Commuting regular bus operation line intelligent early warning system and method

Abstract

The invention belongs to the technical field of intelligent traffic, and particularly relates to a commuting regular bus operation line intelligent early warning system, which comprises a data acquisition module, a data communication module, a processing analysis module, an early warning output module and a feedback module, and a commuting regular bus operation line intelligent early warning method comprises a risk prediction model training process, the specific process is as follows: step 1, a data preparation stage; 2, a model construction stage; 3, a model training stage; 4, a reinforcement learning optimization stage; 5, a verification and deployment stage; and step 6, a closed loop iteration stage. According to the invention, the limitation of a single data source can be broken through, the early warning precision is improved, and accidents such as sudden traffic congestion, vehicle faults, driver misoperation and dangerous behaviors of passengers can be timely handled; a processing feedback continuous iteration model is combined to adapt to a complex traffic scene; emergency resources are automatically matched based on the early warning level, and the operation loss is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent transportation, and particularly relates to an intelligent early warning system and method for the operation route of commuter buses. Background Art

[0002] The intelligent early warning system for the operation route of commuter buses combines real-time data collection, machine learning algorithms, and dynamic scheduling strategies to achieve intelligent prediction and disposal of the operation risks of buses.

[0003] In the prior art, the following problems generally exist in the operation of commuter buses: Data isolation: Traditional scheduling relies on manual experience and lacks multi-dimensional data linkage of on-vehicle devices, road monitoring, and user terminals. Lagged early warning: The response to events such as sudden traffic congestion, vehicle failures, driver operation errors, and passenger dangerous behaviors depends on after-the-fact processing, and it is impossible to achieve pre-event prediction. The alarm is not timely, and emergency measures cannot be taken immediately. Resource waste: Fixed route planning is difficult to adapt to dynamic demand changes, resulting in a high empty load rate. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent early warning system and method for the operation route of commuter buses, which can break through the limitations of a single data source, improve the accuracy of early warning, continuously iterate the model in combination with disposal feedback to adapt to complex traffic scenarios, and automatically match emergency resources based on the early warning level to reduce operation losses.

[0005] The technical solutions adopted by the present invention are specifically as follows: An intelligent early warning system for the operation route of commuter buses, comprising: Data collection module: The data collection module consists of an on-vehicle device, environmental perception, and a user terminal; Among them, the on-vehicle device includes an OBD terminal for collecting vehicle speed, fuel consumption, and vibration frequency, a GPS / Beidou dual-mode positioning device for tracking and positioning, a millimeter-wave radar for detecting the distance to the vehicle ahead, an on-vehicle camera for detecting the behavior of passengers in the carriage, a driver's seat camera for detecting the behavior information of the driver, a steering wheel pressure sensor for detecting abnormal external forces, and a microphone array for capturing high-frequency voiceprint features; Environmental perception includes a platform passenger flow counter for counting the passenger flow, a road meteorological sensor for monitoring temperature, humidity, and road waterlogging, and a traffic signal status interface; The user terminal includes an employee reservation system for counting ride demand data and real-time location feedback on the mobile side; Processing and analysis module: The processing and analysis module consists of a data fusion center unit and an intelligent analysis engine; Among them, the data fusion center unit aligns the spatio-temporal tags of vehicle-mounted, road condition, and user reservation data to construct a multi-dimensional feature matrix; The intelligent analysis engine includes an LSTM neural network and a reinforcement learning module. The LSTM neural network trains a dynamic risk prediction model based on historical accident data, and the reinforcement learning module optimizes the scheduling strategy according to the early warning disposal feedback; Based on the data acquisition module, real-time risk identification of multi-dimensional data is fused; Based on the processing and analysis module, it dynamically responds to demands, intelligently plans routes, and has visual multi-dimensional indicators and a rapid disposal mechanism.

[0006] It also includes: Data communication module: The data communication module consists of an edge computing node unit and a communication protocol; Among them, the edge computing node unit is deployed on the edge servers of vehicles and platforms for data cleaning and preliminary feature extraction; the communication protocol includes a 5G / V2X vehicle-road collaborative data transmission unit and an MQTT protocol sensor data stream transmission unit; Early warning output module: The early warning output module consists of hierarchical early warning output and visual monitoring; Among them, in the hierarchical early warning output, events with a probability < 30% are determined as low risks, events with a probability of 30% - 70% are determined as medium risks, and events with a probability > 70% are determined as high risks; the visual monitoring component displays real-time vehicle trajectories, risk heat maps, and resource scheduling status through scheduling; Feedback module: The feedback module consists of a data precipitation unit and a model iteration unit; Among them, the data precipitation unit is used to store the disposal records of early warning events and associate the operating costs; the model iteration unit dynamically updates the weights of the risk prediction model based on reinforcement learning.

[0007] An intelligent early warning method for the operation route of a commuter bus includes a risk prediction model training process, and the specific process is as follows: Step 1, data preparation stage: 1.2): Raw data collection, the data sources are vehicle-mounted devices, environmental data, and user behaviors; 1.3): Spatio-temporal alignment, synchronize the GPS timestamp with the meteorological data API; 1.4): Feature engineering, extract time series features, spatial feature fusion, and passenger dangerous behaviors; 1.5): Dataset division; Step 2, model construction stage: 2.1): Model selection, select an appropriate model architecture according to the data type; 2.2): Hyperparameter initialization to solve the "long-term dependence loss" problem of traditional RNNs; 2.3): Input layer design to capture short-term fluctuations and long-term trends; Step 3: Model training stage: 3.1): Propagation prediction to generate the spatio-temporal correlation for capturing emergencies; 3.2): Loss calculation to solve the class imbalance problem; 3.3): Backpropagation; 3.4): Parameter update; Step 4: Reinforcement learning optimization stage: 4.1): Simulation environment construction to build a safe training sandbox; 4.2): Policy network training to balance exploration of trying new policies and exploitation of optimizing known policies; 4.3): Dynamic threshold adjustment to calibrate the problems of over-sensitivity and dullness in early warnings; Step 5: Verification and deployment stage: 5.1): Cross-line cross-validation to adapt to different road topologies, passenger flow patterns, and extreme scenario simulations; 5.2): A / B testing to quantify the actual benefits of model upgrades; 5.3): Edge-side model lightweighting; Step 6: Closed-loop iteration stage: 6.1): Online data collection; 6.2): Incremental training; 6.3): Model version management.

