An artificial intelligence-based logistics management system

Abstract

The present invention discloses an artificial intelligence-based logistics management system for improving the efficiency and intelligence level of logistics scheduling and resource allocation, which includes the following steps: By collecting various types of logistics data such as positioning, sensors, and operation logs, a unified data set is established, and a data structure containing spatio-temporal and transportation status information is formed through a modeling method. The system uses a causal analysis model to identify the key impact relationships between logistics events, and combines cross-modal feature extraction technology to generate representation vectors such as transportation trajectories, loads, and statuses. Through an adaptive strategy, causal information and multi-modal features are fused to generate decision-making bases, and the fusion ratio is dynamically adjusted according to actual operation feedback to achieve optimized transportation routes, distribution task planning, and reasonable resource allocation. This system has the ability of continuous learning and can maintain the accuracy and flexibility of scheduling decisions in an ever-updating data environment.

Description

Technical Field

[0001] The present invention relates to the technical field of logistics management, and particularly to a logistics management system based on artificial intelligence. Background Art

[0002] With the continuous development of artificial intelligence and information technology, modern logistics systems are facing unprecedented challenges in data complexity and decision-making real-time. Especially in the context of high-frequency flow of multi-source heterogeneous data, traditional logistics management systems have shown obvious limitations in scheduling response, path optimization, and resource allocation. The events in the logistics process have high spatio-temporal dynamics and multi-modal interactivity, and single information processing strategies and static rules are no longer sufficient to meet the intelligent decision-making needs of complex operation scenarios.

[0003] In the prior art, traditional logistics management solutions generally adopt rule-based scheduling systems and static optimization algorithms. These methods have the following main defects when dealing with rapidly changing transportation environments and multi-source data fusion processing:

[0004] 1. Lack of causal modeling ability: Existing methods mostly make scheduling judgments based on correlation analysis, making it difficult to identify the true causal relationships between logistics events, resulting in structural misjudgments in resource scheduling and path selection.

[0005] 2. Insufficient understanding of multi-modal data: Traditional systems are difficult to effectively process heterogeneous data from GPS, sensors, and operation logs simultaneously, relying on a large amount of manual annotation and lacking an automatic semantic extraction mechanism.

[0006] 3. The fusion mechanism is rigid and non-self-adaptive: Most methods adopt fixed weights or static fusion methods, and are unable to dynamically adjust the weight ratio between causal information and feature expression according to actual operation feedback.

[0007] Therefore, how to provide a logistics management system based on artificial intelligence is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a logistics management system based on artificial intelligence. By modeling the causal relationships between logistics events, automatically extracting and dynamically fusing cross-modal feature vectors, and combining a real-time feedback mechanism for joint decision generation, the present invention has the advantages of fast response speed, high scheduling accuracy, and strong self-adaptive ability.

[0009] A logistics management system based on artificial intelligence according to an embodiment of the present invention includes the following steps:

[0010] S1. A data acquisition module, used to collect sensors, And multi-source logistics data of operation logs, perform standardized preprocessing on the collected multi-source logistics data, and construct a logistics data set;

[0011] S2. A data fusion module, which is used to perform spatio-temporal alignment and fusion on the logistics data set, and generate fusion data based on a spatio-temporal dynamic logistics network modeling method, including logistics event features, transportation status features, resource usage features, and current network status data;

[0012] S3. A causal inference module, based on a deep causal inference algorithm, uses a multi-layer neural network to quantitatively model the causal relationship between each logistics event in the fusion data, and outputs causal inference parameters;

[0013] S4. A self-supervised multi-modal representation learning module, constructs a cross-modal self-supervised pre-training task, and automatically extracts feature vectors of the fusion data through a cross-modal feature interaction mechanism, including spatio-temporal trajectory features, vehicle load features, and transportation status features;

[0014] S5. A fusion decision module, adopts an adaptive dynamic decision weight adjustment strategy, adaptively integrates causal inference parameters and feature vectors according to real-time online feedback, and generates logistics scheduling and resource allocation decisions, including the optimal transportation path, dynamic distribution plan, and resource allocation plan.

[0015] Optionally, the S3 specifically includes:

[0016] S31. An event feature preprocessing unit, receives the fusion data, performs standardized processing on the original features of each logistics event in the fusion data, extracts the event occurrence time, geographical location, status information, and numerical indicators, and constitutes an event feature vector matrix ;

[0017] S32. A multi-layer neural network mapping unit, based on a deep causal inference algorithm, performs layer-by-layer non-linear mapping on the event feature vector matrix using a multi-layer neural network. Each hidden layer outputs an abstract feature representation through a combination of linear transformation and activation function. Let the non-linear mapping function be , and obtain an abstract feature matrix;

[0018] S33. A causal impact calculation unit, constructs a quantitative model of the causal relationship between events based on the abstract feature matrix, and calculates the causal impact weight of the logistics event on the logistics event through the following formula:

[0019] ;

[0020] Among them, is the logistics event on the logistics event The causal influence weight, is a non - linear mapping function composed of the abstract feature matrix of logistics event and logistics event ; is the weight matrix, is the bias vector, is the activation function;

[0021] S34, the causal inference parameter generation unit, quantifies the causal relationship between each logistics event according to the causal influence weight, and generates causal inference parameters , where the causal inference parameters include event influence weight, resource scheduling influence factor, and transportation path adjustment factor;

[0022] S35, the model parameter optimization unit, uses the back - propagation algorithm to dynamically adjust the weight matrix and the bias vector of the quantitative model of the causal relationship between logistics events, so as to ensure that the causal inference module stably and accurately outputs the above - mentioned causal inference parameters in different logistics scenarios.

