Air-light truck identification and differential charging pricing system based on deep learning

Abstract

The invention relates to the technical field of intelligent transportation, particularly discloses an air-light truck identification and differentiated charging pricing system based on deep learning, and aims to solve the problems of low identification accuracy of air-light trucks and poor implementation effect of differentiated charging policies in the prior art. The system comprises a multi-source data fusion processing module, a driving stage adaptive segmentation module, a spatial-temporal feature dynamic weighted extraction module, a historical behavior mode memory and fusion module, a confidence integration decision module and a differentiated charging pricing execution module. Through adaptive segmentation of driving stages, dynamic weighting of key features and fusion of historical behavior modes and confidence decision, accurate identification of empty and light load states of a truck is realized, differential charging is automatically executed based on an identification result, and the identification accuracy and the charging policy implementation effect are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, specifically relating to a deep learning-based system for identifying empty and lightly loaded freight vehicles and implementing differentiated pricing. Background Technology

[0002] In the fields of intelligent transportation and highway operation and management, improving road network efficiency and economic benefits through technological means has become an important development direction. Among them, implementing precise differentiated toll policies based on vehicle operating status is a key measure to achieve optimal resource allocation and guide traffic flow. As the main users of highways, the accurate identification of the empty and lightly loaded status of trucks constitutes the technological foundation for the implementation of such policies.

[0003] Among these, truck load status recognition based on deep learning is a current research hotspot. Its goal is to automatically determine the actual load status of trucks by analyzing vehicle trajectory data, providing a basis for toll collection system decisions. This technological direction aims to overcome the limitations of traditional manual interpretation or simple rule-based methods, achieving more efficient and automated status recognition.

[0004] Existing technologies mainly consist of two representative solutions, both of which face significant challenges. Rule-based static recognition methods rely on fixed thresholds for judgment, failing to effectively handle misclassification during stationary loading and unloading phases. They also exhibit poor adaptability to trajectory deviations caused by GPS drift, resulting in low accuracy in real-world testing. More seriously, these methods completely ignore historical vehicle behavior patterns, leading to the continuous misclassification of vehicles with specific operational patterns. While traditional deep learning models can process sequential data, their architecture assigns equal weight to noisy segments in the trajectory data. When data is continuously missing or interfered with, the model's confidence level drops sharply. Its fundamental flaw lies in its failure to distinguish the physical differences between different driving phases (such as acceleration and stationary phases), diluting the weights of key features. Furthermore, the model relies solely on the current trajectory segment for judgment, severing the vehicle's inherent historical loading behavior patterns and failing to effectively compensate for poor data quality.

[0005] The aforementioned technical defects directly result in a persistently low accuracy rate for identifying empty freight trucks in complex real-world road network environments. As a result, a large number of empty vehicles that should have enjoyed toll discounts have failed to benefit due to system misjudgments, which seriously weakens the effectiveness and attractiveness of the differentiated toll policy and constitutes a technical problem that urgently needs to be solved. Summary of the Invention

[0006] This invention aims to provide a deep learning-based system for identifying empty and lightly loaded freight vehicles and implementing differentiated pricing, in order to solve the technical problems in the prior art, such as the inability to effectively distinguish the physical meaning of vehicle driving stages, the fragmentation of historical behavior patterns, and the difficulty in dealing with noise interference in trajectory data, resulting in low accuracy in identifying empty and lightly loaded freight vehicles and poor implementation of differentiated pricing policies.

[0007] The technical solution of the present invention consists of a multi-source data fusion processing module, a driving phase adaptive segmentation module, a spatiotemporal feature dynamic weighted extraction module, a historical behavior pattern memory and fusion module, a confidence integration decision module, and a differentiated pricing execution module.

[0008] The multi-source data fusion processing module receives and preprocesses raw data from roadside sensing units and vehicle-mounted terminals, outputting standardized continuous trajectory data. This module specifically performs data alignment, outlier removal, and standardization operations. The data alignment process synchronizes GPS positioning data, vehicle axle load sensor data, vehicle contour scan data, and gantry transaction record data spatiotemporally based on a unified timestamp and vehicle identification identifier. Outlier removal employs an outlier detection algorithm based on statistical distribution, filtering data points that exceed a preset reasonable physical range. Standardization converts sensor readings of different dimensions and magnitudes to a unified numerical range, providing standardized input for subsequent analysis.

[0009] The adaptive segmentation module for driving phases is connected to the multi-source data fusion processing module to divide the preprocessed continuous trajectory data into phases. This module first identifies stationary, constant-speed, accelerating, and decelerating states by analyzing vehicle velocity and acceleration sequences. Furthermore, it introduces trajectory curvature-based analysis to distinguish between straight-line and curved driving phases. The segmentation process does not use a fixed time window but rather adaptively segments based on significant changes in vehicle kinematic characteristics, ensuring consistent physical properties within each segmented driving phase.

