Intelligent road driving time prediction method based on multi-source data fusion and deep learning

Abstract

The invention discloses a smart road driving time prediction method based on multi-source data fusion and deep learning, and the method comprises the steps: constructing a multi-source data collection and fusion platform, developing a spatial-temporal feature deep learning model, deploying a real-time data processing and anomaly detection system, building a self-adaptive prediction optimization mechanism, and constructing a smart traffic application service platform. Comprehensive acquisition, deep analysis, real-time processing, accurate prediction and intelligent service of traffic data are realized. The method specifically comprises the following steps: deploying an Internet of Things sensor network to collect multi-source data, and designing an adaptive weighted fusion algorithm to construct a multi-dimensional dynamic traffic database; constructing an LSTM-GNN hybrid neural network architecture to extract spatio-temporal features; real-time data processing and anomaly detection are carried out by adopting a streaming computing framework and various technologies; designing a multi-objective optimization algorithm and a reinforcement learning model to realize adaptive prediction optimization; and developing a visual decision support system and a personalized navigation service. The system effectively improves the traffic management efficiency and the public travel experience.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation, and in particular to a method for predicting travel time on intelligent roads based on multi-source data fusion and deep learning. Background Art

[0002] In today's society, the rapid development of urban transportation has brought about a series of complex problems.

[0003] In traditional traffic management, data collection methods are relatively simple, usually relying on limited fixed sensors such as loop coils, which makes it difficult to fully and accurately obtain information on traffic flow, vehicle speed, weather conditions, road events, etc. This leads to incompleteness and lag in traffic data, and it is unable to reflect the real traffic situation in a timely manner.

[0004] At the same time, there is a lack of effective fusion processing methods for the collected heterogeneous data. Different types of data often operate independently, and it is impossible to effectively align and integrate them in time and space, making it difficult to build a comprehensive and accurate traffic status model.

[0005] In terms of traffic status analysis and prediction, traditional methods have the problems of low accuracy and poor adaptability. They cannot accurately capture the time series characteristics of traffic flow and the spatial correlation of road networks, and are difficult to cope with complex and changing traffic conditions.

[0006] In addition, the existing traffic management system also has deficiencies in real-time data processing and anomaly detection capabilities. The identification of abnormal events such as sudden accidents and extreme weather is not timely and accurate enough, and effective response measures cannot be taken quickly, which leads to further aggravation of traffic congestion.

[0007] In terms of traffic prediction and optimization, traditional methods are usually based on simple statistical models or rules, which cannot comprehensively consider the impact of multiple factors, making it difficult to achieve accurate predictions and effective optimization, and it is difficult to meet the needs of modern urban traffic management. Summary of the invention

[0008] To solve the above problems, the present invention proposes a smart road travel time prediction method based on multi-source data fusion and deep learning. The specific steps are as follows, which are characterized by:

[0009] Step 1: Build a multi-source data collection and fusion platform, deploy an IoT sensor network, integrate multi-source data collection equipment such as traffic flow detectors, road state sensing equipment, meteorological monitoring stations, and vehicle-mounted GPS terminals; perform real-time data preprocessing through edge computing nodes, and use an adaptive weighted fusion algorithm to align heterogeneous data such as traffic flow, real-time vehicle speed, weather conditions, and road events in time and space, and establish a multi-dimensional dynamic traffic database covering the entire road network;

[0010] Step 2: Develop a deep learning model for spatiotemporal features and construct an LSTM-GNN hybrid neural network architecture, where the LSTM module is responsible for capturing the time series characteristics of traffic flow, and the GNN module extracts spatial correlation features based on the road network topology structure, and implements dynamic weight allocation of features through the attention mechanism;

[0011] Step 3: Deploy a real-time data processing and anomaly detection system, establish a streaming computing framework in the cloud, use empirical mode decomposition technology to perform multi-scale decomposition of real-time traffic data, and use the isolation forest algorithm to identify abnormal events such as sudden accidents and extreme weather; develop a dynamic correction module to automatically trigger the online update mechanism of model parameters when an abnormal event is detected to ensure the system's anti-interference ability;

[0012] Step 4: Establish an adaptive prediction optimization mechanism, design a multi-objective optimization algorithm, comprehensively consider the influence of multiple factors such as historical average speed, real-time floating vehicle data, and signal light timing, and build a dynamic weight adjustment model based on reinforcement learning; continuously update the road state feature vector through the online learning module to predict the travel time of the entire road network;

[0013] Step 5: Build a smart transportation application service platform that integrates real-time prediction API interfaces and digital twin road network models to provide traffic management departments with intelligent decision-making tools such as congestion warning, signal optimization, and emergency dispatch. Develop personalized navigation services for public travel, provide dynamic route planning based on vehicle characteristics and driving preferences, and push the optimal route and ETA prediction in real time through mobile APP.

