Parcel clamping stagnation diagnosis and processing system and method based on large language model

Abstract

The invention discloses a parcel clamping stagnation diagnosis and processing system and method based on a large-scale language model, and the system comprises an acquisition module, a multi-modal fusion layer, a prompt construction module, an LLM large-scale language model, a decision module, an execution module, a feedback collection module and a reinforcement learning module. The matching method comprises the steps of monitoring, sorting, parcel state diagnosis, automatic or human intervention execution of corresponding decisions according to diagnosis results, summarization and feedback, and enhancement of LLM large-scale language model learning. The parcel sorting condition is monitored in a data fusion mode, abnormal numerical values are responded and diagnosed, the model judges whether clamping stagnation occurs or not, measures such as confidence analysis, decision execution of distribution of affected target groups, manual troubleshooting and continuous monitoring are given, an operator is allowed to know diagnosis information and then conduct manual intervention, and the accuracy and the reliability of parcel sorting are improved. And the feedback is recorded in time for the model to update the strategy, the architecture is optimized, and the accuracy and response efficiency of subsequent diagnosis are improved.

Description

Technical Field

[0001] This invention relates to the field of automated parcel sorting systems, and more specifically, to a parcel jam diagnosis and handling system and method based on a large language model. Background Technology

[0002] With the rapid development of e-commerce and the increasing demand for fast, accurate, and low-cost delivery, automated parcel sorting systems have become a core component of modern logistics. These systems rely on conveyor belts, barcode scanners, size and weight detection stations, and various mechanical sorting equipment (such as cross belts, picks, and ejector wheels) to guide parcels to their destination.

[0003] While these technologies have significantly improved processing efficiency and reduced reliance on manual labor, they still present numerous challenges in handling package jams and stalls. Modern sorting systems can process up to 24,000 packages per hour and are designed to handle a variety of packaging types, including cartons, plastic bags, flat items, and irregularly shaped goods. Subsequent developments such as active roller conveyors, cross-belt sorters, and multi-directional worktables have provided modularity and scalability in this field. However, these systems often prioritize speed and throughput over the timely handling and adaptation to anomalies.

[0004] Common factors that cause jams include: irregular package shape or size; fragile or underweight packaging; mechanical wear or equipment misalignment; sensor misreading or barcode recognition failure.

[0005] Although some systems have integrated anti-jamming functions, such as adding breakable trays and cabinet sensors, most still rely on rule logic or manual intervention to solve jamming problems, resulting in frequent system downtime and decreased operational efficiency.

[0006] Based on the investigation and summary, the existing system has the following limitations: it lacks real-time adaptive diagnostic capabilities, cannot dynamically analyze sensor data, and cannot identify the root cause of jamming; its contextual understanding capabilities are limited, and the rule-based system has difficulty handling edge cases, such as overlapping packages or fuzzy sensor data; it relies on manual intervention, requiring manual inspection and troubleshooting, which increases labor costs and system latency; and it has poor adaptability to diverse packaging, with fragile, irregular, or lightweight packages being more prone to stagnation or missorting.

[0007] Therefore, there is an urgent need for a more intelligent sorting system with context awareness capabilities that can autonomously detect, diagnose, and respond to jamming events. Summary of the Invention

[0008] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a package jam diagnosis and processing system and method based on a large language model, so as to solve one or more of the above-mentioned problems.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for diagnosing and handling package jams based on a large language model, including S1. Real-time sensing and monitoring of sorting. Collect and organize package data, and continuously monitor package movement and sorting dynamics; S2, Package Status Diagnosis Multimodal fusion of package data, combined with operation log information, is used to construct diagnostic suggestions. The diagnostic prompts are analyzed based on a strategy using a large-scale LLM language model, and the diagnostic results are output. S3. Based on the diagnostic results, execute corresponding decisions automatically or with human intervention. S3.1 If no blockage is detected, record the event, continue monitoring, and maintain the observation cycle for subsequent packages; S3.2 If a blockage is detected, perform the corresponding operation to clear the blockage. S4. Summary and feedback Record and collect the results of the system's automatic processing or the operations performed by human intervention; S5. Enhance the learning of large-scale language models using LLM. Based on feedback from previous processing results, the LLM large language model is updated to provide suggestions for construction and decision-making strategies, and to optimize subsequent diagnostics.

