Traffic control period division method based on multi-modal large model and skill arrangement

Abstract

The application discloses a signal control period division method based on a multi-modal large model and skill arrangement, and core technologies include: image understanding and inference ability of the multi-modal large model, a combinable processing flow based on skill arrangement, multi-granularity traffic flow statistics oriented to different optimization targets, a flow classification method based on import roads-turning, a two-stage period division strategy fusing image driving and structured data analysis, and a rule updating mechanism based on a feedback closed loop. The application combines visual understanding ability of the multi-modal large model with structured data analysis, realizes end-to-end automatic processing from user demand input to period division result output through a skill arrangement mechanism, is suitable for period division optimization of a city road intersection signal control scheme, and is suitable for scenes such as cycle optimization, green ratio and cycle optimization, and phase design and timing scheme full-element optimization.

Description

Technical Field

[0001] This invention relates to a signal control time period division method based on multimodal large model and skill orchestration, belonging to the interdisciplinary field of traffic signal control and artificial intelligence. Background Technology

[0002] Traffic signal control is one of the core means of urban traffic management. In practical engineering, signal control schemes typically adopt a time-segmentation strategy, dividing the day into several time periods, each with a different timing scheme (including parameters such as cycle time, green light ratio, and phase sequence) to adapt to the changing characteristics of traffic flow over time. Therefore, the rationality of time period division directly affects the implementation effectiveness of the signal control scheme.

[0003] Currently, existing technologies in the field of signal control time segmentation mainly include the following categories:

[0004] (1) Clustering-based methods Clustering-based methods are the most widely used technical approach in the field of time-segmentation. Relevant patents and literature include: a signal control time-segmentation method based on ordered clustering (e.g., Chinese patent application CN106408960A), which utilizes the sequential regularity of traffic flow to cluster highly similar time series, achieving automatic optimization of the number of clusters; a multi-time-segmentation method for intersection signal control based on the k-shape algorithm, adapting to time series of different lengths through shape similarity matching; a multi-dimensional time series segmentation method for intersection control time-segmentation, employing dynamic programming and correction algorithms based on non-parametric multi-change point analysis, and e-agglo hierarchical aggregation and cyclic goodness-of-fit statistical estimation of the number of segments and segment point locations; a signal control time-segmentation method based on trajectory data, using a bisection k-means clustering algorithm to cluster traffic flow data; and an automatic multi-time-segmentation method for intersection timed signal control, achieving time-segmentation based on data standardization, symbolic aggregation approximation, and KMeans clustering.

[0005] While the clustering-based methods described above have achieved automation of time period segmentation to some extent, they share the following common problems: they are sensitive to abnormal data and short-term noise, and the clustering results are easily affected by single outliers; the clustering algorithms rely on preset parameters (such as the number of clusters, distance metrics, etc.), and the selection of parameters has a significant impact on the results; the clustering methods are mainly based on numerical features for calculation, making it difficult to fully utilize the overall shape information of traffic flow curves and the correlation between multiple curves.

[0006] (2) Method based on manual interpretation In practical engineering applications, traffic engineers typically determine time-period division schemes manually by observing flow-time distribution curves and combining this with their professional experience. This method comprehensively considers the overall trend of the curve, peak and trough characteristics, transition intervals, and the relationship between multi-directional flow, offering good interpretability and engineering applicability. However, this method suffers from low efficiency, high subjectivity, difficulty in standardizing and reusing results, and different engineers may provide different division results for the same set of curves.

[0007] (3) Methods that consider flow direction characteristics A signal control time-sharing method based on precise flow direction identification at intersections (e.g., Chinese invention patent CN117292562B) utilizes electronic eye data and intersection video data to obtain traffic flow data for each turn at the approach lanes. It then uses a traffic demand change feature extraction model in polar coordinates to extract features and construct a time-sharing scheme based on different flow directions within the intersection. While this method considers the traffic demand change characteristics of different flow directions, its core reliance on flow direction identification from video data and polar coordinate feature extraction makes it a traditional feature engineering method. It does not involve multimodal large-scale model image understanding capabilities, nor does it implement a mechanism for automatically switching statistical granularity based on optimization demand type.

[0008] (4) Methods based on large models / AI In recent years, with the development of artificial intelligence technology, patents related to the application of large models to traffic signal control have emerged: Urban traffic signal optimization methods based on AI large models utilize large models to analyze real-time monitoring data time series, but do not specifically focus on time period division; dynamic timing methods for road intersections based on traffic large models optimize intersection timing using large traffic models; and adaptive regional intersection signal timing coordination control methods utilize graph neural networks to learn traffic flow relationships between different intersections. While these methods introduce large models or deep learning techniques, none involve applying multimodal large models to image understanding of flow-time distribution curves, nor do they implement composable processing flows based on skill orchestration.

[0009] Based on the above analysis, the existing technical solutions have the following main problems in terms of signal control time period division: (1) Insufficient automation and strong reliance on manual intervention. Although clustering-based methods have achieved automation to a certain extent, parameter setting and result adjustment still require manual intervention; methods based on manual interpretation rely entirely on the experience of engineers and are inefficient.

[0010] (2) The statistical granularity is too limited to adapt to different optimization objectives. Existing methods usually treat traffic flow data as a uniform input and use a single statistical caliber (such as total traffic flow at the intersection), ignoring the different requirements of different traffic control optimization objectives for data granularity. For example, when only performing periodic optimization, attention is paid to the overall traffic flow change at the intersection; when performing green light ratio and periodic optimization, attention needs to be paid to the traffic flow change of each phase; and when performing phase design and timing scheme full-element optimization, attention needs to be paid to the time-varying differences between each approach lane and each turn.

[0011] (3) Sensitive to abnormal data and short-term noise. Clustering-based methods mainly rely on numerical features for calculation. When the detection equipment experiences occasional anomalies or there are short-term disturbances in traffic flow, the clustering results are easily interfered with, resulting in inaccurate identification of time period boundaries.

[0012] (4) Lack of comprehensive analysis capability of time-varying characteristics of multiple flow directions. Existing methods usually cannot simultaneously take into account the time-varying relationship between multiple entrances, multiple turns or multiple phases. When the peak times of different flow directions are inconsistent or the trends of change are inconsistent, a single total flow curve can easily mask the traffic characteristics of local directions.

[0013] (5) Lack of feedback optimization and continuous improvement mechanisms. Existing methods are usually one-time outputs. When the division results are inconsistent with user experience, there is a lack of systematic adjustment and feedback mechanisms, which makes it impossible to precipitate human experience and actual application effects into the basis for rule updates.

[0014] (6) Insufficient differentiation from general large model / agent orchestration techniques. Although general large model or general agent orchestration techniques can perform task decomposition or multimodal processing, they usually do not include the task templates, statistical granularity selection, channelization constraints, boundary saliency verification, and time period number engineering control rules required for traffic signal control time period division. It is difficult to directly output interpretable time period division results that can be used for signal control timing. Existing technologies do not fully consider the special characteristics of signal control time period division business and effectively combine general AI capabilities with professional business logic. Summary of the Invention

[0015] The present invention aims to solve the following technical problems: (1) How to achieve end-to-end automated processing from user needs to time period segmentation results, reducing human intervention and subjectivity; (2) How to automatically determine the corresponding traffic volume statistics granularity according to different signal control optimization objectives, so that the time period division results are closer to the actual needs of subsequent optimization tasks; (3) How to integrate the image understanding capabilities of multimodal large models with structured data analysis, reduce the impact of abnormal data and short-term noise on time period boundary identification, and improve the accuracy and robustness of the segmentation results; (4) How to construct a combinable processing flow based on skill orchestration so that the processing links under different optimization needs can be dynamically switched and support feedback closed-loop optimization and continuous improvement; (5) How to perform interpretable fine-tuning, merging or splitting of candidate boundaries in scenarios such as peak misalignment in multiple flow directions, local abrupt changes, low-flow CV distortion at night and excessive number of time periods, so as to ensure that the output time periods have engineering usability.

