An automated logistics conveying line scheduling optimization method and device and related medium

Abstract

The application discloses an automatic logistics conveying line scheduling optimization method and device and a related medium, the method comprises the following steps: acquiring multi-source operation data and performing data preprocessing to obtain a standardized operation state sequence; performing parameter mapping on the standardized operation state sequence to construct state characteristics and obtain an environment state vector sequence; performing strategy agent solving on the environment state vector sequence to generate initial scheduling data; inputting the initial scheduling data into a verification model for checking to obtain plan execution data; performing control instruction analysis, device mapping and message coding according to the plan execution data, generating a to-be-downloaded scheduling control message and performing transmission adaptation processing, and outputting a scheduling control result set. The application outputs the to-be-downloaded scheduling control message after corresponding processing of the calculated plan execution data, and performs transmission adaptation processing to obtain the scheduling control result set for controlling the logistics conveying line, so that the collaboration of the logistics system is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical application direction of large models and the field of automated logistics equipment scheduling, in particular to an automated logistics conveying line scheduling optimization method and device and related media. BACKGROUND

[0002] With the rapid development of e-commerce and the increasing complexity of global supply chains, the logistics industry is facing unprecedented challenges. As the core infrastructure of modern warehousing and distribution centers, the running efficiency of automated logistics conveying lines directly affects the throughput and cost of the overall logistics system. Traditional logistics conveying line scheduling systems usually rely on preset rules or manual intervention, which is difficult to effectively respond to real-time changes in logistics demand, sudden events (such as equipment failure, order surge) and complex conveying path optimization problems. This leads to a series of pain points such as conveying line congestion, low resource utilization, backlog of goods, and slow response to scheduling.

[0003] There are some intelligent management systems in the prior art, but it is also difficult to uniformly absorb and make real-time decisions on multi-source heterogeneous operation data under dynamic conditions such as equipment failure and order surge, which can easily cause conveying line congestion, low resource utilization, and slow response. There is currently a lack of end-to-end closed-loop adaptive scheduling mechanism. SUMMARY

[0004] The embodiments of the present application provide an automated logistics conveying line scheduling optimization method, device and related media, aiming to solve the problem of weak data perception of the logistics system in the prior art, which leads to poor coordination of the logistics system.

[0005] In a first aspect, the embodiments of the present application provide an automated logistics conveying line scheduling optimization method, comprising:

[0006] Obtain multi-source operation data and perform data preprocessing to obtain a standardized operation state sequence;

[0007] Map parameters to the standardized operation state sequence to construct state features and obtain an environment state vector sequence;

[0008] Call a multi-agent collaborative decision engine to solve the strategy agent for the environment state vector sequence and generate initial scheduling data;

[0009] Input the initial scheduling data into a verification model for verification and rule-based correction to obtain plan execution data;

[0010] According to the plan execution data, control instruction analysis, device mapping and message encoding are performed respectively to generate a conveying line controller-oriented scheduling control message to be issued;

[0011] The dispatch control result set is outputted after transmission adaptation processing is performed on the to-be-issued dispatch control message.

[0012] In a second aspect, an embodiment of the present application provides an automatic logistics conveying line scheduling optimization device, comprising:

[0013] A data processing unit is configured to acquire multi-source operation data and perform data preprocessing to obtain a standardized operation state sequence.

[0014] A parameter mapping unit is configured to perform parameter mapping on the standardized operation state sequence to construct a state feature and obtain an environment state vector sequence.

[0015] A strategy scheduling unit is configured to call a multi-agent collaborative decision engine to solve a strategy agent for the environment state vector sequence and generate initial scheduling data.

[0016] A data verification unit is configured to input the initial scheduling data into a verification model for verification and perform rule-based correction to obtain plan execution data.

[0017] A message generation unit is configured to perform control instruction analysis, device mapping and message encoding according to the plan execution data to generate to-be-issued dispatch control messages for a conveying line controller.

[0018] A scheduling output unit is configured to perform transmission adaptation processing on the to-be-issued dispatch control message and output a dispatch control result set.

[0019] In a third aspect, an embodiment of the present application provides a computer device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the automatic logistics conveying line scheduling optimization method of the first aspect when executing the computer program.

[0020] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executable on a processor to implement the automatic logistics conveying line scheduling optimization method of the first aspect.

[0021] The embodiment of the present application provides a kind of automatic logistics conveying line scheduling optimization method, comprising obtaining multi-source operation data and carrying out data preprocessing, obtain standardized operating state sequence;Parameter mapping is carried out to the standardized operating state sequence to construct state characteristics, and obtain environmental state vector sequence;Multi-agent collaborative decision engine is called to solve strategy agent to the environmental state vector sequence, generates initial scheduling data;The initial scheduling data is input into verification model to check and is regularized correction, obtains plan execution data;According to the plan execution data, control instruction analysis, equipment mapping and message coding are carried out respectively, generate the scheduling control message of the conveying line controller to be issued;Scheduling control message to be issued is carried out transmission adaptation processing, and output scheduling control result set.The plan execution data calculated by the present application is processed and output to the scheduling control message to be issued, and transmission adaptation processing is carried out, obtains the scheduling control result set for controlling logistics conveying line, so that the collaboration of logistics system is greatly improved.

[0022] The embodiment of the present application also provides an automatic logistics conveying line scheduling optimization device, computer equipment and storage medium, also have the beneficial effects described above. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical scheme of the embodiment of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0024] Figure 1 A flowchart of an automatic logistics conveying line scheduling optimization method provided by the embodiment of the present application is shown.

[0025] Figure 2 A schematic block diagram of an automatic logistics conveying line scheduling optimization device provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0026] The technical scheme in the embodiment of the present application will be described clearly and completely with reference to the drawings in the embodiment of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0027] It should be understood that the terms "comprises" and "comprising," when used in this specification and the following claims, indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0028] It should also be understood that the terms used in the specification of the application herein are used for the purpose of describing particular embodiments only and are not intended to limit the application. As used in the specification and the appended claims of the application, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0029] It should be further understood that the term "and / or" used in the specification of the application and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0030] Please see the following Figure 1 , Figure 1 The flowchart of an automatic logistics conveying line scheduling optimization method provided by the embodiment of the application, specifically comprises steps S101-S106:

[0031] S101, acquire multi-source running data and perform data preprocessing to obtain a standardized running state sequence;

[0032] S102, perform parameter mapping on the standardized running state sequence to construct state characteristics, and obtain an environment state vector sequence;

[0033] S103, call a multi-agent collaborative decision engine to solve a policy agent for the environment state vector sequence, and generate initial scheduling data;

[0034] S104, input the initial scheduling data into a verification model for checking and perform rule-based correction to obtain plan execution data;

[0035] S105, perform control instruction analysis, device mapping and message encoding according to the plan execution data respectively, and generate a to-be-downloaded scheduling control message for a conveying line controller;

[0036] S106, perform transmission adaptation processing on the to-be-downloaded scheduling control message, and output a scheduling control result set.

[0037] The application can carry out end-to-end closed-loop scheduling for complex topologies formed by belts, drums, elevators and split and merge nodes of a warehouse and a sorting center.

[0038] In step S101, the system deploys RFID, photoelectric, visual, weighing, speed, encoder and vibration sensors at key nodes of the conveying line, continuously collects raw data such as cargo position, speed, weight, size, line flow and equipment state, and the sensing side can clean, fuse and extract features from the collected flow, and after unifying the degree and time axis, a standardized running state sequence that can be directly consumed by the upper layer decision is obtained.