[0014] The present invention is based on a multi-source data fusion and deep learning intelligent road travel time prediction method, which has beneficial effects. The technical effects of the present invention are:

[0015] 1. By constructing a multi-source data collection and fusion platform, the present invention can obtain multi-dimensional traffic data in a comprehensive, accurate and real-time manner, providing a rich and reliable data basis for subsequent analysis and decision-making, and greatly improving the integrity and timeliness of traffic data.

[0016] 2. The spatiotemporal feature deep learning model developed by the present invention can accurately capture the time series characteristics of traffic flow and the spatial correlation of the road network, thereby significantly improving the accuracy of traffic status analysis and prediction, and providing a more valuable reference for traffic management and planning.

[0017] 3. The real-time data processing and anomaly detection system deployed by the present invention can timely and accurately identify abnormal events such as sudden accidents and extreme weather, quickly trigger response measures, effectively reduce the impact of abnormal events on traffic, and enhance the anti-interference ability of the traffic system.

[0018] 4. The adaptive prediction optimization mechanism established by the present invention comprehensively considers the influence of multiple factors such as historical average speed, real-time floating vehicle data, and signal light timing, realizes the accuracy of traffic prediction and the effectiveness of optimization, improves the utilization efficiency of traffic resources, and alleviates traffic congestion. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the present invention;

[0020] Figure 2 A flow chart of the deep learning model for developing spatiotemporal features of the present invention;

[0021] Figure 3 The flowchart of the deployment real-time data processing and anomaly detection system of the present invention. DETAILED DESCRIPTION

[0022] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0023] The present invention proposes a method for predicting travel time on intelligent roads based on multi-source data fusion and deep learning. The invention flowchart is as follows: Figure 1 As shown, the steps of the present invention are described in detail below.

[0024] Step 1: Build a multi-source data collection and fusion platform

[0025] Deploy an IoT sensor network, integrate multi-source data collection equipment such as traffic flow detectors, road status perception equipment, meteorological monitoring stations, and vehicle-mounted GPS terminals; perform real-time data preprocessing through edge computing nodes, and use an adaptive weighted fusion algorithm to align heterogeneous data such as traffic flow, real-time vehicle speed, weather conditions, road events, etc. in time and space to establish a multi-dimensional dynamic traffic database covering the entire road network.

[0026] Step 1.1 IoT sensor network deployment

[0027] Traffic flow detectors use circular coils, microwave radars and video recognition composite sensors, which are deployed on main roads / expressways at intervals of 200-500 meters. High-precision magnetoresistive sensors are used at intersections to support bidirectional 12-lane coverage. Fiber Bragg grating strain sensors and infrared thermal imaging cameras are installed as road state sensing devices to detect road cracks, water accumulation and ice conditions. The deployment density is 2 groups per kilometer, which are integrated into signal light poles. Micro-meteorological stations are deployed at key nodes of the road network, integrating temperature, humidity, rainfall, and visibility sensors. The spatial coverage radius is ≤3 kilometers, and the data sampling frequency is 1Hz. OBD-II interface GPS devices are installed on taxis / buses with a sampling frequency of 1Hz, and longitude and latitude, speed, and heading angle data are uploaded in real time. The coverage density is ≥5% of the vehicle ownership. Flow, speed, and meteorological data are obtained through sensors and transmitted to edge nodes through the LoRaWAN protocol at a rate of 50kbps. GPS data is directly transmitted to the cloud using TCP long connections. Define a unified data message format: including device ID, timestamp, numeric field, and checksum;

[0028] Step 1.2 Real-time preprocessing of edge computing nodes

[0029] NVIDIA Jetson AGX Xavier edge servers are deployed on the roadside units with a computing power of 32TOPS. Each node covers a radius of 1 km and is connected to the cloud through 5G CPE. A sliding window mean filter is used with a window width of 60 seconds, and data outside the 3σ range is marked as invalid. Based on GPS time, linear interpolation is performed on sensor data, and the time alignment accuracy is controlled within 100ms.

[0030] Step 1.3 Adaptive weighted fusion algorithm design

[0031] Resolving device sampling frequency differences through dynamic time warping:

[0032] DTW(S,T)=min π Σ (i,j)∈π ||s i -t j || 2

[0033] Where π is the alignment path, S and T are the sequences to be aligned, and s i is the data sequence collected by the i-th sensor, t j For the data sequence collected by the jth sensor, the alignment path is calculated to achieve effective alignment of sequences with different sampling frequencies.

[0034] The inverse distance weighted method is used to perform spatial interpolation of meteorological data:

[0035]

[0036] Among them, p = 2 is the standard IDW, z(x0) is the interpolation point value, is the interpolation result at the interpolation point x0, z(x i ) is the observation point x i The value at d(x0,x i ) is the distance between the interpolation point and the observation point, n represents the number of observation points involved in the interpolation calculation, and P is the distance attenuation coefficient.

