Big data analysis method and system for route data of logistic vehicles

Abstract

The application discloses a kind of big data analysis method and system of route data of logistics vehicle, and the present application relates to the technical field of big data analysis, discrete trajectory point sequence is converted into continuous route data, for historical route data in logistics route database, high-dimensional feature vector is extracted from multiple different dimensions and big data analysis is carried out, the accurate identification of benchmark route mode under different distribution scenarios is realized, the accuracy of benchmark route mode and the reliability of data basis are effectively improved.Current route data is matched with corresponding benchmark route mode in time and space similarity, and external environmental data is associated using big data analysis mechanism, to determine the dynamic performance deviation value of current route data relative to benchmark route mode, mark current route data as to be optimized route data, input it into big data decision engine, and generate recommended route data based on real-time road network weight, realize the dynamic closed-loop optimization from anomaly detection to route recommendation.

Description

Technical Field

[0001] This invention relates to the technical field of big data analysis, and in particular to a big data analysis method and system for logistics vehicle route data. Background Technology

[0002] With the rapid development of the logistics industry, refined analysis and optimization of logistics vehicle routes have become key means to reduce transportation costs and improve delivery efficiency. While existing technologies can collect vehicle trajectory data and perform basic route planning, they still have the following significant shortcomings in practical applications, especially for complex and ever-changing logistics delivery scenarios: Existing technologies often directly use raw GPS data for simple map matching, lacking a preprocessing mechanism for deep timestamp alignment and coordinate correction of trajectory data. This makes it difficult to transform discrete trajectory point sequences into continuous, high-precision route data, resulting in low-quality logistics route databases. Furthermore, existing technologies often limit the analysis of historical route data to a single dimension, lacking the ability to extract high-dimensional feature vectors from multiple different dimensions, thus failing to accurately identify baseline route patterns in different delivery scenarios. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a big data analysis method and system for logistics vehicle route data.

[0004] This invention provides a big data analysis method for logistics vehicle route data, including: The system receives real-time trajectory data streams reported by logistics vehicles, performs time stamp alignment and coordinate correction preprocessing on the trajectory data streams to output preprocessed trajectory data, and performs map matching enhancement on the preprocessed trajectory data to transform discrete trajectory point sequences into continuous route data in order to build a logistics route database. Based on historical route data in the logistics route database, high-dimensional feature vectors are extracted from multiple different dimensions. Big data analysis is performed on the high-dimensional feature vectors to identify the baseline route patterns of logistics vehicles in different delivery scenarios. The system acquires real-time data on the current route of logistics vehicles and external environment data. It performs spatiotemporal similarity matching between the current route data and the corresponding baseline route pattern, and uses big data analysis mechanisms to correlate external environment data, thereby determining the dynamic performance deviation value of the current route data relative to the baseline route pattern. If the dynamic performance deviation value exceeds the preset dynamic threshold, the current route data is marked as route data to be optimized. The route data to be optimized is then input into the big data decision engine, and recommended route data is generated based on the real-time road network weights.

[0005] This invention provides a big data analysis system for logistics vehicle route data, which is applied to the aforementioned big data analysis method for logistics vehicle route data; the big data analysis system for logistics vehicle route data includes: The database module is used to access the trajectory data stream reported by logistics vehicles in real time, perform time stamp alignment and coordinate correction preprocessing on the trajectory data stream to output preprocessed trajectory data, perform map matching enhancement on the preprocessed trajectory data, and transform discrete trajectory point sequences into continuous route data to build a logistics route database. The big data analysis module is used to extract high-dimensional feature vectors from multiple different dimensions of historical route data in the logistics route database, perform big data analysis on the high-dimensional feature vectors, and identify the baseline route patterns of logistics vehicles in different delivery scenarios. The dynamic performance module is used to acquire real-time data on the current route of logistics vehicles and external environment data. The current route data is matched with the corresponding benchmark route pattern in a spatiotemporal similarity. The external environment data is correlated using a big data analysis mechanism to determine the dynamic performance deviation value of the current route data relative to the benchmark route pattern. The recommended route data module is used to mark the current route data as route data to be optimized if the dynamic performance deviation value exceeds the preset dynamic threshold. The route data to be optimized is then input into the big data decision engine, and recommended route data is generated based on the real-time road network weights.

[0006] Compared with the prior art, the beneficial effects of the present invention are: (1) By performing map matching enhancement on the preprocessed trajectory data, the discrete trajectory point sequence is transformed into continuous route data, and a high-quality logistics route database is constructed. On this basis, high-dimensional feature vectors are extracted from multiple different dimensions and big data analysis is performed on the historical route data in the logistics route database. This enables accurate identification of the baseline route pattern under different delivery scenarios. The logistics route database is introduced to control the high-dimensional feature vectors in multiple dimensions, which effectively improves the accuracy of the baseline route pattern and the reliability of the data foundation.

[0007] (2) By acquiring the current route data and external environment data of logistics vehicles in real time, the current route data is matched with the corresponding benchmark route pattern in terms of spatiotemporal similarity, and the external environment data is associated with the big data analysis mechanism to determine the dynamic performance deviation value of the current route data relative to the benchmark route pattern. If the dynamic performance deviation value exceeds the preset dynamic threshold, the current route data is marked as route data to be optimized and input into the big data decision engine. Recommended route data is generated based on the real-time road network weight, realizing dynamic closed-loop optimization from anomaly detection to route recommendation. It fully considers the comprehensive influence of the route data to be optimized, the dynamic performance deviation value and the real-time road network weight, and significantly improves the accuracy of the recommended route data. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the big data analysis method for logistics vehicle route data in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a big data analysis system for logistics vehicle route data in an embodiment of the present invention; Figure 3 This is a hardware diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0010] Please see Figures 1 to 3 A big data analysis method for logistics vehicle route data, applied to big data analysis scenarios; the big data analysis method for logistics vehicle route data includes: Step S11: Real-time access to the trajectory data stream reported by logistics vehicles, preprocessing the trajectory data stream with timestamp alignment and coordinate correction to output preprocessed trajectory data, and performing map matching enhancement on the preprocessed trajectory data to transform discrete trajectory point sequences into continuous route data in order to build a logistics route database. Step S12: Extract high-dimensional feature vectors from multiple different dimensions based on historical route data in the logistics route database, perform big data analysis on the high-dimensional feature vectors, and identify the baseline route patterns of logistics vehicles in different delivery scenarios. Step S13: Real-time acquisition of current route data and external environment data of logistics vehicles; spatiotemporal similarity matching of current route data with corresponding benchmark route patterns; association of external environment data using big data analysis mechanisms; thereby determining the dynamic performance deviation value of current route data relative to benchmark route patterns. Step S14: If the dynamic performance deviation value exceeds the preset dynamic threshold, the current route data is marked as route data to be optimized. The route data to be optimized is input into the big data decision engine, and recommended route data is generated based on the real-time road network weight.

[0011] In step S11, the specific steps are as follows: S111: The trajectory data stream reported by logistics vehicles is accessed in real time through a distributed streaming computing framework. The trajectory data stream is timestamped using a sliding time window, and a trajectory correction mechanism based on adaptive Kalman filtering is combined with the road network topology to perform coordinate correction preprocessing to output the preprocessed trajectory data. S112: Perform map matching enhancement on the preprocessed trajectory data, and combine the road network weight probability map to perform road segment mapping correction on the preprocessed trajectory data to output a discrete trajectory point sequence. Transform the discrete trajectory point sequence into continuous route data, which includes road segment ID, driving direction and road level attributes, and construct a distributed logistics route database that supports spatiotemporal multidimensional indexing.

[0012] In step S111, the specific steps are as follows: At this point, the system constructs a data access layer based on a distributed streaming computing framework, which subscribes to and pulls the raw data stream reported by the logistics vehicle terminal in real time through a high-throughput message middleware; the raw data stream includes at least the terminal's unique identifier, timestamp, longitude, latitude, instantaneous speed, and driving direction angle.

[0013] In the distributed streaming computing framework, a sliding time window is independently set for each logistics vehicle. The window length is fixed at 60 seconds, and the sliding step is set to 10 seconds. This parameter is determined based on the 30-second reporting cycle commonly used by logistics vehicle terminals and the maximum network latency margin. When a data packet enters the window, the system calculates the difference between the server's reception time and the business timestamp reported by the vehicle terminal as the network transmission delay. If the delay exceeds 300 seconds, the data is considered to have lost its real-time nature and is discarded directly. If the delay is within 300 seconds but the data packet arrives out of order, the system inserts it into the correct time sequence position within the window based on its business timestamp. For time segments with missing data within the window, the system uses linear interpolation to fill in the gaps. That is, based on the time and coordinates of the two valid data points before and after the missing segment, the system calculates the position coordinates of the interpolation point according to the time ratio, thereby forming a complete trajectory sequence with equal time intervals.

[0014] After acquiring the raw data stream, the system configures a sliding time window mechanism for "each logistics vehicle: logically grouped according to terminal ID". This mechanism defines the window length and sliding step size. The window length is set to the maximum threshold that can cover the normal communication delay period of this type of vehicle, such as 60 seconds, and the sliding step size determines the frequency of data output.