[0008] In step 1.4, by extracting time series features, the average vehicle speed and the change rate of the congestion index in the past 1 hour are extracted; First step: Data source: Vehicle speed data: The instantaneous vehicle speed is obtained through in-vehicle GPS or roadside millimeter-wave radar, and the sampling frequency is 5 seconds / time; Congestion index: Calculated based on the average speed of the road section, and the formula is: Congestion index = Free flow speed / Current average speed × 100%; Second step: Time window sliding calculation: Average vehicle speed in the past 1 hour: First, sort the vehicle speed data by timestamp to construct a time series dataset, then define the sliding window as 1 hour, and then for each time point t, calculate the arithmetic mean of the vehicle speeds within the window: , and finally output the feature vector: ; Add optimization strategies: Use exponential weighted moving average to reduce the weight of historical data; Parallel computing: Implement real-time sliding window statistics through Spark Streaming; Step 3, congestion index change rate: First, based on the real-time calculated congestion index sequence {Ct}, and secondly, calculate the instantaneous change rate: ; Immediately afterwards, calculate the cumulative change rate within the past 1 hour: , where T is the time window length, and the unit is the same as the sampling interval; finally, output the feature vector: ; Add optimization strategies: Introduce sliding standard deviation: Synchronously calculate the fluctuation range of the change rate to characterize the stability of the traffic state; Combine the road section topological structure: Assign higher weights to key nodes such as intersections and ramps.

[0009] In step 1.4, through spatial feature fusion, obtain the road section topological structure and intersection density information; Step 1, extract the road section topological structure: Data source and preprocessing: Obtain road network vector data from the traffic management department, including road line segments, intersection coordinates, and lane number attributes. Use the ArcGIS topology tool to check for hanging point and pseudo-node errors to ensure that the road section connectivity conforms to the rules of the real road network; Topological network modeling: Convert the road line segments into a graph structure, where the nodes represent intersections or road section endpoints, the edges represent road section entities, and record the direction and speed limit attributes. Count the number of road sections connected to each node to reflect the complexity of the road network, identify closed loops, and mark them as special topological types; Feature encoding and application: Use One-Hot encoding to convert the topological types into numerical features, and use a graph neural network to learn the vector representation of road section nodes to capture topological relationships; Step 2, calculation method of intersection density information: Spatial grid division: Divide the target area into 500m×500m grids, count the number of intersections in each grid, generate a density heat map based on the intersection coordinates to reflect the regional congestion risk, statistically analyze the density changes in different time periods, and match the traffic flow patterns; Multi-scale fusion strategy: Take the target road section as the center, calculate the number of intersections within a radius of 300m, and statistically analyze the average density of intersections in the entire operating line coverage area for risk comparison and analysis; Data visualization verification: Use ArcGIS spatial overlay to overlay the density layer with real-time traffic flow data to verify the correlation between density and accident rate; Step 3, spatial feature and model fusion technology: Feature joint input: Combine the topological connectivity, intersection density, and time series data into a three-dimensional input matrix: sample × time step × feature; Model architecture design: A dual-branch neural network is adopted. LSTM processes the vehicle trajectory sequence, and GNN processes the topological graph structure to output fused features, dynamically allocating weights for topological and density features to enhance the perception of key areas; Real-time update mechanism: When the road network is renovated, the topological feature encoding is updated through online learning.

[0010] In step 1.4, through the dangerous behavior of passengers, identify the aggressive actions of passengers; The first step, data collection: Detect abnormal external force application through the steering wheel pressure sensor; capture abnormal voiceprint features through the microphone array; The second step, data processing and feature extraction: Object detection and tracking: Use the YOLOv7 algorithm to locate human key points in real time, continuously track the moving trajectory of passengers through the Kalman filter algorithm, and judge whether they enter the dangerous radius of the driving area; Construction of an aggressive action feature library: Annotate typical dangerous actions; extract spatio-temporal joint features: hand movement speed, change rate of the distance between the head and the driver; The third step, behavior recognition and hierarchical early warning: Multi-modal fusion recognition model: Use the 3D-CNN + Transformer architecture to analyze the spatio-temporal features of video clips, input 10 consecutive frames of images, output the probability value of aggressive behavior, and cross-validate with sensor data; Hierarchical response mechanism: First-level early warning, when it is detected that a passenger continuously approaches the driving area, trigger a voice prompt "Do not interfere with the driver"; second-level early warning, when a waving action is recognized, automatically reduce the vehicle speed and turn on the hazard lights; third-level early warning, when physical contact or abnormal force on the steering wheel is detected, immediately activate the emergency brake; The fourth step, system optimization and reliability guarantee: Anti-interference design: Simulate complex scenarios through the adversarial generation network to reduce the false alarm rate to less than 3%; set the judgment of behavior continuity, and only trigger an alarm when the dangerous action lasts for more than 3 seconds; Edge-cloud collaborative computing: The in-vehicle edge device performs real-time detection, and the cloud stores video clips of high-risk events for model iterative training.

[0011] In step 2.1, through model selection, select an adapted model architecture according to the data type. Select the LSTM + Attention mechanism to process vehicle trajectories and passenger flows; select the CNN branch structure: extract road section topology and intersection density, and capture long-term and short-term time dependencies; In step 2.2, solve the "long-term dependence loss" problem of traditional RNN through hyperparameter initialization. Adopt the Xavier initialization method to balance gradient propagation and set the Dropout rate to 0.3 to prevent overfitting; In step 2.3, short-term fluctuations and long-term trends are captured through the input layer design, and a three-dimensional input tensor is constructed: number of samples × time steps × feature dimensions. Time steps: The past 1 hour is divided into 12 5-minute time windows. Feature dimensions: Include four types of features: vehicle status, environmental data, user behavior, and passenger dangerous behavior.

[0012] In step 3.1, the spatio-temporal correlation of emergencies is captured through forward propagation and prediction generation. First, the three-dimensional tensor of input data is input, and then model processing is performed. The time series dependence is extracted through the time branch of the LSTM network, and the road section topological features are fused through the space branch, and the risk probability distribution for the next 15 minutes is output. In step 3.2, the class imbalance problem is solved through loss calculation. First, weighted cross-entropy loss is adopted, and higher weights are assigned to high-risk events to enhance the sensitivity of the model to key risks. The loss function is: , where the weight assignment is: is for high-risk events, is for normal events; In steps 3.3 and 3.4, combined with backpropagation and parameter update, a learning rate strategy with cosine annealing scheduling is adopted. The initial learning rate is 3e-4, and it drops to 1e-4 every 50 epochs, accelerating convergence and jumping out of local optima at the same time. In addition, an early stopping mechanism is adopted. Training is terminated when the validation set loss does not decrease for 5 consecutive rounds, and when the fluctuation range of the ROC-AUC index < 0.5%.