[0037] Step 1.4 Construction of multi-dimensional dynamic traffic database

[0038] InfluxDB is used to store raw sensor data, with a write speed of more than 100,000 points / second, and time-sliced ​​storage is performed with 24 hours as one slice. PostGIS is used to store road network topology, supporting KNN queries, and the response time is controlled within 50ms. Near-real-time data is stored in SSD and retained for 7 days, while historical data is transferred to object storage that complies with the S3 protocol. When the speed change rate is detected to be >15% or a new event is generated, database record insertion is immediately started.

[0039] Step 2: Develop a deep learning model for spatiotemporal features

[0040] Construct an LSTM-GNN hybrid neural network architecture, in which the LSTM module is responsible for capturing the time series characteristics of traffic flow, and the GNN module extracts spatial correlation features based on the road network topology structure. The dynamic weight allocation of features is achieved through the attention mechanism, and the flow chart of the deep learning model of spatiotemporal features is developed as shown in the figure. Figure 2 shown.

[0041] Step 2.1 LSTM-GNN hybrid neural network architecture design principle

[0042] Step 2.1.1 LSTM module realizes time series feature extraction

[0043] The module input is 30 minutes of historical data, with a time step of 5 minutes and feature dimension 16, including traffic, speed, weather level, event status, etc. The 3-layer bidirectional stacked LSTM structure is used. The mathematical expression is:

[0044] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0045] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0046]

[0047] O t =σ(W O ·[h t-1 ,x t ]+b O )

[0048] h t =O t tanh(C t )

[0049] Among them, f t For the forget gate, i t is the input gate, C t is a candidate memory unit, C t is the memory unit, C t-1 is the memory unit of the previous moment, O t is the output gate, h t is the hidden state, W f W i W C W O is the weight matrix, b f b i b C b O is the bias term, σ is the sigmoid activation function, t is the time step, [h t-1 ,x t ] means to change the hidden state h of the previous moment t-1 and the current input x t For splicing, input feature x t It includes parameters such as traffic and speed in the time window. The time series sliding window is set to 30 minutes of historical data, and the time step is t = 5 minutes. Through three layers of stacked LSTM, the hidden layer dimension is 128, and multi-granularity time patterns at the hour, half-hour, and minute levels are captured, and the time series feature vector h is finally output. t .

[0050] Step 2.1.2 GNN module builds spatial topological relationship

[0051] Define the road network graph as G = (V, E), node v i ∈V represents the road section and intersection unit, and the edge e ij ∈E represents node v i and v j The edges connecting the nodes, i and j are used to identify the nodes in the graph. i To the neighbor v j The attention weight α ij Calculated as:

[0052]

[0053] Where W is a learnable parameter, || represents vector concatenation, and N i is a node set, h i 、h j 、h k i, j and k are the feature vectors of nodes in the graph, a is a learnable attention vector, LeakyReLU is the activation function, exp is the exponential function, and T is the matrix transpose. ij For node feature h i 'Update, and obtain the weighted sum of neighbor features:

[0054]

[0055] Among them, σ is the sigmoid activation function. After passing through multiple layers of GNN, the features of each node are obtained, and the features of all nodes are stacked to obtain the feature matrix S∈R N×64 , N is the number of nodes.

[0056] Step 2.2 Attention Mechanism Implementation

[0057] Using the feature S output by GNN, construct the node correlation matrix A s :

[0058] A s =sigmoid(SWS T ),W∈R 64×64

[0059] Adaptively enhance the feature matrix:

[0060]

[0061] Among them, ⊙ is the dot product symbol, and the enhanced spatial features are obtained.

[0062] Step 3: Deploy real-time data processing and anomaly detection system

[0063] A streaming computing framework is established in the cloud, and the empirical mode decomposition technology is used to perform multi-scale decomposition of real-time traffic data. The isolation forest algorithm is used to identify abnormal events such as sudden accidents and extreme weather. A dynamic correction module is developed to automatically trigger the online update mechanism of model parameters when an abnormal event is detected to ensure the system's anti-interference ability. The real-time data processing and anomaly detection system flow chart is as follows: Figure 3 shown.

[0064] Step 3.1 Building a cloud streaming computing framework

[0065] Step 3.1.1: At the data source access layer, the pre-processed data stream of the edge computing node is connected through the Kafka message queue, and the "road section ID + time window" partitioning strategy is used to ensure the spatiotemporal continuity of the data.

[0066] Step 3.1.2 Stream processing job design

[0067] Time window definition: Using event time semantics, define a rolling window T = 1 minute, a sliding window step size = 10 seconds, and a window length = 5 minutes, which are used for real-time statistics and historical trend analysis respectively.

