New energy heavy truck battery swap station identification system and method based on deep learning

Abstract

The invention discloses a new energy heavy truck battery swap station identification system and method based on deep learning. The method comprises the following steps of 1, collecting original multi-source behavior data of a new energy heavy truck and performing preprocessing to generate a track behavior sequence; step 2, identifying a stay segment of the new energy heavy truck in the track behavior sequence according to a stay point identification rule; step 3, carrying out space-time constraint density clustering analysis by adopting an improved DTW-DBSCAN algorithm to obtain a stay point cluster; 4, obtaining the first K non-repeated shortest paths of the battery replacement behavior event atlas by adopting a Yen algorithm, and determining a battery replacement behavior event; 5, drawing a thermodynamic distribution map to determine candidate positions of the battery swap station; and step 6, for each battery swap station candidate position, calculating a corresponding average Poisson probability based on Poisson distribution, and identifying the new energy heavy truck battery swap station. According to the invention, an improved DTW-DBSCAN algorithm and a Yen algorithm are fused, and an intelligent identification system for accurately identifying the battery swap station is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent identification of traffic infrastructure, and in particular to a new energy heavy truck battery swap station identification system and method based on deep learning. BACKGROUND

[0002] With the wide application of new energy heavy trucks in long-distance transportation and logistics distribution, battery swap stations, as the core infrastructure to ensure the endurance of new energy heavy trucks, have gradually become key nodes in the transportation energy system. Unlike traditional gas stations or charging piles, battery swap stations need to be reasonably distributed according to the operation rules and actual use behavior of heavy trucks to ensure the efficiency of battery swapping and the rationality of heavy truck operation scheduling. However, due to the frequent temporary parking and non-standardized operation of new energy heavy trucks in different transportation tasks, the trajectory behavior of new energy heavy trucks has high heterogeneity and spatio-temporal uncertainty, which brings challenges to the automatic identification and modeling of battery swap stations.

[0003] Most of the current common battery swap station identification technologies rely on static map information, enterprise record information or manual labeling methods at fixed sampling points, and cannot fully reflect the real battery swapping mode implied by the dynamic behavior data in actual operation. Some studies attempt to identify based on the detection of stopping points in trajectory data, but most of them use simple speed threshold or grid clustering methods, which are difficult to cope with the interference of parking behavior caused by high frequency, non-periodicity and multi-task scheduling of heavy trucks, and are prone to misjudgment of temporary parking behavior as battery swapping behavior. At the same time, existing clustering methods generally ignore the joint constraints of time and space, and cannot effectively mine the real density mode of stopping behavior. Moreover, they lack the ability of structured modeling and graph mining of behavior paths, and cannot extract stable and periodic battery swapping event patterns from trajectory behavior, resulting in low accuracy of the final identification results.

[0004] In addition, existing methods generally lack periodic verification means for battery swapping behavior, and cannot analyze the statistical rules of long-term data by combining probability models, nor can they eliminate temporary stopping points that only appear frequently in individual time periods but do not have long-term service functions. In actual deployment and site planning, there are often problems of resource waste and layout deviation, which cannot meet the precise layout requirements of battery swap infrastructure for large-scale new energy heavy truck operation.

[0005] Therefore, how to provide a new energy heavy truck battery swap station identification system and method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to propose a new energy heavy truck battery swap station recognition system and method based on deep learning, which combines the improved DTW-DBSCAN algorithm and Yen path search algorithm, combines heat map analysis and Poisson distribution modeling, realizes accurate identification and periodic verification of new energy heavy truck battery swap behavior, and constructs a high-robustness and high-precision battery swap station intelligent identification system, and improves the planning and operation efficiency of the battery swap station.

[0007] The new energy heavy truck battery swap station recognition method based on deep learning according to the embodiment of the present application comprises the following steps: Step one: collect the original multi-source behavior data generated by the new energy heavy truck in the running process, and generate the trajectory behavior sequence by preprocessing; Step two: identify the stay segment of the new energy heavy truck in the trajectory behavior sequence according to the stay point identification rule; Step three: for all identified stay segments, perform spatio-temporal constrained density clustering analysis using the improved DTW-DBSCAN algorithm to obtain stay point clustering clusters; Step four: construct a battery swap behavior event graph based on the stay point clustering cluster, obtain the top K non-repeating shortest paths of the battery swap behavior event graph using the Yen algorithm, and determine the battery swap behavior event; Step five: collect all geographic location coordinate points marked as battery swap behavior events, and draw a heat distribution graph to determine the candidate position of the battery swap station; Step six: for each battery swap station candidate position, calculate the corresponding average Poisson probability based on the Poisson distribution, Identify the new energy heavy truck battery swap station.

[0008] Optionally, the original multi-source behavior data includes geographic location positioning data, vehicle speed data, battery state data, load data and timestamp data; the preprocessing step includes respectively performing outlier rejection, missing value filling and normalization processing on different types of data in the original multi-source behavior data, and arranging the trajectory behavior sequence in ascending order according to time.