[0010] Furthermore, in step S1, different package data are collected through sensor detection, metadata tracking, image acquisition, and log retrieval. Sensor and image data include the physical behavior and motion characteristics of packages in the sorting environment; Metadata includes the package's static and dynamic attributes; The operation log information includes a historical record of system time.

[0011] Furthermore, in step S2, the different types of package data obtained are combined with operation log information to perform multimodal fusion and integrate them into a unified package status and context representation; Based on the integrated content, diagnostic prompts are built, structured input is generated, and then sent to the LLM large-scale language model.

[0012] Furthermore, in step S2, the LLM large-scale language model outputs diagnostic assessment results based on diagnostic prompts, including determining whether a pause has occurred, estimating the cause of the pause, and generating a confidence score.

[0013] Furthermore, the diagnostic results in step S3 include high confidence indicating that a jam has occurred, uncertain jam occurrence, and no jam detected; High confidence indicates a jam or uncertainty indicates a jam, proceeding to step S3.1. A high confidence level indicates a blockage. The system then determines if there is a backup conveyor belt. If there is, the affected packages are redirected to the backup conveyor belt. If there is no backup conveyor belt, the operator is notified to request manual intervention, and the event is recorded in the log module for tracking and learning. If a glitch occurs, the system records the event and alerts the operator, allowing for manual intervention and feedback. No jamming was detected in step S3.2.

[0014] Furthermore, in step S4, feedback is captured by the feedback collection module for cases where there is high confidence indicating stagnation and uncertain stagnation. The feedback includes qualitative assessment, quantitative performance indicators, and contextual metadata.

[0015] Furthermore, the update of the LLM large language model in step S5 includes modifying the internal decision-making strategy, optimizing the selection logic involving corrective measures, reallocating the influence weights of sensor data inputs, and improving the interpretation of the LLM confidence score. The updated LLM large language model performs optimization by introducing new features, emphasizing previously underestimated signals, and adjusting the weight ratio between sensor data and historical context. The updated strategy and cueing mechanism are applied to the next round of lag detection and response. The reinforcement learning of large-scale language models in LLM employs one or more standard learning methods, such as Q-learning, gradient strategy, and Actor-Critic method.

[0016] The system used in conjunction with the above method includes The logging module records all operations and generates a log, connecting to all other modules; The data collection module collects information from the packages; A multimodal fusion layer connects the acquisition module and fuses the data from the acquisition module and the log module into a suitable unified representation; The prompt building module connects to the multimodal fusion layer and generates diagnostic prompts for the fused multimodal data; The LLM large-scale language model connects the prompt building module, analyzes based on diagnostic prompts, and outputs corresponding decisions; The decision-making module connects to the large-scale language model of LLM and makes decisions spontaneously or with human intervention. The execution module connects to the decision-making module and executes the corresponding actions based on the decisions made. The feedback collection module connects to the execution module and collects feedback on system decision results or operator behavior. The reinforcement learning module connects to the feedback collection module, updates the strategies used for prompt construction and decision-making based on feedback information, optimizes subsequent diagnostic prompts, and improves detection accuracy.

[0017] Furthermore, the data acquisition module includes The sensor data acquisition module consists of multiple sensors that upload data in real time, collecting objective information about the package. The image acquisition module dynamically captures image data along the transmission path and monitors the movement information of the package in real time. The metadata processing module retrieves aggregated package data and provides package context information.

[0018] Furthermore, the decision-making module also includes a human-machine collaboration interface, which includes an operator dashboard and an intervention decision-making module. The operator dashboard displays diagnostic outputs obtained from the LLM large language model, and the intervention decision-making module executes the decision after the human judges whether to accept or reject the diagnostic output.