[0010] The spatiotemporal feature dynamic weighting extraction module is connected to the driving stage adaptive segmentation module. It is responsible for extracting feature vectors representing the vehicle's load state from each segmented driving stage and applying dynamic weights to the features extracted from different stages. Feature extraction covers temporal features, frequency domain features, and spatial geometric features. Temporal features include the average velocity, velocity variance, maximum acceleration, average acceleration, and statistical distribution moments of acceleration within the stage. Frequency domain features are extracted by performing a Fast Fourier Transform on the acceleration sequence of the stage to extract the main frequency components and their energy distribution. Spatial geometric features include the rate of curvature change and heading angle stability of the trajectory within the stage. The core of the dynamic weighting strategy is to assign different feature weight coefficients based on the type of driving stage. Specifically, for the acceleration stage, acceleration-related features are given higher weights; for the uniform straight-line driving stage, velocity stability and trajectory straightness features are given higher weights; for the stationary stage, the feature weights are significantly reduced or even ignored to effectively avoid interference from the static loading / unloading phase on the overall judgment.

[0011] The historical behavior pattern memory and fusion module, independent of current trip analysis, is used to construct and update the historical load behavior profile of a specific truck. This module maintains a database indexed by vehicle identifiers. For each truck, the module records the frequency of being classified as empty, lightly loaded, or heavily loaded in its historical trips, typical occurrence times, frequently traveled routes, and typical load states on each route. This module uses a sliding time window model to update the historical behavior profile, giving higher update weight to recent behavior data. In the analysis of the current trip, this module performs similarity calculations and pattern matching between the currently extracted spatiotemporal features and the historical behavior profile, outputting a prior probability of load state based on historical behavior patterns.

[0012] The confidence integration decision module receives the current trip feature analysis results from the spatiotemporal feature dynamic weighted extraction module and the historical behavior prior probabilities from the historical behavior pattern memory and fusion module. The core of this module is a deep neural network classifier, whose input is a concatenation of the weighted current spatiotemporal feature vector and the historical behavior prior probabilities. This neural network adopts a multilayer perceptron architecture, including an input layer, multiple hidden layers, and an output layer. The output layer generates the posterior probability distribution of the current vehicle belonging to three categories: empty, lightly loaded, and heavily loaded, using the Softmax function. Further, this module calculates the confidence score for this classification, which is determined by the entropy of the posterior probability distribution and the distribution characteristics of the activation values ​​of the neural network's hidden layers. When the confidence score is higher than a preset high confidence threshold, the classification result is directly output. When the confidence score is lower than the high confidence threshold but higher than the low confidence threshold, a decision delay mechanism is triggered, waiting for data from subsequent driving stages for re-evaluation and decision fusion. When the confidence score is lower than the low confidence threshold, the most probable load state output by the historical behavior pattern memory and fusion module is adopted as the final decision.

[0013] The differentiated pricing execution module connects to the confidence-based integrated decision-making module to implement corresponding toll strategies based on the final determined load status. This module pre-stores rate tables for empty, lightly loaded, and heavily loaded states. Once it receives the load status determination result, it calculates the toll for the current trip based on the vehicle's travel route, mileage information, and the corresponding rate. The toll calculation logic is automatically executed via smart contracts deployed on roadside units or provincial toll centers. The billing result is updated in real-time to the vehicle's user account or associated payment system, generating a transaction voucher. This module also has a data feedback interface, encrypting the identification results, the data features used, and the final toll record before sending them back to the historical behavior pattern memory and fusion module to update the vehicle's behavioral profile.

[0014] In a preferred embodiment of the present invention, the determination of a stationary state in the adaptive segmentation module during the driving phase must simultaneously meet the conditions of a speed of less than 0.5 km / h and a duration of more than 10 seconds. Furthermore, by combining vehicle contour scan data, brief stops with people in the cab are excluded, thereby accurately distinguishing between loading and unloading operations and normal traffic congestion or temporary stops.

[0015] Furthermore, the dynamic weighted extraction module for spatiotemporal features incorporates centrifugal force effects in its dynamic weight allocation strategy for the curve driving phase. This module estimates lateral acceleration using trajectory curvature and vehicle speed, and compares this estimate with the theoretical anti-roll stability threshold of the vehicle under the current load condition. If the estimated value is close to or exceeds the threshold, the feature weights for that curve phase are adaptively attenuated to reduce the impact of noise introduced by aggressive driving behavior on load assessment.

[0016] Furthermore, the historical behavior pattern memory and fusion module employs an attention-based recurrent neural network to model the vehicle's historical sequence behavior. This network can learn the differences in the importance of different historical journeys to the prediction of the current state, automatically assigning higher attention weights to more relevant and recent historical journeys, thereby more accurately capturing the evolution trend of the vehicle's load-bearing behavior.

[0017] Furthermore, the specific operation process of the decision delay mechanism in the confidence integration decision module is as follows: When a delayed decision is triggered, the system does not generate a final load status determination, but continues to collect data from one or more subsequent driving stages of the vehicle. This newly collected data, after adaptive segmentation of driving stages and dynamic weighted extraction of spatiotemporal features, is temporally concatenated and fused with the previously temporarily stored data features. The fused feature sequence is then input again into a deep neural network classifier for calculation, generating a new posterior probability distribution and confidence score. This process can be iterative until the confidence score meets the high confidence threshold requirement, or the vehicle's journey ends. If the confidence score is still insufficient at the end of the journey, a fallback to a decision based on historical behavior patterns is implemented.