[0009] Optionally, the step two is specifically: A time window of a predetermined size is constructed for any two adjacent data points in the trajectory behavior sequence, the average speed value in the time window is calculated, and if the average speed value is lower than the predetermined speed threshold, the trajectory behavior segment of the corresponding time window size is determined as a low-speed behavior segment; The spatial movement distance of any two adjacent data points in the trajectory behavior sequence in the time window of the predetermined size is calculated, and if it is less than the spatial movement distance threshold, the trajectory behavior segment of the corresponding time window size is determined as a position keeping segment; The cumulative time of the low-speed behavior section and the position maintaining section in a plurality of continuous time windows is counted, and if the cumulative time exceeds a preset time threshold, the section is marked as a stay section, and the start time, end time and duration of the stay section are recorded.

[0010] Optionally, the improved DTW-DBSCAN algorithm specifically comprises the following steps: An attribute index table is established for all the stay sections, and the attribute index table records the number, corresponding central geographic coordinates and central time of each stay section. A trajectory section vector is constructed by extracting the behavior attributes corresponding to each stay section, and the behavior attributes include continuous geographic coordinates, stay start time, stay end time and stay duration. The DTW distance between any two trajectory section vectors is calculated, and a trajectory similarity matrix between any two stay sections is constructed, wherein each row and each column in the trajectory similarity matrix corresponds to a stay section, and each element represents the DTW distance between the corresponding trajectory section vectors, and the elements on the diagonal line represent the distance between the same stay section and are fixed as zero. Each group of elements in the trajectory similarity matrix is traversed, the corresponding trajectory number pair is extracted from the attribute index table, and the spatial distance between the central geographic coordinates of the corresponding two stay sections and the time difference between the central time points are calculated. If the spatial distance is less than a preset spatial distance threshold and the time difference is less than a preset time window threshold, the values of the corresponding two elements in the trajectory similarity matrix are kept unchanged, otherwise the corresponding two elements are marked as invalid connection, and a space-time adjacency matrix is obtained. The neighborhood radius threshold and the minimum neighborhood point number are set as clustering parameters, and the reachable neighbor point number of each stay section within the neighborhood radius threshold range is calculated according to the valid connection relationship retained in the space-time adjacency matrix. If the reachable neighbor point number of a stay section is greater than or equal to the minimum neighborhood point number, the stay section is marked as a core section. The stay sections that fail to meet the minimum neighborhood point number requirement do not participate in clustering and are removed as noise points. Each core section is taken as a starting point to search for all adjacent sections having a valid connection relationship with the core section, and a same stay point clustering cluster is formed by extension. Until all the stay sections meeting the density clustering requirement are assigned to the corresponding stay point clustering cluster.

[0011] Optionally, the step four specifically comprises the following steps: For each stay point cluster, the corresponding stay start time, stay end time, stay duration, power change amount before and after the stay, and load change amount before and after the stay are extracted to construct a stay behavior feature vector, which represents a candidate battery swap behavior event; According to a preset battery swap behavior judgment rule, it is judged whether the stay behavior feature vector meets the conditions that the power change amount is greater than a set threshold, the stay duration is within a specified time range, and the load change amount is less than a set threshold. According to all stay behavior feature vectors that meet the battery swap behavior judgment rule, a battery swap behavior event graph is constructed, and the construction step is: Sort the stay behavior feature vectors in ascending order according to the stay start time; Each stay behavior feature vector is taken as a node of the battery swap behavior event graph; The Euclidean distance between each two nodes is calculated, and if the Euclidean distance is less than a preset threshold, an edge is established between the corresponding two nodes, and the weight of the edge is the corresponding Euclidean distance; Yen algorithm is used to search the path of the battery swap behavior event graph, and the path searching step is: From the battery swap behavior event graph, the first K non-repeating shortest paths between the start node and the end node are searched, and the start node and the end node are respectively the two nodes with the earliest and latest time points; The first K non-repeating shortest paths are obtained by: Call the Dijkstra path search method to start from the start node to find a shortest path to the end node, and add the first shortest path to the shortest path list; From the second path, the following steps are executed in turn until there are K non-repeating shortest paths in the shortest path list: Take the last found shortest path as the reference path; Fix the prefix path between the start node and each intermediate node in the reference path as the prefix path, and the intermediate node is a node that does not include the start node and the end node; Remove the connection edge between the last node of the prefix path and the subsequent node in the battery swap behavior event graph; From the last node of the prefix path, re-execute the Dijkstra path search method to find a new suffix path to the end node; Combine the suffix path and the prefix path into a new complete path, and add it to the shortest path list.

[0012] The frequency of all nodes in the shortest path list is counted, and if a node appears in K paths with a frequency greater than a set frequency threshold, it is marked as a battery swap behavior event.

[0013] Optionally, the step five is specifically: Obtain the geographical position coordinate points corresponding to the stay behavior feature vectors marked as battery swap behavior events; Project all the geographical position coordinate points into a two-dimensional space map, and construct a regular space grid on the two-dimensional space map, wherein the space grid is to divide the entire geographical area into a plurality of equal-sized space cells; Statistically count the number of geographical position coordinate points contained in each space cell, and draw a heat distribution map on the two-dimensional space map, wherein the heat distribution map distinguishes different heat intensities with different colors, and the heat intensity is the ratio of the number of geographical position coordinate points contained in the space cell to the number of geographical position coordinate points in the entire geographical area, representing the density of battery swap behavior events in each space cell; Extract all the space cells exceeding the set heat intensity threshold, and take the center position coordinates of each space cell as the candidate positions of battery swap stations.

