Road freight channel identification method and device, computer equipment and readable storage medium

Abstract

The invention discloses a road freight channel identification method and device, computer equipment and a readable storage medium, and relates to the technical field of traffic transportation, and the method comprises the steps: obtaining the original OD travel data of a truck based on high-precision positioning data, extracting the related OD data, and carrying out the cleaning and feature construction; obtaining characteristic variables including district and county-level OD latitude and longitude, travel distance, travel quantity and azimuth angle; a multi-layer perceptron (MLP) is adopted to carry out nonlinear dimension reduction processing on the feature variables, and an optimal low-dimensional embedded vector is obtained; carrying out clustering analysis on the low-dimensional embedded vector by adopting a segmented clustering algorithm and filtering noise; and performing path fitting processing on the filtered line segments to generate a continuous transportation channel, and finally outputting a freight transportation channel identification result after matching with a road network. According to the method, through combination of deep learning dimension reduction and segmented clustering, the spatial distribution characteristics of the freight transport channel are effectively extracted, and the freight transport channel identification precision and efficiency are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of transportation, in particular to a highway freight channel identification method and device, computer equipment and readable storage medium. BACKGROUND

[0002] The highway freight channel is a core component of the regional logistics network, and its identification accuracy directly affects the traffic resource allocation and logistics efficiency optimization. Traditional identification methods mostly rely on manual investigation or toll station data statistics, which have limited coverage, insufficient dynamics, and low identification accuracy. With the popularization of high-precision positioning technology for freight vehicles, massive raw OD travel data provide a data basis for channel identification. However, such data have characteristics such as high dimensionality, nonlinearity, and noise interference, and traditional linear dimensionality reduction methods cannot preserve key features. Direct application of clustering algorithms is susceptible to outliers, resulting in low matching degree between the identification results and the actual road network. SUMMARY

[0003] The present application relates to the field of transportation, in particular to a highway freight channel identification method and device, computer equipment and readable storage medium.

[0004] In a possible implementation, the method comprises the following steps:

[0005] Based on high-precision positioning data, raw OD travel data of freight vehicles are obtained;

[0006] OD data related to the freight channel in the raw OD travel data are extracted, and data cleaning and feature construction are performed to obtain feature variables including county-level OD latitude and longitude, travel distance, travel quantity, and azimuth angle;

[0007] A multi-layer perception MLP is used to perform nonlinear dimensionality reduction processing on the feature variables to obtain optimal low-dimensional embedding vectors;

[0008] A segmented clustering algorithm is used to perform clustering analysis on the low-dimensional embedding vectors, and noise filtering is performed on the clustering results;

[0009] The filtered line segments are subjected to path fitting processing to generate continuous transportation channels;

[0010] After matching the continuous transportation channels with the road network, the freight channel identification results are output.

[0011] In a possible implementation, the method comprises the following steps:

[0012] Data cleaning is performed on the original OD travel data to remove invalid, abnormal and duplicate data;

[0013] OD data with coal as the type of transported goods and with the starting point coordinates located within the range of a preset starting point province are screened out.

[0014] The screened OD data are merged with the county-level administrative unit as the minimum unit, and the county centroid coordinates are used to replace the original starting point and ending point longitude and latitude coordinates.

[0015] The merged OD information is output, and the merged OD information includes the starting point county name, the ending point county name, the starting point county centroid longitude and latitude, the ending point county centroid longitude and latitude, the average travel distance, the travel quantity and the azimuth.

[0016] In a possible implementation, the adopting a multi-layer perception (MLP) to perform nonlinear dimensionality reduction processing on the feature variables to obtain an optimal low-dimensional embedding vector includes:

[0017] The feature variables are subjected to standardization processing.

[0018] The standardized feature variables are input into the trained multi-layer perception model, and the low-dimensional embedding vector is output by the embedding layer of the multi-layer perception model.

[0019] The multi-layer perception model is trained by minimizing a joint loss function, and the joint loss function is a weighted sum of a reconstruction loss and a clustering loss. The reconstruction loss is used to constrain the ability of the reduced data to retain the original information, and the clustering loss is used to constrain the clustering separability of the reduced data.

[0020] In a possible implementation, the feature variables input into the multi-layer perception (MLP) include the starting point longitude, the starting point latitude, the ending point longitude, the ending point latitude, the azimuth, the average travel distance and the travel quantity, and all the feature variables are subjected to Z-score standardization processing before being input.

[0021] In a possible implementation, the segment clustering adopts a small-batch K-means algorithm.

[0022] In a possible implementation, after the clustering is completed, the clustering effect is comprehensively evaluated by calculating the contour coefficient, the Calinski-Harabasz index and the Davies-Bouldin index.

[0023] In a possible implementation, the line segments obtained after filtering are subjected to path fitting processing to generate continuous transport channels, including:

[0024] The end point coordinates of the line segments in the same cluster are subjected to spatial mean smoothing processing by using a sliding window.

[0025] Calculate the distance between adjacent line segments, and determine whether the distance is less than a preset distance threshold;

[0026] If the distance between adjacent line segments is less than the distance threshold, connect them;

[0027] If the distance is greater than the distance threshold, re-match the line segment with the nearest line segment in the same cluster for connection, and generate a continuous transport channel represented by a line connecting the end point centroids of each cluster.

[0028] In a second aspect, an embodiment of the present application provides a highway freight transport channel identification device, comprising:

[0029] An acquisition module is configured to acquire original OD travel data of a truck based on high-precision positioning data.

[0030] An identification module is configured to extract OD data related to a freight transport channel from the original OD travel data, and perform data cleaning and feature construction to obtain feature variables including county-level OD latitude and longitude, travel distance, travel quantity, and azimuth angle. The identification module is further configured to perform non-linear dimension reduction processing on the feature variables by using a multi-layer perception MLP to obtain optimal low-dimensional embedding vectors, perform clustering analysis on the low-dimensional embedding vectors by using a segmented clustering algorithm, and perform noise filtering on the clustering results. The identification module is further configured to perform path fitting processing on the line segments obtained after filtering to generate continuous transport channels, and output a freight transport channel identification result after matching the continuous transport channels with a road network.

[0031] In a third aspect, an embodiment of the present application provides a computer device, which comprises a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device performs the method of the first aspect.

[0032] In a fourth aspect, an embodiment of the present application provides a readable storage medium, which comprises a computer program. When the computer program is executed, the readable storage medium controls a computer device on which the readable storage medium is located to perform the method of the first aspect.

