Intelligent traffic data processing method and system based on edge computing

Abstract

The present invention provides a method and system for processing intelligent traffic data based on edge computing, the method comprising: collecting multi-source traffic data through edge computing nodes, preprocessing the data, and obtaining preprocessed data. The preprocessed data is subjected to spatiotemporal alignment and data association processing to generate an associated data set, and feature extraction and preliminary analysis are completed at the edge computing node to obtain local analysis results. The local analysis results are transmitted to a cloud server for deep data mining and pattern recognition to generate global analysis results. Based on the global analysis results, a local traffic control strategy is generated at the edge computing node, and a global traffic optimization plan is generated at the cloud server. The local traffic control strategy is sent to traffic facilities for execution, and the global optimization plan is distributed to each edge computing node to achieve collaborative control. The present invention combines the advantages of edge computing and cloud processing to effectively improve the real-time performance of traffic data processing and the overall optimization capability of the intelligent transportation system.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and in particular to a smart traffic data processing method and system based on edge computing. Background Art

[0002] With the continuous growth of urban traffic demand, intelligent transportation systems have become an important means to optimize traffic flow and improve management efficiency. The current intelligent transportation data processing methods usually rely on cloud computing, uploading all traffic data to the cloud for analysis and decision-making. However, this centralized processing mode is prone to insufficient response speed due to data transmission delays and bandwidth bottlenecks when facing massive real-time traffic data. Especially in dynamic and complex traffic environments, it cannot meet the real-time requirements of data processing, affecting the accuracy and efficiency of local traffic control. Summary of the invention

[0003] The main purpose of the present invention is to solve the technical problems that the existing intelligent traffic data processing methods rely on cloud computing, resulting in high data transmission delay, insufficient real-time performance, and difficulty in meeting the needs of local traffic control in dynamic and complex traffic environments;

[0004] A first aspect of the present invention provides a method for processing intelligent traffic data based on edge computing, and the method for processing intelligent traffic data based on edge computing includes:

[0005] Collect multi-source traffic data through a preset edge computing node, and perform data preprocessing on the multi-source traffic data to obtain preprocessed data;

[0006] Performing spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and performing feature extraction and preliminary analysis on the associated data set at an edge computing node to obtain a local analysis result;

[0007] Transmitting the local analysis results to a cloud server for deep data mining and pattern recognition to obtain global analysis results, and based on the global analysis results, generating a local traffic control strategy at the edge computing node and a global traffic optimization solution at the cloud server;

[0008] The local traffic control strategy is sent to the corresponding traffic facilities for execution, and the global traffic optimization plan is distributed to each edge computing node for collaborative control.

[0009] Optionally, in a first implementation of the first aspect of the present invention, the performing spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and performing feature extraction and preliminary analysis on the associated data set at an edge computing node to obtain a local analysis result includes:

[0010] Identifying a timestamp of the preprocessed data, and aligning the preprocessed data based on the timestamp to obtain time-synchronized data;

[0011] Based on a predefined spatial grid model, spatially mapping and matching the time synchronization data to obtain spatiotemporal alignment data;

[0012] Utilizing a local knowledge base to contextualize the spatiotemporal alignment data, establish logical connections between data items, and obtain a correlation data set;

[0013] A lightweight machine learning algorithm is applied to the associated data set to extract key features to obtain feature data, and a lightweight traffic state estimation model deployed on the edge node is used to process the feature data and generate preliminary analysis results.

[0014] Optionally, in a second implementation of the first aspect of the present invention, the using of a local knowledge base to contextually associate the spatiotemporal alignment data, establish logical connections between data items, and obtain an associated data set includes:

[0015] Performing spatial mapping on the spatiotemporal alignment data according to a predefined local road network topology structure to obtain topological association data;

[0016] Extracting traffic elements from the topological association data, and applying a semantic annotation algorithm based on a traffic concept dictionary to the traffic elements, giving the traffic elements semantic attributes, and obtaining semantic annotation data;

[0017] Using graph embedding technology to vectorize the traffic elements in the semantically labeled data to obtain a multi-dimensional feature vector;

[0018] Calculating a similarity matrix between traffic elements based on the multidimensional feature vector, and grouping similar traffic elements using a hierarchical clustering algorithm to obtain an element clustering result;

[0019] Association rules are constructed according to the element clustering results, and the association rules are applied to the topological association data to obtain an association data set.

[0020] Optionally, in a third implementation of the first aspect of the present invention, the transmitting the local analysis result to a cloud server for deep data mining and pattern recognition to obtain a global analysis result includes:

[0021] Performing spatiotemporal alignment and data fusion on local analysis results from multiple edge nodes to obtain a global data set, and performing time-frequency domain decomposition on the global data set using a multi-scale wavelet transform to obtain a multi-level feature representation;

[0022] Performing unsupervised learning on the multi-level feature representation through a preset deep belief network, extracting high-order abstract features, and obtaining a deep feature vector;

[0023] Applying a pre-trained dynamic Bayesian network to the deep feature vector to perform traffic state reasoning to obtain a current traffic state estimate;

[0024] The current traffic state estimate is input into the long short-term memory network model to obtain the future traffic state prediction result, and the future traffic state prediction result, the current traffic state estimate and the historical traffic data are combined to perform a multi-dimensional comparative analysis to obtain a global analysis result.

[0025] Optionally, in a fourth implementation of the first aspect of the present invention, generating a local traffic control strategy at an edge computing node and generating a global traffic optimization solution at a cloud server based on the global analysis result includes:

[0026] Applying a multi-objective genetic algorithm to the global analysis result to generate a set of candidate global traffic optimization solutions, and selecting the best solution from the set of candidate global traffic optimization solutions based on predefined evaluation indicators to obtain an initial optimization solution;

[0027] Using traffic flow theory to perform a macroscopic evaluation on the initial optimization scheme, screen out a subset of feasible schemes, obtain optimization candidate schemes, and input the optimization candidate schemes into a pre-trained reinforcement learning model, evaluate each candidate scheme, and obtain a scheme scoring list;

[0028] Based on the scheme scoring list, a Monte Carlo tree search algorithm is used to explore the decision space, select the optimal action sequence, obtain an optimized action plan, and convert the optimized action plan into specific traffic control parameters to obtain a global traffic optimization plan;

[0029] Based on the global traffic optimization scheme, a graph segmentation algorithm is used to decompose the preset global optimization problem into multiple sub-problems to obtain a regional control task, and a model predictive control algorithm is applied to the regional control task at the edge computing node to generate a control strategy that adapts to local traffic characteristics and obtain a local traffic control strategy.

[0030] Optionally, in a fifth implementation of the first aspect of the present invention, based on the global traffic optimization solution, a graph partitioning algorithm is used to decompose the preset global optimization problem into multiple sub-problems to obtain a regional control task, and a model predictive control algorithm is applied to the regional control task at an edge computing node to generate a control strategy adapted to local traffic characteristics, and the local traffic control strategy includes:

[0031] Performing graph-structured processing on the global traffic optimization scheme, representing the traffic network as a weighted graph, and obtaining a traffic network graph model;

[0032] Applying a spectral clustering algorithm to the traffic network graph model, segmenting the graph based on node connectivity and edge weights, and obtaining multiple subgraph structures;

[0033] For each subgraph structure, extract local traffic characteristics and constraints, construct sub-problem descriptions, and obtain regional control tasks;

[0034] Assigning the regional control tasks to corresponding edge computing nodes, and constructing a local state space model on each node to obtain a dynamic system representation;

[0035] A model predictive control algorithm is applied to the dynamic system representation, and a control input sequence is calculated through a rolling horizon optimization to obtain a local traffic control strategy.

[0036] Optionally, in a sixth implementation of the first aspect of the present invention, the sending of the local traffic control strategy to corresponding traffic facilities for execution, and the distributing of the global traffic optimization solution to each edge computing node for collaborative control includes:

[0037] Decomposing the local traffic control strategy into instructions, generating specific control instruction sequences for different types of traffic facilities, and obtaining a device-level control instruction set;

[0038] Using the local communication interface of the edge node, the device-level control instruction set is sent to the corresponding traffic facility execution module to obtain execution confirmation information;

[0039] Distributed encoding is performed on the global traffic optimization plan, and the plan is converted into a compressed data format suitable for edge node processing to obtain an optimization plan data packet;

[0040] Through the preset edge computing network topology, the optimization solution data packet is distributed to each edge computing node using a multi-hop routing method to obtain node optimization data;

[0041] On each edge computing node, the received node optimization data is decoded and localized, and the global optimization objective is converted into local control constraints to obtain collaborative control parameters.