[0039] In step S102, according to the conveying line topology and equipment list, the standardized time series data is aligned to the "node-edge-station" structure (i.e. node-connection), and key quantities such as real-time flow, cargo distribution, queue length, equipment running state and order priority of each section are extracted, after time window statistics and missing interpolation, the environmental state vector sequence is organized according to the step time to meet the input requirements of subsequent strategy solving.

[0040] In an embodiment, the step S102 comprises:

[0041] According to the standardized running state sequence, semantic annotation classification is performed to obtain a type annotation sequence;

[0042] Based on the conveying line topology, each data item of the type annotation sequence is respectively mapped to the node, edge and station identifier to obtain a topology-aligned data sequence;

[0043] The topology-aligned data sequence is subjected to multi-modal processing to obtain a multi-modal state sub-vector set;

[0044] The multi-modal state sub-vector set is aligned and spliced to obtain a time window statistical feature sequence;

[0045] The time window statistical feature sequence is subjected to missing interpolation, and then continuous feature splicing to obtain a fusion feature tensor;

[0046] The fusion feature tensor is organized in time sequence as step input, and subjected to dimension rearrangement and mask construction to output an environmental state vector sequence.

[0047] In the embodiment, the first face-oriented perception layer has collected and pre-processed multi-source data (including RFID, photoelectric, visual, weighing, speed device, encoder and vibration sensor information) to establish a unified semantic and topological reference, so as to ensure that data of different modalities, different sampling granularities and different device sources can be aligned on the same time axis and the same conveying line topology, forming stable time sequence feature input. The topological structure includes but is not limited to the physical layout and connection relationship of belt machines, roller lines, sorting machines, elevators and the like. The physical layout of the conveying line can be planned in advance, and the physical position number of each conveying line device (which can also be referred to as device number on the scheduling system, and the device number is unique in the same project scene) is provided. The above-mentioned perception layer collection object and preprocessing items include: arranging RFID, photoelectric, visual, weighing, speed / encoder and vibration sensors at key nodes of the conveying line, collecting real-time cargo position, speed, weight, flow and device state, and performing data cleaning, fusion, format conversion, outlier detection and feature extraction to output structured high-dimensional perception data stream as the basis input of the step. The nodes include sorting ports, merging points, workstations (such as but not limited to unattended points or manned workstations such as unloading, stacking, picking, inventorying and the like, which have also included workstation attributes) and the like. In the semantic annotation classification stage, the system identifies the types of the standardized running state sequence based on the domain vocabulary and the preset dictionary, classifies the original fields into "position signal (such as RFID positioning, trajectory)", "passing signal (photoelectric interruption counting, interval)", "appearance and volume representation (vision, depth)", "load representation (weighing)", "motion representation (speed device, encoder)", "device health representation (vibration)", and the like semantic categories, to obtain a type annotation sequence, so as to clearly define the role and constraint boundary of each data item in subsequent feature construction. RFID is used for unique identity and trajectory identification, photoelectricity is used for counting and instantaneous speed estimation, vision is used for type, size and anomaly detection, weighing is used for real-time load, speed device, encoder is used for line speed and actual moving speed of goods, vibration is used for key component state monitoring and fault warning, and the above-mentioned semantic division and source are one-to-one correspondence.

[0048] In the topology alignment stage, the system assigns topology coordinates and device identifiers to the semantic data items according to the conveyor lines and the field topology (node-edge-station), realizing the ternary mapping from "data source-physical location-function unit": mapping the RFID read head and its coverage segment to the node, the edge (i.e. the position where each conveyor line is connected to other conveyor lines or devices, which is used to determine the direction of the device execution), aligning the photoelectric pair to the sorting port, the confluence point and the beat segment, aligning the visual camera to the key identification and sorting station, binding the weighing module to the specific weighing segment, associating the speed device, encoder and driving unit, and binding the vibration sensor to the key components such as motor and bearing. After completing the mapping, all data items are written into a unified topology index and clock stamp, and the topology alignment data sequence is output, ensuring that the subsequent cross-modal combination and statistics are performed under the same topology-time reference.

[0049] In the multi-modal processing stage, the system constructs state sub-vectors according to the topology alignment data sequence by modality: extracting displacement, dwell time and repeated scanning ratio for RFID trajectory and state sequence; extracting passing count, beat interval and instantaneous speed for photoelectric sequence; extracting type encoding, three-dimensional size approximation and abnormal event marking for visual sequence; extracting net weight, weight fluctuation and over-limit identification for weighing sequence; extracting line body speed, acceleration and deceleration events and slip indication for speed device and encoder sequence; extracting amplitude, spectral features and health score for vibration sequence. After time synchronization and scale normalization, the multi-modal state sub-vector set is formed.

[0050] In the alignment and splicing stage, a sliding time window can be used to window the multi-modal state sub-vector set for statistical processing, calculate the mean, extreme value, variance, change rate, quantile, kurtosis, skewness, event count and occupancy rate, and stack them according to node, edge, station and time sequence to obtain the time window statistical feature sequence; the window step and length are configured according to the production beat and field control period to ensure that the statistical features can cover the time scale of local congestion, fluctuation and abnormality.

[0051] In the missing interpolation and continuous feature splicing stage, the time window statistical feature sequence is processed for missing data: short-time missing is filled forward or backward and linearly interpolated, long-time missing is estimated by neighborhood topology regression or similar station migration, and abnormal values are robustly truncated or quantile backed off; then the interpolated and corrected features are continuously spliced in the time and topology dimensions to generate a fusion feature tensor organized by "time x topology x feature", which is used to carry the complete state expression before decision-making.

[0052] In the step-in input and mask construction stage, the fused feature tensor is arranged in time sequence as step-in input, and combined with the scheduling period to form equal-length state fragments. To adapt to the subsequent strategy model, the tensor is dimensionally rearranged to make the time dimension, topology dimension and feature dimension meet the expected read-in order of the model; at the same time, multi-class masks are constructed, including missing data masks (identifying the interpolation source and missing data confidence interval), shutdown masks (identifying the maintenance, disabled nodes or edges), no goods masks (identifying the channels without effective goods passing in the current window) and out-of-boundary masks (identifying the specification, load out-of-limit). The step-in input sequence after dimension rearrangement and mask injection is output as the environment state vector sequence for the multi-agent collaborative decision engine to solve the strategy.

[0053] In step S103, the multi-agent collaborative decision engine (including a scheduling agent, an optimization agent and a verification agent, and combined with LLM for strategy generation and correction) is used to solve the strategy agent for the environment state vector. The decision layer can work collaboratively with the scheduling agent, the optimization agent and the verification agent: the scheduling agent estimates and samples the strategy based on the reinforcement learning model for the current state, and outputs candidate atomic actions (such as speed adjustment, split start-stop, priority adjustment); the optimization agent combines operations research and graph theory, heuristic method, solves the minimum delay routing and split, speed coordination and refines the candidate scheme; if necessary, a large language model (LLM) is used to semantically analyze the unstructured instructions and business constraints to provide supplementary information for strategy shaping. After priority shaping, time window alignment and continuity rearrangement, the cross-section resource occupation is combined into structured initial scheduling data.