[0015] Inside the window, a timestamp alignment operation is performed; the system extracts the server reception time and the service reporting time from the data packet, calculates the time difference between the two, i.e., the network transmission delay; if the calculated delay exceeds a preset abnormal threshold, such as 300 seconds, the data point is determined to be "expired data" and discarded to prevent historical data from interfering with real-time calculations; for out-of-order data with delays within the allowable range, it is inserted into the correct time sequence position within the window according to its timestamp; if there is a case of missing data in a specific time segment within the window, the time segment is marked as "to be interpolated" according to the data packet status before and after the window, providing a basis for subsequent processing.

[0016] For the aligned trajectory data stream, an adaptive Kalman filter algorithm is used to model and correct the vehicle's motion state. Given the large mass and inertia of logistics vehicles, their motion state has a low probability of sudden changes; therefore, a nonlinear state-space model considering the vehicle's kinematic characteristics is constructed in this stage. The system initializes the state vector of the Kalman filter, which typically includes two-dimensional planar coordinates and eastward and northward velocity components. The process noise covariance matrix (Q) and the observation noise covariance matrix (R) are defined, where Q describes the uncertainty of the vehicle's own dynamics, and R describes the error of the GPS observations.

[0017] The adaptive adjustment mechanism is implemented as follows: the system calculates the information between the current GPS observation and the one-step prediction of the filter in real time; and dynamically adjusts the weights of the observation noise covariance matrix R according to the statistical characteristics of the information.

[0018] When a vehicle is detected to be "stationary: instantaneous speed remains zero" or in a constant straight-line state, the weight of the position component in the R matrix is ​​reduced, which enhances the reliability of the observations, ensures that the filtering results closely match the GPS positioning points, and eliminates static drift.

[0019] When the vehicle is detected to be turning, changing speed, or having its GPS signal blocked, such as when entering a tunnel, high-rise building, or canyon, causing abnormal Doppler effects, the weight of the R matrix is ​​increased, that is, the reliability of the observation is reduced, and the trajectory is instead derived from the vehicle's motion model.

[0020] Through the iterative prediction and update steps described above, the system outputs a smoothed trajectory coordinate sequence, effectively eliminating high-frequency random noise in the original data, making the trajectory curve more consistent with the actual physical motion trajectory of the vehicle.

[0021] After outputting the filtered trajectory coordinates, the road network topology is introduced for further coordinate correction. This stage relies not only on geometric calculations but also on the connectivity constraints of the road network to correct the position. The system preloads a topology map containing the connection relationships between road network nodes and road segments. For each filtered trajectory point, the system searches for candidate road segments within its surrounding geometric radius, such as 50 meters. For each candidate road segment, the system calculates the projection point and projection distance from the trajectory point to the road segment.

[0022] At the same time, the system calculates the topology cost; specifically, it obtains the road segment ID matched by the trajectory point at the previous time step, and determines whether the current candidate road segment is directly connected to the previous matched road segment on the topology map; if the current candidate road segment has a topology connection relationship with the previous road segment, it is assigned a lower topology cost; if they are not connected, for example, if it needs to cross the median strip or does not turn directly at the intersection, it is assigned an extremely high topology cost or even eliminated.

[0023] By combining geometric and topological costs, the projection point with the minimum total cost is selected as the corrected coordinates at that moment. For trajectory points located in complex topological structures such as intersections, the system combines vehicle turn signal status data or historical driving direction angles to further strengthen the weighting of straight, left-turn, or right-turn segments, thereby determining the vehicle's accurate position in the road network. The system outputs preprocessed trajectory data containing corrected latitude and longitude, high-precision timestamps, matched road segment IDs, and driving direction, and transmits it to the next processing stage.

[0024] Optionally, after the aforementioned filtering process outputs smooth trajectory coordinates, the system further performs coordinate correction based on the road network topology. For each filtered trajectory point, the system searches for all candidate road segments within a 50-meter radius centered on that point; for each candidate road segment, the system calculates the vertical projection distance from the trajectory point to the centerline of that road segment as the geometric cost.

[0025] Simultaneously, the system retrieves the ID of the previously matched road segment and queries the road network topology to determine whether the current candidate road segment shares a common end node with the previously matched road segment (i.e., they are directly connected in the road network). If they are directly connected, the topology cost is set to 0. If there is an intermediate road segment between them and the length of the intermediate road segment is less than 100 meters (i.e., vehicles may complete the transition through a short road segment in a short time), the topology cost is set to 5. If they are not topologically connected and the interval is more than two road segments, the topology cost is set to infinity, and the candidate road segment is directly eliminated. The comprehensive cost is calculated as follows: Total Cost = Geometric Cost (in meters) + Topology Cost × 2, where the weight coefficient 2 of the topology cost is calibrated based on actual road test data to prioritize the road network connectivity of the matching results.

[0026] For trajectory points located at intersections, the system further extracts the vehicle's turn signal status signal at that moment: if the left turn signal is on, the topology priority of the left-turn segment is increased among all candidate road segments, i.e., its topology cost is reduced by 3; if the right turn signal is on, the topology priority of the right-turn segment is increased; if there is no turn signal, the system judges based on the vehicle's historical driving direction angle change rate in the previous 5 seconds—if the cumulative change in direction angle exceeds 15 degrees and continues to deviate in the same direction, the system determines that the vehicle is turning, and adjusts the priority weight of the candidate road segments accordingly. Finally, the system selects the projection point on the candidate road segment with the lowest overall cost as the corrected coordinates of the trajectory point, and outputs the corrected latitude and longitude, high-precision timestamp, matching road segment ID, and driving direction.

[0027] In step S112, the specific steps are as follows: At this point, after acquiring the preprocessed trajectory data, the system constructs a road network weighted probability graph for high-precision map matching. This probability graph abstracts the road network into a graph structure G=(V,E), where nodes V represent road intersections or endpoints, and edges E represent road segments. For each trajectory point, the system identifies several candidate road segments within a preset search radius, such as 100 meters. Unlike traditional matching based solely on geometric distance, this stage utilizes the calculation of observation probability and transition probability. Observation probability calculation: Based on the vertical distance from the trajectory point to the candidate road segment and the preprocessed positioning error variance, calculate the probability that the point appears on the road segment; the closer the distance, the greater the observation probability.

[0028] Transfer probability calculation: Based on the road network weight probability map, the probability of transferring from the matching candidate road segment in the previous time step to the candidate road segment in the current time step is calculated. This process introduces the road network weight as the core constraint factor. The weight is jointly determined by the geometric characteristics of the road segment, such as length and curvature, attribute characteristics, such as road grade, design speed, and vehicle adaptability characteristics.

[0029] The road network weighted probability graph assigns a dynamic weight value to each edge. For logistics vehicles, road segments with "low-level roads, such as rural roads and village roads" or "physical restrictions, such as width and height restrictions" are assigned extremely low weights or even negative weights, while highways and main roads are assigned high weights. The system uses the Viterbi algorithm to find a candidate road segment sequence in the probability graph that maximizes the overall posterior probability, as a preliminary result of map matching, thereby absorbing the discrete trajectory point sequence to the correct road network topology location.

[0030] Optionally, for each edge in the road network, the system calculates its dynamic weight value based on factors in three dimensions. The first dimension is the road grade factor, whereby the system assigns basic weight values ​​of 1.0, 0.8, 0.6, and 0.3 to expressways, Class I highways, Class II highways, Class III highways and below, respectively. This value reflects the degree of preference of logistics vehicles for different grades of roads.

[0031] The second dimension is the logistics traffic restriction factor. The system pre-imports a database of truck traffic restrictions issued by the traffic management department. For road sections with height restrictions (clearance less than 4.2 meters), width restrictions (width less than 2.5 meters), or weight restrictions (load capacity less than 10 tons), their weight value is directly set to a strong negative value of -100 to ensure that these road sections will not be selected in the subsequent matching process. For road sections with truck restricted hours, if the vehicle is currently within the restricted time window, the weight of the road section is also set to -100.

[0032] The third dimension is the historical traffic efficiency factor. The system calculates the average speed of all logistics vehicles on this road segment over the past 30 days. If the average speed is less than 30% of the designed speed for that road segment, it is considered frequently congested, and its weight value is multiplied by a discount factor of 0.5. Conversely, if the average speed is greater than 70% of the designed speed, its weight value is multiplied by a reward factor of 1.2. Ultimately, the comprehensive weight value for each road segment is the product of the corresponding values ​​of the above three factors. This weight value is recalculated every 24 hours based on the latest historical data during system operation to achieve dynamic updates. The weight values ​​of the edges in the road network weight probability graph are not only used for calculating the transition probability in the map matching stage but also directly serve the road network weight assignment in subsequent steps.

[0033] In response to the characteristics of logistics vehicles, such as long wheelbase and large turning radius, the system identifies the distribution characteristics of trajectory points in the intersection area. If the matching result shows that the vehicle made a left turn at an intersection without a left turn sign, or matched a narrow road segment with insufficient width, the system triggers correction logic: combining the vehicle's steering angle change rate with historical traffic data, the candidate road segment is re-evaluated, and the trajectory point is forcibly projected onto an adjacent road segment that meets the vehicle's physical passage capacity.