[0023] Optionally, the S32 specifically includes:

[0024] S321, the input pre - processing unit, is used to receive the output event feature vector matrix , normalize the data in the event feature vector matrix to form a normalized feature matrix ;

[0025] S322, the layer - by - layer mapping unit, based on the deep causal inference algorithm, constructs a multi - layer neural network containing hidden layers. Each hidden layer performs an operation combining a linear transformation and an activation function, and performs layer - by - layer mapping on the normalized feature matrix . Define the initial input as . For the th layer , calculate using the following formula:

[0026] ;

[0027] Among them, represents the hidden representation of the th layer, represents the weight matrix of the th layer, represents the bias vector of the th layer, is the non - linear activation function, represents the hidden representation of the th layer, and the final hidden representation is obtained after layer - by - layer mapping ;

[0028] S323. Feature abstraction unit, with the non-linear mapping function set as , perform non-linear transformation on the final hidden representation output by the layer-by-layer mapping unit to generate an abstract feature matrix.

[0029] Optionally, the S4 specifically includes:

[0030] S41. Modal data construction unit, which is used to receive the generated fusion data and divide the fusion data into multiple modal subsets. The modal subsets include sensor modality, GPS modality, and log modality, corresponding to environmental perception data, trajectory positioning data, and operation scheduling data respectively. After division, a modal input set is formed;

[0031] S42. Modal feature encoding unit, based on the modal input set, constructs a corresponding feature encoding network for each modal subset. The feature encoding network includes several hidden layers, and each layer contains trainable weight parameters and bias parameters to extract its primary representation features. Let the th modal data be , and its encoded representation is:

[0032] ;

[0033] Among them, represents the encoding function of modality , and is the modal feature vector;

[0034] S43. Cross-modal contrast learning unit, which is used to construct a cross-modal self-supervised learning task based on the generated modal feature vectors , and realize the aggregation of feature vector spaces of the same logistics event under different modalities and the separation of feature vector spaces of different logistics events by constructing positive sample pairs and negative sample pairs and introducing a contrast loss function, and generate optimized modal feature vectors;

[0035] S44. Feature interaction and fusion unit, which is used to optimize the modal feature vectors for feature alignment and fusion, calculate the correlation scores between modalities using a cross-attention mechanism, and generate a feature vector in combination with a fusion strategy. The feature vector includes spatio-temporal trajectory features, vehicle load features, and transportation status features.

[0036] Optionally, the S43 specifically includes:

[0037] S431. Sample pair construction unit, which is used to receive the modal feature vectors , construct positive sample pairs for the same logistics event among different modalities, and construct negative sample pairs among different logistics events to form a set of sample pairs , where each represents a pair of feature vectors from different modalities. Positive sample pairs correspond to the same logistics event, and negative sample pairs correspond to different logistics events;

[0038] S432. Contrast loss function definition unit, used to construct a cross-modal self-supervised contrast learning objective based on the set of sample pairs and define the contrast loss function as follows:

[0039] ;

[0040] Among them, is the contrast loss function, is the similarity function between feature vectors, is the temperature coefficient, is the set that forms negative sample pairs with the anchor feature vector , is the total number of sample pairs, is the interference feature vector that forms a negative sample pair with the anchor feature vector , is the matching feature vector that forms a positive sample pair with the anchor feature vector ;

[0041] S433. Feature representation optimization unit, used to adjust the trainable weight parameters and bias parameters of the feature encoding network according to the defined contrast loss function, so that the feature vectors in the positive sample pairs converge in the feature vector space, and the feature vectors in the negative sample pairs are separated in the feature vector space, forming a distinguishable cross-modal embedding representation structure;

[0042] S434. Optimized representation output unit, used to output the optimized modal feature vectors generated after updating the distinguishable cross-modal embedding representation structure to the feature interaction and fusion unit.

[0043] Optionally, the S5 specifically includes:

[0044] S51. Feature integration preparation unit, used to receive the causal inference parameter and the feature vector , and perform format conversion and dimension unification on the causal inference parameter and the feature vector to construct an input alignment vector set ;

[0045] S52. Adaptive weight generation unit, used to generate weights based on the input alignment vector set Establish a fusion strategy to perform weighted integration on the causal inference parameters and feature vectors, and construct a joint representation vector , and the following formula is used for fusion:

[0046] ;

[0047] where, is an adjustable fusion weight coefficient used to control the contribution ratio of the causal inference parameters and feature vectors;

[0048] S53. A feedback regulation unit constructs a logistics execution system and generates real-time feedback information. The real-time feedback information includes the task completion time, resource usage status, and execution deviation index. A weight update function is constructed based on the real-time feedback information, and the adjustable fusion weight coefficient is dynamically adjusted to adapt to the current operating state of the logistics system;

[0049] S54. A decision generation unit is used to receive the joint representation vector , and perform policy inference by combining the current network state data to generate logistics scheduling and resource allocation decisions including the optimal transportation path, dynamic distribution plan, and resource allocation scheme.

[0050] Optionally, the S52 specifically includes:

[0051] S521. A feature similarity calculation unit is used to receive the output set of input alignment vectors , and calculate the feature representation similarity of the input alignment vector set. The feature representation similarity uses the normalized cosine similarity function:

[0052] ;

[0053] where, represents the feature representation similarity between the causal inference parameters and the fused feature vectors, represents the vector dot product operation, represents the L2 norm of the vector;

[0054] S522. A weight coefficient initialization unit is used to determine the adjustable fusion weight coefficient based on the calculated feature representation similarity , and perform the calculation through a monotonic mapping function . The is a monotonic mapping function defined in the interval , satisfying , and is used to control the initial fusion ratio between the causal information and the fused features;

[0055] S523. A fusion representation generation unit is used to receive the adjustable fusion weight coefficient , the input alignment vector set , and construct a joint representation vector .