[0016] To address the aforementioned technical problems, the present invention discloses a method for dividing information control time periods based on a multimodal large model and skill orchestration. The method is characterized by employing a layered architecture design, with the overall technical architecture comprising, from top to bottom: a user interaction and application layer, a requirement understanding and task construction layer, a skill orchestration and scheduling layer, a business skill execution layer, a model and task template support layer, and a data and infrastructure support layer. User interaction and application are geared towards business users, providing input for credit control time period segmentation requirements, parameter supplementation interaction, result viewing and confirmation, and analysis report output; The requirements understanding and task construction layer is used to transform user input into standardized tasks that the system can execute and to construct task parameters; The skill orchestration and scheduling layer is used to organize skill calls according to task parameters and business processes, and to perform parameter verification and completion, skill call order control, branch selection under different optimization requirements, exception handling and result verification. The business skills execution layer is used to handle data processing capabilities during the credit control time period segmentation process. The model and task template support layer provides a multimodal large model and time period segmentation task templates. The multimodal large model is used to identify trends, peaks and troughs, inflection points, and stable intervals in the traffic-time distribution curve. The task templates are used to configure corresponding time period segmentation rules and analysis priorities according to different optimization needs. The data and infrastructure support layer serves as the source of basic data. The overall business process of the credit control time period division method includes: User demand identification and intent parsing: Receive the user's signal control time period segmentation demand input in natural language form, and parse it to obtain the target intersection, optimization date type and optimization demand type; Historical traffic data query: Automatically determines the skill set and call path based on optimization demand type and skill prerequisites; Traffic flow classification statistics and traffic flow-time distribution curve plotting: Obtain lane-level 5-minute historical traffic flow data, intersection channelization information, phase scheme data and holiday calendar data from traffic data sources, determine the statistical granularity according to the optimization demand type, and generate traffic flow-time distribution curves; Time period segmentation and result presentation: Input the traffic-time distribution curve and task template into the multimodal large model to obtain candidate time period segments, and then use the structured traffic sequence to perform comprehensive traffic construction, smoothing, difference calculation, coefficient of variation calculation and significance calculation, refine the candidate boundary, and form the final time period segmentation result and overlay display map.

[0017] Preferably, the implementation of user demand identification and intent parsing includes: A pre-defined standardized natural language input template is used to guide users to perform standardized input. Users can either fill in the blanks based on the standardized natural language input template or express themselves freely in colloquial language without using the standardized natural language input template. The large-scale model is used for semantic understanding and key information extraction to complete target intersection name recognition, optimized date type recognition, and optimized demand type recognition. Specifically: when recognizing target intersection names, the large-scale model processes aliases, abbreviations, and ambiguous expressions; when recognizing optimized date types, the large-scale model understands time descriptions and makes judgments based on a date type configuration library; when recognizing optimized demand types, the large-scale model matches and classifies user input with preset demand types based on its semantic understanding capabilities. The system performs legality and completeness checks on the extracted target intersection name, optimization date type, and optimization requirement type according to preset rules. When key parameters are missing, an interactive follow-up questioning mechanism is automatically triggered to guide the user to supplement the necessary information.

[0018] Preferably, the implementation of traffic classification statistics and traffic-time distribution curve plotting includes: the step of determining the classification statistics dimension, the step of entrance-turning classification, and the step of plotting the traffic-time distribution curve, wherein: The step of determining the classification and statistical dimensions automatically determines the corresponding traffic volume statistical granularity based on the type of optimization requirement input by the user. According to the steps of the approach lane-turning classification, the lanes of each approach lane are classified for turning, and then the traffic volume is counted according to the approach lane-turning. The classification is based on the functional markings of each lane in the intersection channelization design, while also considering the constraints of lanes that need to be released simultaneously due to channelization restrictions. The steps for drawing the flow-time distribution curve are based on the classification statistics results. With time as the horizontal axis and traffic volume as the vertical axis, the flow-time distribution curves of the classification statistics are drawn, including intersection-level flow-time distribution curves, phase-level flow-time distribution curves, and approach lane-turning-level flow-time distribution curves, corresponding to different optimization demand types.

[0019] Preferably, in the step of determining the classification statistical dimension: When the optimization requirement type is to support time-segmented periodic optimization: Calculate the traffic volume every 5 minutes at the intersection, and sum the traffic flow of all lanes within the intersection at the same time granularity to obtain the intersection-level traffic volume every 5 minutes; When the optimization requirement type is to support time-segmented green light ratio and cycle optimization: traffic volume is counted every 5 minutes according to phase, the lane set corresponding to each phase is determined according to the intersection phase design scheme, and the traffic flow of each lane in the same phase within the same time granularity is summed to obtain the phase-level traffic volume every 5 minutes. When the optimization requirement type is to support the full-element optimization of time-segment phase design and timing scheme: calculate the traffic volume every 5 minutes according to the approach lane-turning statistics, and complete the lane turning classification and traffic volume summation according to the intersection channelization design and approach lane-turning classification method to obtain the approach lane-turning level traffic volume every 5 minutes.

[0020] Preferably, the steps of the import lane-steering classification specifically include: Step 1: Initialize the collection With sets : Traversing the import lane All lane function signs: If a lane contains two or more turns, then the current lane is added to the set. ; Otherwise, if the lane is a single-turn lane and not a right turn, then the current lane is added to the set. ; Step 2, if set If empty, then there is no multi-turn shared lane, and the entrance lane is empty. Directly by set The individual steering elements are categorized, with each different steering element forming a separate category, and the process ends here. If set If not empty, then for the set Perform clustering and grouping of sets Each lane is considered a set of steering functions, and they are grouped according to the principle of "connected lanes with the same steering": If the turning sets of two lanes intersect, then the two lanes belong to the same group. By transitive closure, the collection Divide into several non-overlapping groups , , …, , ; Each group Steering combination Take the union of all lane turning sets within the group. If a lane intersects with multiple groups, then these groups are merged. Ultimately, the intersection of any two different groups' turning combinations is empty; Step 3: Process the set Merging with groups: If set If the value is empty, proceed directly to step 4 to determine the final category; If set Not empty, iterate through the set For each lane in the system, the current lane's turning direction is recorded as... For each lane, the following judgment is made; Check steering Is it included in a group? Steering combination middle: If such a group exists, then the current lane will be removed from the set. Remove from the list and add to the corresponding group. ; If no group contains a steering Then the current lane will be retained in the set. middle; After the traversal is complete, we obtain the updated groups and the updated sets. ; Step 4: Determine the final steering category of the approach lane: Case 1: If set Not empty: set The lanes are classified according to their single steering function, with each different steering direction forming a separate category; Import channel The steering categories include: each group Corresponding composite steering classes and sets Each different single steering type corresponds to a single steering class; Case 2: If the set Empty: Import channel The steering category is determined solely by the updated grouping: if the number of groups is 1, there is only one category, namely the composite steering category of that group; if the number of groups is greater than 1, there are multiple composite steering categories.

