Intelligent logistics full-link management and control system based on multi-modal AI collaborative decision

Abstract

The invention discloses an intelligent logistics full-link management and control system based on multi-modal AI collaborative decision, and relates to the technical field of logistics and artificial intelligence. The system adopts a'sensing-preprocessing-decision-execution-feedback 'closed-loop architecture, and comprises a sensing layer, a data preprocessing layer, a core decision-making layer, an execution layer and a feedback optimization layer. The core decision-making layer solves the problems that an existing logistics management and control module is isolated, data fusion is insufficient and the like through cooperative linkage of dynamic path planning, intelligent inventory prediction, cargo state intelligent monitoring and a resource optimal scheduling sub-module in combination with multi-modal data deep fusion, online iterative optimization and scene self-adaptive configuration. The logistics full-link global optimal management and control is realized, the path planning efficiency is improved by more than 40%, the inventory prediction accuracy rate reaches more than 96%, the abnormal response reaches the millisecond level, the logistics comprehensive cost is reduced by 20%-25%, and the method is suitable for multiple scenes such as power distributors, cold chains and industries, and has remarkable practicability and industrialization value.

Description

Technical Field

[0001] This invention relates to the field of supply chain management technology, and in particular to an intelligent logistics end-to-end control system based on multimodal AI collaborative decision-making. Background Technology

[0002] With the increasing complexity of global e-commerce and industrial supply chains, modern logistics systems face unprecedented challenges. Traditional logistics management systems largely rely on pre-set rules and single-point information solutions, which have the following inherent drawbacks:

[0003] (1) Information fragmentation and decision isolation: Data in warehousing, transportation, and distribution are independent of each other, forming "information silos". Decision systems (such as route planning and inventory management) are usually optimized based on local information, lacking a full-link perspective, resulting in low overall efficiency.

[0004] (2) Slow response and static nature: Existing systems are unable to respond in real time to dynamically changing external environments (such as sudden traffic congestion, order surges, and cargo anomalies). Path planning is mostly static or semi-static, and cannot be recalculated and scheduled within minutes.

[0005] (3) Reliance on human experience and rules: The core decision-making process relies heavily on the experience of managers and lacks data-driven intelligent analysis capabilities. It is difficult to find the optimal balance in anomaly handling and multi-objective optimization (such as cost, timeliness, and service quality).

[0006] (4) Single perception dimension: Traditional systems mainly process structured data (such as order number and inventory quantity), and do not make sufficient use of the large amount of unstructured data (such as cargo images and surveillance videos) and semi-structured data (such as real-time traffic conditions), thus failing to achieve a comprehensive and accurate perception of logistics status.

[0007] Although existing logistics systems utilize artificial intelligence, they typically target only a single aspect (such as optimizing routes or predicting inventory), and lack effective collaboration mechanisms between AI modules. This prevents the formation of a complete closed loop from perception and decision-making to execution and feedback optimization, leaving significant room for improvement in their intelligence level and overall efficiency. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to provide an intelligent logistics end-to-end management and control system based on multimodal AI collaborative decision-making. This system aims to achieve adaptive and intelligent management and control of the entire logistics chain through multimodal data fusion, deep collaboration of core AI modules, and closed-loop feedback optimization, thereby comprehensively improving logistics efficiency, reducing operating costs, and enhancing system resilience.

[0009] Technical solution:

[0010] A smart logistics end-to-end control system based on multimodal AI collaborative decision-making is characterized by comprising a perception layer, a data preprocessing layer, a core decision-making layer, an execution layer, and a feedback optimization layer connected in sequence, wherein the output of the feedback optimization layer is connected to the core decision-making layer, thereby forming a closed-loop intelligent control system.

[0011] The perception layer includes a multimodal data acquisition module, used to collect structured data, unstructured data and semi-structured data in the entire logistics chain;

[0012] The data preprocessing layer includes an AI heterogeneous data fusion module, which is used to clean, align and extract features from the multi-source heterogeneous data collected by the perception layer to generate standardized multimodal feature vectors.

[0013] The core decision-making layer includes a multi-module collaborative AI decision-making module, which consists of a dynamic path planning sub-module, an intelligent inventory prediction sub-module, an intelligent cargo status monitoring sub-module, and a resource optimal scheduling sub-module. Each sub-module interacts and links with the other through a data interface.

[0014] The execution layer includes an execution control module, which is used to convert the collaborative decision-making results of the core decision-making layer into executable instructions and drive downstream logistics equipment and systems to complete warehousing, transportation and anomaly handling operations.