[0054] Specifically, the multi-agent collaborative decision engine is a multi-agent collaborative decision engine based on the ReAct framework. The ReAct (Reasoning and Acting) framework is an AI agent (AI Agent) design paradigm (not a specific framework project or product, but a design pattern or development philosophy category), and the design idea of the framework is to realize dynamic solution of complex tasks through alternating "reasoning" (Reasoning) and "action" (Acting) steps. The general execution process of the framework is as follows: in each iteration cycle, the ReAct agent first generates an internal reasoning trajectory based on chain-of-thought (Chain-of-Thought, abbreviated as CoT) to analyze the current state, evaluate the candidate scheme and make an action plan; then, specific actions are performed, such as calling external APIs, querying knowledge bases or operating environment interfaces (for example, this place may operate: call relevant algorithm models for decision-making, call task instructions to PLC controllers, etc.); the agent then observes the action result and feeds back the feedback to the next round of reasoning, thereby forming a closed-loop adaptive process.

[0055] In an embodiment, the step S103 comprises:

[0056] construct a Markov decision process based on the sequence of environment state vectors, to obtain a Markov decision parameter set;

[0057] uniformly process the sequence of environment state vectors by using the Markov decision parameter set, to obtain a policy input tensor;

[0058] input the policy input tensor into a scheduling agent network for estimation and policy sampling, to output a candidate atomic action sequence;

[0059] perform device quota constraint based on the candidate atomic action sequence, to obtain a candidate scheduling sequence;

[0060] perform priority shaping, time window alignment and continuity rearrangement on the candidate scheduling sequence, to generate a set of structured scheduling segments;

[0061] perform cross-segment resource occupation merging on the set of structured scheduling segments, to output initial scheduling data.

[0062] In this embodiment, the state space can be composed of real-time traffic, cargo distribution, queue length of key nodes (such as sorting ports and merging points), device status, order priority and historical congestion patterns of each segment of the conveying line; the action space includes discrete or continuous control actions such as adjusting the speed of a specific segment (acceleration, deceleration), opening or closing a diversion channel, changing the priority of goods, and pausing or starting a specific conveying segment; the reward function can adopt “shortening the in-transit time, improving the utilization rate of the conveying line, and ensuring timely delivery” as the positive target, and set penalty items for congestion, cargo damage and device failure, to obtain a cumulative return optimization target (i.e., the Markov decision parameter set) oriented to maximize throughput and minimize delay.

[0063] The sequence of environment state vectors is uniformly processed by using the Markov decision parameter set to obtain a policy input tensor. The uniform processing includes performing scale normalization, time alignment and missing data mask injection on the state dimensions of different topological positions and time slices, and arranging the tensor into equal-length step segments according to the decision period; for unstructured special instructions from the operator (such as temporary urgent, special pallet processing), perform intent recognition and semantic translation by LLM to solidify them into computable priority prompts, forbidden or must-arrive constraints or time window restrictions, which are spliced into the policy input tensor as additional context features, to ensure that the policy solving has the ability to perceive business side constraints.

[0064] Furthermore, the policy input tensor is input into the scheduling agent network (such as DQN) for estimation and policy sampling, outputting candidate atomic action sequences. The scheduling agent employs a deep reinforcement learning (DRL) algorithm, preferentially selecting a deep Q-network (DQN) or its variants (such as Double DQN, Dueling DQN, priority experience replay, etc.), calculating action-value estimates for each step segment and using trade-offs for policy sampling to obtain candidate atomic action sequences including speed adjustment, traffic splitting, priority adjustment, and segment-level start / stop control. In terms of training, offline pre-training is performed using historical and simulation data, followed by online small-batch fine-tuning based on real-time feedback to continuously adapt to on-site dynamics. Historical and simulation data include historical equipment feedback data, historical algorithm weight parameter information, and efficiency comparisons between historical feedback and post-tuning. Based on real-time operational data feedback, the weight parameter data can be automatically and adaptively fine-tuned, continuously approaching the optimal solution. Based on the candidate atomic action sequences, equipment quota constraints are applied to obtain candidate scheduling sequences. The constraints include at least: concurrent quotas, bandwidth and cycle time limits per device / channel / PLC cycle, safety intervals and occupancy conflict constraints for critical nodes, and derating operation boundaries triggered by equipment health status. For cross-segment coupled resource contention, the scheduling agent and optimization agent collaborate to implement capacity constraints and feasible region pruning on candidate atomic actions to obtain candidate scheduling sequences that meet physical and process constraints. The scheduling agent network draws inspiration from Deep Q-Network (DQN), with the specific formula as follows:

[0065] ;

[0066] Where Q represents a function, and the reward R is... t =α·throughput increment - β·delay penalty, where α and β are weights, throughput increment represents the increase in the number of goods passing through the conveyor line (positive reward), and delay penalty represents the time it takes for goods to queue through the conveyor line or the deviation in transit time (negative penalty). The weight parameters are custom values ​​and can be balanced through optimization.

[0067] Further, the candidate scheduling sequence is priority shaped, time window aligned and continuity rearranged to generate a structured scheduling segment set. The priority shaping is hierarchically ordered according to order urgency, estimated time of arrival (ETA) and service SLA; the time window alignment is boundary adsorbed and aligned according to control period and device response time to ensure that the message issuing and physical execution beats are consistent; the continuity rearrangement is used to eliminate unnecessary switching and jitter, and to ensure that the control actions on the same topology path have the minimum switching number and reasonable residence time, so that the structured scheduling segment set is organized according to the time slice-topology segment-control instruction. The structured scheduling segment set is merged across the segment resource occupation to output the initial scheduling data. During the merging process, the system merges and resolves conflicts of the resource occupation of adjacent time slices and connected topology segments to ensure that the actions across the distribution / flow nodes are globally consistent under the capacity and safety constraints; for the segments involving multi-device linkage, the cooperative speed adjustment and path consistency verification are calculated to obtain the initial scheduling data that can directly enter the subsequent verification link, realizing the complete strategy generation closed loop from estimation-sampling-trimming-shaping-merging.

[0068] In an embodiment, the step S103 further comprises:

[0069] A multi-objective evaluation vector is constructed using the environment state vector sequence, and a multi-objective optimization model is obtained.

[0070] A weighted flow network is established based on the multi-objective optimization model, and the environment state vector sequence is mapped to corresponding path variables, distribution variables and speed variables to obtain a graph model instance.

[0071] The graph model instance is subjected to path set screening and delay weight calculation respectively to obtain a candidate path set.

[0072] The candidate path set and the multi-objective optimization model are jointly established to solve a linear programming problem, and a parameter optimization solution is obtained.

[0073] The parameter optimization solution is subjected to heuristic iterative fine-tuning, and local recalculation is performed according to local congestion indication to obtain an iterative optimization solution.

[0074] The iterative optimization solution is mapped to an optimized candidate scheduling sequence, and is time-aligned with the environment state vector sequence to output initial scheduling data.