[0021] Preferably, a two-stage segmentation strategy is adopted for time period segmentation. The first stage is the image-driven preliminary segmentation stage based on a multimodal large model. Taking the flow-time distribution curve as input, the visual understanding capability of the multimodal large model is used to analyze the overall shape and micro-changes of the curve and output candidate time period segmentation schemes. The second stage is the boundary point fine-tuning stage based on the fusion of large model inference and data analysis. Taking the preliminary segmentation results obtained in the first stage and the original structured flow data as input, the boundary of the preliminary segmentation is fine-tuned through data quantitative analysis.

[0022] Preferably, in the first stage, the large model performs inference according to the following analytical steps: Observe the overall shape of the curve, identify obvious single-peak, double-peak or multi-peak structures, determine the approximate time range of the morning peak and evening peak, and determine the nighttime low peak period; Identify key parameters including the starting point, peak point and plateau period, pullback point, and fluctuation threshold, among which: The starting point is the point in time when the flow rate suddenly begins to rise continuously from a low value; Peak points and plateau periods are the time periods when the highest traffic occurs. If the peak lasts for a long time and there is a significant dip in the middle, consider whether to split it. The pullback point is the turning point where the flow rate drops significantly from a high level and then remains at a new level. By using fluctuation thresholds, we can focus on significant trend changes lasting longer than 15-20 minutes to ignore spikes; Perform time period merging and segmentation, further including: Perform peak-hour merging: If traffic changes are minimal from late night to early morning, merge them into a longer time period; Peak traffic segmentation: During morning and evening peak hours, traffic changes rapidly. The traffic flow is segmented according to the characteristics of "climbing-peak-drop" with a time granularity of 15-30 minutes. Peak-hour consolidation: During off-peak hours, if traffic fluctuates but the overall volume remains consistent, it can be merged into one or two large time periods. Separate transition periods: If the flow rate decreases linearly during the transition from peak to off-peak periods, a separate transition period is defined. When making comprehensive decisions based on multiple curves, if the graph contains multiple curves, the "greatest common divisor" or "bottleneck-oriented" approach is used: if any critical curve undergoes a significant change in shape, the current time point is taken as a potential dividing point; if the peak times in each direction are misaligned, the current time point is retained to ensure that all directions can obtain a suitable control scheme.

[0023] Preferably, the second stage includes the steps of data preprocessing and indicator definition, the step of stepwise reasoning and correction based on data, the step of total quantity control and mandatory merging, and the step of generating the final scheme, wherein: The data preprocessing and indicator definition steps are used to achieve: Comprehensive traffic flow calculation: Calculate the comprehensive traffic flow at the intersection every 5 minutes. This is the sum of traffic for each category every 5 minutes:

[0024] in, Indicates the first Traffic volume for a specific category within 5 minutes This indicates the total number of categories.

[0025] Smoothing: For the overall flow rate every 5 minutes and each The sequence is processed using a 3-point moving average. Calculate key metrics, which further include: difference The difference between the smoothed combined flow and the flow of each category is used to identify abrupt changes in flow. ,in, for The sum of the flow rate every 5 minutes after the 3-point moving average processing; coefficient of variation Standard deviation of flow rate within a time period with the mean The ratio, , Considered highly stable, The price fluctuations are considered extreme and a split should be considered. Significance threshold: Calculate the rate of change of the mean over adjacent time periods. , ,like Furthermore, a duration greater than 15 minutes is considered a valid segmentation criterion, and if the average change rate of adjacent time periods is... These are considered strong significant differences that must be separated. The step-by-step reasoning and correction based on data includes: Nighttime / off-peak period treatment: Discard the coefficient of variation Instead, use absolute range. For low-flow periods, if the absolute range is within an acceptable range, it should be merged into a long period. Find the moment when the flow rate increases monotonically for three consecutive cycles and the total flow rate exceeds a specific threshold, and use this as the end point of the low-peak period.

[0026] Peak start-up and ramp-up period: Check the difference If a significant jump occurs, and the initial boundary is drawn too early or too late, the boundary should be fine-tuned according to the principle of "continuous monotonic change for 3 consecutive periods". Peak period segmentation judgment: Calculate the coefficient of variation within the peak period. ,like Furthermore, there are obvious "double peaks" or "misaligned peaks". When assessing whether to split the peaks, the principle of engineering simplification should be followed: if splitting the peaks would result in too many time periods and the overall load is high, it is preferable not to split the peaks, but the start and end times must be accurately defined. Determining the merging of peak and decline periods: Calculate the mean difference rate between adjacent periods. If the mean difference is <15% and the overall trend is stable, it is recommended to merge them to simplify the scheme; if the mean difference is >25%, they must be separated. The total quantity control and mandatory merging steps are executed according to the following priority: The mandatory merging strategy is followed by the following priority: Prioritize merging adjacent time periods with the smallest difference in average traffic volume; The second-best approach is to merge the "early morning rush hour" with the "climbing period". The second merger will combine the "decline period" with the subsequent "off-peak period"; The final measure is to merge "evening peak start" and "evening peak peak"; The final time period segmentation points are determined through the steps of generating the final solution, and each time period is assigned a specific project name.

[0027] Preferably, the skill invocation path is automatically determined through a skill orchestration and scheduling mechanism based on task parameters, optimization requirement type, skill prerequisites, and intermediate result status, including the following steps: The system receives user input regarding traffic control time period segmentation requirements. Through user requirement identification and intent parsing skills, it analyzes the target intersection, optimizes the date type and requirement type, and forms standardized task parameters. The orchestration mechanism verifies the completeness of the task parameters, and when parameters are missing, it calls parameter verification and completion skills to supplement them. The scheduling mechanism determines the set of skills required for this task based on the task parameters. For all tasks, it calls the historical data query skill to obtain lane-level 5-minute historical traffic data. For tasks that need to be statistically analyzed by phase, it further calls the intersection phase design scheme query skill. Different statistical paths are selected based on the type of optimization needs. Among them, periodic optimization corresponds to intersection-level statistics, green ratio and periodic optimization correspond to phase-level statistics, and all-element optimization corresponds to entrance lane-turning level statistics. The skills are connected through standardized input and output. The output of the upstream skill serves as the input of the downstream skill. After each skill is executed, its output result is judged. Only when the current data meets the input requirements of the subsequent skill will the next skill be triggered.

[0028] Preferably, it also includes an exception handling mechanism and a feedback optimization mechanism, wherein: During the skill arrangement process, an exception handling mechanism is used to handle exceptions based on the intermediate result status: when the user inputs missing key parameters, the system returns a parameter supplementation prompt; when historical traffic data is missing or of insufficient quality, the system prompts a data exception and can replace it with the same type of historical date or use existing data to supplement it; when the curve fluctuates abnormally or the candidate boundaries are too dense, the system can merge the boundaries according to the minimum time period length and traffic stability rules; when the division result is inconsistent with the user's experience, the system submits the result to the user or expert for confirmation and correction. After users or experts adjust the time period division results, the adjustment content and reasons are recorded using a feedback optimization mechanism and used as feedback information. The feedback information is used to analyze the differences between the system division results and the manually confirmed results, and to further optimize the task templates, threshold parameters and skill arrangement rules.

[0029] This invention relates to core technologies such as multimodal large-scale model image understanding and reasoning capabilities, skill-based composable processing flow, multi-granularity traffic flow statistics for different optimization objectives, traffic flow classification method based on approach lanes and turns, two-stage time period segmentation strategy integrating image-driven and structured data analysis, and a rule update mechanism based on feedback loop. The data resources utilized include lane-level 5-minute historical traffic flow data, intersection channelization information, phase scheme data, holiday calendar or date type data, etc. Through the multimodal large-scale model's ability to understand the trend of traffic flow-time distribution curves and its ability to numerically correct structured traffic flow sequences, interpretable signal control time period segmentation results are generated.