[0056] The beneficial effects of the present invention are as follows:

[0057] (1) By constructing a logistics management system that integrates causal reasoning, self-supervised multimodal representation learning, and an adaptive fusion strategy, the present invention realizes the modeling of causal relationships between logistics events, the efficient expression of multi-source heterogeneous data, and the intelligent decision-making of scheduling strategies, enabling the system to identify key logistics influencing factors and dynamically generate transportation routes, distribution plans, and resource allocation schemes, thereby effectively solving the problems of lagging response and low scheduling accuracy of traditional logistics systems in dynamic environments.

[0058] (2) By introducing a feedback perception mechanism and an online parameter update strategy, the present invention constructs a weight regulation model driven by operation feedback, enabling the fusion decision result to adapt to the current logistics network state in real time, improving the scheduling flexibility and stability of the system in different transportation scenarios, and having the ability of continuous learning and self-optimization, effectively improving the resource utilization efficiency and scheduling response speed.

[0059] (3) By adaptively constructing a joint representation vector from causal reasoning parameters and multimodal features and forming a closed-loop decision-making system in combination with online feedback, the system can automatically adapt to the logistics operation characteristics in different scenarios without relying on a large number of manual rules or annotations, providing more intelligent, accurate, and dynamic logistics management capabilities, and significantly enhancing the system decision-making ability and operation efficiency in complex environments. Description of the Drawings

[0060] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used in conjunction with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0061] Figure 1 is a schematic diagram of the overall architecture of a logistics management system based on artificial intelligence proposed by the present invention. Detailed Embodiments

[0062] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0063] Refer to Figure 1 , a logistics management system based on artificial intelligence, including the following steps:

[0064] S1. A data acquisition module for collecting sensors, And multi-source logistics data of operation logs, perform standardized preprocessing on the collected multi-source logistics data, and construct a logistics data set;

[0065] In this embodiment, by deploying a data collection module, multi-source logistics data from sensors, GPS, and operation logs is collected in real time. The multi-source logistics data includes environmental perception data, trajectory positioning data, and operation scheduling data. Among them, the environmental perception data includes the temperature inside the carriage collected by the on-vehicle temperature and humidity sensor, the impact intensity recorded by the vibration sensor, the detection of the carriage opening state by the light sensor, the detection of the gas concentration inside the vehicle by the air quality sensor, and the instant speed recorded by the vehicle speed sensor. The trajectory positioning data includes GPS real-time positioning coordinates, a set of driving trajectory points, docking event records, driving state identifiers, and route deviation warning data. The operation scheduling data includes transportation task orders, vehicle allocation records, scheduling plan timetables, distribution priorities, and real-time task statuses; and unified standardized preprocessing is performed, including time alignment, outlier removal, and feature normalization, so as to construct a logistics data set with a unified structure and consistent semantics. This module, as the information entry point of the system, ensures the input quality of causal modeling and multi-modal learning, enables the subsequent analysis model to accurately extract spatio-temporal, load, and operation characteristics on the basis of high-quality data, realizes a comprehensive perception of key behaviors and resource states in the logistics scenario, and improves the adaptability of the system in complex environments and the reliability of the decision-making basis.

[0066] S2. A data fusion module, which is used to perform spatio-temporal alignment and fusion on the logistics data set, and generate fusion data based on a spatio-temporal dynamic logistics network modeling method, including logistics event characteristics, transportation state characteristics, resource usage characteristics, and current network state data;

[0067] In this embodiment, a spatio-temporal dynamic logistics network modeling method is adopted to perform spatio-temporal alignment and fusion processing on the collected logistics data set, uniformly encode heterogeneous information such as GPS trajectories, sensor states, and operation logs according to timestamps and geographical locations, and model the relationships between nodes, edges, and resource flows in the network graph structure. Finally, fusion data is generated, including logistics event characteristics, transportation state characteristics, resource usage characteristics, and current network state data. This processing method realizes the organic integration of multi-source logistics information in a unified space and time dimension, provides structured and high-density input data support for subsequent causal reasoning and scheduling decisions, and effectively enhances the system's understanding ability of the overall logistics state and response ability to real-time changes.

[0068] S3. A causal reasoning module, based on a deep causal reasoning algorithm, uses a multi-layer neural network to quantitatively model the causal relationships between various logistics events in the fusion data, and outputs causal reasoning parameters;

[0069] S4. Self-supervised multimodal representation learning module, which constructs a cross-modal self-supervised pre-training task and automatically extracts the feature vectors of the fused data through a cross-modal feature interaction mechanism. The feature vectors include spatio-temporal trajectory features, vehicle load features, and transportation status features;

[0070] S5. Fusion decision-making module, which adopts an adaptive dynamic decision weight adjustment strategy to adaptively integrate causal inference parameters and feature vectors according to real-time online feedback, and generates logistics scheduling and resource allocation decisions, including the optimal transportation path, dynamic distribution plan, and resource allocation plan.