[0019] In summary, the present invention has the following beneficial effects: it monitors the parcel sorting situation by fusing sensor data, image data, and metadata, reacts to and diagnoses abnormal values, uses an LLM (Large Language Model) to determine whether there is a bottleneck and provides confidence analysis, and then the system decides to implement measures such as diverting affected target groups, manually troubleshooting, and continuing monitoring. It allows operators to intervene manually after understanding the diagnostic information and records feedback in a timely manner so that the LLM can update its strategy and optimize its architecture, thereby improving the accuracy and response efficiency of subsequent diagnoses. Attached Figure Description

[0020] Figure 1 A structural flowchart of one embodiment of the present invention; Figure 2 A diagnostic flowchart is provided as one embodiment of the present invention; Figure 3 This is a flowchart illustrating human-machine collaboration in one embodiment of the present invention.

[0021] In the diagram: 101, Sensor Data Acquisition Module; 102, Image Acquisition Module; 103, Metadata Processing Module; 201, Multimodal Fusion Layer; 301, Prompt Construction Module; 401, LLM Large Language Model; 501, Decision Module; 502, Operator Dashboard; 503, Intervention Decision Module; 601, Execution Module; 701, Feedback Collection Module; 801, Reinforcement Learning Module; 901, Log Module. Detailed Implementation

[0022] Example: The following is in conjunction with the appendix Figure 1-3 The present invention will be described in further detail below.

[0023] The terms used herein (including, but not limited to, “package,” “stuck,” “diagnostics,” “sensor fusion,” “LLM,” “confidence score,” and “human intervention”) should be interpreted broadly and encompass all technical equivalents and variations known to those skilled in the art. References to specific implementations, configurations, or operational procedures are for illustrative purposes only and should not be construed as restrictive.

[0024] Embodiments of this invention can be implemented through hardware, software, firmware, or any combination thereof. Components and processes can be distributed across multiple physical devices, computing nodes, or logical modules. The invention is not limited to any particular architecture, protocol, or deployment environment. All variations and modifications within the scope of this description are to be covered.

[0025] Package jamming diagnosis and processing system based on large-scale language models, such as Figure 1 As shown, the system mainly includes a log module 901, an acquisition module, a multimodal fusion layer 201, a prompting construction module 301, an LLM (Large Language Model) 401, a decision-making module 501, an execution module 601, a feedback collection module 701, and a reinforcement learning module 801. The log module 901 records possible operations or historical information from the other modules, including historical system time records such as previous glitches, specific operator interventions, LLM outputs, and corrective actions taken. The log can also record confidence scores and timestamps related to the LLM output to support auditability, model optimization, system transparency, and provide timestamps and contextual basis for current observations. The acquisition module includes a sensor data acquisition module 101, an image acquisition module 102, and a metadata processing module 103. The decision-making module 501 also includes a human-machine interface, which includes an operator dashboard 502 and an intervention decision-making module 503.

[0026] The data acquisition module collects information from the packages. The image acquisition module 102 is a special type of sensor data acquisition module 101. The sensor data acquisition module 101 consists of one or more sensors that upload data in real time. These sensors are used to collect data related to the packages, including but not limited to weight sensors and motion detectors. They collect objective information about the packages, reflecting their physical behavior and motion characteristics in the sorting environment, such as actual package weight, actual package size, and actual package transport speed. Sensor data has multimodal characteristics, used to characterize package behavior, detect anomalies, and provide input to the diagnostic engine. The image acquisition module 102 dynamically captures image data along the transport path, monitoring package movement in real time. The metadata processing module 103 is generally associated with the sensor data acquisition module 101, aggregating relevant metadata for each package. This metadata is mostly obtained from the package tracking system, i.e., the information stored in the package picking input, mainly focusing on the package's static and dynamic attributes, which may include package identifier, size, weight, destination, priority, and processing instructions. The metadata information collected by the metadata processing module 103 provides context for sensor readings and supports subsequent decision-making. The multimodal fusion layer 201 connects to the acquisition module and fuses the data from the acquisition module and the log module to form a suitable unified representation. This primarily involves integrating sensor data, package metadata, and historical operation log entries into a unified package status and context representation. The multimodal fusion layer 201 relates to a multimodal fusion architecture for implementing LLM (Large Language Model) 401-assisted package jam diagnosis in a sorting system. It integrates heterogeneous data sources into a unified representation, serving as input for the LLM 401 for inference and decision support. Before multimodal fusion, sensor data, image data, and metadata all require standardization, filtering, and structuring as preprocessing steps. The system receives preprocessed sensor data, accesses information records in the operation log, integrates package metadata, and finally integrates the data in the multimodal fusion layer 201. This process aligns, correlates, and synthesizes data from different modalities to form a unified representation suitable for semantic interpretation. The fusion process may involve feature extraction, temporal alignment, and context embedding. The output of the fusion layer is fed into the prompting construction module 301, and finally reaches the LLM large language model 401, which is the key diagnostic module. It performs inference to determine whether a bottleneck has occurred, identifies possible causes, generates a confidence score, and can even suggest corrective measures or mark uncertain cases for operator review. This multimodal fusion architecture, by combining the advantages of sensor data, historical logs, and wrapped specific metadata, works with the LLM large language model 401 to achieve robust and context-aware judgments, supporting real-time and retrospective analysis. The entire function can be implemented through software, hardware, or a hybrid configuration.