[0033] Compared with the prior art, the present application provides the beneficial effects including: by using the highway freight channel identification method, device, computer equipment and readable storage medium disclosed by the present application, the original OD travel data of the truck is obtained based on high-precision positioning data, the relevant OD data is extracted and cleaned and the feature variables including the county-level OD longitude and latitude, travel distance, travel quantity and azimuth angle are obtained; the multi-layer perception MLP is used to perform nonlinear dimension reduction processing on the feature variables to obtain the optimal low-dimensional embedding vector; the low-dimensional embedding vector is clustered and analyzed by using the segmented clustering algorithm and the noise is filtered; the filtered line segment is subjected to path fitting processing to generate a continuous transportation channel, and finally the highway freight channel identification result is output after matching with the road network. The method combines deep learning dimension reduction and segmented clustering to effectively extract the spatial distribution characteristics of the freight channel, and improves the accuracy and efficiency of the freight channel identification. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0035] Figure 1 The step flowchart of the highway freight channel identification method provided by the embodiment of the present application;

[0036] Figure 2 The multi-provincial area coal highway channel identification result schematic diagram provided by the embodiment of the present application;

[0037] Figure 3 The multi-provincial area coal highway channel identification path fitting schematic diagram provided by the embodiment of the present application;

[0038] Figure 4 The structure schematic diagram of the highway freight channel identification device provided by the embodiment of the present application;

[0039] Figure 5 The structure schematic diagram of the computer equipment provided by the embodiment of the present application. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0041] The specific embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0042] To solve the technical problems in the foregoing background art, Figure 1 The flowchart of the highway freight channel identification method provided by the embodiments of the present disclosure is shown below, and the highway freight channel identification method is described in detail.

[0043] Step S201, based on high-precision positioning data, obtaining original OD travel data of a truck;

[0044] Step S202, extracting OD data related to the freight channel in the original OD travel data, and performing data cleaning and feature construction to obtain feature variables including county-level OD latitude and longitude, travel distance, travel quantity, and azimuth angle;

[0045] Step S203, using a multi-layer perception MLP to perform nonlinear dimensionality reduction processing on the feature variables to obtain an optimal low-dimensional embedding vector;

[0046] Step S204, using a piecewise clustering algorithm to perform clustering analysis on the low-dimensional embedding vector, and performing noise filtering on the clustering result;

[0047] Step S205, performing path fitting processing on the filtered line segment to generate a continuous transportation channel;

[0048] Step S206, matching the continuous transportation channel with the road network, and outputting the freight channel identification result.

[0049] In the embodiments of the present application, for example, the present embodiment takes the identification of a specific freight class highway freight channel in a certain region (including multiple provincial administrative units and having typical cross-regional freight characteristics) as an application scenario, and details the complete process of the server performing the highway freight channel identification method. As the execution subject, the server realizes the automatic identification of the specific freight class highway freight channel in the region by integrating high-precision positioning data, deep learning models, and spatial clustering algorithms. The specific operation steps are described below:

[0050] The server first receives high-precision positioning data through the dynamic monitoring platform for trucks in the region. The data comes from the real-time positioning system of about 4.5 million freight vehicles in the region, covering January, April, July, and October of three consecutive natural years (covering typical transportation characteristics of four seasons to capture the regularity of freight transportation in different seasons). The coordinate system of the original OD travel data uses the national geodetic coordinate system, and the data fields include: vehicle unique code (ID), starting longitude, starting latitude, ending longitude, ending latitude, departure time, arrival time, transportation goods class, load, emission standard, driving distance, energy supply mode, etc. The server stores the received raw data in a distributed database (such as HBase), indexes it by "vehicle ID + timestamp", and ensures efficient data retrieval to facilitate fast access in subsequent steps.

[0051] The server calls the original OD travel data from the distributed database, performs data filtering, cleaning, and feature construction processes, with the following specific operations:

[0052] Data cleaning stage: The server first performs integrity check on the original data, and removes records with missing vehicle ID, starting and ending point longitude and latitude, and transportation goods class fields (about 3.5% of the original data). Through timestamp verification, it removes abnormal records with departure time later than arrival time (about 0.7%). Based on the regional administrative boundary vector data (Shapefile format), the Point-in-Polygon (PIP) algorithm is used to remove abnormal records with coordinates outside the region range (about 1.3%). For missing driving distance values, the server calculates the average value by grouping records with the same starting and ending points and similar times (e.g., for a missing distance record, the server retrieves 10 valid records of the same county pair within 3 days, and fills in the average distance of 450 kilometers), and directly removes records without valid records (about 0.4%). Finally, through vehicle ID, OD pair, and timestamp field deduplication, the most complete record is retained (deduplication ratio about 2.0%).

[0053] Data filtering stage: The server filters OD data with transportation goods class as "specific goods class" (such as coal, ore, etc. with bulk and long-distance transportation characteristics) according to the identification target of freight channel (about 65% of the cleaned data), and confirms that the starting county belongs to the region through the PIP algorithm (removes records with starting points outside the region, about 7% of the specific goods class data).

[0054] Feature construction stage: The server merges OD data by county as the minimum unit: records with the same "origin county-destination county" combination are grouped, and the original origin and destination latitude and longitude are replaced by the county centroid (calculated by administrative vector data) (for example, the centroid coordinates of county A in a certain area are E1° east and N1° north, and the centroid coordinates of county B are E2° east and N2° north); the average travel distance (for example, 100 OD records from county A to county B have an average travel distance of 420 kilometers) and the number of trips (i.e., the number of records in the group, for example, the number of trips in this group is 100) are calculated; the azimuth is calculated clockwise from the origin centroid as the polar coordinate center, with the positive east direction (0°) as the reference (for example, county B is located in the southwest direction of county A, and the azimuth is 230°). The final output includes the feature variables "origin county centroid longitude, origin county centroid latitude, destination county centroid longitude, destination county centroid latitude, average travel distance, number of trips, and azimuth" fields, stored in CSV format for subsequent dimensionality reduction processing.

[0055] The server calls the feature variable data and implements nonlinear dimensionality reduction through a multilayer perceptron (MLP). The specific process is as follows:

[0056] Feature standardization: The server performs Z-score standardization on the input features (origin longitude, origin latitude, destination longitude, destination latitude, azimuth, average travel distance, and number of trips) with the formula (where is the feature mean and is the standard deviation). For example, the original value range of the azimuth feature is 0°-360°, and after standardization, it is converted to a distribution with a mean of 0 and a standard deviation of 1 (e.g., 230° is standardized to 0.92), avoiding dimensional differences that interfere with model training.