[0014] Optionally, the step six is specifically: Divide the time stamps of all battery swap behavior events according to a set time period for each candidate position of battery swap station, and construct a plurality of continuous time window sequences; Statistically count the number of battery swap behavior events in each time window, and form an event frequency sequence corresponding to the candidate position of battery swap station; Map the event frequency sequence to the observation samples of a Poisson process, and calculate the occurrence probability under the current Poisson model parameters according to the number of battery swap behavior events in each time window; Summarize the probability values in all time windows, and calculate the average Poisson probability; If the average Poisson probability is greater than a set Poisson probability threshold, it indicates that the corresponding candidate position of battery swap station stably exists battery swap behavior events in multiple periods, which meets the periodic behavior characteristics, and is identified as a real existing new energy heavy truck battery swap station; If the average Poisson probability is lower than the set Poisson probability threshold, it is determined as a temporary stay behavior point.

[0015] The new energy heavy truck battery swap station identification system based on deep learning according to the embodiment of the application comprises the following modules: A data acquisition and preprocessing module is configured to collect original multi-source behavior data generated by new energy heavy trucks during operation, and perform outlier rejection, missing value filling and normalization processing on different types of data respectively, and arrange the trajectory behavior sequence in ascending order of time; A stay segment identification module is configured to identify low-speed behavior segments, position maintaining segments, and determine stay segments; A space-time constraint density clustering analysis module is configured to perform space-time constraint density clustering analysis on all stay segments by using an improved DTW-DBSCAN algorithm; A battery replacement behavior event identification module is configured to construct a stay behavior feature vector according to each stay point cluster, and obtain the first K non-repeated shortest paths of a battery replacement behavior event graph by using a Yen algorithm, so as to determine the battery replacement behavior event. A heat distribution graph analysis module is configured to map the geographic position coordinates of all battery replacement behavior events to a space grid, count the number of geographic position coordinate points in each space cell, generate a heat distribution graph, and extract the center position coordinate points of the space cells with heat intensity exceeding a set threshold as candidate battery replacement stations. A periodic behavior verification module is configured to count the number of battery replacement behavior events of each candidate battery replacement station within a set time period, construct an event frequency sequence, map the event frequency sequence to an observation sample of a Poisson process, calculate the average value of Poisson probabilities in all time windows, determine whether the periodic behavior feature is met, and finally confirm the new energy heavy truck battery replacement station that actually exists.

[0016] The present application has the following advantages: The present application proposes a battery replacement station identification method by fusing space-time constraint, density clustering, behavior graph construction and periodic verification, which can effectively improve the accuracy and stability of the battery replacement station positioning. Compared with the existing technology which depends on a static map or a simple trajectory stay point extraction method, the present application introduces behavior feature vector construction, an improved DTW-DBSCAN algorithm and Yen algorithm path analysis, which can accurately separate the real battery replacement behavior events from the complex running trajectory, avoiding the error of mistaking temporary stay or accidental parking as battery replacement behavior. At the same time, the rule space grid statistics and Poisson probability modeling method are used to identify the periodicity of the battery replacement event frequency distribution, further eliminate the non-periodic or accidental high-frequency stay points, and ensure that the identified station has long-term stability and real service ability. The method realizes multi-level processing and verification from the original trajectory to the battery replacement station, effectively improves the automation and intelligence level of the battery replacement station identification, and provides strong data support and algorithm basis for new energy heavy truck operation path planning and energy infrastructure scientific deployment. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, illustrate the application, and are used to explain the application together with the embodiments of the application, and do not constitute a limitation on the application. In the drawings:

[0018] Figure 1 The overall flowchart of the new energy heavy truck battery replacement station identification method based on deep learning proposed by the present application; Figure 2 A structure diagram of a new energy heavy truck battery swap station recognition system based on deep learning is provided for the present application. DETAILED DESCRIPTION

[0019] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only show the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0020] REFERENCE Figure 1 The new energy heavy truck battery swap station recognition method based on deep learning comprises: Step one: Collecting original multi-source behavior data generated by new energy heavy trucks during operation and preprocessing to generate trajectory behavior sequences; Step two: Identifying the stay segments of new energy heavy trucks in the trajectory behavior sequences according to the stay point identification rules; Step three: For all identified stay segments, using the improved DTW-DBSCAN algorithm for spatio-temporal constraint density clustering analysis to obtain stay point clustering clusters; Step four: Building a battery swap behavior event graph based on the stay point clustering clusters, using the Yen algorithm to obtain the top K non-repeating shortest paths of the battery swap behavior event graph, and determining the battery swap behavior event; Step five: Collecting all geographic location coordinate points marked as battery swap behavior events and drawing a heat distribution graph to determine the candidate positions of the battery swap station; Step six: For each battery swap station candidate position, calculating the corresponding average Poisson probability based on the Poisson distribution, Recognizing the new energy heavy truck battery swap station.