[0030] This invention combines the visual understanding capabilities of multimodal large models with structured data analysis, and achieves end-to-end automated processing from user demand input to time period segmentation result output through a skill orchestration mechanism. It is applicable to time period optimization of signal control schemes at urban road intersections, including scenarios such as cycle optimization, green ratio and cycle optimization, and phase design and timing scheme full-element optimization. Compared with existing technical solutions, it has the following specific advantages: 1) End-to-end automation reduces labor costs. This invention realizes a complete automated processing process from user input to time period segmentation results output. Users only need to describe their needs in natural language, and the system can automatically complete all steps such as data query, traffic statistics, curve plotting, time period segmentation and result presentation, which significantly reduces the workload of manual interpretation and parameter tuning. 2) Demand-driven multi-granularity flow statistics mechanism This invention automatically determines the corresponding traffic volume statistics granularity based on the type of optimization requirement input by the user: periodic optimization corresponds to intersection-level statistics, green ratio and periodic optimization correspond to phase-level statistics, and all-element optimization corresponds to entrance lane-turning level statistics. This mechanism makes the time period division results closer to the actual needs of subsequent signal control optimization tasks and avoids the problem that a single statistical caliber cannot support different optimization scenarios. 3) Integrating image understanding and numerical analysis to improve segmentation accuracy. This invention combines the image understanding capabilities of multimodal large models with structured data analysis. Through a two-stage partitioning strategy, it fully utilizes curve morphology information (overall trend, peak and valley positions, transition intervals, and relationships between multiple curves) and refines the boundaries through data quantification analysis (difference, coefficient of variation, and significance test). This reduces the abnormal sensitivity caused by relying solely on clustering algorithms and also reduces the subjectivity of manual interpretation. 4) Comprehensive analytical capabilities for time-varying characteristics of multiple flow directions This invention generates multiple classified traffic flow curves through phase-level or approach lane-turning level traffic flow statistics, and comprehensively analyzes the changing relationships between different curves during time period division, including features such as peak misalignment between different flow directions, changes in traffic proportion, directional imbalance and local abrupt changes, which can more accurately identify the process of traffic state change. 5) Adjustable processes based on skills choreography This invention breaks down the signal control time period division process into multiple independent skills, and dynamically organizes the calling order and processing path of each skill according to user input and intermediate results through an orchestration and scheduling mechanism. When business rules, statistical standards, division requirements or intersection characteristics change, the method can be adapted by adjusting skill parameters, task templates or orchestration rules, thereby improving the method's adaptability to actual business changes. 6) Continuous optimization capability of feedback loop This invention introduces a feedback optimization mechanism into the skill scheduling process, which can back-accumulate human experience, expert judgment, and actual application effects into the rule base, template base, and case base. This makes the time period division of similar intersections, similar dates, or similar optimization tasks more in line with the actual traffic operation characteristics, and has the ability to continuously correct and iteratively optimize. Attached Figure Description

[0031] Figure 1 The overall business process diagram illustrates the complete business process of the method disclosed in this invention from user demand input to time period segmentation result output, including four main stages: user demand identification and intent parsing, historical traffic data query, traffic classification statistics and curve plotting, and time period segmentation and result presentation. Figure 2The overall technical architecture diagram illustrates the six-layer architecture design adopted by the method disclosed in this invention, including the user interaction and application layer, the requirement understanding and task construction layer, the skill orchestration and scheduling layer, the business skill execution layer, the model and task template support layer, and the data and basic support layer. Figure 3 The flowchart for user needs identification illustrates the business process of user needs identification and intent parsing, including steps such as guiding and receiving user input, semantic understanding, intent parsing, parameter validation and completion, and structured output. Figure 4 Flowchart for historical traffic data query; Figure 5 A flowchart is created to illustrate the business process of traffic flow classification and statistics and curve plotting, including steps such as determining the classification and statistics dimensions, traffic flow classification and statistics, and plotting the traffic flow-time distribution curve. Figure 6 The flowchart for time period segmentation based on a multimodal large model illustrates the two-stage approach to time period segmentation. The first stage is the initial segmentation of time periods driven by images based on a multimodal large model, and the second stage is the fine-tuning of boundary points based on the fusion of large model inference and data analysis. Figure 7 The flowchart illustrates the complete process of the skill orchestration and scheduling mechanism, including steps such as task parameter parsing, skill set determination, prerequisite check, branch path selection, skill input and output binding, execution status verification, and result transmission. Figure 8 A flowchart for traffic classification and statistics. Detailed Implementation

[0032] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0033] The information disclosed in this invention provides a method for dividing time periods based on a multimodal large model and skill orchestration. In its implementation, it adopts a layered architecture design, organizing user interaction, requirement analysis, skill orchestration, business skill execution, model and task template support, and basic data support in layers. This achieves a complete automated processing flow from user requirement input to time period division result output. The overall technical architecture comprises six layers from top to bottom: (1) User Interaction and Application Layer: This layer is designed for business users and provides functions such as inputting credit control time period division requirements, parameter supplementation interaction, result viewing and confirmation, and analysis report output. Users can submit time period division requirements to the user interaction and application layer through natural language or business forms.

[0034] (2) Demand understanding and task construction layer: Transform user input into standardized tasks that the system can execute, complete user demand identification, target intersection parsing, optimization date type identification and optimization demand type identification, and construct task parameters containing target intersection, date type, optimization demand type and other content.

[0035] (3) Skill orchestration and scheduling layer: organizes skill calls according to task parameters and business processes, and is responsible for parameter verification and completion, skill call order control, branch selection under different optimization requirements, exception handling and result verification.

[0036] (4) Business Skills Execution Layer: It carries the main processing capabilities in the process of dividing the traffic control time period, including historical data query, traffic volume classification statistics, traffic flow-time distribution curve drawing, task template matching, image-driven preliminary division based on multimodal large model, comprehensive traffic construction, smoothing processing, difference calculation, coefficient of variation calculation, significance calculation, boundary point fine correction, and drawing of time period division results and traffic flow-time distribution curve overlay diagram.

[0037] (5) Model and Task Template Support Layer: This includes a multimodal large model and time period segmentation task templates. The multimodal large model is used to identify trends, peaks and troughs, inflection points and stable intervals in the flow-time distribution curve; the task templates are used to configure corresponding time period segmentation rules and analysis priorities according to different optimization needs.

[0038] (6) Data and basic support layer: the source of basic data, including lane-level 5-minute historical traffic flow data, intersection channelization information, phase scheme data, holiday calendar or date type data, etc.

[0039] The overall business process of the method disclosed in this invention includes the following main steps: user demand identification and intent parsing → historical traffic data query → traffic classification statistics and traffic-time distribution curve drawing → time period division and result presentation. Each step is sequentially connected through a skill orchestration mechanism, with the output of the upstream step serving as the input of the downstream step, forming a continuous business processing link.

[0040] 1) User Needs Identification and Intent Parsing Methods User demand identification and intent parsing is the first business step of this invention. Its goal is to transform user-expressed requests for traffic control time slot allocation into standardized instructions that the system can understand and execute. This step requires identifying and parsing three types of key parameters: (1) Target intersection name: The signal-controlled intersection where the user wants to divide the time period.

[0041] (2) Optimize date types: include weekdays, weekends and holidays.

[0042] (3) Optimization of demand types: including only changing the cycle, changing the green ratio and cycle, and comprehensive optimization (phase design, green ratio and cycle can all be changed).

