Logistics vehicle intelligent scheduling method and system based on big data

Abstract

The invention relates to the technical field of logistics big data, and discloses a logistics vehicle intelligent scheduling method and system based on big data. According to the method, multi-source heterogeneous data from a vehicle-mounted sensor, a traffic monitoring platform and the like are collected in real time, and after distributed cleaning, standardization processing and outlier detection are performed, a vehicle-environment state feature vector is generated by using a data fusion technology. Furthermore, a vehicle scheduling decision model is constructed based on a deep reinforcement learning framework, a scheduling sequence is generated and optimized through a strategy network and a multi-objective optimization algorithm, and finally the scheduling sequence is coded into an executable instruction set to drive a vehicle to execute. The system also realizes closed-loop dynamic optimization, ensures that the scheduling strategy adapts to the environmental change, and improves the logistics efficiency.

Description

Technical Field

[0001] This invention relates to the field of logistics big data technology, specifically to a method and system for intelligent scheduling of logistics vehicles based on big data. Background Technology

[0002] Traditional logistics scheduling relies on manual experience or simple rules, making it difficult to integrate multi-source heterogeneous data such as vehicle status, traffic environment, and cargo demand in real time, resulting in unreasonable route planning and uneven resource allocation.

[0003] Existing methods mostly use static models, which cannot dynamically adapt to sudden factors such as traffic congestion and weather changes, resulting in high transportation costs and poor timeliness.

[0004] Although some systems have introduced basic data analysis, they lack deep learning and multi-objective optimization capabilities, making it difficult to achieve global optimization in terms of cost, efficiency, and resource utilization.

[0005] As logistics scales up, traditional methods can no longer meet the needs of intelligent scheduling in complex scenarios. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a big data-based intelligent scheduling method and system for logistics vehicles, solving the problems of low efficiency, poor resource utilization, and weak adaptability of existing logistics scheduling methods.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for intelligent scheduling of logistics vehicles based on big data, the method comprising: Real-time data related to logistics vehicle scheduling is collected from multiple heterogeneous data sources, including vehicle-mounted sensors, traffic monitoring platforms, cargo management databases, and historical operation record databases, to obtain real-time vehicle location data, cargo delivery demand data, traffic flow data, and environmental factor data. The collected data undergoes distributed cleaning and standardization, outlier detection algorithms are used to eliminate data anomalies, and data fusion technology is employed to integrate multi-source data into a unified vehicle-environment state feature vector. The vehicle-environment state feature vector includes vehicle state features, environmental features, and location features. The vehicle state features include vehicle load rate, driving speed, and remaining fuel. The environmental features include traffic congestion index and weather condition indicators. The location features include real-time latitude and longitude coordinates and road segment identification. Based on vehicle-environment state feature vectors, a vehicle scheduling decision model is constructed using a deep reinforcement learning framework. Candidate strategies for vehicle paths and scheduling sequences are generated through a policy network, and a multi-objective optimization algorithm is used to optimize and select the candidate strategies. The multi-objective optimization algorithm considers transportation cost, time efficiency and resource utilization simultaneously. The scheduling sequence is encoded into an executable instruction set and sent to the on-board control unit of the corresponding logistics vehicle through a wireless communication network to drive the vehicle to execute the scheduling task. Real-time monitoring of vehicle location data, task execution status, and changes in the external environment; collection of feedback data; and updating of parameters in the vehicle dispatching decision model based on time difference error to achieve closed-loop dynamic optimization.

[0008] Preferably, in one possible implementation of the first aspect, the outlier detection algorithm specifically comprises: A local outlier factor algorithm based on dynamic time warping is used to process time-series data streams, which include real-time vehicle location data and traffic flow data. The local outlier factor algorithm divides the data stream into continuous segments using a sliding window and uses a combination of k-nearest neighbor algorithm to calculate the local density deviation of each data point within a segment; The detection threshold is initialized based on the distribution characteristics of historical data and is adaptively updated through an exponentially weighted dynamic mechanism; For the identified outliers, a probabilistic interpolation method based on Gaussian mixture model is used to repair the data and maintain the integrity of the data sequence.