[0018] Furthermore, the differentiated toll pricing execution module integrates a cost-effectiveness evaluation submodule. This submodule periodically analyzes toll collection for vehicles with different load conditions, changes in road network traffic distribution, and system recognition accuracy. Based on these statistical analysis results, the submodule can generate toll adjustment suggestions, aiming to further optimize the distribution of freight vehicle load structure on the road network through price levers and improve overall operational efficiency. The generation of toll adjustment suggestions follows a preset optimization objective function, which comprehensively considers toll revenue, road wear costs, traffic congestion index, and environmental indicators.

[0019] Furthermore, the present invention also provides an empty vehicle route optimization planning module, which is connected to the confidence integration decision module and the differentiated toll pricing execution module. Based on the empty status identification results, combined with real-time traffic information, road toll policies, road conditions and historical behavior patterns, it generates differentiated route planning suggestions for empty trucks and optimizes the selection of national highways or expressways for driving.

[0020] The empty truck route optimization and planning module uses a multi-objective optimization algorithm to comprehensively consider factors such as toll costs, travel time, road wear, environmental indicators, and user preferences, outputs the optimal route plan, and pushes real-time suggestions to drivers through vehicle terminals or mobile applications, thereby further improving the operational efficiency and economic benefits of empty trucks.

[0021] In summary, this application includes at least one of the following beneficial technical effects: 1. By introducing an adaptive segmentation mechanism for driving stages and a dynamic weighted extraction mechanism for spatiotemporal features, this invention fundamentally distinguishes the physical meaning of different driving stages and applies differentiated weights to the features. This effectively amplifies the contribution of features strongly related to load in key stages such as acceleration, while significantly suppressing interference from non-key stages such as stationary and noisy stages. It solves the core defect of traditional deep learning methods where key features are diluted due to equal weights, thereby greatly improving the discriminative ability of feature representation.

[0022] 2. By constructing an independent historical behavior pattern memory and fusion module, this invention combines the instantaneous judgment of the current journey with the long-term historical behavior patterns of the vehicle. When the quality of the current trajectory data is poor or there is interference, it can effectively compensate and correct deviations by utilizing prior historical behavior. This overcomes the limitations of existing technologies that sever historical connections and rely solely on single journey segments for judgment, and enhances the robustness and judgment continuity of the system in complex real-world scenarios.

[0023] 3. By designing an integrated decision-making mechanism and decision delay strategy based on confidence scores, this invention achieves dynamic quality control of the identification process. The system can autonomously assess the reliability of a single judgment, choosing to wait for more evidence rather than forcibly outputting a potentially erroneous result when confidence is insufficient. This prudent decision-making mechanism effectively avoids misjudgments caused by a sharp drop in confidence under conditions of missing data or continuous interference, significantly improving the accuracy and reliability of the final output. This provides a solid technical guarantee for empty and light-loaded freight vehicles to enjoy accurate toll discounts, ensuring the precise implementation of differentiated toll policies and the effective guidance of these policies. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the fusion of confidence-integrated decision-making and historical behavior patterns in this invention. Detailed Implementation

[0025] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on specific implementation methods of the present invention.

[0026] Example 1 Please refer to the attached document. Figure 1 This embodiment details a deep learning-based system for identifying empty and lightly loaded freight vehicles and implementing differentiated pricing, including the following steps: Step S1: The multi-source data fusion processing module receives raw data streams from the roadside sensing unit and the vehicle terminal, and performs data alignment, outlier removal and standardization operations to generate continuous trajectory data.

[0027] The roadside sensing unit is deployed at highway toll stations, gantries, and key road sections, and includes a GPS signal receiver, a dynamic weighing instrument, and a 3D LiDAR scanning device; the vehicle-mounted terminal is integrated inside the truck and continuously collects vehicle identification information, engine operating parameters, and auxiliary positioning information.

[0028] First, a data alignment operation is performed. Based on nanosecond-level precision timestamps and globally unique vehicle identification identifiers, GPS latitude and longitude coordinates from different physical locations, vehicle axle load measurements, vehicle outer contour point cloud data, and gantry electronic toll transaction records are strictly synchronized in time and space. The data alignment process uses a sliding window matching algorithm to ensure that all data attributes at the same time belong to the same truck.

[0029] Subsequently, outlier removal is performed using an outlier detection algorithm based on the three-standard-deviation criterion. This algorithm automatically identifies and filters data points that exceed a preset reasonable physical range, such as GPS coordinate jumps, instantaneous axle load exceeding limits, and sudden changes in profile dimensions. The preset reasonable physical range is pre-defined based on the maximum design gross vehicle weight of trucks, national standards for vehicle outline dimensions, and road alignment design specifications.

[0030] Finally, a standardization process is performed to convert sensor readings of different dimensions and magnitudes, such as the kilonewton values ​​output by the axle load sensor, the latitude and longitude offsets output by the GPS, and the point cloud density values ​​output by the contour scan, into a unified numerical range of zero to one using the max-min normalization method, thus providing a standardized input data matrix for subsequent analysis modules.

