Method, Apparatus, and Program for Continuously Learning-Based Vehicle Arrival Time Prediction Using Driving Data

Abstract

A method for predicting a vehicle arrival time based on driving data and continuous learning according to various embodiments of the present invention is disclosed. The method may include: a step of inputting a driver's driving pattern, physical characteristics of a vehicle, and road characteristics into a prediction model to calculate an arrival time correction coefficient, and applying the arrival time correction coefficient to a base arrival prediction time to predict a personalized vehicle arrival time; a step of calculating an error between the actual arrival time of the vehicle and the personalized vehicle arrival time at the time of completion of the vehicle's operation, and storing metadata along with the error; a step of detecting a repeated error pattern; a step of adjusting the weights of the prediction model when a repeated error pattern is detected; and a step of predicting the personalized vehicle arrival time for the next vehicle operation based on the adjusted prediction model. In addition, the present invention is a technology developed through the Seoul Metropolitan Government Seoul Economic Promotion Agency's 2024 Artificial Intelligence Technology Commercialization Support Project (CY240219) 'Development of AI Cargo Control Solution for Real-time Logistics Distribution Optimization'.

Description

Technology Field

[0001] The present invention relates to a driving data-based continuous learning type vehicle arrival time prediction method, apparatus, and program. Specifically, it relates to a method, apparatus, and program that predicts the arrival time by continuously learning the difference between the predicted estimated arrival time and the actual arrival time and correcting the error. Background Technology

[0003] Technology for predicting the Estimated Time of Arrival (ETA) of vehicles is widely utilized in various fields, including traffic guidance, logistics, ride-hailing services, and public transportation scheduling. By calculating the estimated time of arrival for users at their destinations, this ETA prediction enhances the convenience of schedule management and serves as a key factor contributing to operational optimization and improved customer satisfaction in the provision of services such as transportation and delivery.

[0004] Estimated arrival time prediction is generally achieved by combining map-based route search systems with traffic flow data. Prediction algorithms typically calculate the time required for each driving segment by reflecting road distance, speed limits, road types, historical average speeds, traffic volume statistics for specific time periods, and real-time traffic sensor information, and then calculate the total arrival time by accumulating these figures.

[0005] However, existing estimated arrival time prediction systems operate in a fixed manner based on initially trained models and have limitations in that prediction accuracy is not maintained even when traffic patterns, road environments, and driver characteristics change over time. For example, since the model's adaptability to external variables such as increased traffic volume, road construction, seasonal pattern changes, and signal system modifications is poor, a problem arises where prediction errors actually increase as the frequency of use increases.

[0006] Improving the performance of prediction models in existing systems requires retraining the entire model, a process that is performed manually offline, hindering real-time capabilities and operational efficiency. Furthermore, existing models lack structures that effectively reflect the difference between predictions and actual driving results or incorporate cumulative learning; consequently, it is difficult to continuously improve prediction accuracy, particularly in logistics environments requiring real-time route calculation.

[0007] Recently, deep learning-based estimated arrival time prediction models have also emerged, but they often rely on the uniform application of pre-trained models, which may limit their adaptability to error patterns occurring in repeated routes or changes in the dynamic environment.

[0008] Therefore, there is a growing need for a self-learning (online learning-based) architecture for estimated arrival time prediction systems that continuously incorporates actual operational data and can improve its performance autonomously through a feedback loop of prediction, execution, verification, and learning. In particular, the ability to automatically detect and reflect error patterns occurring in repeatedly operated routes and segments represents a technical gap not sufficiently supported by existing systems, requiring a new prediction structure to address this issue.

[0009] Based on this background, the present invention aims to provide a technology that incorporates a feedback loop within an estimated arrival time prediction system, automatically stores and analyzes the error between the predicted value and the actual value at each end of operation, detects sections where error patterns repeat, and improves the prediction model for the corresponding section or condition online.

[0010] In addition, the present invention is a technology developed through the Seoul Metropolitan Government Seoul Economic Promotion Agency's 2024 Artificial Intelligence Technology Commercialization Support Project (CY240219) 'Development of AI Cargo Control Solution for Real-time Logistics Distribution Optimization'. The problem to be solved

[0012] The present invention, devised in response to the aforementioned background technology, aims to provide a driving data-based continuous learning type vehicle arrival time prediction method, apparatus, and program.

[0013] The technical problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0015] According to an embodiment of the present invention for solving the problem described above, a driving data-based continuous learning method for predicting vehicle arrival times is disclosed. The method may include: a step of inputting a driver's driving pattern, physical characteristics of a vehicle, and road characteristics into a prediction model to calculate an arrival time correction coefficient, and applying the arrival time correction coefficient to a basic arrival prediction time to predict a personalized vehicle arrival time; a step of calculating an error between the actual arrival time of the vehicle and the personalized vehicle arrival time at the time of completion of the vehicle's operation, and storing metadata along with the error; a step of detecting a repeated error pattern based on the metadata; a step of adjusting the weights of the prediction model when a repeated error pattern is detected; and a step of predicting a personalized vehicle arrival time based on the adjusted prediction model during the next vehicle operation.

[0016] In one embodiment of the present disclosure, the metadata may include at least one of driver information, vehicle information, route information, and time information.

[0017] In one embodiment of the present disclosure, the step of detecting a repeated error pattern based on metadata may include: obtaining error data accumulated more than a certain number of times under the same operating conditions; calculating the average value and variance value of the error from the accumulated error data; and detecting a repeated error pattern when the average value of the error is greater than or equal to a preset threshold and the variance value of the error is less than or equal to a preset reference value.

[0018] In one embodiment of the present disclosure, the same operating conditions may be characterized as having the same driver, the same vehicle, and the same route conditions.

[0019] In one embodiment of the present disclosure, the prediction model may include a fusion module and a regression prediction module.

[0020] In one embodiment of the present disclosure, when a repeated error pattern is detected, the step of adjusting the weights of the prediction model that calculates the arrival time correction coefficient may be characterized by adjusting the weights of the regression prediction module.

[0021] In one embodiment of the present disclosure, the weight adjustment of the regression prediction module can be performed by keeping the encoder layer, which is the front block, fixed and performing fine-tuning only on the rear block to adjust the arrival time correction coefficient.

[0022] In one embodiment of the present disclosure, when a repeated error pattern is detected, the step of adjusting the weights of the prediction model that calculates the arrival time correction coefficient may be characterized by adjusting the weights of the fusion module.

[0023] In one embodiment of the present disclosure, the weight adjustment of the fusion module may adjust the weights between the corresponding input feature attentions within the fusion module to readjust the degree of reflection of the interrelationships between the input features.

[0024] According to one embodiment of the present invention for solving the above-described problem, an apparatus is disclosed. The apparatus comprises: a memory for storing one or more instructions; and a processor for executing the one or more instructions stored in the memory, and the processor can perform the above-described methods by executing the one or more instructions.

[0025] According to one embodiment of the present invention for solving the above-described problem, a computer program stored on a recording medium readable by a computer is disclosed, which is combined with a computer, which is hardware, to perform the above-described methods.

[0026] Other specific details of the present invention are included in the detailed description and drawings. Effects of the invention

[0028] The present invention can continuously improve the accuracy of vehicle arrival time prediction by embedding a feedback loop in the arrival time prediction system and performing real-time online learning.

[0029] The effects of the present invention are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below. Brief explanation of the drawing

[0031] FIG. 1 is a drawing illustrating a system according to one embodiment of the present invention. FIG. 2 is a hardware configuration diagram of a computing device according to one embodiment of the present invention. FIGS. 3 to 6 are drawings for explaining a personalized vehicle arrival time prediction method according to a first embodiment of the present invention. FIGS. 7 and 8 are drawings for explaining a driving data-based continuous learning type vehicle arrival time prediction method according to a second embodiment of the present invention. FIG. 9 is a schematic diagram showing one or more network functions related to an embodiment of the present invention. Specific details for implementing the invention

[0032] Various embodiments are now described with reference to the drawings. In this specification, various descriptions are provided to facilitate an understanding of the invention. However, it is evident that these embodiments can be practiced without such specific descriptions.

[0033] As used herein, terms such as “component,” “module,” “system,” etc. refer to computer-related entities, hardware, firmware, software, combinations of software and hardware, or executions of software. For example, a component may be, but is not limited to, a procedure executed on a processor, a processor, an object, an execution thread, a program, and / or a computer. For example, both an application executed on a computing device and the computing device itself may be a component. One or more components may reside within a processor and / or an execution thread. A component may be localized within a single computer. A component may be distributed among two or more computers. Additionally, these components may be executed from various computer-readable media having various data structures stored therein. Components may communicate through local and / or remote processes, for example, according to signals having one or more data packets (e.g., data from a component interacting with another component in a local system or distributed system, and / or data transmitted through signals to other systems and networks such as the Internet).

[0034] Furthermore, the term "or" is intended to mean an implicit "or" rather than an exclusive "or." That is, unless otherwise specified or evident from the context, "X uses A or B" is intended to mean one of the natural implicit substitutions. In other words, if X uses A; if X uses B; or if X uses both A and B, "X uses A or B" may apply to any of these cases. Additionally, the term "and / or" as used herein should be understood to refer to and include all possible combinations of one or more of the enumerated related items.

[0035] Additionally, the terms “comprising” and / or “comprising” should be understood to mean that such features and / or components are present. However, the terms “comprising” and / or “comprising” should be understood not to exclude the presence or addition of one or more other features, components and / or groups thereof. Furthermore, unless otherwise specified or clearly evident from the context to indicate a singular form, the singular in this specification and claims should generally be interpreted to mean “one or more.”

[0036] Those skilled in the art should recognize that the various exemplary logical blocks, configurations, modules, circuits, means, logics, and algorithmic steps described in connection with the embodiments disclosed herein may be implemented in electronic hardware, computer software, or a combination of both. To clearly exemplify the interchangeability of hardware and software, various exemplary components, blocks, configurations, means, logics, modules, circuits, and steps have been generally described above in terms of their functionality. Whether such functionality is implemented in hardware or software depends on the specific application and design constraints imposed on the overall system. Skilled technicians may implement the described functionality in various ways for each specific application. However, such decisions regarding implementation should not be interpreted as moving out of the scope of the invention.