[0042] A second aspect of the present invention provides an edge computing-based smart traffic data processing system, the edge computing-based smart traffic data processing system comprising:

[0043] A preprocessing module, used to collect multi-source traffic data through a preset edge computing node, and perform data preprocessing on the multi-source traffic data to obtain preprocessed data;

[0044] A local analysis module is used to perform spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and perform feature extraction and preliminary analysis on the associated data set at an edge computing node to obtain a local analysis result;

[0045] A cloud processing module is used to transmit the local analysis results to a cloud server, perform deep data mining and pattern recognition, obtain global analysis results, and generate local traffic control strategies at edge computing nodes and global traffic optimization solutions at the cloud server based on the global analysis results;

[0046] The collaborative control module is used to send the local traffic control strategy to the corresponding traffic facilities for execution, and distribute the global traffic optimization plan to each edge computing node for collaborative control.

[0047] The above-mentioned intelligent traffic data processing method and system based on edge computing collects multi-source traffic data through edge computing nodes, pre-processes the data, and obtains pre-processed data. The pre-processed data is subjected to spatiotemporal alignment and data association processing to generate an associated data set, and feature extraction and preliminary analysis are completed at the edge computing node to obtain local analysis results. The local analysis results are transmitted to the cloud server for deep data mining and pattern recognition to generate global analysis results. Based on the global analysis results, a local traffic control strategy is generated at the edge computing node, and a global traffic optimization plan is generated at the cloud server. The local traffic control strategy is sent to the traffic facilities for execution, and the global optimization plan is distributed to each edge computing node to achieve collaborative control. The present invention combines the advantages of edge computing and cloud processing to effectively improve the real-time nature of traffic data processing and the overall optimization capability of the intelligent transportation system.

[0048] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic diagram of a first embodiment of a method for processing intelligent traffic data based on edge computing in an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of an embodiment of an intelligent traffic data processing system based on edge computing in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0054] To facilitate understanding of this embodiment, a method for processing intelligent traffic data based on edge computing disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 As shown, the method comprises the following steps:

[0055] 101. Collect multi-source traffic data through a preset edge computing node, and perform data preprocessing on the multi-source traffic data to obtain preprocessed data;

[0056] In one embodiment of the present invention, the process of collecting multi-source traffic data through a preset edge computing node first involves the deployment and configuration of the edge node. These edge computing nodes are usually installed in a traffic light control box, a roadside unit or a dedicated edge server, and have certain computing power and storage space. Each node is equipped with a variety of data acquisition interfaces, including wired network interfaces (such as Ethernet) and wireless communication modules (such as 4G / 5G, Wi-Fi, ZigBee, etc.) to adapt to different types of data sources. During the data collection process, the edge node first establishes a connection with various traffic sensing devices, including but not limited to vehicle detectors, cameras, radars, signal light controllers, and environmental sensors. After the connection is established, the node continues to receive data streams through a predefined communication protocol (such as MQTT, CoAP, or a custom lightweight protocol). In order to handle a large number of concurrent data streams, the node adopts an asynchronous I / O model, such as the epoll mechanism under Linux, to achieve efficient multiplexing. After the data is received, preliminary formatting is performed immediately to uniformly convert the raw data from different sources into a standard intermediate format, such as JSON or Protocol Buffers, for subsequent processing. During this process, the edge node will also perform a real-time integrity check on the data, discard obviously erroneous or incomplete data packets, and ensure the data quality for subsequent analysis. Data preprocessing of multi-source traffic data is to improve data quality and reduce the complexity of subsequent processing. The preprocessing process first performs data cleaning, including removing duplicate data, processing missing values ​​and outliers. For duplicate data, the system uses sliding window technology combined with hash algorithm to quickly identify and delete data entries that appear repeatedly in a short period of time. Missing value processing uses multiple interpolation methods, and selects appropriate interpolation strategies according to the data type, such as linear regression interpolation for numerical data and nearest neighbor interpolation for categorical data. Outlier detection uses a combination of statistical-based methods and machine learning methods, such as using the Z-score method to identify values ​​far from the mean, and using the isolation forest algorithm to detect outliers in the multidimensional feature space. Detected outliers are corrected or marked according to predefined rules. Data standardization is another key step in preprocessing. After standardization, data from different sources are converted to the same scale range, usually using Min-Max scaling or Z-score standardization methods. For time series data, the system also performs time alignment to unify data with different sampling rates to a fixed time interval, usually using linear interpolation or spline interpolation methods. In addition, for high-dimensional data, preprocessing also includes preliminary dimensionality reduction operations, such as using principal component analysis (PCA) or autoencoders to extract main features to reduce the computational burden of subsequent processing. Finally, preprocessing also includes data encoding, converting categorical variables into numerical representations. Common methods include one-hot encoding and label encoding.After these preprocessing steps, the raw multi-source traffic data is converted into a structured, clean, and standardized dataset, namely the preprocessed data.

[0057] 102. Based on the standardized data structure, the camera parameters and the lighting environment parameters of the scene where the building block model is located are set respectively to obtain the scene rendering parameters;

[0058] In one embodiment of the present invention, the pre-processed data is subjected to spatiotemporal alignment and data association to obtain an associated data set, and the associated data set is subjected to feature extraction and preliminary analysis at an edge computing node to obtain a local analysis result, including: identifying the timestamp of the pre-processed data, and aligning the pre-processed data based on the timestamp to obtain time synchronization data; spatially mapping and matching the time synchronization data based on a predefined spatial grid model to obtain spatiotemporal aligned data; contextually associating the spatiotemporal aligned data using a local knowledge base to establish logical connections between data items to obtain an associated data set; applying a lightweight machine learning algorithm to the associated data set to extract key features to obtain feature data, and using a lightweight traffic state estimation model deployed on the edge node to process the feature data and generate preliminary analysis results.

[0059] Specifically, the process of identifying the timestamps of preprocessed data and aligning them first involves data parsing. The system uses regular expression matching methods to extract time information from data in different formats. For example, for data in JSON format, JSONPath expressions such as "$.timestamp" are used to directly extract the time field; for CSV format, time information is obtained through predefined column indexes. The system uses Python's datetime library to uniformly convert these different formats into UTC time. The time alignment process uses fixed-size time windows, such as 5-minute intervals, to group data. For each time window, the system calculates the average timestamp of all data in the window as the representative time of the window. For data sources with lower sampling rates, linear interpolation methods are used to fill in missing time points. Specifically, if data is missing in a time window, the values ​​of the data source in adjacent windows are used for interpolation calculations. This method ensures the consistency of data with different sampling rates in the time dimension. The resulting time-synchronized data is a data structure containing multiple time windows, and the data items in each window have a unified timestamp.

[0060] Specifically, the process of spatial mapping and matching based on the predefined spatial grid model first requires the construction of a spatial index structure. The system uses a quadtree data structure to recursively divide the urban area into grids of different sizes. Each leaf node represents a minimum grid cell, which is usually set to 100 meters by 100 meters. For each data item, the system first checks whether it contains accurate GPS coordinates. If so, the spatial index structure is directly used to locate the corresponding grid cell. For data that lacks accurate coordinates, the system uses a fuzzy matching algorithm. For example, for data with only a section ID, the system maintains a mapping table from the section ID to the grid cell, and determines the spatial position of the data item by looking up the table. When processing intersection data, the system considers that an intersection may span multiple grid cells, so the data is copied to all relevant grids. In order to deal with the GPS drift problem, the system implements a trajectory smoothing algorithm based on Kalman filtering. The algorithm uses historical location information and road network topology to correct the position of each data point and improve the accuracy of spatial matching. After these processes, each data item is accurately located in one or more spatial grid cells to form spatiotemporal alignment data.

[0061] Specifically, the process of using local knowledge bases for context association involves complex semantic processing and reasoning. The local knowledge base is stored in a graph database (such as Neo4j) and contains information such as the topological structure of the traffic network, road attributes, and historical traffic patterns. The system first matches the elements in the spatiotemporal aligned data with the entities in the knowledge base through an entity linking algorithm. For example, the text description "Wudaokou intersection" is linked to the corresponding node in the knowledge base. This process uses the TF-IDF algorithm to calculate text similarity and combines the edit distance metric to handle synonyms and spelling variants. Next, the system applies a graph traversal algorithm to explore the relationship between entities. Specifically, the depth-first search (DFS) algorithm is used to explore the neighbor nodes in the knowledge graph from each matched entity, with a maximum depth of 3. This process can discover implicit relationships between data items, such as congestion in the upstream section may affect the traffic conditions of the downstream intersection. In addition, the system also implements a rule-based reasoning engine, which contains a series of predefined IF-THEN rules for inferring complex traffic conditions. For example, "If the traffic volume on the main road suddenly decreases, and the traffic volume on the peripheral branches increases, there may be traffic control." These rules are executed on the data set through the forward chaining algorithm to generate new association information. The resulting association data set is a network structure rich in contextual information, where each node represents a data item and the edges represent various relationships between data items.