[0009] Preferably, in one possible implementation of the first aspect, the data fusion technology includes: After eliminating data anomalies, the data is normalized to eliminate dimensional differences. The normalized data is divided into time-series dynamic feature data and static feature data; The time-series dynamic feature data includes real-time vehicle location data, traffic flow data, and dynamic environmental features derived from environmental factor data; the static feature data includes vehicle inherent attribute data and cargo fixed information. For time-series dynamic feature data, a multi-head self-attention network is input, and the weight coefficients of each feature in the fusion process are dynamically allocated by calculating the correlation scores between features. The multi-head self-attention network integrates a position encoder to retain the time-series attributes of the data. For static feature data, a fully connected neural network is used for feature transformation and fusion; The processed temporal dynamic features and static features are concatenated and compressed into a unified vehicle-environment state feature vector.

[0010] Preferably, in one possible implementation of the first aspect, the vehicle scheduling decision model adopts a deep reinforcement learning-based framework, including a feature extraction network, a policy network, and a value network; The feature extraction network is used to extract spatiotemporal feature vectors from vehicle-environment state feature vectors. It is composed of a one-dimensional convolutional neural network and a long short-term memory network connected in series. The one-dimensional convolutional neural network is used to extract the local spatial patterns of the vehicle-environment state feature vectors, and the long short-term memory network is used to capture the temporal dependencies of the features. The policy network is used to generate vehicle scheduling policies, which include vehicle routes and scheduling sequences. The value network is used to evaluate the long-term benefits of the scheduling strategy; The multi-objective optimization algorithm employs an improved non-dominated sorting genetic algorithm, which maintains a Pareto optimal solution set in each generation of evolution and preserves the diversity of solutions by calculating the crowding degree of the solutions. Finally, it selects the scheduling sequence with the highest comprehensive utility from the solution set as the output.

[0011] Preferably, in one possible implementation of the first aspect, the policy network adopts an encoder-decoder architecture, wherein its input layer receives a spatiotemporal feature vector extracted by a feature extraction network; The encoder consists of multiple fully connected layers, each followed by a batch normalization layer and a ReLU activation function. The decoder uses an attention-based recurrent neural network structure to gradually generate vehicle path sequences and scheduling sequences; The output layer of the policy network uses the Softmax function to output the probability distribution of vehicle path selection and the probability distribution of vehicle scheduling sequence arrangement at each decision time.

[0012] Preferably, in one possible implementation of the first aspect, the input to the value network includes a spatiotemporal feature vector extracted by the feature extraction network and a vehicle scheduling policy output by the policy network, and the output is a long-term value estimate scalar for the current state-action pair. The value network contains multiple fully connected layers to achieve a non-linear mapping from input features to value estimation; The value network is trained using a time difference learning algorithm to minimize the mean square error between the predicted value and the target value.

[0013] Preferably, in one possible implementation of the first aspect, the operation of the improved non-dominated sorting genetic algorithm includes: Initialize the population, perform non-dominated sorting of the individuals in the population, and calculate the crowding degree; Tournament selection is based on an individual's front-end ranking and congestion level; The offspring population is generated using simulated binary crossover and polynomial mutation operations. Among them, the crossover probability and mutation probability are adaptively adjusted based on the diversity index of the current generation of the population; After each generation of evolution, a local search operation is applied to the individuals in the Pareto optimal solution set, and the local search employs a perturbation operator related to the problem domain. The algorithm calculates the comprehensive utility value of each solution from the Pareto optimal solution set of the last generation population according to the preset preference rules or weight coefficients, and selects the scheduling sequence with the highest utility value as the output.