[0027] The prompting construction module 301, connected to the multimodal fusion layer 201, is configured to generate diagnostic prompts based on the fused multimodal data. It generates structured input for processing by the LLM large language model 401. These prompts may include sensor anomalies, package attributes, and relevant historical patterns. The LLM large language model 401, connected to the prompting construction module 301, analyzes the generated diagnostic prompts and outputs corresponding jamming detection decisions. These decisions are generally categorized into three types: confirmed jamming, uncertain jamming, and no jamming detected. This is the core functional module. After receiving data and forming diagnostic prompts, the prompting construction module 301 further processes the data to obtain diagnostic results, inferring the existence of jamming, identifying possible causes, and generating a score reflecting the confidence level of the inference. The diagnostic results can be presented to a human operator or directly used to trigger an automatic response. The LLM large language model 401 can be continuously updated and optimized under the intervention of the reinforcement learning module 801.

[0028] The decision module 501 connects to the LLM large language model 401, making decisions spontaneously or with human intervention. It serves to reconfirm the decisions generated by the model, making this part most suitable for human intervention. Therefore, an external human-machine collaboration interface is connected to demonstrate a human-participatory decision-making process. This interface manages the package delay times diagnosed by the LLM large language model 401 in the sorting system, allowing for human supervision and intervention when automatic diagnosis may require manual verification, correction, or contextual judgment. Figure 3 As shown, the system specifically includes an operator dashboard 502 and an intervention decision module 503. The operator dashboard 502 enables human-computer interaction, displaying diagnostic outputs obtained from the LLM large language model 401. This interface allows operators to view the diagnostic outputs of the LLM large language model 401, including relevant confidence scores, sensor data corroboration, and historical background information, and provides command issuance functionality. The intervention decision module 503 allows for human judgment in accepting or rejecting the diagnostic outputs made by the model. The execution module 601 is connected to the decision module 501 and executes corresponding actions based on the confirmed decisions, including actions such as redirecting packages, stopping the conveyor belt segment, or initiating maintenance protocols. This human-participatory mechanism ensures that automated decisions can be manually reviewed in cases of low confidence, ambiguous sensor data, or operational anomalies, supporting safety, accountability, and adaptability in high-risk environments.

[0029] The feedback collection module 701 connects to the execution module 601, collecting feedback on system decision results or operator behavior. It is a pre-module based on the LLM large-scale language model 401 for learning and upgrading. By integrating operational result feedback, it continuously improves the accuracy of model diagnosis and decision-making efficiency. For example... Figure 2As shown, the operation results are generally divided into successfully resolving jamming situations, partially alleviating jamming phenomena, and failing to resolve abnormal jamming. These results are correlated and analyzed together with the original diagnosis and the corrective measures taken, which are obtained from operation logs, operator notes, or downstream performance metrics. The feedback content generally includes self-qualitative assessment, quantitative performance metrics, and context metadata. The reinforcement learning module 801 is connected to the feedback collection module 701, and the feedback information is updated to be used for hint construction and decision-making strategies, optimizing subsequent diagnostic hints, improving detection accuracy, and establishing a diagnostic loop to achieve continuous learning and adaptation of the model in a closed-loop architecture, enhancing the generalization ability and long-term performance of the system in diverse operation scenarios.