[0057] MLP model construction and training: The server constructs a 3-layer hidden layer MLP network with 32 neurons in each layer, an input layer dimension of 7 (corresponding to 7 features), and an output layer dimension of 4 (target embedding vector dimension). The activation function uses ReLU, the optimizer is Adam (learning rate 0.001, decay coefficient 0.9), and the loss function is defined as the weighted sum of "reconstruction error + clustering loss" (weight ratio 1:1): reconstruction error is calculated by mapping the embedding vector back to the original feature space to calculate the mean square error, and clustering loss is calculated based on the distance between the embedding vector and the preset cluster center. The server trains the model using 5-fold cross-validation, iterates 500 times, adjusts the hidden layer weights each time (e.g., updates the weight matrix of the second hidden layer through backpropagation), and finally outputs the optimal 4-dimensional embedding vector (each OD sample corresponds to a 4-dimensional vector, e.g., [0.25, -0.48, 1.05, -0.12]) when the minimum loss (e.g., loss value 0.035) is achieved on the validation set.

[0058] The server performs segment clustering and noise filtering based on the spatiotemporal characteristics of the embedding vectors, in the following manner:

[0059] Spatial segmentation: The server segments the OD line segments according to their actual spatial length, with a threshold of 500 kilometers (e.g., an OD line segment of 1300 kilometers is segmented into three segments: 0-500 kilometers, 500-1000 kilometers, and 1000-1300 kilometers), ensuring that each segment represents a local transportation mode.

[0060] MiniBatchKMeans clustering: The server calls the MiniBatchKMeans algorithm in the Scikit-learn library, setting the number of clusters to 50 (determined through preliminary experiments, with the Calinski-Harabasz index reaching a maximum value of 2900 when the number of clusters = 50), the batch size to 1000 (randomly selecting 1000 samples to update the cluster centers each time), and the number of iterations to 300. The 4-dimensional embedding vectors are input into the model, and the server calculates the Euclidean distance between the samples and the cluster centers in batches, assigning cluster labels (e.g., a line segment is assigned to cluster 15). After clustering, the server calculates the SilhouetteCoefficient coefficient (0.73, close to 1, indicating good clustering results), the Calinski-Harabasz index (2900, a large value), and the Davies-Bouldin index (0.34, close to 0), comprehensively verifying the effectiveness of the clustering.

[0061] Noise filtering: The server performs post-processing on the clustering results, removing line segments with a cluster label of "-1" (marked by the algorithm as noise points) or fewer than 5 samples in the cluster (considered as isolated noise), and retaining valid clustered line segments (about 90% of the total line segments), ensuring the reliability of subsequent path fitting.

[0062] The server performs path smoothing and connection processing on the filtered valid clustered line segments, in the following manner:

[0063] Sliding window smoothing: The server sorts the line segments in the same cluster according to their spatial positions, and uses a sliding mean smoothing window of size 3 to smooth the endpoint coordinates (e.g., the start longitude of line segments 1, 2, and 3 is E3°, E4°, and E5°, respectively, so the smoothed start longitude of line segment 2 is (E3°+E4°+E5°) / 3=E4°), reducing the impact of coordinate fluctuations on path continuity.

[0064] Line segment connection and matching: the server calculates the distance between the endpoints of adjacent line segments (e.g., the straight-line distance between the endpoint of line segment A and the starting point of line segment B), and if the distance is ≤ 10 kilometers (threshold), it directly connects as a continuous line segment; if the distance > 10 kilometers, the server searches for the closest other line segment within the same cluster (e.g., the starting point of line segment C is 8 kilometers away from the endpoint of line segment A), and re-matches the connection. For example, a certain cluster contains 6 discrete line segments, after smoothing, line segment 1 is directly connected to line segment 2 with a distance of 9 kilometers, and line segment 2 is connected to line segment 3 with a distance of 16 kilometers. The server matches to line segment 4 in the same cluster (6 kilometers away from the endpoint of line segment 2), and finally connects as a continuous path of "line segment 1 - line segment 2 - line segment 4 - line segment 3".

[0065] Generating transport corridors: the server takes the cluster center trajectory as the skeleton, fits the continuous line segments into a smooth curve (e.g., using B-spline curve fitting), and generates transport corridors with spatial coherence, each corridor containing geometric shape (sequence of longitude and latitude), weight information (total number of trips, e.g., the number of trips on a certain corridor is 15000).

[0066] The server calls high-precision road network data in this area (including highway, national highway, provincial highway, etc. vector information), and aligns the fitted transport corridor with the actual road network through topological matching algorithm (e.g., adjusting the corridor path to fit the main trunk highway direction), to ensure that the recognition result conforms to the real road layout. Finally, the freight transport channel recognition result is output: including the channel geometric shape (e.g., a polyline composed of 120 longitude and latitude points), weight information (number of trips, average traffic volume), covered counties (e.g., passing through counties C and D in the area), and OD attribution data (e.g., the channel contains OD records of "county E - county F" and other county pairs), and is stored in Shapefile format, which can be used by regional transportation planning departments to optimize freight transport network (e.g., adding service facilities, optimizing traffic organization).

[0067] In the embodiment of the present application, the original OD travel data related to the freight transport channel is extracted, and data cleaning and feature construction are performed to obtain feature variables including county-level OD longitude and latitude, trip distance, trip number, and azimuth. The implementation can be performed through the following examples.

[0068] Data cleaning is performed on the original OD travel data to eliminate invalid, abnormal, and duplicate data;

[0069] OD data with coal as the type of transported goods and with the starting point coordinates located within the preset starting point province range are screened out;

[0070] The screened OD data is merged in the smallest unit of county-level administrative unit, and the county centroid coordinates are used to replace the original starting point and ending point longitude and latitude coordinates;

[0071] The merged OD information is outputted, and the merged OD information includes a start county name, an end county name, a start county centroid longitude and latitude, an end county centroid longitude and latitude, an average travel distance, a travel quantity, and an azimuth.