[0014] Preferably, in one possible implementation of the first aspect, the process of encoding the scheduling sequence into an executable instruction set includes: Based on the preset scheduling operation specification library, the scheduling sequence is parsed into a series of atomic operation instructions, each atomic operation instruction corresponding to a basic action that a vehicle can perform; According to the standard communication protocol of the vehicle control unit, atomic operation instructions are serialized into structured data packets. The data packets are encapsulated in JSON format and appended with timestamps and vehicle identification identifiers. Data packets are compressed using a data compression algorithm, and forward error correction codes are added to ensure transmission reliability, resulting in the final executable instruction set.

[0015] Preferably, in one possible implementation of the first aspect, the closed-loop dynamic optimization process specifically includes: After receiving and parsing the executable instruction set, the vehicle control unit drives the vehicle to perform the corresponding actions; At the same time, vehicle positioning data, engine status parameters and real-time traffic information are continuously collected by onboard sensors as feedback data. The feedback data is sent back to the central processing system. Based on the time difference error between the actual reward signal and the value network prediction value, the online learning process of the model is triggered. The reinforcement learning algorithm is used to incrementally update the parameters of the policy network and the value network to achieve closed-loop optimization of the scheduling strategy.

[0016] Secondly, the present invention provides a big data-based intelligent dispatching system for logistics vehicles, the system comprising: The data acquisition module collects logistics vehicle scheduling data in real time from multiple heterogeneous data sources; The data preprocessing and fusion module cleans, standardizes, and detects outliers in the collected data, and performs multi-source data fusion to generate the vehicle-environment state feature vector. The scheduling decision calculation module runs the vehicle scheduling decision model and multi-objective optimization algorithm to generate scheduling sequences. The instruction encoding and communication module encodes the scheduling sequence into an executable instruction set and sends it out via a wireless communication network. The monitoring and closed-loop feedback module monitors the vehicle's execution status, collects feedback data, and drives the dynamic update of the scheduling decision model.

[0017] The beneficial effects of this invention are as follows: This invention uses big data technology to achieve real-time decision-making in dynamic environments through multi-source data fusion and deep reinforcement learning.

[0018] The system integrates vehicle sensors, traffic platforms, and historical data to construct a unified vector that includes vehicle status, environmental features, and location features, thereby improving data utilization.

[0019] Multi-objective optimization based on an improved non-dominated sorting genetic algorithm can simultaneously reduce transportation costs, improve time efficiency, and optimize resource allocation.

[0020] The closed-loop dynamic optimization mechanism adjusts model parameters through real-time feedback to ensure that the scheduling strategy adapts to traffic changes and task requirements.

[0021] In summary, this method can reduce empty runs and delays, thereby improving overall logistics efficiency. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This application provides a flowchart of a big data-based intelligent scheduling method for logistics vehicles. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1: As Figure 1 As shown, this invention provides a big data-based intelligent scheduling method for logistics vehicles, comprising: Real-time data related to logistics vehicle scheduling is collected from multiple heterogeneous data sources, including vehicle-mounted sensors, traffic monitoring platforms, cargo management databases, and historical operation record databases, to obtain real-time vehicle location data, cargo delivery demand data, traffic flow data, and environmental factor data.

[0026] In this embodiment, the data acquisition process is implemented through a distributed data acquisition module. For vehicle-mounted sensors, the acquisition module receives real-time vehicle status data streams generated by the sensors via the communication interface built into the vehicle control unit. This includes latitude and longitude coordinates, driving speed output by the speed sensor, cargo load rate detected by the load sensor, and remaining fuel level reported by the fuel level sensor. This data is continuously transmitted using a high-frequency sampling method to ensure the timeliness of location and status information. For the traffic monitoring platform, the acquisition module uses an asynchronous communication protocol to access the real-time data service provided by the platform, acquiring traffic flow data covering the entire road network, such as average vehicle speed, congestion index, and accident alarm information.