[0071] In this embodiment, the specific content of S3 is as follows:

[0072] S31. Event feature preprocessing unit, which receives the fused data, standardizes the original features of each logistics event in the fused data, and extracts the event occurrence time, geographical location, status information, and numerical indicators to form an event feature vector matrix ;

[0073] S32. Multilayer neural network mapping unit, which, based on the deep causal inference algorithm, performs layer-by-layer non-linear mapping on the event feature vector matrix using a multilayer neural network. Each hidden layer outputs an abstract feature representation through a combination of linear transformation and activation function. Let the non-linear mapping function be , and an abstract feature matrix is obtained;

[0074] S33. Causal impact calculation unit, which constructs a quantitative model of causal relationship between events based on the abstract feature matrix, and calculates the causal impact weight of logistics event on logistics event through the following formula:

[0075] ;

[0076] where, is the causal impact weight of logistics event on logistics event , is a non-linear mapping function composed of the abstract feature matrices of logistics event and logistics event , is the weight matrix, is the bias vector, is the activation function;

[0077] This formula is used to quantify the intensity of causal relationships between different logistics events, reflecting the direct force between events. Its principle is based on the linear mapping and non-linear activation mechanisms in deep neural networks, which can model the relationships of complex high-dimensional abstract features, enabling the system to not only identify the co-occurrence between events but also explore their potential causal driving mechanisms, thus providing a structured causal basis for subsequent scheduling optimization.

[0078] S34. Causal inference parameter generation unit, which quantifies the causal relationships between each logistics event according to the causal influence weights and generates causal inference parameters , where the causal inference parameters include event influence weights, resource scheduling influence factors, and transportation path adjustment factors. Among them, the event influence weight reflects the intensity of the causal effect of a certain type of logistics event on other event nodes in the scheduling, the resource scheduling influence factor is used to describe the matching and dependence relationships of tasks on resources such as vehicles and manpower, and the path adjustment factor indicates the degree of perturbation of external change factors to the feasibility of the current path plan, which is used to assist in dynamic path reconstruction;

[0079] S35. Model parameter optimization unit, which uses the backpropagation algorithm to dynamically adjust the weight matrix and bias vector of the quantitative model of the causal relationship between logistics events, ensuring that the causal inference module stably and accurately outputs the above-mentioned causal inference parameters under different logistics scenarios.

[0080] In this embodiment, by constructing an event modeling process based on a deep causal inference network, first, the logistics event features in the fusion data are standardized, information such as time, location, and status is extracted and formed into a feature vector matrix, and then the feature vectors are non-linearly mapped layer by layer using a multi-layer neural network to generate an abstract feature representation, construct a quantitative model of the causal relationship between events, calculate the causal influence weights between each logistics event, and further generate causal inference parameters for logistics scheduling and path planning, including event influence degree, scheduling factor, and path adjustment amount. By introducing the backpropagation mechanism to continuously optimize the model parameters, the system can dynamically adapt to the changes in event logic under different scenarios, effectively improving the modeling accuracy of complex logistics relationships and the reliability of causal decision-making.

[0081] In this embodiment, the S32 specifically includes:

[0082] S321. Input preprocessing unit, which is used to receive the output event feature vector matrix , and perform normalization processing on the data in the event feature vector matrix to form a normalized feature matrix ;

[0083] S322. Layer-by-layer mapping unit, which constructs a network containing A multi-layer neural network with hidden layers, each hidden layer performs a linear transformation and activation function combination operation, and normalizes the feature matrix Perform layer-by-layer mapping and define the initial input as , for the layer , calculated using the following formula:

[0084] ;

[0085] in, Indicates The hidden representation of the layer, Indicates The weight matrix of the layer, Indicates The bias vector of the layer, is a nonlinear activation function, Indicates The hidden representation of the layer is mapped layer by layer to obtain the final hidden representation ;

[0086] This formula is used to perform nonlinear mapping of event feature vectors layer by layer in a multi-layer neural network, and to achieve gradual abstraction and extraction of features through a combination of weight transformation and activation function. The principle is to superimpose linear transformations and nonlinear activation operations of multiple hidden layers to enable the model to have stronger feature expression capabilities, thereby mining the implicit causal structure and high-dimensional feature relationships in logistics events. This process provides a high-quality feature input basis for subsequent causal reasoning parameter calculations.

[0087] S323, feature abstraction unit, assuming that the nonlinear mapping function is , the final hidden representation of the output of the layer-by-layer mapping unit Perform nonlinear transformation to generate an abstract feature matrix.

[0088] This implementation method constructs a multi-layer neural network model based on a deep causal reasoning architecture, performs a nonlinear mapping operation of layer-by-layer linear transformation and activation function combination on the normalized event feature vector matrix, gradually extracts hidden semantic information at different levels, and finally generates a high-level abstract feature matrix for causal relationship modeling through an abstract feature transformation function. This method effectively retains the deep feature representation of logistics events in different spatial, temporal and resource states, enables the model to accurately capture the potential causal structure between logistics events, improves the ability to understand event-driven mechanisms, and provides a reliable feature foundation for subsequent causal impact quantification and intelligent decision-making.

[0089] In this implementation manner, S4 specifically includes:

[0090] S41. Modal data construction unit, which is used to receive the generated fusion data and divide the fusion data into multiple modal subsets. The modal subsets include sensor modality, GPS modality, and log modality, corresponding to environmental perception data, trajectory positioning data, and operation scheduling data respectively. After division, a modal input set is formed.

[0091] S42. Modal feature encoding unit. Based on the modal input set, a corresponding feature encoding network is constructed for each modal subset. The feature encoding network includes several hidden layers, and each layer contains trainable weight parameters and bias parameters to extract its primary representation features. Let the th modal data be , and its encoded representation is:

[0092] ;

[0093] Among them, represents the encoding function of modality , and is the modal feature vector.