[0068] Complex event detection: The accident mode is defined through the Flink CEP complex event processing engine. When the speed of five consecutive vehicles drops by more than 50% and lasts for 30 seconds, the abnormal event label is triggered.

[0069] State management: Use RocksDB as the state backend to persistently store sliding window statistics and support state consistency after fault recovery.

[0070] Step 3.1.3 Elastic resource scheduling: Dynamically expand computing nodes based on Kubernetes and configure automatic expansion and reduction strategies.

[0071] Step 3.2 EMD multi-scale decomposition and feature extraction

[0072] Empirical mode decomposition (EMD) decomposes non-stationary signals into the sum of a finite number of intrinsic mode functions (IMFs):

[0073]

[0074] Among them, x(t) is the original signal, t is the time variable, and IMF k is the kth eigenmode function, r n (t) is the residual signal, and n is the number of decompositions.

[0075] Step 3.2.1 Sliding window processing: intercept the real-time vehicle speed data stream with a 60-second window, and use the mirror extension method to eliminate the endpoint effect to ensure signal continuity.

[0076] Step 3.2.2 Multi-scale feature generation, high-frequency IMF reflects sudden disturbances, and the variance threshold is set to 0.8σ 2 , σ is the standard deviation of historical normal data. When the variance exceeds the threshold, it is marked as a sudden disturbance.

[0077] Low-frequency IMF characterizes periodic congestion and extracts energy entropy features E j :

[0078]

[0079] Step 3.3 Isolation Forest Anomaly Detection

[0080] Step 3.3.1 Combine the 6-dimensional IMF energy entropy after EMD decomposition and the original vehicle speed standard deviation into a 7-dimensional feature vector, and add the enhanced spatial feature The feature vector is formed to provide multi-dimensional input features for the isolation forest anomaly detection model.

[0081] Step 3.3.2 Model training: 100 isolated trees are constructed on historical normal data, and 256 instances are sampled for each tree. The mean sample path length h(x) and the anomaly score s(x) are calculated using the formula:

[0082]

[0083] Where c(n) is the normalization factor and h(x) is the mean path length:

[0084]

[0085] Among them, H(n-1) is the harmonic number, and n represents the number of instances sampled by each isolated tree.

[0086] Step 3.3.3 Set the dynamic threshold s th , when s(x)>s th When an abnormal alarm is triggered, the isolation forest outputs the classification results of abnormal event types: accident / congestion / weather impact, and associates them with the specific road section ID and timestamp.

[0087] Step 3.4 Dynamic correction module design

[0088] When the isolation forest detects an anomaly and the duration exceeds 3 window periods, the model correction is started and the LSTM-GNN model weights are updated using the online gradient descent algorithm:

[0089]

[0090] Among them, W t , W t+1 is the weight matrix of the LSTM-GNN model at time steps t and t+1, is the loss function L with respect to weight W t The gradient of x t Model input data for time step t, y t is the true label corresponding to time step t, and the learning rate η is adaptively adjusted according to the error change rate:

[0091]

[0092] Among them, ε is a very small positive number to prevent the denominator from being 0, and is the gradient of the loss function at time steps t and t-1.

[0093] Step 4: Establish an adaptive prediction optimization mechanism

[0094] A multi-objective optimization algorithm is designed to comprehensively consider the influence of multiple factors such as historical average speed, real-time floating vehicle data, and traffic light timing, and a dynamic weight adjustment model based on reinforcement learning is constructed; the road state feature vector is continuously updated through the online learning module to predict the travel time of the entire road network.

[0095] Step 4.1 Multi-objective optimization algorithm design

[0096] Multi-objective optimization aims to optimize multiple conflicting objectives at the same time, and uses the Pareto optimal solution set to represent the optimal trade-off solution. The objective function is defined as:

[0097]

[0098] Among them, f1(w) is the prediction error (MAE), f2(w) is the model calculation delay, and f3(w) is the CPU / memory consumption.

[0099] The optimization variables of the model are set to w = [w1, w2, w3, w4], which are the weight coefficients of historical average speed, real-time floating vehicle data, traffic light timing, and enhanced spatial features, respectively.

[0100] The constraints are weight normalization w1+w2+w3+w4=1, real-time constraint calculation delay f2(w)≤200ms, and resource constraint CPU utilization f3(w)≤80%.

[0101] Solving the Pareto frontier method:

[0102] Step 4.1.1 Initialization, randomly generate 100 weight combinations w (i) .

[0103] Step 4.1.2 Non-dominated sorting: sort the solution set in layers according to the objective function value.

[0104] Step 4.1.3: Calculate the crowding degree, evaluate the distribution density of individuals in the solution set, and retain diversity.

[0105] Step 4.1.4 Selection and crossover, generate offspring through tournament selection and simulated binary crossover.