[0013] In step 4.1, a safe training sandbox is constructed through a simulation environment. First, a virtual traffic simulation environment is constructed based on historical operation data to simulate dynamic events during the operation of the shuttle bus. The state space is defined: including the real-time position of the vehicle, the road congestion index, and the passenger occupancy rate. The action space is defined: covering detour path selection, spare vehicle scheduling, and warning level adjustment; At the same time, a reward function is constructed: Reward function = reduction ratio of operating cost + improvement value of passenger satisfaction. Among them, the operating cost: Cost = 0.3 * detour mileage + 0.5 * passenger delay minutes + 0.2 * number of spare vehicle startups; In step 4.2, the PPO algorithm is adopted to balance exploring new strategies and exploiting and optimizing known strategies. The Actor network is used to generate actions, the Critic network is used to evaluate the long-term value of the action, and the strategy parameters are updated through gradient ascent to increase the probability of high-reward actions; In step 4.3, the problems of over-sensitivity and sluggishness are calibrated through dynamic thresholds. First, a fixed risk probability threshold is initially set; according to the time period characteristics and external factors, the Q-learning is used to dynamically adjust the threshold. When the morning peak congestion becomes normal, the threshold is increased to 75% to reduce false alarms; in rainy weather, the threshold is decreased to 65% to respond in advance.

[0014] In step 5.1, cross-line cross-validation is used to adapt to different road topologies, passenger flow patterns, and extreme scenario simulations. First, the trained model is applied to the shuttle bus line data that has not participated in the training to verify its spatio-temporal generalization ability, and then the fluctuation range of the prediction accuracy of different lines is compared; In step 5.2, the actual benefits of model upgrade are compared and quantified through A / B testing. First, the new model is deployed on some vehicles, and the old rule system is still used for the rest. Secondly, the warning response time and false alarm rate of the two groups of vehicles are collected within two weeks; In step 5.3, the TensorRT engine is used to quantize and compress the LSTM model and deploy it to the on-vehicle edge computing device; In step 6.1, the vehicle operation data is uploaded in real time through the on-vehicle terminal, and the traffic event API is synchronously accessed to form an incremental data set; In step 6.2, the rolling time window strategy is adopted: only the data of the most recent 12 months is retained, and the outdated samples are eliminated; based on the elastic weight consolidation algorithm, the new data is trained on the old model parameters; In step 6.3, the model registry is used to record the version hash value, training data set, and performance metrics, and an automatic rollback mechanism is set: when the false alarm rate of the new version increases by more than 15% in the A / B test, the old version is triggered to be restored.

[0015] The technical effects achieved by the present invention are as follows: 1. Improve safety and accident prevention capabilities Real-time risk identification: Through the fusion analysis of on-vehicle sensors (such as vibration frequency monitoring), road meteorological data (water accumulation depth), driver behavior analysis (fatigue / distraction detection), and passenger behavior analysis (passenger dangerous behavior), the abnormal state of the vehicle or the risk of harsh environment can be discovered in advance, reducing the accident rate; Long-term pattern mining: Based on the historical accident data, a model (such as an LSTM network) is trained to identify high-risk scenarios such as morning and evening peaks and specific road sections, and the scheduling strategy is optimized accordingly; 2. Optimize operation efficiency and resource utilization Dynamic demand response: The model analyzes the user reservation data and the passenger flow at the platform to predict the future travel demand, realizes the adjustment of the flexible departure interval, and reduces the empty load rate; Intelligent route planning: Combined with real-time traffic flow simulation (such as congestion index change rate prediction), it automatically recommends detour routes to reduce fuel waste and time delays caused by congestion; 3. Enhance data-driven decision support Cross-scenario generalization capability: Through pre-trained model transfer learning technology (such as cross-city traffic flow prediction), experience from areas with sufficient data can be transferred to new routes to solve the cold start problem; Multi-dimensional indicator visualization: The model outputs indicators such as risk level and cost saving rate, providing managers with quantitative decision-making basis (such as Banan Rural Traffic Risk Early Warning System); 4. Agility in responding to emergencies Rapid response mechanism: When the model predicts high-risk events (such as waterlogging due to heavy rain, driver fatigue, and dangerous passenger behavior), it automatically triggers the dispatch of backup vehicles and passenger notifications to avoid operational interruptions; Continuous learning and iteration: Through the reinforcement learning module, the warning effect is associated with the actual cost (such as loss of work time), and the response threshold and disposal strategy are dynamically optimized; 5. Promote intelligent upgrade of traffic management Technology ecosystem construction: Integrate emerging technologies such as 5G, edge computing, and V2X into model training to lay the foundation for vehicle-road collaborative systems; Extended social benefits: Reduce indirect losses from traffic accidents (such as medical expenses and road repairs) and improve public travel satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a system architecture diagram provided by an embodiment of the present invention; Figure 2 is a risk prediction model training flow chart provided by an embodiment of the present invention; Figure 3 is a risk prediction model training flow chart provided by an embodiment of the present invention; Figure 4 It is a logical tree diagram of warning classification and disposal strategy provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0018] like Figure 1 As shown, an intelligent early warning system for commuter bus operation routes includes: Data acquisition module: The data acquisition module consists of in-vehicle devices, environmental perception, and user terminals; Among them, the in-vehicle devices include an OBD terminal for collecting vehicle speed, fuel consumption, and vibration frequency, a GPS / Beidou dual-mode positioning device for tracking and positioning, a millimeter-wave radar for detecting the distance to the vehicle ahead, an in-vehicle camera for detecting the behavior of passengers in the carriage, a driver's seat camera for detecting the behavior information of the driver, a steering wheel pressure sensor for detecting abnormal external forces, and a microphone array for capturing high-frequency voiceprint features; Environmental perception includes a platform passenger flow counter for counting the passenger flow, a road meteorological sensor for monitoring temperature, humidity, and road waterlogging, and a traffic signal status interface; The user terminal includes an employee reservation system for counting ride demand data and real-time location feedback on the mobile side; Data communication module: The data communication module consists of an edge computing node unit and a communication protocol; Among them, the edge computing node unit is deployed on the edge servers of vehicles and platforms for data cleaning and preliminary feature extraction. Data cleaning is to filter out outliers and align timestamps; the communication protocol includes a 5G / V2X vehicle-road collaborative data transmission unit and an MQTT protocol sensor data stream transmission unit; Processing and analysis module: The processing and analysis module consists of a data fusion center unit and an intelligent analysis engine; Among them, the data fusion center unit aligns the spatio-temporal tags of in-vehicle, road conditions, and user reservation data to construct a multi-dimensional feature matrix (such as congestion index, overload rate, abnormal vibration frequency); The intelligent analysis engine includes an LSTM neural network and a reinforcement learning module. The LSTM neural network trains a dynamic risk prediction model based on historical accident data (input: vehicle state + environmental data; output: risk probability in the next 15 minutes), and the reinforcement learning module optimizes the scheduling strategy according to the early warning disposal feedback (such as adjusting the start threshold of standby vehicles); Early warning output module: The early warning output module consists of hierarchical early warning output and visual monitoring; Among them, events with a probability < 30% in the hierarchical early warning output are determined as low risks, events with a probability of 30% - 70% are determined as medium risks, and events with a probability > 70% are determined as high risks; visual monitoring shows the real-time vehicle trajectory, risk heat map, and resource scheduling status through scheduling. In addition, the early warning output adopts multi-terminal collaborative synchronization, including in-vehicle screens, scheduling platforms, and passenger APPs; Feedback module: The feedback module consists of a data precipitation unit and a model iteration unit; Among them, the data sedimentation unit is used to store the warning event handling records and associated operating costs (such as detour fuel consumption and loss of work time); the model iteration unit dynamically updates the risk prediction model weights based on reinforcement learning.