[0094] S43. Cross-modal contrastive learning unit, which is used to construct a cross-modal self-supervised learning task based on the generated modal feature vectors . By constructing positive sample pairs and negative sample pairs and introducing a contrastive loss function, it realizes the aggregation of feature vector spaces of the same logistics event under different modalities and the separation of feature vector spaces of different logistics events, and generates optimized modal feature vectors.

[0095] S44. Feature interaction and fusion unit, which is used to optimize the modal feature vectors for feature alignment and fusion. It calculates the correlation scores between modalities using a cross-attention mechanism and generates a feature vector in combination with a fusion strategy. The feature vector includes spatio-temporal trajectory features, vehicle load features, and transportation status features.

[0096] In this embodiment, by dividing the fusion data into sensor modality, GPS modality, and log modality, corresponding to environmental perception, trajectory positioning, and operation scheduling data respectively, a modal input set is constructed. And a feature encoding network is designed for each modal subset to extract the primary feature representation. By constructing positive and negative sample pairs through a cross-modal contrastive learning mechanism and introducing a contrastive loss function, it realizes the representation aggregation of the same logistics event under different modalities and the distinction of different events, generates optimized modal features, and uses a cross-attention mechanism to fuse the optimized feature vectors to generate a unified feature representation including spatio-temporal trajectory, vehicle load, and transportation status and other information. This method effectively solves the problems of weak processing ability of traditional methods for multi-source heterogeneous data, insufficient feature expression, and difficult semantic alignment, and improves the adaptability of the logistics scheduling system to complex scenarios and the multi-modal data understanding ability.

[0097] In this embodiment, step S43 specifically includes:

[0098] S431, a sample pair construction unit, which is configured to receive modal feature vectors , construct positive sample pairs for the same logistics event among different modalities, construct negative sample pairs among different logistics events, and form a sample pair set , where each represents a pair of feature vectors from different modalities. The positive sample pairs correspond to the same logistics event, and the negative sample pairs correspond to different logistics events;

[0099] S432, a contrastive loss function definition unit, which is configured to construct a cross-modal self-supervised contrastive learning objective based on the sample pair set , and define the contrastive loss function as follows:

[0100] ;

[0101] where, is the contrastive loss function, is the similarity function between feature vectors, is the temperature coefficient, is the set that forms negative sample pairs with the anchor feature vector , is the total number of sample pairs, is that the interference feature vector and the anchor feature vector form a negative sample pair, is that the matching feature vector and the anchor feature vector form a positive sample pair;

[0102] The principle of this formula is to maximize the similarity between positive sample pairs while minimizing the similarity between the anchor and all negative sample pairs, thereby guiding the model to align the multi-modal features of the same logistics event in the shared embedding space and distinguish the features of different events. This mechanism can effectively improve the consistency and discriminability of the representations between modalities, providing a more stable feature basis for subsequent fusion and scheduling.

[0103] S433, a feature representation optimization unit, which is configured to adjust the trainable weight parameters and bias parameters of the feature encoding network according to the defined contrastive loss function, so that the feature vectors in the positive sample pairs converge in the feature vector space, and the feature vectors in the negative sample pairs are separated in the feature vector space, forming a distinguishable cross-modal embedding representation structure;

[0104] S434, an optimized representation output unit, which is configured to output the optimized modal feature vectors generated after updating the distinguishable cross-modal embedding representation structure to the feature interaction and fusion unit.

[0105] In this embodiment, by constructing a set of positive and negative sample pairs and introducing a cross-modal contrastive loss function, consistent modeling and enhanced discriminability of the feature representations of logistics events in different modalities are achieved. Specifically, the system receives the feature vectors generated by the multi-modal encoding network, automatically constructs positive sample pairs of the same event and negative sample pairs of different events, and optimally adjusts the similarity between the anchor feature and the positive and negative samples through the contrastive loss function. During the training process, the weight parameters and bias parameters in the encoding network are dynamically updated, enabling the model to converge similar event representations and distance different event representations in a unified feature space, and finally generating optimized modal feature vectors. This mechanism significantly improves the discriminability of feature representations and the cross-modal alignment effect, providing a more stable and accurate input expression for subsequent scheduling fusion and path decision-making.

[0106] In this embodiment, S5 specifically includes:

[0107] S51. A feature integration preparation unit for receiving causal inference parameters and feature vectors , and performing format conversion and dimension unification on the causal inference parameters and feature vectors to construct a set of input alignment vectors ;

[0108] S52. An adaptive weight generation unit for establishing a fusion strategy based on the set of input alignment vectors to perform weighted integration on the causal inference parameters and feature vectors, and construct a joint representation vector , and the following formula is used for fusion:

[0109] ;

[0110] where is an adjustable fusion weight coefficient for controlling the contribution ratio of the causal inference parameters and feature vectors;

[0111] This formula realizes the linear weighted combination of causal information and fused features. By adjusting the value of , the importance of causal relationships and perceptual features in the final decision can be dynamically balanced, ensuring that the model has flexible adaptation and representation capabilities in different scenarios. The principle of this mechanism is based on the idea of linear superposition of features, organically integrating the structural explanatory information provided by causal inference and data-driven multi-modal features, thereby improving the expression integrity and discriminability of the joint features in complex logistics scheduling tasks.