[0106] Step 4.1.5 Iterative optimization, repeat the above steps until convergence, the maximum number of iterations = 100.

[0107] Step 4.2 reinforces the learning dynamic weight adjustment model, automatically adjusts the weight coefficient based on the real-time traffic status, and realizes adaptive optimization of the prediction model.

[0108] Step 4.2.1 Construct real-time traffic state vector Among them, v avg is the historical average speed, v real is the real-time speed of the floating vehicle, Δt signal is the remaining time of the signal light, To enhance the spatial characteristics of the node average speed and congestion index

[0109] Step 4.2.2 Set the action space and weight adjustment action to a t =[Δw1,Δw2,Δw3,Δw4]

[0110] Step 4.2.3 Set the reward function, based on the prediction error reward r t :

[0111]

[0112] Among them, α=0.7, β=0.3 are weight coefficients, MAE t is the prediction error at time t, latency t Compute the delay for the model at time t.

[0113] Step 4.2.4 Model training

[0114] Step 4.2.4.1 Q-learning update rule:

[0115]

[0116] Among them, s t+1 is the traffic state vector at time t+1, a t+1 is the weight adjustment action at time t+1, Q(s t+1 ,a t+1 ) is state s t+1 Take action a t+1 The Q value, For state s t+1 The maximum Q value that can be obtained among all possible actions under the learning rate η = 0.1 and the discount factor γ = 0.9.

[0117] Step 4.2.4.2 Exploration strategy, adopt the ∈ greedy strategy (∈=0.1).

[0118] Step 4.2.4.3 Experience playback, storage of historical transfers (s t ,a t ,r t ,s t+1 ), 32 samples are sampled each time.

[0119] Step 4.3: Online learning module, which updates the road state characteristics and model parameters in real time to ensure the predictive adaptability.

[0120] Step 4.3.1 Feature vector update:

[0121] Update the road status feature vector s every 5 minutes t .

[0122] Step 4.3.2: Incremental model update, using Mini-batch gradient descent, updating the Q network parameters every 10 minutes:

[0123]

[0124] Among them, θ is the Q network parameter, η Q is the learning rate, ← is the assignment symbol, is the gradient of parameter θ, L(θ) is the loss function, and the loss function expression is:

[0125]

[0126] Among them, i is the sample index value, Q(s i ,a i ; θ) is in state s i Next, perform action a i The Q value estimated by the model with parameter θ is y i The target value is:

[0127] y i =r i +γmax a (s i+1 ,a;θ - )

[0128] Among them, r i For state s i Next, perform action a i The reward value immediately obtained by the agent, γ is the discount factor, max a To find the maximum value for all possible actions a, θ - is the target network parameter, s i+1 For state s i Next, perform action a i Then, transfer to the next state.

[0129] Step 4.4 Calculate the predicted travel time for the grid cells

[0130] Step 4.4.1 Discretize the road network and divide the urban road network into grid units of 500m×500m.

[0131] Step 4.4.2 Prediction value calculation:

[0132] For each grid cell G k , calculate the predicted travel time t k :

[0133]

[0134] Where k is the grid cell index, d' k is the total length of roads in the grid, f(Δt signal ) is the signal light timing influence function, is the average value of the enhanced spatial features within the grid.

[0135] Step 5: Build a smart transportation application service platform

[0136] The platform integrates real-time prediction API interfaces and digital twin road network models to provide traffic management departments with intelligent decision-making tools such as congestion warning, signal optimization, and emergency dispatch. It develops personalized navigation services for public travel, provides dynamic route planning based on vehicle characteristics and driving preferences, and pushes the optimal route and ETA prediction in real time through mobile APP.

[0137] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A smart road travel time prediction method based on multi-source data fusion and deep learning, the specific steps are as follows, characterized by: Step 1: Build a multi-source data collection and fusion platform, deploy an IoT sensor network, and integrate multi-source data collection equipment such as traffic flow detectors, road state sensing equipment, meteorological monitoring stations, and vehicle-mounted GPS terminals; Real-time data preprocessing is performed through edge computing nodes, and an adaptive weighted fusion algorithm is used to align heterogeneous data such as traffic flow, real-time vehicle speed, weather conditions, and road events in time and space to establish a multi-dimensional dynamic traffic database covering the entire road network. Step 2: Develop a deep learning model for spatiotemporal features and construct an LSTM-GNN hybrid neural network architecture, where the LSTM module is responsible for capturing the time series characteristics of traffic flow, and the GNN module extracts spatial correlation features based on the road network topology structure, and implements dynamic weight allocation of features through the attention mechanism; Step 3: Deploy a real-time data processing and anomaly detection system, establish a streaming computing framework in the cloud, use empirical mode decomposition technology to perform multi-scale decomposition of real-time traffic data, and use the isolation forest algorithm to identify sudden accidents and extreme weather anomalies; Develop a dynamic correction module that automatically triggers the online update mechanism of model parameters when an abnormal event is detected to ensure the system's anti-interference ability; Step 4: Establish an adaptive prediction optimization mechanism, design a multi-objective optimization algorithm, comprehensively consider the influence of historical average speed, real-time floating vehicle data, and signal light timing, and build a dynamic weight adjustment model based on reinforcement learning; continuously update the road state feature vector through the online learning module to predict the travel time of the entire road network; Step 5: Build a smart transportation application service platform that integrates real-time prediction API interfaces and digital twin road network models to provide traffic management departments with intelligent decision-making tools for congestion warning, signal optimization, and emergency dispatch; Develop personalized navigation services for public travel, provide dynamic route planning based on vehicle characteristics and driving preferences, and push the optimal route and ETA prediction in real time through mobile APP.