[0037] The description of the presented embodiments is provided to enable those skilled in the art to use or practice the present invention. Various modifications to these embodiments will be apparent to those skilled in the art. The general principles defined herein may be applied to other embodiments without departing from the scope of the present invention. Thus, the present invention is not limited to the embodiments presented herein. The present invention should be interpreted in the broadest possible scope consistent with the principles and novel features presented herein.

[0038] In this specification, the term "computer" refers to any type of hardware device comprising at least one processor, and may be understood to include software configurations operating on said hardware device according to the embodiments. For example, the term "computer" may be understood to include smartphones, tablet PCs, desktops, laptops, and user clients and applications running on each of these devices, but is not limited thereto.

[0039] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0040] Each step described in this specification is described as being performed by a computer, but the subject of each step is not limited thereto, and depending on the embodiment, at least some of each step may be performed on different devices.

[0042] FIG. 1 is a drawing illustrating a system according to one embodiment of the present invention.

[0043] Referring to FIG. 1, a system according to one embodiment of the present invention may include a computing device (100), a user terminal (200), and an external server (300). The system illustrated in FIG. 1 is according to one embodiment, and its components are not limited to the embodiment illustrated in FIG. 1 and may be added, changed, or deleted as needed.

[0044] In one embodiment, the computing device (100) can perform a driving data-based continuous learning-type vehicle arrival time prediction. For example, the computing device (100) can perform a series of processes such as identifying the driver and the vehicle, calculating arrival time correction coefficients and predicting the final arrival time, calculating the error with the actual arrival time, detecting repeated error patterns, and adjusting the weights of the prediction model.

[0045] Specifically, the computing device (100) inputs the driver's driving pattern, the physical characteristics of the vehicle, and the road characteristics into a prediction model to calculate an arrival time correction coefficient, and can predict a personalized vehicle arrival time by applying the arrival time correction coefficient to a basic arrival prediction time.

[0046] Additionally, the computing device (100) can calculate the error between the actual arrival time of the vehicle and the personalized vehicle arrival time at the time the vehicle's operation is completed, and store metadata along with the error.

[0047] Additionally, the computing device (100) can detect repeated error patterns based on metadata, and if repeated error patterns are detected, it can adjust the weights of the prediction model.

[0048] Accordingly, the computing device (100) of the present invention can provide a driving data-based continuous learning type vehicle arrival time prediction method with improved real-time performance and operational efficiency compared to existing arrival time prediction systems by calculating the error between the predicted personalized vehicle arrival time and the actual arrival time, detecting repeated error patterns, and adjusting the weights of the prediction model in real-time while online.

[0049] Hereinafter, an example of a method in which a computing device (100) performs a driving data-based continuous learning type vehicle arrival time prediction is described with reference to FIGS. 5 to 7.

[0050] In various embodiments, the computing device (100) may provide Web or Application-based services. However, it is not limited thereto.

[0051] The computing device (100) may include any type of computer system or computer device, such as, for example, a microprocessor, a mainframe computer, a digital processor, a portable device, and a device controller. However, it is not limited thereto.

[0052] Hereinafter, the hardware configuration of the computing device (100) will be described with reference to FIG. 2.

[0053] Meanwhile, the user terminal (200) may be connected to the computing device (100) via the network (400) and may be a user terminal that uses a driving data-based continuous learning type vehicle arrival time prediction function provided by the computing device (100). For example, the user terminal (200) may include a vehicle driver's terminal.

[0054] Here, the user terminal (200) may include, for example, various types of computer devices. Specifically, for example, the user terminal (200) may refer to various terminal devices such as smartphones, tablet PCs, desktops, and laptops.

[0055] The user terminal (200) includes a display on at least a part of the terminal and may include an operating system for running applications or extension-based services provided by the computing device (100). For example, the user terminal (200) may be a smartphone, but is not limited thereto, and the user terminal (200) may include all kinds of handheld-based wireless communication devices such as navigation, PCS (Personal Communication System), GSM (Global System for Mobile communications), PDC (Personal Digital Cellular), PHS (Personal Handyphone System), PDA (Personal Digital Assistant), IMT (International Mobile Telecommunication)-2000, CDMA (Code Division Multiple Access)-2000, W-CDMA (W-Code Division Multiple Access), Wibro (Wireless Broadband Internet) terminal, smartpad, tablet PC, etc., as wireless communication devices that ensure portability and mobility.

[0056] An external server (300) can be connected to a computing device (100) via a network (400), and can transmit and receive various information / data necessary for the computing device (100) to perform a driving data-based continuous learning type vehicle arrival time prediction, and can store and manage various information / data generated as the computing device (100) performs the driving data-based continuous learning type vehicle arrival time prediction.

[0057] For example, the external server (300) may be a database server that stores information used in the driving data-based continuous learning vehicle arrival time prediction method. As another example, the external server (300) may be a server that provides information used in the driving data-based continuous learning vehicle arrival time prediction method.

[0058] The network (400) may refer to a connection structure capable of exchanging information between each node, such as computing devices, multiple terminals, and servers. For example, the network (400) includes a Local Area Network (LAN), a Wide Area Network (WAN), the World Wide Web (WWW), a wired and wireless data network, a telephone network, a wired and wireless television network, etc.

[0059] Wireless data communication networks include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth network, NFC (Near-Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc.

[0061] FIG. 2 is a hardware configuration diagram of a computing device according to one embodiment of the present invention.

[0062] Referring to FIG. 2, a computing device (100) according to one embodiment of the present invention may include one or more processors (110), a memory (120) for loading a computer program (151) executed by the processor (110), a bus (130), a communication interface (140), and a storage (150) for storing the computer program (151). Here, FIG. 2 illustrates only the components related to the embodiment of the present invention. Therefore, a person skilled in the art to which the present invention pertains will understand that other general-purpose components may be included in addition to the components illustrated in FIG. 2.

[0063] The processor (110) controls the overall operation of each component of the computing device (100). The processor (110) may be composed of one or more cores and may include processors for data analysis and deep learning, such as a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), or a tensor processing unit (TPU) of the computing device. Alternatively, it may be configured to include any type of processor well known in the art of the present invention.

[0064] Additionally, the processor (110) can perform operations for at least one application or program for executing the method according to embodiments of the present invention, and the computing device (100) may have one or more processors.

[0065] In various embodiments, the processor (110) may further include RAM (Random Access Memory, not shown) and ROM (Read-Only Memory, not shown) for temporarily and / or permanently storing signals (or data) processed within the processor (110). Additionally, the processor (110) may be implemented in the form of a system-on-chip (SoC) comprising at least one of a graphics processing unit, RAM, and ROM.

[0066] Memory (120) stores various data, instructions and / or information. Memory (120) may load a computer program (151) from storage (150) to execute a method / operation according to various embodiments of the present invention. When the computer program (151) is loaded into memory (120), the processor (110) may perform the method / operation by executing one or more instructions constituting the computer program (151). Memory (120) may be implemented as volatile memory such as RAM, but the technical scope of the present invention is not limited thereto.

[0067] The bus (130) provides communication functions between components of the computing device (100). The bus (130) can be implemented as various types of buses, such as an address bus, a data bus, and a control bus.

[0068] The communication interface (140) supports wired and wireless internet communication of the computing device (100). Additionally, the communication interface (140) may support various communication methods other than internet communication. To this end, the communication interface (140) may be configured to include a communication module well known in the art of the present invention. In some embodiments, the communication interface (140) may be omitted.

[0069] Storage (150) can store a computer program (151) non-temporarily. When performing a process according to an embodiment of the present invention through a computing device (100), storage (150) can store various information necessary to perform a method according to the disclosed embodiment or to provide a service.

[0070] The storage (150) may be configured to include non-volatile memory such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium well known in the art to which the present invention belongs.

[0071] A computer program (151) may include one or more instructions that cause a processor (110) to perform a method / operation according to various embodiments of the present invention when loaded into memory (120). That is, the processor (110) may perform the method / operation according to various embodiments of the present invention by executing the one or more instructions.

[0072] In one embodiment, the computer program (151) may include one or more instructions to perform various methods related to various tasks related to learning a neural network model.

[0073] The steps of the method or algorithm described in connection with embodiments of the present invention may be implemented directly in hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. The software module may reside in RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), Flash Memory, a hard disk, a removable disk, a CD-ROM, or any form of computer-readable recording medium well known in the art to which the present invention belongs.

[0074] The components of the present invention may be implemented as a program (or application) and stored on a medium to be executed in combination with a computer, which is hardware. The components of the present invention may be implemented as software programming or software elements, and similarly, embodiments may be implemented in programming or scripting languages ​​such as C, C++, Java, assembler, etc., including various algorithms implemented as combinations of data structures, processes, routines, or other programming configurations. Functional aspects may be implemented as algorithms executed on one or more processors.

[0076] FIGS. 3 and 4 are drawings for explaining a personalized vehicle arrival time prediction method according to a first embodiment of the present invention.

[0077] Referring to FIGS. 3 and 4, the computing device (100) can obtain identification information and path information for identifying each of the driver and the vehicle driven by the driver (S110).

[0078] Here, the identification information may include identification values ​​used to distinguish the unique entities of drivers and vehicles, and to search for relevant data or input it into a model based on this distinction. For example, in the case of a driver, this may be a user account ID, driver's license number, mobile app integration identifier (e.g., UUID), or an authentication token from a user terminal associated with the vehicle. In the case of a vehicle, this may include a vehicle registration number, VIN (Vehicle Identification Number), a unique identifier received from a vehicle terminal (OBD, TCU, etc.), or a unique key value received from sensors and communication modules attached to the vehicle. This identification information can be linked with a pre-registered user and vehicle database and utilized as a reference value to extract driving history, specifications, and driving style analysis results for the corresponding driver or vehicle.