[0030] Based on the above-mentioned package jamming diagnosis preprocessing method supporting the system, as Figure 1-3 shown, it includes: S1. Real-time perception, monitoring sorting, collecting and organizing package data, and continuously observing the movement and sorting dynamics of packages; collecting different package data through means such as sensor detection, metadata tracking, image acquisition, and log retrieval; The sensor and image data include the physical behaviors and motion characteristics of packages in the sorting environment, such as actual weight, transmission speed, actual posture, and comparison of historical postures; The metadata includes the static and dynamic attributes of packages, such as original size, original weight, destination, priority, and processing requirements; The operation log information includes historical records of system time, such as sensor historical readings, sensor anomalies, detected or suspected jamming events, diagnostic results generated by the LLM, manual intervention records, and system responses, confidence scores related to LLM outputs, and timestamps.

[0031] S2. Package status diagnosis, performing multimodal fusion on the obtained different types of package data in combination with the operation log information, and integrating them into a unified package status and context representation; constructing diagnostic hints based on the integrated content, generating structured inputs, and transmitting them to the LLM large language model 401; constructing diagnostic hints, analyzing the diagnostic hints based on the strategy of the LLM large language model 401, and outputting the diagnostic results; The LLM large language model 401 outputs diagnostic evaluation results according to the diagnostic hints, including judging whether jamming occurs, estimating the cause of jamming, and generating a confidence score.

[0032] S3. According to the diagnostic results, automatically or with human intervention, execute corresponding decisions. The diagnostic results include high confidence indicating jamming, uncertainty about jamming, and no detected jamming; High confidence indicates a jam or uncertainty indicates a jam, proceeding to step S3.1. A high confidence level indicates a blockage. The system then determines if there is a backup conveyor belt. If there is, the affected packages are redirected to the backup conveyor belt. If there is no backup conveyor belt, the operator is notified to request manual intervention. The event is recorded in the log module 901 for tracking and learning. If a glitch occurs, the system records the event and alerts the operator, allowing for manual intervention and feedback. No stuck entry into step S3.2 was detected; S3.1 If no blockage is detected, record the event, continue monitoring, and maintain the observation cycle for subsequent packages; S3.2 If a blockage is detected, perform the corresponding operation to clear the blockage. S4. Summary and feedback Record and collect the results of the system's automatic processing or the operations performed by human intervention; For situations where high confidence indicates stagnation or uncertainty indicates stagnation, feedback is captured by the feedback collection module 701. The feedback includes qualitative assessments, quantitative performance indicators, and contextual metadata.

[0033] S5, Enhanced learning of LLM large-scale language model 401, Based on feedback from previous processing results, the LLM large language model 401 was updated to provide strategies for prompting construction and decision-making, and subsequent diagnostics were optimized. The LLM Large Language Model 401 update includes modifications to the internal decision-making strategy, optimization of the selection logic involving corrective actions, reallocation of the influence weights of sensor data inputs, and improvement of the interpretation of LLM confidence scores. The updated LLM large language model 401 performs optimization, introducing new features, emphasizing previously underestimated signals, and adjusting the weight ratio between sensor data and historical context. The updated strategy and cueing mechanism are applied to the next round of lag detection and response. The reinforcement learning of the LLM large language model 401 employs one or more standard techniques, such as Q-learning, gradient strategies, and Actor-Critic methods.