[0015] The feedback optimization layer includes a model iteration optimization module, which is used to collect the running effect data of the execution layer and update and iteratively optimize the parameters of the AI ​​model in the core decision layer.

[0016] The decision-making linkage mechanism of the core decision-making layer specifically includes:

[0017] The dynamic route planning submodule feeds back output data, including real-time traffic conditions and vehicle occupancy status, to the resource optimal scheduling submodule; the resource optimal scheduling submodule dynamically adjusts the capacity allocation scheme for subsequent orders based on the feedback data.

[0018] The intelligent inventory forecasting submodule synchronously pushes SKU-level inventory forecast data to the warehouse scheduling unit of the execution layer and the resource optimal scheduling submodule; the warehouse scheduling unit pre-generates sorting paths based on the forecast data, and the resource optimal scheduling submodule matches corresponding transportation resources based on the forecast data.

[0019] The intelligent cargo status monitoring submodule synchronizes cargo anomaly level information to the dynamic path planning submodule and the optimal resource scheduling submodule. When the anomaly level reaches a preset threshold, the dynamic path planning submodule plans an emergency handling path, and the optimal resource scheduling submodule initiates the backup resource allocation process.

[0020] The multimodal data acquisition module includes:

[0021] The IoT sensor unit is used to collect physical environment data, including temperature and humidity sensors, vibration sensors, GPS positioning modules and shelf pressure sensors.

[0022] The visual acquisition unit is used to acquire image and video data, including cameras deployed in warehouses, transport vehicles, and cargo loading and unloading points;

[0023] The data interface unit is used to access data from external systems, including traffic management department road condition platforms, e-commerce order systems, and market demand data platforms.

[0024] The AI ​​heterogeneous data fusion module employs an adaptive noise filtering and feature fusion algorithm.

[0025] The adaptive noise filtering algorithm employs differentiated filtering strategies for data of different modalities, including Kalman filtering for sensor data, median filtering for image data, and regular expression matching for text data. The feature fusion algorithm adopts a Transformer-based architecture to extract and fuse numerical features of structured data, visual features of unstructured data, and semantic features of semi-structured data, and outputs a feature vector of a unified dimension.

[0026] The dynamic route planning submodule uses a reinforcement learning algorithm for modeling. Its state space includes real-time traffic conditions, order attributes, vehicle status, and constraints. Its reward function integrates route time, transportation cost, and order fulfillment rate. This submodule is configured to recalculate and output the optimized route within a specified time when preset trigger conditions are met. The resource optimal scheduling submodule uses a model combining mixed integer programming and deep learning to construct and solve an optimization problem with transportation timeliness and operating cost as objectives.

[0027] The intelligent inventory prediction submodule adopts a recurrent neural network model that incorporates an attention mechanism;

[0028] The model takes historical inventory data, order data, market characteristics, and environmental characteristics as inputs, and dynamically adjusts the weights of each input feature through the attention mechanism. The submodule is also equipped with a prediction error feedback mechanism, which automatically triggers the feature reconstruction and retraining process of the model when the continuous prediction error exceeds the preset tolerance.

[0029] The intelligent cargo status monitoring submodule adopts a technical solution that combines computer vision and information fusion.

[0030] The technical solution first identifies visual anomalies in cargo images using a target detection algorithm, then combines sensor data and employs DS evidence theory to fuse and reason about multi-source anomaly information, ultimately outputting a quantified anomaly level. The anomaly level is associated with a graded response strategy, with different levels triggering different levels of system alarms and handling instructions.

[0031] The mechanism continuously collects actual operational metrics data from the execution layer and uses incremental learning algorithms to update the parameters of the AI ​​model in the core decision layer; when the performance metrics of any sub-module are continuously lower than a preset threshold, the model structure optimization process of that sub-module is automatically triggered.

[0032] The dynamic route planning submodule and the resource optimal scheduling submodule form a capacity-route collaborative optimization loop; the capacity allocation result of the resource optimal scheduling submodule serves as the input of the dynamic route planning submodule, and the route time prediction data output by the dynamic route planning submodule is fed back to the resource optimal scheduling submodule; the resource optimal scheduling submodule judges the order fulfillment risk based on the route time prediction data, and makes a new decision on the capacity allocation scheme accordingly.

[0033] The system also includes a scenario adaptation module; the scenario adaptation module allows users to adjust the decision weights and thresholds of each sub-module in the core decision layer according to different logistics business scenarios through a parameter configuration interface; the different logistics business scenarios include at least e-commerce logistics, cold chain logistics and industrial logistics.