[0079] Furthermore, route information is a concept that includes spatial data related to a vehicle's driving plan, and may include not only simple origin and destination coordinates but also a route sequence composed of specific road segments connecting them. Route information may include a list of road segments generated from map-based route search results, the distance of each segment, speed limit, gradient, road type (e.g., expressway, general road, two-lane road, etc.), intersection density, real-time traffic flow information, and traffic event information such as accidents or construction; this information can be collected through TPEG, VDS, DSRC, GPS-based real-time traffic APIs, or route search engines. Additionally, route information is represented as a continuous sequence of road units with a time-series structure rather than a simple connection between points, and can be utilized as input for subsequent road characteristic analysis and the calculation of basic estimated arrival times.

[0080] Specifically, the computing device (100) can receive a driver ID and a vehicle ID through integration with a user's mobile terminal, a vehicle's OBD device, or a server, and then retrieve driver and vehicle profiles corresponding to the identifiers from a server or local storage. Additionally, regarding route information, when a user sets a destination, the computing device (100) can calculate an optimal route between a starting point and a destination through a route search engine integrated with a map API or a navigation system, and accordingly obtain static and dynamic information regarding road sections that constitute the route.

[0081] For example, the computing device (100) can obtain detailed route data including real-time average speed, road congestion level, elevation change amount, etc. for the route from an external server (e.g., Google Maps API, HERE API, Tmap) or a self-developed traffic information system.

[0082] These identification and route information can be used in subsequent stages to reflect driver and vehicle characteristics, analyze road conditions, and calculate arrival time correction coefficients.

[0083] In one embodiment, the computing device (100) can recognize the driving pattern of the driver and the physical characteristics of the vehicle based on identification information, and recognize the road characteristics of a plurality of road sections included in the driving path based on path information (S120).

[0084] In one embodiment, the computing device (100) can acquire driving history data of a driver based on a driver identifier included in identification information and generate a driver embedding vector based on the driving history data.

[0085] Specifically, the computing device (100) can arrange driving history data in a time series. Here, the driving history data may consist of detailed items such as the average speed per driving session, the ratio of driving speed to the speed limit, the frequency of occurrence of sudden acceleration or sudden deceleration events, stopping and restarting patterns, the time taken to pass through an intersection, the trend of speed change by driving time period, and the signal disregard rate. This time series data can be preprocessed to reflect the dynamic driving characteristics of the driver and input into a transformer-based encoder.

[0086] Additionally, the computing device (100) can input driving history data arranged in a time series into a transformer-based encoder to extract the driving characteristics of the driver. Here, the transformer-based encoder learns the interrelationships between driving behaviors that change over time through a self-attention mechanism and can extract a feature vector representing the driver's unique driving style.

[0087] And, the computing device (100) can generate a driver embedding vector based on driving characteristics.

[0088] More specifically, the computing device (100) can filter only driving records that satisfy certain conditions from driving logs of a preset period (e.g., the last 30 days) and use them as targets for analysis, and if there are missing items, it can apply a missing value correction technique and then perform embedding processing. In addition, the computing device (100) can construct an input vector in a state where noise is removed by sampling at the same time unit or by applying an outlier detection algorithm to ensure the stability of the time series analysis.

[0089] Additionally, the computing device (100) can obtain vehicle specification data based on a vehicle identifier included in identification information and generate a vehicle embedding vector based on the specification data.

[0090] Here, the vehicle specification data may consist of detailed items such as vehicle weight, engine output, load condition, and drive mode. Specifically, the vehicle specification data may include physical factors that affect responsiveness to road conditions, such as the vehicle's acceleration performance, climbing ability, braking distance, and fuel efficiency.

[0091] And, the computing device (100) can normalize the vehicle's specification data and input it into a multilayer perceptron-based embedding network to generate a vehicle embedding vector.

[0092] For example, a computing device (100) can first select data such as whether it is an electric vehicle, curb weight, maximum load capacity, output / torque curve, drivetrain configuration, etc., and then, depending on the data type, convert continuous values ​​into normalized forms and categorical values ​​into one-hot or embedding forms. Subsequently, a multilayer perceptron learns weights to reflect the relative influence of each specification on vehicle performance, and the output layer can output a fixed-dimensional vector summarizing vehicle performance.

[0093] Additionally, the computing device (100) can acquire road characteristic data for each of the plurality of road sections based on the path information and generate road embedding vectors for each of the plurality of road sections based on the road characteristic data.

[0094] Specifically, the computing device (100) can sort road characteristic data for each of a plurality of road sections in a path order. Then, the computing device (100) can input the road characteristic data sorted in a path order into a transformer encoder to generate road embedding vectors for each of the plurality of road sections.

[0095] Here, road characteristic data includes road length, speed limit, road type, number of lanes, and real-time traffic congestion for each section, and the computing device (100) can vectorize this data to construct an input sequence. Then, the transformer encoder can learn the characteristics of each section and the mutual influence with adjacent sections to output a road embedding vector that reflects the overall flow and difficulty of the road conditions. This road embedding vector can subsequently be used as an input value for personalized prediction.

[0096] As a result, since driver, vehicle, and road embedding vectors are learned independently, even if a new driver or vehicle is added, only the corresponding profile needs to be embedded to be integrated into the prediction model.

[0097] In one embodiment, the computing device (100) can calculate an arrival time correction coefficient based on the driver's driving pattern, the physical characteristics of the vehicle, and the road characteristics (S130).

[0098] Specifically, the computing device (100) inputs a driver embedding vector, a vehicle embedding vector, and a plurality of road embedding vectors for each road section into a pre-trained prediction model to obtain an arrival time correction coefficient from the prediction model.

[0099] Here, the prediction model can consist of a fusion module and a regression prediction module.

[0100] The prediction model of the present invention receives driver embedding vectors, vehicle embedding vectors, and road embedding vectors for multiple road sections as inputs, integrates the relationships between them through a multi-head attention-based fusion module, and can calculate arrival time correction coefficients based on this.

[0101] Specifically, the prediction model receives driver, vehicle, and road embedding vectors as inputs, respectively, and can learn the relationships between the driver, vehicle, and road conditions through a multi-head attention structure while maintaining their individual characteristics. Here, the fusion module reflects the combined effects between each element through attention heads from various perspectives and can generate an integrated representation vector by assigning weights to input elements with high importance under specific conditions.

[0102] As a result, the fusion module highlights important features by cross-referencing each embedding vector with other embeddings through attention. For example, if a high attention value is measured between a driver embedding vector and a specific road segment embedding, the model can detect that the driver behaves particularly differently on that type of road segment.

[0103] For example, if the attention score between the urban road section and the driver embedding is relatively high, and the vehicle is a large truck that significantly affects the estimated arrival time during acceleration and deceleration within the city, a high weight is assigned to the urban section embedding. Consequently, the generated integrated representation vector is configured to emphasize the characteristics of the driver and vehicle when driving in the urban section. The fusion module of the present invention learns meaningful interactions between input elements in this manner, thereby supporting the calculation of a more precise arrival time correction coefficient in a subsequent step.

[0104] Additionally, the prediction model can input the integrated representation vector into the regression prediction module to calculate the arrival time correction coefficient to be applied to the base arrival time. To explain further, the prediction model receives the integrated representation vector through a multilayer perceptron-based regression prediction module and outputs an arrival time correction coefficient (f) in the form of a real value; this arrival time correction coefficient can be used to calculate the final personalized vehicle arrival time (ETA_pred) by multiplying it by the route-based base arrival time (ETA_base).

[0105] For example, if the driver characteristics include a large deviation in average driving speed and a distinct tendency for high-speed driving, the vehicle characteristics include a lightweight, high-output sedan, and the road characteristics consist mainly of highways, the prediction model determines that this combination is likely to result in an actual arrival time faster than the estimated time based on the basic route, and can calculate an arrival time correction factor of 0.85. This is an example of numerically reflecting the potential for time reduction based on the driver-vehicle-route combination, based on results learned from similar past cases.

[0106] As another example, if the driver exhibits a conservative driving pattern with a high rate of driving below the speed limit and few sudden accelerations or decelerations, the vehicle is a large van with heavy cargo, and the road environment consists mainly of congested urban areas with dense signal intersections, the prediction model can calculate an arrival time correction factor of 1.25 by reflecting the interaction of complex factors. In other words, the prediction model can calculate an arrival time correction factor by reflecting the possibility that the actual arrival time will be delayed compared to the base arrival prediction time.

[0107] As such, the prediction model of the present invention is designed to uniquely reflect each input information and to integrally interpret the relationships between different information such as the driver, vehicle, and road environment, thereby generating personalized vehicle arrival prediction results that closely approximate real driving situations.

[0109] In one embodiment, the computing device (100) can calculate a base arrival prediction time based on road characteristics (S140).

[0110] First, the computing device (100) can acquire road characteristic data for each of a plurality of road sections, including at least one of information such as road length, speed limit, road type, and real-time traffic congestion.

[0111] Specifically, the computing device (100) can obtain static information (e.g., speed limit, road length, number of lanes, road type) and dynamic information (e.g., average driving speed, traffic density, accident or construction information, etc.) for each road segment constituting the route search result by linking with a map database and a real-time traffic information provision system. This information can be obtained by API calls, DSRC or TPEG data received from a vehicle terminal, or by periodically synchronizing from a cloud-based traffic infrastructure, and can be stored in the form of road characteristic vectors organized by segment.

[0112] For example, if a specific path consists of 5 sections, the computing device (100) can construct a vector including the distance (e.g., 1.2 km), speed limit (e.g., 60 km / h), road type (e.g., 4 lanes in each direction), and real-time traffic congestion level (e.g., congestion level 2) for each section, and perform subsequent time calculation operations based thereon.

[0113] Additionally, the computing device (100) can calculate the time to pass each of a plurality of road sections based on road characteristic data. And, the computing device (100) can calculate a basic arrival prediction time by accumulating the time to pass.