[0072] In the embodiment of the present application, for the task of identifying a coal freight channel in a certain area, the server performs the following processing procedure on the original OD travel data:

[0073] Data cleaning stage: the server calls 120 million pieces of original OD data stored in the distributed database, first removes records with missing vehicle ID, start and end point longitude and latitude (a total of 3.84 million, accounting for 3.2%) through field non-empty verification; deletes abnormal records (960,000, accounting for 0.8%) with departure time (e.g. 2023-10-01 14:30) later than arrival time (2023-10-01 13:15) based on timestamp logical verification; removes records with start point coordinates outside the region (e.g. an OD start longitude of 102.3° and a latitude of 28.5°, which is beyond the region boundary range, a total of 1.8 million, accounting for 1.5%) through a spatial point-face matching algorithm (PIP) and region administrative division vector data comparison. For the 240,000 records with missing travel mileage, the server retrieves valid mileage data within the last 7 days for the same OD pair (e.g. county A-county B), takes the average to fill in (e.g. for an OD pair with missing mileage, the server matches 12 valid records with an average mileage of 520 kilometers, which is filled in), and directly removes records without matching (360,000, accounting for 0.3%). Finally, through vehicle ID+OD pair+timestamp combination deduplication, 21.6 million repeated records are retained, with the most complete version of the field (deduplication ratio 1.8%).

[0074] Data screening stage: the server screens records with a transportation cargo class field marked as "coal" (a total of 68.04 million, accounting for 65%) from the cleaned data, and matches the PIP algorithm with the preset A, B and C three-province administrative division boundary data to confirm that the start county belongs to the three-province range (remove 544,000 records outside the province, accounting for 8% of the coal data).

[0075] Data merging and feature construction stage: the server merges data in units of counties: records with the same "starting county-end county" are divided into a group (for example, the "county C-County D" combination contains 1200 records), and the geometric center (centroid) is calculated by calling the county administrative division vector data to replace the original starting and ending point longitude and latitude (for example, the centroid coordinates of county C are E1° east and N1° north, and the centroid coordinates of county D are E2° east and N2° north); the average travel distance in the group (the total mileage of 1200 records is 624000 kilometers, and the average is 520 kilometers) and the number of trips (1200) are calculated; the starting centroid is taken as the origin, and the arctangent function is used to calculate the azimuth (the clockwise angle of the ending centroid relative to the starting point, for example, county D is located in the southwest direction of county C, and the azimuth is 225°).

[0076] Result output stage: the server finally outputs the feature variables containing the fields of "starting county name (such as county C), ending county name (such as county D), starting centroid longitude (E1°), starting centroid latitude (N1°), ending centroid longitude (E2°), ending centroid latitude (N2°), average travel distance (520 kilometers), number of trips (1200), and azimuth (225°)", and stores them as a structured data table for subsequent MLP dimensionality reduction processing.

[0077] In the embodiment of the application, the MLP is used to perform nonlinear dimensionality reduction processing on the feature variables to obtain an optimal low-dimensional embedding vector, which can be implemented by the following examples.

[0078] The feature variables are standardized;

[0079] The standardized feature variables are input into the trained MLP model, and the low-dimensional embedding vector is output by the embedding layer of the MLP model.

[0080] The MLP model is trained by minimizing a joint loss function, and the joint loss function is a weighted sum of a reconstruction loss and a clustering loss. The reconstruction loss is used to constrain the ability of the reduced data to retain original information, and the clustering loss is used to constrain the clustering separability of the reduced data.

[0081] In the embodiment of the application, for example, the server performs the following multi-layer perception (MLP) nonlinear dimensionality reduction process on the cleaned and feature-constructed coal freight OD data:

[0082] The server calls a structured data table containing 7 feature variables (origin county centroid longitude, latitude, destination county centroid longitude, latitude, average trip distance, trip quantity, azimuth), and performs standardization on each feature. The specific operation is as follows: the server first calculates the average value and data fluctuation range (standard deviation) of each feature, for example, the average value of the azimuth feature is 180 degrees, and the data fluctuation range is 100 degrees, and the average value of the average trip distance is 520 kilometers, and the data fluctuation range is 180 kilometers. Then each original feature value is converted into the fluctuation degree relative to the average value, so that all features are scaled to a similar numerical range (average value is 0, and data fluctuation range is 1). For example, the azimuth of a certain OD sample is 225 degrees, which is converted to 0.45 after standardization; the average trip distance is 680 kilometers, which is converted to 0.89 after standardization. The standardized feature matrix is stored as a 7-dimensional vector (for example, the standardized vector of a certain sample is [0.32, -0.41, -0.57, 0.26, 0.89, 1.25, 0.45]), avoiding the interference of dimension difference in model training.

[0083] The server constructs an MLP model for dimension reduction, with the following structure: the input layer receives 7-dimensional standardized features, and is connected to 3 hidden layers (32 neurons per layer, using ReLU activation function) in turn, and the output layer (embedding layer) outputs a 4-dimensional low-dimensional embedding vector. The model additionally contains a decoder module (3 layers of hidden layers symmetric to the encoder + 7-dimensional output layer) for reconstructing the embedding vector to the original feature dimension, to verify the information retention ability of the dimension reduction process. The server divides the standardized feature matrix into a training set and a validation set in a ratio of 8:2, trains the model using the Adam optimizer (learning rate 0.001), and iterates 500 times, adjusting the network weights through back propagation each time.

[0084] The server optimizes the model parameters through a joint loss function composed of reconstruction loss and clustering loss (weight ratio 1:1):

[0085] Reconstruction loss: measures the difference between the reconstructed features output by the decoder and the original standardized features. The smaller the difference, the more complete the original information retained by the embedding vector. For example, the original standardized features of a certain sample are [0.32, -0.41, -0.57, 0.26, 0.89, 1.25, 0.45], and the reconstructed features output by the decoder are [0.30, -0.43, -0.59, 0.24, 0.87, 1.23, 0.47], when the overall difference between the two is small, the reconstruction loss value is low.

[0086] Clustering loss: By performing clustering analysis on embedding vectors, the average distance between sample embedding vectors and their corresponding cluster centers is measured. The smaller the distance, the better the clustering separation of embedding vectors. For example, the embedding vector of a cluster center is [0.20, -0.50, 1.00, -0.20], and the sample embedding vector is [0.23, -0.51, 1.04, -0.16]. When the distance between them is small, the clustering loss value is low.

[0087] The server continuously monitors the change of the joint loss value during the training process. When the validation set loss reaches the minimum value (the sum of the reconstruction loss and the clustering loss is 0.032) at iteration 420, the training is stopped and the model parameters are saved.

[0088] The server inputs the entire standardized feature matrix into the trained MLP model, and the 4-dimensional low-dimensional embedding vector is directly output by the embedding layer. For example, after the aforementioned sample is processed by the model, the output embedding vector is [0.23, -0.51, 1.04, -0.16]. The embedding vectors of all samples form an N x 4 matrix (N is the total number of samples), which is stored for subsequent clustering analysis.