[0021] This step realizes the accurate identification and spatial clustering of new energy heavy truck battery swap behavior by modeling multi-source behavior data, integrating the improved DTW-DBSCAN density clustering algorithm and the Yen path search algorithm, combining heat map analysis and Poisson probability modeling, effectively mining the distribution positions of the real existing battery swap stations, and significantly improving the accuracy and reliability of the identification.

[0022] In this embodiment, the original multi-source behavior data includes geographic location positioning data, vehicle speed data, battery state data, load data, and timestamp data; the preprocessing step includes outlier rejection, missing value filling, and normalization processing for different types of data in the original multi-source behavior data, and arranging the trajectory behavior sequences in ascending order of time.

[0023] The step one cleans and uniformly processes key parameters such as geographic position, speed, battery state and load in multi-source behavior data, ensures the integrity and consistency of the data, significantly improves the accuracy and robustness of subsequent trajectory modeling and behavior recognition, and lays a high-quality data foundation for accurate identification of new energy heavy truck battery replacement behavior.

[0024] In the embodiment, the step two is specifically: A time window of a preset size is constructed for any two adjacent data points in the trajectory behavior sequence, an average speed value in the time window is calculated, and if the average speed value is lower than a preset speed threshold, the trajectory behavior segment of the corresponding time window size is determined as a low-speed behavior segment; The spatial moving distance of any two adjacent data points in the trajectory behavior sequence in a time window of a preset size is calculated, and if the spatial moving distance is less than a spatial moving distance threshold, the trajectory behavior segment of the corresponding time window size is determined as a position keeping segment; The cumulative time of low-speed behavior segments and position keeping segments in a plurality of continuous time windows is counted, and if the cumulative time exceeds a preset time threshold, the segment is marked as a stay segment, and the start time, end time and duration of the stay segment are recorded.

[0025] The step one cleans and uniformly processes key parameters such as geographic position, speed, battery state and load in multi-source behavior data, ensures the integrity and consistency of the data, significantly improves the accuracy and robustness of subsequent trajectory modeling and behavior recognition, and lays a high-quality data foundation for accurate identification of new energy heavy truck battery replacement behavior.

[0026] In the embodiment, the improved DTW-DBSCAN algorithm is specifically: An attribute index table is established for all stay segments, and the attribute index table records the number, corresponding central geographic coordinates and central time of each stay segment; A trajectory segment vector is constructed by extracting the behavior attributes corresponding to each stay segment, and the behavior attributes include continuous geographic position coordinates, stay start time, stay end time and stay duration; The DTW distance between any two trajectory segment vectors is calculated, a trajectory similarity matrix between any two stay segments is constructed, each row and each column in the trajectory similarity matrix corresponds to a stay segment, each element represents the DTW distance between the corresponding trajectory segment vectors, and the elements on the diagonal line represent the distance between the same stay segment, which is fixed as zero; Each group of elements in the trajectory similarity matrix is traversed, the corresponding trajectory number pair is extracted from the attribute index table, and the spatial distance between the central geographic coordinates of the corresponding two stay segments and the time difference between the central time points are calculated; If the spatial distance is less than the preset spatial distance threshold and the time difference is less than the preset time window threshold, then the values ​​of the two corresponding elements in the trajectory similarity matrix remain unchanged; otherwise, the two corresponding elements are marked as invalid connections, and the spatiotemporal adjacency matrix is ​​obtained. The neighborhood radius threshold and the minimum number of neighborhood points are set as clustering parameters. Based on the effective connection relationships retained in the spatiotemporal adjacency matrix, the number of reachable neighbor points for each dwell segment within the neighborhood radius threshold range is calculated. If the number of reachable neighbor points of a dwell segment is greater than or equal to the number of minimum neighbor points, it is marked as a core segment; Fragments that fail to meet the minimum neighborhood point requirement are not included in clustering and are removed as noise points. Starting from each core segment, search for all adjacent segments that have a valid connection with the core segment, and expand to form clusters with the same dwell point; This continues until all dwell segments that meet the density clustering requirements are assigned to the corresponding dwell point clusters.

[0027] This step introduces an improved DTW-DBSCAN algorithm, incorporating triple constraints of trajectory similarity, spatial proximity, and temporal proximity into cluster analysis. This significantly enhances the accuracy and clustering effect of identifying the dwelling behavior of new energy heavy-duty trucks. Compared to traditional clustering methods based on a single distance metric, this approach can more effectively filter temporary and atypical dwelling behaviors, thereby accurately extracting clusters of dwelling points with spatiotemporal consistency. This provides high-quality foundational data for subsequent battery swapping event modeling.