[0034] The entire response process involves the acquisition module monitoring package sorting behavior, data transmission being preprocessed and then transformed into a suitable unified representation by the multimodal fusion layer 201, followed by diagnostic prompts generated by the prompt construction module 301. These prompts are then analyzed and diagnosed by the LLM large-scale language model 401. After secondary confirmation by the decision module 501, the execution module 601 directly processes the data according to different strategies to complete the package delay response. The decision results or intervention actions are centrally pushed to the reinforcement learning module 801 by the feedback collection module 701. The reinforcement learning module 801 ingests historical operation logs, operator feedback, and result data to retrain or fine-tune the LLM large-scale language model 401, thereby improving future diagnostic accuracy and response capabilities. After processing, the reinforcement learning module 801 updates the model strategy and optimizes the model method. All operations are stored in the log module 901, allowing for reference of historical information during both data processing and diagnostic stages. The system architecture is modular and can be implemented through software, hardware, or a combination of both. Components are distributed across multiple physical devices or integrated into a unified control platform. The system is designed to run in real time and can be deployed in various parcel sorting environments, including warehouses, distribution centers, and logistics hubs.

[0035] The diagnostic workflow begins with monitoring. The system starts, continuously observing package movement and system status, collecting sensor data, image data, and package metadata. It acquires data from multiple sensors in real time and obtains specific package information. The system then autonomously preprocesses the data, performing multimodal fusion. The fused data is used to construct diagnostic prompts. In the model, jam detection corresponds to the evaluation of the LLM output. If the diagnostic result indicates jamming with high confidence, the system activates a backup conveyor belt, redirecting the affected package to it. If no backup conveyor belt is available, the operator is notified for manual intervention. The event is then recorded by the log module 901 for future learning and tracking. If the diagnostic result is uncertain, a person can record the event and notify the operator, allowing for manual intervention and feedback. Feedback from both high-probability and uncertain jamming scenarios is captured by the feedback collection module 701 and used by the reinforcement learning module 801 to update the strategy and optimize the model's future diagnostic capabilities. If no jamming is detected, the system remains healthy, continuing to observe subsequent packages. The entire system is a closed loop, integrating real-time perception, LLM inference, human-machine collaborative decision-making, and adaptive learning capabilities to support automated and semi-automated responses to sorting anomalies, indicating operational resilience and diagnostic accuracy.

[0036] The reinforcement learning (RL) feedback process supports adaptive optimization of the model for stuck event diagnosis. It begins by observing operational outcomes—the results of responses taken by the system or operator to diagnosed stuck events, which may include successful resolution, partial mitigation, or failure to resolve the anomaly. These outcomes are then correlated with the original diagnosis and the corrective actions taken. The feedback collection module 701 collects structured and unstructured feedback related to operational outcomes. Feedback sources may include system logs, operator comments, or downstream performance metrics. Feedback content may include qualitative assessments, quantitative performance metrics, and contextual metadata. The collected feedback is used to perform reinforcement learning policy updates, which modify the system's internal decision-making strategy. This may involve the selection of corrective actions, the weighting of sensor inputs, or the interpretation of LLM confidence scores. Reinforcement learning updates can employ standard techniques such as Q-learning, policy gradients, or Actor-Critic methods. Subsequently, the system performs LLM cue policy updates, optimizing the structure, content, or contextual framework of cue submitted to the LLM in future diagnostic cycles. The updated cue policy may introduce new features, emphasize previously underestimated signals, or adjust the weighting ratio between sensor data and historical context. The process concludes with a diagnostic loop, where the updated strategies and alerting mechanisms are applied to the next round of lag detection and response. This closed-loop architecture enables continuous learning and adaptation, enhancing the system's generalization ability and long-term performance across diverse operational scenarios.

[0037] In the event of uncertainties, jams, or other unexpected situations, a human-involved decision-making process is introduced. This involves human supervision and intervention when automatic diagnosis may require verification, correction, or contextual judgment. Starting with the model's diagnostic output, which is based on multimodal fusion data, the diagnostic results may include whether jamming has occurred, possible causes, and recommended corrective actions. The diagnostic results are presented to the operator via an operator dashboard 502, which allows the operator to view the LLM model's diagnostic output, including relevant confidence scores, sensor evidence, and historical context information. At the intervention decision module 503, the operator evaluates whether to accept the model's diagnostic recommendations or human intervention. If intervention is selected, the operator performs an alternative operation or provides corrective input; if the diagnosis is accepted, the system makes an automatic decision. Regardless of acceptance, the system's execution module 601 and decision module 501 complete the relevant operations, which may include redirecting packages, stopping conveyor belt segments, or initiating maintenance protocols.