[0027] Access to the cargo management database is achieved through a Structured Query Language (SCL) interface. The data acquisition module periodically performs database queries to extract details of the current delivery task, including required data such as cargo type, destination, priority, and estimated delivery time. Simultaneously, the acquisition module batch-retrieves historical vehicle dispatch records, route selection patterns, and resource utilization efficiency data from the historical operations record database. Environmental factor data is derived from the meteorological service interface and road condition monitoring system. The acquisition module integrates multi-source meteorological reports and real-time environmental sensor readings to generate weather condition indicators such as precipitation, visibility, and temperature impact coefficients.

[0028] All data acquisition operations adhere to real-time principles, and data streams are timestamped to ensure temporal consistency. Protocol differences between heterogeneous data sources are handled by the acquisition module's adaptation layer. During acquisition, the module employs a heartbeat mechanism to monitor connection status and, in the event of a network interruption, enables local caching to temporarily store data, which is then retransmitted upon network recovery, ensuring data integrity.

[0029] The collected data undergoes distributed cleaning and standardization, outlier detection algorithms are used to eliminate data anomalies, and data fusion technology is employed to integrate multi-source data into a unified vehicle-environment state feature vector. The vehicle-environment state feature vector includes vehicle state features, environmental features, and location features. Vehicle state features include vehicle load rate, driving speed, and remaining fuel. Environmental features include traffic congestion index and weather condition indicators. Location features include real-time latitude and longitude coordinates and road segment identification.

[0030] In this embodiment, the data preprocessing and fusion module processes the collected multi-source heterogeneous data to generate vehicle-environment state feature vectors. The processing flow begins with distributed data cleaning. The module uses a distributed computing framework to distribute the data stream to multiple parallel processing units, performing operations such as removing duplicate records, correcting format errors, and filling in missing values. For missing values, a time-series-based linear interpolation method is used to fill in the missing values, ensuring data continuity.

[0031] To ensure the accuracy of subsequent analysis, the cleaned data was standardized to eliminate dimensional differences, and the min-max scaling method was used to normalize various numerical features to a unified numerical range.

[0032] Building upon this foundation, the module performs outlier detection and anomaly removal. This process employs a local outlier factor algorithm based on dynamic time warping, used to process time-series data such as real-time vehicle locations and traffic flow. The algorithm divides the data into continuous segments using a sliding window and combines this with the k-nearest neighbor algorithm to calculate local density deviations for each data point to identify anomalies. The detection threshold is initialized based on historical data distribution and dynamically updated using an exponential weighting mechanism. For identified outliers, a probabilistic interpolation method based on a Gaussian mixture model is used for repair, maintaining the integrity of the data sequence.

[0033] After anomaly handling, the data enters the data fusion stage. The module first divides the data into two main categories: time-series dynamic feature data and static feature data. Time-series dynamic feature data includes real-time vehicle location, traffic flow, and derived dynamic environmental features; static feature data includes inherent vehicle attributes and fixed cargo information.

[0034] For temporally dynamic features, a multi-head self-attention network integrating a position encoder is input. This network dynamically assigns weight coefficients to each feature during the fusion process by calculating the correlation scores between features, capturing long-term dependencies and preserving temporal attributes. For static features, a fully connected neural network is used for feature transformation and fusion.

[0035] Finally, the two types of processed features are concatenated and dimensionality reduced through a compression layer, integrating them into a unified vehicle-environment state feature vector. This vector includes vehicle state features such as load rate, driving speed, and remaining fuel, environmental features such as traffic congestion index and weather conditions, and location features such as real-time latitude and longitude coordinates and road segment identification.

[0036] Based on vehicle-environment state feature vectors, a vehicle scheduling decision model is constructed using a deep reinforcement learning framework. Candidate strategies for vehicle paths and scheduling sequences are generated through a policy network, and a multi-objective optimization algorithm is used to optimize and select the candidate strategies. The multi-objective optimization algorithm considers transportation cost, time efficiency, and resource utilization simultaneously.