2. The method for predicting travel time on smart roads based on multi-source data fusion and deep learning according to claim 1 is characterized by: The construction of a multi-source data acquisition and fusion platform in step 1 can be expressed as: Step 1.1 IoT sensor network deployment Traffic flow detectors use circular coils, microwave radars and video recognition composite sensors, which are deployed on main roads / expressways at intervals of 200-500 meters. High-precision magnetoresistive sensors are used at intersections to support bidirectional 12-lane coverage. Fiber Bragg grating strain sensors and infrared thermal imaging cameras are installed as road state sensing devices to detect road cracks, water accumulation and ice conditions. The deployment density is 2 groups per kilometer, which are integrated into signal light poles. Micro-meteorological stations are deployed at key nodes of the road network, integrating temperature, humidity, rainfall, and visibility sensors, with a spatial coverage radius of ≤3 kilometers and a data sampling frequency of 1Hz. OBD-II interface GPS devices are installed in taxis / buses with a sampling frequency of 1Hz, and longitude and latitude, speed, and heading angle data are uploaded in real time. The coverage density is ≥5% of the vehicle ownership. Flow, speed, and meteorological data are obtained through sensors and transmitted to edge nodes through the LoRaWAN protocol at a rate of 50kbps. GPS data is directly transmitted to the cloud using TCP long connections. A unified data message format is defined: including device ID, timestamp, value field, and checksum. Step 1.2 Real-time preprocessing of edge computing nodes NVIDIA Jetson AGX Xavier edge servers are deployed on roadside units with a computing power of 32TOPS. Each node covers a radius of 1 km and is connected to the cloud through 5G CPE. A sliding window mean filter is used with a window width of 60 seconds, and data outside the 3σ range is marked as invalid. Based on GPS time, linear interpolation is performed on sensor data, and the time alignment accuracy is controlled within 100ms. Step 1.3 Adaptive weighted fusion algorithm design Resolving device sampling frequency differences through dynamic time warping: DTW(S,T)=min π ∑ (i,j)∈π ||s i -the j || 2 Where π is the alignment path, S and T are the sequences to be aligned, and s i is the data sequence collected by the i-th sensor, t j For the data sequence collected by the jth sensor, the alignment path is calculated to achieve effective alignment of sequences with different sampling frequencies; The inverse distance weighted method is used to perform spatial interpolation of meteorological data: Among them, p = 2 is the standard IDW, z(x0) is the interpolation point value, is the interpolation result at the interpolation point x0, z(x i ) is the observation point x i The value at d(x0,x i ) is the distance between the interpolation point and the observation point, n represents the number of observation points involved in the interpolation calculation, and P is the distance attenuation coefficient; Step 1.4 Construction of multi-dimensional dynamic traffic database InfluxDB is used to store raw sensor data, with a write speed of more than 100,000 points per second, and time-sliced ​​storage is performed with 24 hours as one slice. PostGIS is used to store road network topology relationships, supporting KNN queries, and the response time is controlled within 50ms. Near-real-time data is stored in SSD and retained for 7 days, while historical data is transferred to object storage that complies with the S3 protocol. When the speed change rate is detected to be >15% or a new event is generated, database record insertion is immediately initiated.