[0028] In this embodiment, step four specifically includes: For each dwell point cluster, the corresponding dwell start time, dwell end time, dwell duration, change in battery power before and after the dwell, and change in load before and after the dwell are extracted to construct a dwell behavior feature vector, which represents a candidate battery swapping behavior event. Based on the preset battery swapping behavior judgment rules, determine whether the dwell behavior feature vector meets the conditions that the change in battery power is greater than a set threshold, the dwell time is within a specified time range, and the change in load is less than a set threshold. A battery swapping behavior event map is constructed based on the feature vectors of all dwelling behaviors that satisfy the battery swapping behavior determination rules. The construction steps are as follows: Sort the dwell behavior feature vectors in ascending order by the start time of dwell time; Each dwell behavior feature vector is used as a node in the battery swapping behavior event graph; Calculate the Euclidean distance between any two nodes. If the Euclidean distance is less than a preset threshold, establish an edge connection between the corresponding two nodes, with the weight of the edge being the corresponding Euclidean distance. The path search step is: K shortest paths between the start node and the end node are searched from the battery swap behavior event graph, the start node and the end node being the two nodes with the earliest and latest time points respectively; The K shortest paths obtaining step is: A Dijkstra path search method is called to find a shortest path from the start node to the end node as the first shortest path and added to the shortest path list; From the second path, the following steps are sequentially executed until there are K shortest paths in the shortest path list: The last shortest path found is taken as a reference path; The path segment between the start node and each intermediate node in the reference path is fixed as a prefix path, the intermediate node being a node excluding the start node and the end node; The connection edge between the last node of the prefix path and the subsequent node in the battery swap behavior event graph is removed; A Dijkstra path search method is re-executed to find a new suffix path from the last node of the prefix path to the end node; The suffix path and the prefix path are combined into a new complete path and added to the shortest path list.

[0029] The frequency of occurrence of all nodes in the shortest path list is counted, and if the frequency of occurrence of a node in the K paths exceeds a set frequency threshold, the node is marked as a battery swap behavior event.

[0030] This step effectively identifies the battery swap behavior mode in the new energy heavy truck running track by constructing a battery swap behavior event graph and introducing a Yen algorithm for multi-path search. Compared with the existing battery swap behavior recognition method based on only a single feature threshold, the graph construction method fully considers the internal structural relationship of the battery swap event in time, space and behavior attributes, and improves the robustness and accuracy of recognition through multi-path redundancy verification, achieving high confidence labeling of real battery swap behavior events.

[0031] In this embodiment, the step five is specifically: All geographical location coordinate points corresponding to the stay behavior feature vectors marked as battery swap behavior events are obtained; All geographical location coordinate points are projected into a two-dimensional space map, and a regular space grid is constructed on the two-dimensional space map, the space grid being a space cell that divides the entire geographical area into several equal-sized space cells; counting the number of geographic position coordinate points contained in each spatial cell, and drawing a heat distribution map on a two-dimensional spatial map, the heat distribution map distinguishing different heat intensities with different colors, the heat intensity being a ratio of the number of geographic position coordinate points contained in the spatial cell to the number of geographic position coordinate points of the entire geographic region, representing the density of battery swap behavior events in each spatial cell; extracting all spatial cells exceeding a set heat intensity threshold, and taking the center position coordinates of each spatial cell as a battery swap station candidate position.

[0032] This step can significantly improve the spatial aggregation recognition ability of the high-frequency distribution area of the battery swap station by constructing a battery swap event heat distribution map through spatial gridding processing of the marked battery swap behavior events. By introducing the density index of heat intensity, the sporadic or discrete interference behavior points are effectively eliminated, and the hotspot area with significant aggregation characteristics is focused on, and the preliminary battery swap station candidate point is extracted based on the center coordinates, providing accurate spatial input basis for subsequent station periodic verification and real battery swap station determination.

[0033] In this embodiment, the step six is specifically: dividing all battery swap behavior event timestamps according to a set time period for each battery swap station candidate position, to construct a plurality of continuous time window sequences; counting the number of battery swap behavior events in each time window to form an event frequency sequence corresponding to the battery swap station candidate position; mapping the event frequency sequence to an observation sample of a Poisson process, and calculating the occurrence probability under the current Poisson model parameter according to the number of battery swap behavior events in each time window; summarizing the probability values in all time windows and calculating the average Poisson probability; If the average Poisson probability is greater than a set Poisson probability threshold, it indicates that the corresponding battery swap station candidate position has stable battery swap behavior events in multiple periods, which meets the periodic behavior characteristics, and is identified as a real new energy heavy truck battery swap station; If the average Poisson probability is lower than the set Poisson probability threshold, it is determined as a temporary stay behavior point.

[0034] This step can effectively depict the stability and repeatability of the battery swap behavior events in the time dimension by constructing the event frequency sequence and mapping it to the observation sample of the Poisson process, and introducing the average Poisson probability as the quantitative criterion of periodic behavior. Compared with the traditional method relying on empirical threshold judgment, this method has stronger statistical robustness and universal adaptability, and can accurately distinguish between long-term fixed battery swap stations and sporadic temporary stop points, thereby significantly improving the accuracy and robustness of new energy heavy truck battery swap station recognition, and providing reliable support for battery swap infrastructure planning.