[0037] In this embodiment, the construction process of the vehicle scheduling decision model first involves the deployment of a feature extraction network. The feature extraction network consists of a one-dimensional convolutional neural network and a long short-term memory network connected in series. The one-dimensional convolutional neural network is responsible for extracting local spatial patterns from the vehicle-environment state feature vector, while the long short-term memory network is used to capture the temporal dependencies of the features, thereby outputting a spatiotemporal feature vector. This spatiotemporal feature vector serves as the basic input for subsequent networks, ensuring that the model can simultaneously consider the influence of spatial distribution and temporal evolution.

[0038] Next, the policy network employs an encoder-decoder architecture. The encoder consists of multiple fully connected layers, each followed by a batch normalization layer and a ReLU activation function to enhance the network's stability and non-linear expressive power. After receiving the spatiotemporal feature vectors, the encoder performs feature transformation and dimensionality adjustment through the fully connected layers. The decoder uses a recurrent neural network structure based on an attention mechanism. This structure progressively generates vehicle path sequences and scheduling sequences. The attention mechanism allows the model to dynamically focus on key information in the input sequence when generating each decision point, improving the accuracy of sequence generation. The output layer of the policy network uses the Softmax function to output the probability distribution of vehicle path selection and the probability distribution of vehicle scheduling sequence arrangement at each decision time. These probability distributions provide the basis for generating candidate policies.

[0039] The value network is used to evaluate the long-term benefits of scheduling policies generated by the policy network. The inputs to the value network include the spatiotemporal feature vectors output by the feature extraction network and the vehicle scheduling policies output by the policy network. Multiple fully connected layers implement a non-linear mapping from input features to a scalar value estimate. The value network is trained using a temporal difference learning algorithm to minimize the mean squared error between the predicted and target values, thus ensuring the accuracy of the value estimate. Temporal difference learning uses the error between the actual reward signal and the predicted value for incremental updates, enabling the model to adapt to dynamic environmental changes.

[0040] After generating candidate strategies, an improved non-dominated sorting genetic algorithm is used for multi-objective optimization. The algorithm first initializes a population, where each individual represents a possible scheduling strategy. Then, the individuals are non-dominatedly sorted, and a frontier rank is assigned based on their performance across multiple objectives, including transportation cost, time efficiency, and resource utilization. The crowding level of each individual is also calculated to maintain the diversity of the solution set. Based on the frontier rank and crowding level, the algorithm performs a tournament selection operation, choosing superior individuals for the next generation. Next, simulated binary crossover and polynomial mutation operations are used to generate the offspring population, where the crossover and mutation probabilities are adaptively adjusted based on the diversity index of the current generation to balance exploration and utilization. After each generation, a local search operation is applied to the individuals in the Pareto optimal solution set. This local search employs perturbation operators relevant to the problem domain, such as slightly adjusting path order or resource allocation, to improve the quality of the solutions. From the Pareto optimal solution set of the last generation, the algorithm calculates the comprehensive utility value of each solution according to preset preference rules or weight coefficients, and finally selects the scheduling sequence with the highest utility value as the output.

[0041] The entire model training process employs a deep reinforcement learning framework, achieving policy optimization through the interaction of the policy network and the value network. In the early stages of training, the model explores the solution space using a random policy. As training progresses, the value estimate provided by the value network guides the policy network towards better updates. A multi-objective optimization algorithm ensures that the final policy achieves a balance among multiple conflicting objectives, meeting the complex requirements of logistics scheduling.

[0042] The scheduling sequence is encoded into an executable instruction set and sent to the on-board control unit of the corresponding logistics vehicle through a wireless communication network to drive the vehicle to execute the scheduling task.

[0043] In this embodiment, the instruction encoding and communication module is responsible for executing this process. First, the module parses the generated scheduling sequence into a series of atomic operation instructions based on a preset scheduling operation specification library. Each atomic operation instruction corresponds to a basic action that a vehicle can perform, such as starting the engine, controlling the steering mechanism, adjusting the throttle opening and braking pressure, or performing loading and unloading operations.