[0019] The working principle of this system is as follows: Please refer to the risk prediction model training process below for details.

[0020] like Figures 2 - 4 As shown, a commuter bus operation route intelligent early warning method includes a risk prediction model training process, and the specific process is as follows: Step 1: Data preparation stage: 1.2): Raw data collection, data sources include vehicle-mounted equipment, environmental data, and user behavior; 1.3): Time and space alignment, GPS timestamp synchronization with weather data API; 1.4): Feature engineering, extracting time series features, spatial feature fusion and passenger dangerous behaviors; 1.5): Dataset division; Step 2: Model building phase: 2.1): Model selection, select the appropriate model architecture according to the data type (time series, spatial features); 2.2): Hyperparameter initialization to solve the "long-term dependency loss" problem of traditional RNN; 2.3): Input layer design to capture short-term fluctuations and long-term trends; Step 3: Model training phase: 3.1): Propagation prediction, generating and capturing the spatiotemporal correlation of emergencies; 3.2): Loss calculation to solve the problem of category imbalance; 3.3): Back propagation; 3.4): Parameter update; Step 4: Reinforcement learning optimization phase: 4.1): Simulate the environment and build a safe training sandbox; 4.2): Policy network training, balancing exploration and trying new strategies with optimizing known strategies; 4.3): Dynamic threshold adjustment, calibration and warning oversensitivity and slowness issues; Step 5: Verification and deployment phase: 5.1): Cross-route cross-validation to adapt to different road topologies, passenger flow patterns and extreme scenario simulations; 5.2): A / B testing to quantify the actual benefits of model upgrades; 5.3): Lightweight edge models; Step 6: Closed-loop iteration phase: 6.1): Online data collection; 6.2): Incremental training; 6.3): Model version management.

[0021] Furthermore, in step 1.4, by extracting time series features, the average vehicle speed in the past 1 hour and the change rate of the congestion index are extracted; The first step, data source: Vehicle speed data: The instantaneous speed of the vehicle is obtained through in-vehicle GPS or roadside millimeter-wave radar, and the sampling frequency is 5 seconds / time; Congestion index: Calculated based on the average speed of the road section, and the formula is: Congestion index = free flow speed / current average speed × 100%; The second step, time window sliding calculation: Average vehicle speed in the past 1 hour: First, sort the vehicle speed data by timestamp to construct a time series data set, then define the sliding window as 1 hour (i.e., window length = 720 5-second sampling points), and then for each time point t, calculate the arithmetic mean of the vehicle speed within the window: , and finally output the feature vector: (Updated every 5 seconds); Add optimization strategies: Use exponential weighted moving average (EWMA) to reduce the weight of historical data and enhance the sensitivity to recent changes; Parallel computing: Realize real-time sliding window statistics through Spark Streaming; The third step, congestion index change rate: First, based on the real-time calculated congestion index sequence {Ct}, secondly calculate the instantaneous change rate (differential method): ; Immediately calculate the cumulative change rate within the past 1 hour (integration method): (T is the time window length, and the unit is the same as the sampling interval); Finally, output the feature vector: ; Add optimization strategies: Introduce sliding standard deviation: Synchronously calculate the fluctuation range of the change rate to characterize the stability of the traffic state; Combine the road section topological structure: Assign higher weights to key nodes such as intersections and ramps.

[0022] Furthermore, in step 1.4, through spatial feature fusion, the road section topological structure and intersection density information are obtained; The first step, extract the road section topological structure: Data source and preprocessing: Obtain road network vector data (Shapefile format) from the traffic management department, which contains attributes such as road segments, intersection coordinates, and number of lanes. Use ArcGIS topological tools to check errors such as dangle nodes and pseudo nodes to ensure that the road section connectivity conforms to the rules of the real road network; Topological Network Modeling: Convert road segments into a graph structure (Graph), where nodes represent intersections or road segment endpoints, edges represent road segment entities, and attributes such as direction and speed limit are recorded. Count the number of road segments connected to each node to reflect the complexity of the road network (e.g., the connection degree of an intersection is 4), identify closed loops (e.g., roundabouts), and mark them as special topological types; Feature Encoding and Application: Use One-Hot encoding to convert topological types (straight, branch, circular) into numerical features, and use a Graph Neural Network (GNN) to learn the vector representation of road segment nodes to capture topological relationships; Step 2. Intersection Density Information Calculation Method: Spatial Grid Division: Divide the target area into 500m×500m grids, count the number of intersections within each grid, generate a density heat map based on intersection coordinates to reflect the regional congestion risk, and count the density changes in different time periods (morning rush hour / evening rush hour) to match the traffic flow pattern; Multi-Scale Fusion Strategy: Centered on the target road segment, calculate the number of intersections within a radius of 300m, and count the average density of intersections in the area covered by the entire operation line for risk comparison and analysis; Data Visualization Verification: Use ArcGIS spatial overlay to overlay the density layer with real-time traffic flow data to verify the correlation between density and accident rate Step 3. Spatial Feature and Model Fusion Technology: Joint Feature Input: Combine topological connectivity, intersection density, and time series data (vehicle speed, passenger flow) into a three-dimensional input matrix (sample×time step×feature); Model Architecture Design: Adopt a dual-branch neural network, use LSTM to process vehicle trajectory sequences, use GNN to process topological graph structures, output fused features, dynamically allocate weights for topological and density features, and enhance the perception of key areas; Real-Time Update Mechanism: When the road network is renovated (new intersections are added), update the topological feature encoding through online learning.