3. The method for predicting travel time on smart roads based on multi-source data fusion and deep learning according to claim 1 is characterized in that: The deep learning model for spatiotemporal features developed in step 2 can be expressed as follows: Step 2.1 LSTM-GNN hybrid neural network architecture design principle Step 2.1.1 LSTM module realizes time series feature extraction The module input is 30 minutes of historical data, with a time step of 5 minutes and feature dimension 16 of traffic, speed, weather level, and event status data. A 3-layer bidirectional stacked LSTM structure is used, and the mathematical expression is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) The t =σ(W O ·[h t-1 ,x t ]+b O ) h t =O t ·tanh(C t ) Among them, f t For the forget gate, i t is the input gate, is a candidate memory unit, C t is the memory unit, C t-1 is the memory unit of the previous moment, O t is the output gate, h t is the hidden state, W f W i W C W O is the weight matrix, b f b i b C b O is the bias term, σ is the sigmoid activation function, t is the time step, [h t-1 ,x t ] means to change the hidden state h of the previous moment t-1 and the current input x t For splicing, input feature x t It contains the flow and speed parameters in the time window. The time series sliding window is set to 30 minutes of historical data, and the time step is t = 5 minutes. Through three layers of stacked LSTM, the hidden layer dimension is 128, and the multi-granularity time patterns at the hour, half-hour, and minute levels are captured. Finally, the time series feature vector h is output. t ; Step 2.1.2 GNN module builds spatial topological relationship Define the road network graph as G = (V, E), node v i ∈V represents the road section and intersection unit, and the edge e ij ∈E represents node v i and v j The edges connecting the nodes, i and j are used to identify the nodes in the graph; node v i To neighbor v j The attention weight α ij Calculated as: Where W is a learnable parameter, || represents vector concatenation, and N i is a node set, h i 、h j 、h k i, j and k are the feature vectors of nodes in the graph, a is a learnable attention vector, LeakyReLU is the activation function, exp is the exponential function, and T is the matrix transpose; using the attention weight α ij For node feature h i 'Update, and obtain the weighted sum of neighbor features: Among them, σ is the sigmoid activation function. After passing through multiple layers of GNN, the features of each node are obtained, and the features of all nodes are stacked to obtain the feature matrix S∈R N×64 , N is the number of nodes; Step 2.2 Attention Mechanism Implementation Using the feature S output by GNN, construct the node correlation matrix A s : A s =sigmoid(SWS T ),W∈R 64×64 Adaptively enhance the feature matrix: Among them, ⊙ is the dot product symbol, and the enhanced spatial features are obtained.

4. The method for predicting travel time on smart roads based on multi-source data fusion and deep learning according to claim 1 is characterized in that: The deployment of the real-time data processing and anomaly detection system in step 3 can be expressed as follows: Step 3.1 Building a cloud streaming computing framework Step 3.1.1 Data source access layer: connect the pre-processed data stream of the edge computing node through the Kafka message queue, and ensure the spatiotemporal continuity of data according to the "road section ID + time window" partition strategy; Step 3.1.2 Stream processing job design Time window definition: Using event time semantics, define a rolling window T = 1 minute, a sliding window step size = 10 seconds, and a window length = 5 minutes, which are used for real-time statistics and historical trend analysis respectively; Complex event detection: The Flink CEP complex event processing engine is used to define the accident mode. When the speed of five consecutive vehicles drops by more than 50% and lasts for 30 seconds, the abnormal event label is triggered. State management: Use RocksDB as the state backend to persistently store sliding window statistics and support state consistency after fault recovery; Step 3.1.3 Elastic resource scheduling: dynamically expand computing nodes based on Kubernetes and configure automatic expansion and contraction strategies; Step 3.2 EMD multi-scale decomposition and feature extraction Empirical mode decomposition (EMD) decomposes non-stationary signals into the sum of a finite number of intrinsic mode functions (IMFs): Among them, x(t) is the original signal, t is the time variable, and IMF k is the kth eigenmode function, r n (t) is the residual signal, n is the number of decompositions; Step 3.2.1 Sliding window processing: intercept the real-time vehicle speed data stream with a 60-second window, and use the mirror continuation method to eliminate the endpoint effect to ensure signal continuity; Step 3.2.2 Multi-scale feature generation, high-frequency IMF reflects sudden disturbances, and the variance threshold is set to 0.8σ 2 , σ is the standard deviation of historical normal data. When the variance exceeds the threshold, it is marked as a sudden disturbance; Low-frequency IMF characterizes periodic congestion and extracts energy entropy features E j : Step 3.3 Isolation Forest Anomaly Detection Step 3.3.1 Combine the 6-dimensional IMF energy entropy after EMD decomposition and the original vehicle speed standard deviation into a 7-dimensional feature vector, and add the enhanced spatial feature The feature vector is formed to provide multi-dimensional input features for the isolation forest anomaly detection model; Step 3.3.2 Model training: 100 isolated trees are constructed on historical normal data, and 256 instances are sampled for each tree. The mean sample path length h(x) and the anomaly score s(x) are calculated using the formula: Where c(n) is the normalization factor and h(x) is the mean path length: Among them, H(n-1) is the harmonic number, and n represents the number of instances sampled by each isolated tree; Step 3.3.3 Set the dynamic threshold s th , when s(x)>s th When an abnormal alarm is triggered, the Isolation Forest outputs the classification results of abnormal event types: accident / congestion / weather impact, and associates it with the specific road section ID and timestamp; Step 3.4 Dynamic correction module design When the isolation forest detects an anomaly and the duration exceeds 3 window periods, the model correction is started and the LSTM-GNN model weights are updated using the online gradient descent algorithm: Among them, W t , W t+1 is the weight matrix of the LSTM-GNN model at time steps t and t+1, is the loss function L with respect to weight W t The gradient of x t Model input data for time step t, y t is the true label corresponding to time step t, and the learning rate η is adaptively adjusted according to the error change rate: Among them, ε is a very small positive number to prevent the denominator from being 0, and is the gradient of the loss function at time steps t and t-1.