[0034] The corrected discrete trajectory point sequence is transformed into continuous route data; this process includes data aggregation and attribute extraction: trajectory compression: multiple consecutive sampling points falling on the same road segment ID and the same driving direction are merged, and only the key nodes entering and leaving the road segment are retained.

[0035] Attribute assignment: Extract the static attributes corresponding to the road segment, including the globally unique road segment ID (LinkID), "driving direction: represented by the azimuth angle along the road direction or the forward and reverse identification", and road level attributes, such as expressway, first-class highway, urban arterial road, etc.

[0036] Feature completion: Calculate derived features such as the average driving speed and dwell time of vehicles in this road segment, and associate them with the continuous route data of this segment.

[0037] The output is no longer an unordered set of coordinate points, but a logical link composed of a series of structured road segment objects, which clearly describes the road network path that the vehicle experiences from the starting point to the destination.

[0038] The system persistently stores the generated continuous route data and builds efficient query indexes. It employs a distributed columnar storage database (such as HBase) or a distributed search engine (such as Elasticsearch) as the underlying storage medium. To address the retrieval needs of massive amounts of logistics vehicle trajectory data, a composite RowKey or routing index strategy is designed: Time-dimensional index: "Time stamp, such as by day or hour" is used as the index prefix to ensure that historical data queries based on time ranges can directly locate the physical storage shard.

[0039] Spatial dimension index: The spatial location of a vehicle or road segment is encoded into a string index using spatial filling curves (such as Geohash encoding or Z-order curves), supporting fast retrieval based on "geofencing, such as querying all vehicles in a certain area".

[0040] Business Dimension Index: Build an inverted index for vehicle ID, road segment ID, and road level.

[0041] Through the index construction described above, the database can support complex multidimensional queries such as "querying the driving trajectory of a logistics vehicle on all highway sections in the past year" or "querying the congestion situation of a specific road section during morning and evening rush hours".

[0042] In step S12, the specific steps are as follows: S121: Based on the detection of the logistics route database, historical route data is determined. For the historical route data in the logistics route database, a deep feature extraction network is used to extract high-dimensional feature vectors from multiple different dimensions such as trajectory geometry, driving speed distribution, stop point distribution features and road network topology attributes. The high-dimensional feature vectors are iterated at multiple levels, and a route feature embedding space is constructed during the iteration process. S122: Spatial clustering is performed on the route feature embedding space. During the spatial clustering process, big data analysis is performed on the high-dimensional feature vectors, and pattern labeling is performed in combination with the delivery business scenario labels to identify the baseline route pattern of logistics vehicles in different delivery scenarios. The baseline route pattern covers time patterns, path preferences and energy consumption characteristics.

[0043] In step S121, the specific steps are as follows: At this point, the system retrieves and determines the set of historical route data based on the logistics route database; for each historical route data, a deep feature extraction network is constructed. This deep feature extraction network adopts a multi-branch parallel architecture to extract features from four dimensions: trajectory geometry, driving speed distribution, stop point distribution features, and road network topology attributes.

[0044] Trajectory geometry branch: The coordinate point sequence of the route is processed using a one-dimensional convolutional neural network (1D-CNN) or a long short-term memory network (LSTM); the network extracts geometric features such as the curvature change, turning angle distribution, winding degree and road shape coverage of the route. This branch aims to capture the motion pattern of logistics vehicles in physical space.

[0045] At this point, the input to the trajectory geometry branch is a sequence of route coordinate points. Each route is sampled at equal time intervals, resulting in a fixed set of 128 trajectory points. Each trajectory point contains three channels: longitude, latitude, and instantaneous velocity. The network contains three convolutional blocks: the first uses 64 kernels of size 3, outputting 64 channels; the second uses 128 kernels of size 3, outputting 128 channels; and the third uses 256 kernels of size 3, outputting 256 channels. Each convolutional block is followed by a batch normalization layer and a modified linear unit activation function, and a max-pooling layer with a stride of 2 is applied after each convolutional block to gradually reduce the sequence length. The output of the third convolutional block is aggregated into a 256-dimensional fixed-length vector by a global average pooling layer. This vector represents the geometric features of the route in physical space, such as curvature changes, turning angle distribution, and degree of meandering.

[0046] Driving speed distribution branch: Construct a speed feature encoder based on attention mechanism to process the time-speed sequence corresponding to the route; the network extracts statistical features such as average speed, speed variance, acceleration change frequency and idling ratio, focusing on the acceleration and deceleration behavior patterns of vehicles under different road conditions.

[0047] At this point, the input to the speed distribution branch is the time-speed sequence corresponding to the route, sampled at 128 time points, each containing two channels: instantaneous speed value and acceleration value. The network first maps the input to a 128-dimensional embedding representation through a fully connected layer, followed by two stacked multi-head self-attention layers, each with four attention heads. The key and query vectors of each attention head are both 32-dimensional. After the self-attention layers, a feedforward neural network with 256-dimensional hidden layers is applied, and the output undergoes layer normalization. This branch finally outputs a 128-dimensional feature vector through global average pooling. This vector focuses on statistical features of the vehicle under different road conditions, such as average speed, speed variance, acceleration mutation frequency, and idling percentage.

[0048] Stop point distribution branch: Using density clustering algorithm combined with time series analysis, stop points in the route are identified, such as service area rest areas and loading and unloading points; the network extracts features such as the duration of stop points, the distance interval between stop points, and "stop point location type: based on POI association" to form a description of the vehicle operation rhythm.

[0049] At this point, the stop point distribution branch first uses a density-based spatial clustering algorithm to identify stop points on the route trajectory. The clustering parameters include a neighborhood radius of 50 meters and a minimum number of sampling points of 5. Clusters of points with a continuous stop time exceeding 120 seconds are marked as valid stop points. After identifying stop points, the system extracts the duration of each stop point, the distance interval between the stop point and the previous stop point, and queries the surrounding point of interest type labels (such as logistics parks, gas stations, service areas, warehouse gates, etc.) through a reverse geocoding interface. These stop point features are then concatenated and input into a three-layer fully connected network with 64, 32, and 16 neurons in each layer, outputting a 16-dimensional stop point distribution feature vector that describes the vehicle's operational rhythm and stopping behavior patterns.

[0050] Road network topology attribute branch: Graph Neural Network (GNN) or Graph Attention Network (GAT) is used to process the road network structure through which the route passes; with road segments as nodes and connectivity as edges, the network extracts topological features such as "frequency of road grade change, such as the frequency of highway ramps", intersection complexity, and "turning characteristics, such as the proportion of left turns", to quantify the route's occupation of road network resources.

[0051] At this point, the road network topology attribute branch constructs the sequence of road segments traversed by the route into a graph structure, where each road segment is a node in the graph, and the connectivity between road segments is the edge of the graph. The initial feature vector of each node consists of four dimensions: the length of the road segment, the road class code (5 for expressways, 4 for Class I highways, and so on up to 2 for Class III highways), the number of lanes, and the design speed. The graph attention network contains two graph attention layers. The first layer sets up 8 attention heads, each outputting 8-dimensional features, which are concatenated to obtain 64-dimensional features. The second layer sets up 1 attention head, outputting 64-dimensional features. This branch finally outputs a 64-dimensional road network topology feature vector, quantifying topological attributes such as the frequency of road class changes, intersection complexity, and turning characteristics in the route.

[0052] The deep feature extraction network fuses the outputs from the four dimensions using a fully connected layer, outputting an initial high-dimensional feature vector that comprehensively represents the combined features of the historical route. Optionally, the feature vectors output from the four branches are 256-dimensional, 128-dimensional, 16-dimensional, and 64-dimensional, respectively. The system concatenates these to form an initial fused vector of 464 dimensions, which is then passed through a fully connected hidden layer with 128 neurons and an output layer with 64 neurons, ultimately generating a 64-dimensional initial high-dimensional feature vector. All layers use rectified linear units as activation functions, and no activation function is applied to the output layer to preserve the numerical magnitude information of the features.

[0053] After obtaining the initial high-dimensional feature vectors, the system introduces a multi-level iterative mechanism to optimize and reduce the dimensionality of the feature vectors, and constructs a route feature embedding space.

[0054] Iterative training is performed using a stacked autoencoder structure or a deep metric learning network; in each iteration, the network attempts to update the weight parameters of the feature vector by reconstructing the input features or minimizing the distance loss between samples.

[0055] Primary iteration: mainly focuses on feature denoising and integrity restoration, filling feature gaps caused by missing data.

[0056] Intermediate iteration: Introduce an optimization objective by constructing "anchor sample-positive sample-negative sample" triples. For example, different samples of the same OD route are positive samples, and samples of different OD routes are negative samples. This forces the network to narrow the distance between similar routes in the feature space and widen the distance between dissimilar routes.

[0057] Advanced iteration: Nonlinear mapping and dimensionality reduction are performed on the feature vectors to remove redundant information and retain the most distinctive semantic features.

[0058] After multiple iterations and convergence, the system outputs the final optimized high-dimensional feature vector and maps the feature vectors of all historical routes to a unified Euclidean space, namely the route feature embedding space. In this space, each point represents a logistics route, and the distance between points, such as cosine similarity or Euclidean distance, directly reflects the similarity of routes in behavioral patterns and physical paths. This space can support efficient nearest neighbor search and similarity measurement.