[0035] Reference Figure 2 The new energy heavy truck battery swap station recognition system based on deep learning comprises the following modules: A data acquisition and preprocessing module is configured to acquire original multi-source behavior data generated by a new energy heavy truck during operation, and perform outlier rejection, missing value filling and normalization processing on different types of data respectively, and arrange the trajectory behavior sequence in ascending order of time; A stay segment identification module is configured to identify low-speed behavior segments, position keeping segments, and determine stay segments; A spatiotemporal constraint density clustering analysis module is configured to perform spatiotemporal constraint density clustering analysis on all stay segments using an improved DTW-DBSCAN algorithm; A battery swap behavior event identification module is configured to construct a stay behavior feature vector according to each stay point clustering cluster, and use the Yen algorithm to obtain the top K non-repeating shortest paths of the battery swap behavior event graph, and determine the battery swap behavior event; A heat distribution graph analysis module is configured to map the geographic location coordinates of all battery swap behavior events to a spatial grid, count the number of geographic location coordinate points in each spatial cell, generate a heat distribution graph, and extract the spatial cell center location coordinate points with heat intensity exceeding a set threshold as battery swap station candidate points; A periodic behavior verification module is configured to count the number of battery swap behavior events of each battery swap station candidate point within a set time period, construct an event frequency sequence, and map it to an observation sample of a Poisson process, calculate the average value of the Poisson probability in all time windows, and determine whether the periodic behavior feature is satisfied, and finally confirm the existence of a real new energy heavy truck battery swap station.

[0036] The system integrates multiple deep learning and data mining modules such as data acquisition, stay identification, density clustering, path analysis, heat mapping and Poisson probability modeling, forms an end-to-end battery swap station recognition closed loop process, and can realize the whole process automatic identification of new energy heavy truck battery swap behavior from original trajectory data analysis to real battery swap station confirmation. The logic of each module in the system is closely connected, which not only guarantees the data integrity of the processing chain, but also enhances the spatiotemporal accuracy of the identification and the statistical reliability of the behavior discrimination, significantly improves the accuracy, coverage and adaptability of the battery swap station identification, and provides efficient tool support for intelligent planning of battery swap infrastructure.

[0037] Example 1: In order to verify the feasibility of the application in implementation, the application is applied to a certain province's high-speed logistics trunk line along the multiple heavy truck regular operation area as the research object, and the actual operation data of new energy heavy truck is collected and analyzed. More than 50 new energy heavy trucks are deployed in this area for heavy trunk transportation, and these vehicles continuously record the original multi-source behavior data such as geographic position data, speed, battery power, load and time stamp during continuous operation. The battery swap station identification method proposed by the application aims to accurately identify the real existing battery swap station position in actual operation, in order to solve the problem of low identification accuracy and high false positive rate caused by traditional dependence on manual annotation or inefficient clustering means.

[0038] In the actual application process, first, the collected new energy heavy truck behavior data is preprocessed, including removing abnormal GPS drift data, filling in short missing power or load data, and normalizing the speed, power and other numerical values, finally forming a trajectory behavior sequence containing more than 1.2 million records. Based on the time window sliding mechanism, the low speed section and position keeping section appearing in the running process of the vehicle are identified, and about 9000 representative stay behavior segments are further identified. These segments are input into the improved DTW-DBSCAN clustering model, which combines DTW trajectory similarity and spatiotemporal adjacency constraints, successfully removes noise stay points, classifies stay segments with high frequency, stable appearance in the same position and obvious time distribution regularity into the same clustering cluster, and obtains 271 effective stay point clustering clusters.

[0039] Further, the behavior feature vector is constructed by combining the stay time, power drop amplitude and load change of the vehicle in each clustering cluster, and 264 groups of suspected battery swap behavior events are marked by setting threshold judgment rules. After constructing the battery swap behavior event graph, the Yen algorithm is used to mine the top 20 non-repeating shortest paths between the earliest and latest nodes, and the nodes with high frequency in the path are confirmed as key event points representing battery swap behavior. After projecting these event points in space and rasterizing, a heat distribution map is generated, from which the center coordinates of the grid cells with heat intensity higher than 0.75 are extracted, and 22 candidate positions of battery swap stations are obtained.

[0040] Subsequently, based on the historical battery swap behavior data of the 22 candidate points for 6 consecutive months, the event frequency sequence is constructed, and it is mapped to a Poisson process, and the average Poisson probability of each point is calculated. Finally, the average Poisson probability of 17 candidate points exceeds the threshold value of 0.82, indicating that the battery swap behavior of these points has significant periodicity and stability, and they can be identified as real new energy heavy truck battery swap stations. The remaining 5 points are determined as temporary stay areas due to large frequency fluctuations or unstable event occurrence.

[0041] In order to verify the effect of the method, the method is compared with three typical battery swap station identification methods, which are manual rule matching method, K-means spatial clustering method and discrimination method based on LSTM behavior prediction model. The comparison results are shown in the following table:

[0042] Table 1 Comparison table of new energy heavy truck battery swap station identification effect.

[0043]

[0044] As can be seen from Table 1, the method of the application can not only significantly improve the identification accuracy when processing larger scale data, but also effectively reduce the false positive and false negative rates, and has higher computing efficiency. Among them, the improved DTW-DBSCAN clustering integrates the space-time constraint on the basis of maintaining the shape feature of the trajectory, effectively avoiding the problem that the traditional K-means cannot process irregular trajectories; and the Yen algorithm combined with the path frequency analysis of the behavior feature spectrum also significantly enhances the discrimination ability of the battery swap behavior event. The periodic modeling of Poisson distribution further improves the stability verification ability of the candidate battery swap station, and finally realizes the identification target of new energy heavy truck battery swap station with high reliability and high precision.