[0034] Beneficial effects:

[0035] 1. Through deep collaboration among sub-modules within the core decision-making layer, decision-making barriers in various logistics links have been broken down, achieving integrated collaborative optimization across the entire supply chain from warehousing and distribution to transportation, and realizing a systemic effect of "1+1>2".

[0036] 2. Based on reinforcement learning-based path planning and online learning mechanisms, the system can respond to changes in the external environment in seconds, significantly improving the agility and robustness of the logistics chain.

[0037] 3. Through multimodal data fusion and advanced AI algorithms, the system achieves accurate prediction and optimization of inventory, routes, cargo status, and resource scheduling, reducing reliance on subjective human experience and significantly enhancing the scientific nature of decision-making.

[0038] 4. The unique feedback optimization layer design enables the system to continuously learn from practice. The AI ​​model can continuously iterate and optimize as business data accumulates, making the system "smarter" the more it is used, and the long-term maintenance cost is low.

[0039] 5. Through parameterized configuration, the system can flexibly adapt to the core needs of different scenarios such as e-commerce logistics, cold chain logistics, and industrial logistics, making it highly versatile. Attached Figure Description

[0040] Figure 1 This is a system architecture diagram of the present invention;

[0041] Figure 2 This is the architecture diagram of the core decision-making layer of this invention;

[0042] Figure 3 This is an architecture diagram of the multi-module acquisition module of the present invention. Detailed Implementation

[0043] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example 1

[0045] like Figure 1-3 As shown, the perception layer is specifically constructed through a multimodal data acquisition module. This module integrates three functional units: an IoT sensor unit, a visual acquisition unit, and a data interface unit. The IoT sensor unit consists of a high-precision temperature and humidity sensor, a triaxial vibration sensor, and a GPS module with integrated 4G communication. They transmit data to the edge gateway via low-power protocols such as LoRaWAN. The visual acquisition unit uses network cameras supporting the RTSP protocol, densely deployed between warehouse shelves, inside transport vehicle cargo boxes, and loading and unloading platforms, capturing visual information at a rate of 30 frames per second. The data interface unit, as the portal for the system to connect with external data, securely connects to Gaode Maps' real-time traffic conditions, e-commerce platform order systems, and market data analysis services through an enterprise-level API gateway, using the OAuth2.0 protocol to ensure the security and reliability of data exchange. The design intent of this layer is to build a fully covered neural network, seamlessly transforming the physical state of the entire logistics chain, including the micro-environment of goods, the macro-position of vehicles, and the external social environment, into a standardized data stream that can be understood by the computing system, laying a solid data foundation for upper-level intelligent decision-making.

[0046] The core task of the data preprocessing layer is undertaken by the AI ​​heterogeneous data fusion module, whose workflow consists of two closely linked stages. The first stage is adaptive noise filtering, which applies differentiated processing to different data characteristics: Kalman filtering is used for state prediction and smoothing of sensor time-series data to suppress random fluctuations; mean filtering is applied to image data to effectively remove salt-and-pepper noise while preserving edge details; and carefully designed regular expression rules are used to accurately remove invalid characters and formatting tags from text data. The second stage is feature fusion. This module uses a Transformer-based architecture, employing a pre-trained VisionTransformer model to extract deep visual features from images, while simultaneously using a BERT model to encode rich semantics in the text. Finally, these heterogeneous features are concatenated and aligned with the numerical features of structured data, mapping them to a unified 1024-dimensional feature vector space. The fundamental purpose of this series of processes is to solve the problem of heterogeneity in multi-source data. By refining the chaotic raw data into clean, aligned, and semantically rich standardized features, it provides high-quality, computable input for the subsequent core decision-making layer, directly determining the upper limit of the entire system's intelligence level.

[0047] The core decision-making layer's collaborative mechanism is implemented in engineering through a central event bus architecture. In this architecture, each AI submodule acts as both a decision-making unit and an information publisher. When a submodule, such as the dynamic route planning submodule, completes its calculations, it encapsulates key metadata such as route time into a specific event and publishes it to the bus. Simultaneously, submodules such as resource optimization scheduling, as subscribers, continuously listen for related events. Once an event is received, they trigger their own collaborative computation logic, such as reassessing order fulfillment risks and adjusting capacity allocation based on new route times—the entire process is asynchronous and parallel. This loosely coupled design based on a publish-subscribe model allows highly specialized AI modules to operate independently and efficiently, while also achieving deep strategic interaction through lightweight information transmission. At the system level, it ensures that warehousing, transportation, and delivery are no longer isolated information silos, but rather function as an organic whole, responding in unison and enabling globally optimized intelligent behaviors when facing complex and ever-changing logistics scenarios.