[0059] In step S122, the specific steps are as follows: At this point, after obtaining the route feature embedding space, the system uses a density peak-based spatial clustering algorithm to perform big data analysis on the high-dimensional feature vectors. The system calculates the similarity distance between any two feature vectors in the embedding space, such as cosine distance or Euclidean distance. Since the embedding space has been optimized through deep metric learning, similar routes exhibit a tightly clustered pattern in the space, while routes of different patterns are far apart from each other.

[0060] The system sets clustering parameters, including the neighborhood radius (Epsilon) and the minimum number of neighborhood points; the algorithm traverses the embedding space, identifies core objects, that is, points that contain enough feature vectors within their neighborhood radius, and merges the core objects and the points in their neighborhood into the same cluster through density reachability.

[0061] During the clustering process, the system automatically filters out outliers that cannot be classified into any high-density clusters. These outliers usually represent abnormal driving trajectories or data noise. The clustering results are output as several feature vector sets, each set representing a group of highly similar trajectories.

[0062] To imbue the clustering results with practical business meaning, the system introduces delivery business scenario tags to semantically annotate each generated cluster. The system predefines or extracts business metadata associated with historical route data in real time, and the tags include, but are not limited to: "Transportation type: trunk transportation, branch line delivery, last-mile delivery", "Cargo attributes: cold chain, general cargo, hazardous chemicals", "Origin and destination type: port-warehouse, warehouse-store", and "Time window: night delivery, daytime delivery".

[0063] For each cluster, the system statistically analyzes the distribution of business labels corresponding to all route samples within it; calculates the frequency of each label within the cluster, and selects the label with the highest proportion as the dominant business scenario label for that cluster based on maximum likelihood estimation or a voting mechanism.

[0064] In addition, the system also uses POI (Point of Interest) data to assist in labeling; for example, by analyzing the types of POIs around the starting and ending points of routes within a cluster, if the starting points are mostly "large logistics distribution centers" and the ending points are mostly "highway service areas," then it is labeled as "trunk line supply scenario." Through this step, the abstract mathematical clusters are transformed into scenario categories with clear business semantics, such as "inter-provincial long-distance trunk line transportation," "short-distance transshipment around cities," or "port container transportation."

[0065] For each labeled business scenario cluster, the system uses a big data statistical analysis mechanism to calculate key indicators for all route samples within that cluster, thereby constructing a baseline route pattern for that scenario. The baseline route pattern not only includes centralized feature vectors but also specifically covers the following three dimensions of statistical characteristics: temporal patterns: calculating the total travel time distribution, departure time distribution, and variance of time points passing key nodes within the cluster; the system extracts the standard time consumption model and time-efficiency features for that scenario.

[0066] Path preference: Statistical analysis of route selection preferences within a cluster; for example, calculating highway usage, frequency of passage of "specific road sections, such as river-crossing bridges and tunnels", and average number of turns, which defines the "optimal physical path" that logistics vehicles tend to choose in this scenario.

[0067] Energy consumption characteristics: Based on vehicle CAN bus data or fuel consumption model, calculate the average fuel consumption per 100 kilometers of the route within the cluster, or the proportion of electricity consumption, idling fuel consumption and climbing energy loss, which establishes the standard energy consumption baseline in this scenario.

[0068] The system outputs a series of structured baseline route pattern objects, each object uniquely corresponding to a delivery business scenario, and encapsulates the baseline parameters of time, route and energy consumption mentioned above.

[0069] In step S13, the specific steps are as follows: S131: Real-time acquisition of current route data of logistics vehicles, and synchronous access to external environmental data, including meteorological data, traffic event data and road restriction data; use a multi-scale dynamic time warping mechanism to calculate the spatiotemporal similarity between the high-dimensional feature vector of the current route data and the corresponding baseline route pattern, and construct the corresponding similarity measurement matrix based on the spatiotemporal similarity. S132: Using big data analysis mechanisms, external environmental data and similarity measurement matrices are mapped to the road network topology layer. The influence weight of environmental factors on route efficiency is calculated, and then the dynamic efficiency deviation of the current route data relative to the baseline route pattern in terms of time cost, mileage cost, and fuel consumption is determined through a weighted fusion mechanism.

[0070] In step S131, the specific steps are as follows: At this time, the system acquires the current route data of logistics vehicles in real time through the streaming computing pipeline. This data stream includes real-time updated latitude and longitude coordinates, instantaneous speed, driving direction and vehicle status information. At the same time, the system starts the multi-source data access adapter to synchronously access external environmental data.

[0071] External environment data specifically includes: Meteorological data: Obtain information on visibility, precipitation, wind speed, and road surface temperature along the vehicle's current location and predicted path from the meteorological service interface.

[0072] Traffic incident data: Information on accident locations, construction sections, congestion queue lengths, and average traffic speeds obtained from traffic management departments or internet map service providers.

[0073] Road traffic restriction data: Obtain real-time traffic restriction policies for logistics vehicles issued by traffic management departments, including time-based restrictions, area restrictions, and odd-even license plate restrictions.

[0074] The system performs strict spatiotemporal alignment processing on all incoming data; it constructs a spatiotemporal query window based on the timestamp and spatial coordinates currently reported by the vehicle; within this window, meteorological data is interpolated to the vehicle's location, traffic event data is mapped to the buffer zone in front of the vehicle, and traffic restriction data is associated with the ID of the road segment where the vehicle is currently located; after alignment, the current route data is expanded into an enhanced real-time state stream containing multi-dimensional attributes of "vehicle-road-loop".

[0075] To eliminate the micro-timeline differences caused by traffic light waiting, brief yielding, or road condition fluctuations during the operation of logistics vehicles, the system adopts a multi-scale dynamic time warping mechanism to calculate the spatiotemporal similarity between the current route data and the baseline route pattern.

[0076] The system utilizes the aforementioned deep feature extraction network to perform sliding window feature extraction on the real-time accessed current route data, generating a real-time high-dimensional feature vector sequence. Simultaneously, it retrieves the standard feature vector sequence from the corresponding baseline route pattern. Unlike traditional single-scale DTW, this step sets different scaling scales in the time and spatial dimensions: time scale regularization: allowing non-linear bending between the time axis of the current trajectory and the time axis of the baseline pattern; for example, if a vehicle is stuck in a traffic jam for 5 minutes on a certain road segment, while the baseline pattern only requires 2 minutes to travel on that segment, the DTW algorithm finds the optimal path and flexibly aligns the "congested waiting segment" of the current sequence with the "normal passage segment" of the baseline sequence, rather than directly determining a mismatch.

[0077] Spatial scale regularity: Considering the large turning radius of logistics vehicles and the wide distribution of trajectory points at intersections, the algorithm introduces dynamic bandwidth in spatial distance calculation, relaxes the penalty for local micro position offset, and focuses on matching the consistency of macro path trends.

[0078] The system calculates the cumulative distance between the two sequences after elastic alignment. This distance is normalized and converted into a spatiotemporal similarity index. The similarity value is between 0 and 1. The closer the value is to 1, the more the current driving state matches the baseline mode.

[0079] Based on the calculated spatiotemporal similarity, the system constructs a similarity metric matrix to support subsequent deviation analysis. This similarity metric matrix is ​​a multi-dimensional data structure, where the row indices correspond to time series windows, and the column indices correspond to different feature dimensions, such as path geometric similarity, speed distribution similarity, and environmental adaptability similarity.

[0080] During the matrix construction process, the system not only records the final similarity value, but also the optimal path information during the DTW alignment process; the matrix element Mi,j represents the degree of fit between the current route and the baseline pattern in the j-th feature dimension under the i-th time window.

[0081] In addition, the system injects external environmental data as a weighting factor into the matrix. If the current external environment is severe, such as in a blizzard, the system will reduce the weight of the speed dimension and increase the weight of the safety dimension, so that the metric matrix can intelligently reflect the "true" similarity under different environmental backgrounds. The matrix is ​​updated in real time, providing a quantitative numerical basis for determining whether the route deviates from the baseline.

[0082] In step S132, the specific steps are as follows: At this point, the system utilizes the spatial correlation analysis mechanism of big data to accurately map external environmental data and similarity measurement matrices onto a unified digital road network topology layer. The system uses the "edge: i.e., road segment" of the road network topology map as the basic unit; for external environmental data, it performs spatial interpolation and gridded mapping; for example, "meteorological data, such as rainfall and visibility" is interpolated to the affected road segment set using the inverse distance weighting method based on the coverage of meteorological stations; "traffic event data, such as accident sites and congestion queues" is associated with the corresponding topological edges and adjacent upstream edges based on their station numbers or geographical coordinates.

[0083] For the similarity metric matrix, the spatiotemporal feature vectors contained therein are aligned to the actual trajectory edges traveled by the vehicle according to the time series; the system assigns the similarity value of each time slice in the matrix as an attribute value to the road segment node where the vehicle is located at that moment.