[0031] Step S2: The adaptive segmentation module for driving phases is connected to the multi-source data fusion processing module via a high-speed data bus to perform physically meaningful segmentation of the preprocessed continuous trajectory data.

[0032] The adaptive segmentation module for the driving phase incorporates a multi-state Kalman filter to smooth and estimate the vehicle's speed and acceleration sequences. Based on the processed kinematic sequence, the adaptive segmentation module for the driving phase first identifies the stationary state. The determination condition is that the instantaneous speed of the vehicle is below 0.5 km / h and the duration of this low-speed state is more than 10 seconds.

[0033] To further improve the accuracy of static state determination, the adaptive segmentation module for the driving phase calls vehicle contour scan data. By analyzing the point cloud clustering features and heat map distribution of the cab area, it eliminates brief stops caused by the driver's activities inside the vehicle, thereby accurately distinguishing between the loading and unloading operation phase and the normal traffic congestion or temporary parking phase.

[0034] By analyzing the point cloud clustering characteristics and heatmap distribution of the cab area, the following methods can be used: First, 3D point cloud data of the cab area is extracted based on vehicle contour scan data. Second, density-based clustering analysis (such as the DBSCAN algorithm) is performed on this point cloud data to identify clusters that may represent the driver. Next, a heat map of the cab area is generated to reflect the point cloud density distribution. Finally, based on the number, size, and positional changes of the clusters, as well as the center offset and dispersion of the heat map, combined with preset thresholds, it is determined whether there is continuous human activity in the cab. If so, it is judged as a temporary stop; otherwise, it is judged as loading and unloading operations. For example, if a cluster of more than X points is detected in the cab area and its spatial position is stable in region Y, it is judged that there is human activity in the cab, ruling out the possibility of loading and unloading operations.

[0035] For motion states, the adaptive segmentation module for driving phases identifies uniform driving state, accelerating driving state, and decelerating driving state by analyzing the first-order and second-order differences of the acceleration sequence.

[0036] The condition for determining a constant speed driving state is that the absolute value of the acceleration is continuously below 0.3 m / s²; the condition for determining an accelerating driving state is that the acceleration is continuously above 0.5 m / s²; and the condition for determining a decelerating driving state is that the acceleration is continuously below -0.5 m / s².

[0037] Furthermore, the adaptive segmentation module for driving phases introduces a geometric analysis algorithm based on trajectory curvature, which distinguishes between straight-line driving phases and curve driving phases by calculating the rate of change of path angles formed by continuous GPS coordinate points.

[0038] The criteria for determining straight-line driving is an average curvature of less than 0.01 per meter. The criteria for determining curve driving is an average curvature of more than 0.05 per meter.

[0039] The segmentation process abandons the fixed time window strategy and instead adopts an adaptive segmentation algorithm based on the significant changes in vehicle kinematic characteristics. It calculates the variance of the sliding window of speed and acceleration in real time. When the variance value exceeds the adaptive threshold, it triggers the segmentation of the stage boundary, ensuring that each segmented driving stage has highly consistent dynamic and geometric characteristics.

[0040] In step S3, the spatiotemporal feature dynamic weighted extraction module is connected to the driving stage adaptive segmentation module through a memory sharing interface. It is responsible for extracting high-dimensional feature vectors representing the vehicle load status from each segmented driving stage and applying dynamic weight coefficients to the features according to the stage type.

[0041] Feature extraction operations cover three dimensions: time-domain features, frequency-domain features, and spatial geometric features.

[0042] The time-domain characteristics include the average velocity value, velocity sequence variance, maximum acceleration, average acceleration, acceleration skewness, kurtosis, and other statistical distribution moments within this stage.

[0043] Frequency domain features were obtained by performing a 256-point Fast Fourier Transform on the acceleration sequence of this stage, extracting the amplitude proportion and energy concentration index of the top ten main frequency components.

[0044] Spatial geometric features include the standard deviation of the rate of change of curvature of the trajectory during this stage and the variance of the heading angle sequence.

[0045] The core of the dynamic weighting strategy lies in assigning different feature weight coefficients based on the physical characteristics of each driving stage. Specifically, for the acceleration stage, acceleration-related features such as maximum and average acceleration are given a higher weight of 0.8, while the weight of speed-related features is reduced to 0.3; for the uniform straight-line driving stage, speed stability features such as speed variance and trajectory straightness features such as the rate of change of curvature are given a higher weight of 0.7; for the stationary stage, the feature weights are uniformly set to 0.01 to minimize the interference of the stationary loading / unloading section on the overall load assessment; for the curve driving stage, the dynamic weight allocation strategy additionally considers the centrifugal force effect.

[0046] The spatiotemporal feature dynamic weighted extraction module estimates the lateral acceleration value by using the trajectory curvature radius and the vehicle's instantaneous speed. This estimated value is then compared with the theoretical anti-roll stability threshold of the vehicle under the current load condition. The theoretical anti-roll stability threshold is pre-calculated based on the vehicle's wheelbase, center of gravity height, and suspension stiffness parameters. If the estimated lateral acceleration value exceeds 80% of the threshold, all feature weights for that curve phase are attenuated by 50% to reduce the negative impact of dynamic noise introduced by aggressive driving behavior on load judgment.