[0089] Through the above process, the server compresses high-dimensional data into low-dimensional space while preserving the original freight feature correlation, providing efficient input for subsequent clustering analysis.

[0090] In an embodiment of the present application, the feature variables input into the multi-layer perceptron MLP include: starting longitude, starting latitude, ending longitude, ending latitude, azimuth angle, average travel distance, and travel quantity; and all feature variables are subjected to Z-score standardization processing before input.

[0091] In an embodiment of the present application, the server performs the following standardization process on the feature variables input into the multi-layer perceptron (MLP) to ensure that the model input data meets the training requirements:

[0092] The server calls the structured data table after data cleaning and merging, which contains 7 feature variables to be input into the MLP, specifically: starting county centroid longitude (calculated based on the geometric center of the county administrative division), starting county centroid latitude (calculated based on the geometric center of the county administrative division), ending county centroid longitude (geometric center longitude of the ending county), ending county centroid latitude (geometric center latitude of the ending county), azimuth angle (direction angle of the ending centroid center clockwise from the starting centroid center, taking values from 0° to 360°), average travel distance (average travel distance of all OD records for the same "starting county-end county" combination), and travel quantity (total number of OD records for the same "starting county-end county" combination).

[0093] To eliminate the influence of dimensional differences on the training of the MLP model, the server performs Z-score standardization processing on the above seven characteristic variables: first, the server traverses all sample data to calculate the mean and standard deviation (data fluctuation range) of each characteristic. For example, the mean of the starting longitude characteristic is the average of the longitude of the center of mass of all counties in the region, and the standard deviation is the overall fluctuation degree of the characteristic value; the mean of the azimuth characteristic is the average of the azimuth of all OD samples, and the standard deviation is the dispersion degree of the azimuth value.

[0094] Subsequently, the server converts each characteristic value according to the Z-score standardization rule: subtract the mean of the characteristic from the original characteristic value, and then divide by the standard deviation, so that the converted characteristic value fluctuates around 0, and the overall fluctuation range is uniform. For example, the original value of the starting longitude of a certain OD sample is the mean of the characteristic plus 1.2 times the standard deviation, which is converted to 1.2 after standardization; the original value of the average travel distance is the mean of the characteristic minus 0.8 times the standard deviation, which is converted to -0.8 after standardization.

[0095] Finally, the server combines the standardization results of the seven characteristic variables into a seven-dimensional feature vector (such as [0.3, -0.5, 0.7, -0.2, 1.1, 0.6, -0.4]), forming a standardized characteristic matrix as input data for the MLP model, ensuring that the influence weight of each characteristic on parameter updating during model training is consistent.

[0096] In an embodiment of the present application, the segment clustering adopts a mini-batch K-means algorithm.

[0097] In an embodiment of the present application, for example, the server inputs the four-dimensional low-dimensional embedding vector (which has been segmented according to the 500 km threshold) obtained by MLP dimensionality reduction into the mini-batch K-means algorithm: set the number of clustering clusters to 50 (determined based on preliminary experiments to ensure the distinction of transportation modes), the batch size to 1000 (randomly select 1000 line segment samples to update the cluster center each time), and the iteration to 300 rounds. The server calculates the Euclidean distance between the sample embedding vector and the current cluster center by batch, assigns the line segment to the nearest cluster (for example, a line segment embedding vector [0.23, -0.51, 1.04, -0.16] is assigned to cluster 15), and updates the cluster center coordinates in each iteration. Finally, the clustering labels of each line segment are output, which are used for subsequent noise filtering and path fitting.

[0098] In an embodiment of the present application, after clustering is completed, the clustering effect is comprehensively evaluated by calculating the contour coefficient, the Calinski-Harabasz index, and the Davies-Bouldin index.

[0099] In the embodiment of the present application, the server evaluates the clustering effect after completing the small batch K-means clustering (cluster number 50) by the following steps: first, calculate the silhouette coefficient, traverse all the cluster segments, calculate the average distance of each sample to other samples in its own cluster (intra-cluster distance) and the average distance to the nearest other cluster samples (inter-cluster distance), and obtain the coefficient by the difference value and the maximum value (the result is 0.72, close to 1, indicating that the sample clustering is tight and well separated); then calculate the Calinski-Harabasz index, evaluate the clustering separation degree by the ratio of intra-cluster dispersion sum of squares to inter-cluster dispersion sum of squares (the result is 2800, the value is large, indicating that the cluster difference is significant); finally, calculate the Davies-Bouldin index, and calculate the average value of the ratio of the distance between cluster centers to the cluster diameter (the result is 0.35, close to 0, indicating that the cluster structure is compact). The three comprehensively indicate that the clustering result validity meets the subsequent path fitting requirements.

[0100] In the embodiment of the present application, the path fitting processing is performed on the line segments obtained after filtering to generate a continuous transportation channel, which can be implemented by the following examples.

[0101] The spatial mean smoothing processing is performed on the line segment endpoint coordinates in the same cluster by using a sliding window;

[0102] The distance between adjacent line segments is calculated, and it is judged whether the distance is less than a preset distance threshold;

[0103] If the distance between adjacent line segments is less than the distance threshold, they are connected;

[0104] If it is greater than the distance threshold, the nearest line segment is matched for connection in the same cluster, and a continuous transportation channel represented by the center line connecting the end points of each cluster is generated.

[0105] In the embodiment of the present application, the server performs path fitting processing on the effective cluster line segments (transportation line segment set in the same cluster) filtered by noise, generates a continuous transportation channel by smoothing the endpoints, distance judgment and line segment connection, and the specific process is as follows:

[0106] Sliding window spatial mean smoothing processing: The server first sorts the line segments in the same cluster according to the spatial position (in the direction from the starting point to the ending point, according to the spatial distribution order of the line segments in the cluster), and uses a sliding window with a size of 3 to perform mean smoothing on the line segment endpoint coordinates (longitude, latitude). Specifically, for the sorted line segment sequence, take the corresponding endpoint coordinates (starting point or ending point) of the current line segment and its adjacent line segments before and after it, calculate the arithmetic mean of the longitude and latitude, and take it as the smoothed coordinate of the endpoint of the current line segment. For example, a cluster contains line segment 1, line segment 2, and line segment 3 arranged in spatial order, with starting point longitudes of E105.2°, E105.5°, and E105.8°, and starting point latitudes of N36.1°, N36.3°, and N36.5°, respectively. The server performs smoothing processing on the starting point coordinates of line segment 2, and the smoothed starting point longitude is (E105.2°+E105.5°+E105.8°) / 3=E105.5°, and the smoothed starting point latitude is (N36.1°+N36.3°+N36.5°) / 3=N36.3°. Through this processing, the dispersion of the endpoint coordinates caused by the fluctuation of the original data is reduced, and the endpoint of the line segment is closer to the overall spatial distribution trend in the cluster.