[0084] Through this process, the road network topology layer is transformed into a dynamic weighted graph. Each edge in the graph not only contains static geometric attributes, but also dynamically carries the intensity of environmental disturbances, traffic flow, and the degree of matching between vehicle trajectories and baseline patterns at the current moment. This provides a spatial data foundation for subsequent refined calculations.

[0085] After mapping, the system calculates the impact weights of different environmental factors on route efficiency indicators based on the characteristics of logistics vehicles. The system constructs an environment-efficiency coupled model, which fits the nonlinear relationship between changes in environmental parameters and the decline in vehicle operating efficiency by analyzing massive historical data.

[0086] Weighting the impact of time cost: Calculate the resistance coefficients of traffic congestion level, "visibility: affecting safe driving speed" and "road surface friction coefficient: affecting braking and acceleration" on traffic speed; for example, in heavy rain, the safe driving speed of logistics vehicles needs to be significantly reduced, and at this time the weight of the impact of "meteorological factors" on "time cost" increases significantly.

[0087] The weighting of the impact on mileage cost is mainly calculated based on the detour factor caused by road construction and "traffic control: road closure". If the current route involves a detour, the weighting of the impact of "road restriction / construction data" on "mileage cost" is set to the maximum value.

[0088] The weighting of the impact on fuel consumption is as follows: the energy loss coefficients of gradient, road conditions and idling in congestion are calculated; for D-type heavy trucks, fuel consumption during start-up and hill climbing is much higher than that during constant speed driving, so "traffic events, frequent start-stop" and "terrain data" have a high impact weight on the "fuel consumption" dimension.

[0089] The system outputs a set of dynamic weight vectors W={wtime,wdist,wfuel} through multi-objective regression analysis or deep learning prediction models. This vector reflects in real time the contribution of each environmental factor to the final performance evaluation under the current complex environment.

[0090] The system uses the calculated impact weights, combined with the current route data and the baseline route pattern, to determine the dynamic performance deviation value through a weighted fusion mechanism.

[0091] The system calculates the basic deviation in each of the three dimensions: Time cost deviation: ΔT=(Tcurrent-Tbase) / Tbase, where Tcurrent is the estimated travel time after adjusting for environmental weights, and Tbase is the standard travel time in the baseline mode.

[0092] Mileage cost deviation: ΔD=(Dcurrent-Dbase) / Dbase, reflects the degree of deviation of the actual driving distance from the reference distance.

[0093] Fuel consumption deviation: ΔF=(Fcurrent-Fbase) / Fbase, reflects the degree of deviation of actual fuel consumption from the benchmark fuel consumption.

[0094] The system applies a weighted fusion formula: Edeviation = wtime⋅ΔT + wdist⋅ΔD + wfuel⋅ΔF, where Edeviation is the dynamic performance deviation value. This formula is not a simple linear superposition but introduces an environmental factor weight w. If the external environment is extremely severe, such as a blizzard causing road closures, even if vehicles do not take detours and ΔD≈0, the calculated Edeviation will still increase significantly due to "severe time delay: maximum T" and the increased environmental weight wtime, accurately reflecting the substantial decline in route performance. The system outputs this dynamic performance deviation value as the core quantitative indicator for determining whether the current route needs optimization.

[0095] In step S14, the specific steps are as follows: S141: Obtain the historical deviation statistical distribution and determine the preset dynamic threshold based on the historical deviation statistical distribution. If the dynamic performance deviation value exceeds the preset dynamic threshold, it is determined that the current route is abnormal and marked as route data to be optimized. S142: Input the route data to be optimized into the big data decision engine based on the reinforcement learning mechanism. In the big data decision engine, based on the real-time road network weight, vehicle load status and delivery time requirements, the recommended route data with the shortest time and lowest cost is generated using the multi-objective path planning mechanism. The recommended route data is then pushed to the vehicle terminal for assisted navigation, thereby realizing the dynamic closed-loop optimization of the logistics delivery route.

[0096] In step S141, the specific steps are as follows: At this point, the system retrieves and extracts a massive amount of dynamic performance deviation values ​​samples generated by logistics vehicles during their long-term operation from the historical database. These samples cover historical operating data for different seasons, time periods, road conditions, and different vehicle types, especially logistics vehicles.

[0097] For the obtained set of deviation values, the system uses kernel density estimation or Gaussian mixture model to model the statistical distribution. The system analyzes the probability density function of the deviation values, identifies the distribution characteristics of the data, such as whether it follows a normal distribution, a log-normal distribution, or has a long tail effect, and calculates the key numerical characteristics of the statistical distribution, including the mean (μ), standard deviation (σ), quantiles, such as the 75th quantile and the 95th quantile, and extreme points.

[0098] To improve the accuracy of the model, the system processes historical data in layers according to business scenarios. For logistics vehicles, the system constructs independent deviation statistical distribution sub-models for different scenarios such as "long-distance highway transportation", "urban trunk line delivery" and "port transshipment" to eliminate benchmark drift caused by scenario differences.

[0099] Through the above modeling, the system obtains a probability distribution curve that reflects the range of historical normal fluctuations. This curve describes the natural fluctuation characteristics of vehicle performance deviation values ​​under the condition of no abnormal interference.

[0100] After obtaining the statistical distribution model, the system calculates and determines a preset dynamic threshold based on the distribution. This threshold is not a fixed constant, but a judgment boundary that is dynamically adjusted according to business needs and environmental sensitivity.

[0101] The system sets a confidence level parameter, such as 95% or 99%, and inversely calculates the corresponding critical value based on the cumulative distribution function (CDF) of the historical deviation statistical distribution. If the deviation value follows a normal distribution N(μ,σ²), the dynamic threshold δ is usually set to μ+k⋅σ, where k is an adjustment coefficient determined according to the business tolerance, such as k=2 or k=3.

[0102] The system incorporates a sliding window mechanism to update the parameters μ and σ of the distribution model "regularly, such as weekly," using the latest historical deviation data. For example, if winter arrives and road icing becomes frequent, causing the overall mean μ of the historical deviation to shift upward, the system automatically adjusts the dynamic threshold δ to increase accordingly, avoiding false alarms caused by seasonal factors. The final preset dynamic threshold divides the system into a "normal fluctuation zone" and an "abnormal deviation zone." Only when the deviation value exceeds this boundary is it considered statistically significant.

[0103] The system acquires the dynamic performance deviation value Edeviation of the current route calculated in step S132 in real time and compares it with the preset dynamic threshold δ in real time.

[0104] The decision logic is as follows: If Edeviation≤δ, the system determines that although the current route has some deviation, it is still within the normal fluctuation range of historical statistics and belongs to an acceptable normal driving state, and maintains the current monitoring state.

[0105] If Edeviation > δ, the system determines that there is an anomaly in the current route, indicating that the actual operating efficiency of the vehicle has deviated significantly from the historical normal of the baseline mode.

[0106] Once an anomaly is confirmed, the system immediately marks the current route data as "route data to be optimized". This marking operation will append a specific anomaly label to the data stream, along with a trigger timestamp, deviation value, and anomaly type, such as time delay or cost increase. The marked data will be routed to the highest priority processing queue and transmitted to the big data decision engine in real time, triggering the route optimization recommendation process in step S142.

[0107] In step S142, the specific steps are as follows: At this point, the system inputs the marked route data to be optimized into the big data decision engine in real time. This decision engine is built on deep reinforcement learning (DeepQ-Network, DQN) or its variants (such as PPO, A3C) and includes an agent, an environment state space, an action space and a reward function.

[0108] During the input processing phase, the system vectorizes the route data to be optimized, real-time road network data, vehicle status data, and external environment data to construct the current state space St of the decision engine.

[0109] Road network feature vectorization: Extract the topology of the road network in the area to be optimized, including node connection relationships, road segment length, and real-time road network weights, such as congestion index and average vehicle speed.

[0110] Vehicle status feature vectorization: Extract the real-time load status of logistics vehicles, which is estimated by load sensors or historical waybills, remaining fuel / electricity, current vehicle location and physical attributes, such as length, width, height and turning radius.

[0111] Task constraint feature vectorization: Extract delivery time requirements, such as promised delivery time window and loading and unloading attributes at the destination.

[0112] The state space St is input into the policy network of the reinforcement learning network as the basic perceptual input for decision-making.

[0113] Based on the current state space St, the decision engine uses a multi-objective path planning mechanism to output the optimal action sequence, i.e., generate recommended route data. Addressing the core requirements of logistics and delivery, a multi-objective reward function R is constructed, incorporating time cost, economic cost, and safety penalty.

[0114] Time cost Ctime: Based on real-time road network weights, predict the travel time of each candidate road segment, and focus on penalizing congested road segments that cause delays.

[0115] Economic cost Ccost: Taking into account fuel consumption, the operating cost is calculated based on the logistics vehicle load and slope model, toll fees, and road wear and tear depreciation.

[0116] Psafe, a safety and compliance penalty system, assigns high negative rewards to restricted road sections, height-restricted road sections, sharp bends, and low-grade roads, taking into account the characteristics of logistics vehicles.