[0043] This step employs a three-tiered, progressive implementation strategy: "preset template guidance, large-scale model semantic parsing, and rule verification with follow-up questions as a fallback." The first layer: Pre-set input templates guide users to input in a standardized manner. The system pre-sets standardized natural language input templates. The basic template sentence is: "Divide the traffic control time periods for the ×× intersection into [weekdays / weekends / holidays] to support subsequent [cycle-only optimization / green ratio + cycle optimization / phase sequence and other full-element optimization]". Users can fill in the blanks based on the templates or express themselves freely in conversational language without using the templates.

[0044] The second layer: The large model is responsible for semantic understanding and intent parsing. Utilizing the large model for semantic understanding and key information extraction, it can accurately identify user intent even when expressions are incomplete, contain colloquialisms, or use non-standard technical terms. The large model performs the following core parsing tasks: intersection name recognition (handling aliases, abbreviations, and ambiguous expressions), optimized date type recognition (understanding time descriptions and making judgments based on a date type configuration library), and optimized demand type recognition (matching and classifying user input with preset demand types based on semantic understanding capabilities).

[0045] The third layer: Parameter validation and follow-up mechanism. The system validates the legality and completeness of the extracted parameters according to preset rules. When key parameters are missing, the system automatically triggers an interactive follow-up mechanism to guide the user to supplement the necessary information.

[0046] This step ultimately outputs structured data in JSON format, containing three fields: target intersection, optimization date type, and optimization requirement type, which serve as input parameters for subsequent skill orchestration.

[0047] (ii) Traffic Classification Statistics and Traffic-Time Distribution Curve Plotting 1) Determining the Dimensions of Classification Statistics The method disclosed in this invention automatically determines the corresponding traffic volume statistics granularity based on the type of optimization requirement input by the user. When the optimization requirement type is to support time-segmented periodic optimization: Calculate the traffic volume every 5 minutes at the intersection, and sum the traffic flow of all lanes within the intersection at the same time granularity to obtain the intersection-level traffic volume every 5 minutes; When the optimization requirement type is to support time-segmented green light ratio and cycle optimization: traffic volume is counted every 5 minutes according to phase, the lane set corresponding to each phase is determined according to the intersection phase design scheme, and the traffic flow of each lane in the same phase within the same time granularity is summed to obtain the phase-level traffic volume every 5 minutes. When the optimization requirement type is to support the full-element optimization of time-segment phase design and timing scheme: calculate the traffic volume every 5 minutes according to the approach lane-turning statistics, and complete the lane turning classification and traffic volume summation according to the intersection channelization design and approach lane-turning classification method to obtain the approach lane-turning level traffic volume every 5 minutes.

[0048] 2) Classification method of approach lane - steering When calculating traffic volume by approach lane and turning lane, it is necessary to first classify the lanes of each approach lane by turning lane. The core basis for classification is the functional identification of each lane in the intersection channelization design, while also considering the constraints of lanes that need to be released simultaneously due to channelization restrictions. In this embodiment of the invention, the classification method may specifically include the following steps (traversing each approach lane according to the following rules): Step 1: Initialize the collection With sets

[0049] Traversing the import lane All lane function signs: If a lane contains two or more turns (such as straight-left, straight-right, left U-turn, straight-left-right, etc.), then the current lane will be added to the set. .

[0050] Otherwise, if the lane is a single-turn lane and not a right turn (i.e., left turn, straight ahead, U-turn, etc.), then the current lane will be added to the set. (Right-turn lanes are not included in this category and can be ignored.) Step 2: If set empty There are no multi-turn shared lanes, and the entrance lanes are not included. Directly by set The individual steering elements are categorized (each different steering element is in one category), and the process ends.

[0051] Step 3: If set Not empty, for the set Clustering set Each lane is considered a set of steering functions, and they are grouped according to the principle of "connected lanes with the same steering": If the turning sets of two lanes intersect, then they belong to the same group; By transitive closure, the collection Divide into several non-overlapping groups , , …, , ; Each group Steering combination Take the union of all lane turning sets within the group. (If a lane intersects with multiple groups, these groups are merged; ultimately, the intersection of any two different groups' turning combinations is empty.) Step 4: Process the collection Merging with groups If set If the value is empty, proceed directly to step 5 to determine the final category.

[0052] If set Not empty, iterate through the set Each lane in the system (whose turning direction is denoted as ) ): examine Is it included in a group? Steering combination China (i.e.) ).

[0053] If such a group exists, then remove the lane from the set. Remove from the list and add to the corresponding group. ( Unchanged, because (Already included).

[0054] If no group contains Then the lane will be retained in the set. middle.

[0055] After the traversal is complete, we obtain the updated groups and the updated sets. .

[0056] Step 5: Determine the final steering category for the approach lane Case 1: Set Not empty set The lanes are classified according to their single steering function (each different steering is a separate category).

[0057] Import channel The steering categories include: each group The corresponding composite steering class (steering combination is) ) and sets Each different single steering corresponds to a single steering class.

[0058] Case 2: Set empty Import channel The steering category is determined solely by the updated grouping: if the number of groups is 1, there is only one category, namely the composite steering category of that group; if the number of groups is greater than 1, there are multiple composite steering categories.

[0059] 3) Plotting the flow-time distribution curve Based on the classification statistics, traffic flow-time distribution curves are plotted with time on the horizontal axis and traffic volume on the vertical axis. The curve types corresponding to different optimization needs are: intersection-level traffic flow-time distribution curve, phase-level traffic flow-time distribution curve, and approach lane-turning-level traffic flow-time distribution curve.

[0060] (iii) Two-stage time-sharing method based on multimodal large model The method disclosed in this invention adopts a two-stage segmentation strategy, combining the image understanding capabilities of multimodal large models with structured data analysis to achieve progressive time segmentation from initial segmentation to refined correction.

[0061] Phase 1: Preliminary segmentation of image-driven time periods based on a multimodal large model The first stage uses the flow-time distribution curve as input and leverages the visual understanding capabilities of the multimodal large model to analyze the overall shape and micro-changes of the curve, and outputs candidate time period division schemes.

[0062] In one embodiment of the present invention, an optional implementation method is that the large model performs inference according to the following analysis steps: Step 1: Global Trend Scan. Observe the overall shape of the curve, identify obvious single-peak, double-peak, or multi-peak structures, determine the approximate time range of the morning and evening peaks, and determine the nighttime low peak period (the period when traffic is lowest and most stable).

[0063] Step Two: Key Inflection Point Identification. In this step, the key inflection points identified include: Starting point – find the point in time when traffic suddenly starts to rise continuously from a low value (usually at the beginning of the morning rush hour). Peak points and plateau periods – Identify the time periods when the highest traffic occurs. If the peak lasts for a long time and has a significant dip in the middle (double peak), consider whether to split it. The point of decline – finding the turning point where traffic drops significantly from a high level and then maintains a new level (end of peak). Fluctuation threshold – Ignore minor spikes (noise) and focus on significant trend changes lasting more than 15-20 minutes.

[0064] Step 3: Time Period Merging and Segmentation Principles. The principles adopted in this step include: Peak hours merging – If traffic changes are minimal from late night to early morning, they should be merged into a single longer period. Peak traffic segmentation – During morning and evening peak hours, traffic flow changes rapidly and should be segmented more finely according to the characteristics of “climbing-peak-fall” (granularity of about 15-30 minutes). Off-peak consolidation – During off-peak hours during the day, if the traffic flow fluctuates but the overall volume remains consistent, it can be combined into one or two large time periods; Independent transition period – The period from peak to off-peak traffic, if the flow rate decreases linearly, can be divided into a separate transition period.