[0023] Furthermore, in step 1.4, identify aggressive actions of passengers through their dangerous behaviors; Step 1. Data Collection: Detect abnormal external forces applied through the steering wheel pressure sensor, such as when the snatching force exceeds the 5kg threshold; capture abnormal voiceprint features through a microphone array, such as high-frequency quarreling sounds and hitting sounds; Step 2. Data Processing and Feature Extraction: Object Detection and Tracking: Use the YOLOv7 algorithm to real-time locate human key points, and continuously track the moving trajectory of passengers through the Kalman filtering algorithm to determine whether they enter the dangerous radius of the driving area (<1 meter); Construction of Aggressive Action Feature Library: Annotation of typical dangerous actions, such as waving and hitting, lunging forward, and continuous pushing or physical entanglement; Extraction of spatio-temporal joint features: Hand movement speed (judged as a rapid wave when >2m / s), Rate of change of the distance between the head and the driver (a sudden drop of more than 50% triggers an alarm); Step 3, Behavior Recognition and Graded Warning: Multi-modal Fusion Recognition Model: Use the 3D-CNN+Transformer architecture to analyze the spatio-temporal features of video clips, input 10 consecutive frames of images (time resolution 0.5 seconds), output the probability value of aggressive behavior (0-100%), and cross-validate with sensor data (e.g., the weight is increased by 20% when a scream is detected by the microphone); Graded Response Mechanism: Level 1 Warning (Low Risk) - When it is detected that a passenger continuously approaches the driving area, trigger a voice prompt "Do not interfere with the driver"; Level 2 Warning (Medium Risk) - When a waving action is recognized, automatically reduce the vehicle speed and turn on the hazard lights; Level 3 Warning (High Risk), when physical contact or abnormal force on the steering wheel is detected, immediately activate the emergency brake; Step 4, System Optimization and Reliability Assurance: Anti-interference Design: Simulate complex scenarios (such as accidental touches in a crowded carriage) through a Generative Adversarial Network (GAN) to reduce the false alarm rate to less than 3%; Set the judgment of behavior continuity, and only trigger an alarm when a dangerous action lasts for more than 3 seconds (excluding instantaneous interference); Edge-Cloud Collaborative Computing: In-vehicle edge devices perform real-time detection (delay <200ms), and the cloud stores video clips of high-risk events for model iterative training.

[0024] Furthermore, in step 2.1, through model selection, select an appropriate model architecture according to the data type (time series, spatial features), such as selecting the LSTM+Attention mechanism to process time series data such as vehicle trajectories and passenger flows; Select the CNN branch structure: Extract spatial features such as road section topology and intersection density, and capture long-term and short-term time dependencies; Break through the limitation of the traditional statistical model's ability to express non-linear relationships, and realize the collaborative analysis of features in the time dimension (vehicle speed change) and the spatial dimension (road section structure).

[0025] Furthermore, in step 2.2, solve the "long-term dependence loss" problem of traditional RNNs through hyperparameter initialization, use the Xavier initialization method to balance gradient propagation and set the Dropout rate to 0.3 to prevent overfitting, use bidirectional LSTM to solve the "long-term dependence loss" problem of traditional RNNs, improve the memory ability of historical accident features, and the Attention mechanism highlights the influence weight of high-risk time periods (such as specific time windows under heavy rain weather) on the prediction results.

[0026] Furthermore, in step 2.3, short-term fluctuations and long-term trends are captured through the input layer design to construct a three-dimensional input tensor: number of samples × time steps × feature dimensions. For example, time steps: the past 1 hour is divided into 12 5-minute time windows, and feature dimensions: include four types of features such as vehicle status (speed, fuel consumption), environmental data (congestion index, water depth), user behavior (number of reservations), and passenger dangerous behavior; short-term fluctuations (such as sudden congestion) and long-term trends (such as morning rush hour patterns) are captured through the time window sliding mechanism, and multi-modal feature fusion enhances the model's perception ability for complex scenarios.

[0027] Furthermore, in step 3.1, the spatio-temporal correlation of emergencies is captured through forward propagation and prediction generation. First, a three-dimensional tensor of input data is input, including features such as vehicle status (such as speed, vibration frequency), environmental data (such as congestion index, water depth), user behavior (such as number of reservations), and passenger dangerous behavior. Then, model processing is carried out. The time series dependence relationship is extracted through the time branch of the LSTM network, and the road segment topological features are fused through the space branch to output the risk probability distribution for the next 15 minutes; realizing mapping multi-modal data to a unified feature space to capture the spatio-temporal correlation of emergencies (such as a sharp increase in risk on a specific road segment due to heavy rain, a sharp increase in driving risk due to passenger dangerous behavior).

[0028] Furthermore, in step 3.2, the problem of class imbalance is solved through loss calculation. First, weighted cross-entropy loss is adopted, and higher weights (such as weight coefficient 3.0) are assigned to high-risk events (probability > 70%) to enhance the model's sensitivity to key risks. The loss function is: , where the weight assignment is: (high-risk events), (regular events); realizing the solution to the problem of class imbalance (the proportion of low-risk samples usually reaches more than 80%), avoiding the model prediction from biasing towards the majority class; in addition, the AdamW optimizer is used, combined with weight decay to prevent overfitting and minimize the false negative rate of high-risk events.

[0029] Furthermore, in steps 3.3 and 3.4, combined with backpropagation and parameter update, a learning rate strategy with cosine annealing scheduling is adopted. The initial learning rate is 3e-4, and it drops to 1e-4 every 50 epochs, accelerating convergence and jumping out of local optima at the same time; realizing the balance between the model convergence speed and generalization ability to adapt to the high dynamics of traffic data (such as the differences in morning and evening rush hour patterns); in addition, an early stopping mechanism is adopted, and the training is terminated when the validation set loss does not decrease for 5 consecutive rounds, or when the fluctuation range of the ROC-AUC index < 0.5% to prevent the model from overfitting on noisy data (such as abnormal labels caused by sensor false alarms) and improve the prediction accuracy.

[0030] Furthermore, in step 4.1, a secure training sandbox is constructed through a simulation environment. First, a virtual traffic simulation environment is built based on historical operation data (such as vehicle trajectories, accident records, weather conditions, and passenger dangerous behavior records) to simulate dynamic events during the operation of the shuttle bus (such as sudden congestion, vehicle breakdowns, and passenger dangerous behavior); the state space is defined, including variables such as the real-time position of the vehicle, the road congestion index, and the passenger occupancy rate; the action space is defined, covering decision options such as detour path selection, spare vehicle scheduling, and warning level adjustment, so as to avoid the risks brought by trial and error directly in actual operation, support high-frequency scenario simulations (such as heavy rain, accidents, and passenger dangerous behavior), and cover low-probability and high-impact events; at the same time, a reward function is constructed: Reward function = proportion of reduced operating costs + increased passenger satisfaction value, where the operating cost: Cost = 0.3 * detour mileage + 0.5 * passenger delay minutes + 0.2 * number of times of starting spare vehicles, which is used to guide the model to find a balance between reducing operating costs and improving service quality.