[0117] The reinforcement learning agent performs policy search in a simulated road network environment. It does not rely on a single static algorithm (such as Dijkstra's algorithm), but instead learns from the driving experience of historically successful drivers and statistical patterns in large datasets to find the path that maximizes the R-value while satisfying time window constraints. Load-adaptive planning: Based on the input load status, the agent prefers roads with gentler gradients and smoother surfaces to avoid extra fuel consumption caused by heavy-load uphill climbing. Time-sensitive planning: Based on the input delivery time requirements, the agent dynamically adjusts time weights; if time is tight, it tends to choose toll highways; if time is ample, it tends to choose free but slightly longer national highways. The system output includes recommended route data containing a new path node sequence, turning instructions for each road segment, estimated arrival time, and estimated cost.

[0118] The system encapsulates the generated recommended route data into a standard navigation command format, such as GPX or a proprietary protocol, and pushes it to the in-vehicle terminal via a 4G / 5G communication network. After receiving the data, the in-vehicle terminal parses it and renders the recommended route on an electronic map. The terminal system guides the driver to switch to the recommended route through voice announcements, such as "Congestion ahead, the system is replanning a highway route for you," and visual highlighting. If the driver confirms acceptance, the terminal begins providing real-time guidance according to the recommended route.

[0119] The system continuously monitors whether the driver adopts the recommended route and the actual driving effect after adoption; if the driver deviates from the recommended route, the system records the reason for the deviation as negative feedback to update the policy network of the reinforcement learning model.

[0120] Through a continuous cycle of "perception-decision-execution-feedback", the decision engine constantly evolves, making the generated recommended routes increasingly consistent with the actual driving habits of logistics vehicles and the cost reduction and efficiency improvement goals of logistics companies, thereby achieving dynamic closed-loop optimization of logistics delivery routes.

[0121] Specific application scenarios and corresponding implementation methods of this application: In step S11, a big data analysis scenario for route data of "logistics vehicles, such as heavy semi-trailer tractors" is performed. Logistics vehicles typically perform long-distance trunk transportation tasks and frequently cross mobile communication network coverage blind spots, such as mountainous areas and cross-river bridges. In the "first stage", the sliding time window mechanism can effectively handle the backlog of historical data reported in batches after the vehicle leaves the blind spot. For example, when a logistics vehicle reports 100 trajectory points in the past 5 minutes immediately after leaving the tunnel, the system restores this part of the data to the timeline according to the actual occurrence time through timestamp alignment logic, and removes duplicate or invalid drift points to ensure the temporal continuity of subsequent analysis.

[0122] Logistics vehicles, when fully loaded, can weigh up to 49 tons, exhibiting significant inertia. Furthermore, the GPS antennas, typically mounted on the roof of the cab, can cause noticeable "fishtailing" when turning due to vehicle tilt. In the "second phase," adaptive Kalman filtering leverages the vehicle's large mass to detect anomalies such as sudden GPS location jumps to the roadside non-motorized vehicle lane. For example, when a vehicle is traveling at a constant speed of 80 km / h on a highway, if a GPS point jumps into the roadside median due to signal error, the filter, based on a high-confidence motion model prediction, will smoothly correct the point towards the lane centerline instead of directly outputting erroneous jump coordinates. This avoids generating false "illegal lane change" or "deviation from the route" warnings.

[0123] Logistics vehicles are extremely sensitive to road attributes, subject to strict height, weight, and traffic restrictions, and their turning radius is much larger than that of ordinary passenger cars. In the "third stage," correction based on the road network topology is particularly crucial. For example, when a logistics vehicle travels under an urban overpass, its GPS signal drifts due to reflection, and its coordinates fall onto a side road parallel to the main road. This side road may have a height restriction of 3 meters, prohibiting trucks from passing. By searching the road network topology, the system finds that the main road and the side road are physically isolated and have no connecting intersections. The vehicle was on the main road at the previous moment, so the main road is determined to be the candidate road segment with the minimum topological cost. Based on this, the system forcibly corrects the coordinates back to the main road position, avoiding incorrectly matching the logistics vehicle to an impassable side road, and thus preventing unreasonable route planning or violation warnings based on incorrect locations in subsequent steps.

[0124] Logistics vehicles are generally taller than 4 meters and are strictly prohibited from entering urban centers or low-grade roads with insufficient load-bearing capacity. In the "first phase," when constructing the road network weight probability map, the system will overlay the electronic fences and height restriction data of logistics vehicles. For example, when a heavy semi-trailer is driving on the auxiliary lane of an expressway, the GPS signal may drift to a nearby alley with a width of only 3 meters due to errors. Ordinary algorithms may make incorrect matches due to the close proximity. However, in this solution, since the weight of the alley in the probability map is set to extremely low, or even infinitely high, indicating that it is impassable, the transition probability calculated by the HMM model will approach zero, thereby forcing the algorithm to keep the vehicle matched on the auxiliary lane of the expressway, effectively avoiding the incorrect judgment of logistics vehicles as "illegally entering the restricted area."

[0125] When a heavy semi-trailer truck makes a right turn at an intersection, the rear wheel trajectory deviates significantly inward, a phenomenon known as "inner wheel difference." In the "second stage," if the GPS trajectory points at the rear of the vehicle are mapped solely based on geometric distance, they might be mapped onto the pedestrian walkway or non-motorized vehicle lane on the right side of the intersection. The system corrects this by combining the vehicle's direction of travel with the road network topology connectivity, recognizing that the trajectory sequence is essentially a complete right turn. Therefore, it corrects the discrete offset points as a whole and maps them onto the segment ID of the target motor vehicle lane, ensuring that the generated continuous route data accurately reflects the vehicle's legal driving path, rather than a false "driving in the wrong lane" record.

[0126] In the "third stage," the amount of data generated by logistics vehicles is enormous and the query scenarios are complex. For example, the logistics dispatch center needs to analyze "the average speed of all logistics vehicles from Beijing to Shanghai that have traveled through the Jinan section of the G2 expressway in the past three months." Based on the constructed spatiotemporal multidimensional index, the system performs spatial index filtering through "road grade = expressway" and "road segment ID = G2 Jinan section", then locks the data for the past three months through time range index, and finally aggregates and calculates the average speed. This index structure enables complex analysis of hundreds of millions of trajectory data to be responded to in seconds, thereby providing real-time data support for logistics companies' capacity scheduling, route planning optimization, and cost accounting.

[0127] In step S12, logistics vehicles are extremely long and must occupy a large lane width when turning, and cannot make sharp turns at intersections like cars. In the "trajectory geometry branch", the deep network focuses on learning the features of large-radius smooth curves. For example, when the network analyzes a historical route through a complex urban intersection, it extracts the geometric features of "gradual curvature" and "large-angle turning", while suppressing the "small curvature sharp turn" feature of cars as noise or abnormal patterns. This ensures that the extracted feature vectors accurately reflect the unique spatial motion trajectory of logistics vehicles.

[0128] The speed of logistics vehicles is extremely sensitive to their load status; empty logistics vehicles accelerate quickly and travel at high speeds; while fully loaded vehicles accelerate slowly and their speed drops significantly when climbing long hills. In the "speed distribution branch," the network can identify specific load characteristics and road condition characteristics by analyzing the fluctuation patterns of the speed sequence. For example, when encountering long uphill sections, the speed feature vector of logistics vehicles will show a "low mean and low variance" pattern. The network will mark this as a heavy-load uphill mode rather than a normal congestion mode, thus avoiding misjudging normal heavy-load slow travel as route anomalies.

[0129] Logistics vehicles are strictly "restricted by road grade, such as being prohibited from entering low-grade roads below level 3" and height and weight restrictions. In the "road network topology attribute branch", the graph neural network will prioritize learning the topological connection patterns of highways, national highways and urban expressways. For example, if a historical route includes a culvert section with a height restriction of 3.5 meters, the network will encode the attribute of that section as a specific topological risk feature. In the constructed embedding space, routes containing such low-passability sections will be mapped to areas far away from the "standard trunk logistics mode", helping the system to automatically identify potentially substandard routes that do not meet the safety regulations for logistics vehicles.

[0130] Logistics vehicles typically perform point-to-point trunk transportation or branch delivery, with their stops usually corresponding to logistics parks, distribution centers, or large warehouses. In the "stop point distribution branch," the network extracts features of "long-term stops: >30 minutes" and fixed locations within logistics parks. For example, for a route from Shanghai to Beijing, the network strengthens the feature vector with the feature pattern of "long-distance direct connection between two points with only brief stops at service areas in between." These feature vectors are highly aggregated in the embedding space, forming a feature cluster of "long-distance trunk transportation," thus distinguishing it from the "urban delivery" model with frequent stops at multiple points. This achieves automated grouping of different logistics business scenarios.

[0131] Logistics vehicles are commonly used for long-distance trunk line trailer-swapping transportation and fixed-point delivery of large items within cities. In the "first phase," density clustering based on feature embedding space can effectively distinguish between these two distinct patterns. For example, for long-distance transportation from Shanghai to Chengdu, due to its large trajectory span, high proportion of highways, and only brief stops at service areas along the way, its feature vectors will form a high-density cluster in the embedding space. On the other hand, the trajectory of "delivery to building materials markets within Shanghai," although short in distance, with many traffic lights and frequent stops, will form another independent cluster of feature vectors. The clustering algorithm automatically separates these two types of data, avoiding pattern confusion caused by similar physical distances but different business logics.