[0047] Step S4: The historical behavior pattern memory and fusion module is independently deployed on a cloud server cluster to build and update the long-term historical load behavior profile of a specific truck.

[0048] The historical behavior pattern memory and fusion module maintains a distributed time-series database with vehicle identification identifiers as the primary key. The database table structure includes fields such as vehicle identification identifier, timestamp, trip identifier, load status label, feature vector, geographic location coordinates, and time period identifier.

[0049] For each truck, the historical behavior pattern memory and fusion module calculates the following statistical indicators by aggregating its historical trip data: The percentage of trips that were determined to be unloaded, lightly loaded, and heavily loaded in the past 30 days. Typical load distribution for four time periods each day: 0:00-6:00, 6:00-12:00, 12:00-18:00, and 18:00-24:00; A set of highway numbers for frequently traveled routes and the frequency of empty, lightly loaded, and heavily loaded conditions on each route.

[0050] The historical behavior pattern memory and fusion module uses a sliding time window model to update historical behavior profiles. The window size is set to 30 days. The behavior data of the most recent 7 days is given an update weight of 0.6, the data from day 8 to day 15 has a weight of 0.3, and the data from day 16 to day 30 has a weight of 0.1.

[0051] In the analysis of the current journey, the historical behavior pattern memory and fusion module performs cosine similarity calculation and dynamic time warping matching on the currently extracted spatiotemporal feature vector and the feature templates of the same route and time period in the historical behavior profile, and outputs a load status prior probability vector based on historical behavior patterns. The load status prior probability vector includes three components: empty load prior probability, light load prior probability, and heavy load prior probability.

[0052] In practice, the distinction between "light load" and "heavy load" can be made by setting a threshold M for the vehicle's total mass (or axle load). For example, if the identified vehicle load is below M, it is considered light load; if it is above or equal to M, it is considered heavy load. The threshold M can be dynamically adjusted or preset based on different vehicle models, operational needs, vehicle design parameters, and empirical data.

[0053] In step S5, the confidence integration decision module receives the current journey weighted feature vector from the spatiotemporal feature dynamic weighted extraction module and the historical behavior prior probability vector from the historical behavior pattern memory and fusion module via gigabit Ethernet.

[0054] The core of the confidence ensemble decision module is a deep neural network classifier with 256 input layer nodes, corresponding to the total dimension obtained by concatenating the 200 dimensions of the weighted spatiotemporal feature vector with the three dimensions of the historical behavior prior probability vector.

[0055] The hidden layers adopt a three-layer fully connected architecture. The first hidden layer contains 128 neurons and uses the modified linear unit activation function. The second hidden layer contains 64 neurons and also uses the modified linear unit activation function. The third hidden layer contains 32 neurons and uses the hyperbolic tangent activation function.

[0056] The output layer contains three neurons, corresponding to the three categories of unloaded, lightly loaded, and heavily loaded, respectively. The Softmax function is used to generate the posterior probability distribution of the current vehicle's category. The calculation process of this posterior probability distribution is expressed as the mathematical formula: Where vector z represents the original score of the output layer, and c represents the category index.

[0057] The confidence score ensemble decision module further calculates the confidence score for this classification. The confidence score is determined by the entropy of the posterior probability distribution and the distribution characteristics of the activation values ​​of the hidden layer of the neural network. The entropy confidence component is obtained by calculating the Shannon entropy of the posterior probability distribution, and the hidden layer confidence component is obtained by calculating the variance of the activation values ​​of the neurons in the third hidden layer.

[0058] The final confidence score is the weighted sum of the two components, with weighting coefficients of 0.7 and 0.3 respectively. The system presets a high confidence threshold of 0.85 and a low confidence threshold of 0.6.

[0059] When the confidence score is higher than 0.85, the classification result is directly output as the final load status determination.

[0060] When the confidence score is below 0.85 but above 0.6, a decision delay mechanism is triggered. The system temporarily stores all current feature data and intermediate results and continues to collect data from one or more subsequent driving stages of the vehicle for re-evaluation.

[0061] When the confidence score is below 0.6, the load state corresponding to the maximum prior probability output by the historical behavior pattern memory and fusion module is adopted as the final decision.

[0062] The training process for a deep neural network classifier is as follows: First, a labeled dataset containing a large amount of historical driving data of trucks under different load conditions is constructed. Each data point includes raw sensor data, the corresponding driving stage division, spatiotemporal feature vector, prior probability of historical behavior, and an actual load condition label determined by manual verification or weighbridge weighing. Second, during the training phase, cross-entropy is used as the loss function, and the Adam optimizer is selected for epoch-wise updates of model parameters. During training, mini-batch gradient descent is used, with a batch size set to, for example, 32, and an initial learning rate of 0.001. The learning rate is dynamically adjusted during loss plateaus using a learning rate scheduler. To prevent overfitting, a Dropout layer (e.g., setting a Dropout rate of 0.5 between hidden layers) and an early stopping mechanism are introduced. Training stops when the validation set loss no longer decreases for 10 consecutive epochs. Through the above training, the network can accurately predict the load condition based on input features and output classification results with reliable confidence scores.