[0075] In this embodiment, according to the business SLA and the on-site operation strategy, delay targets (in-transit time, waiting queue time), throughput targets (throughput per unit time), energy consumption targets (estimated impact of speed start-stop on energy consumption), and switching costs (penalty terms for frequent start-stop and shunting switching) are defined; normalization and weight setting are performed on each target item (weights can be given by historical statistics or policy configuration), so as to obtain a multi-objective evaluation vector for solving, and a multi-objective optimization model equivalent to "weighted summation" or "epsilon-constraint" is established accordingly. Then, a weighted flow network is established based on the multi-objective optimization model, and the environment state vector sequence is mapped to corresponding path variables, shunting variables, and speed variables, to obtain a graph model instance. Taking the conveying line topology as the graph base, the workstation equipment as the node, and the conveying section as the directed edge, attributes such as capacity, nominal speed, energy consumption coefficient, and safety interval are added to the edge; the demand in the scheduling period and the cargo flow demand are mapped to source-sink pairs; path variables are defined to select candidate routes, shunting variables are used to allocate flow among parallel branches, and speed variables are used to give coordinated speed suggestions on adjustable speed edges; meanwhile, capacity constraints, edge occupation conflict constraints, and minimum or maximum speed boundaries are introduced, to obtain a graph model instance with a physically feasible region.

[0076] Further, path set screening and delay weight calculation are performed on the graph model instance respectively, to obtain a candidate path set. Preferably, a multi-objective variant of Dijkstra or A* is used to calculate K-shortest paths under the combined weight of "delay-energy consumption-switching cost"; for each candidate path, expected edge delay (including the influence of congestion coefficient and speed setting) and energy consumption cost are estimated according to the environment state vector sequence, and the reachability and safety interval of the path across shunting or merging nodes are checked to exclude infeasible or inferior paths, and the candidate path set that meets the capacity or safety constraints is retained.

[0077] On this basis, the candidate path set and the multi-objective optimization model are combined to establish a linear programming (or multi-objective linear programming) solving problem, to obtain a parameter optimization solution. The objective function can be a multi-objective weighted summation minimization: ·total delay + ·energy consumption + ·switching cost− • Throughput; constraints include: ① flow conservation and source-sink satisfaction; ② edge capacity and concurrent quota constraints; ③ device start-stop cadence and safety interval constraints; ④ coupling constraints of speed variables and path selection variables (e.g. linearized Big-M constraints). The solution is a parameter optimization solution of path selection, split proportion, and speed setting within the current time window. Then the parameter optimization solution is iteratively fine-tuned heuristically, and local recalculation is performed according to local congestion indicators, to obtain an iterative optimization solution. Specifically: use genetic algorithm or ant colony algorithm to perturb the split proportion and speed setting at a fine granularity, evaluate the improvement of the objective value and retain the superior individuals; when a local congestion indicator (such as the queue length or utilization of a certain edge exceeding the threshold) is monitored, trigger incremental recalculation of the affected subgraph, and only reselect K-shortcuts and redistribute the split proportion in the local area, so as to achieve rapid congestion relief at a lower computational cost. The above process is aligned with the scheduling period until the objective function converges or the set number of iterations is reached, and the iterative optimization solution is output. The specific formula of multi-objective linear programming is as follows:

[0078] ;

[0079] Where (x) represents the decision variable vector, and the variables include path variables, split variables, and speed variables; f i (x) represents the i-th objective function, f1(x) represents the delay, and f2(x) represents the energy consumption; w i represents the weight coefficient, A represents the constraint matrix, and b represents the constraint right-hand side vector (capacity constraint, device quota boundary). The multi-objective problem can be converted into a single-objective linear programming by referring to this formula.

[0080] Finally, the iterative optimization solution is mapped to the optimized candidate scheduling sequence, and is time-aligned with the environmental state vector sequence to output the initial scheduling data. The mapping rules include: translating path variables into specific split start-stop and routing instructions, translating split variables into traffic or release quotas for each branch, and translating speed variables into executable segment speed settings; performing boundary adsorption and cadence correction on adjacent time slices, merging consecutive control segments on the same route, forming a structured scheduling segment of "time slice-topological segment-control instruction", and aligning with the perception side clock stamp to output the initial scheduling data for verification agent and regularization module processing.

[0081] In step S104, the verification agent performs parallel simulation evaluation on the initial scheme and implements multi-model voting, outputs consistent evaluation results for segments that may cause congestion or violations; combines the safety process rule library to impose regularization correction on abnormal segments, and at the same time the LLM performs semantic translation on temporary job requests or special pallet handling requirements involving natural language and incorporates them into the constraint set, and then performs constraint synthesis and update on the initial scheduling data to output plan execution data that can be directly issued.

[0082] In an embodiment, the step S104 comprises:

[0083] A verification input object is constructed using the initial scheduling data to obtain a verification input set;

[0084] The verification input set is input into corresponding model parallel simulation to generate a multi-model evaluation result set;

[0085] Consistency metrics are calculated using the multi-model evaluation result set to obtain a consistency evaluation report;

[0086] The consistency evaluation report is input into a regularization correction module for processing to obtain a correction rule set;

[0087] Based on the correction rule set, a large language model is called to perform semantic analysis and policy text translation on unstructured requests, and is merged with the correction rule set to obtain a correction constraint set;

[0088] The initial scheduling data is constrained and synthesized using the correction constraint set to update the output plan execution data.

[0089] In this embodiment, a verification input object is constructed using initial scheduling data to obtain a verification input set. Specifically, the time window, topology segment, control instruction (such as segment speed setting, shunt start-stop, priority adjustment), and its boundary condition (capacity, concurrency quota, safety interval, device health status) of the scheduling segment are extracted; at the same time, the environmental state vector slice in the same window is loaded as the scene context for driving simulation and prediction. The verification input set completely expresses the ternary relationship of “scheme-environment-constraint”, providing a unified input structure for subsequent parallel evaluation. The verification input set is then input into corresponding model parallel simulation to generate a multi-model evaluation result set. The verification agent maintains multiple complementary evaluators, i.e., discrete event simulation for replaying queue evolution and in-transit time; rule-based checker for fast screening of strong constraints (such as hard boundaries of capacity, beat, safety interval); and prediction model for estimating congestion rate, delay distribution, and energy consumption trend. Each model runs independently and outputs key indicators (such as: average waiting time, node utilization, congestion probability, energy consumption estimate, switching frequency) to obtain a multi-model evaluation result set. The multi-model evaluation result set is then used to calculate consistency metrics to obtain a consistency evaluation report. The outputs of each evaluator can be measured for consistency according to the principle of “multi-model voting + threshold consistency”: when the core indicators such as congestion rate and delay time are within the safety threshold and the cross-model difference is within the tolerance interval, it is recorded as consistent; if there is a significant deviation, the topology segment and time window where the divergence occurs are located and the inconsistency reason is marked. The report also lists the hard constraint hit conditions and risk levels triggered, which are used to guide subsequent regularization correction.

[0090] Further, the consistency evaluation report is input into a regularization correction module for processing to obtain a correction rule set. The regularization correction module generates structured correction suggestions for the inconsistent or out-of-bound scheduling segments according to the compliance class hard constraints and the experiential soft constraints, such as reducing the local segment speed, delaying the release time window, limiting the concurrent quota, replacing the path passing through the congested node, and the like, and outputs the correction rule set in the form of machine-readable rule entries (condition-action-range- effective time window). Subsequently, based on the correction rule set, a large language model is called to perform semantic analysis and policy text translation on the unstructured request, and the correction rule set is merged to obtain a correction constraint set. The unstructured request can include an operator's natural language operation instruction, such as a context such as "the tray number XXX needs to be adjusted to the highest priority", and a voice instruction, etc. For example, the whisper voice model is used to process the operator's voice input to convert it into natural language text, and the natural language is converted into an unstructured text request to the large language model, thereby continuing to form a closed loop. For the temporary task, special cargo handling requirement or abnormal disposal instruction (natural language description) from the operator, the LLM performs context understanding and intent recognition to convert it into structured constraints such as priority adjustment, must-reach or forbidden path, time-limited arrival, maximum switching times, and the like, and performs conflict detection and merging with the correction rule set to output a unified correction constraint set, ensuring that the business side intent is accurately and executable expressed at the scheduling end.