[0132] In the "second phase," it is crucial to label clusters based on delivery business scenario tags. When transporting cold chain fresh produce, logistics vehicles typically require higher "timeliness: higher average speed" and are strictly prohibited from stopping for extended periods to prevent abnormal temperature control. The system analyzed the cluster tags and found that the cargo type tags for routes within a certain cluster were all "cold chain food," and the time window tags were mostly "nighttime transportation to avoid daytime congestion." The system clearly labeled this cluster as a "cold chain trunk express delivery scenario."

[0133] Logistics vehicles, due to their large size, are usually highly dependent on road grade, prefer high speeds, avoid complex intersections, and their fuel consumption is greatly affected by road conditions. In the "third stage," when constructing the baseline route pattern, the system will accurately quantify these characteristics.

[0134] Regarding route preferences: For the baseline mode labeled "long-distance general freight", statistical analysis shows that its highway utilization rate is as high as 95%, and the probability of passing through specific congested road sections is only 5%, which has become the "standard route preference" for this scenario.

[0135] Regarding energy consumption characteristics: The system calculates that the average fuel consumption of logistics vehicles in the "plain highway" benchmark mode is 32 liters per 100 kilometers, while the fuel consumption rises to 42 liters per 100 kilometers in the "mountainous national highway" benchmark mode.

[0136] Practical application: By establishing these benchmarks, when a logistics vehicle is performing a "long-distance general freight" task, if its real-time route deviates from "highway: route preference mismatch" or its fuel consumption abnormally increases to 40 liters, indicating an energy consumption mismatch, the system can determine that the vehicle has deviated from the benchmark route pattern, trigger an abnormal warning, and indicate that there may be risks of detour, fuel theft, or vehicle malfunction.

[0137] In step S13, logistics vehicles are greatly affected by policies and regulations. In the "first stage", when a vehicle enters a city ring road, the road restriction data accessed by the system simultaneously indicates that "yellow-plate trucks are prohibited from passing through the section during the morning rush hour (7:00-9:00)". The system aligns this strong constraint information with the vehicle's current route data and marks it as a "high-risk environment". At the same time, if the accessed meteorological data shows that the current road surface is icy, the system associates the environmental characteristic of low road friction coefficient with the vehicle's "logistics vehicle: large mass, long braking distance" attribute, setting a "safety first" tone for subsequent similarity calculations.

[0138] Logistics vehicles accelerate slowly when starting fully loaded, and their turning trajectories at wide intersections are large arcs, which differ greatly from those of cars. In the "second stage," traditional algorithms may misjudge anomalies because the vehicle's start is slower than the baseline time. However, by adopting the multi-scale DTW mechanism, the algorithm identifies that the vehicle is in a "heavy-load start" state, allowing the time axis to be moderately stretched in this stage, successfully aligning the current slow acceleration sequence with the normal acceleration sequence in the baseline mode. Similarly, at intersections, the spatial scale warping mechanism allows the vehicle trajectory to deviate slightly from the lane centerline. Because it requires a larger width, as long as the macroscopic direction is consistent, high spatiotemporal similarity can still be calculated, thus avoiding frequent false alarms.

[0139] In the "third stage," the constructed similarity metric matrix can sensitively capture real anomalies. For example, a logistics vehicle is traveling on a highway. Although its geometric path is highly consistent with the baseline pattern, with a path geometric similarity close to 0.98, the traffic event data accessed by the system shows that an accident occurred 5 kilometers ahead, causing the vehicle to stop for a long time. At this time, the value of the "speed distribution similarity" dimension in the matrix drops sharply to 0.2. After integrating the matrix data, the system calculates that the comprehensive spatiotemporal similarity is below the threshold, accurately identifying that although the vehicle did not "take a detour," its "operational efficiency" deviated significantly from the baseline pattern due to external traffic events. This triggers subsequent dynamic efficiency deviation calculations, prompting the decision engine to consider recommending detour routes.

[0140] The braking distance of logistics vehicles increases significantly in heavy rain or snow, and drivers must reduce their speed to ensure safety. In the "second phase", when external environmental data detects moderate to heavy snow in the route area and the road surface temperature is below zero, the system will increase the weight of the environmental factor on "time cost" wtime from 0.4 under normal circumstances to 0.7. This is because for logistics vehicles, the speed reduction caused by weather is the primary factor affecting efficiency, far more important than the distance.

[0141] Logistics vehicles are extremely heavy, and frequent starts and stops lead to a sharp increase in fuel consumption. In the "first stage" and "third stage," assuming a logistics vehicle encounters traffic congestion on an urban expressway, although its mileage deviation ΔD is almost zero and it does not take a detour, the impact weight wfuel of the "traffic event" on "fuel consumption" calculated by the system increases significantly because the external environmental data mapped to the road network layer shows severe congestion. At the same time, combined with the current trajectory characteristics of frequent starts and stops, the fuel consumption deviation ΔF calculated by the system will far exceed the time deviation ΔT. The final dynamic efficiency deviation value Edeviation will keenly reflect that although the distance has not changed, the ineffective fuel consumption caused by poor road conditions has seriously deviated from the baseline pattern, indicating the need for route optimization to avoid congested areas and thus save high fuel costs.

[0142] When logistics vehicles enter an area with temporary traffic control, such as a city center where trucks are restricted from entering during the day, in the "second phase," the impact weights of "road restriction data" on "mileage cost" and "time cost" take effect simultaneously. In the "third phase," drivers are forced to choose to detour via the ring expressway. At this time, the system calculates that the mileage deviation ΔD increases by 15%, and due to the longer detour distance, the time deviation ΔT also increases by 10%. After weighted fusion, the dynamic performance deviation value Edeviation quickly exceeds the threshold. The system identifies this as performance degradation caused by external limiting factors, rather than driver driving problems.

[0143] In step S14, the fuel consumption and speed of logistics vehicles are greatly affected by the load; the efficiency benchmarks for empty and fully loaded driving are completely different; in the "first stage", when the system obtains the historical deviation statistical distribution, it will combine vehicle load sensor data or historical waybill information to perform classification modeling; for example, the system will calculate that the average deviation distribution in the "fully loaded highway" scenario is 5%, and the average deviation distribution in the "empty return trip" scenario is 3%. Through this detailed modeling, the system avoids misjudging the normal deviation when empty by using the high standard of full load, and ensures the objectivity of the threshold setting.

[0144] Logistics vehicles often transport goods across provinces and are significantly affected by seasons and climate. In the "second phase," assuming the onset of winter, frequent snowfall in northern regions leads to highway closures or traffic restrictions. Historical data shows that the dynamic performance deviation of all logistics vehicles generally increases during this period. Based on the sliding window update distribution model, the system calculates that the dynamic threshold δ for winter is automatically adjusted from 0.15 (15% deviation) in summer to 0.25 (25% deviation). This dynamic adjustment mechanism prevents a large number of logistics vehicles from being incorrectly marked as abnormal due to force majeure environmental changes, reducing the ineffective load on the decision engine.

[0145] In the "third stage," a logistics vehicle, while performing a transportation task, entered a low-grade rural road due to the driver's unfamiliarity with the road conditions or an attempt to evade tolls. Although the distance did not increase significantly, the dynamic performance deviation value Edeviation soared to 0.40 because the logistics vehicle was traveling at extremely slow speeds and frequently braking on the narrow road. The system compared this value with the current scenario's "dynamic threshold δ: assumed to be 0.20," determined that 0.40 > 0.2, and immediately marked the route as "route data to be optimized." This marking triggered the system's intervention mechanism, which promptly alerted the dispatch center, indicating that the vehicle might have violated regulations by taking a detour or was in trouble, and that immediate route correction was necessary.

[0146] Please see Figure 2 The big data analysis system for logistics vehicle route data is applied to the aforementioned big data analysis method for logistics vehicle route data; the big data analysis system for logistics vehicle route data includes: Database module 21 is used to access the trajectory data stream reported by logistics vehicles in real time, perform time stamp alignment and coordinate correction preprocessing on the trajectory data stream to output preprocessed trajectory data, perform map matching enhancement on the preprocessed trajectory data, and transform discrete trajectory point sequences into continuous route data to build a logistics route database. The big data analysis module 22 is used to extract high-dimensional feature vectors from multiple different dimensions of historical route data in the logistics route database, perform big data analysis on the high-dimensional feature vectors, and identify the baseline route patterns of logistics vehicles in different delivery scenarios. The dynamic performance module 23 is used to acquire the current route data and external environment data of logistics vehicles in real time. The current route data is matched with the corresponding benchmark route pattern in a time and space. The external environment data is associated with the data using a big data analysis mechanism to determine the dynamic performance deviation value of the current route data relative to the benchmark route pattern. The recommended route data module 24 is used to mark the current route data as route data to be optimized if the dynamic performance deviation value exceeds the preset dynamic threshold, input the route data to be optimized into the big data decision engine, and generate recommended route data based on the real-time road network weight.

[0147] Please see Figure 3 See below for reference. Figure 3 To describe an electronic device 30 according to this embodiment of the present invention. Figure 3 The electronic device 30 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0148] like Figure 3 As shown, the electronic device 30 is presented in the form of a general-purpose computing device. The components of the electronic device 30 may include, but are not limited to: at least one processing unit 31, at least one storage unit 32, and a bus 33 connecting different system components.