[0062] Specifically, the process of feature extraction and preliminary analysis of the associated data set uses a combination of multiple lightweight algorithms. First, the system uses a sliding window method (window size is 1 hour, step size is 5 minutes) to segment the time series data. For the data in each window, statistical features such as mean, variance, kurtosis, skewness, etc. are calculated. These features are efficiently calculated through vectorized operations of the NumPy library. Then, the system applies fast Fourier transform (FFT) to extract frequency domain features to capture the periodic pattern of traffic flow. To reduce the dimensionality, the system implements an approximate principal component analysis algorithm based on random projection, which has a time complexity of O(nd log d), where n is the number of samples and d is the number of features. In the feature selection stage, the system uses a lightweight decision tree model based on the CART algorithm to calculate the feature importance score and selects the top k features with the highest scores (k is usually set to 20% of the original number of features). For time series features, the system implements a simplified version of the ARIMA model, uses the Levenberg-Marquardt algorithm to optimize the model parameters, and sets the prediction window to 15 minutes. These extracted features are organized into a feature vector as the input of a lightweight traffic state estimation model. The model is a three-layer fully connected neural network that uses the ReLU activation function and the output layer uses the Softmax function to classify the current traffic state. The model is deployed on edge devices using the TensorFlow Lite framework, and quantization technology is used to compress the model size to ensure real-time inference performance. The output of the model includes discrete state estimates such as traffic flow level, average speed range, congestion level, and the probability distribution of these state changes in the next 15 minutes. These preliminary analysis results provide key data support for subsequent traffic management decisions.

[0063] Furthermore, the use of a local knowledge base to contextually associate the spatiotemporal alignment data and establish logical connections between data items to obtain an associated data set includes: spatially mapping the spatiotemporal alignment data according to a predefined local road network topology structure to obtain topologically associated data; extracting traffic elements from the topologically associated data, and applying a semantic annotation algorithm based on a traffic concept dictionary to the traffic elements to assign semantic attributes to the traffic elements to obtain semantically annotated data; using graph embedding technology to vectorize the traffic elements in the semantically annotated data to obtain a multidimensional feature vector; calculating a similarity matrix between traffic elements based on the multidimensional feature vector, and using a hierarchical clustering algorithm to group similar traffic elements to obtain an element clustering result; constructing association rules based on the element clustering result, and applying the association rules to the topologically associated data to obtain an associated data set.

[0064] Specifically, the process of spatial mapping of spatiotemporal aligned data according to the predefined local road network topology first involves building an efficient spatial index. The system uses an R-tree data structure to represent the road network topology, with each leaf node corresponding to a road segment or intersection. During the construction of the R-tree, the geographic space is divided into hierarchical rectangular regions using the minimum bounding rectangle (MBR) algorithm to ensure fast spatial query performance. During spatial mapping, for each spatiotemporal aligned data point, the system uses the range query function of the R-tree to find the nearest road element. Specifically, the system first calculates the vertical distance from the data point to each candidate road segment and selects the road segment with the smallest distance as the mapping target. For data points near intersections, the system sets a buffer (usually 20 meters), and the points falling within the buffer are mapped to the corresponding intersection nodes. When processing data of moving targets (such as vehicle trajectories), the system uses a hidden Markov model (HMM) for map matching, considering the spatial correlation of consecutive time points to improve mapping accuracy. For data points with GPS drift, the system implements a trajectory smoothing algorithm based on particle filtering to correct the current position estimate through historical position information and road network constraints. The output of this spatial mapping process is topologically linked data, in which each data point is associated with a specific road network element (road segment or intersection), while retaining the original spatiotemporal information and attribute data.

[0065] Specifically, the process of extracting traffic elements from topologically associated data and semantically annotating them involves complex natural language processing and knowledge representation techniques. First, the system uses predefined rule templates to identify and extract key traffic elements from topologically associated data, such as vehicles, pedestrians, signal light status, road conditions, etc. These rules are implemented based on regular expressions and syntax parsers, and can handle structured and semi-structured data formats. The extracted traffic elements are then input into a semantic annotation algorithm based on a traffic concept dictionary. The algorithm uses a pre-built traffic domain ontology as the knowledge basis, which defines the hierarchy, attributes, and relationships of traffic concepts. The semantic annotation process uses a conditional random field (CRF) model, which captures contextual information through feature functions and assigns the most appropriate semantic label to each traffic element. Feature functions include word form, part of speech, location information, and matching degree with concepts in the ontology. In order to deal with ambiguity, the system also integrates a word sense disambiguation module, which uses context windows and attention mechanisms to select the most appropriate concept interpretation. During the annotation process, the system not only assigns category labels to elements, but also assigns attribute values ​​according to the ontology definition, such as vehicle type, speed, direction, etc. For complex traffic scene descriptions, the system uses dependency parsing technology to build a relationship graph between elements and capture more fine-grained semantic information. The resulting semantically annotated data is a structured representation rich in domain knowledge, with each traffic element being given clear semantic attributes and relationships.

[0066] Specifically, the use of graph embedding technology to vectorize the traffic elements in the semantically annotated data is to convert the semantic information into a numerical form that can be directly processed by the machine learning algorithm. The system uses the DeepWalk algorithm as the basic graph embedding method, which generates node sequences by performing random walks on the graph structure, and then uses the Skip-gram model to learn the distributed representation of the nodes. In order to better capture the characteristics of the traffic network, the system improves DeepWalk and introduces a weighted random walk strategy based on the attention mechanism. Specifically, the attributes of the edges (such as road grade and capacity) are considered during the walk to adjust the transition probability to ensure that the generated sequence is more consistent with the actual traffic flow characteristics. For each traffic element, the system performs multiple random walks, each with a length of 100 steps, to generate a large number of sequences representing local structures. These sequences are input into a Skip-gram model with hierarchical Softmax for training. The hidden layer size of the model is set to 128, and negative sampling technology is used to accelerate the training process. In order to integrate semantic annotation information, the system adds a multi-task learning framework based on the original DeepWalk, which simultaneously predicts the context and semantic attributes of nodes, so that the learned embedding can represent both the network structure and the semantic features. For dynamically changing traffic data, the system implements an incremental learning mechanism to regularly update the embedding representation to adapt to changes in traffic patterns. The resulting multidimensional feature vector is a 128-dimensional real-valued vector, where each dimension encodes a potential feature of a traffic element in the network structure and semantic space.

[0067] Specifically, the process of calculating the similarity matrix between traffic elements based on multidimensional feature vectors and grouping them using a hierarchical clustering algorithm aims to discover the intrinsic structure and patterns in the traffic network. First, the system calculates the pairwise distances between feature vectors, using cosine similarity as the main metric because it is insensitive to vector length and suitable for comparing similarities in directions. To improve computational efficiency, the system uses an approximate nearest neighbor search algorithm (such as local sensitive hashing LSH) to accelerate large-scale similarity calculations. The resulting similarity matrix is ​​an N×N symmetric matrix, where N is the number of traffic elements and each element represents the degree of similarity between two traffic elements. Next, the system applies a hierarchical clustering algorithm to group these elements. Considering the complexity of the traffic network, a bottom-up agglomerative hierarchical clustering method is selected. The algorithm starts with each element as a separate cluster and gradually merges the most similar clusters until a predetermined number of clusters is reached or a stopping condition is met. Ward's method is used as the inter-cluster distance metric during the merging process, which tends to produce clusters of similar size and is suitable for balanced traffic partitioning. In order to handle large-scale data, the system implements an approximate hierarchical clustering algorithm based on a minimum spanning tree, which significantly reduces the computational complexity. During the clustering process, the system dynamically adjusts the number of clusters and uses the Silhouette Coefficient as an evaluation indicator to find a balance between intra-cluster similarity and inter-cluster differences. The final element clustering result is a hierarchical structure that contains traffic element groupings of different granularities, reflecting the functional areas and similar patterns in the traffic network.

[0068] Specifically, the process of constructing association rules based on the results of element clustering and applying them to topological association data is to discover and utilize the potential relationships between traffic elements. First, the system performs feature analysis on each cluster and extracts the typical attributes and patterns of the category. These features are converted into a series of IF-THEN rules through the decision tree algorithm, and each rule describes the common characteristics of the elements within the cluster. For example, "If the road type is a main road and the average speed is less than 20km / h, it is marked as a congested section." In order to deal with the uncertainty of the rules, the system introduces fuzzy logic and uses membership functions to represent the applicability of the rules. After the rules are generated, the system uses the Apriori algorithm to mine frequent item sets within and between clusters to discover the association patterns between elements. These association patterns are converted into more advanced rules, such as "If section A is congested, section B has an 80% probability of being congested within 30 minutes." The constructed rule set is pruned to remove redundant and low-confidence rules to improve the simplicity and effectiveness of the rule base. When applying rules, the system uses the Rete algorithm to achieve efficient rule matching. The algorithm greatly reduces the computational complexity of rule evaluation by building a decision network. For real-time data streams, the system uses sliding window technology to apply rules at fixed time intervals to capture dynamically changing association relationships. The results of rule application are used to enrich the original topological association data, adding new association tags and derived attributes. The resulting association data set not only contains the original topological and semantic information, but also integrates advanced association knowledge based on data-driven discovery, providing a rich information foundation for subsequent traffic analysis and decision-making.