[0063] Please refer to the attached document. Figure 2 The specific operational process of the decision-making delay mechanism is as follows: When a delayed decision is triggered, the system creates a temporary session context in memory, storing all feature vectors, prior probabilities, posterior probabilities, and confidence scores at the current moment.

[0064] The system continues to run the adaptive segmentation module for driving phases and the dynamic weighted extraction module for spatiotemporal features to process new data from the next 3 to 5 driving phases. These newly generated feature vectors are then spliced ​​with the temporarily stored feature vectors in a temporal sequence to form an extended feature sequence.

[0065] The extended feature sequence is input again into the deep neural network classifier for computation. The network structure remains unchanged, but the input dimensions are dynamically adjusted. The new computation produces an updated posterior probability distribution and confidence score. If the updated confidence score meets the high confidence threshold requirement of 0.85, the classification result is output and the temporary session is cleared. If the confidence score is still between 0.6 and 0.85, the system continues to wait for data from subsequent stages, iterating a maximum of 5 times. If the confidence score is still less than 0.85 after 5 iterations or the vehicle journey ends, the system reverts to a decision-making strategy based on historical behavior patterns, selecting the load state with the highest historical prior probability as the output.

[0066] In step S6, the differentiated pricing execution module connects to the confidence integration decision module through a secure application programming interface to execute the corresponding charging strategy based on the final determined load status.

[0067] The differentiated pricing execution module has three pre-stored rate tables, corresponding to empty, lightly loaded, and heavily loaded states respectively. The rate table structure includes fields such as highway number, starting station number of mileage segment, ending station number of mileage segment, base rate, and load factor.

[0068] The load factor for the no-load rate table is 0.8, the load factor for the light-load rate table is 0.9, and the load factor for the heavy-load rate table is 1.2.

[0069] Once the load status determination result is received, the differentiated toll pricing execution module calculates the toll for this trip based on the highway number, mileage information, and corresponding load coefficient of the actual route traveled by the vehicle. The toll calculation logic is automatically executed through a smart contract deployed in the provincial toll center. The smart contract uses blockchain technology to ensure the immutability of the calculation process. The billing result is updated in real time to the user account bound to the vehicle or the associated electronic payment system, and a digital transaction voucher containing elements such as transaction time, route, mileage, load status, and toll amount is generated.

[0070] The differentiated pricing execution module also has a two-way data feedback interface. After asymmetric encryption, the final result of this identification, all feature vectors used, decision confidence scores, and the final charging record are securely transmitted back to the database server of the historical behavior pattern memory and fusion module to update the long-term behavior profile of the vehicle.

[0071] The system in this embodiment also integrates a cost-effectiveness evaluation submodule, which is embedded in the differentiated pricing execution module.

[0072] The cost-effectiveness evaluation submodule regularly analyzes the total toll collection of vehicles under different load conditions, the traffic flow distribution curves of each road segment, and the system identification accuracy index on a weekly basis.

[0073] The accuracy rate index is calculated by comparing manual sampling audits with weighbridge weighing data. Based on these multi-dimensional statistical analysis results, the cost-effectiveness evaluation submodule uses a linear programming algorithm to generate rate adjustment suggestions, aiming to further optimize the distribution of freight vehicle load structure on the road network through price leverage and improve overall operational efficiency.

[0074] The generation of toll rate adjustment recommendations follows a preset multi-objective optimization function. This function simultaneously maximizes toll revenue, minimizes road wear costs, controls traffic congestion index within the second-level service level, and ensures that exhaust emissions meet national standards. The specific form of the optimization function is a weighted summation model, and the weight coefficients of the four objectives are dynamically adjusted according to the road network management strategy.

[0075] Step S7: The empty vehicle route optimization planning module is connected to the confidence integration decision module and the differentiated pricing execution module through the data interface, and is used to perform route optimization planning based on the empty vehicle status identification results.

[0076] The empty vehicle route optimization planning module has a built-in multi-objective optimization engine, and the input parameters include: Real-time traffic data, including real-time congestion index, average vehicle speed, and accident information for highways and national roads, obtained from traffic management departments; Road toll data, obtained from the differentiated toll pricing execution module, includes highway toll rates under empty load conditions, and national highway toll policies (if applicable). Road condition data, including road slope, curve radius, road surface type, height and weight restrictions; Historical behavior patterns: The vehicle's frequently traveled routes and typical time period preferences are obtained from the historical behavior pattern memory and fusion module. User preference settings, such as shortest time, lowest cost, optimal comfort, etc.

[0077] The multi-objective optimization engine uses a weighted summation model, with the objective function as follows: Among them, the weighting coefficient , , , It can be dynamically adjusted based on user preferences or system strategies. Costs include tolls and fuel costs; time is based on real-time traffic flow prediction; road wear is estimated based on vehicle axle load and road type; environmental score is calculated based on exhaust emission models.

[0078] The route planning output includes: Recommended routes: highways, national roads, or a combination of both; Expected costs: Total toll and fuel costs; Expected time: travel time and possible delays; Environmental assessment: Carbon emission level.