[0107] Adjacent line segment distance calculation and connection judgment: After smoothing, the server calculates the endpoint distance between adjacent line segments (the straight line distance between the endpoint coordinates of the previous line segment and the starting point coordinates of the next line segment, calculated based on the longitude and latitude coordinates through the geodetic distance formula, with a unit of kilometers), and compares it with a preset distance threshold (set to 10 kilometers according to the average distance of regional road network and the continuity requirement of transportation channel). For example, after smoothing, the endpoint coordinates of line segment A are (E106.0°, N36.8°), and the starting point coordinates of line segment B are (E106.2°, N36.9°). The server calculates the straight line distance between them as 8 kilometers, which is less than the 10-kilometer threshold, and determines that it meets the direct connection condition, connects the endpoint of line segment A with the starting point of line segment B, and forms a continuous line segment "line segment A-line segment B".

[0108] Re-matching and connection of distance beyond the threshold: If the distance between adjacent line segments is greater than the 10-kilometer threshold, the server re-matches the current line segment with the nearest line segment in the same cluster for connection. Specifically, all unconnected line segments in the cluster are traversed, the straight-line distance between the end point (the starting point or the terminal point) of the current line segment and the end point of other line segments is calculated, and the line segment with the smallest distance is selected as the matching object. For example, the terminal point coordinates of line segment C are (E107.5°, N37.2°), the starting point coordinates of the original adjacent line segment D are (E108.0°, N37.8°), and the distance between them is 15 kilometers (greater than the threshold); the server retrieves other line segments in the same cluster and finds that the starting point coordinates of line segment E are (E107.6°, N37.3°), and the distance between the terminal point of line segment C and the starting point of line segment E is only 7 kilometers (the smallest distance). Therefore, line segment C is connected with line segment E to form “line segment C-line segment E”.

[0109] Generating a continuous transportation channel: After all the line segments are connected, the server takes the overall starting point centroid (the mean of the starting point coordinates of all line segments in the cluster) and the overall terminal point centroid (the mean of the terminal point coordinates of all line segments in the cluster) of each cluster as the end points, fits the connected continuous line segments into a smooth curve with the “starting point centroid-terminal point centroid” line as the skeleton, and takes it as the continuous transportation channel corresponding to the cluster. For example, the starting point centroid of a certain cluster is (E105.0°, N36.0°) and the terminal point centroid is (E109.0°, N38.0°). After fitting, the connected continuous line segments form a continuous channel from the starting point centroid to the terminal point centroid, which runs through the main transportation path in the cluster and has a clear spatial trend and continuity.

[0110] Through the above process, the server converts discrete clustered line segments into spatially continuous transportation channels, providing structured path data for subsequent road network matching.

[0111] In order to more clearly describe the scheme provided by the embodiments of the present application, a relatively complete implementation manner is provided below.

[0112] S11, obtaining truck OD travel data based on high-precision positioning data;

[0113] S12, extracting OD data related to the freight channel in the data and preliminarily constructing feature variables; first, determining the starting point province of the freight channel identification, screening out the OD freight data whose transportation goods type is coal and whose starting point coordinates are in the starting point province, merging the screened OD data in the county as the minimum unit and replacing the original coordinates with the county-level administrative area centroid, and finally outputting the OD information containing the county-level OD latitude and longitude, travel distance, travel quantity, and azimuth angle.

[0114] S13, using a multi-layer perceptron to perform nonlinear dimensionality reduction on the feature variables; the MLP multi-layer perceptron is used to reduce the dimensionality of the feature variables, the high-dimensional information is standardized, and the embedding vector after dimensionality reduction is obtained by taking the optimal loss function as the target. This step includes the optimization of the number of hidden layers and the number of hidden layers, and the embedding vector is the hidden layer vector when the loss function is optimal.

[0115] S14, segment clustering and path fitting are performed on the data after dimensionality reduction; according to the spatial continuity of the OD data, the low-dimensional embedding vector is segmented according to the preset distance threshold, the MiniBatchKMeans algorithm is used to cluster the segmented low-dimensional embedding features, and the local transportation mode and similar line segment corridors are mined. After clustering, the effective clustering results are retained, and the line segments in the cluster are smoothed and connected, and the discrete segmented line segments are fitted into continuous channel curves, improving the spatial continuity and expression consistency of the freight channel.

[0116] S15, output the freight channel recognition result; the line segment result after clustering and path fitting is matched with the road network, and the final freight channel recognition result is generated, and the geometric shape, weight information and related OD attribution data of each channel are output as a reference for subsequent highway transportation organization and transportation corridor optimization.

[0117] Specifically, please refer to Figure 2 and Figure 3 , Figure 2 the multi-province region coal highway channel recognition result schematic diagram provided by the embodiment of the present application, Figure 3 the multi-province region coal highway channel recognition path fitting schematic diagram provided by the embodiment of the present application:

[0118] S21, based on high-precision positioning data, obtain truck OD travel data; the OD data used in this embodiment comes from about 4.64 million trucks in multiple provinces (hereinafter referred to as multiple provinces) OD data, and the time range is January, April, July and October in 2019, 2021 and 2023.

[0119] S21, obtain county-level OD latitude and longitude information and travel distance, quantity information; extract the OD latitude and longitude coordinate information of the corresponding county level, and extract the travel distance, travel quantity and freight volume of each OD corresponding key indicators, to provide spatial position basis for subsequent feature extraction and spatial analysis.

[0120] S23, obtain the optimal embedding vector after dimension reduction; a multi-layer perceptron (MLP) is used for nonlinear dimension reduction of the above multi-dimensional feature variables, the MLP network structure is 3 layers of hidden layers, each layer has 32 neurons, the loss function L is a weighted sum of the additional task loss and the clustering target loss according to a weight coefficient, the activation function uses a rectified linear unit (ReLU), and the optimizer uses an adaptive moment estimation (Adam) optimization algorithm. Through the design, the embedding vector can retain the original data information while improving the separability and consistency of subsequent clustering analysis. The input feature variables include the starting longitude, the starting latitude, the ending longitude, the ending latitude, the azimuth angle, the average travel distance, and the travel distance quantity information. For numerical data, Z-score standardization processing is uniformly used, so that the mean of each feature is 0 and the standard deviation is 1, thereby reducing the influence of extreme values on the clustering result. The optimal embedding vector output after dimension reduction is used for subsequent segmented clustering and overall clustering analysis, and the Silhouette Coefficient and Calinski-Harabasz index are calculated to evaluate the dimension reduction effect.