[0031] Furthermore, in step 4.2, the PPO algorithm is adopted to balance exploring new strategies and exploiting and optimizing known strategies. The Actor network is used to generate actions (such as "trigger a high-risk warning and dispatch a spare vehicle"), the Critic network is used to evaluate the long-term value of this action, and the strategy parameters are updated through gradient ascent to increase the probability of high-reward actions; it realizes the upgrade from static risk prediction to dynamic decision-making, adapts to real-time traffic changes, and approximates the optimal strategy through multiple rounds of iteration. For example, the response speed of the detour plan in the heavy rain scenario of a certain line is increased by 40%.

[0032] Furthermore, in step 4.3, the problems of over-sensitivity and dullness of early warnings are calibrated through dynamic thresholds. First, a fixed risk probability threshold is initially set (such as >70% to trigger a high-risk warning); according to the time period characteristics (morning peak / flat peak) and external factors (holidays / harsh weather), the Q-learning is used to dynamically adjust the threshold. For example, when congestion is normal during the morning peak, the threshold is increased to 75% to reduce false alarms; in heavy rain weather, the threshold is decreased to 65% to respond in advance; it realizes the solution to the problems of over-sensitivity or dullness of early warnings caused by traditional fixed thresholds and greatly reduces the false alarm rate.

[0033] Furthermore, in step 5.1, cross-line cross-validation is used to adapt to different road topologies, passenger flow patterns, and extreme scenario simulations. First, the trained model is applied to the shuttle bus line data that has not participated in the training (such as the spare line from Park B to C) to verify its spatio-temporal generalization ability, and then the fluctuation ranges of the prediction accuracies of different lines are compared (such as 92.3% for the core line vs 87.6% for the new line); it can avoid the overfitting of the model to specific line characteristics, ensure the adaptability to different road topologies and passenger flow patterns, and is conducive to identifying potential weak links (such as insufficient recognition of sharp turns in mountain roads).

[0034] Further, in step 5.2, the actual benefits of the quantitative model upgrade are compared through A / B testing. First, the new model is deployed in some vehicles (experimental group), and the old rule system is still used for the rest (control group). Second, the differences in indicators such as the early warning response time and false alarm rate of the two groups of vehicles within two weeks are collected; the actual benefits of the model upgrade can be quantified, and the impact of the model on the operating cost is verified through real operation data.

[0035] Further, in step 5.3, the TensorRT engine is used to quantize and compress the LSTM model and deploy it to the in-vehicle edge computing device; the model inference latency is reduced from 210ms to 65ms, meeting the real-time early warning requirements, reducing the dependence on cloud computing power, and ensuring local operation during network interruption.

[0036] Further, in step 6.1, the vehicle operation data (such as sudden hard braking, detour trajectory deviation, etc.) is uploaded in real time through the in-vehicle terminal, and synchronized with the traffic event API (such as accident alarm, temporary road closure information) to form an incremental data set; it is used to capture extreme scenarios not covered in historical training (such as tire skidding characteristics in blizzard weather), make up for the model blind spot, dynamically expand data diversity, and avoid the attenuation of prediction performance caused by traffic pattern changes.

[0037] Further, in step 6.2, the rolling time window strategy is adopted: only the data of the most recent 12 months is retained, and the outdated samples are eliminated; based on the elastic weight consolidation algorithm (EWC), new data is trained on the old model parameters to prevent catastrophic forgetting; it is used to reduce storage and computing costs, balance old and new knowledge, and maintain long-term memory of seasonal patterns (such as the Spring Festival travel rush).

[0038] Further, in step 6.3, the model registry is used to record the version hash value, training data set, and performance metrics, and an automatic rollback mechanism is set: when the false alarm rate of the new version increases by more than 15% in the A / B test, the old version is triggered to be restored; it is used to achieve the traceability of the model life cycle, meet the compliance requirements of the transportation industry, and avoid systematic operation risks caused by single iteration mistakes.

[0039] The working principle of the present invention is as follows: First, improve safety and accident prevention capabilities Real-time risk identification: Through the fusion analysis of in-vehicle sensors (such as vibration frequency monitoring), road meteorological data (water depth), driver behavior analysis (fatigue / distraction detection), and passenger behavior analysis (passenger dangerous behavior), abnormal vehicle states or harsh environment risks can be discovered in advance, reducing the accident rate; Long-term pattern mining: Based on historical accident data, a model (such as an LSTM network) is trained to identify high-risk scenarios such as morning and evening rush hours and specific road sections, and the scheduling strategy is optimized accordingly; 2. Optimize operational efficiency and resource utilization Dynamic demand response: The model analyzes user reservation data and platform passenger flow to predict future travel demand, achieve flexible departure interval adjustment, and reduce empty load rate; Intelligent route planning: Combined with real-time traffic flow simulation (such as congestion index change rate prediction), it automatically recommends detour routes to reduce fuel waste and time delays caused by congestion; 3. Enhance data-driven decision support Cross-scenario generalization capability: Through pre-trained model transfer learning technology (such as cross-city traffic flow prediction), experience from areas with sufficient data can be transferred to new routes to solve the cold start problem; Multi-dimensional indicator visualization: The model outputs indicators such as risk level and cost saving rate, providing managers with quantitative decision-making basis (such as Banan Rural Traffic Risk Early Warning System); 4. Agility in responding to emergencies Rapid response mechanism: When the model predicts high-risk events (such as waterlogging due to heavy rain, driver fatigue, and dangerous passenger behavior), it automatically triggers the dispatch of backup vehicles and passenger notifications to avoid operational interruptions; Continuous learning and iteration: Through the reinforcement learning module, the warning effect is associated with the actual cost (such as loss of work time), and the response threshold and disposal strategy are dynamically optimized; 5. Promote intelligent upgrade of traffic management Technology ecosystem construction: Integrate emerging technologies such as 5G, edge computing, and V2X into model training to lay the foundation for vehicle-road collaborative systems; Extended social benefits: Reduce indirect losses from traffic accidents (such as medical expenses and road repairs) and improve public travel satisfaction.