[0038] In recent years, Large Language Models (LLMs) 401 have demonstrated great potential in parsing multimodal data (such as text, images, and sensor streams) and making context-sensitive decisions. Although LLMs have been applied in fields such as grocery packaging (e.g., LLM-Pack uses visual language models to simulate manual packing strategies), their application in parcel sorting is still in its early stages. Currently, parcel sorting technologies focus on innovations in mechanical structures (such as multi-layer sorting systems, volumetric loading, and container-based processing). These systems improve parcel throughput and space efficiency, but have not yet introduced LLM or AI-driven diagnostic mechanisms. Therefore, this solution introduces LLMs 401 to analyze sensor anomalies and operation logs, predict and classify jamming scenarios, recommend or execute corrective measures, and learn from historical events to improve future performance. In other words, it introduces a parcel jamming detection and response framework based on LLMs 401, realizing a shift from a passive rule-based system to proactive intelligent automation. Traditional rule-based systems often struggle to adapt to the complex and ever-changing real-world sorting environments. Package jams can be caused by factors such as overweight, irregular sizes, fragile packaging, or mechanical failures, severely disrupting logistics operations. This application uses a large-scale language model (LLM) 401 to detect, diagnose, and respond to package jam events in automated sorting facilities.

[0039] The entire system uses an LLM (Large Language Model) 401 to analyze multimodal sensor data (such as weight sensors, visual images, and motion detectors), package metadata (such as package type), and operation logs to identify possible characteristics of package stagnation or jamming. After training or fine-tuning, the LLM 401 can recognize contextual cues and anomalies, enabling multiple functions: real-time detection of package jamming; diagnosing the root cause of jamming based on package characteristics, conveyor belt behavior, or environmental conditions by analyzing sensor flow, package metadata, and operation logs; recommending or automatically executing corrective measures; redirecting packages; adjusting conveyor belt parameters; or notifying operators for manual intervention; and continuously learning from historical events to improve the accuracy and timeliness of future detections. A separate manual intervention pathway, coupled with a feedback loop, continuously optimizes the LLM 401 using operator intervention behaviors and processing results, thus enabling continuous learning. The entire solution supports integration with existing warehouse management systems and robot control platforms, enhancing operational flexibility and system throughput.

[0040] Abnormal package behavior is detected by aggregating and interpreting heterogeneous circular data. Stuck events are identified through deterministic thresholds, probabilistic inference, or diagnostic results generated by the LLM Large Language Model 401. The LLM Large Language Model 401 can generate diagnostic results with accompanying confidence scores, which can be used to guide automated or semi-automated decision-making processes. The system can also display diagnostic results to operators via an operator dashboard 502, allowing for manual intervention in the operation of the LLM Large Language Model 401. Intervention decisions are recorded and used for model optimization. The model itself has an adaptive learning mechanism that improves the accuracy and response efficiency of future diagnoses through operator feedback and historical event data.

[0041] It should be noted that this specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A parcel stuck diagnosis and handling method based on a large language model, characterized in that: Comprising S1, real-time perception, monitoring sorting, Collect and organize package data, continuously observe package movement and sorting dynamics; S2, package state diagnosis, Multi-modal fusion of package data, combined with operation log information, to construct diagnostic prompts, LLM large language model-based strategy analyzes diagnostic prompts, and diagnostic results are output; S3, according to the diagnostic results, automatically or with human intervention, execute the corresponding decision, S3.1, if the diagnosis does not occur, record the event, continue to monitor, and maintain the observation cycle of the subsequent package; S3.2, if the diagnosis occurs, execute the corresponding operation, eliminate the package jam; S4, summary feedback, Record and collect the results of system self-processing or human intervention operations; S5, strengthen the learning of LLM large language model, According to the feedback of the previous processing results, update the LLM large language model for prompt construction and decision-making strategy, optimize subsequent diagnosis.