[0069] 103. Transmitting the local analysis results to the cloud server for deep data mining and pattern recognition to obtain global analysis results, and based on the global analysis results, generating local traffic control strategies at the edge computing node and generating global traffic optimization solutions at the cloud server;

[0070] In one embodiment of the present invention, the local analysis results are transmitted to a cloud server for deep data mining and pattern recognition to obtain global analysis results, including: performing spatiotemporal alignment and data fusion on local analysis results from multiple edge nodes to obtain a global data set, and using a multi-scale wavelet transform to decompose the global data set in the time-frequency domain to obtain a multi-level feature representation; performing unsupervised learning on the multi-level feature representation through a preset deep belief network to extract high-order abstract features to obtain a deep feature vector; applying a pre-trained dynamic Bayesian network to the deep feature vector to perform traffic state inference to obtain a current traffic state estimate; inputting the current traffic state estimate into a long short-term memory network model to obtain a future traffic state prediction result, and combining the future traffic state prediction result, the current traffic state estimate and historical traffic data to perform a multi-dimensional comparative analysis to obtain a global analysis result.

[0071] Specifically, the process of spatiotemporal alignment and data fusion of local analysis results from multiple edge nodes first involves data transmission and synchronization mechanisms. Edge nodes regularly upload local analysis results to cloud servers through secure communication protocols (such as HTTPS or dedicated VPN channels). To ensure data integrity and consistency, the system adopts a two-phase commit protocol to implement transaction management during data transmission. The data arriving at the cloud is first timestamped and the network time protocol (NTP) is used to ensure time synchronization between different edge nodes. Subsequently, the system applies the dynamic time warping (DTW) algorithm to align the time series data of different nodes to deal with time inconsistencies caused by network delays or device clock deviations. In the spatial dimension, the system uses geohash technology to map data from different regions into a unified spatial grid to facilitate subsequent spatial correlation analysis. The Dempster-Shafer evidence theory is used in the data fusion stage to comprehensively consider the reliability and conflict of different data sources to generate a unified global view. For conflicting data, the system uses weighted average or median methods to reconcile. During the fusion process, the system also applies the Kalman filter algorithm to smooth the continuous time series data to reduce the impact of noise. The resulting global dataset is a large-scale data structure that is consistent in time and space and information-fused, and contains comprehensive status information of the entire transportation network.

[0072] Specifically, the use of multi-scale wavelet transform to decompose the global data set in the time-frequency domain is to capture the characteristics of traffic data in different time scales and frequency ranges. The system uses discrete wavelet transform (DWT) as the basic algorithm, and the specific implementation uses Mallat fast algorithm to recursively decompose the signal through high-pass and low-pass filters. Considering the non-stationarity of traffic data, the system uses Daubechies wavelet as the mother wavelet, whose good localization characteristics are suitable for analyzing the sudden changes and periodic changes of traffic flow. During the decomposition process, the system performs 5-level decomposition on each time series to generate a tree structure containing approximate coefficients and detail coefficients. In order to process multi-dimensional data, the system implements tensor wavelet transform and decomposes it in both time and space dimensions. In frequency domain analysis, the system applies wavelet packet transform technology to provide more flexible frequency band division, which is particularly suitable for analyzing high-frequency components in traffic signals. In order to improve computational efficiency, the system uses lifting scheme to implement wavelet transform, which significantly reduces computational complexity and memory requirements. For long time series, the system uses sliding window technology combined with wavelet transform to achieve real-time multi-scale analysis. The decomposed multi-level feature representation contains the characteristics of traffic data at different time scales and frequency ranges, which not only retains the global trend information but also captures the local detail changes, providing rich input for subsequent deep learning and pattern recognition.

[0073] Specifically, the unsupervised learning of multi-level feature representations through a preset deep belief network is to extract high-level abstract features. The structure of the deep belief network (DBN) is composed of multiple layers of restricted Boltzmann machines (RBMs) stacked together, and each layer of RBM is a generative model that can learn the probability distribution of data. The system adopts a greedy layer-by-layer pre-training strategy to train each layer of RBM from bottom to top. Specifically, the first layer of RBM directly takes the multi-level feature representation as input, and each subsequent layer takes the hidden unit activation value of the previous layer of RBM as input. The training process uses the contrastive divergence algorithm (CD-k), and k is usually set to 1 to balance computational efficiency and model performance. In order to speed up the training process, the system implements batch training and GPU acceleration, and significantly improves the computing speed by using the CUDA parallel computing framework. Considering the temporal characteristics of traffic data, the system introduces time-delayed connections on the basis of DBN to form a dynamic DBN structure that can capture temporal dependencies. In order to handle features of different scales, the system designs a multi-resolution DBN, and features of different levels are input to network layers of different depths. After training, the system uses the back-propagation algorithm to fine-tune the entire network to further optimize the feature representation. The resulting deep feature vector is a highly abstract and compressed representation that encodes the essential characteristics and potential patterns of traffic data, and the dimension is usually set to 128 or 256, which is sufficient to capture complex traffic dynamics.

[0074] Specifically, the process of applying a pre-trained dynamic Bayesian network to deep feature vectors for traffic state reasoning involves the construction and reasoning of a probabilistic graphical model. The dynamic Bayesian network (DBN) model is pre-trained through the structural learning and parameter learning stages, where the structural learning uses a scoring search algorithm to determine the network topology, and the parameter learning uses the expectation maximization (EM) algorithm to estimate the conditional probability table. In practical applications, the system uses a sliding window method to process continuous deep feature vector sequences, and the window size is usually set to 5-10 time steps to balance real-time and contextual information capture. For each time window, the system first inputs evidence and maps the deep feature vector to the corresponding observation node in the DBN. Then, the system uses the joint tree algorithm for precise reasoning to calculate the posterior probability distribution of all hidden variables in the network. For large-scale networks, the system uses approximate reasoning methods such as particle filtering to improve computational efficiency. The reasoning process takes into account time dependency and uses a forward algorithm to pass time information. In order to deal with uncertainty, the system implements a soft evidence update mechanism that allows the input of observations with confidence. The reasoning results are post-processed, including thresholding and smoothing, to obtain a discretized traffic state estimate. The final current traffic state estimate includes the probability distribution of various key traffic parameters (such as flow, speed, density), as well as the classification results of the overall traffic conditions (such as smooth, light congestion, and severe congestion).

[0075] Specifically, the process of inputting the current traffic state estimate into the long short-term memory network model for future state prediction takes advantage of the LSTM network's advantages in dealing with long-term dependency problems. The system adopts a stacked LSTM structure, which usually contains 2-3 layers of LSTM units. The number of neurons in each layer is set according to the complexity of the problem, generally 128 or 256. The input layer receives the current traffic state estimate, and each time step corresponds to a state vector. The LSTM memory unit selectively updates and maintains long-term state information through the input gate, forget gate, and output gate mechanism. In order to improve the generalization ability of the model, the system applies dropout technology during training, and usually sets a dropout rate of 0.5 between LSTM layers. The model training adopts the back propagation through time (BPTT) algorithm, and uses the Adam optimizer to adaptively adjust the learning rate. In order to handle multi-step prediction problems, the system implements a sequence-to-sequence (seq2seq) architecture, and the encoder-decoder structure can generate outputs of different prediction durations. Considering the variability of traffic status, the system also integrates an attention mechanism to dynamically adjust the focus on historical status. The model output is post-processed, including denormalization and confidence interval calculation, to obtain specific future traffic state prediction results. These prediction results include traffic parameter estimates at multiple time points in the future (such as 15 minutes, 30 minutes, and 1 hour), as well as corresponding uncertainty measures.

[0076] Specifically, the multi-dimensional comparative analysis of future traffic state prediction results, current traffic state estimation and historical traffic data is to obtain comprehensive global analysis results. The system first constructs a multi-dimensional data cube with time, space and traffic parameters as different dimensions. Using OLAP (Online Analytical Processing) technology, the system can flexibly slice, dice and roll up and drill down to analyze data from different angles. During the comparative analysis, the system applies time series decomposition technology to decompose traffic data into trend, seasonality and residual components for comparison. Trend analysis uses exponential smoothing methods to identify long-term change patterns; seasonal analysis uses Fourier transform to detect periodic patterns; residual analysis is used to detect abnormal events. In the spatial dimension, the system uses a geographically weighted regression model to analyze the correlation of traffic states between different regions. To evaluate the accuracy of predictions, the system calculates multiple error indicators, such as mean absolute percentage error (MAPE) and root mean square error (RMSE). The system also implements a rule-based anomaly detection engine to identify potential traffic anomalies by comparing current status, prediction results and historical patterns. Finally, the system generates a comprehensive analysis report, including a key performance indicator (KPI) dashboard, trend charts, abnormal event alerts, and optimization suggestions. This global analysis result not only summarizes the current traffic situation, but also provides future trend forecasts and early warnings of potential problems, providing strong support for traffic management decisions.

