basic decay step size and response lag Jointly determined: ; in, This represents the base decay step size, which is a preset positive constant. This represents the response hysteresis sensitivity coefficient, which is a preset normal value. This formula indicates that the slower the driver's response, the faster the recovery margin decreases when the risk worsens.

[0084] Meanwhile, the recovery coefficient Based on the basic recovery coefficient Operating range and deviation from the rate of descent Jointly determined: ; in, This represents the baseline recovery coefficient, which is a preset normal value. This represents the sensitivity coefficient to the amplitude of the operation, which is a preset normal value. This represents the sensitivity coefficient for deviation from the descent rate, which is a preset normal value. This formula indicates that the greater the driver's corrective action and the faster the deviation from the descent rate, the faster the recovery margin recovers upon risk recovery.

[0085] In asymmetric recursion, the basic recovery coefficient Significantly smaller than the basic decay step size , typically, The possible values ​​are The margin is 0.2 to 0.5 times that of the risk, which makes the recovery margin decrease rapidly when the risk worsens and increase slowly when the risk recovers, which is in line with the engineering principle of rapid response and slow release of risk in the field of driving safety. In actual calculations, it must satisfy 0. Boundary constraints are applied, and the boundary value is taken when the recursive result exceeds the boundary.

[0086] S3.53. The abnormal threshold of the inconsistent state of the operation response is used as the basis for determining the establishment of risk, and the driving risk recovery margin is lower than the safety threshold as the basis for determining the intervention trigger. Combined with the operation load stability occupancy factor to limit the range of permissible intervention methods, the risk level is jointly classified and the intervention urgency is calculated. The output is a driving behavior risk classification result that includes risk level, intervention urgency and permissible intervention method constraints.

[0087] In this step, the inconsistent job response status output by S3.3 is considered. S3.52 output driving risk recovery margin and the job load stability occupancy factor output by S3.41 Risk classification is completed through the following three-level judgment logic: The first level of judgment is the risk assessment. This involves determining the inconsistency in the current operational response. Compared with the preset abnormal threshold Compare them. The settings are based on historical data statistical analysis or the experience of safety experts. The typical setting method is to take the normal operating conditions. The 95th percentile of a statistical distribution. Less than If no persistent driving risk is identified, the risk level is set to Level I, and no further intervention procedures are triggered; this round of risk assessment ends. Greater than or equal to If the driving behavior is deemed to pose a risk, the process proceeds to the second level of assessment.

[0088] The second level of assessment involves intervention triggering and urgency calculation. If a risk is established, the driving risk recovery margin is determined. With preset safety threshold Compare them. The typical value range is 0.3 to 0.5, determined by statistical analysis of vehicle safe braking distance and driver average reaction time. Greater than At this time, it indicates that the vehicle still has sufficient recovery margin, and the risk level is set to Level II, maintaining monitoring without triggering active intervention. Less than or equal to At that time, the need for proactive intervention is triggered, and based on Below Calculate the urgency of intervention The calculation method is as follows: ; in, This indicates the urgency of the intervention, with values ​​ranging from (0,1). A higher value indicates a more urgent intervention. According to... The value of categorizes the urgency of intervention into three levels: The interval (0, 0.3) is classified as Level 3 Urgency. The interval (0.3, 0.7) is considered level two urgent. The interval (0.7, 1) is considered to be of level one urgency.

[0089] The third level of judgment is based on intervention method constraints. This is determined by the operational load stability occupancy factor. This is based on limiting the scope of permissible intervention methods. First load safety threshold. Second load safety threshold As a preset threshold, both satisfy Less than ,in The typical value range is 0.3 to 0.5. The typical value range is 0.6 to 0.8, and the specific value is determined by vehicle stability tests or dynamic simulation analysis. When Less than At this time, the load stabilization requirement is relatively small, allowing for any one or a combination of three intervention methods: longitudinal deceleration, lateral steering, and working mechanism retraction. Greater than or equal to and less than or equal to At this time, lateral steering intervention is restricted or prohibited, allowing only longitudinal deceleration and working mechanism retraction. Greater than At the same time, longitudinal deceleration and lateral steering intervention are strictly limited, and only the working mechanism return is allowed as the main intervention method to avoid load instability caused by braking or steering actions.