[0091] Finally, the initial scheduling data is constrained and synthesized using the correction constraint set to update the output plan execution data. The constraint synthesis process checks and executes the initial scheduling segments one by one, and forcibly rewrites or replaces the segments that do not meet the hard constraints; the soft constraints are fine-tuned in the form of cost shaping to adjust the priority and time window; the capacity is redistributed and the beat is aligned for the cross-segment resource occupation to ensure global feasibility and beat consistency. After the synthesis is completed, the plan execution data consistent with the field control cycle is formed, which can be directly encoded as control messages as input for the subsequent execution layer.

[0092] In step S105, the system parses the plan execution data into instruction units for the PLC channel, completes device identification matching and address mapping; constructs a load structure and serializes the code according to the control theme and interface specification to obtain a control message frame set; finally, the PLC channel is merged and grouped to obtain a compatible dispatch control message to be issued for the field control network, realizing accurate mapping from the strategy to the physical execution interface.

[0093] In an embodiment, the step S105 comprises:

[0094] Based on the plan execution data parsing, the corresponding control target is parsed and the instruction granularity is divided to obtain an instruction unit sequence;

[0095] According to the PLC controller channel, the instruction unit sequence is respectively subjected to equipment identification matching and address label mapping, to obtain an equipment mapping instruction table;

[0096] The equipment mapping instruction table is subjected to subject mapping, to obtain a protocol mapping table;

[0097] According to the protocol mapping table, a control payload is constructed, and the control payload is organized into a load structure set according to an interface specification;

[0098] The load structure set is subjected to serialization encoding, to obtain a control message frame set;

[0099] According to the PLC controller channel, the control message frame set is merged and grouped, to output a dispatch control message to be issued for a conveyor controller.

[0100] In the embodiment, the structured scheduling segments in the plan execution data are expanded piece by piece, and a topology segment identifier, a control type and a target parameter are extracted, and abstract actions such as "segment speed setting", "split start and stop", "path switching", "lifting positioning" and the like are refined into single and atomic control targets; in combination with a field control beat and a device response time, adjacent control segments are subjected to boundary adsorption and time alignment, to generate an instruction unit sequence meeting a field cycle, to provide an explicit control semantic and a time sequence anchor point for subsequent message generation. The analysis ensures that subsequent issuance can drive the PLC to perform fine-grained dynamic adjustment on speed, split and path, thereby serving the congestion avoidance and flow balancing goals. The channel configuration is re-searched, a unique equipment identifier and an address label (such as a register, a coil, a data block index) of the equipment in the corresponding PLC are allocated for each instruction unit, and necessary safety and interlock flag bits are simultaneously loaded. When the planning is completed, the physical space position of the equipment (for example, derived from a CAD drawing or the like) is marked with a unique number, which is the default equipment unique identifier. This part of data is an information set that can be obtained before the system is implemented. For segments involving multi-device linkage, an occupation conflict check and a capacity boundary check are performed, to ensure that the mapped instructions meet the device concurrent quota and safety interval constraints. After the mapping is completed, an equipment mapping instruction table of "equipment-address-action-parameter-time window" is formed, as a direct input of protocol layer encapsulation. According to field integration selection, a standardized communication protocol and a subject, a node and an endpoint identifier are selected for different PLC channels: when a message type channel is adopted, the mapping is an MQTT subject hierarchy; when an industrial internet channel is adopted, the mapping is an OPC UA node path; when an HTTP, gateway channel is adopted, the mapping is a RESTful endpoint and a resource path. The subject mapping maintains a discriminable structure of "control object-action type-site line-time slice", and records transmission metadata such as quality of service, timeout and retry strategy, to obtain a protocol mapping table, to support subsequent payload construction and orderly issuance.

[0101] Further, according to the protocol mapping table, the control payload is constructed, and the control payload is organized into a load structure set according to the interface specification. For speed setting type actions, the load body at least contains device address, target speed, effective time window and out-of-bound processing strategy; for split start-stop and path switching type actions, the load body contains split identification, switch state, target path, minimum holding time and switching cost identification; for lifting positioning type actions, the load body contains target layer and position tolerance. All load bodies are uniformly attached with time stamp, sequence number and check field, and strictly follow the data model and field constraints of the selected protocol, and the load structure set is obtained by combination, so as to ensure that the instructions can be reliably executed by the built-in logic of PLC after reaching PLC. For MQTT / REST channels, serialization can be performed according to the agreed JSON and binary structure; for OPC UA channels, encoding and batch packaging are performed according to the node writing model. The encoding process maintains the time sequence order and idempotent identification to avoid repeated execution; for multi-frame instructions of cross-device linkage, the same batch mark is set to ensure the arrival order and execution atomicity, and finally the control message frame set which can be directly transmitted into the network is generated. Finally, the control message frame set is merged and grouped according to the PLC controller channel to output the dispatch control message to the conveyor controller. The merging strategy takes "channel-site-control cycle" as the primary key, and combines multiple frames under the same cycle and the same PLC channel into a batch for delivery; for messages with sequential dependencies, sequence relationships are added and confirmation or retransmission strategies are configured. The merged dispatch control message is delivered to the PLC through the standardized API, and the PLC converts the digital instructions into electrical signals for driving motors, splitters, mergers and elevators, realizes dynamic and accurate adjustment of line parameters, and returns feedback information such as actual speed, split state and response time after execution for subsequent closed-loop control and continuous learning.

[0102] In step S106, the message is transmitted and adapted and then delivered to the PLC and the actuator; the actual speed, split state, device response time and other execution feedback generated by the execution layer are returned to the perception layer and the feedback layer, and are bound with the original scheduling record; the system writes the key feedback elements into the vector database to establish a similarity index, which is used for subsequent retrieval and experience playback, and calculates the structured performance items (such as in-transit time, throughput, congestion rate and energy consumption), associates the delivery record and outputs the dispatch control result set, closing the self-evolution loop of "perception-decision-execution-feedback". It should be noted that the PLC controller or other related similar controllers control the physical logistics conveyor line, and the dispatch control result set is delivered to the corresponding control unit for execution.

[0103] In an embodiment, the step S106 comprises:

[0104] A transmission adaptation object is established to identify and configure a correlation key for message metadata in the to-be-issued scheduling control message, and generate an issued record for result binding;

[0105] Actual operation data is obtained according to the issued record, and an execution feedback original set is obtained by aggregation;

[0106] Required interaction parameters in the execution feedback original set are extracted to obtain a feedback element set;

[0107] The feedback element set is written into a vector database and a similarity retrieval index is established to obtain a feedback vector index set;

[0108] Based on the feedback vector index set, a cargo parameter index is calculated to obtain a structured result item;

[0109] The structured result item is associated with the issued record to obtain a scheduling control result set.