[0065] Step 4: Multi-curve integrated decision-making. When the graph contains multiple curves, it is necessary to choose the "greatest common divisor" or "bottleneck-oriented" approach—any time point where any critical curve undergoes a significant morphological change should be considered a potential dividing point. If the peak times in different directions are misaligned, these time points should be retained separately to ensure that a suitable control scheme can be obtained for all directions.

[0066] The first phase outputs a preliminary time period division scheme, including a feature analysis summary and a time period division table (including sequence number, start time, end time, time period name, and reason for division).

[0067] Phase Two: Refined Boundary Point Correction Based on the Fusion of Large Model Inference and Data Analysis The second stage uses the preliminary division results and raw structured traffic data as input to refine the boundaries of the preliminary division through data quantitative analysis.

[0068] In one embodiment of the present invention, an optional implementation includes the following steps in the second stage: Step 1: Data preprocessing and indicator definition.

[0069] Comprehensive traffic flow calculation: Calculate the comprehensive traffic flow at the intersection every 5 minutes. This is the sum of traffic for each category every 5 minutes:

[0070] in, Indicates the first Traffic volume for a specific category within 5 minutes This indicates the total number of categories.

[0071] Smoothing: For and each The sequence is processed using a 3-point moving average (i.e., the average of the current time and the previous and next 5-minute points) to filter out occasional spikes and retain the true trend. Let the total flow rate for a given 5-minute period be denoted as... The combined traffic for the five minutes before and after that five minutes were respectively and The smoothing result is The smoothing method is the same for all traffic categories.

[0072] Key performance indicators (KPIs) calculations include: (1) Difference The difference between the smoothed combined flow and the flow of each category is used to identify abrupt changes in flow. ; (2) Coefficient of variation Standard deviation of flow rate within a time period with the mean The ratio, , Considered highly stable, The price fluctuations are considered extreme and a split should be considered. (3) Significance threshold: the rate of change of the mean between adjacent time periods Furthermore, a duration greater than 15 minutes is considered a valid segmentation criterion; the rate of change of the mean of adjacent time periods... , are considered to be strong significant differences that must be split.

[0073] Step 2: Stepwise reasoning and correction based on data.

[0074] Nighttime / off-peak period processing: Discontinue Instead, use absolute extreme range. For low-traffic periods, even... If the absolute range is high, but the absolute range is within an acceptable range (e.g., range < 5-10 vehicles), it should be combined into a longer period. Find the moment when the traffic volume increases monotonically for three consecutive cycles, and the total traffic volume exceeds a specific threshold (e.g., 20 vehicles / 5 minutes), as the end point of the low peak.

[0075] Peak start and ramp-up period: Inspection Points showing significant jumps. If the initial boundary delineation is too early or too late, the boundary should be fine-tuned according to the principle of "continuous monotonic change over 3 cycles".

[0076] Peak period segmentation judgment: Calculate the peak period segmentation judgment. ,like Furthermore, there are obvious "double peaks" or "misaligned peaks". When assessing whether to split the peaks, the principle of engineering simplification should be followed: if splitting the peaks would result in too many time periods and the overall load would be high, it is preferable not to split the peaks, but the start and end times should be precisely defined.

[0077] Determining the merging of peak and trough periods: Calculate the mean difference rate between adjacent time periods. If the mean difference is <15% and the overall trend is stable, it is recommended to merge them to simplify the scheme; if the mean difference is >25% (such as midday trough and peak periods before and after), they must be separated.

[0078] Step 3: Total quantity control and mandatory merger.

[0079] The final number of time slots should be strictly controlled to around 10 (range 8-12). If the number of time slots exceeds the limit after initial correction, a forced merging strategy will be implemented according to the following priority: Prioritize merging adjacent time periods with the smallest difference in average traffic flow (especially at night or during transition periods). The second-best approach is to combine the "early morning rush hour" with the "climbing period" (if the difference in mean values ​​is acceptable, such as <35%). The next merge will combine the "decline period" with the subsequent "peak period" (if the averages are close). The final measure is to merge "evening peak start" and "evening peak peak" (if...). Although the temperature is high, the overall load is high.

[0080] Note: It is strictly forbidden to merge core time periods with a mean difference greater than 25%.

[0081] Step 4: Final solution generation.

[0082] Based on the above reasoning, the final time period segmentation points are determined. Each time period needs to be given a clear project name (such as nighttime low peak, morning peak uphill, morning peak peak, morning peak followed by flat peak, midday low peak, afternoon flat peak, evening peak peak, evening peak decline, evening transition, late night low peak).

[0083] (iv) Skills arrangement and scheduling mechanism 1) Scheduling and scheduling mechanism The skill orchestration and scheduling mechanism in the method disclosed in this invention is used to automatically determine the skill invocation path based on task parameters, optimization requirement type, skill prerequisites and intermediate result status.

[0084] In one optional implementation of this invention, the basic method for implementing the orchestration and scheduling mechanism includes the following steps: Step 1: Task parameter parsing and verification.

[0085] The system receives user input regarding traffic control time period segmentation requirements. Using user requirement identification and intent parsing skills, it analyzes the target intersection, optimization date type, and optimization requirement type, generating standardized task parameters. An orchestration mechanism verifies the completeness of these parameters, invoking parameter verification and completion skills to supplement any missing parameters.

[0086] Step 2: Determine the skill set.

[0087] The scheduling mechanism determines the set of skills required for the current task based on the task parameters. For all tasks, the historical data query skill must be invoked to obtain lane-level 5-minute historical traffic flow data; for tasks that require phase-based statistics, the intersection phase design scheme query skill is further invoked.

[0088] Step 3: Branch path selection.

[0089] Different statistical paths are selected based on the type of optimization needs: periodic optimization corresponds to intersection-level statistics, green light ratio and periodic optimization correspond to phase-level statistics, and all-element optimization corresponds to entrance lane-turning-level statistics.

[0090] Step 4: Skill input / output binding and execution status verification.

[0091] Skills are interconnected through standardized input and output, with the output of an upstream skill serving as the input for a downstream skill. After each skill is executed, the system performs a status check on its output. Only if the current data meets the input requirements of the subsequent skill will the next skill be triggered.

[0092] 2) Skills System The skills involved in the methods disclosed in the embodiments of this invention are classified into five categories according to their functions: (1) Requirements analysis skills: including user requirements identification and intent analysis skills, parameter verification and completion skills.

[0093] (2) Data query skills: including historical data query skills and intersection phase design scheme query skills.

[0094] (3) Classification statistics and curve drawing skills: including traffic volume classification statistics skills and traffic flow-time distribution curve drawing skills.

[0095] (4) Skills in time period division and boundary correction: including task template matching skills, image-driven preliminary division skills based on multimodal large models, comprehensive traffic construction skills, smoothing skills, difference calculation skills, coefficient of variation calculation skills, significance calculation skills, and boundary point fine correction skills based on the fusion of large model reasoning and data analysis.

[0096] (5) Results presentation skills: including skills in drawing overlay graphs of time period division results and flow-time distribution curves, and results output skills.

[0097] (v) Anomaly Handling and Feedback Optimization Mechanism 1) Exception handling mechanism During the skill arrangement process, the system can handle anomalies based on the intermediate result status: when the user inputs missing key parameters, the system returns a parameter supplementation prompt; when historical traffic data is missing or of insufficient quality, the system prompts a data anomaly and can replace it with the same type of historical date or use existing data to supplement it; when the curve fluctuates abnormally or the candidate boundaries are too dense, the system can merge the boundaries according to the minimum time period length and traffic stability rules; when the division result is inconsistent with the user's experience, the system submits the result to the user or expert for confirmation and correction.