5. The method for predicting travel time on smart roads based on multi-source data fusion and deep learning according to claim 1 is characterized by: The adaptive prediction optimization mechanism established in step 4 can be expressed as follows: Step 4.1 Multi-objective optimization algorithm design Multi-objective optimization aims to optimize multiple conflicting objectives at the same time, and the Pareto optimal solution set is used to represent the optimal trade-off solution; the objective function is defined as: Among them, f1(w) is the prediction error (MAE), f2(w) is the model calculation delay, and f3(w) is the CPU / memory consumption; The optimization variables of the model are set to w = [w1, w2, w3, w4], which are the weight coefficients of historical average speed, real-time floating vehicle data, signal light timing, and enhanced spatial features respectively; The constraints are weight normalization w1+w2+w3+w4=1, real-time constraint calculation delay f2(w)≤200ms, resource constraint CPU utilization f3(w)≤80%; Solving the Pareto frontier method: Step 4.1.1 Initialization, randomly generate 100 weight combinations w (i) ; Step 4.1.2 Non-dominated sorting: sort the solution set in layers according to the objective function value; Step 4.1.3: Calculate the crowding degree, evaluate the distribution density of individuals in the solution set, and retain diversity; Step 4.1.4 selection and crossover, generating offspring through tournament selection and simulated binary crossover; Step 4.1.5 Iterative optimization, repeat the above steps until convergence, maximum number of iterations = 100; Step 4.2 Strengthen the learning dynamic weight adjustment model, automatically adjust the weight coefficient based on the real-time traffic status, and realize adaptive optimization of the prediction model; Step 4.2.1 Construct real-time traffic state vector Among them, v avg is the historical average speed, v real is the real-time speed of the floating vehicle, Δt signal is the remaining time of the signal light, To enhance the spatial characteristics of the node average speed and congestion index Step 4.2.2 Set the action space and weight adjustment action to a t =[Δw1,Δw2,Δw3,Δw4] Step 4.2.3 Set the reward function, based on the prediction error reward r t : Among them, α=0.7, β=0.3 are weight coefficients, MAE t is the prediction error at time t, latency t Calculate the delay for the model at time t; Step 4.2.4 Model training Step 4.2.4.1 Q-learning update rule: Among them, s t+1 is the traffic state vector at time t+1, a t+1 is the weight adjustment action at time t+1, Q(s t+1 ,a t+1 ) is state s t+1 Take action a t+1 Q value, m a axQ(s t+1 ,a) is state s t+1 The maximum Q value that can be obtained from all possible actions under the learning rate η = 0.1 and the discount factor γ = 0.9; Step 4.2.4.2 Exploration strategy, adopt ∈ greedy strategy (∈=0.1); Step 4.2.4.3 Experience playback, storage of historical transfers (s t ,a t ,r t ,s t+1 ), 32 samples are sampled each time; Step 4.3: Online learning module, which updates road state characteristics and model parameters in real time to ensure prediction adaptability; Step 4.3.1 Feature vector update: Update the road status feature vector s every 5 minutes t ; Step 4.3.2: Incremental model update, using Mini-batch gradient descent, updating the Q network parameters every 10 minutes: Among them, θ is the Q network parameter, η Q is the learning rate, ← is the assignment symbol, is the gradient of parameter θ, L(θ) is the loss function, and the loss function expression is: Among them, i is the sample index value, Q(s i ,a i ; θ) is in state s i Next, perform action a i The Q value estimated by the model with parameter θ, y i The target value is: y i =r i +γmax a (s i+1 ,a;θ - ) Among them, r i For state s i Next, perform action a i The reward value immediately obtained by the agent, γ is the discount factor, max a To find the maximum value for all possible actions a, θ - is the target network parameter, s i+1 For state s i Next, perform action a i After that, transfer to the next state; Step 4.4 Calculate the predicted travel time for the grid cells Step 4.4.1 Discretize the road network and divide the urban road network into grid cells of 500m×500m; Step 4.4.2 Prediction value calculation: For each grid cell G k , calculate the predicted travel time t k : Where k is the grid cell index, d' k is the total length of roads in the grid, f(Δt signal ) is the signal light timing influence function, is the average value of the enhanced spatial features within the grid.

Images