[0044] Next, the module serializes atomic operation instructions into structured data packets according to the standard communication protocol of the vehicle control unit. The data packets are encapsulated in JSON format and embedded with timestamps and vehicle identification. Subsequently, lossless compression algorithms are used to compress the packets, and forward error correction codes are added to improve transmission reliability, ultimately forming an executable instruction set.

[0045] The instruction set is transmitted via a wireless communication network. After receiving it, the vehicle control unit performs decompression and error correction decoding, parses the JSON message to extract the atomic operation instruction sequence, and verifies the instruction's identity, timeliness, and logical rationality. Once verified, the vehicle control unit drives the vehicle's drive system, steering system, and other actuators to automatically execute the scheduling task.

[0046] Real-time monitoring of vehicle location data, task execution status, and changes in the external environment; collection of feedback data; and updating of parameters in the vehicle dispatching decision model based on time difference error to achieve closed-loop dynamic optimization.

[0047] In this embodiment, after receiving and parsing the executable instruction set, the onboard control unit drives the vehicle to perform corresponding scheduling actions. Simultaneously, it continuously collects vehicle positioning data, engine status parameters, and real-time traffic information as feedback data through the onboard sensor network. The feedback data includes the vehicle's actual location coordinates, speed changes, fuel consumption, engine operating status, and dynamic road condition information from the traffic monitoring platform.

[0048] This feedback data is transmitted back to the central processing system in real time via a wireless communication network. The central processing system preprocesses the feedback data, including data verification and time alignment, to ensure data accuracy and consistency. Subsequently, the system calculates the time difference error based on the actual reward signal and the value network's predicted value. The actual reward signal is generated based on the achievement of scheduling objectives, including multi-dimensional indicators such as transportation cost savings, time efficiency improvements, and resource utilization optimization. The time difference error reflects the deviation between actual and predicted benefits and is used to evaluate the performance of the current scheduling strategy.

[0049] When the time difference error exceeds a preset threshold, the system triggers the online learning process of the model. The online learning employs a reinforcement learning algorithm, specifically the asynchronous advantage actor-critic algorithm, to incrementally update the parameters of the policy network and the value network. The policy network update focuses on adjusting the probability distribution of policies for vehicle path selection and scheduling sequence generation, improving decision-making accuracy. The value network update aims to optimize long-term value estimation, making predictions closer to actual rewards. The update process uses stochastic gradient descent to minimize the mean squared loss function of the time difference error.

[0050] Through continuous iterative monitoring, feedback, and parameter updates, the vehicle dispatching decision model can adapt to environmental changes and task requirements, gradually optimizing dispatching strategies. This closed-loop dynamic optimization mechanism ensures the system's robustness and efficiency in complex logistics scenarios, achieving intelligent and adaptive dispatching processes.

[0051] Example 2: This invention provides a big data-based intelligent dispatching system for logistics vehicles, comprising: The data acquisition module collects logistics vehicle scheduling data in real time from multiple heterogeneous data sources; The data preprocessing and fusion module cleans, standardizes, and detects outliers in the collected data, and performs multi-source data fusion to generate vehicle-environment state feature vectors. The scheduling decision calculation module runs the vehicle scheduling decision model and multi-objective optimization algorithm to generate scheduling sequences. The instruction encoding and communication module encodes the scheduling sequence into an executable instruction set and sends it out via a wireless communication network. The monitoring and closed-loop feedback module monitors the vehicle's execution status, collects feedback data, and drives the dynamic update of the scheduling decision model.