[0040] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. An intelligent early warning system for commuter bus operation routes, characterized in that: include: Data acquisition module: The data acquisition module is composed of vehicle-mounted equipment, environmental perception and user terminals; Among them, the on-board equipment includes an OBD terminal for collecting vehicle speed, fuel consumption, and vibration frequency, a GPS / Beidou dual-mode positioning device for tracking and positioning, a millimeter-wave radar for detecting the distance to the vehicle in front, an on-board camera for detecting the behavior of passengers in the car, a driving camera for detecting driver behavior information, a steering wheel pressure sensor for detecting abnormal external forces, and a microphone array for capturing high-frequency voiceprint features; Environmental perception includes platform passenger flow counters for counting passenger flow, road weather sensors for monitoring temperature, humidity and road water accumulation, and traffic light status interfaces; The user terminal includes an employee reservation system for collecting ride demand data and real-time location feedback on the mobile terminal; Processing and analysis module: The processing and analysis module is composed of a data fusion center unit and an intelligent analysis engine; Among them, the data fusion center unit aligns the spatiotemporal labels of vehicle-mounted, road condition, and user reservation data to construct a multi-dimensional feature matrix; The intelligent analysis engine includes LSTM neural network and reinforcement learning module. LSTM neural network trains dynamic risk prediction model based on historical accident data, and reinforcement learning module optimizes scheduling strategy according to early warning and disposal feedback. Based on the data collection module, real-time risk identification is achieved by integrating multi-dimensional data; Based on the processing and analysis module, it dynamically responds to demands, intelligently plans paths, and has both visual multi-dimensional indicators and a rapid disposal mechanism.

2. The intelligent early warning system for commuter bus operation routes according to claim 1 is characterized in that: Also includes: Data communication module: The data communication module is composed of an edge computing node unit and a communication protocol; Among them, the edge computing node unit is deployed on the edge server of the vehicle and the platform for data cleaning and preliminary feature extraction; the communication protocol includes the 5G / V2X vehicle-road cooperative data transmission unit and the MQTT protocol sensor data stream transmission unit; Warning output module: The warning output module consists of hierarchical warning output and visual monitoring; Among them, in the graded warning output, events with a probability of <30% are judged as low risk, events with a probability of 30%-70% are judged as medium risk, and events with a probability of >70% are judged as high risk; the visual monitoring component displays real-time vehicle trajectories, risk heat maps, and resource scheduling status through scheduling; Feedback module: The feedback module consists of a data precipitation unit and a model iteration unit; Among them, the data sedimentation unit is used to store warning event handling records and associated operating costs; the model iteration unit dynamically updates the risk prediction model weights based on reinforcement learning.

3. A commuter bus operation line intelligent early warning method, used for using the commuter bus operation line intelligent early warning system as claimed in claim 1, characterized in that: Including the risk prediction model training process, the specific process is as follows: Step 1: Data preparation stage: 1.2): Raw data collection, data sources include vehicle-mounted equipment, environmental data, and user behavior; 1.3): Time and space alignment, GPS timestamp synchronization with weather data API; 1.4): Feature engineering, extracting time series features, spatial feature fusion and passenger dangerous behaviors; 1.5): Dataset division; Step 2: Model building phase: 2.1): Model selection, select the appropriate model architecture according to the data type; 2.2): Hyperparameter initialization to solve the "long-term dependency loss" problem of traditional RNN; 2.3): Input layer design to capture short-term fluctuations and long-term trends; Step 3: Model training phase: 3.1): Propagation prediction, generating and capturing the spatiotemporal correlation of emergencies; 3.2): Loss calculation to solve the problem of category imbalance; 3.3): Back propagation; 3.4): Parameter update; Step 4: Reinforcement learning optimization phase: 4.1): Simulate the environment and build a safe training sandbox; 4.2): Policy network training, balancing exploration and trying new strategies with optimizing known strategies; 4.3): Dynamic threshold adjustment, calibration and warning oversensitivity and slowness issues; Step 5: Verification and deployment phase: 5.1): Cross-route cross-validation to adapt to different road topologies, passenger flow patterns and extreme scenario simulations; 5.2): A / B testing to quantify the actual benefits of model upgrades; 5.3): Lightweight edge models; Step 6: Closed-loop iteration phase: 6.1): Online data collection; 6.2): ​​Incremental training; 6.3): Model version management.

4. The intelligent early warning method for commuter bus operation routes according to claim 3 is characterized by: In step 1.4, the average vehicle speed and congestion index change rate in the past hour are extracted by extracting time series features; Step 1: Data source: Vehicle speed data: The vehicle's instantaneous speed is obtained through the vehicle's onboard GPS or roadside millimeter-wave radar, with a sampling frequency of 5 seconds per time; Congestion index: calculated based on the average speed of the road section, the formula is: Congestion index = free flow speed / current average speed free flow speed × 100%; Step 2: Time window sliding calculation: Average speed in the past hour: First, sort the speed data by timestamp to construct a time series data set, then define the sliding window as 1 hour, and then calculate the arithmetic mean of the speed in the window for each time point t: , and finally output the feature vector: ; Add optimization strategy: use exponentially weighted moving average to reduce the weight of historical data; parallel computing: achieve real-time sliding window statistics through Spark Streaming; Step 3: Congestion Index Change Rate: First, based on the real-time calculated congestion index sequence {Ct}, the instantaneous rate of change is calculated: ; Then calculate the cumulative rate of change in the past hour: , T is the time window length, the unit is consistent with the sampling interval; finally, the feature vector is output: ; Add optimization strategies: Introduce sliding standard deviation: Synchronously calculate the fluctuation amplitude of the rate of change to characterize the stability of the traffic state; Combine the road section topology structure: Give higher weights to key nodes of intersections and ramps.

5. The intelligent early warning method for commuter bus operation routes according to claim 3 is characterized by: In step 1.4, the road section topology and intersection density information are obtained through spatial feature fusion; The first step is to extract the road section topology: Data source and preprocessing: Obtain road network vector data from the traffic management department, including road segments, intersection coordinates, and lane number attributes. Use ArcGIS topology tools to check for hanging points and pseudo node errors to ensure that the connectivity of the road segments complies with the real road network rules. Topological network modeling: convert road segments into graph structures, with nodes representing intersections or segment endpoints and edges representing segment entities. It also records direction and speed limit attributes, counts the number of segments connected to each node, reflects the complexity of the road network, identifies closed loops, and marks them as special topological types. Feature encoding and application: Use One-Hot encoding to convert topological types into numerical features, and use graph neural networks to learn vector representations of road segment nodes to capture topological relationships; Step 2: Calculation method of intersection density information: Spatial grid division: Divide the target area into 500m×500m grids, count the number of intersections in each grid, generate density heat maps based on intersection coordinates to reflect regional congestion risks, count density changes by time period, and match traffic flow patterns; Multi-scale fusion strategy: With the target road section as the center, the number of intersections within a radius of 300m is calculated, and the average density of intersections in the entire operating line coverage area is counted for risk comparison analysis; Data visualization verification: Using ArcGIS spatial overlay to overlay density layers with real-time traffic flow data, verify the correlation between density and accident rate; Step 3: Spatial feature and model fusion technology: Feature joint input: Combine topological connectivity, intersection density and time series data into a three-dimensional input matrix: sample × time step × feature; Model architecture design: using a dual-branch neural network, LSTM to process vehicle trajectory sequences, GNN to process topological graph structures, output fusion features, dynamically allocate weights of topological and density features, and enhance key area perception; Real-time update mechanism: When the road network is modified, the topological feature encoding is updated through online learning.