[0112] S53. A feedback control unit constructs a logistics execution system and generates real-time feedback information, where the real-time feedback information includes task completion time, resource usage status, and execution deviation metrics. A weight update function is constructed based on the real-time feedback information, and the adjustable fusion weight coefficient is dynamically adjusted to adapt to the current operating state of the logistics system;

[0113] S54. A decision generation unit is used to receive the joint representation vector and perform policy inference in combination with the current network status data to generate logistics scheduling and resource allocation decisions including the optimal transportation path, dynamic distribution plan, and resource allocation plan.

[0114] In this embodiment, by receiving the causal inference parameters and the fusion feature vectors, an input alignment vector set is constructed after unifying the formats, and based on this set, the information integration of the two is realized through the adjustable fusion weight coefficient to generate a joint representation vector for scheduling decisions. The system constructs a feedback control mechanism to collect real-time feedback information such as task completion time, resource usage status, and execution deviation, and dynamically adjusts the fusion weight to adapt to the changes in the operating state of the logistics system. The system infers and generates the optimal transportation path, dynamic distribution plan, and resource allocation plan based on the joint representation vector and the current network status data, effectively improving the response speed, adaptability, and execution accuracy of the decision-making, and realizing the intelligent logistics scheduling optimization supported by the joint support of causal drive and data features.

[0115] In this embodiment, the S52 specifically includes:

[0116] S521. A feature similarity calculation unit is used to receive the output input alignment vector set and calculate the feature representation similarity of the input alignment vector set. The feature representation similarity uses the normalized cosine similarity function:

[0117] ;

[0118] where represents the feature representation similarity between the causal inference parameters and the fusion feature vectors, represents the vector dot product operation, represents the L2 norm of the vector;

[0119] This formula is based on the principle of normalized cosine similarity, measures the degree of consistency of the directions of two vectors in the representation space, and the output value ranges from , the closer the value is to 1, the more consistent the directions of the two vectors are, and the stronger the semantic relevance. Through this calculation, the correlation between causal information and fused features can be effectively characterized, which is used to drive the adaptive initialization of subsequent fusion weights, and helps to improve the ability of the fused representation to retain key features and the pertinence of decision-making expression.

[0120] S522. A weight coefficient initialization unit, configured to determine adjustable fusion weight coefficients based on the calculated feature representation similarity to determine adjustable fusion weight coefficients , through a monotonic mapping function for calculation, where the is a monotonic mapping function defined in the interval and satisfies , and is used to control the initial fusion ratio between causal information and fused features;

[0121] S523. A fused representation generation unit, configured to receive the adjustable fusion weight coefficients , the input alignment vector set , and construct a joint representation vector .

[0122] In this embodiment, the feature similarity calculation unit calculates the normalized cosine similarity between the causal inference parameter vector and the fused feature vector to obtain the correlation measure between them in the feature space; then, the weight coefficient initialization unit dynamically generates the fusion weight coefficients through the monotonic mapping function according to the similarity value to achieve the ratio control of causal information and fused features. The fused representation generation unit performs a weighted combination of the above weight coefficients and the input vectors to construct a joint representation vector for subsequent decision-making inference. This method realizes a weight initialization mechanism driven by feature similarity, enabling the system to automatically determine the fusion ratio of causal information and multi-modal features under unsupervised conditions, improving the adaptability and expression accuracy of the fusion strategy, and providing a more reasonable decision-making basis for subsequent intelligent scheduling and resource allocation. Embodiment

[0123] To verify the feasibility of the present invention in implementation, the present invention is applied to a large regional logistics center, which is responsible for the express sorting and main line transportation scheduling of multiple cities in the Yangtze River Delta region. This logistics center processes more than 200,000 packages per day, covering more than 120 transportation routes and more than 300 transportation vehicles. The logistics data sources are extensive, including multiple data sources such as GPS trajectories, warehouse status, order scheduling records, and vehicle load information. The traditional scheduling system mainly relies on fixed routes, manual experience judgment, and static allocation rules, and it is difficult to adapt to the operating environment with large order fluctuations and frequent changes in regional distribution loads. The scheduling accuracy is not high, the path planning lags behind, and the resource utilization efficiency significantly decreases.

[0124] The logistics center has deployed the system proposed by the present invention in the scheduling platform, and has focused on testing the adaptive weight generation unit. Based on historical causal reasoning parameters and real-time fusion feature vectors, the system automatically generates initial fusion weights through feature similarity calculation, and constructs a joint representation vector for subsequent scheduling optimization and path planning. During the actual operation process, the system automatically collects and cleans nearly 60GB of multimodal logistics data every day, including vehicle trajectories, traffic congestion data, real-time parcel distribution density, etc., and uniformly performs format conversion and feature alignment.

[0125] For example, in a trunk transportation task, the system detected that due to a temporary large-scale promotional event in a certain area, the inbound volume of the sorting station increased sharply, and there was serious congestion on the highway along the original planned path. The traditional system failed to respond in time, resulting in a delay of nearly 2 hours for some transportation tasks. After receiving abnormal data from a certain sorting station, the system integrated the event impact factor of "the promotional event increases the pressure on path resources" identified by the causal reasoning module, combined with the latest traffic conditions and vehicle load information, dynamically adjusted the original scheduling plan, generated a new joint representation, and adaptively adjusted the fusion coefficient through an online feedback mechanism. Finally, it output the optimal path plan, significantly alleviating the transportation pressure.