[0114] Specifically, the computing device (100) can calculate the average driving time for each of a plurality of road sections based on the speed limit and section length included in the road characteristic data. Additionally, the computing device (100) can calculate the passing time for the road section by applying a traffic correction coefficient corresponding to real-time traffic congestion to the average driving time.

[0115] More specifically, the computing device (100) can calculate a basic average driving time based on speed limit and section length data for each road section. For example, the computing device (100) can calculate a basic average driving time by dividing the section length by the speed limit. Subsequently, the computing device (100) can apply a traffic correction factor to the average driving time that reflects the actual traffic conditions of each road section based on real-time traffic congestion information. The traffic correction factor is set as a quantified value according to the level of traffic congestion, for example, 1.0 when traffic is smooth, 1.2 for normal congestion, and 1.5 or higher for congestion, so that the predicted passing time can vary depending on the situation even under the same distance and speed limit conditions.

[0116] For example, if a specific road section is 1.5 km long and the speed limit is 60 km / h, the computing device (100) can calculate the reference average travel time as 1.5 km ÷ 60 km / h = 0.025 hours (= 1.5 minutes). Then, if the computing device (100) determines that the traffic information received in real-time for the section is congestion level 2 (e.g., moderate congestion), it multiplies by a predefined traffic correction factor (e.g., 1.3) to calculate the actual travel time as 1.5 minutes × 1.3 = 1.95 minutes. In this way, the computing device (100) can calculate the base estimated time of arrival (ETA_base) for the entire route by repeating the same calculation for all road sections on the route and accumulating the travel time for each section.

[0117] Additionally, when calculating the basic arrival prediction time, the computing device (100) may apply a separate correction factor or heuristic model to a specific section type by considering the specific characteristics of urban sections, such as connecting signals between multiple road sections, intersection waiting times, and high probability of congestion. For example, for sections where signal waiting occurs repeatedly, the computing device (100) may reflect the average waiting time according to the signal cycle or apply an average value by learning the distribution of additional travel time in the section from statistically analyzed past driving data. Through this, the computing device (100) can calculate a more realistic basic arrival prediction time by reflecting dynamic and contextual information rather than relying solely on static road information.

[0118] In one embodiment, the computing device (100) can predict a personalized vehicle arrival time by applying an arrival time correction factor to a base arrival prediction time (S150).

[0119] Specifically, the computing device (100) can calculate a personalized vehicle arrival time (ETA_Pred) by multiplying or adding an arrival time correction factor f, which integrally reflects driver characteristics, vehicle characteristics, and road characteristics, to a base arrival time prediction time (ETA_Base) calculated based on road characteristic data. In this case, the arrival time correction factor is a real value and is a quantified value based on the interaction with the driver's driving style (e.g., actual average speed relative to the speed limit), vehicle performance (e.g., acceleration ability based on load condition or output), and road characteristics.

[0120] For example, the computing device (100) can calculate ETA_PRED = 30 minutes × 1.1 = 33 minutes when the base arrival prediction time is 30 minutes and the arrival time correction factor f=1.1. As another example, the computing device (100) can calculate the personalized vehicle arrival time to be reduced to 27 minutes when the driver drives faster than average and the vehicle performance is excellent, so f=0.9. This method can provide high prediction accuracy in terms of driver customization and realism compared to existing systems that use only static route-based prediction times.

[0121] Accordingly, the computing device (100) of the present invention is not limited to simple route distance or average speed-based prediction, but calculates the arrival time by integratively considering driving habits of each driver, the actual performance of the vehicle, and characteristics of each road section, thereby providing personalized arrival time prediction information with minimized error even in various driving environments. This function enhances utility value in application fields where the reliability of real-time arrival information is important, such as logistics delivery, mobility services, and car-sharing platforms, and can simultaneously improve user satisfaction and system efficiency.

[0122] According to an additional embodiment of the present invention, the computing device (100) may include a learning structure based on fusion with external factors in addition to driver, vehicle, and road information to further improve the precision and adaptability of driving prediction.

[0123] Specifically, the computing device (100) can collect external data such as weather information (temperature, precipitation, visibility, etc.), time zone information (commute time, whether driving at night, etc.), and event information (construction zones, accident-prone areas, etc.), in addition to existing driver driving history, vehicle specifications, and road characteristics, and include this data in the analysis target. Through this, it is possible to accurately reflect the fact that even with the same driver and the same vehicle, driving patterns may change depending on the time zone or weather conditions.

[0124] More specifically, the computing device (100) may not process these external data as fixed input variables, but rather vectorize them through a separate encoder module and include them in an integrated representation along with existing driver embeddings, vehicle embeddings, and road embeddings. For example, a weather encoder can learn how changes in the presence and intensity of precipitation interact with driver deceleration tendencies or vehicle braking distances, and a time zone encoder can determine whether a particular driver tends to drive more conservatively at night.

[0125] For example, the computing device (100) can calculate a personalized arrival time that reflects time-based characteristics by applying different time coefficients to the prediction results for departures at 8:00 AM and 9:00 PM, even if the same vehicle and the same route are used. Alternatively, if a specific driver's deceleration rate tends to be significantly higher than usual on rainy days, this can be learned and reflected in the arrival time correction coefficient to reduce the prediction error with respect to the actual arrival time.

[0126] Accordingly, the computing device (100) can achieve multidimensional personalized predictions that go beyond simple driving history-based correction predictions and reflect time, environment, and event-centered contextual information, thereby enabling more precise and reliable arrival time predictions even under complex driving conditions.

[0128] FIGS. 5 to 7 are drawings for explaining a driving data-based continuous learning type vehicle arrival time prediction method according to a second embodiment of the present invention.

[0129] Referring to FIGS. 5 and 6, the computing device (100) inputs the driver's driving pattern, the physical characteristics of the vehicle, and the road characteristics into a prediction model to calculate an arrival time correction coefficient, and can predict a personalized vehicle arrival time by applying the arrival time correction coefficient to a basic arrival prediction time (S210).

[0130] Here, an example of a personalized vehicle arrival time prediction method was described above with reference to FIGS. 3 to 6.

[0131] In one embodiment, the computing device (100) can calculate the error between the actual arrival time of the vehicle and the personalized vehicle arrival time at the time the vehicle's operation is completed, and store metadata along with the error (S220).

[0132] To explain further, the time difference between the vehicle's actual arrival time and the personalized vehicle arrival time predicted prior to the trip can be calculated as an error value at the time the vehicle's trip is completed.

[0133] The completion of a trip can be recognized, for example, when a vehicle arrives at its final destination and the location is confirmed based on the location, or when a specific event occurs on a transportation management application, such as delivery completion or trip termination.

[0134] The error value calculated at this time is determined as a time difference, and metadata may be stored along with the error value. Here, the metadata may include driver information (driver ID, driver's driving history identifier, etc.), vehicle information (vehicle ID, vehicle type, vehicle loading status, etc.), route information (origin, destination, major waypoints, etc.), time information (departure time, arrival time, date and day of the week of operation, etc.), weather information (weather conditions for the relevant time period, etc.), road characteristic information (road type, traffic pattern, etc.), and other related information that may affect operation (accident information, etc.).

[0135] These error values ​​and metadata are recorded in a training data repository linked to the prediction model and can be used in subsequent steps to detect recurring error patterns and adjust the prediction model weights online.

[0137] In one embodiment, the computing device (100) can detect a repeated error pattern based on metadata (S230).

[0138] Referring to FIG. 7, the computing device (100) can obtain error data accumulated more than a certain number of times under the same operating conditions to detect repeated error patterns based on metadata (S231). The average value and variance value of the error can be calculated from the error data (S232). If the average value of the error is greater than or equal to a threshold and the variance value of the error is less than or equal to a reference value, it can be detected as a repeated error pattern (S233).

[0139] First, driving data corresponding to identical driving conditions is accumulated and stored. At this time, identical driving conditions refer to a set of driving where the driver identifier, vehicle identifier, and driving route identifier are identical, and whether the conditions match can be determined based on driving metadata within the system.

[0140] Subsequently, the system can determine whether a specific number of operations (e.g., 10 or more) have been accumulated among the sets of operations corresponding to the same operation conditions.

[0141] When operation data exceeding a certain number of times is secured, error data between the estimated arrival time and the actual arrival time of the vehicle can be extracted from the corresponding operation set, and the average and variance values ​​of the error can be calculated.

[0142] If the average value of the calculated error is greater than or equal to a set threshold (e.g., 10% of the personalized estimated time of arrival) and the variance of the error is less than or equal to a set standard (e.g., the standard deviation value is less than or equal to 20% of the average value of the error), that is, if the error shows a continuous and stable pattern above a certain level, it can be detected that a repeated error pattern exists for the condition.

[0143] By applying these judgment criteria, errors caused by temporary or exceptional operating conditions (e.g., sudden accidents, one-time construction, etc.) can be filtered out, and only structurally recurring errors under specific operating conditions can be identified.

[0144] If a repeated error pattern is detected, weight adjustment of the prediction model may be performed in a subsequent step.

[0145] In one embodiment, the computing device (100) can adjust the weights of the prediction model when a repeated error pattern is detected (S240).

[0146] The above prediction model may include a fusion module and a regression prediction module.

[0147] Hereinafter, a detailed description of the fusion module and the regression prediction module has been given with reference to FIGS. 3 and FIGS. 4. When a repeated error pattern according to the present invention is detected, the step of adjusting the weights of the prediction model may be configured to adjust the weights of at least one of the fusion module and the regression prediction module.

[0148] In one embodiment, the computing device (100) performs fine-tuning targeting only the rear block of the regression prediction module, while keeping the encoder layer, which is the front block, fixed and adjusting the arrival time correction coefficient.

[0149] In the present invention, the regression prediction module can be largely composed of a front-end block and a back-end block.