[0077] Furthermore, based on the global analysis results, generating a local traffic control strategy at the edge computing node and generating a global traffic optimization plan at the cloud server includes: applying a multi-objective genetic algorithm to the global analysis results to generate a set of candidate global traffic optimization plans, and selecting the optimal plan from the set of candidate global traffic optimization plans based on predefined evaluation indicators to obtain an initial optimization plan; using traffic flow theory to perform a macro-evaluation of the initial optimization plan, screening out a subset of feasible plans, obtaining an optimized candidate plan, and inputting the optimized candidate plan into a pre-trained reinforcement learning model, evaluating each candidate plan, and obtaining a plan scoring list; based on the plan scoring list, using a Monte Carlo tree search algorithm to explore the decision space, select the optimal action sequence, obtain an optimized action plan, and convert the optimized action plan into specific traffic control parameters to obtain a global traffic optimization plan; based on the global traffic optimization plan, using a graph segmentation algorithm to decompose the preset global optimization problem into multiple sub-problems to obtain a regional control task, and applying a model predictive control algorithm to the regional control task at the edge computing node to generate a control strategy adapted to local traffic characteristics to obtain a local traffic control strategy.

[0078] Specifically, the process of applying a multi-objective genetic algorithm to the global analysis results to generate a set of candidate global traffic optimization solutions first involves the encoding of the problem and the definition of the objective function. The system encodes the control parameters of the traffic network (such as signal timing, lane allocation, speed limit, etc.) as chromosomes, and each gene represents a specific control variable. The objective function usually includes multiple dimensions, such as minimizing the overall travel time, maximizing the network throughput, minimizing emissions, etc. The algorithm uses the non-dominated sorting genetic algorithm II (NSGA-II) to deal with multi-objective optimization problems, which can effectively find the Pareto optimal solution set in the objective space. In the population initialization stage, the system generates an initial population based on the current traffic status and historical optimization experience to ensure the diversity and quality of the initial solution. During the evolution process, the system uses simulated binary crossover (SBX) and polynomial mutation operators to generate offspring. The selection operation adopts the tournament selection method to balance the selection pressure and population diversity. In order to accelerate convergence, the algorithm integrates a local search strategy to locally optimize the non-dominated solution after every few generations. The algorithm runs iteratively until the predetermined number of generations or convergence criteria are reached. The final set of candidate global traffic optimization solutions is a set of solutions distributed on the Pareto front, each of which represents a possible optimization solution under different objective trade-offs.

[0079] Specifically, the process of selecting the best solution from the set of candidate global traffic optimization solutions involves multi-criteria decision-making technology. The system first defines a set of evaluation indicators, including quantitative indicators (such as average travel time, network throughput) and qualitative indicators (such as the feasibility of solution implementation and environmental impact). For quantitative indicators, the system directly uses simulation results or mathematical models to calculate; for qualitative indicators, the system quantifies them through expert scoring or fuzzy evaluation methods. The evaluation process uses the analytic hierarchy process (AHP) to determine the weights of each indicator, taking into account the relative importance of the indicators. On this basis, the system uses the TOPSIS (top-to-bottom solution distance method) method to calculate the degree of closeness of each candidate solution to the ideal solution. The TOPSIS method first constructs a normalized decision matrix, then calculates the weighted normalized decision matrix, determines the positive ideal solution and the negative ideal solution, and finally calculates the Euclidean distance of each solution to the ideal solution. The solution with the highest score is selected as the initial optimization solution. This initial optimization solution represents the best compromise under multiple objectives and constraints, providing a good starting point for subsequent refinement and evaluation.

[0080] Specifically, the process of macro-evaluation of the initial optimization scheme using traffic flow theory involves the application of multiple traffic flow models. The system first uses macro traffic flow models, such as the LWR model (Lighthill-Whitham-Richards model), to simulate the traffic flow dynamics after the scheme is implemented. The LWR model is based on the principles of fluid dynamics and describes the spatiotemporal evolution of traffic flow through partial differential equations. In order to improve the applicability of the model, the system integrates the Fundamental Diagram of multiple types of road sections, including sections with different characteristics such as urban roads and highways. At key intersections and bottleneck locations, the system applies the queuing theory model to evaluate delays and queues. At the same time, the system discretizes the LWR model using the cellular transport model (CTM) to improve computational efficiency and numerical stability. During the evaluation process, the system also considers the dynamic changes in traffic demand and uses the dynamic traffic assignment model (DTA) to simulate the driver's path selection behavior. Through the comprehensive application of these models, the system can predict network performance indicators after the scheme is implemented, such as average speed, flow distribution, congestion level, etc. Based on the predefined threshold conditions, the system screens out a subset of feasible schemes that meet the performance requirements to form an optimization candidate scheme set. These candidate solutions are not only feasible in theory, but also verified by macro-traffic flow dynamics, laying the foundation for the next step of micro-evaluation and optimization.

[0081] Specifically, the process of inputting the optimized candidate solutions into the pre-trained reinforcement learning model for evaluation takes advantage of the advantages of deep reinforcement learning in complex decision-making problems. The system uses the deep Q network (DQN) as the basic model, which can handle high-dimensional state space and large-scale action space. The network structure of DQN includes a multi-layer convolutional neural network for feature extraction, followed by a fully connected layer for Q value estimation. The input of the model is the encoded traffic state and candidate solution parameters, and the output is the Q value corresponding to each action (i.e., control decision). In order to improve the stability and performance of the model, the system has implemented a number of improvements, including dual DQN, priority experience replay, and noise network technologies. The model training adopts temporal difference learning and uses the ε-greedy strategy to balance exploration and exploitation. The training data comes from a large number of traffic simulation experiments, covering various traffic scenarios and control strategies. During the evaluation process, the system converts each candidate solution into a series of discrete actions and gradually evaluates the value of each action through the DQN model. Finally, the system calculates the cumulative return of the entire solution and forms a list of solution scores. This scoring list not only reflects the overall performance of the solution, but also includes an evaluation of the robustness of the solution under different traffic conditions, providing a quantitative basis for subsequent decision-making.

[0082] Specifically, the process of exploring the decision space using the Monte Carlo Tree Search (MCTS) algorithm based on the scheme scoring list combines the planning ability of tree search with the efficiency of random sampling. The MCTS algorithm initializes the current traffic state at the root node, and each child node represents a possible control decision. The algorithm iterates through four main steps: selection, expansion, simulation, and backpropagation. The selection phase uses the UCT (Upper Confidence Bound for Trees) strategy to balance exploration and exploitation. The expansion phase generates new child nodes based on the predefined action space. The simulation phase uses a fast traffic flow model or heuristic strategy to quickly evaluate the return from the new node to the terminal state. The backpropagation phase updates the number of visits and cumulative returns of the node. In order to improve the search efficiency, the system implements asynchronous MCTS, allowing multiple simulations to be executed in parallel. At the same time, the system integrates domain knowledge and uses traffic expert rules to prune the search tree to reduce invalid searches. The number of iterations of MCTS is usually set to 10,000-100,000, depending on the complexity of the problem and the available computing resources. After the search is complete, the system selects the most visited child of the root node as the optimal first action, and recursively selects subsequent actions to form a complete optimized action plan. This plan represents the sequence of control strategies that performed best in a large number of simulations, taking into account long-term rewards and adapting to short-term traffic fluctuations.

[0083] Specifically, converting the optimized action plan into specific traffic control parameters is a key step in achieving the global traffic optimization plan. The system first maps the abstract action sequence output by MCTS to actual traffic control variables, such as signal light phase and period, ramp control rate, variable speed limit, etc. This mapping process uses a predefined action-parameter comparison table to ensure that each abstract action can be converted into an accurate control instruction. For continuous-valued parameters, the system uses interpolation techniques to generate smooth parameter curves between discrete actions. Considering the physical constraints of the actual traffic system, the system also applies constraint satisfaction techniques, such as penalty function method, to ensure that the generated parameters meet hard constraints (such as minimum green light time). In order to adapt to different traffic equipment and control systems, the system implements converters for multiple standard protocols, such as NTCIP (National Traffic Control Equipment Communication Protocol). In addition, the system also generates a series of alternative plans to deal with obstacles that may be encountered during the implementation of the main plan. The final global traffic optimization plan includes not only specific control parameter values, but also a schedule for parameter adjustment, execution priority, and contingency plan, forming a comprehensive traffic management strategy package.

[0084] Specifically, the process of task decomposition and local strategy generation based on the global traffic optimization scheme involves distributed control theory and edge computing technology. The system first uses the spectral clustering algorithm to segment the traffic network and decompose the global optimization problem into multiple relatively independent sub-problems. The segmentation process takes into account the road topology, traffic flow characteristics and computing resource distribution to ensure that each sub-problem has sufficient independence and maintains the necessary coordination. For each sub-problem, the system generates a regional control task, which contains local objective functions, constraints and available control variables. These tasks are distributed to the corresponding edge computing nodes through secure communication channels. On the edge nodes, the system deploys an algorithm based on model predictive control (MPC) to generate local control strategies. The MPC algorithm uses the local traffic flow model as a prediction model and considers the evolution of traffic states in the next few time steps. The objective function usually includes minimizing local delays, balancing road network loads, etc. In order to deal with model uncertainty and external interference, the system implements robust MPC variants, such as tunnel MPC and stochastic MPC. The calculation process of the control strategy uses real-time optimization techniques, such as fast gradient method or interior point method, to ensure that control decisions can be generated in time under limited computing resources. The resulting local traffic control strategy is a set of time-varying control parameter sequences that can be dynamically adjusted according to real-time traffic conditions while remaining consistent with the global optimization objective.