[0048] The dynamic route planning submodule is constructed using a near-end policy optimization reinforcement learning algorithm. Its state space is a high-dimensional vector that integrates various constraints such as real-time traffic conditions, order priority, vehicle real-time load, and fuel consumption. Its reward function is carefully designed as a multi-objective trade-off function, simultaneously considering negative total transportation time, negative total operating costs, and positive order fulfillment rate. Deployed as a cloud-native microservice, this module automatically invokes the service immediately when the system detects a congestion index increase exceeding 30% within 5 minutes or receives a new urgent order, requiring it to return a replanned optimal route within 10 seconds. Essentially, this method simulates an agent learning optimal strategies through continuous interaction with its environment. It elevates route planning from static computation based on fixed rules to an autonomous decision-making capability that dynamically adapts to the environment and comprehensively weighs the interests of multiple parties, thereby optimizing operating costs while ensuring timeliness.

[0049] The intelligent cargo status monitoring submodule's execution flow comprises two sequential intelligent analysis stages. In the first stage, the system utilizes a YOLOv8 object detection model pre-trained on a large-scale dataset to perform real-time analysis of key video frames captured by cameras, accurately identifying various visual anomalies such as packaging damage, cargo tilting, and water immersion. In the second stage, the confidence levels of these visual anomalies, along with data from sensors such as excessive temperature and humidity readings and vibration exceeding thresholds, are input as multi-source evidence into the DS evidence theory model for information fusion calculation. This model calculates the comprehensive confidence level of each anomaly hypothesis through Bayesian inference and outputs a quantified anomaly level accordingly. This technical approach, combining cutting-edge computer vision with classical information fusion theory, significantly improves the robustness and accuracy of status determination. It effectively avoids the risk of false alarms or missed alarms from single sensors and provides an objective and accurate quantification and classification of cargo risks in transit, thus providing irrefutable decision-making basis for the execution layer to take differentiated, timely, and precise handling measures, greatly enhancing the safety and reliability of the logistics process.

[0050] The feedback optimization layer acts as a self-evolutionary engine within the system. It continuously collects real-world operational feedback from the execution layer, including key performance indicators such as actual path latency, inventory prediction errors, and anomaly detection accuracy. This massive amount of data from the production line drives online learning algorithms, silently and continuously updating the parameters of each AI model in the core decision-making layer. Furthermore, this layer establishes a comprehensive performance monitoring system. When the system detects that the evaluation metrics of a submodule are below a preset threshold for several consecutive periods, it automatically triggers a deep optimization process for that module, and may even initiate adjustments to the model structure. The existence of this layer allows the entire system to transcend the traditional, fixed paradigm of software deployment, transforming it into a living organism capable of continuously learning and calibrating from real business flows. By continuously fine-tuning its AI models using feedback data, the system can proactively adapt to changes in business models and fluctuations in the external environment. Its decision-making intelligence accumulates and improves over time, thereby achieving sustained evolution in system efficiency and continuous consolidation of core competitiveness.

[0051] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An intelligent logistics full-link management and control system based on multi-modal AI collaborative decision-making, characterized in that, The intelligent management and control system comprises a perception layer, a data preprocessing layer, a core decision layer, an execution layer and a feedback optimization layer connected in sequence, and an output end of the feedback optimization layer is connected to the core decision layer, so as to form a closed-loop intelligent management and control system; The perception layer comprises a multi-modal data acquisition module for acquiring structured data, unstructured data and semi-structured data in a logistics chain; The data preprocessing layer comprises an AI heterogeneous data fusion module for cleaning, aligning and feature extraction of multi-source heterogeneous data acquired by the perception layer, and generating a standardized multi-modal feature vector; The core decision layer comprises a multi-module collaborative AI decision module composed of a dynamic path planning submodule, an intelligent inventory prediction submodule, a cargo state intelligent monitoring submodule and a resource optimal scheduling submodule, and each submodule interacts and links through a data interface; The execution layer comprises an execution control module for converting the collaborative decision result of the core decision layer into an executable instruction and driving downstream logistics equipment and systems to complete warehouse, transportation and abnormal handling operations; The feedback optimization layer comprises a model iteration optimization module for collecting running effect data of the execution layer and updating and iteratively optimizing AI models in the core decision layer.