[0090] After completing the above three-level assessment, the risk level and the urgency of intervention will be determined. The final classification result is output by combining the constraints of the intervention method with the overall constraint. Greater than or equal to , Less than or equal to and When the value is less than or equal to 0.7, the risk level is Level III, indicating that proactive intervention is needed but not yet urgent. Greater than or equal to , Less than or equal to and When the value is greater than 0.7, the risk level is Level IV, indicating that the strongest intervention needs to be implemented immediately.

[0091] Based on the above three-level joint assessment, this step outputs a driving behavior risk classification result, which includes the following information in structured data form: risk level and intervention urgency. The quantitative values ​​and corresponding urgency levels, as well as the constraints on permissible intervention methods, are determined. The permissible intervention method constraints specifically include the enabling flags for each intervention channel and the upper limit of permissible execution intensity. This grading result serves as input to step S4, used to dynamically match the grading warning strategy with the issuance of intervention commands from the onboard equipment.

[0092] S4. Tiered Early Warning and Active Intervention Control: Based on the results of driving behavior risk grading, this step dynamically matches tiered early warning strategies with onboard equipment intervention commands and issues them for execution, aiming to reduce driving safety risks during port mobile machinery operations. Specifically, this step takes the driving behavior risk grading results output in S3.53 as input, analyzes the risk level, intervention urgency, and allowed intervention method constraints, dynamically matches the corresponding tiered early warning strategies with onboard equipment intervention commands, and issues the warnings and commands to the onboard execution terminal to achieve tiered control over the driving behavior risks of port mobile machinery. This includes the following steps: S4.1 Analyze the driving behavior risk classification results and extract the corresponding risk level, intervention urgency, and allowable intervention method constraints; specifically, the classification results output by S3.53 are structured data, including risk level and intervention urgency. The risk level consists of three fields: a quantitative value, a corresponding risk level, and permissible intervention methods. This step analyzes the risk level results field by field: the risk level field takes values ​​of Level I, Level II, Level III, or Level IV; the intervention urgency field contains a quantitative value. The corresponding urgency level; the allowable intervention method constraint field includes the enable flags and upper limits of the allowed execution intensity for the longitudinal deceleration channel, lateral steering channel, and working mechanism callback channel. The parsed fields are used by S4.2 and S4.3 respectively.

[0093] S4.2. Based on the risk level and the urgency of intervention, match the corresponding level of early warning strategy and determine the early warning triggering method and output sequence; Specifically, the early warning strategy is divided into multiple levels based on risk level and intervention urgency: When the risk level is Level I, no warning is triggered, and the system remains in normal monitoring status.

[0094] When the risk level is Level II, a warning at the "Attention Level" is triggered. The warning is displayed as a text message or indicator light on the vehicle screen, without an audible alarm. It is used to remind the driver that the current driving behavior is risky but still within a controllable range. The warning output is a single trigger and is updated when the status changes.

[0095] When the risk level is Level III, a warning is triggered. The warning is issued through a combination of an audible alarm and a text prompt on the vehicle-mounted display screen, with the audible alarm using an intermittent buzzer. Simultaneously, the warning information is reported to the port safety monitoring backend via a wireless communication module. The warning output is periodically triggered, repeating at preset time intervals until the risk level is downgraded.

[0096] When the risk level is Level IV, an emergency warning is triggered. The warning method includes a continuous audible alarm, a flashing indicator on the vehicle's display screen, and a warning light in the driver's cab. The warning information is simultaneously reported to the port's safety monitoring backend, along with the vehicle's location and current operational status. The warning output is continuously triggered until the risk level is downgraded or confirmed by manual intervention.

[0097] S4.3. Based on the constraints of permissible intervention methods, select on-board equipment intervention commands that are suitable for the current load conditions, and clarify the action channel and execution intensity of the intervention. Specifically, when S4.2 triggers the early warning, the urgency of intervention... When the value is greater than 0, this step filters executable intervention instructions based on the allowed intervention methods.