[0121] S24, segmented clustering of the embedding vector; after obtaining the low-dimensional embedding vector, it is input into the Mini Batch KMeans clustering algorithm for segmented clustering analysis. In this embodiment, the Mini Batch KMeans clustering parameters include: the number of clustering clusters is set to 50, the batch size is set to 1000, the random seed is set to 42, the sample data input for clustering is the embedding vector after MLP dimension reduction, the dimension is 4, and all have been processed by Z-score standardization. The clustering output result includes the clustering label and the cluster center coordinates corresponding to each segmented line segment, and the spatial segmentation length of the segmented clustering is set to 500 kilometers in this embodiment. The evaluation indexes are Silhouette Coefficient index, Calinski-Harabasz index and Davies-Bouldin index

[0122] S25, determine whether it is a noise line segment; based on the clustering result, noise filtering is performed on the segmented line segment to retain the effective clustering line segment. Specifically, for the line segment with a valid class label, the corresponding clustering unit and related attribute information are recorded, and if the line segment clustering label is determined to be invalid or noise, it is removed, so as to ensure that the retained candidate channel result has good local consistency and spatial continuity.

[0123] S26, path fitting; on the basis of the segmented clustering result, the embodiment adopts a method based on a sliding window and an adjacent distance threshold to perform path smoothing and connection processing on the clustered effective line segments. The line segments in the same cluster are first subjected to sliding mean smoothing processing according to spatial positions, and the window size is 3. Then, whether adjacent segments are continuously connected or the same cluster nearest line segment is re-matched is judged according to a segment distance threshold of 10 kilometers, so as to generate a transport corridor path with spatial coherence. In the corridor output, a straight line generated by the start and end point centroids of each clustering cluster is taken as a representative expression of the corridor path, and is saved as a channel information table.

[0124] S27, output coal freight channel identification result.

[0125] The specific steps of obtaining the county-level OD latitude and longitude information and the travel distance and quantity information include the following steps:

[0126] S31, truck OD travel data; original OD travel records of coal transport trucks in the region are collected, the OD data coordinate system is CGCS2000, and the data fields include vehicle unique code (ID), vehicle starting longitude, vehicle starting latitude, vehicle ending longitude, vehicle ending latitude, vehicle departure time, vehicle arrival time, vehicle transportation type, vehicle load, vehicle emission standard, vehicle driving mileage data, and vehicle energy supply mode.

[0127] S32, data cleaning; first, the original OD travel data obtained is subjected to data cleaning, and data with incomplete data fields of vehicle unique code, vehicle starting longitude, vehicle starting latitude, vehicle ending longitude, vehicle ending latitude, and vehicle transportation type is removed. Abnormal records with time stamp confusion and geographical coordinates beyond the research region or the specified range are removed. At the same time, the missing values of the vehicle driving mileage data are filled according to the average value of other effective records in the similar time period with the same start-end combination. If there is no effective record, the record is removed. For repeated records, the vehicle ID, OD pair, and time stamp field are compared and de-duplicated. If there are multiple same OD records for the same vehicle at the same time, only the one with the most complete data fields is retained.

[0128] S33, screening coal freight; the OD data with the vehicle transportation type being coal is screened out, and the non-coal freight OD data is removed.

[0129] S34, screen starting point coordinates; the starting point longitude and latitude coordinates and the ending point longitude and latitude in the cleaned and screened coal freight OD travel data are firstly subjected to geographic position inverse analysis, in this embodiment, based on the district county boundary administrative division boundary vector data in the Shapefile format, a spatial point surface matching (Point-in-Polygon, PIP) algorithm is adopted to superimpose each coordinate on the administrative division surface, the district county unit where the coordinate is located is determined, the abnormal coordinate records that fail in inverse analysis and cannot match the administrative division unit are removed, the starting point district county name and the ending point district county name are recorded for the data records that succeed in inverse analysis, and finally the records whose district county names are not in the multi-province area are removed through field matching;

[0130] S35, merge OD travel; firstly, the data whose starting point and ending point are in the same district county is removed, the OD data is grouped with the starting point district county name and the ending point district county name as the joint primary key, each group is a new OD data, the starting point longitude and the starting point latitude of the OD data are the centroid longitude and latitude coordinates of the district county where the starting point and the ending point are located, the travel distance is the average value of the left and right vehicle OD travels in the group, the travel quantity is the record number in the group, and the azimuth angle is calculated in the clockwise direction with the starting point of the OD line as the polar coordinate center, taking the equatorial parallel direction (i.e. the positive east direction) as the 0° reference line.

[0131] S36, output district county level OD information; the output district county level OD information content includes the starting point and ending point district county name, the starting point and ending point district county centroid coordinates, the average travel distance, the travel quantity and the azimuth angle, and the output format is csv data, and the data is used for subsequent dimension reduction processing.

[0132] Please refer to Figure 4 , Figure 4 A structure schematic view of a highway freight transport channel identification device 110 provided by the embodiment of the application, comprising:

[0133] An acquisition module 1101 is configured to acquire original OD travel data of a truck based on high-precision positioning data.

[0134] An identification module 1102 is configured to extract OD data related to a freight transport channel in the original OD travel data, and perform data cleaning and feature construction to obtain feature variables including district county level OD longitude and latitude, travel distance, travel quantity and azimuth angle; adopt a multi-layer perception MLP to perform nonlinear dimension reduction processing on the feature variables to obtain an optimal low-dimensional embedding vector; adopt a segmented clustering algorithm to perform clustering analysis on the low-dimensional embedding vector, and perform noise filtering on the clustering result; perform path fitting processing on the filtered line segments to generate continuous transport channels; and after matching the continuous transport channels with a road network, output a freight transport channel identification result.