[0045] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A new energy heavy truck battery swap station recognition method based on deep learning, characterized in that, The application relates to a method for identifying a new energy heavy truck battery swap station. The method comprises the following steps: step 1, collecting original multi-source behavior data generated by a new energy heavy truck in a running process and performing pretreatment to generate a trajectory behavior sequence; step 2, identifying a stay segment of the new energy heavy truck in the trajectory behavior sequence according to a stay point identification rule; step 3, performing time and space constraint density clustering analysis on all the identified stay segments by using an improved DTW-DBSCAN algorithm to obtain a stay point clustering cluster; step 4, constructing a battery swap behavior event graph based on the stay point clustering cluster, obtaining K non-repeated shortest paths of the battery swap behavior event graph by using a Yen algorithm, and determining a battery swap behavior event; step 5, collecting all the geographic position coordinate points corresponding to the battery swap behavior events marked as the battery swap behavior events, and drawing a heat distribution graph to determine a battery swap station candidate position; and step 6, calculating an average Poisson probability corresponding to each battery swap station candidate position based on a Poisson distribution. The original multi-source behavior data comprises geographic position positioning data, vehicle speed data, battery state data, load data and timestamp data; the pretreatment step comprises respectively performing outlier rejection, missing value filling and normalization processing on different types of data in the original multi-source behavior data, and arranging the data in an ascending order of time to form the trajectory behavior sequence. In the step 2, the following steps are specifically performed: a time window of a preset size is constructed for any two adjacent data points in the trajectory behavior sequence, an average speed value in the time window is calculated, if the average speed value is lower than a preset speed threshold value, then a trajectory behavior segment of the corresponding time window size is determined as a low-speed behavior segment; a space moving distance of any two adjacent data points in the trajectory behavior sequence in the time window of the preset size is calculated, if the space moving distance is lower than a space moving distance threshold value, then the trajectory behavior segment of the corresponding time window size is determined as a position keeping segment; cumulative time of the low-speed behavior segment and the position keeping segment in continuous time windows is counted, if the cumulative time exceeds a preset time length threshold value, then the trajectory behavior segment is marked as a stay segment, and a start time, an end time and a duration of the stay segment are recorded. In the improved DTW-DBSCAN algorithm, the following steps are specifically performed:

2. The deep learning-based new energy heavy truck battery swap station identification method according to claim 1, characterized in that, an attribute index table is established for all the stay segments, the attribute index table records a number of each stay segment, corresponding central geographic coordinates and central time; 3. The deep learning-based new energy heavy truck battery swap station identification method of claim 1, characterized in that, a trajectory segment vector is constructed by extracting behavior attributes corresponding to each stay segment, the behavior attributes comprise continuous geographic position coordinates, a stay start time, a stay end time and a stay duration; a DTW distance between any two trajectory segment vectors is calculated, a trajectory similarity matrix between any two stay segments is constructed, each row and each column in the trajectory similarity matrix correspond to a stay segment, each element represents a DTW distance between corresponding trajectory segment vectors, and elements on a diagonal line represent distances between the same stay segments and are fixed as zero; each group of elements in the trajectory similarity matrix is traversed, a trajectory number pair corresponding to the elements is extracted from the attribute index table, and a space distance between the central geographic coordinates of the corresponding two stay segments and a time difference between the central time points are calculated. ​ 4. The new energy heavy truck battery swap station recognition method based on deep learning according to claim 1, characterized in that, ​ ​ ​ ​ ​ If the spatial distance is less than the preset spatial distance threshold and the time difference is less than the preset time window threshold, the values of the two corresponding elements in the trajectory similarity matrix are kept unchanged, otherwise the two corresponding elements are marked as invalid connection, and a space-time adjacency matrix is obtained; The neighborhood radius threshold and the minimum neighborhood point number are set as clustering parameters, and the reachable neighbor point number of each stay segment within the neighborhood radius threshold is calculated according to the valid connection relationship reserved in the space-time adjacency matrix; If the reachable neighbor point number of a stay segment is greater than or equal to the minimum neighborhood point number, the stay segment is marked as a core segment; The stay segments that do not meet the minimum neighborhood point number requirement are not involved in clustering and are removed as noise points; Taking each core segment as a starting point, all adjacent segments having valid connection relationship with the core segment are searched and expanded to form a stay point clustering cluster; Until all stay segments meeting the density clustering requirement are assigned to the corresponding stay point clustering cluster.