[0150] Among these, the shear block serves as an encoder layer that transforms input data into high-dimensional feature vectors and primarily performs the role of extracting meaningful representations from time series or context information. Since this shear block contributes to the learning of generalized representations and plays a key role in overall model performance, the weights of the corresponding shear block are maintained in a fixed state even during the local correction process for repeated error patterns in this invention. Specifically, it is configured so that backpropagation to the encoder layer does not occur during training by freezing the gradient.

[0151] Meanwhile, the rear block of the regression prediction module may be composed of a fully connected layer or an output layer that calculates the final predicted value based on the feature vector transmitted from the front block. In this invention, when repeated error patterns are detected in data with specific metadata, instead of retraining the entire model, fine-tuning is performed only on the rear block to ensure computational efficiency and is designed to be suitable for real-time or near-real-time online learning environments. This structure has the advantage of improving prediction accuracy by introducing local correction coefficients while maintaining the representation learning capabilities of existing encoders.

[0152] For example, it can be assumed that in a specific urban section (section A to section B), the same driver P repeatedly experiences a pattern of estimated arrival time delays of about 5 minutes or more during the same time period (8:00 AM to 9:00 AM). After detecting the error pattern, the system can construct a feedback dataset filtered by the conditions of section A, driver P, and time period (8:00 AM to 9:00 AM).

[0153] Subsequently, training can be performed only on the rear block of the regression prediction module using the dataset. In this case, the encoder layer, which is the front block, remains fixed, and as the weights of the rear block are updated, it becomes possible to predict the estimated time of arrival by reflecting correction coefficients specialized for the corresponding section, driver, and time period. As a result, during the next operation, an estimated time of arrival is predicted with a delay of approximately 5 minutes compared to the previous one under the same conditions, and as operations are repeated, the results of fine-tuning accumulate, enabling more precise predictions.

[0154] This method offers the following advantages. First, it allows for the rapid incorporation of local error correction while maintaining the existing model structure and the learning capabilities of key representations. Second, due to the low computational load, online learning can be applied in real-time or near-real-time. Third, when repeated error patterns are limited to a few conditions (such as specific routes or drivers), local precision can be improved without compromising the generalization of overall model performance.

[0155] In one embodiment, the computing device (100) can adjust the attention weights within the fusion module to readjust the degree of reflection of the interrelationships between input features.

[0156] If repeated error patterns are detected, weight adjustments can be performed on the fusion module. This weight adjustment can be implemented by readjusting the corresponding attention weights. In other words, the importance and the degree of reflection of interrelationships between input features can be readjusted; this can be particularly effective when the cause of the error stems from the inaccurate reflection of interactions between specific input features.

[0157] For example, it can be assumed that when the same truck (fixed vehicle ID) travels the same delivery route, a pattern is detected in which the estimated arrival time error occurs repeatedly depending on the cargo load. Specifically, when the same vehicle is driven empty (empty load state), it showed excellent accuracy with an error of within ±1 minute between the predicted estimated arrival time and the actual arrival time; however, when the cargo load is 80% or more, a repeated pattern is observed in which the estimated arrival time is consistently predicted 5 to 7 minutes early (i.e., optimistic estimated arrival time) in the hilly sections of the same route. This can be analyzed as being due to the fact that the existing fusion module did not sufficiently reflect the interaction between the vehicle's physical characteristics (load) and the road characteristics (hill gradient).

[0158] Accordingly, the fusion module of the present invention utilizes an attention-based structure to perform adjustments by increasing the cross-attention weight between the input features of vehicle physical characteristics (load) and road characteristics (slope, curvature, etc.) for conditions where repeated error patterns are detected (cargo load of 80% or more and specific hill sections). The updated fusion module is subsequently trained so that changes in vehicle driving speed according to the load under the same conditions are more accurately reflected in the prediction of the estimated time of arrival, and as a result, the problem of early prediction of the estimated time of arrival that occurred in the relevant hill section during high-load operation in the future can be resolved.

[0159] In addition, the attention weight adjustment within the fusion module according to the present invention can also be applied in a direction that weakens the relationship between specific input features.

[0160] For example, consider a case where the same vehicle repeatedly travels on a specific highway section under a high load (80% or more). Analysis reveals that the highway section in question is a flat highway with minimal gradient or curvature; despite the fact that the actual change in driving speed due to the vehicle's cargo load is negligible, existing models may perform unnecessary speed corrections by excessively reflecting the relationship between load characteristics and road properties. As a result, a phenomenon in which the estimated time of arrival is predicted to be excessively conservative (i.e., the estimated time of arrival is predicted to be later than the actual time) may occur repeatedly.

[0161] When such repeated error patterns are detected, the present invention may perform an update in the attention layer of the fusion module to weaken the cross-attention weights between cargo load and road characteristics. In other words, through the learning of the attention layer, the weights can be adjusted to decrease so that the load characteristics have almost no effect on the estimated time of arrival prediction in the relevant highway section. By reducing the reflection of the load under the same conditions (highway section and high load operation), the updated attention structure can subsequently calculate an estimated time of arrival prediction that better matches actual driving conditions without unnecessary speed correction.

[0162] As such, the present invention enables bidirectional updates that strengthen or weaken the importance of relationships between features in the attention layer within the fusion module, thereby effectively improving the problem of overfitting or underfitting in the estimated time of arrival prediction model.

[0163] In one embodiment, the computing device (100) can predict the personalized vehicle arrival time for the next vehicle trip based on an adjusted prediction model (S250).

[0164] After repeated error patterns are detected and the weights of the prediction model are adjusted online, the updated prediction model is applied to calculate a personalized vehicle arrival time for the next vehicle trip. This step serves as the final stage of a cycle for the system's continuous adaptation and improvement of prediction accuracy, and is characterized by the fact that the adjusted model is immediately reflected in the real-time service to provide users with the latest improved estimated arrival time information.

[0165] Specifically, at the time a new vehicle operation request is received or a vehicle route is established, input data such as driver and vehicle identifier information, route information, and time-based traffic patterns can be input into the prediction model. At this time, since the input data is processed according to the latest weights and correction coefficients of the updated model, a corrected prediction result can be produced by reflecting specific error patterns that have occurred repeatedly in the past.

[0166] The arrival time correction coefficient derived from the processing of the prediction model is applied to the base arrival prediction time, and the result can be converted into a personalized vehicle arrival time. This is provided to the driver or delivery system manager through a user interface (UI), and monitoring can subsequently be performed during the actual operation based on the aforementioned corrected estimated arrival time.

[0167] Through this, the present invention can provide structural advantages in which the model automatically improves precision as operations are repeated, and newly detected error patterns are immediately reflected starting from the next operation. Furthermore, the newly adjusted prediction model can be cumulatively applied not only to the next operation but also to future operations under the same conditions, allowing it to evolve into a personalized and localized high-precision arrival time prediction engine in the long term.

[0168] FIG. 8 is a diagram illustrating a method for predicting the arrival time of a vehicle according to a loading state and road characteristics according to a third embodiment of the present invention.

[0169] Referring to FIG. 8, the computing device (100) can receive vehicle identification information and route information (S310).

[0170] Here, vehicle identification information may include an identification value for distinguishing the unique entity of a vehicle and for searching for related data or inputting it into a model based on this. For example, it may include a vehicle registration number, a VIN (Vehicle Identification Number), a unique identifier received from a vehicle terminal (OBD, TCU, etc.), or a unique key value received from sensors and communication modules attached to the vehicle. This identification information can be linked with a pre-registered vehicle database and utilized as a reference value to extract driving history, specification data analysis results, and the like for the corresponding vehicle.

[0171] Route information is a concept that includes spatial data related to a vehicle's driving plan. It may include not only simple origin and destination coordinates but also a route sequence composed of specific road segments connecting them. Route information may include a list of road segments generated from map-based route search results, the distance of each segment, speed limits, road types (e.g., expressways, general roads, two-lane roads, etc.), real-time traffic flow information, and traffic event information such as accidents or construction. This information can be collected through TPEG, VDS, DSRC, GPS-based real-time traffic APIs, or route search engines. Furthermore, route information is represented as a continuous sequence of road units with a time-series structure rather than a simple connection of points, and can be utilized as input for subsequent road characteristic analysis and the calculation of basic estimated arrival times.

[0172] Additionally, when a user sets a destination, the computing device (100) calculates an optimal route between the starting point and the destination through a route search engine linked to a map API or a navigation system, and accordingly obtains static and dynamic information about the road sections that constitute the route.

[0173] For example, the computing device (100) can obtain detailed route data including real-time average speed, road congestion level, elevation change amount, etc. for the route from an external server (e.g., Google Maps API, HERE API, Tmap) or a self-developed traffic information system.

[0174] These identification and route information can be used in subsequent stages to reflect vehicle characteristics, analyze road conditions, and calculate arrival time correction coefficients.

[0175] In one embodiment, the computing device (100) can obtain a matching driving profile that matches the input vehicle identification information and route information (S320).

[0176] To explain further, the computing device (100) receives a matching driving profile from at least one of a transportation management system (TMS) or a driving management system (FMS) through an application programming interface (API), and the matching driving profile may include vehicle specification data consisting of the type and loading status of the vehicle and road characteristic data consisting of at least one of gradient, intersection density, and traffic light density.

[0177] Vehicle specification data may include static attributes such as the type of target vehicle (e.g., truck, van, trailer), maximum allowable load capacity per vehicle type, cargo box volume, and drive method, as well as dynamic attributes such as weight (kg), volume (m³), load rate (%), and load volume ratio (%) corresponding to the load condition at the time of operation. This information may be constructed through dispatch history, cargo registration information, or data collected from in-vehicle OBD terminals or load sensors stored in a transportation management system (TMS) or freight management system (FMS).

[0178] Road characteristic data may include structural or environmental characteristics of multiple road sections included in the route actually traveled by the vehicle. For example, it may include the gradient (%), intersection density (number of intersections / 1km), traffic light density (number of traffic lights / 1km) of each section, and this information may be generated through GPS-based trajectory analysis or integration with an external map engine.