[0052] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for intelligent scheduling of logistics vehicles based on big data, characterized in that, The method includes: Real-time data related to logistics vehicle scheduling is collected from multiple heterogeneous data sources, including vehicle-mounted sensors, traffic monitoring platforms, cargo management databases, and historical operation record databases, to obtain real-time vehicle location data, cargo delivery demand data, traffic flow data, and environmental factor data. The collected data undergoes distributed cleaning and standardization, outlier detection algorithms are used to eliminate data anomalies, and data fusion technology is employed to integrate multi-source data into a unified vehicle-environment state feature vector. The vehicle-environment state feature vector includes vehicle state features, environmental features, and location features. The vehicle state features include vehicle load rate, driving speed, and remaining fuel. The environmental features include traffic congestion index and weather condition indicators. The location features include real-time latitude and longitude coordinates and road segment identification. Based on vehicle-environment state feature vectors, a vehicle scheduling decision model is constructed using a deep reinforcement learning framework. Candidate strategies for vehicle paths and scheduling sequences are generated through a policy network, and a multi-objective optimization algorithm is used to optimize and select the candidate strategies. The multi-objective optimization algorithm considers transportation cost, time efficiency and resource utilization simultaneously. The scheduling sequence is encoded into an executable instruction set and sent to the on-board control unit of the corresponding logistics vehicle through a wireless communication network to drive the vehicle to execute the scheduling task. Real-time monitoring of vehicle location data, task execution status, and changes in the external environment; collection of feedback data; and updating of parameters in the vehicle dispatching decision model based on time difference error to achieve closed-loop dynamic optimization.

2. The intelligent scheduling method for logistics vehicles based on big data as described in claim 1, characterized in that, The outlier detection algorithm is specifically as follows: A local outlier factor algorithm based on dynamic time warping is used to process time-series data streams, which include real-time vehicle location data and traffic flow data. The local outlier factor algorithm divides the data stream into continuous segments using a sliding window and uses a combination of k-nearest neighbor algorithm to calculate the local density deviation of each data point within a segment; The detection threshold is initialized based on the distribution characteristics of historical data and is adaptively updated through an exponentially weighted dynamic mechanism; For the identified outliers, a probabilistic interpolation method based on Gaussian mixture model is used to repair the data and maintain the integrity of the data sequence.

3. The intelligent scheduling method for logistics vehicles based on big data as described in claim 2, characterized in that, The data fusion technology includes: After eliminating data anomalies, the data is normalized to eliminate dimensional differences. The normalized data is divided into time-series dynamic feature data and static feature data; The time-series dynamic feature data includes real-time vehicle location data, traffic flow data, and dynamic environmental features derived from environmental factor data; the static feature data includes vehicle inherent attribute data and cargo fixed information. For time-series dynamic feature data, a multi-head self-attention network is input, and the weight coefficients of each feature in the fusion process are dynamically allocated by calculating the correlation scores between features. The multi-head self-attention network integrates a position encoder to retain the time-series attributes of the data. For static feature data, a fully connected neural network is used for feature transformation and fusion; The processed temporal dynamic features and static features are concatenated and compressed into a unified vehicle-environment state feature vector.

4. The intelligent scheduling method for logistics vehicles based on big data as described in claim 1, characterized in that, The vehicle dispatching decision model adopts a deep reinforcement learning-based framework, including a feature extraction network, a policy network, and a value network. The feature extraction network is used to extract spatiotemporal feature vectors from vehicle-environment state feature vectors. It is composed of a one-dimensional convolutional neural network and a long short-term memory network connected in series. The one-dimensional convolutional neural network is used to extract the local spatial patterns of the vehicle-environment state feature vectors, and the long short-term memory network is used to capture the temporal dependencies of the features. The policy network is used to generate vehicle scheduling policies, which include vehicle routes and scheduling sequences. The value network is used to evaluate the long-term benefits of the scheduling strategy; The multi-objective optimization algorithm employs an improved non-dominated sorting genetic algorithm, which maintains a Pareto optimal solution set in each generation of evolution and preserves the diversity of solutions by calculating the crowding degree of the solutions. Finally, it selects the scheduling sequence with the highest comprehensive utility from the solution set as the output.