5. The new energy heavy truck battery swap station recognition method based on deep learning according to claim 1, characterized in that, The step four is specifically: For each stay point clustering cluster, the corresponding stay start time, stay end time, stay duration, power change amount before and after stay, and load change amount before and after stay are extracted to construct a stay behavior feature vector, which represents a candidate battery replacement behavior event; According to a preset battery replacement behavior judgment rule, it is judged whether the stay behavior feature vector meets the conditions that the power change amount is greater than a set threshold, the stay duration is within a specified time range, and the load change amount is less than a set threshold; A battery replacement behavior event graph is constructed according to all stay behavior feature vectors meeting the battery replacement behavior judgment rule, and the construction step is: The stay behavior feature vectors are sorted in ascending order according to the stay start time; Each stay behavior feature vector is taken as a node of the battery replacement behavior event graph; The Euclidean distance between each two nodes is calculated, and if the Euclidean distance is less than a preset threshold, an edge connection is established between the corresponding two nodes, and the weight of the edge is the corresponding Euclidean distance; Yen algorithm is used to search the path of the battery replacement behavior event graph, and the path searching step is: The first K non-repeating shortest paths between the start node and the end node in the battery replacement behavior event graph are searched, and the start node and the end node are respectively the two nodes with the earliest and latest time points; The first K non-repeating shortest paths are obtained by: Dijkstra path search method is called to start from the start node to find a shortest path to the end node, which is taken as the first shortest path and added to the shortest path list; From the second path, the following steps are executed in turn until there are K non-repeating shortest paths in the shortest path list: Take the last found shortest path as a reference path; Fix the path segment between the start node and each intermediate node in the reference path as a prefix path, and the intermediate node is a node excluding the start node and the end node; Remove the connection edge between the last node of the prefix path and the subsequent node in the battery replacement behavior event graph; A new suffix path to the end node is found by re-executing the Dijkstra path search method from the last node of the prefix path. Combining the suffix path with the prefix path into a new complete path, and adding to the shortest path list.

6. Statistics the frequency of all nodes in the shortest path list, if a node appears more than a set frequency threshold in K paths, it is marked as a battery swap behavior event.

7. The new energy heavy truck battery swap station recognition method based on deep learning according to claim 1, characterized in that, The step five is specifically: Obtain the geographic location coordinate points corresponding to the stay behavior feature vectors marked as battery swap behavior events; Project all geographic location coordinate points into a two-dimensional space map, and construct a regular space grid on the two-dimensional space map, the space grid is to divide the entire geographic area into several equal size space cells; Statistical the number of geographic location coordinate points contained in each space cell, and draw a heat distribution map on the two-dimensional space map, the heat distribution map distinguishes different heat intensity with different colors, the heat intensity is the ratio of the number of geographic location coordinate points contained in the space cell to the number of geographic location coordinate points in the entire geographic area, which represents the density of battery swap behavior events in each space cell; Extract all space cells that exceed the set heat intensity threshold, and take the center position coordinates of each space cell as the candidate location of the battery swap station.

8. The new energy heavy truck battery swap station recognition method based on deep learning according to claim 1, characterized in that, The step six is specifically: Divide the time stamps of all battery swap behavior events according to a set time period for each battery swap station candidate location, and construct multiple continuous time window sequences; Statistical the number of battery swap behavior events in each time window, and form the event frequency sequence corresponding to the battery swap station candidate location; Map the event frequency sequence to the observation sample of the Poisson process, and calculate the occurrence probability under the current Poisson model parameters according to the number of battery swap behavior events in each time window; Summarize the probability values in all time windows, and calculate the average Poisson probability; If the average Poisson probability is greater than a set Poisson probability threshold, it indicates that the corresponding battery swap station candidate location has stable battery swap behavior events in multiple periods, which meets the periodic behavior characteristics, and is identified as a real existing new energy heavy truck battery swap station; If the average Poisson probability is lower than the set Poisson probability threshold, it is determined as a temporary stay behavior point. 9.The deep learning-based new energy heavy truck battery swap station identification system of claim 1, which performs the deep learning-based new energy heavy truck battery swap station identification method of any one of claims 1 to 7, characterized in that, Comprise the following modules: A data acquisition and preprocessing module for collecting original multi-source behavior data generated by new energy heavy trucks during operation, and performing outlier rejection, missing value filling and normalization processing on different types of data respectively, and arranging into trajectory behavior sequence in ascending order of time; A stay segment identification module for identifying low-speed behavior segments, position keeping segments, and determining stay segments; A spatio-temporal constraint density clustering analysis module for performing spatio-temporal constraint density clustering analysis on all stay segments using an improved DTW-DBSCAN algorithm; A battery swap behavior event identification module for constructing a stay behavior feature vector according to each stay point cluster, and using Yen algorithm to obtain the top K shortest paths of the battery swap behavior event graph, and determining the battery swap behavior event; The heat distribution map analysis module is configured to map geographical position coordinates of all battery swap behavior events to a spatial grid, count the number of geographical position coordinate points in each spatial cell, generate a heat distribution map, and extract a spatial cell center position coordinate point with a heat intensity exceeding a set threshold as a battery swap station candidate point. The periodic behavior verification module is configured to count the number of battery swap behavior events of each battery swap station candidate point within a set time period, construct an event frequency sequence, map the event frequency sequence to an observation sample of a Poisson process, calculate a Poisson probability average value in all time windows, determine whether the periodic behavior characteristic is met, and finally confirm a new energy heavy truck battery swap station that actually exists.

Images