6. The intelligent early warning method for commuter bus operation routes according to claim 3 is characterized by: In step 1.4, the passenger's aggressive actions are identified through the passenger's dangerous behavior; Step 1: Data collection: Detect abnormal external force through the steering wheel pressure sensor; capture abnormal voiceprint features through the microphone array; Step 2: Data processing and feature extraction: Target detection and tracking: The YOLOv7 algorithm is used to locate the key points of the human body in real time, and the Kalman filter algorithm is used to continuously track the movement trajectory of passengers to determine whether they have entered the dangerous radius of the driving area; Aggressive action feature library construction: typical dangerous action annotation; extraction of spatiotemporal joint features: hand movement speed, rate of change of distance between head and driver; Step 3: Behavior identification and graded warning: Multimodal fusion recognition model: Use 3D-CNN+Transformer architecture to analyze the spatiotemporal features of video clips, input 10 consecutive frames of images, output the probability value of attack behavior, and combine sensor data for cross-validation; Gradual response mechanism: Level 1 warning: When a passenger is detected approaching the driving area, a voice prompt "Do not disturb the driver" is triggered; Level 2 warning: When a waving gesture is recognized, the vehicle speed is automatically reduced and the hazard lights are turned on; Level 3 warning: When physical contact or abnormal force on the steering wheel is detected, emergency braking is immediately initiated; Step 4: System optimization and reliability assurance: Adversarial interference design: Simulate complex scenarios through adversarial generative networks to reduce the false alarm rate to less than 3%; Set the behavior continuity judgment, and the alarm will be triggered only when the dangerous action lasts for more than 3 seconds; Edge-cloud collaborative computing: Onboard edge devices perform real-time detection, and the cloud stores video clips of high-risk events for iterative model training.

7. The intelligent early warning method for commuter bus operation routes according to claim 3 is characterized by: In step 2.1, through model selection, select the appropriate model architecture according to the data type, select the LSTM+Attention mechanism to process vehicle trajectories and passenger flows; select the CNN branch structure: extract the road section topology and intersection density, and capture the long-term and short-term time dependencies; In step 2.2, the "long-term dependency loss" problem of traditional RNN is solved by initializing hyperparameters, using the Xavier initialization method to balance gradient propagation and setting the Dropout rate to 0.3 to prevent overfitting; In step 2.3, the input layer is designed to capture short-term fluctuations and long-term trends, and a three-dimensional input tensor is constructed: number of samples × time step × feature dimension, time step: the past hour is divided into 12 5-minute time windows, feature dimension: contains 4 types of features: vehicle status, environmental data, user behavior, and passenger dangerous behavior.

8. The intelligent early warning method for commuter bus operation routes according to claim 3 is characterized by: In step 3.1, the spatiotemporal correlation of emergencies is captured through forward propagation and prediction generation. First, the three-dimensional data tensor is input, and then the model is processed. The time series dependency is extracted through the time branch of the LSTM network, and the spatial branch integrates the topological features of the road section to output the risk probability distribution in the next 15 minutes; In step 3.2, the class imbalance problem is solved by loss calculation. First, weighted cross entropy loss is used to give higher weights to high-risk events and strengthen the model's sensitivity to key risks. The loss function is: , where the weight distribution is: For high-risk events, For routine events; In step 3.3 and step 3.4, back propagation and parameter update are combined, and the learning rate strategy of cosine annealing scheduling is adopted. The initial learning rate is 3e-4, and it decreases to 1e-4 every 50 epochs to accelerate convergence and escape the local optimum. In addition, the early stopping mechanism is adopted to terminate the training when the validation set loss does not decrease for 5 consecutive rounds, and when the fluctuation range of the ROC-AUC indicator is <0.5%.

9. The intelligent early warning method for commuter bus operation routes according to claim 3 is characterized by: In step 4.1, a safe training sandbox is built through a simulation environment. First, a virtual traffic simulation environment is built based on historical operation data to simulate dynamic events during the operation of the shuttle bus. The state space is defined, including the real-time location of the vehicle, the road congestion index, and the passenger load factor. The action space is defined, including the detour path selection, spare vehicle scheduling, and warning level adjustment. At the same time, a reward function is constructed: reward function = operating cost reduction ratio + passenger satisfaction improvement value, where operating cost: Cost = 0.3*detour mileage + 0.5*passenger delay minutes + 0.2*number of standby vehicle starts; In step 4.2, the PPO algorithm is used to balance the exploration and trial of new strategies with the optimization of known strategies. The Actor network is used to generate actions, the Critic network is used to evaluate the long-term value of the action, and the strategy parameters are updated through gradient ascent to increase the probability of high-reward actions. In step 4.3, the oversensitivity and slowness of early warning are corrected through dynamic threshold calibration. First, a fixed risk probability threshold is initially set. The threshold is dynamically adjusted using Q-learning according to time characteristics and external factors. When morning rush hour congestion becomes normalized, the threshold is raised to 75% to reduce false alarms. In heavy rain, the threshold is lowered to 65% to respond in advance.

10. The intelligent early warning method for commuter bus operation routes according to claim 3, characterized in that: In step 5.1, the trained model is first applied to the shuttle bus route data that has not participated in the training to verify its spatiotemporal generalization ability through cross-route cross-validation to adapt to different road topologies, passenger flow patterns and extreme scenario simulations, and then the prediction accuracy fluctuation range of different routes is compared; In step 5.2, the actual benefits of the quantitative model upgrade are compared through A / B testing. First, the new model is deployed on some vehicles, and the rest use the old rule system. Then, the warning response time and false alarm rate of the two groups of vehicles are collected within two weeks. In step 5.3, the LSTM model is quantized and compressed using the TensorRT engine and deployed to the vehicle-mounted edge computing device; In step 6.1, the vehicle operation data is uploaded in real time through the vehicle terminal, and the traffic event API is simultaneously accessed to form an incremental data set; In step 6.2, a rolling time window strategy is adopted: only the data of the last 12 months is retained and outdated samples are eliminated; based on the elastic weight merging algorithm, new data training is superimposed on the old model parameters; In step 6.3, use the model registry to record the version hash value, training data set, and performance indicators, and set up an automatic rollback mechanism: when the false positive rate of the new version in the A / B test increases by more than 15%, the old version is restored.

Images