[0110] In this embodiment, a transmission adaptation object is established to identify and configure a correlation key for message metadata of a to-be-issued scheduling control message, and generate an issued record for result binding. For each batch of issued control message, the channel, station, control period, topological section device mapping, instruction type and timestamp and other key information are recorded, so as to correspond one by one with the subsequent execution feedback, and realize the traceable binding relationship from “message-device-time window”. The issued record serves as an anchor point for result aggregation and evaluation, and penetrates through the execution layer return and the feedback layer database storage process. Actual operation data is obtained according to the issued record, and an execution feedback original set is obtained by aggregation. The execution layer returns the key operation data such as actual speed, shunting state and device response time after the instruction is landed; the system extracts and aggregates according to the time window and topological position indicated by the issued record to obtain an execution feedback original set corresponding to the message batch, which is used for subsequent data cleaning, evaluation and database storage. The return mechanism ensures that the instruction execution effect can enter the feedback link in real time, supporting closed-loop control.

[0111] Further, the required interaction parameters in the feedback original set are extracted to obtain a feedback element set. The extracted elements include at least: in-transit time, throughput, congestion rate of each segment, equipment utilization rate, energy consumption, and abnormal failure markers, and other key performance indicators (KPIs) closely related to scheduling quality, and the context bound to the original message (decision segment, action type, device topology index) is retained. Through the element extraction process, standardized and structured representations are provided for subsequent similarity retrieval and strategy review. Then, the feedback element set is written into a vector database and a similarity retrieval index is established to obtain a feedback vector index set. The system stores the features processed by the sensor, scheduling action records, execution results, reward signals, and necessary LLM interaction records in the form of high-dimensional vectors, and constructs a similarity index based on a nearest neighbor search structure, which is used to quickly retrieve similar scenarios and successful handling strategies in history when a new congestion pattern or abnormal scenario occurs, providing data support for experience replay and case learning for reinforcement learning and strategy optimization.

[0112] On this basis, the cargo parameter indicators are calculated based on the feedback vector index set to obtain structured result entries. The cargo parameter indicators include cargo in-transit time deviation, congestion probability, etc. Combined with the similar historical entries indexed and the current feedback elements, the key parameters (such as cargo in-transit time and its deviation, path segment delay and congestion probability, unit throughput, energy consumption estimation, and equipment utilization, etc.) are calculated and stored for a single cargo, a single time window, and a single topology segment, and structured result entries that can be audited are generated for direct consumption by operation and maintenance reports, alarm threshold evaluation, and subsequent model training. Finally, the structured result entries are data-associated with the issued records to obtain a scheduling control result set. The system uses the message batch and the device topology index as the key to complete the closed-loop binding of “plan instruction-feedback evaluation indicators”, and forms a result set output oriented to multiple perspectives (time, device, path, order); the result set is written back to the feedback layer to provide a continuous data source for offline training, online fine-tuning, and model optimization of AI Agent, and to promote the system to achieve self-adaptation and self-evolution.

[0113] The application is based on a multi-agent automated logistics conveying line scheduling optimization system that operates according to the closed-loop mechanism of "perception-decision-execution-feedback-self-evolution", which corresponds to steps S101-S106 in the foregoing technical solution one by one, ensuring that the whole process from data collection to instruction landing and then to effect backflow is computable, verifiable, and traceable. The system first uniformly collects and pre-processes multi-source running data at the perception layer to form a standardized running state sequence that can be directly consumed by algorithms; then maps it to an environment state vector sequence at the decision layer to drive multi-agent collaboration to generate executable scheduling instructions; then converts abstract instructions into physical control of devices through standardized APIs and PLC controllers at the execution layer; finally, the feedback layer precipitates the execution effect in a structured manner and supports continuous learning, forming a stable cycle of "data-strategy-control-evaluation-optimization".

[0114] In the perception link, IoT sensors (such as RFID, photoelectric, vision, weighing, etc.) collect real-time data such as cargo location, speed, weight, size, conveying line flow, and device status, etc. After cleaning, fusion, and feature extraction, structured perception data is obtained. This data is aligned on the time axis and topological coordinates (nodes, edges, stations), and input into the parameter mapping process to construct an environment state vector sequence that can fully represent the current field operation situation, providing accurate and timely input for subsequent intelligent decision-making.

[0115] In the intelligent decision-making and collaboration link, the scheduling agent evaluates the environment state and samples strategies based on a reinforcement learning model, giving a preliminary scheduling scheme aimed at improving throughput and reducing delay; the optimization agent combines operations research, graph theory, and heuristic algorithms to globally refine and weigh path selection, shunting ratio, and segment speed setting; the verification agent uses multi-model parallel simulation and voting mechanisms to check safety, feasibility, and constraint consistency; when there are unstructured business requests or abnormal scenarios, the LLM analyzes natural language intent and translates strategies, allowing business-side constraints to be injected into the decision-making link in a structured form. The final output is plan execution data that can directly enter the control link.

[0116] In the instruction execution and control link, plan execution data is transmitted to PLC controllers through standardized APIs (such as MQTT, OPC UA, RESTful API), and is converted into specific electrical signals for motor, shunt, merger, elevator, and other actuators, achieving dynamic and accurate adjustment of speed, direction, and shunt switch. Actual speed, shunting state, device response time, and other information generated during execution are synchronously fed back, providing a basis for closed-loop control and rapid handling of abnormalities.

[0117] In the effect feedback and continuous learning stage, the system writes the execution results and their key performance indicators (such as in-transit time, throughput, congestion rate, equipment utilization, energy consumption, etc.), together with the original perception data, policy decision records, and reward signals, into a vector database and establishes a similarity index. On the one hand, this enables the system to quickly retrieve similar situations in history and their successful strategies as "experience replay" when encountering new congestion or abnormal patterns; on the other hand, it also provides a unified and reusable data base for offline training, online fine-tuning, and strategy optimization of the model.

[0118] In the self-evolution stage, the AI Agent of the decision layer triggers retraining periodically or when performance degrades or new patterns appear: offline large-scale training using accumulated data to obtain a more optimal global strategy; online small-batch fine-tuning using incremental data to quickly adapt to short-term disturbances and emergencies. With the above closed-loop and learning mechanism, the system continuously improves scheduling efficiency and robustness in long-term operation, and stably achieves the goals of autonomy and flexibility of production.

[0119] The system of the present application has a wide range of application scenarios. In the automated logistics integration scenario, for complex conveyor topologies in large warehouses, intelligent factories, and sorting centers, the system can realize seamless collaboration and global scheduling between different conveyor equipment and sorting equipment, optimize the overall process, and improve material turnover efficiency and throughput. In the e-commerce warehouse and distribution center scenario, the system can cope with the business characteristics of large order volume, obvious peak and valley, and strong timeliness, support fast sorting, precise merging, and efficient outbound, shorten the fulfillment cycle, and improve service quality. In the manufacturing production line material distribution scenario, the system can intelligently allocate according to production rhythm, real-time demand, and workstation inventory, ensure "just-in-time" delivery of materials, reduce work-in-process occupation, and maintain continuous and efficient operation of the production line.