2. The intelligent logistics full-link management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The decision linkage mechanism of the core decision layer comprises: The dynamic path planning submodule feeds back output data containing real-time road conditions and vehicle occupancy state to the resource optimal scheduling submodule; the resource optimal scheduling submodule dynamically adjusts the transportation capacity allocation scheme of subsequent orders based on the feedback data; The intelligent inventory prediction submodule synchronously pushes SKU-level inventory prediction data to the warehouse scheduling unit of the execution layer and the resource optimal scheduling submodule; the warehouse scheduling unit generates a sorting path in advance according to the prediction data, and the resource optimal scheduling submodule matches corresponding transportation capacity resources according to the prediction data; The cargo state intelligent monitoring submodule synchronously feeds back cargo abnormality level information to the dynamic path planning submodule and the resource optimal scheduling submodule; when the abnormality level reaches a preset threshold, the dynamic path planning submodule plans an emergency handling path, and the resource optimal scheduling submodule starts a backup resource allocation process. 3.The intelligent logistics whole-chain management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The multi-modal data acquisition module comprises: An IoT sensor unit for acquiring physical environment data, including a temperature and humidity sensor, a vibration sensor, a GPS positioning module and a shelf pressure sensor; A visual acquisition unit for acquiring image and video data, including cameras deployed in warehouses, transportation vehicles and cargo loading and unloading points; A data interface unit for accessing data of external systems, including a road condition platform of a traffic management department, an e-commerce order system and a market demand data platform.

4. The intelligent logistics full-link management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The AI heterogeneous data fusion module adopts an adaptive noise filtering and feature fusion algorithm; The adaptive noise filtering algorithm adopts differentiated filtering strategies for data of different modalities, including Kalman filtering for sensor data, median filtering for image data, and regular expression matching for text data; the feature fusion algorithm adopts a Transformer-based architecture to extract and fuse numerical features of structured data, visual features of unstructured data, and semantic features of semi-structured data, and outputs a feature vector of uniform dimension.

5. The intelligent logistics full-link management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The dynamic path planning submodule adopts a reinforcement learning algorithm for modeling, with a state space including real-time road conditions, order attributes, vehicle states, and constraint conditions, and a reward function that comprehensively considers path time consumption, transportation cost, and order fulfillment rate; the submodule is configured to recalculate and output an optimized path within a specified time when a preset triggering condition is met; The resource optimal scheduling submodule adopts a model combining mixed integer programming and deep learning to construct and solve an optimization problem targeting transportation timeliness and operating cost.

6. The intelligent logistics full-link management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The intelligent inventory prediction submodule adopts a recurrent neural network model that fuses an attention mechanism; The model takes historical inventory data, order data, market characteristics, and environmental characteristics as inputs, and dynamically adjusts the weights of each input feature through the attention mechanism; the submodule is also provided with a prediction error feedback mechanism that automatically triggers feature reconstruction and retraining of the model when consecutive prediction errors exceed a preset tolerance.

7. The intelligent logistics full-link management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The intelligent cargo state monitoring submodule adopts a technical solution combining computer vision and information fusion; The technical solution first identifies visual abnormalities in cargo images through a target detection algorithm, then combines sensor data to perform fusion reasoning on multi-source abnormal information using D-S evidence theory, and finally outputs a quantified abnormality level; the abnormality level is associated with a graded response strategy, with different levels triggering different levels of system alarms and disposal instructions. 8.The intelligent logistics whole-chain management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The mechanism continuously collects actual operation indicator data from the execution layer and updates the parameters of AI models in the core decision layer using an incremental learning algorithm; when the performance indicators of any submodule continuously fall below a preset threshold, the model structure optimization process of the submodule is automatically triggered. 9.The intelligent logistics whole-chain management and control system based on multi-modal AI collaborative decision according to claim 5, characterized in that, The dynamic path planning submodule and the resource optimal scheduling submodule form a transport capacity-path collaborative optimization loop; the transport capacity allocation results of the resource optimal scheduling submodule serve as inputs for the dynamic path planning submodule, and the path time consumption prediction data output by the dynamic path planning submodule are fed back to the resource optimal scheduling submodule; based on the path time consumption prediction data, the resource optimal scheduling submodule judges the order fulfillment risk and accordingly redecides the transport capacity allocation scheme.

10. The intelligent logistics full-link management and control system based on multi-modal AI collaborative decision according to claim 1, characterized in that, The system also includes a scenario adaptation module; the scenario adaptation module allows users to adjust the decision weights and thresholds of each submodule in the core decision layer according to different logistics business scenarios through a parameter configuration interface; the different logistics business scenarios at least include e-commerce logistics, cold-chain logistics, and industrial logistics.

Images