[0098] The intervention command library contains various pre-set intervention commands for onboard equipment, categorized by their application channel into longitudinal deceleration commands, lateral steering commands, and work mechanism reversal commands. Longitudinal deceleration commands include engine torque limiting commands and brake assist boost commands; lateral steering commands include steering rate limiting commands; and work mechanism reversal commands include lifting lock commands, tilt reset commands, and lateral shift return commands. Each command has a defined maximum permissible intensity limit.

[0099] During the screening process, the enable flags of each intervention instruction's channel within the allowed intervention method constraints are checked one by one. Channels with an enable flag indicating permission are included in the executable instruction set; channels with an enable flag indicating prohibition are excluded from their corresponding instructions. For instructions included in the executable instruction set, their execution strength is the smaller value between the instruction's preset strength and the upper limit of allowed execution strength.

[0100] Furthermore, based on the urgency of intervention The severity level determines the execution scope and intensity of the commands. At Level 3 urgency, only the lifting lock command within the working mechanism callback category is executed to reduce the load height and improve stability. At Level 2 urgency, in addition to executing Level 3 urgency commands, an engine torque limit command within the longitudinal deceleration category is added to restrict vehicle acceleration. At Level 1 urgency, all executable commands are executed, including engine torque limit commands, brake assist boost commands, and corresponding working mechanism callback commands, achieving maximum intensity of active intervention. If all channels corresponding to the current urgency level are restricted by allowed intervention methods, hardware intervention commands are not executed; only the corresponding level of warning output is retained, and this abnormal condition information is simultaneously reported to the port safety monitoring backend.

[0101] S4.4. The graded early warning strategy and the intervention instructions of the vehicle equipment are sent to the corresponding execution terminal for execution, so as to realize the graded control of driving behavior risks.

[0102] Specifically, the output targets of the early warning strategy include the vehicle-mounted display terminal, the vehicle-mounted audible and visual alarm device, and the wireless communication module. The vehicle-mounted display terminal receives the warning text information and prompt type, the vehicle-mounted audible and visual alarm device receives the sound alarm trigger signal and the warning light trigger signal, and the wireless communication module receives the reported data packets.

[0103] The output targets of the on-board equipment intervention commands include the engine controller, brake controller, and work mechanism controller. The engine controller receives torque limiting commands and adjusts the upper limit of throttle opening; the brake controller receives brake assist boost commands and pre-builds brake line pressure; and the work mechanism controller receives lifting lock commands, tilt reset commands, or lateral shift return commands and executes the corresponding mechanism actions.

[0104] Upon receiving the instruction, each execution terminal executes it according to the instruction content and timing. The system continuously monitors the risk classification results output by S3. When the risk level is downgraded or the intervention urgency reaches zero, the warning output and intervention instructions are gradually lifted, restoring the vehicle's normal operating authority. Through the closed-loop control of the above-mentioned graded warning and active intervention, the risk of driving behavior of port mobile machinery is managed step by step.

[0105] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0106] 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 claimed invention.

Claims

1. A method for AI-based monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Real-time acquisition of onboard operation data, driver status and behavior data, port operation environment and spatial positioning data, load and operation mechanism status data of port mobile machinery under different operating conditions, and time-series alignment and standardization preprocessing of multi-source data to obtain standardized multi-source monitoring datasets. S2. Multidimensional feature extraction and fusion: Based on the standardized multi-source monitoring dataset, multidimensional features reflecting the evolution of driving behavior risks are extracted through spatiotemporal dimension alignment, multidimensional feature mapping, and feature-level fusion processing to obtain a driving behavior risk fusion feature set; S3. Driving behavior risk reasoning and classification: Based on the driving behavior risk fusion feature set, the driving behavior risk is judged by a joint reasoning mechanism of work response inconsistency state recursion, driving risk recovery margin recursion, and work load stability occupancy factor correction, and the driving behavior risk classification result includes risk level, intervention urgency and allowable intervention method constraints. S4. Tiered early warning and active intervention control: Based on the risk grading results of the driving behavior, the tiered early warning strategy and the intervention instructions of the on-board equipment are dynamically matched and issued for execution, in order to reduce the driving safety risks during the operation of mobile machinery in the port.

2. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 1, characterized in that, In step S1, time-series alignment and standardization preprocessing are performed on the multi-source data to obtain a standardized multi-source monitoring dataset, including the following steps: S1.1 Collect on-board operation data, driver status and behavior data, port operation environment and spatial positioning data, and load and operation mechanism status data of port mobile machinery under different operating conditions through corresponding sensing terminals according to data categories, and obtain raw multi-source monitoring data. S1.

2. Use a unified clock reference to perform timestamp alignment and sampling granularity unification processing on the original multi-source monitoring data to generate time-synchronized multi-source monitoring data; S1.

3. Perform outlier identification and jump data correction on the time-synchronized multi-source monitoring data, remove invalid data segments, and obtain cleaned multi-source monitoring data. S1.4 Perform dimensional unification and standardization mapping on the cleaned multi-source monitoring data to output a standardized multi-source monitoring dataset.

3. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 1, characterized in that, In step S2, multi-dimensional features reflecting the evolution of driving behavior risks are extracted based on the standardized multi-source monitoring dataset to obtain a driving behavior risk fusion feature set, including the following steps: S2.1 Using the standardized multi-source monitoring dataset as the processing object, unify the spatial coordinate system, complete the spatiotemporal benchmark calibration of the multi-source data, and generate spatiotemporal dimension aligned multi-source basic feature data; S2.2 Map the spatiotemporal aligned multi-source basic feature data to the operation requirement feature subset, driving response feature subset, vehicle execution feature subset, and load stability feature subset according to the representation attributes, and complete the dimensional feature classification; S2.3 Extract time-series fluctuation features and cross-dimensional correlation and interaction features for each feature subset to generate derived features that reflect the evolution of driving behavior risks; S2.

4. Perform feature-level fusion and splicing on the subset of operation requirements features, subset of driving response features, subset of vehicle execution features, subset of load stability features, and derived features according to the time sequence dimension to output the fusion feature set of driving behavior risks.

4. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 1, characterized in that, In S3, the driving behavior risk reasoning and classification includes the following steps: S3.1 Match the current job process with the corresponding job micro-stage and generate a scenario-based driving response expectation range and vehicle execution expectation range; S3.2 Calculate the degree of mismatch in the driving response dimension and the vehicle execution dimension respectively, and assess the degree of transmission of operational abnormalities to operational risks; S3.

3. Generate inconsistent operation response states through time-series recursion to quantify the formation and persistence of risks; S3.4 Calculate the working load stability occupancy factor based on load and motion conditions, and correct the vehicle's safety correction capability boundary; S3.

5. Generate driving risk recovery margin and combine it with the inconsistent state of operation response, driving risk recovery margin and load constraints to complete the joint classification and output the driving behavior risk classification result.

5. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion according to claim 4, characterized in that, In step S3.1, matching the job micro-stage corresponding to the current job process and generating a scenario-based expected range includes the following steps: S3.

11. Extract vehicle spatial location, driving direction, operating mechanism status, load status and time sequence change features from the driving behavior risk fusion feature set, and obtain the belonging probability of each operation micro-stage at the current time through time sequence feature matching and probabilistic inference, and generate operation micro-stage probability vector. S3.

12. Using the probability of each stage in the operation micro-stage probability vector as the weight, the driving response expectation template and vehicle execution expectation template corresponding to each pre-configured operation micro-stage are weighted and fused to generate a driving response expectation interval and a vehicle execution expectation interval that match the current operation process.

6. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 5, characterized in that, In step S3.2, calculating the degree of mismatch in the two dimensions and assessing the degree of risk transmission includes the following steps: S3.

21. The driver's state and behavior features in the driving behavior risk fusion feature set are compared with the expected range of driving response feature by feature. For features that exceed the expected range, the standardized deviation is calculated. The deviation is then weighted and summed with the feature weights corresponding to the current operation stage to obtain the comprehensive driving response deviation. S3.