[0120] In summary, the application constructs a closed-loop system of "perception-decision-execution-feedback-self-evolution", deeply integrates multi-source high-precision perception data and multi-agent collaborative decision-making, and can realize unmanned and autonomous scheduling in a complex and dynamic conveying environment. Specifically: first, through the linkage of the ReAct framework of the AI Agent and the reinforcement learning model, the system can automatically plan the path, adjust the flow and segment speed under the driving of real-time data, continuously learn and iteratively optimize, significantly improve the throughput, reduce the in-transit and queuing delay, and reduce the cost of manual intervention and rule maintenance; second, with the division of labor and cooperation architecture composed of "scheduling agent-optimization agent-verification agent-LLM assistance", the global fine-tuning of the preliminary scheme is carried out by using operational research, graph theory and heuristic algorithm, and multi-model voting verification and safety constraints are used to make the decision more comprehensive and reliable, and to avoid local optimization and incorrect issuance; third, the LLM is introduced to perform semantic analysis and strategy translation on unstructured business requests and abnormal descriptions, enhance the explainability of the strategy and the human-machine cooperation ability, so that the scenes such as temporary urgent and special goods handling can also be quickly and accurately implemented; fourth, the high-precision perception and standardized preprocessing formed by the IoT sensors such as RFID, vision and photoelectricity ensure the accurate and timely state representation, and realize the millisecond-level response and precise control of congestion clues; fifth, the vector database is used to deposit the full-link data of "perception-strategy-execution-reward", support fast retrieval and experience playback of similar scenes, speed up model training and online fine-tuning, and ensure the performance to be continuously improved over time; sixth, the standardized API and PLC are deeply integrated to stably map the abstract strategy to the executable control message, flexibly adjust the speed, path and flow strategy, and balance efficiency, energy consumption and equipment safety.

[0121] In combination Figure 2 As shown in the accompanying drawings, Figure 2 A schematic block diagram of an automatic logistics conveying line scheduling optimization device provided by an embodiment of the application is shown in the accompanying drawings. The automatic logistics conveying line scheduling optimization device 200 comprises:

[0122] A data processing unit 201 is configured to acquire multi-source operation data and perform data preprocessing to obtain a standardized operation state sequence.

[0123] A parameter mapping unit 202 is configured to perform parameter mapping on the standardized operation state sequence to construct state features and obtain an environment state vector sequence.

[0124] A strategy scheduling unit 203 is configured to call a multi-agent collaborative decision-making engine to solve a strategy agent for the environment state vector sequence and generate initial scheduling data.

[0125] A data verification unit 204 is configured to input the initial scheduling data into a verification model for verification and perform rule-based correction to obtain plan execution data.

[0126] The message generation unit 205 is configured to perform control instruction analysis, device mapping and message coding according to the plan execution data respectively, and generate a to-be-downloaded scheduling control message for a conveying line controller.

[0127] The scheduling output unit 206 is configured to perform transmission adaptation processing on the to-be-downloaded scheduling control message, and output a scheduling control result set.

[0128] In the embodiment, the data processing unit 201 acquires multi-source operation data and performs data preprocessing to obtain a standardized operation state sequence; the parameter mapping unit 202 performs parameter mapping on the standardized operation state sequence to construct state features, and obtains an environment state vector sequence; the strategy scheduling unit 203 calls a multi-agent collaborative decision engine to solve a strategy agent for the environment state vector sequence, and generates initial scheduling data; the data verification unit 204 inputs the initial scheduling data into a verification model for verification and rule-based correction, and obtains plan execution data; the message generation unit 205 performs control instruction analysis, device mapping and message coding according to the plan execution data respectively, and generates a to-be-downloaded scheduling control message for a conveying line controller; and the scheduling output unit 206 performs transmission adaptation processing on the to-be-downloaded scheduling control message, and outputs a scheduling control result set.

[0129] In an embodiment, the data processing unit 201 is specifically configured to:

[0130] perform semantic annotation and classification on the standardized operation state sequence to obtain a type annotation sequence;

[0131] map each data item of the type annotation sequence to a node, an edge and a station identifier based on a conveying line topology, and obtain a topology-aligned data sequence;

[0132] perform multi-modal processing on the topology-aligned data sequence to obtain a multi-modal state sub-vector set;

[0133] align and splice the multi-modal state sub-vector set to obtain a time window statistical feature sequence;

[0134] perform missing value interpolation on the time window statistical feature sequence, and then perform continuous feature splicing to obtain a fusion feature tensor;

[0135] organize the fusion feature tensor into step-by-step input in chronological order, and perform dimension rearrangement and mask construction to output the environment state vector sequence.

[0136] In an embodiment, the strategy scheduling unit 203 is specifically configured to:

[0137] construct a Markov decision process based on the environment state vector sequence, and obtain a Markov decision parameter set.

[0138] uniformize the environment state vector sequence by using the Markov decision parameter set, to obtain a policy input tensor;

[0139] input the policy input tensor into a scheduling agent network for estimation and policy sampling, to output a candidate atomic action sequence;

[0140] perform device quota constraint based on the candidate atomic action sequence, to obtain a candidate scheduling sequence;

[0141] perform priority shaping, time window alignment and continuity rearrangement on the candidate scheduling sequence, to generate a set of structured scheduling segments;

[0142] perform cross-segment resource occupation merging on the set of structured scheduling segments, to output initial scheduling data.

[0143] In an embodiment, the policy scheduling unit 203 is further specifically configured to:

[0144] construct a multi-objective evaluation vector by using the environment state vector sequence, to obtain a multi-objective optimization model;

[0145] establish a weighted flow network based on the multi-objective optimization model, and map the environment state vector sequence into corresponding path variables, shunt variables and speed variables, to obtain a graph model instance;

[0146] perform path set screening and delay weight calculation on the graph model instance respectively, to obtain a candidate path set;

[0147] jointly establish a linear programming solving problem for the candidate path set and the multi-objective optimization model, to obtain a parameter optimization solution;

[0148] perform heuristic iterative fine-tuning on the parameter optimization solution, and perform local recalculation according to a local congestion indication, to obtain an iterative optimization solution;

[0149] map the iterative optimization solution into an optimized candidate scheduling sequence, and perform time sequence alignment with the environment state vector sequence, to output initial scheduling data.

[0150] In an embodiment, the data verification unit 204 is specifically configured to:

[0151] construct a verification input object by using the initial scheduling data, to obtain a verification input set;

[0152] input the verification input set into corresponding model parallel simulation running respectively, to generate a multi-model evaluation result set;

[0153] calculate consistency measure by using the multi-model evaluation result set, to obtain a consistency evaluation report;

[0154] inputting the consistency evaluation report into a regularized correction module for processing to obtain a correction rule set;

[0155] invoking a large language model based on the correction rule set to perform semantic analysis and policy text translation on the unstructured request, and merging with the correction rule set to obtain a correction constraint set;

[0156] constraining and synthesizing the initial scheduling data using the correction constraint set to update and output plan execution data.

[0157] In an embodiment, the message generation unit 205 is specifically configured to:

[0158] based on the plan execution data analysis corresponding control target, and performing instruction granularity segmentation to obtain an instruction unit sequence;

[0159] According to the PLC controller channel, the instruction unit sequence is respectively matched with the device identifier and mapped with the address label to obtain a device mapping instruction table;

[0160] mapping the device mapping instruction table to obtain a protocol mapping table;

[0161] According to the protocol mapping table, a control payload is constructed, and the control payload is organized into a load structure set according to the interface specification;

[0162] serializing and encoding the load structure set to obtain a control message frame set;

[0163] According to the PLC controller channel, the control message frame set is merged and grouped to output a to-be-downloaded scheduling control message for the conveyor controller.