[0126] In the experimental stage, the system ran for 14 days respectively during peak and off-peak periods. The traditional system and the system of the present invention were set to run in parallel for testing, and simulation experiments on the operating performance of the system were conducted by comparing indicators such as scheduling response time, path replanning frequency, resource utilization rate, and order delivery timeliness rate. The experiment used preset multimodal logistics data, including simulated GPS trajectory points, sensor status data, and task scheduling records. The data types were the same as those collected by the data acquisition module of the S1 module in the specification. After preprocessing, they were uniformly input into the fusion system. After the system completed reasoning and fusion, it output the scheduling results, and statistics were made on indicators such as scheduling response time, path adjustment frequency, resource allocation rate, and task completion time to form the comparison data in Table 1. The experimental data are as follows:

[0127] Table 1 Comparison of the effects of the traditional logistics scheduling system and the system of the present invention (14-day experimental period)

[0128] Index Traditional scheduling system System of the present invention Average scheduling response time (seconds) 135 47 Path replanning frequency (times / day) 1.2 3.8 Vehicle resource utilization rate (%) 71.4 89.6 Order delivery timeliness rate (%) 85.2 96.7 Average distance saved per single path (km) 0 11.3 Real-time exception handling success rate (%) 63.5 91.8

[0129] As can be seen from the table, the system of the present invention shortened the average scheduling response time from 135 seconds to 47 seconds, and the scheduling response speed increased by more than 65%. The system greatly enhanced the path replanning ability through a dynamic fusion decision-making strategy and was able to adjust more timely to emergencies. The resource utilization rate increased from 71.4% to 89.6%, reflecting that the system performed more efficiently in vehicle allocation, path design, and transportation batch integration. At the same time, the on-time rate of order delivery increased from 85.2% to 96.7%, indicating that the system had significant advantages in service quality guarantee.

[0130] Further analysis of the correlation between "fusion weight change" and system output shows that during the peak promotion period, the average similarity value calculated by the system is 0.43, and the weight after mapping is stable at 0.64, indicating that causal information has a significant impact on decision-making; during non-peak hours, the similarity value increases to 0.76, and the weight drops to 0.42, and the system relies more on fusion feature expression for scheduling. This strategy of adaptively adjusting decision weights based on similarity effectively avoids the single dependence of causal drive or feature drive, and improves the overall scheduling robustness of the system.

[0131] In addition, in response to the historical records of frequent peaks in a certain type of night orders but a high delivery failure rate, the system conducted similarity clustering modeling on hundreds of historical tasks and found that there was a key causal path between the "vehicle's nighttime remaining load rate" and the "order accumulation density". The system automatically injected it as a parameter into the fusion vector, and ultimately reduced the delayed delivery rate of this type of orders from 8.7% to 2.3%.

[0132] In summary, this embodiment shows that the adaptive fusion weight generation mechanism based on feature similarity driven proposed in the present invention can automatically adjust the information fusion strategy according to the correlation between causal parameters and multimodal features, and combined with the feedback optimization mechanism, it can effectively improve the logistics scheduling efficiency, resource allocation rationality and path reconstruction capabilities in complex scenarios, and verify the practicality and advancement of the present invention in a real logistics operation environment.

[0133] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A logistics management system based on artificial intelligence, characterized in that: The steps include: S1, data acquisition module, used to collect sensor, and multi-source logistics data of operation logs, perform standardized preprocessing on the collected multi-source logistics data, and construct a logistics data set; S2, data fusion module, used to align and fuse logistics data sets in time and space, and generate fused data based on the spatiotemporal dynamic logistics network modeling method, including logistics event characteristics, transportation status characteristics, resource usage characteristics and current network status data; S3, causal reasoning module, based on deep causal reasoning algorithm, uses multi-layer neural network to quantitatively model the causal relationship between logistics events in the fusion data and output causal reasoning parameters; S4, self-supervised multimodal representation learning module, constructs cross-modal self-supervised pre-training tasks, and automatically extracts feature vectors of fused data through cross-modal feature interaction mechanisms, including spatiotemporal trajectory features, vehicle load features, and transportation status features; S5, the fusion decision module, adopts an adaptive dynamic decision weight adjustment strategy, adaptively integrates causal reasoning parameters and feature vectors based on real-time online feedback, and generates logistics scheduling and resource allocation decisions, including optimal transportation routes, dynamic distribution plans, and resource allocation plans; The S4 specifically includes: S41, a modality data construction unit, used to receive the generated fusion data, and divide the fusion data into a plurality of modality subsets, wherein the modality subsets include sensor modality, GPS modality and log modality, which correspond to environmental perception data, trajectory positioning data and operation scheduling data, respectively, to form a modality input set after division; S42, a modal feature encoding unit, based on the modal input set, constructs a corresponding feature encoding network for each modal subset, the feature encoding network includes several hidden layers, each layer contains trainable weight parameters and bias parameters, extracts its primary representation features, and sets The modal data is , which is coded as: ； in, Representing modality The encoding function, is the modal eigenvector; S43, cross-modal contrastive learning unit, for generating modal feature vectors based on Construct a cross-modal self-supervised learning task. By constructing positive and negative sample pairs and introducing a contrast loss function, the feature vector space aggregation of the same logistics event in different modes and the feature vector space separation of different logistics events are achieved to generate an optimized modal feature vector. S44, feature interaction fusion unit, is used to optimize the modal feature vector for feature alignment and fusion, use the cross-attention mechanism to calculate the correlation score between the modalities, and combine the fusion strategy to generate the feature vector , the feature vector Including spatiotemporal trajectory characteristics, vehicle load characteristics and transportation status characteristics.