[0085] Furthermore, based on the global traffic optimization scheme, a graph partitioning algorithm is used to decompose the preset global optimization problem into multiple sub-problems to obtain regional control tasks, and a model predictive control algorithm is applied to the regional control tasks at the edge computing nodes to generate a control strategy that adapts to local traffic characteristics and obtains a local traffic control strategy, including: performing graph structuring on the global traffic optimization scheme, representing the traffic network as a weighted graph, and obtaining a traffic network graph model; applying a spectral clustering algorithm to the traffic network graph model, performing graph partitioning based on node connectivity and edge weights, and obtaining multiple sub-graph structures; for each sub-graph structure, extracting local traffic characteristics and constraints, constructing a sub-problem description, and obtaining a regional control task; assigning the regional control tasks to corresponding edge computing nodes, and constructing a local state space model on each node to obtain a dynamic system representation; applying a model predictive control algorithm to the dynamic system representation, calculating the control input sequence through rolling time domain optimization, and obtaining a local traffic control strategy.

[0086] Specifically, the process of graph-structuring the global traffic optimization scheme first involves abstracting the complex traffic network into a mathematical representation. The system adopts a weighted graph model in graph theory, in which nodes represent key traffic elements (such as intersections and road sections) and edges represent the connection relationship between them. Node attributes include geographic coordinates, traffic flow, signal control parameters, etc., while the weight of the edge reflects factors such as road section length, traffic capacity, and current congestion level. The graph construction process starts with digital map data and uses spatial databases (such as PostGIS) for geographic information processing. The system uses the Dijkstra algorithm to calculate the shortest path between nodes and determine the topological structure of the graph. In order to reflect dynamic traffic conditions, the system updates the weight of the edge in real time and uses a sliding time window to calculate the average travel time as the weight basis. For large-scale networks, the system adopts a hierarchical graph structure, dividing roads into two layers: main roads and secondary roads according to functional levels, simplifying the computational complexity. The graph storage adopts an adjacency list structure, which takes into account both space efficiency and access speed. The final traffic network graph model is a dynamically updated data structure that not only reflects the static road network topology but also contains real-time traffic status information, providing a basis for subsequent graph segmentation and optimization calculations.

[0087] Specifically, the process of applying the spectral clustering algorithm to the traffic network graph model for graph segmentation utilizes the Laplacian matrix characteristics of the graph to discover the natural grouping of the network. The system first constructs the adjacency matrix and degree matrix of the graph, and then calculates the normalized Laplacian matrix. The Laplacian matrix is ​​subjected to eigenvalue decomposition, and the eigenvectors corresponding to the k smallest non-zero eigenvalues ​​are selected (the k value is determined by indicators such as the silhouette coefficient). These eigenvectors form a low-dimensional representation space, in which the system applies the K-means algorithm for clustering. In order to handle large-scale networks, the system uses the Lanczos algorithm to calculate the eigenvalues ​​of sparse matrices, which significantly improves efficiency. The clustering process takes into account the connectivity of nodes and the weight of edges to ensure that the subgraphs after segmentation are closely connected internally and the connections between subgraphs are weak. In order to balance the size of subgraphs, the system introduces size constraints and uses penalty functions to adjust the clustering targets. For important roads that span multiple subgraphs, the system implements an overlapping segmentation strategy, allowing these key nodes to belong to multiple subgraphs at the same time. The segmentation results are evaluated by the modularity indicator to ensure the quality of segmentation. The resulting multiple subgraph structures represent relatively independent traffic subnetworks. Each subgraph contains a set of interrelated nodes and edges, laying the foundation for subsequent distributed optimization.

[0088] Specifically, the process of extracting local traffic features and constraints for each subgraph structure involves an in-depth analysis of the segmented network. The system first performs a topological analysis on each subgraph to identify key nodes (such as major intersections) and bottleneck sections. The importance of nodes is quantified using centrality indicators (such as betweenness centrality and closeness centrality). Then, the system extracts traffic flow features of the subgraph, including average flow, speed distribution, OD (origin-destination) demand matrix, etc. These features capture the patterns and trends of historical data through time series analysis methods (such as ARIMA models). The construction of constraints takes into account physical limitations (such as road capacity, minimum green light time) and management requirements (such as emergency vehicles have priority). The system uses fuzzy logic to handle some fuzzy constraints, such as "maintaining reasonable waiting time". For traffic flows across subgraph boundaries, the system defines boundary conditions to ensure coordination and consistency between subproblems. The objective functions of subproblems usually include minimizing total travel time, maximizing throughput, etc., and multi-objective optimization methods are used to deal with potential conflicts. The system also constructs a simplified traffic flow model, such as a macro basic graph model, for rapid evaluation of the impact of control strategies. The final regional control task is a structured optimization problem description, which includes the objective function, decision variables, constraints and simplified models, and provides a clear problem definition for the local optimization of edge nodes.

[0089] Specifically, the process of assigning regional control tasks to corresponding edge computing nodes and constructing local state space models involves resource scheduling and dynamic modeling of distributed systems. Task allocation uses a heuristic algorithm that takes into account the computing power, storage capacity, and network connection quality of edge nodes. The system implements a dynamic load balancing mechanism to adjust task allocation according to the real-time load of the nodes. To ensure data security, the task transmission process uses end-to-end encryption and uses digital signatures to verify integrity. On the edge node, the system first parses the received regional control tasks into a locally executable format. Then, based on the extracted local traffic characteristics, a discrete time state space model is constructed. State variables include the number of vehicles on the road section, average speed, etc., and control inputs include signal timing, ramp control rate, etc. The system uses a Kalman filter to estimate state variables and handle the noise and uncertainty of sensor data. The state transition equation of the model is based on simplified traffic flow theory, such as the cellular transmission model (CTM), which captures the dynamic characteristics of traffic flow. The output equation corresponds to the actual observable traffic parameters. In order to adapt to nonlinear characteristics, the system uses extended Kalman filtering or unscented Kalman filtering when necessary. The model parameters are dynamically updated through online parameter estimation methods, such as recursive least squares, to ensure that the model can adapt to changing traffic conditions. The resulting dynamic system representation is a state-space model that is updated in real time.

[0090] Specifically, the process of applying the model predictive control algorithm to the dynamic system representation uses the idea of ​​rolling horizon optimization to generate control strategies. The system first defines the prediction horizon and the control horizon, which are usually 15-30 minutes for the prediction horizon and 5-10 minutes for the control horizon. At the beginning of each control cycle, the system uses the current state estimate and the dynamic model to predict the traffic state for several time steps in the future. The prediction process takes into account known external factors, such as scheduled large-scale events or weather forecasts. Based on the prediction results, the system constructs a finite horizon optimization problem, and the objective function usually includes indicators such as minimizing overall delay and maximizing throughput. Constraints include physical limitations (such as maximum green light time) and coordination requirements (such as green waves at adjacent intersections). The optimization problem is solved by quadratic programming or nonlinear programming methods, such as interior point method or sequential quadratic programming. To improve real-time performance, the system implements fast MPC algorithms, such as explicit MPC, to pre-calculate the control law lookup table. Considering model uncertainty, the system adopts robust MPC strategies, such as tube-based MPC, to ensure stability even in the worst case. In the optimized control sequence, only the control action of the first time step is executed, and then the system moves to the next time step, updates the state estimate, and repeats the whole process. This rolling optimization strategy can adapt to rapidly changing traffic conditions and adjust control decisions in time. The final local traffic control strategy is a series of time-varying control parameters, such as signal timing schemes, ramp control rates, etc.

[0091] 104. The local traffic control strategy is sent to the corresponding traffic facilities for execution, and the global traffic optimization plan is distributed to each edge computing node for collaborative control.

[0092] In one embodiment of the present invention, the local traffic control strategy is sent to the corresponding traffic facilities for execution, and the global traffic optimization plan is distributed to each edge computing node for collaborative control, including: decomposing the local traffic control strategy, generating specific control instruction sequences for different types of traffic facilities, and obtaining a device-level control instruction set; using the local communication interface of the edge node, sending the device-level control instruction set to the corresponding traffic facility execution module to obtain execution confirmation information; performing distributed encoding on the global traffic optimization plan, converting it into a compressed data format suitable for edge node processing, and obtaining an optimization plan data packet; through a preset edge computing network topology structure, the optimization plan data packet is distributed to each edge computing node using a multi-hop routing method to obtain node optimization data; on each edge computing node, the received node optimization data is decoded and localized, and the global optimization goal is converted into a local control constraint to obtain collaborative control parameters.