[0179] In one embodiment, the loading state may be expressed as at least one of weight, volume, loading rate, and loading volume ratio.

[0180] To explain further, the loading status information included in the matching driving profile may include weight-based information or volume-based information of the vehicle. Specifically, it may be expressed as at least one indicator among the weight (kg) of the actual cargo loaded on the vehicle, the loading volume (m³), the loading ratio (%; loading weight χ maximum allowable weight), or the loading volume ratio (%; loading volume χ maximum cargo capacity).

[0181] For example, in the case of an 11-ton cargo truck, if the actual load is 8,800 kg under conditions where the maximum allowable load is 11,000 kg, the load factor is calculated as 80%. As such, the loading condition is a key indicator that reflects the load conditions at the time of operation and affects various physical characteristics such as the vehicle's acceleration / deceleration performance, fuel efficiency, and speed on inclines; therefore, it must be reflected to improve the accuracy of the arrival time prediction model.

[0182] The above loading status information may be information received at the time of dispatch planning or cargo registration in the transportation management system (TMS), or data collected in real time from a load sensor or a vehicle terminal linked within the flight management system (FMS) (e.g., OBD-II, CAN communication device, etc.). By integrating and analyzing such loading status data together with vehicle specifications, the computing device can generate a high-dimensional vector representation of the physical characteristics of the vehicle.

[0183] In one embodiment, at least one of the slope, intersection density, and traffic light density can be calculated by machine learning based on at least one of the sensor information and GPS information of the vehicle included in the matching driving profile.

[0184] Specifically, the average slope of each road section can be estimated based on the rate of change of altitude values ​​and the difference in altitude relative to distance on the vehicle's GPS trajectory, and precise slope calculation is possible through the pitch angle (longitudinal slope) value provided by the IMU sensor (inertial measurement unit) attached to the vehicle.

[0185] For example, the gradient can be calculated based on changes in elevation values ​​identified in the vehicle's GPS trajectory data and the length of the section. For instance, if the elevation value increases by 5 meters while the vehicle travels approximately 100 meters, the average gradient of that section can be estimated as 5%. In this case, if the pitch angle value of the IMU sensor collected in the same section averages 2.8 degrees, a more precise calculation of the gradient is possible by cross-verifying the GPS-based estimation results with the sensor-based data.

[0186] In addition, intersection density can be estimated based on locations where stop-and-go patterns occur repeatedly in specific road sections. For example, if a vehicle's sudden stop followed by restart event occurs an average of four or more times within a 1km section and is accompanied by a change in steering angle, it can be determined that there is a high probability of at least three intersections existing in this section. This information is extracted by analyzing the driving logs of multiple vehicles in the past using a clustering technique.

[0187] Similarly, traffic light density can also be estimated based on driving logs. For example, if a pattern is repeatedly observed where different vehicles stop at the same location and then start at intervals of approximately 20 to 30 seconds, it can be determined that there is a traffic light at that point where control is performed according to the signal cycle. If this stop-start interval is repeated five or more times within a 1km section, that section can be considered a high-density traffic light control section and classified as a section with high traffic light density.

[0188] Consequently, one embodiment of the present invention can provide a vehicle arrival time prediction structure capable of precisely reflecting real driving conditions by combining the reception of structural data through a transportation management system (TMS) or a flight management system (FMS) with statistical learning-based road characteristic estimation based on GPS, sensor data, etc.

[0189] In one embodiment, the computing device (100) can recognize physical characteristics of the vehicle according to the loading state of the vehicle based on the matching driving profile and recognize road characteristics of a plurality of road sections included in the driving path according to the path information (S330).

[0190] In one embodiment, the computing device (100) can generate a vehicle embedding vector based on the vehicle's specification data and generate a plurality of road embedding vectors for each road section based on the road characteristic data.

[0191] Here, the vehicle specification data may consist of detailed items such as vehicle type, load condition, vehicle weight, engine output, and drive method. Specifically, the vehicle specification data may include physical factors that affect responsiveness to road conditions, such as the vehicle's acceleration performance, climbing ability, braking distance, and fuel efficiency.

[0192] And, the computing device (100) can normalize the vehicle's specification data and input it into a multilayer perceptron-based embedding network to generate a vehicle embedding vector.

[0193] For example, a computing device (100) can first select data such as whether it is an electric vehicle, curb weight, maximum load capacity, output / torque curve, drivetrain configuration, etc., and then, depending on the data type, convert continuous values ​​into normalized forms and categorical values ​​into one-hot or embedding forms. Subsequently, a multilayer perceptron learns weights to reflect the relative influence of each specification on vehicle performance, and the output layer can output a fixed-dimensional vector summarizing vehicle performance.

[0194] Additionally, the computing device (100) can sort road characteristic data for each of a plurality of road sections in a path order. Then, the computing device (100) can input the road characteristic data sorted in a path order into a transformer encoder to generate road embedding vectors for each of the plurality of road sections.

[0195] Here, road characteristic data includes road length, speed limit, road type, number of lanes, and real-time traffic congestion for each section, and the computing device (100) can vectorize this data to construct an input sequence. Then, the transformer encoder can learn the characteristics of each section and the mutual influence with adjacent sections to output a road embedding vector that reflects the overall flow and difficulty of the road conditions.

[0196] These vehicle and road embedding vectors can subsequently be used as input values ​​for predicting vehicle arrival times based on load status and road characteristics.

[0197] In one embodiment, the computing device (100) can calculate an arrival time correction coefficient based on the physical characteristics of the vehicle and the road characteristics (S340).

[0198] To explain further, the vehicle embedding vector and the road embedding vectors for each of the plurality of road sections are input into a pre-trained prediction model, and arrival time correction coefficients can be obtained from the prediction model.

[0199] Here, the prediction model comprises a fusion module and a regression prediction module, and the vehicle embedding vector and a plurality of road embedding vectors for each road section are input into the fusion module to generate an integrated representation vector reflecting the interrelationship between the vehicle and road characteristics, and the integrated representation vector is input into the regression prediction module to calculate an arrival time correction coefficient to be applied to the basic arrival prediction time.

[0200] Specifically, the prediction model receives vehicle embedding vectors and road embedding vectors as inputs, respectively, and can learn the relationship between load status and road characteristics through a multi-head attention structure while maintaining their individual characteristics. Here, the fusion module reflects the combined effects between each element through attention heads from various perspectives and can generate an integrated representation vector by assigning weights to input elements with high importance under specific conditions.

[0201] To explain further, the fusion module highlights important features by cross-referencing each embedding vector with other embeddings through attention. For example, when a characteristic recognized as a large cargo truck with a high load capacity in a vehicle embedding vector is combined with a characteristic of an urban road section with many slopes and high density of traffic lights and intersections in a road embedding vector, the attention score between the combinations can be measured as relatively high. In this case, the model can learn an interaction relationship indicating that there is a high probability of speed reduction and arrival time delays when a heavily loaded vehicle travels on a road with slopes and frequent stop-and-go traffic.

[0202] The fusion module reflects meaningful correlations between these input characteristics to assign high weights to the corresponding road segment embeddings, and the resulting integrated representation vector can be configured to emphasize the arrival time delay characteristics that occur when a heavy-load vehicle travels under complex urban road conditions. In this manner, the fusion module of the present invention learns meaningful interactions between input elements, and this integrated representation can be passed to a regression prediction module to support the calculation of more precise arrival time correction coefficients.

[0203] In addition, the prediction model can input the integrated representation vector into the regression prediction module to calculate an arrival time correction coefficient to be applied to the base arrival time. Specifically, the prediction model receives the integrated representation vector through a multilayer perceptron-based regression prediction module and outputs an arrival time correction coefficient (f) in the form of a real value; this arrival time correction coefficient can be used to calculate the vehicle's arrival time (ETA_pred) based on the final loading status and road characteristics by multiplying it by the route-based base arrival time (ETA_base).

[0204] For example, if the vehicle characteristics are those of a medium-sized truck with a low load and the road characteristics consist of smooth sections centered on highways, the prediction model may determine that the vehicle is likely to drive at a stable speed without being significantly affected by road conditions. Under these conditions, it is assessed that the actual arrival time may be slightly faster than the baseline prediction time, so the arrival time correction factor may be calculated as a value lower than 1, such as 0.90. This is an example where the prediction model numerically reflects that low load and a smooth highway environment contribute to reducing arrival time, based on results learned from past similar driving conditions.

[0205] As another example, if the vehicle is a large truck with a high load capacity and the road environment consists mainly of national highway sections with many slopes and traffic lights, the prediction model may determine that the actual arrival time is likely to be delayed compared to the estimate due to the overlap of high load conditions, frequent stops and starts, and deceleration situations. Under these conditions, the arrival time correction factor is calculated as a value greater than 1, such as 1.20, so the final estimated arrival time is conservatively adjusted. This is an example where the model learns the possibility of speed reduction resulting from the interaction between load conditions and road structural complexity, and numerically reflects the potential for delay for that combination.

[0206] In one embodiment, the computing device (100) can calculate a base arrival prediction time and apply the arrival time correction coefficient to the base arrival prediction time to predict the arrival time of a vehicle according to the loading condition and road characteristics (S350).

[0207] A computing device (100) can acquire road characteristic data for each of a plurality of road sections, including at least one of information such as road length, speed limit, road type, gradient, and real-time traffic congestion.

[0208] Additionally, the computing device (100) can calculate the time to pass each of a plurality of road sections based on road characteristic data. And, the computing device (100) can calculate a basic arrival prediction time by accumulating the time to pass.

[0209] Specifically, the computing device (100) can calculate the average driving time for each of a plurality of road sections based on the speed limit and section length included in the road characteristic data. Additionally, the computing device (100) can calculate the passing time for the road section by applying a traffic correction coefficient corresponding to real-time traffic congestion to the average driving time.