5. The intelligent scheduling method for logistics vehicles based on big data as described in claim 4, characterized in that, The policy network adopts an encoder-decoder architecture, and its input layer receives spatiotemporal feature vectors extracted by the feature extraction network. The encoder consists of multiple fully connected layers, each followed by a batch normalization layer and a ReLU activation function. The decoder uses an attention-based recurrent neural network structure to gradually generate vehicle path sequences and scheduling sequences; The output layer of the policy network uses the Softmax function to output the probability distribution of vehicle path selection and the probability distribution of vehicle scheduling sequence arrangement at each decision time.

6. The intelligent scheduling method for logistics vehicles based on big data as described in claim 5, characterized in that, The input to the value network includes spatiotemporal feature vectors extracted by the feature extraction network and vehicle scheduling policies output by the policy network, and the output is a long-term value estimate scalar for the current state-action pair. The value network contains multiple fully connected layers to achieve a non-linear mapping from input features to value estimation; The value network is trained using a time difference learning algorithm to minimize the mean square error between the predicted value and the target value.

7. The intelligent scheduling method for logistics vehicles based on big data as described in claim 6, characterized in that, The operation process of the improved non-dominated sorting genetic algorithm includes: Initialize the population, perform non-dominated sorting of the individuals in the population, and calculate the crowding degree; Tournament selection is based on an individual's front-end ranking and congestion level; The offspring population is generated using simulated binary crossover and polynomial mutation operations. Among them, the crossover probability and mutation probability are adaptively adjusted based on the diversity index of the current generation of the population; After each generation of evolution, a local search operation is applied to the individuals in the Pareto optimal solution set, and the local search employs a perturbation operator related to the problem domain. The algorithm calculates the comprehensive utility value of each solution from the Pareto optimal solution set of the last generation population according to the preset preference rules or weight coefficients, and selects the scheduling sequence with the highest utility value as the output.

8. The intelligent scheduling method for logistics vehicles based on big data as described in claim 1, characterized in that, The process of encoding the scheduling sequence into an executable instruction set includes: Based on the preset scheduling operation specification library, the scheduling sequence is parsed into a series of atomic operation instructions, each atomic operation instruction corresponding to a basic action that a vehicle can perform; According to the standard communication protocol of the vehicle control unit, atomic operation instructions are serialized into structured data packets. The data packets are encapsulated in JSON format and appended with timestamps and vehicle identification identifiers. Data packets are compressed using a data compression algorithm, and forward error correction codes are added to ensure transmission reliability, resulting in the final executable instruction set.

9. The intelligent scheduling method for logistics vehicles based on big data as described in claim 8, characterized in that, The closed-loop dynamic optimization process is specifically as follows: After receiving and parsing the executable instruction set, the vehicle control unit drives the vehicle to perform the corresponding actions; At the same time, vehicle positioning data, engine status parameters and real-time traffic information are continuously collected by onboard sensors as feedback data. The feedback data is sent back to the central processing system. Based on the time difference error between the actual reward signal and the value network prediction value, the online learning process of the model is triggered. The reinforcement learning algorithm is used to incrementally update the parameters of the policy network and the value network to achieve closed-loop optimization of the scheduling strategy.

10. A big data-based intelligent dispatching system for logistics vehicles, characterized in that, The system is used to implement a big data-based intelligent scheduling method for logistics vehicles as described in any one of claims 1 to 9, comprising: The data acquisition module collects logistics vehicle scheduling data in real time from multiple heterogeneous data sources; The data preprocessing and fusion module cleans, standardizes, and detects outliers in the collected data, and performs multi-source data fusion to generate the vehicle-environment state feature vector. The scheduling decision calculation module runs the vehicle scheduling decision model and multi-objective optimization algorithm to generate scheduling sequences. The instruction encoding and communication module encodes the scheduling sequence into an executable instruction set and sends it out via a wireless communication network. The monitoring and closed-loop feedback module monitors the vehicle's execution status, collects feedback data, and drives the dynamic update of the scheduling decision model.

Images