[0093] Specifically, the process of decomposing instructions for local traffic control strategies involves converting high-level control decisions into specific device operation instructions. The system first identifies the various types of traffic facilities involved in the control strategy, such as traffic lights, variable message signs, ramp controllers, etc. For each type of facility, the system maintains a command template library containing all the operation commands and parameter ranges supported by this type of equipment. The instruction decomposition process uses a decision tree algorithm to map each decision point in the control strategy to the corresponding device instruction sequence. For example, for traffic light control, the system converts the phase timing scheme into a series of switching instructions, each of which contains a phase ID, duration, and switching time point. For variable speed limit signs, the system generates speed value update instructions and their effective time. The system also considers the physical limitations and response time of the equipment and inserts appropriate delays and buffers in the instruction sequence. To ensure the executability of the instructions, the system implements an instruction verification module to check whether the generated instructions meet the equipment specifications and safety standards. The final device-level control instruction set is a set of structured command sequences, each of which contains precise operation instructions and execution schedules for a specific type of traffic facility, providing standardized input for the next step of instruction issuance.

[0094] Specifically, the process of using the local communication interface of the edge node to send the device-level control instruction set to the corresponding traffic facility execution module involves complex communication protocols and device management. The system first determines the communication mode of each traffic facility based on the device registration information. For each communication mode, the system implements the corresponding protocol stack, such as NTCIP (National Traffic Control Equipment Communication Protocol) or a custom lightweight protocol. The instruction issuance process adopts an asynchronous communication model and uses message queues to manage a large number of concurrent instruction transmissions. The system implements a reliable transmission mechanism, including message confirmation, timeout retransmission, and error recovery. To ensure communication security, all instructions are encrypted before transmission and digital signatures are attached to verify integrity and source. The system also implements a real-time monitoring module to track the sending status and execution of each instruction. After receiving the execution confirmation information of the device, the system updates the instruction status and records the results to the log system. If an execution failure or timeout occurs, the system triggers a predefined fault handling process, such as trying an alternative communication channel or starting an emergency control plan. The final execution confirmation information includes a detailed record of the execution status, completion time and any abnormal conditions of each instruction, providing key data for the system to evaluate control effects and perform fault diagnosis.

[0095] Specifically, the process of distributed encoding of the global traffic optimization scheme aims to convert complex optimization results into an efficient and compact data format to adapt to the resource constraints of the edge computing environment. The system first performs a structured analysis of the optimization scheme to identify key parameters and control strategies. Then, a hierarchical encoding strategy is adopted to divide the optimization scheme into three levels: global parameters, regional strategies, and local control. Fixed-length encoding is used for global parameters to ensure fast parsing; variable-length encoding is used for regional strategies to dynamically adjust the data length according to complexity; and incremental encoding is used for local control to only transmit the change relative to the current state. The system implements an adaptive compression algorithm to select the most suitable compression method, such as Huffman coding or arithmetic coding, according to the characteristics of the data. To process time series data, the system uses differential encoding technology to transmit only the difference between adjacent time steps. Considering the heterogeneity of edge nodes, the system generates multiple versions of encoded data to adapt to nodes with different computing capabilities and storage capacities. The encoding process also includes the generation of metadata, such as version numbers, timestamps, and checksums to ensure data consistency and integrity. The final optimization solution data packet is a highly compressed and structured binary data stream that contains complete optimization solution information. It also greatly reduces data transmission and storage requirements, providing efficiency guarantees for subsequent distribution and processing.

[0096] Specifically, the process of distributing optimization scheme packets using multi-hop routing through a preset edge computing network topology involves complex network planning and routing algorithms. The system first builds a logical network topology that reflects the connection relationship and communication quality between edge nodes. This topology is represented by a graph structure, where nodes represent edge computing devices, edges represent available communication links, and edge weights reflect link quality (such as bandwidth, latency). The system implements a dynamic routing protocol similar to OSPF (Open Shortest Path First), but optimized for edge computing environments. The routing algorithm uses the Dijkstra algorithm to calculate the optimal path, taking into account factors such as link quality, node load, and energy consumption. To improve reliability, the system adopts a multi-path routing strategy to spread packets across multiple paths for transmission. Considering the dynamic nature of the edge network, the system implements an adaptive routing mechanism that can respond to network topology changes and link status fluctuations in real time. Packet transmission uses fragmentation and reassembly technology to split large optimization scheme packets into small pieces suitable for network transmission. Each fragment contains a sequence number and checksum information to ensure that it can be correctly reassembled at the receiving end. The system also implements congestion control and traffic shaping mechanisms to avoid network overload. In terms of security, the system uses end-to-end encryption to protect the data packet content and implements a lightweight intrusion detection system to monitor abnormal network behavior. Ultimately, the node optimization data received by each edge computing node is a complete, secure and node-optimized data packet, providing the necessary input for subsequent local processing.

[0097] Specifically, the process of decoding and localizing the received node optimization data on each edge computing node is a key step in converting the global optimization goal into specific local control constraints. First, the system uses a decoding algorithm that matches the encoding process to restore the compressed binary data to a structured optimization solution. The decoding process includes data integrity verification, version checking, and format conversion. Then, the system extracts the parts related to the current node, including local control parameters and coordination information of adjacent areas. For global parameters, the system performs local processing and decomposes the global objective function into local objective functions and constraints. This process uses Lagrangian relaxation to convert global constraints into local penalty terms. The system also implements an adaptive mapping mechanism to adjust control parameters based on local traffic characteristics and resource conditions. To deal with potential conflicts between local and global goals, the system uses multi-objective optimization techniques, such as Pareto front analysis, to find the balance point. Considering the real-time requirements, the system implements an incremental update mechanism that only processes the parts that are significantly different from the current state. To ensure consistency among multiple edge nodes, the system uses a consensus algorithm, such as a simplified version of Raft or Paxos, to coordinate the update of key parameters. The resulting collaborative control parameters are a set of control instructions and constraints that adapt to local conditions and are consistent with global goals, providing precise guidance for edge nodes to perform intelligent traffic control.

[0098] In this embodiment, multi-source traffic data is collected through edge computing nodes, and the data is preprocessed to obtain preprocessed data. The preprocessed data is subjected to spatiotemporal alignment and data association processing to generate an associated data set, and feature extraction and preliminary analysis are completed at the edge computing node to obtain local analysis results. The local analysis results are transmitted to the cloud server for deep data mining and pattern recognition to generate global analysis results. Based on the global analysis results, a local traffic control strategy is generated at the edge computing node, and a global traffic optimization plan is generated at the cloud server. The local traffic control strategy is sent to the traffic facilities for execution, and the global optimization plan is distributed to each edge computing node to achieve collaborative control. The present invention combines the advantages of edge computing and cloud processing to effectively improve the real-time nature of traffic data processing and the overall optimization capability of the intelligent transportation system.

[0099] The above describes the smart traffic data processing method based on edge computing in the embodiment of the present invention. The following describes the smart traffic data processing system based on edge computing in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an intelligent traffic data processing system based on edge computing includes:

[0100] The preprocessing module 201 is used to collect multi-source traffic data through a preset edge computing node, and perform data preprocessing on the multi-source traffic data to obtain preprocessed data;

[0101] A local analysis module 202 is used to perform spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and perform feature extraction and preliminary analysis on the associated data set at an edge computing node to obtain a local analysis result;

[0102] The cloud processing module 203 is used to transmit the local analysis results to the cloud server, perform deep data mining and pattern recognition, obtain global analysis results, and generate local traffic control strategies at the edge computing node and global traffic optimization solutions at the cloud server based on the global analysis results;

[0103] The collaborative control module 204 is used to send the local traffic control strategy to the corresponding traffic facilities for execution, and distribute the global traffic optimization plan to each edge computing node for collaborative control.

[0104] In an embodiment of the present invention, the edge computing-based smart traffic data processing system runs the above-mentioned edge computing-based smart traffic data processing method, and the edge computing-based smart traffic data processing system collects multi-source traffic data through edge computing nodes, pre-processes the data, and obtains pre-processed data. The pre-processed data is subjected to spatiotemporal alignment and data association processing to generate an associated data set, and feature extraction and preliminary analysis are completed at the edge computing node to obtain local analysis results. The local analysis results are transmitted to the cloud server for deep data mining and pattern recognition to generate global analysis results. Based on the global analysis results, a local traffic control strategy is generated at the edge computing node, and a global traffic optimization solution is generated at the cloud server. The local traffic control strategy is sent down to the traffic facilities for execution, and the global optimization solution is distributed to each edge computing node to achieve collaborative control. The present invention combines the advantages of edge computing and cloud processing to effectively improve the real-time performance of traffic data processing and the overall optimization capability of the smart traffic system.

[0105] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device, or unit can refer to the corresponding process in the aforementioned method embodiment and will not be repeated here.