2. The large language model based parcel stick diagnosis and processing method according to claim 1, characterized in that: In step S1, different package data are collected by means of sensor detection, metadata tracking, image acquisition, and log retrieval; Sensor and image data include the physical behavior and motion characteristics of the package in the sorting environment; Metadata includes the static and dynamic attributes of the package; Operation log information includes historical records of system time.

3. The parcel stick diagnosis and handling method based on large language model of claim 1, wherein: In step S2, different types of package data obtained are combined with operation log information for multi-modal fusion, integrated into a unified package state and context representation; Based on the integrated content, diagnostic prompts are constructed, structured inputs are generated, and transmitted to the LLM large language model.

4. The parcel stick diagnosis and handling method based on large language model of claim 3, wherein: In step S2, the LLM large language model outputs diagnostic evaluation results according to the diagnostic prompts, including judging whether a jam occurs, estimating the cause of the jam, and generating a confidence score.

5. The large language model based parcel stick diagnosis and processing method of claim 1, wherein: The diagnostic results include high confidence indicating that a jam occurs, uncertainty indicating that a jam occurs, and no jam detected; High confidence indicating that a jam occurs and uncertainty indicating that a jam occurs enter step S3.1, High confidence indicating that a jam occurs, the system determines whether there is a backup conveyor belt, if there is a backup conveyor belt, the affected package is redirected to the backup conveyor belt; if there is no backup conveyor belt, the operator is notified, prompting human intervention, and the event is recorded to the log module for tracking and learning; Uncertainty indicating that a jam occurs, the system records the event and reminds the operator, allowing human intervention and feedback; No jam detected enters step S3.

2.

6. The parcel stick diagnosis and handling method based on large language model of claim 5, wherein: In step S4, feedback is captured for high confidence indicating that a jam occurs and uncertainty indicating that a jam occurs through the feedback collection module, and the feedback content includes qualitative evaluation, quantitative performance indicators, and context metadata.

7. The large language model based parcel stick diagnosis and processing method of claim 1, wherein: In step S5, the update of the LLM large language model includes modifying the internal decision-making strategy, optimizing the selection logic involving corrective measures, redistributing the influence weight of sensor data input, and improving the explanation method of LLM confidence score; After updating, the LLM large language model performs optimization, introduces new features, emphasizes previously underestimated signals, adjusts the weight ratio between sensor data and historical context, and the updated strategy and prompt mechanism is applied to the next round of jam detection and response; The reinforcement learning of the LLM large language model adopts one or more of the following standard learning methods: Q-learning method, gradient policy, and Actor-Critic method.

8. A system implementing a large language model based parcel stuck diagnosis and processing method, characterized in that: Comprise A log module that records all operations and forms a log, connecting all other modules; A collection module that collects information about the package; A multi-modal fusion layer that connects the collection module and fuses the data from the collection module and the log module into a suitable unified representation; A prompt construction module that connects the multi-modal fusion layer and generates diagnostic prompts from the fused multi-modal data; An LLM large language model that connects the prompt construction module, analyzes based on the diagnostic prompts, and outputs corresponding decisions; A decision module that connects the LLM large language model and judges decisions autonomously or with human intervention; An execution module that connects the decision module and executes corresponding actions based on the judged decisions; A feedback collection module that connects the execution module and collects feedback on system decision results or operator behavior; A reinforcement learning module that connects the feedback collection module and updates the strategies for prompt construction and decision making based on feedback information, optimizing subsequent diagnostic prompts and improving detection accuracy.

9. The system for implementing a large language model-based parcel stalemate diagnosis and handling method according to claim 8, wherein: The collection module comprises A sensor data collection module composed of multiple sensors that upload data in real time, collecting objective information about the package; An image collection module that dynamically captures image data of the transmission path and monitors package movement information in real time; A metadata processing module that retrieves and aggregates package data to provide context information about the package.

10. The system for implementing a large language model-based parcel stalemate diagnosis and processing method of claim 8, wherein: The decision module also includes a human-machine collaboration interface, which includes an operator dashboard that displays diagnostic outputs from the LLM large language model and an intervention decision module that executes decisions by the execution module after human judgment accepts or rejects diagnostic outputs.

Images