2. According to claim 1, a logistics management system based on artificial intelligence is characterized in that: The S3 specifically includes: S31, event feature preprocessing unit, receives fused data, standardizes the original features of each logistics event in the fused data, extracts the event occurrence time, geographic location, status information and numerical indicators, and forms an event feature vector matrix ; S32, multi-layer neural network mapping unit, based on deep causal reasoning algorithm, event feature vector matrix A multi-layer neural network is used for layer-by-layer nonlinear mapping. Each hidden layer outputs abstract feature representation through a combination of linear transformation and activation function. The nonlinear mapping function is assumed to be , get the abstract feature matrix; S33, causal impact calculation unit, builds a quantitative model of causal relationships between events based on the abstract feature matrix, and calculates logistics events using the following formula Logistics events The causal influence weight of is: ； in, For logistics events Logistics events The causal influence weight of Logistics incident Logistics events The nonlinear mapping function composed of the abstract feature matrix, is the weight matrix, is the bias vector, is the activation function; S34, causal reasoning parameter generation unit, quantifies the causal relationship between each logistics event according to the causal influence weight, and generates causal reasoning parameters ,The causal reasoning parameters include event impact weights, resource scheduling influencing factors, and transportation path adjustment factors; S35, model parameter optimization unit, uses the back propagation algorithm to optimize the weight matrix of the quantitative model of causal relationship between logistics events With the bias vector Dynamic adjustments are made to ensure that the causal reasoning module can stably and accurately output the above causal reasoning parameters in different logistics scenarios.

3. The artificial intelligence-based logistics management system according to claim 2 is characterized in that: The S32 specifically includes: S321, input preprocessing unit, used to receive the output event feature vector matrix , for the event feature vector matrix The data in are normalized to form a normalized feature matrix ; S322, layer-by-layer mapping unit, based on deep causal reasoning algorithm, builds A multi-layer neural network with hidden layers, each hidden layer performs a linear transformation and activation function combination operation, and normalizes the feature matrix Perform layer-by-layer mapping and define the initial input as , for the layer , calculated using the following formula: ； in, Indicates The hidden representation of the layer, Indicates The weight matrix of the layer, Indicates The bias vector of the layer, is a nonlinear activation function, Indicates The hidden representation of the layer is mapped layer by layer to obtain the final hidden representation ; S323, feature abstraction unit, assuming that the nonlinear mapping function is , the final hidden representation of the output of the layer-by-layer mapping unit Perform nonlinear transformation to generate an abstract feature matrix.

4. The artificial intelligence-based logistics management system according to claim 3 is characterized in that: The S43 specifically includes: S431, sample pair construction unit, used to receive modal feature vector , construct positive sample pairs for the same logistics event between different modes, and construct negative sample pairs between different logistics events to form a sample pair set , where each Represents pairs of feature vectors from different modalities, where positive sample pairs correspond to the same logistics event and negative sample pairs correspond to different logistics events; S432, contrast loss function definition unit, used to define the loss function based on the sample pair set Construct a cross-modal self-supervised contrastive learning objective and define the contrastive loss function as follows: ； in, is the contrast loss function, is the similarity function between feature vectors, is the temperature coefficient, is the anchor feature vector Constitute the set of negative sample pairs, is the total number of sample pairs, is the interference feature vector and the anchor feature vector Constitute a negative sample pair, To match the feature vector with the anchor feature vector Constitute a positive sample pair; S433, a feature representation optimization unit, used to adjust the trainable weight parameters and bias parameters of the feature encoding network according to the defined contrast loss function, so that the feature vectors in the positive sample pairs converge in distance in the feature vector space, and the feature vectors in the negative sample pairs separate in distance in the feature vector space, so as to form a distinguishable cross-modal embedding representation structure; S434, optimized representation output unit, used to embed the distinguishable cross-modal representation structure to generate the optimized modal feature vector Output to the feature interaction fusion unit.

5. The artificial intelligence-based logistics management system according to claim 4, characterized in that: The S5 specifically includes: S51, feature integration preparation unit, used to receive causal reasoning parameters and the eigenvector , and the causal inference parameters and the eigenvector Perform format conversion and dimension unification to construct an input alignment vector set ; S52, an adaptive weight generating unit, for aligning a vector set based on an input Establish a fusion strategy to perform weighted integration of causal inference parameters and feature vectors to construct a joint representation vector , the following formula is used for fusion: ； in, is an adjustable fusion weight coefficient, which is used to control the contribution ratio of causal inference parameters and feature vectors; S53, feedback control unit, builds a logistics execution system and generates real-time feedback information, which includes task completion time, resource usage status and execution deviation indicators, builds a weight update function based on the real-time feedback information, and dynamically adjusts the adjustable fusion weight coefficient , to adapt to the current operation status of the logistics system; S54, decision generation unit, used to receive the joint representation vector , and combines the current network status data for strategic reasoning to generate logistics scheduling and resource allocation decisions including optimal transportation routes, dynamic distribution plans and resource allocation plans.

6. The artificial intelligence-based logistics management system according to claim 5, characterized in that: The S52 specifically includes: S521, feature similarity calculation unit, used to receive the output input alignment vector set , and calculate the feature representation similarity of the input alignment vector set, the feature representation similarity uses the normalized cosine similarity function: ； in, represents the feature representation similarity between the causal inference parameters and the fused feature vector, represents the vector dot product operation, Represents the L2 norm of the vector; S522, weight coefficient initialization unit, used to represent similarity based on the calculated feature Determine the adjustable fusion weight coefficient , through the monotone mapping function Calculate, the For the interval The monotone mapping function defined in , used to control the initial fusion ratio between causal information and fusion features; S523, fusion representation generation unit, used to receive adjustable fusion weight coefficient , input alignment vector set , and construct the joint representation vector .

Images