22. The on-board operation and work mechanism features in the driving behavior risk fusion feature set are compared with the vehicle execution expectation range feature by feature. For features that exceed the expectation range, the standardized deviation is calculated. The deviation is then weighted and summed with the feature weights corresponding to the current operation stage to obtain the comprehensive vehicle execution deviation. S3.

23. Within the preset time window, calculate the temporal transmission matching degree between driving response deviation and vehicle execution deviation through the deviation correlation matrix corresponding to the current operation stage, and generate the transmission correlation quantity of driving response deviation to vehicle execution result.

7. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 6, characterized in that, In S3.3, generating an inconsistent job response state specifically includes: The inconsistency state of the operation response at the previous moment is obtained. Combined with the risk memory coefficient corresponding to the current operation stage, the comprehensive driving response deviation, the comprehensive vehicle execution deviation, and the transmission correlation, the inconsistency state of the operation response at the current moment is calculated using a time-series recursive rule with memory decay. When the vehicle is in reversing positioning, load transfer, densely populated areas, or near the operation boundary, the value of the risk memory coefficient is increased accordingly to prolong the duration of the risk state's continuous memory, thereby quantifying the formation and persistence of the risk of mismatch between driving response and operation requirements.

8. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 7, characterized in that, In step S3.4, calculating the operational load stability occupancy factor and correcting the vehicle safety correction capability boundary includes the following steps: S3.41 Extract four types of normalized features from the driving behavior risk fusion feature set: load amount, load height, load lateral offset, and vehicle steering motion intensity. Obtain the working load stability occupancy factor through multi-factor coupling calculation to quantify the degree of occupancy of the vehicle's stability margin by the current load and motion conditions. S3.

42. Using the aforementioned work load stability occupancy factor as a correction coefficient, the basic correctable capabilities of the three correction channels—longitudinal deceleration, lateral steering, and work mechanism retraction—are attenuated and corrected to obtain the effective correctable capabilities of the vehicle corresponding to each channel under the current load condition.

9. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion according to claim 8, characterized in that, In S3.5, generating the driving risk recovery margin and completing the joint classification output includes the following steps: S3.

51. Based on the comprehensive vehicle execution deviation, calculate the necessary correction amounts for the three channels of longitudinal deceleration, lateral steering, and working mechanism reversal required to restore to a safe state, and calculate the normalized recovery margin for each correction channel in combination with the vehicle's effective correctability. S3.

52. Extract the response hysteresis, operation amplitude, and deviation descent rate features of the driver's corrective operation from the driving behavior risk fusion feature set. Use the minimum normalized recovery margin of each correction channel as the recursive base value. Adjust the decay step size of risk deterioration with response hysteresis and adjust the recovery coefficient of risk recovery with operation amplitude and deviation descent rate. Use the asymmetric recursive rule of rapid decline when risk deteriorates and slow rise when risk recovers to calculate the driving risk recovery margin at the current moment. S3.

53. The abnormal threshold of the inconsistent operation response is used as the basis for determining the establishment of risk, and the driving risk recovery margin is lower than the safety threshold as the basis for determining the intervention trigger. Combined with the operation load stability occupancy factor to limit the range of permissible intervention methods, the risk level is jointly classified and the intervention urgency is calculated. The result of driving behavior risk classification, which includes risk level, intervention urgency and permissible intervention method constraints, is output.

10. The method for AI monitoring, early warning, and proactive intervention of port mobile machinery driving behavior based on multi-source information fusion as described in claim 1, characterized in that, In step S4, the dynamic matching and execution of the graded early warning strategy and the intervention command from the on-board equipment includes the following steps: S4.1 Analyze the driving behavior risk classification results and extract the corresponding risk level, intervention urgency and allowable intervention method constraints; S4.

2. Based on the risk level and the urgency of intervention, match the corresponding level of early warning strategy and determine the early warning triggering method and output sequence; S4.

3. Based on the constraints of permissible intervention methods, select on-board equipment intervention commands that are suitable for the current load conditions, and clarify the action channel and execution intensity of the intervention. S4.

4. The graded early warning strategy and the intervention instructions of the vehicle equipment are sent to the corresponding execution terminal for execution, so as to realize the graded control of driving behavior risks.

Images