[0098] 2) Feedback optimization mechanism After users or experts adjust the time period division results, the system records the adjustments and reasons, and uses this as feedback information. This feedback information can be used to analyze the differences between the system's division results and manually confirmed results, further optimizing task templates, threshold parameters, and skill scheduling rules. For example, when the peak-hour boundary is consistently late at similar intersections, the peak-hour identification rules can be adjusted appropriately; when the time period division in a certain type of task is too fine, the minimum time period length or adjacent boundary merging rules can be adjusted. Through this feedback optimization mechanism, the system can accumulate the experience of manual adjustments in historical tasks and the actual application effects as a basis for rule updates, assisting in template updates and scheduling strategy adjustments in subsequent optimization tasks.

[0099] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this invention can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this invention can be combined or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure.

[0100] The embodiments of the present invention have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. The scope of the present invention is defined by the appended claims and their equivalents. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the present invention, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A method for dividing signal control time periods based on a multimodal large model and skill orchestration, characterized in that, The aforementioned information control time period segmentation method adopts a layered architecture design. The overall technical architecture, from top to bottom, includes a user interaction and application layer, a requirement understanding and task construction layer, a skill orchestration and scheduling layer, a business skill execution layer, a model and task template support layer, and a data and infrastructure support layer, wherein: User interaction and application are geared towards business users, providing input for credit control time period segmentation requirements, parameter supplementation interaction, result viewing and confirmation, and analysis report output; The requirements understanding and task construction layer is used to transform user input into standardized tasks that the system can execute and to construct task parameters; The skill orchestration and scheduling layer is used to organize skill calls according to task parameters and business processes, and to perform parameter verification and completion, skill call order control, branch selection under different optimization requirements, exception handling and result verification. The business skills execution layer is used to handle data processing capabilities during the credit control time period segmentation process. The model and task template support layer provides a multimodal large model and time period segmentation task templates. The multimodal large model is used to identify trends, peaks and troughs, inflection points, and stable intervals in the traffic-time distribution curve. The task templates are used to configure corresponding time period segmentation rules and analysis priorities according to different optimization needs. The data and infrastructure support layer serves as the source of basic data. The overall business process of the credit control time period division method includes: User demand identification and intent parsing: Receive the user's signal control time period segmentation demand input in natural language form, and parse it to obtain the target intersection, optimization date type and optimization demand type; Historical traffic data query: Automatically determines the skill set and call path based on optimization demand type and skill prerequisites; Traffic flow classification statistics and traffic flow-time distribution curve plotting: Obtain lane-level 5-minute historical traffic flow data, intersection channelization information, phase scheme data and holiday calendar data from traffic data sources, determine the statistical granularity according to the optimization demand type, and generate traffic flow-time distribution curves; Time period segmentation and result presentation: Input the traffic-time distribution curve and task template into the multimodal large model to obtain candidate time period segments, and then use the structured traffic sequence to perform comprehensive traffic construction, smoothing, difference calculation, coefficient of variation calculation and significance calculation, refine the candidate boundary, and form the final time period segmentation result and overlay display map.

2. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 1, characterized in that, The implementation of user demand identification and intent parsing includes: A pre-defined standardized natural language input template is used to guide users to perform standardized input. Users can either fill in the blanks based on the standardized natural language input template or express themselves freely in colloquial language without using the standardized natural language input template. The large-scale model is used for semantic understanding and key information extraction to complete target intersection name recognition, optimized date type recognition, and optimized demand type recognition. Specifically: when recognizing target intersection names, the large-scale model processes aliases, abbreviations, and ambiguous expressions; when recognizing optimized date types, the large-scale model understands time descriptions and makes judgments based on a date type configuration library; when recognizing optimized demand types, the large-scale model matches and classifies user input with preset demand types based on its semantic understanding capabilities. The system performs legality and completeness checks on the extracted target intersection name, optimization date type, and optimization requirement type according to preset rules. When key parameters are missing, an interactive follow-up questioning mechanism is automatically triggered to guide the user to supplement the necessary information.

3. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 1, characterized in that, The implementation of the traffic classification statistics and traffic-time distribution curve plotting includes: the step of determining the classification statistics dimension, the step of inlet-turn classification, and the step of plotting the traffic-time distribution curve, wherein: The step of determining the classification and statistical dimensions automatically determines the corresponding traffic volume statistical granularity based on the type of optimization requirement input by the user. According to the steps of the approach lane-turning classification, the lanes of each approach lane are classified for turning, and then the traffic volume is counted according to the approach lane-turning. The classification is based on the functional markings of each lane in the intersection channelization design, while also considering the constraints of lanes that need to be released simultaneously due to channelization restrictions. The steps for drawing the flow-time distribution curve are based on the classification statistics results. With time as the horizontal axis and traffic volume as the vertical axis, the flow-time distribution curves of the classification statistics are drawn, including intersection-level flow-time distribution curves, phase-level flow-time distribution curves, and approach lane-turning-level flow-time distribution curves, corresponding to different optimization demand types.

4. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 3, characterized in that, In the step of determining the classification statistical dimension: When the optimization requirement type is to support time-segmented periodic optimization: Calculate the traffic volume every 5 minutes at the intersection, and sum the traffic flow of all lanes within the intersection at the same time granularity to obtain the intersection-level traffic volume every 5 minutes; When the optimization requirement type is to support time-segmented green light ratio and cycle optimization: traffic volume is counted every 5 minutes according to phase, the lane set corresponding to each phase is determined according to the intersection phase design scheme, and the traffic flow of each lane in the same phase within the same time granularity is summed to obtain the phase-level traffic volume every 5 minutes. When the optimization requirement type is to support the full-element optimization of time-segment phase design and timing scheme: calculate the traffic volume every 5 minutes according to the approach lane-turning statistics, and complete the lane turning classification and traffic volume summation according to the intersection channelization design and approach lane-turning classification method to obtain the approach lane-turning level traffic volume every 5 minutes.

5. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 3, characterized in that, The specific steps of the import lane-steering classification include: Step 1: Initialize the collection With sets : Traversing the import lane All lane function signs: If a lane contains two or more turns, then the current lane is added to the set. ; Otherwise, if the lane is a single-turn lane and not a right turn, then the current lane is added to the set. ; Step 2, if set If empty, then there is no multi-turn shared lane, and the entrance lane is empty. Directly by set The individual steering elements are categorized, with each different steering element forming a separate category, and the process ends here. If set If not empty, then for the set Perform clustering and grouping of sets Each lane is considered a set of steering functions, and they are grouped according to the principle of "connected lanes with the same steering": If the turning sets of two lanes intersect, then the two lanes belong to the same group. By transitive closure, the collection Divide into several non-overlapping groups , , …, , ; Each group Steering combination Take the union of all lane turning sets within the group. If a lane intersects with multiple groups, then these groups are merged. Ultimately, the intersection of any two different groups' turning combinations is empty; Step 3: Process the set Merging with groups: If set If the value is empty, proceed directly to step 4 to determine the final category; If set Not empty, iterate through the set For each lane in the system, the current lane's turning direction is recorded as... For each lane, the following judgment is made; Check steering Is it included in a group? Steering combination middle: If such a group exists, then the current lane will be removed from the set. Remove from the list and add to the corresponding group. ; If no group contains a steering Then the current lane will be retained in the set. middle; After the traversal is complete, we obtain the updated groups and the updated sets. ; Step 4: Determine the final steering category of the approach lane: Case 1: If set Not empty: set The lanes are classified according to their single steering function, with each different steering direction forming a separate category; Import channel The steering categories include: each group Corresponding composite steering classes and sets Each different single steering type corresponds to a single steering class; Case 2: If the set Empty: Import channel The steering category is determined solely by the updated grouping: if the number of groups is 1, there is only one category, namely the composite steering category of that group; if the number of groups is greater than 1, there are multiple composite steering categories.

6. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 1, characterized in that, A two-stage segmentation strategy is adopted for time period division. The first stage is the image-driven preliminary segmentation stage based on a multimodal large model. Taking the traffic-time distribution curve as input, the visual understanding capability of the multimodal large model is used to analyze the overall shape and micro-changes of the curve and output candidate time period segmentation schemes. The second stage is the boundary point fine-tuning stage based on the fusion of large model inference and data analysis. Taking the preliminary segmentation results obtained in the first stage and the original structured traffic data as input, the boundary of the preliminary segmentation is fine-tuned through data quantitative analysis.

7. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 6, characterized in that, In the first stage, the large model reasones according to the following analytical steps: Observe the overall shape of the curve, identify obvious single-peak, double-peak or multi-peak structures, determine the approximate time range of the morning peak and evening peak, and determine the nighttime low peak period; Identify key parameters including the starting point, peak point and plateau period, pullback point, and fluctuation threshold, among which: The starting point is the point in time when the flow rate suddenly begins to rise continuously from a low value; Peak points and plateau periods are the time periods when the highest traffic occurs. If the peak lasts for a long time and there is a significant dip in the middle, consider whether to split it. The pullback point is the turning point where the flow rate drops significantly from a high level and then remains at a new level. By using fluctuation thresholds, we can focus on significant trend changes lasting longer than 15-20 minutes to ignore spikes; Perform time period merging and segmentation, further including: Perform peak-hour merging: If traffic changes are minimal from late night to early morning, merge them into a longer time period; Peak traffic segmentation: During morning and evening peak hours, traffic changes rapidly. The traffic flow is segmented according to the characteristics of "climbing-peak-falling" with a time granularity of 15-30 minutes. Peak-hour consolidation: During off-peak hours, if traffic fluctuates but the overall volume remains consistent, it can be merged into one or two large time periods. Separate transition periods: If the flow rate decreases linearly during the transition from peak to off-peak periods, a separate transition period is defined. When making comprehensive decisions based on multiple curves, if the graph contains multiple curves, the "greatest common divisor" or "bottleneck orientation" is used: if any critical curve undergoes a significant change in shape, the current time point is taken as a potential dividing point; if the peak times in each direction are misaligned, the current time point is retained to ensure that all directions can obtain a suitable control scheme.

8. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 6, characterized in that, The second stage includes steps of data preprocessing and indicator definition, steps of stepwise reasoning and correction based on data, steps of total quantity control and mandatory merging, and steps of generating the final scheme, wherein: The data preprocessing and indicator definition steps are used to achieve: Comprehensive traffic flow calculation: Calculate the comprehensive traffic flow at the intersection every 5 minutes. This is the sum of traffic for each category every 5 minutes: in, Indicates the first Traffic volume for a specific category within 5 minutes This indicates the total number of categories. Smoothing: For the overall flow rate every 5 minutes and each The sequence is processed using a 3-point moving average. Calculate key metrics, which further include: difference The difference between the smoothed combined flow and the flow of each category is used to identify abrupt changes in flow. ,in, for The sum of the flow rate every 5 minutes after the 3-point moving average processing; coefficient of variation Standard deviation of flow rate within a time period with the mean The ratio, , Considered highly stable, This is considered highly volatile and requires consideration of splitting. Significance threshold: Calculate the rate of change of the mean between adjacent time periods. , ,like Furthermore, a duration greater than 15 minutes is considered a valid segmentation criterion, and if the average change rate of adjacent time periods is... These are considered strong significant differences that must be separated. The step-by-step reasoning and correction based on data includes: Nighttime / off-peak period treatment: Discard the coefficient of variation Instead, use absolute range. For low-flow periods, if the absolute range is within an acceptable range, it should be merged into a long period. Find the moment when the flow rate increases monotonically for three consecutive cycles and the total flow rate exceeds a specific threshold, and use this as the end point of the low-peak period. Peak start-up and ramp-up period: Check the difference If a point of significant increase occurs, and the initial boundary is drawn too early or too late, the boundary should be fine-tuned according to the principle of "continuous monotonic change for 3 consecutive periods". Peak period segmentation judgment: Calculate the coefficient of variation within the peak period. ,like Furthermore, there are obvious "double peaks" or "misaligned peaks". When assessing whether to split the peaks, the principle of engineering simplification should be followed: if splitting the peaks would result in too many time periods and the overall load is high, it is preferable not to split the peaks, but the start and end times must be accurately defined. Determining the merging of peak and decline periods: Calculate the mean difference rate between adjacent periods. If the mean difference is <15% and the overall trend is stable, it is recommended to merge them to simplify the scheme; if the mean difference is >25%, they must be separated. The total quantity control and mandatory merging steps are executed according to the following priority: The mandatory merging strategy is followed by the following priority: Prioritize merging adjacent time periods with the smallest difference in average traffic volume; The second-best approach is to merge the "early morning rush hour" with the "climbing period". The second merger will combine the "decline period" with the subsequent "off-peak period"; The final measure is to merge "evening peak start" and "evening peak peak"; The final time period segmentation points are determined through the steps of generating the final solution, and each time period is assigned a specific project name.

9. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 1, characterized in that, The skill orchestration and scheduling mechanism automatically determines the skill invocation path based on task parameters, optimization requirement type, skill prerequisites, and intermediate result status, including the following steps: The system receives user input regarding traffic control time period segmentation requirements. Through user requirement identification and intent parsing skills, it analyzes the target intersection, optimizes the date type and requirement type, and forms standardized task parameters. The orchestration mechanism verifies the completeness of the task parameters, and when parameters are missing, it calls parameter verification and completion skills to supplement them. The scheduling mechanism determines the set of skills required for this task based on the task parameters. For all tasks, it calls the historical data query skill to obtain lane-level 5-minute historical traffic data. For tasks that need to be statistically analyzed by phase, it further calls the intersection phase design scheme query skill. Different statistical paths are selected based on the type of optimization needs. Among them, periodic optimization corresponds to intersection-level statistics, green ratio and periodic optimization correspond to phase-level statistics, and all-element optimization corresponds to entrance lane-turning level statistics. The skills are connected through standardized input and output. The output of the upstream skill serves as the input of the downstream skill. After each skill is executed, its output result is judged. Only when the current data meets the input requirements of the subsequent skill will the next skill be triggered.

10. The signal control time period segmentation method based on multimodal large model and skill orchestration as described in claim 9, characterized in that, It also includes exception handling mechanisms and feedback optimization mechanisms, among which: During the skill arrangement process, an exception handling mechanism is used to handle exceptions based on the intermediate result status: when the user inputs missing key parameters, the system returns a parameter supplementation prompt; when historical traffic data is missing or of insufficient quality, the system prompts a data exception and can replace it with the same type of historical date or use existing data to supplement it; when the curve fluctuates abnormally or the candidate boundaries are too dense, the system can merge the boundaries according to the minimum time period length and traffic stability rules; when the division result is inconsistent with the user's experience, the system submits the result to the user or expert for confirmation and correction. After users or experts adjust the time period division results, the adjustment content and reasons are recorded using a feedback optimization mechanism and used as feedback information. The feedback information is used to analyze the differences between the system division results and the manually confirmed results, and to further optimize the task templates, threshold parameters and skill arrangement rules.

Images