[0135] It should be noted that the implementation principle of the foregoing road freight channel identification device 110 can refer to the implementation principle of the foregoing road freight channel identification method, which will not be described here. It should be understood that the division of each module of the above device is only a logical functional division, and all or part of the actual implementation can be integrated into one physical entity, or can be physically separated. And these modules can all be implemented in the form of software called by a processing element; all can be implemented in the form of hardware; some modules can be implemented in the form of software called by a processing element, and some modules can be implemented in the form of hardware. For example, the road freight channel identification device 110 can be a separately set processing element, or can be integrated into a chip of the above device, in addition, it can also be stored in the form of program code in the memory of the above device, and the function of the above road freight channel identification device 110 is called and executed by a processing element of the above device. The implementation of other modules is similar. In addition, all or part of these modules can be integrated together, or can be independently implemented. The processing element described here can be an integrated circuit with signal processing capability. In the implementation process, each step of the above method or each module can be completed by the integrated logic circuit of the hardware in the processor element or the instructions in the form of software.

[0136] For example, the above modules can be one or more integrated circuits configured to implement the above method, such as: one or more application specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), etc. For another example, when a certain module above is implemented in the form of scheduling program code by a processing element, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules can be integrated together to implement in the form of system on a chip (SOC).

[0137] The embodiment of the application provides a computer device 100, which comprises a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the foregoing road freight channel identification device 110. As shown in the figure, Figure 5 Figure 5 ​A structural block diagram of the computer device 100 is provided for the embodiments of the present application. The computer device 100 comprises a highway freight channel identification apparatus 110, a memory 111, a processor 112 and a communication unit 113.

[0138] To realize the transmission or interaction of data, the memory 111, the processor 112 and the communication unit 113 are electrically connected with each other directly or indirectly. For example, the electrical connection between these elements can be realized by one or more communication buses or signal lines. The highway freight channel identification apparatus 110 comprises at least one software function module which can be stored in the memory 111 in the form of software or firmware or solidified in the operating system (OS) of the computer device 100. The processor 112 is used to execute the highway freight channel identification apparatus 110 stored in the memory 111, such as the software function module and the computer program comprised by the highway freight channel identification apparatus 110.

[0139] The embodiments of the present application provide a readable storage medium, which comprises a computer program. When the computer program is run, it controls the computer device where the readable storage medium is located to execute the aforementioned highway freight channel identification apparatus 110.

[0140] The foregoing description is made with reference to specific embodiments for purposes of illustration only. The illustrative discussion above is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teachings. The embodiments were chosen and described in order to best explain the principles of the disclosure and its practical application and to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.

Claims

1. A method for identifying highway freight corridors, characterized in that, include: Based on high-precision positioning data, obtain the original origin-destination (OD) travel data of trucks; Extract OD data related to freight corridors from the original OD travel data, and perform data cleaning and feature construction to obtain feature variables including county-level OD latitude and longitude, travel distance, number of trips and azimuth angle; The feature variables are nonlinearly reduced using a multilayer perceptron (MLP) to obtain the optimal low-dimensional embedding vector. The low-dimensional embedding vectors are subjected to a segmented clustering algorithm for cluster analysis, and the clustering results are filtered for noise. The filtered line segments are then subjected to path fitting processing to generate continuous transportation channels. After matching the continuous transportation channels with the road network, the freight channel identification result is output.

2. The method according to claim 1, characterized in that, The process involves extracting freight corridor-related OD data from the original OD travel data, performing data cleaning and feature construction, and obtaining feature variables including county-level OD latitude and longitude, travel distance, number of trips, and azimuth, including: The original OD travel data is cleaned to remove invalid, abnormal, and duplicate data; Filter out OD data where the transported cargo type is coal and the origin coordinates are within the preset origin province; The filtered OD data are merged with district / county-level administrative units as the smallest unit, and the original starting and ending latitude and longitude coordinates are replaced with the centroid coordinates of the district / county. Output the merged OD information, which includes the name of the starting county / district, the name of the ending county / district, the latitude and longitude of the centroid of the starting county / district, the latitude and longitude of the centroid of the ending county / district, the average travel distance, the number of travels, and the azimuth angle.

3. The method according to claim 1, characterized in that, The step of using a multilayer perceptron (MLP) to perform nonlinear dimensionality reduction on the feature variables to obtain the optimal low-dimensional embedding vector includes: The feature variables are standardized. The standardized feature variables are input into a trained multilayer perceptron model, and the low-dimensional embedding vector is output by the embedding layer of the multilayer perceptron model. The multilayer perceptron model is trained by minimizing a joint loss function, which is a weighted sum of reconstruction loss and clustering loss. The reconstruction loss is used to constrain the ability of the dimensionality-reduced data to retain the original information, and the clustering loss is used to constrain the clustering separation of the dimensionality-reduced data.

4. The method according to claim 3, characterized in that, The feature variables input to the multilayer perceptron (MLP) include: starting longitude, starting latitude, ending longitude, ending latitude, azimuth, average travel distance, and number of travels; and all feature variables are Z-score standardized before input.

5. The method according to claim 1, characterized in that, The segmented clustering uses the mini-batch K-means algorithm.

6. The method according to claim 5, characterized in that, After clustering is completed, the clustering effect is comprehensively evaluated by calculating the silhouette coefficient, Kalinsky-Harabas index, and Davis-Bourdin index.

7. The method according to claim 1, characterized in that, The step of performing path fitting processing on the filtered line segments to generate a continuous transportation channel includes: A sliding window is used to perform spatial mean smoothing on the endpoint coordinates of line segments within the same cluster. Calculate the distance between adjacent line segments and determine whether the distance is less than a preset distance threshold; If the distance between adjacent line segments is less than the distance threshold, then connect them; If the distance exceeds the threshold, the nearest line segment is reconnected within the same cluster to generate a continuous transport channel represented by the centroid line connecting the start and end points of each cluster.

8. A highway freight lane identification device, characterized in that, include: The acquisition module is used to acquire the original OD travel data of trucks based on high-precision positioning data; The identification module is used to extract OD data related to freight corridors from the original OD travel data, and perform data cleaning and feature construction to obtain feature variables including county-level OD latitude and longitude, travel distance, number of trips, and azimuth angle; the feature variables are subjected to nonlinear dimensionality reduction processing using a multilayer perceptron (MLP) to obtain the optimal low-dimensional embedding vector; the low-dimensional embedding vector is subjected to cluster analysis using a piecewise clustering algorithm, and the clustering results are filtered for noise; the filtered line segments are subjected to path fitting processing to generate continuous transportation corridors; After matching the continuous transportation channels with the road network, the freight channel identification result is output.

9. A computer device, characterized in that, The computer device includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device performs the method according to any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium includes a computer program, which, when executed, controls the computer device on which the readable storage medium is located to perform the method described in any one of claims 1-7.

Images