[0149] The storage unit stores program code, which can be executed by the processing unit 31 to perform the steps described in the "Embodiment Methods" section of this specification according to various exemplary embodiments of the present invention.

[0150] Storage unit 32 may include readable media in the form of volatile storage units, such as random access storage unit 321 and / or cache storage unit 322, and may further include read-only storage unit 323.

[0151] Storage unit 32 may also include program 324 having a set of program modules 325, such program modules 325 including but not limited to: operating system, one or more application programs, other program modules and program data, each of these examples or some combination of these may include implementation of network environment.

[0152] Bus 33 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0153] Electronic device 30 can also communicate with one or more external devices, and with one or more devices that enable a user to interact with it, and / or with any device that enables it to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed through I / O interface 35. Furthermore, electronic device 30 can also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 36. Figure 3 As shown, network adapter 36 communicates with other modules of electronic device 30 via bus 33. It should be understood that, although... Figure 3 As not shown, other hardware and / or software modules may be used in conjunction with electronic device 30, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0154] It should be noted that although multiple modules have been mentioned in the detailed description above, the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the protection scope of the present invention; therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims

1. A big data analysis method for route data of logistics vehicles, characterized in that, include: The system receives real-time trajectory data streams reported by logistics vehicles, performs time stamp alignment and coordinate correction preprocessing on the trajectory data streams to output preprocessed trajectory data, and performs map matching enhancement on the preprocessed trajectory data to transform discrete trajectory point sequences into continuous route data in order to build a logistics route database. Based on historical route data in the logistics route database, high-dimensional feature vectors are extracted from multiple different dimensions. Big data analysis is performed on the high-dimensional feature vectors to identify the baseline route patterns of logistics vehicles in different delivery scenarios. The system acquires real-time data on the current route of logistics vehicles and external environment data. It performs spatiotemporal similarity matching between the current route data and the corresponding baseline route pattern, and uses big data analysis mechanisms to correlate external environment data, thereby determining the dynamic performance deviation value of the current route data relative to the baseline route pattern. If the dynamic performance deviation value exceeds the preset dynamic threshold, the current route data is marked as route data to be optimized. The route data to be optimized is then input into the big data decision engine, and recommended route data is generated based on the real-time road network weights.

2. The big data analysis method for logistics vehicle route data according to claim 1, characterized in that, The real-time access to the trajectory data stream reported by logistics vehicles involves preprocessing the trajectory data stream with timestamp alignment and coordinate correction to output preprocessed trajectory data. Map matching enhancement is then applied to the preprocessed trajectory data to transform discrete trajectory point sequences into continuous route data, thereby constructing a logistics route database. This includes: The system uses a distributed streaming computing framework to access the trajectory data stream reported by logistics vehicles in real time. It uses a sliding time window to align the trajectory data stream with timestamps and employs a trajectory correction mechanism based on adaptive Kalman filtering combined with the road network topology to perform coordinate correction preprocessing, so as to output the preprocessed trajectory data.

3. The big data analysis method for logistics vehicle route data according to claim 2, characterized in that, The real-time access to the trajectory data stream reported by logistics vehicles includes preprocessing the trajectory data stream with timestamp alignment and coordinate correction to output preprocessed trajectory data. Map matching enhancement is then performed on the preprocessed trajectory data to transform discrete trajectory point sequences into continuous route data, thereby constructing a logistics route database. The system also includes: Map matching enhancement is performed on the preprocessed trajectory data, and road segment mapping correction is performed on the preprocessed trajectory data in combination with the road network weight probability map to output discrete trajectory point sequences. The discrete trajectory point sequences are then transformed into continuous route data, which includes road segment ID, driving direction and road level attributes. A distributed logistics route database supporting spatiotemporal multidimensional indexing is constructed.

4. The big data analysis method for logistics vehicle route data according to claim 1, characterized in that, The process involves extracting high-dimensional feature vectors from multiple dimensions from historical route data in the logistics route database, performing big data analysis on these high-dimensional feature vectors, and identifying baseline route patterns for logistics vehicles in different delivery scenarios, including: Based on the detection of the logistics route database, historical route data is determined. For the historical route data in the logistics route database, a deep feature extraction network is used to extract high-dimensional feature vectors from multiple different dimensions such as trajectory geometry, driving speed distribution, stop point distribution features and road network topology attributes. The high-dimensional feature vectors are iterated at multiple levels, and a route feature embedding space is constructed during the iteration process.

5. The big data analysis method for logistics vehicle route data according to claim 4, characterized in that, The process of extracting high-dimensional feature vectors from multiple dimensions based on historical route data in the logistics route database, performing big data analysis on these high-dimensional feature vectors, and identifying baseline route patterns for logistics vehicles in different delivery scenarios also includes: Spatial clustering is performed on the embedded space of route features. During the spatial clustering process, big data analysis is conducted on the high-dimensional feature vectors, and pattern labeling is performed in combination with delivery business scenario labels to identify the baseline route patterns of logistics vehicles in different delivery scenarios. The baseline route patterns cover time patterns, path preferences, and energy consumption characteristics.

6. The big data analysis method for logistics vehicle route data according to claim 1, characterized in that, The real-time acquired logistics vehicle current route data and external environment data are used to perform spatiotemporal similarity matching between the current route data and the corresponding baseline route pattern. A big data analysis mechanism is then used to correlate the external environment data to determine the dynamic performance deviation value of the current route data relative to the baseline route pattern, including: The system acquires real-time route data of logistics vehicles and simultaneously accesses external environmental data, including meteorological data, traffic event data, and road restriction data. It employs a multi-scale dynamic time warping mechanism to calculate the spatiotemporal similarity between the high-dimensional feature vector of the current route data and the corresponding baseline route pattern, and constructs a corresponding similarity metric matrix based on the spatiotemporal similarity.

7. The big data analysis method for logistics vehicle route data according to claim 6, characterized in that, The real-time acquisition of current route data and external environment data of logistics vehicles, including spatiotemporal similarity matching between the current route data and the corresponding baseline route pattern, and correlation of external environment data using big data analysis mechanisms to determine the dynamic performance deviation value of the current route data relative to the baseline route pattern, also includes: By using big data analytics, external environmental data and similarity measurement matrices are mapped onto the road network topology layer. The impact weights of environmental factors on route efficiency are calculated, and then a weighted fusion mechanism is used to determine the dynamic efficiency deviation of the current route data relative to the baseline route pattern in terms of time cost, mileage cost, and fuel consumption.

8. The big data analysis method for logistics vehicle route data according to claim 1, characterized in that, If the dynamic performance deviation value exceeds a preset dynamic threshold, the current route data is marked as route data to be optimized. The route data to be optimized is then input into the big data decision engine, and recommended route data is generated based on real-time road network weights, including: Obtain the historical deviation statistical distribution and determine the preset dynamic threshold based on the historical deviation statistical distribution. If the dynamic performance deviation value exceeds the preset dynamic threshold, it is determined that the current route is abnormal and marked as route data to be optimized.

9. The big data analysis method for logistics vehicle route data according to claim 8, characterized in that, If the dynamic performance deviation value exceeds a preset dynamic threshold, the current route data is marked as route data to be optimized. The route data to be optimized is input into the big data decision engine, and recommended route data is generated based on real-time road network weights. The process also includes: The data of the route to be optimized is input into a big data decision engine based on a reinforcement learning mechanism. In the big data decision engine, based on the real-time road network weight, vehicle load status and delivery time requirements, a multi-objective path planning mechanism is used to generate recommended route data with the shortest time and lowest cost. The recommended route data is then pushed to the vehicle terminal for assisted navigation, thereby realizing dynamic closed-loop optimization of logistics delivery routes.

10. A big data analysis system for route data of logistics vehicles, characterized in that, The big data analysis system for the route data of the logistics vehicles is applied to the big data analysis method for the route data of the logistics vehicles as described in any one of claims 1-9; The big data analysis system for the route data of the logistics vehicles includes: The database module is used to access the trajectory data stream reported by logistics vehicles in real time, perform time stamp alignment and coordinate correction preprocessing on the trajectory data stream to output preprocessed trajectory data, perform map matching enhancement on the preprocessed trajectory data, and transform discrete trajectory point sequences into continuous route data to build a logistics route database. The big data analysis module is used to extract high-dimensional feature vectors from multiple different dimensions of historical route data in the logistics route database, perform big data analysis on the high-dimensional feature vectors, and identify the baseline route patterns of logistics vehicles in different delivery scenarios. The dynamic performance module is used to acquire real-time data on the current route of logistics vehicles and external environment data. The current route data is matched with the corresponding benchmark route pattern in a spatiotemporal similarity. The external environment data is correlated using a big data analysis mechanism to determine the dynamic performance deviation value of the current route data relative to the benchmark route pattern. The recommended route data module is used to mark the current route data as route data to be optimized if the dynamic performance deviation value exceeds the preset dynamic threshold. The route data to be optimized is then input into the big data decision engine, and recommended route data is generated based on the real-time road network weights.

Images