[0079] Recommendations are pushed to drivers via in-vehicle terminals or smartphone apps through a graphical interface, and real-time route adjustment is provided.

[0080] Furthermore, the empty vehicle route optimization and planning module also has learning capabilities. By recording the driver's actual route selection and feedback, it dynamically updates the weight coefficients and personalizes the recommendation strategy.

[0081] Example 2 This embodiment provides an alternative to historical behavior modeling based on attention mechanisms, with a focus on a detailed explanation of the historical behavior pattern memory and fusion module.

[0082] In this embodiment, the historical behavior pattern memory and fusion module uses an attention-based recurrent neural network to dynamically model the historical sequence behavior of the vehicle. The network input consists of the load status sequence, feature vector sequence, and spatiotemporal context sequence of all trips made by the vehicle in the past 30 days.

[0083] The load status sequence consists of one-hot encoded labels for empty, lightly loaded, and heavily loaded. The feature vector sequence consists of a 200-dimensional feature vector output by the spatiotemporal feature dynamic weighted extraction module for each trip. The spatiotemporal context sequence includes the trip start time, main route number, and weather condition code.

[0084] The recurrent neural network (RNN) employs a two-layer long short-term memory (LSTM) network structure, with each layer containing 64 hidden units. The LSM network outputs the hidden state of the last time step as a preliminary representation of the entire historical sequence. The attention mechanism layer receives this hidden state and the real-time feature vector of the current process as input. It calculates the correlation score between the current state and the state of each historical time step using a trainable parameter matrix. After the correlation score is normalized by the Softmax function, it forms an attention weight distribution. The attention weight distribution automatically assigns higher attention weights to more relevant and recent historical processes. The weighted sum of the historical representations is concatenated with the current feature vector, and finally, a fully connected layer outputs the prior probability vector of historical behavior.

[0085] Attention-based models can more accurately capture the long-term dependencies and evolution trends of vehicle load behavior, especially when vehicle operation modes undergo gradual changes, and can adapt to new behavioral patterns more quickly than the statistical model in Example 1.

[0086] Regarding the spatiotemporal feature dynamic weighted extraction module, this embodiment introduces a feature enhancement method based on empirical mode decomposition.

[0087] For acceleration sequences, empirical mode decomposition is performed before fast Fourier transform. The original signal is adaptively decomposed into multiple intrinsic mode function components. The first three intrinsic mode function components containing the main energy information are selected for Hilbert transform, and the instantaneous frequency and instantaneous amplitude are calculated as supplementary frequency domain features.

[0088] These frequency domain features can more precisely characterize the nonlinear vibration characteristics of the suspension and power systems of vehicles under different load conditions, thereby improving the distinguishability of feature representation.

[0089] Regarding the confidence-integrated decision module, the deep neural network classifier in this embodiment uses a hybrid architecture combining a one-dimensional convolutional neural network and a long short-term memory network to process temporal features.

[0090] The input feature sequence is first processed through two one-dimensional convolutional layers for local feature extraction, with kernel sizes of 5 and 3, and filter numbers of 32 and 16, respectively. The output of the convolutional layers is then max-pooled and fed into a bidirectional long short-term memory network containing 50 units to capture the forward and backward dependencies of the feature sequence. The output of the long short-term memory network at the last time step is concatenated with the prior probability vector of historical behavior, and finally the posterior probability distribution is generated through a fully connected layer and the Softmax function.

[0091] This hybrid architecture is particularly well-suited for handling time-dependent features generated during vehicle operation, and can make fuller use of dynamic correlation information between multiple consecutive driving stages.

[0092] In this embodiment, the differentiated pricing execution module enhances the security audit function. All rate table update operations, fee calculation logs, and transaction records are stored through the blockchain network. The stored data includes timestamps, operator digital signatures, and transaction hash values.

[0093] The audit subsystem can trace and verify any historical transaction throughout the entire process, ensuring the openness, transparency and tamper-proof nature of the charging process. At the same time, the differentiated pricing execution module has added a user objection handling interface. When vehicle owners have objections to the determination of load status or the calculation of fees, they can submit a review application through this interface.

[0094] After receiving the application, the system automatically retrieves the complete data chain of the trip, including raw sensor data, intermediate results of each module, decision confidence scores, etc., and generates a highly readable review report for manual review.

[0095] In this embodiment, the cost-effectiveness evaluation submodule introduces a reinforcement learning algorithm for dynamic toll rate optimization. This algorithm abstracts the road network as a Markov decision process. The state space includes truck traffic volume, average speed, toll revenue, and road wear index for different load states on each road segment. The action space represents the fine-tuning range of the toll rate for each load state. The reward function integrates revenue, efficiency, and environmental objectives. By interacting with the environment, the reinforcement learning algorithm continuously learns the optimal toll rate strategy, achieving adaptive optimization of toll pricing. This more accurately guides the distribution of truck loads and improves the overall operational efficiency of the road network.