[0210] More specifically, the computing device (100) can calculate a basic average driving time based on speed limit and section length data for each road section. For example, the computing device (100) can calculate a basic average driving time by dividing the section length by the speed limit. Subsequently, the computing device (100) can apply a traffic correction factor to the average driving time that reflects the actual traffic conditions of each road section based on real-time traffic congestion information. The traffic correction factor is set as a quantified value according to the level of traffic congestion, for example, 1.0 when traffic is smooth, 1.2 for normal congestion, and 1.5 or higher for congestion, so that the predicted passing time can vary depending on the situation even under the same distance and speed limit conditions.

[0211] For example, if a specific road section is 1.5 km long and the speed limit is 60 km / h, the computing device (100) can calculate the reference average travel time as 1.5 km ÷ 60 km / h = 0.025 hours (= 1.5 minutes). Then, if the computing device (100) determines that the traffic information received in real-time for the section is congestion level 2 (e.g., moderate congestion), it multiplies by a predefined traffic correction factor (e.g., 1.3) to calculate the actual travel time as 1.5 minutes × 1.3 = 1.95 minutes. In this way, the computing device (100) can calculate the base estimated time of arrival (ETA_base) for the entire route by repeating the same calculation for all road sections on the route and accumulating the travel time for each section.

[0212] Additionally, the computing device (100) can calculate a personalized vehicle arrival time (ETA_Pred) by multiplying or adding an arrival time correction factor f, which integrally reflects vehicle characteristics and road characteristics, to the calculated base arrival time (ETA_Base). In this case, the arrival time correction factor is a real value and is a quantified value based on the interaction between the vehicle's loading status and road characteristics.

[0213] For example, the computing device (100) can calculate ETA_PRED = 30 minutes × 1.1 = 33 minutes when the base arrival prediction time is 30 minutes and the arrival time correction factor f=1.1. As another example, the computing device (100) can calculate the personalized vehicle arrival time to be reduced to 27 minutes when the driver drives faster than average and the vehicle performance is excellent, so f=0.9. This method can provide high prediction accuracy in terms of driver customization and realism compared to existing systems that use only static route-based prediction times.

[0214] Accordingly, the computing device (100) of the present invention can provide logistics vehicle arrival time prediction information that minimizes errors in logistics delivery by calculating the arrival time by reflecting the vehicle's loading status and road section characteristics, rather than limiting the prediction to simple route distance or average speed-based prediction.

[0215] In one embodiment, the computing device (100) can perform feedback learning on a prediction model based on the error between the vehicle's arrival time and the actual arrival time according to the loading status and road characteristics. Additionally, when a repeated error pattern is detected while performing feedback learning on the prediction model, the computing device (100) can adjust the weight of at least one of the fusion module and the regression prediction module included in the prediction model.

[0216] First, the computing device (100) can calculate the error between the actual arrival time of the vehicle and the arrival time of the vehicle according to the loading status and road characteristics at the time the vehicle's operation is completed, and store metadata along with the error.

[0217] Specifically, at the time the vehicle's operation is completed, the time difference between the vehicle's actual time of arrival and the estimated time of arrival reflecting the pre-operation loading status can be calculated as an error value.

[0218] The completion of a trip can be recognized, for example, when a vehicle arrives at its final destination and the location is confirmed based on the location, or when a specific event occurs on a transportation management application, such as delivery completion or trip termination.

[0219] The error value calculated at this time is determined as a time difference, and metadata may be stored along with the error value. Here, the metadata may include driver information (driver ID, driver's driving history identifier, etc.), vehicle information (vehicle ID, vehicle type, vehicle loading status, etc.), route information (origin, destination, major waypoints, etc.), time information (departure time, arrival time, date and day of the week of operation, etc.), weather information (weather conditions for the relevant time period, etc.), road characteristic information (road type, traffic pattern, etc.), and other related information that may affect operation (accident information, etc.).

[0220] These error values ​​and metadata are recorded in a training data repository linked to the prediction model and can be used in subsequent steps to detect recurring error patterns and adjust the prediction model weights online.

[0221] Next, the computing device (100) can detect repeated error patterns based on metadata.

[0222] Specifically, the computing device (100) can obtain error data accumulated more than a certain number of times under the same operating conditions to detect repeated error patterns based on metadata, and calculate the average value and variance value of the error from the error data. If the average value of the error is greater than or equal to a threshold and the variance value of the error is less than or equal to a reference value, it can be detected as a repeated error pattern.

[0223] First, driving data corresponding to identical driving conditions is accumulated and stored. At this time, identical driving conditions refer to a set of driving where the driver identifier, vehicle identifier, and driving route identifier are identical, and whether the conditions match can be determined based on driving metadata within the system.

[0224] Subsequently, the system can determine whether a specific number of operations (e.g., 10 or more) have been accumulated among the sets of operations corresponding to the same operation conditions.

[0225] When operation data exceeding a certain number of times is secured, error data between the estimated arrival time and the actual arrival time of the vehicle can be extracted from the corresponding operation set, and the average and variance values ​​of the error can be calculated.

[0226] If the average value of the calculated error is greater than or equal to a set threshold (e.g., 10% of the personalized estimated time of arrival) and the variance of the error is less than or equal to a set standard (e.g., the standard deviation value is less than or equal to 20% of the average value of the error), that is, if the error shows a continuous and stable pattern above a certain level, it can be detected that a repeated error pattern exists for the condition.

[0227] By applying these judgment criteria, errors caused by temporary or exceptional operating conditions (e.g., sudden accidents, one-time construction, etc.) can be filtered out, and only structurally recurring errors under specific operating conditions can be identified.

[0228] If a repeated error pattern is detected, the computing device (100) can perform weight adjustment of at least one of the fusion module and regression prediction module of the prediction model in a subsequent step.

[0229] The computing device (100) performs fine-tuning targeting only the rear block of the regression prediction module, while keeping the encoder layer, which is the front block, fixed and can adjust the arrival time correction coefficient.

[0230] In the present invention, the regression prediction module may be largely composed of a front-end block and a back-end block. Among these, the front-end block is an encoder layer that transforms input data into high-dimensional feature vectors and primarily performs the role of extracting meaningful representations from time series or context information. Since this front-end block contributes to the learning of generalized representations and plays a key role in the overall model performance, the weights of the corresponding front-end block are maintained in a fixed state even during the local correction process for repeated error patterns in the present invention. Specifically, it is configured so that backpropagation to the encoder layer does not occur during training by freezing the gradient.

[0231] Meanwhile, the rear block of the regression prediction module may be composed of a fully connected layer or an output layer that calculates the final predicted value based on the feature vector transmitted from the front block. In this invention, when repeated error patterns are detected in data with specific metadata, instead of retraining the entire model, fine-tuning is performed only on the rear block to ensure computational efficiency and is designed to be suitable for real-time or near-real-time online learning environments. This structure has the advantage of improving prediction accuracy by introducing local correction coefficients while maintaining the representation learning capabilities of existing encoders.

[0232] For example, it can be assumed that in a specific urban section (section A to section B), the same driver P repeatedly experiences a pattern of estimated arrival time delays of about 5 minutes or more during the same time period (8:00 AM to 9:00 AM). After detecting the error pattern, the system can construct a feedback dataset filtered by the conditions of section A, driver P, and time period (8:00 AM to 9:00 AM).

[0233] Subsequently, training can be performed only on the rear block of the regression prediction module using the dataset. In this case, the encoder layer, which is the front block, remains fixed, and as the weights of the rear block are updated, it becomes possible to predict the estimated time of arrival by reflecting correction coefficients specialized for the corresponding section, driver, and time period. As a result, during the next operation, an estimated time of arrival is predicted with a delay of approximately 5 minutes compared to the previous one under the same conditions, and as operations are repeated, the results of fine-tuning accumulate, enabling more precise predictions.

[0234] This method offers the following advantages. First, it allows for the rapid incorporation of local error correction while maintaining the existing model structure and the learning capabilities of key representations. Second, due to the low computational load, online learning can be applied in real-time or near-real-time. Third, when repeated error patterns are limited to a few conditions (such as specific routes or drivers), local precision can be improved without compromising the generalization of overall model performance.

[0235] Additionally, the computing device (100) can adjust the attention weights within the fusion module to readjust the degree of reflection of the interrelationships between input features.

[0236] If repeated error patterns are detected, weight adjustments can be performed on the fusion module. Weight adjustments can be implemented by readjusting the corresponding attention weights. Specifically, the importance and the degree of reflection of interrelationships between input features can be readjusted, which can be particularly effective when the cause of the error stems from the inaccurate reflection of interactions between specific input features.

[0237] For example, it can be assumed that when the same truck (fixed vehicle ID) travels the same delivery route, a pattern is detected in which the estimated arrival time error occurs repeatedly depending on the cargo load. Specifically, when the same vehicle is driven empty (empty load state), it showed excellent accuracy with an error of within ±1 minute between the predicted estimated arrival time and the actual arrival time; however, when the cargo load is 80% or more, a repeated pattern is observed in which the estimated arrival time is consistently predicted 5 to 7 minutes early (i.e., optimistic estimated arrival time) in the hilly sections of the same route. This can be analyzed as being due to the fact that the existing fusion module did not sufficiently reflect the interaction between the vehicle's physical characteristics (load) and the road characteristics (hill gradient).

[0238] Accordingly, the fusion module of the present invention utilizes an attention-based structure to perform adjustments by increasing the cross-attention weight between the input features of vehicle physical characteristics (load) and road characteristics (slope, curvature, etc.) for conditions where repeated error patterns are detected (cargo load of 80% or more and specific hill sections). The updated fusion module is subsequently trained so that changes in vehicle driving speed according to the load under the same conditions are more accurately reflected in the prediction of the estimated time of arrival, and as a result, the problem of early prediction of the estimated time of arrival that occurred in the relevant hill section during high-load operation in the future can be resolved.

[0239] In addition, the attention weight adjustment within the fusion module according to the present invention can also be applied in a direction that weakens the relationship between specific input features.