[0106] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0107] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart traffic data processing method based on edge computing, characterized in that: The smart traffic data processing method based on edge computing includes: Collect multi-source traffic data through a preset edge computing node, and perform data preprocessing on the multi-source traffic data to obtain preprocessed data; Performing spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and performing feature extraction and preliminary analysis on the associated data set at an edge computing node to obtain a local analysis result; The local analysis results are transmitted to a cloud server for deep data mining and pattern recognition to obtain global analysis results, and based on the global analysis results, a local traffic control strategy is generated at the edge computing node and a global traffic optimization solution is generated at the cloud server; the local analysis results are transmitted to a cloud server for deep data mining and pattern recognition to obtain global analysis results, including: performing spatiotemporal alignment and data fusion on the local analysis results from multiple edge nodes to obtain a global data set, and using a multi-scale wavelet transform to perform time-frequency domain decomposition on the global data set to obtain a multi-level feature representation; performing unsupervised learning on the multi-level feature representation through a preset deep belief network to extract high-order abstract features to obtain a deep feature vector; applying a pre-trained dynamic Bayesian network to the deep feature vector to perform traffic state inference to obtain a current traffic state estimate; inputting the current traffic state estimate into a long short-term memory network model to obtain a future traffic state prediction result, and combining the future traffic state prediction result, the current traffic state estimate and historical traffic data to perform a multi-dimensional comparative analysis to obtain a global analysis result; The generating of a local traffic control strategy at an edge computing node and a global traffic optimization plan at a cloud server based on the global analysis result includes: applying a multi-objective genetic algorithm to the global analysis result to generate a set of candidate global traffic optimization plans, and selecting the best plan from the set of candidate global traffic optimization plans based on predefined evaluation indicators to obtain an initial optimization plan; using traffic flow theory to perform a macroscopic evaluation on the initial optimization plan, screening out a subset of feasible plans, obtaining an optimized candidate plan, and inputting the optimized candidate plan into a pre-trained reinforcement learning model, evaluating each candidate plan, and obtaining a plan scoring list; based on the plan scoring list, using a Monte Carlo tree search algorithm to explore the decision space, select the best action sequence, obtain an optimized action plan, and convert the optimized action plan into specific traffic control parameters to obtain a global traffic optimization plan; based on the global traffic optimization plan, using a graph segmentation algorithm to decompose the preset global optimization problem into multiple sub-problems to obtain a regional control task, and applying a model predictive control algorithm to the regional control task at the edge computing node to generate a control strategy adapted to local traffic characteristics, and obtain a local traffic control strategy; The local traffic control strategy is sent to the corresponding traffic facilities for execution, and the global traffic optimization plan is distributed to each edge computing node for collaborative control.

2. The method for processing intelligent traffic data based on edge computing according to claim 1 is characterized in that: The performing of spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and performing feature extraction and preliminary analysis on the associated data set at the edge computing node to obtain a local analysis result includes: Identifying a timestamp of the preprocessed data, and aligning the preprocessed data based on the timestamp to obtain time-synchronized data; Based on a predefined spatial grid model, spatially mapping and matching the time synchronization data to obtain spatiotemporal alignment data; Utilizing a local knowledge base to contextualize the spatiotemporal alignment data, establish logical connections between data items, and obtain a correlation data set; A lightweight machine learning algorithm is applied to the associated data set to extract key features to obtain feature data, and a lightweight traffic state estimation model deployed on the edge node is used to process the feature data and generate preliminary analysis results.

3. The method for processing intelligent traffic data based on edge computing according to claim 2 is characterized in that: The using of the local knowledge base to contextually associate the spatiotemporal alignment data and establish logical connections between data items to obtain an associated data set includes: Performing spatial mapping on the spatiotemporal alignment data according to a predefined local road network topology structure to obtain topological association data; Extracting traffic elements from the topological association data, and applying a semantic annotation algorithm based on a traffic concept dictionary to the traffic elements, giving the traffic elements semantic attributes, and obtaining semantic annotation data; Using graph embedding technology to vectorize the traffic elements in the semantically labeled data to obtain a multi-dimensional feature vector; Calculating a similarity matrix between traffic elements based on the multidimensional feature vector, and grouping similar traffic elements using a hierarchical clustering algorithm to obtain an element clustering result; Association rules are constructed according to the element clustering results, and the association rules are applied to the topological association data to obtain an association data set.

4. The method for processing intelligent traffic data based on edge computing according to claim 1 is characterized in that: Based on the global traffic optimization scheme, the preset global optimization problem is decomposed into multiple sub-problems using a graph segmentation algorithm to obtain a regional control task, and a model predictive control algorithm is applied to the regional control task at the edge computing node to generate a control strategy adapted to local traffic characteristics. The local traffic control strategy includes: Performing graph-structured processing on the global traffic optimization scheme, representing the traffic network as a weighted graph, and obtaining a traffic network graph model; Applying a spectral clustering algorithm to the traffic network graph model, segmenting the graph based on node connectivity and edge weights, and obtaining multiple subgraph structures; For each subgraph structure, extract local traffic characteristics and constraints, construct sub-problem descriptions, and obtain regional control tasks; Assigning the regional control tasks to corresponding edge computing nodes, and constructing a local state space model on each node to obtain a dynamic system representation; A model predictive control algorithm is applied to the dynamic system representation, and a control input sequence is calculated through a rolling horizon optimization to obtain a local traffic control strategy.

5. The method for processing intelligent traffic data based on edge computing according to claim 1 is characterized in that: The sending of the local traffic control strategy to corresponding traffic facilities for execution, and the distribution of the global traffic optimization solution to each edge computing node for collaborative control includes: Decomposing the local traffic control strategy into instructions, generating specific control instruction sequences for different types of traffic facilities, and obtaining a device-level control instruction set; Using the local communication interface of the edge node, the device-level control instruction set is sent to the corresponding traffic facility execution module to obtain execution confirmation information; Distributed encoding is performed on the global traffic optimization plan, and the plan is converted into a compressed data format suitable for edge node processing to obtain an optimization plan data packet; Through the preset edge computing network topology, the optimization solution data packet is distributed to each edge computing node using a multi-hop routing method to obtain node optimization data; On each edge computing node, the received node optimization data is decoded and localized, and the global optimization objective is converted into local control constraints to obtain collaborative control parameters.

6. A smart traffic data processing system based on edge computing, characterized in that: The edge computing-based intelligent traffic data processing system includes: A preprocessing module, used to collect multi-source traffic data through a preset edge computing node, and perform data preprocessing on the multi-source traffic data to obtain preprocessed data; A local analysis module is used to perform spatiotemporal alignment and data association on the preprocessed data to obtain an associated data set, and perform feature extraction and preliminary analysis on the associated data set at an edge computing node to obtain a local analysis result; A cloud processing module is used to transmit the local analysis results to a cloud server, perform deep data mining and pattern recognition, obtain global analysis results, and generate local traffic control strategies at edge computing nodes and global traffic optimization solutions at cloud servers based on the global analysis results; the local analysis results are transmitted to a cloud server, perform deep data mining and pattern recognition, and obtain global analysis results, including: performing spatiotemporal alignment and data fusion on local analysis results from multiple edge nodes to obtain a global data set, and using a multi-scale wavelet transform to perform time-frequency domain decomposition on the global data set to obtain a multi-level feature representation; performing unsupervised learning on the multi-level feature representation through a preset deep belief network to extract high-order abstract features to obtain a deep feature vector; applying a pre-trained dynamic Bayesian network to the deep feature vector to perform traffic state inference to obtain a current traffic state estimate; inputting the current traffic state estimate into a long short-term memory network model to obtain a future traffic state prediction result, and combining the future traffic state prediction result, the current traffic state estimate and historical traffic data to perform a multi-dimensional comparative analysis to obtain a global analysis result; The generating of a local traffic control strategy at an edge computing node and a global traffic optimization plan at a cloud server based on the global analysis result includes: applying a multi-objective genetic algorithm to the global analysis result to generate a set of candidate global traffic optimization plans, and selecting the best plan from the set of candidate global traffic optimization plans based on predefined evaluation indicators to obtain an initial optimization plan; using traffic flow theory to perform a macroscopic evaluation on the initial optimization plan, screening out a subset of feasible plans, obtaining an optimized candidate plan, and inputting the optimized candidate plan into a pre-trained reinforcement learning model, evaluating each candidate plan, and obtaining a plan scoring list; based on the plan scoring list, using a Monte Carlo tree search algorithm to explore the decision space, select the best action sequence, obtain an optimized action plan, and convert the optimized action plan into specific traffic control parameters to obtain a global traffic optimization plan; based on the global traffic optimization plan, using a graph segmentation algorithm to decompose the preset global optimization problem into multiple sub-problems to obtain a regional control task, and applying a model predictive control algorithm to the regional control task at the edge computing node to generate a control strategy adapted to local traffic characteristics, and obtain a local traffic control strategy; The collaborative control module is used to send the local traffic control strategy to the corresponding traffic facilities for execution, and distribute the global traffic optimization plan to each edge computing node for collaborative control.

Images