[0096] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0097] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A deep learning-based empty truck identification and differentiated toll pricing system, characterized in that, Comprise: A multi-source data fusion processing module receives and pre-processes raw data from roadside perception units and vehicle-mounted terminals, including data alignment, outlier rejection, and standardization operations, and outputs continuous trajectory data; A driving phase adaptive segmentation module connected to the multi-source data fusion processing module performs phase division on the pre-processed continuous trajectory data, including the identification of static state, uniform speed driving state, acceleration driving state, deceleration driving state, straight line driving phase, and curved driving phase; A spatiotemporal feature dynamic weighting extraction module connected to the driving phase adaptive segmentation module extracts feature vectors representing the vehicle load state from each segmented driving phase and applies dynamic weights to the features extracted in different phases; A historical behavior pattern memory and fusion module constructs and updates the historical load behavior profile of a specific truck and outputs the load state prior probability; A confidence integrated decision module receives current trip feature analysis results from the spatiotemporal feature dynamic weighting extraction module and historical behavior prior probability from the historical behavior pattern memory and fusion module, generates posterior probability distribution of the current vehicle belonging to empty, light, and heavy loads through a deep neural network classifier, and triggers a decision delay mechanism or adopts the load state output by the historical behavior pattern memory and fusion module as the final decision based on the confidence score; A differentiated toll pricing execution module connected to the confidence integrated decision module executes the corresponding toll strategy according to the final determined load state.

2. The deep learning-based empty truck identification and differentiated toll pricing system of claim 1, wherein, The data alignment process performed by the multi-source data fusion processing module synchronizes GPS positioning data, vehicle axle load sensor data, vehicle contour scanning data, and gantry transaction record data in time and space based on a unified timestamp and vehicle identifier; Outlier rejection uses a statistical distribution-based outlier detection algorithm; Standardization operations convert sensor readings of different dimensions and orders of magnitude to a unified numerical interval.

3. The deep learning-based empty truck identification and differentiated toll pricing system of claim 1, wherein The segmentation process of the driving phase adaptive segmentation module is based on the significant change points of vehicle kinematic characteristics for adaptive segmentation; For the determination of the static state, the conditions of speed less than 0.5 km / h and duration more than 10 seconds must be met, and the vehicle contour scanning data is combined to exclude short-term parking in the presence of a driver.

4. The deep learning-based empty truck identification and differentiated toll pricing system of claim 1, wherein, The features extracted by the spatiotemporal feature dynamic weighting extraction module include time domain features, frequency domain features, and spatial geometric features; The dynamic weighting strategy assigns different feature weight coefficients to different driving phases, with higher weights for acceleration-related features in the acceleration driving phase, higher weights for speed stability and trajectory flatness features in the uniform straight line driving phase, and significantly reduced weights for the static phase.

5. The deep learning-based empty truck identification and differentiated toll pricing system of claim 4, wherein, For the curved driving phase, the dynamic weight allocation strategy additionally considers the centrifugal effect by estimating the lateral acceleration from the trajectory curvature and vehicle speed, and comparing it with the theoretical anti-roll stability threshold of the vehicle under the current load state. If the estimated value approaches or exceeds the threshold, the feature weight of the curved phase is adaptively attenuated.

6. The deep learning-based empty truck identification and differentiated toll pricing system of claim 1, wherein The historical behavior pattern memory and fusion module maintains a database indexed by vehicle identity, recording the frequency ratio of empty, light or heavy load in historical trips, typical time period, common travel routes and typical load status on each route; The historical behavior portrait is updated using a sliding time window model, with recent behavior data being given higher update weight.

7. The deep learning-based empty truck identification and differentiated toll pricing system of claim 6, wherein, The historical behavior pattern memory and fusion module uses an attention mechanism-based recurrent neural network to model the historical sequence behavior of the vehicle, including a long short-term memory network. The attention mechanism layer calculates the relevance score of the current state and each historical time step state, automatically assigning higher attention weight to more relevant and more recent historical trips.

8. The deep learning-based empty truck identification and differentiated toll pricing system of claim 1, wherein, The deep neural network classifier of the confidence integrated decision module uses a multi-layer perceptron architecture, with the input being the concatenation of the weighted current spatio-temporal feature vector and the historical behavior prior probability, and the output layer producing a posterior probability distribution through the Softmax function. The confidence score is determined based on the entropy value of the posterior probability distribution and the distribution characteristics of the neural network hidden layer activation values.

9. The deep learning-based empty truck identification and differentiated toll pricing system of claim 8, wherein, The decision delay mechanism of the confidence integrated decision module operates as follows: When the delay decision is triggered, the system does not generate the final load status determination immediately, but continues to collect data for one or more subsequent driving stages. After the newly collected data is extracted through driving stage adaptive segmentation and spatio-temporal feature dynamic weighting, it is spliced and fused with the previously stored data features in time sequence. The fused feature sequence is then input into the deep neural network classifier for calculation.

10. The deep learning-based empty truck identification and differentiated toll pricing system of claim 1, wherein, It also includes an empty vehicle route optimization planning module connected to the confidence integrated decision module and the differentiated charging pricing execution module. Based on the empty state recognition result, combined with real-time traffic information, road charging policy, road conditions and historical behavior patterns, the module generates differentiated route planning suggestions through a multi-objective optimization algorithm to provide empty trucks with optimized selection of national highways or expressways.

Images