[0240] For example, consider a case where the same vehicle repeatedly travels on a specific highway section under a high load (80% or more). Analysis reveals that the highway section in question is a flat highway with minimal gradient or curvature; despite the fact that the actual change in driving speed due to the vehicle's cargo load is negligible, existing models may perform unnecessary speed corrections by excessively reflecting the relationship between load characteristics and road properties. As a result, a phenomenon in which the estimated time of arrival is predicted to be excessively conservative (i.e., the estimated time of arrival is predicted to be later than the actual time) may occur repeatedly.

[0241] When such repeated error patterns are detected, the present invention may perform an update in the attention layer of the fusion module to weaken the cross-attention weights between cargo load and road characteristics. In other words, through the learning of the attention layer, the weights can be adjusted to decrease so that the load characteristics have almost no effect on the estimated time of arrival prediction in the relevant highway section. By reducing the reflection of the load under the same conditions (highway section and high load operation), the updated attention structure can subsequently calculate an estimated time of arrival prediction that better matches actual driving conditions without unnecessary speed correction.

[0242] As such, the present invention enables bidirectional updates that strengthen or weaken the importance of relationships between features in the attention layer within the fusion module, thereby effectively improving the problem of overfitting or underfitting in the estimated time of arrival prediction model.

[0244] FIG. 9 is a schematic diagram showing one or more network functions related to an embodiment of the present invention.

[0245] Throughout this specification, artificial intelligence model, neural network model, neural network, network function, and neural network may be used interchangeably. A neural network may generally consist of a set of interconnected computational units that may be referred to as “nodes.” These “nodes” may also be referred to as “neurons.”

[0246] A deep neural network (DNN) may refer to a neural network that includes multiple hidden layers in addition to input and output layers. Using a deep neural network allows for the identification of latent structures in data. That is, it is possible to identify the latent structures of photos, text, videos, voice, and music (e.g., what objects are present in a photo, what the content and emotions of a text are, what the content and emotions of a voice are, etc.). Deep neural networks may include convolutional neural networks (CNN), recurrent neural networks (RNN), autoencoders, Generative Adversarial Networks (GAN), restricted Boltzmann machines (RBM), deep belief networks (DBN), Q networks, U networks, Siamese networks, etc. The description of deep neural networks described above is merely illustrative and the present invention is not limited thereto.

[0247] Neural networks can be trained using at least one of supervised learning, unsupervised learning, and semi-supervised learning. The purpose of training a neural network is to minimize the error in the output. The training process involves repeatedly inputting training data into the network, calculating the error between the network's output and the target for the training data, and updating the weights of each node in the network by backpropagating the error from the output layer to the input layer in a direction that reduces the error. In the case of supervised learning, training data is used where the correct answer is labeled for each data point (i.e., labeled training data), whereas in the case of unsupervised learning, the correct answer may not be labeled for each training data point. For instance, in the case of supervised learning regarding data classification, the training data may consist of data where each training data point is labeled with a category. Labeled training data is input into the neural network, and the error can be calculated by comparing the network's output (category) with the label of the training data. As another example, in the case of unsupervised learning for data classification, the error can be calculated by comparing the input training data with the neural network output. The calculated error is backpropagated within the neural network (i.e., from the output layer to the input layer), and this backpropagation can update the connection weights of each node in each layer of the neural network. The amount of change in the connection weights of each updated node can be determined by the learning rate. The neural network's calculation of the input data and the backpropagation of the error can constitute a learning cycle (epoch). The learning rate can be applied differently depending on the number of iterations of the neural network's learning cycle.For example, a high learning rate can be used during the early stages of neural network training to quickly achieve a certain level of performance and increase efficiency, while a low learning rate can be used during the later stages to improve accuracy.

[0248] In the training of neural networks, the training data is generally a subset of the real-world data (i.e., the data intended to be processed by the trained neural network). Consequently, a training cycle may exist where errors decrease on the training data but increase on the real-world data. Overfitting is a phenomenon where the network learns excessively on the training data, leading to increased errors on real-world data. For example, a neural network trained on yellow cats might fail to recognize cats when seeing anything other than yellow, which can be considered a type of overfitting. Overfitting can act as a cause for increased errors in machine learning algorithms. Various optimization methods can be used to prevent this overfitting. To prevent overfitting, methods such as increasing the training data, regularization, or dropout—which involves omitting some nodes from the network during the training process—can be applied.

[0249] The steps of the method or algorithm described in connection with embodiments of the present invention may be implemented directly in hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. The software module may reside in RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), Flash Memory, a hard disk, a removable disk, a CD-ROM, or any form of computer-readable recording medium well known in the art to which the present invention belongs.

[0250] The components of the present invention may be implemented as a program (or application) and stored on a medium to be executed in combination with a computer, which is hardware. The components of the present invention may be implemented as software programming or software elements, and similarly, embodiments may be implemented in programming or scripting languages ​​such as C, C++, Java, assembler, etc., including various algorithms implemented as combinations of data structures, processes, routines, or other programming configurations. Functional aspects may be implemented as algorithms executed on one or more processors.

[0251] Those skilled in the art will understand that the various exemplary logic blocks, modules, processors, means, circuits, and algorithm steps described in connection with the embodiments disclosed herein may be implemented by electronic hardware, various forms of programs or design code (referred to herein as “software”), or a combination of all such. To clearly illustrate this interoperability between hardware and software, various exemplary components, blocks, modules, circuits, and steps have been generally described above in relation to their functions. Whether such functions are implemented as hardware or software depends on the design constraints imposed on the specific application and the overall system. Those skilled in the art may implement the functions described in various ways for each specific application, but such implementation decisions should not be interpreted as being outside the scope of the invention.

[0252] The various embodiments presented herein may be implemented as methods, devices, or articles of manufacture using standard programming and / or engineering techniques. The term “article of manufacture” includes a computer program, carrier, or medium accessible from any computer-readable device. For example, computer-readable media include, but are not limited to, magnetic storage devices (e.g., hard disks, floppy disks, magnetic strips, etc.), optical discs (e.g., CDs, DVDs, etc.), smart cards, and flash memory devices (e.g., EEPROMs, cards, sticks, key drives, etc.). Additionally, the various storage media presented herein include one or more devices and / or other machine-readable media for storing information. The term “machine-readable media” includes, but is not limited to, various other media capable of storing and / or holding instruction(s) and / or data.

[0253] It should be understood that the specific order or hierarchy of steps in the presented processes is merely an example of exemplary approaches. It should be understood that, based on design priorities, the specific order or hierarchy of steps in the processes may be rearranged within the scope of the invention. The appended method claims provide various step elements in a sample order, but do not imply limitation to the specific order or hierarchy presented.

[0255] Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention may be implemented in other specific forms without altering its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive.

[0256] delete Explanation of the symbols delete

Claims

Claim 1 A method performed by a computing device comprising at least one processor, comprising: inputting a driver's driving pattern, physical characteristics of a vehicle, and road characteristics into a prediction model to calculate an arrival time correction coefficient, and applying the arrival time correction coefficient to a base arrival prediction time to predict a personalized vehicle arrival time; at the time of completion of vehicle operation, calculating an error between the actual arrival time of the vehicle and the personalized vehicle arrival time, and storing metadata along with the error; based on the metadata, obtaining error data accumulated more than a certain number of times under the same driving conditions, and detecting a repeated error pattern if the average value of the accumulated error data is greater than or equal to a preset threshold and the variance value is less than or equal to a preset reference value; if a repeated error pattern is detected, adjusting the weights of the prediction model; and predicting a personalized vehicle arrival time based on the adjusted prediction model during the next vehicle operation; wherein the prediction model comprises a fusion module and a regression prediction module, and the fusion module fuses the input driver's driving pattern data, vehicle physical characteristic data, and road characteristic data, and reflects the combination effect between each element through multiple perspective attention heads under specific driving conditions A driving data-based continuous learning type vehicle arrival time prediction method, wherein a combined representation vector is generated by assigning weights to input elements of high importance, the prediction model inputs the combined representation vector to a regression prediction module to calculate an arrival time correction coefficient to be applied to the basic arrival time, and when the repeated error pattern is detected, the step of adjusting the weights of the prediction model involves adjusting the weights of at least one of the fusion module and the regression prediction module. Claim 2 A driving data-based continuous learning method for predicting vehicle arrival time, wherein, in claim 1, the metadata includes at least one of driver information, vehicle information, route information, and time information. Claim 3 delete Claim 4 A driving data-based continuous learning method for predicting vehicle arrival times, characterized in that, in claim 2, the same driving conditions are the same driver, the same vehicle, and the same route conditions. Claim 5 delete Claim 6 A driving data-based continuous learning type vehicle arrival time prediction method, wherein, in claim 1, when a repeated error pattern is detected, the step of adjusting the weights of the prediction model that calculates the arrival time correction coefficient is characterized by adjusting the weights of the regression prediction module. Claim 7 A driving data-based continuous learning type vehicle arrival time prediction method, wherein in claim 6, the weight adjustment of the regression prediction module fixes the weight of the encoder layer in the front end responsible for feature extraction among the multiple layers constituting the regression prediction module, and performs fine-tuning only on the regression layer in the rear end responsible for calculating the arrival time correction coefficient. Claim 8 A driving data-based continuous learning type vehicle arrival time prediction method, wherein, in claim 1, when a repeated error pattern is detected, the step of adjusting the weights of the prediction model that calculates the arrival time correction coefficient is characterized by adjusting the weights of the fusion module. Claim 9 In claim 8, the weight adjustment of the fusion module adjusts the weights between the corresponding input feature attentions within the fusion module to readjust the degree of reflection of the interrelationship between input features, a driving data-based continuous learning type vehicle arrival time prediction method. Claim 10 A device comprising: a memory for storing one or more instructions; and a processor for executing the one or more instructions stored in the memory, wherein the processor performs the method of claim 1 by executing the one or more instructions. Claim 11 A computer program stored on a computer-readable recording medium that is combined with a computer, which is hardware, to perform the method of claim 1.

Images