[0164] In an embodiment, the scheduling output unit 206 is configured to:

[0165] establishing a transmission adaptation object to identify and associate key configuration of message metadata in the to-be-downloaded scheduling control message, and generating a downlink record for result binding;

[0166] According to the downlink record, actual operation data is obtained, and an execution feedback original set is obtained;

[0167] extracting the required interaction parameters in the execution feedback original set to obtain a feedback element set;

[0168] write the feedback element set into a vector database and establish a similarity retrieval index to obtain a feedback vector index set;

[0169] Based on the feedback vector index set, a cargo parameter index is calculated to obtain a structured result item;

[0170] The structured result entry is data-associated with the issuing record to obtain a scheduling control result set.

[0171] Since the embodiments of the device part correspond to the embodiments of the method part, the embodiments of the device part are described in the description of the embodiments of the method part, which will not be described here.

[0172] The embodiments of the present application also provide a computer readable storage medium, which has a computer program stored thereon, and the computer program can implement the steps provided by the above embodiments when executed. The storage medium can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various storage medium capable of storing program codes.

[0173] The embodiments of the present application also provide a computer device, which can include a memory and a processor, the memory has a computer program stored therein, and the processor can implement the steps provided by the above embodiments when calling the computer program in the memory. Of course, the computer device can also include various network interfaces, power supplies, graphics card devices, etc., and the performance of the graphics card can be used to operate the model, such as inference and training.

[0174] The embodiments in the description are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts are referred to the method part. It should be pointed out that, for ordinary skilled in the art, without departing from the principles of the present application, some improvements and modifications can be made to the present application, and these improvements and modifications also fall within the protection scope of the claims of the present application.

[0175] It should also be noted that in the specification, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

Claims

1. An automated logistics conveyor line dispatch optimization method, characterized by, The method comprises the following steps: acquiring multi-source operation data and performing data preprocessing to obtain a standardized operation state sequence; performing parameter mapping on the standardized operation state sequence to construct state features and obtain an environment state vector sequence, including: performing semantic labeling classification on the standardized operation state sequence to obtain a type labeling sequence; mapping each data item of the type labeling sequence to a node, an edge, and a station identifier based on the conveyor line topology to obtain a topology alignment data sequence; performing multi-modal processing on the topology alignment data sequence to obtain a multi-modal state sub-vector set; aligning and splicing the multi-modal state sub-vector set to obtain a time window statistical feature sequence; performing missing value interpolation on the time window statistical feature sequence, and then performing continuous feature splicing to obtain a fusion feature tensor; organizing the fusion feature tensor in chronological order as a step input, and performing dimension rearrangement and mask construction to output the environment state vector sequence; calling a multi-agent collaborative decision engine to solve the policy agent for the environment state vector sequence to generate initial scheduling data; inputting the initial scheduling data into a verification model for checking and rule-based correction to obtain plan execution data; performing control instruction analysis, device mapping, and message encoding based on the plan execution data to generate a to-be-downloaded scheduling control message for the conveyor line controller, including: analyzing the corresponding control target based on the plan execution data and performing instruction granularity segmentation to obtain an instruction unit sequence; performing device identifier matching and address label mapping on the instruction unit sequence based on the PLC control channel to obtain a device mapping instruction table; performing theme mapping on the device mapping instruction table to obtain a protocol mapping table; constructing a control payload based on the protocol mapping table, and organizing the control payload into a load structure set according to the interface specification; serializing and encoding the load structure set to obtain a control message frame set; grouping the control message frame set according to the PLC control channel to output the to-be-downloaded scheduling control message for the conveyor line controller; performing transmission adaptation processing on the to-be-downloaded scheduling control message to output a scheduling control result set, including: establishing a transmission adaptation object to configure identification and association keys for the message metadata in the to-be-downloaded scheduling control message to generate a downlink record for result binding; acquiring actual operation data based on the downlink record to obtain an execution feedback original set; extracting required interaction parameters in the execution feedback original set to obtain a feedback element set; writing the feedback element set into a vector database and establishing a similarity retrieval index to obtain a feedback vector index set; calculating a cargo parameter index based on the feedback vector index set to obtain a structured result item; associating the structured result item with the downlink record to obtain the scheduling control result set.

2. The automated logistics conveyor line dispatch optimization method of claim 1, wherein, The calling of the multi-agent collaborative decision engine to solve the policy agent for the environment state vector sequence to generate the initial scheduling data comprises: constructing a Markov decision process based on the environment state vector sequence to obtain a Markov decision parameter set; Uniformize the environment state vector sequence by using the Markov decision parameter set to obtain a policy input tensor; Input the policy input tensor into a scheduling agent network for estimation and policy sampling to output a candidate atomic action sequence; Perform device quota constraint based on the candidate atomic action sequence to obtain a candidate scheduling sequence; Perform priority shaping, time window alignment, and continuity rearrangement on the candidate scheduling sequence to generate a set of structured scheduling segments; Merge the set of structured scheduling segments across segment resource occupancy to output initial scheduling data.

3. The method of claim 1, wherein, The calling multi-agent collaborative decision engine for policy agent solving of the environment state vector sequence to generate initial scheduling data further includes: Construct a multi-objective evaluation vector using the environment state vector sequence to obtain a multi-objective optimization model; Based on the multi-objective optimization model, a weighted flow network is established, and the environment state vector sequence is mapped to corresponding path variables, shunt variables, and speed variables to obtain a graph model instance; Perform path set screening and delay weight calculation on the graph model instance to obtain a candidate path set; Jointly establish a linear programming solving problem for the candidate path set and the multi-objective optimization model to obtain a parameter optimization solution; Perform heuristic iterative fine-tuning on the parameter optimization solution, and perform local recalculation according to local congestion indications to obtain an iterative optimization solution; Map the iterative optimization solution to an optimized candidate scheduling sequence, and perform time sequence alignment with the environment state vector sequence to output initial scheduling data.

4. The method of claim 1, wherein, The initial scheduling data is input into a verification model for checking and rule-based correction to obtain plan execution data, including: Construct a verification input object using the initial scheduling data to obtain a verification input set; Input the verification input set into corresponding model parallel simulation running to generate a multi-model evaluation result set; Calculate the consistency measure using the multi-model evaluation result set to obtain a consistency evaluation report; Input the consistency evaluation report into a rule-based correction module for processing to obtain a correction rule set; Call a large language model based on the correction rule set to perform semantic analysis and policy text translation on unstructured requests, and merge with the correction rule set to obtain a correction constraint set; Constrain and synthesize the initial scheduling data using the correction constraint set to update and output plan execution data.

5. An automated logistics conveyor line dispatch optimization apparatus, characterized by, It includes: A data processing unit for acquiring multi-source running data and performing data preprocessing to obtain a standardized running state sequence; A parameter mapping unit for parameter mapping of the standardized running state sequence to construct state features to obtain an environment state vector sequence; A policy scheduling unit for calling a multi-agent collaborative decision engine to solve the policy agent of the environment state vector sequence to generate initial scheduling data; A data checking unit for inputting the initial scheduling data into a verification model for checking and rule-based correction to obtain plan execution data; A message generation unit for control instruction analysis, device mapping, and message encoding based on the plan execution data to generate a scheduling control message to be issued for a conveyor controller. The scheduling output unit is configured to perform transmission adaptation processing on the to-be-transmitted scheduling control message, and output a scheduling control result set. The parameter mapping unit is specifically configured to perform semantic labeling and classification based on the standardized operation state sequence to obtain a type labeling sequence; map each data item of the type labeling sequence to a node, an edge, and a station identifier based on a conveying line topology to obtain a topology alignment data sequence; perform multi-modal processing on the topology alignment data sequence to obtain a multi-modal state sub-vector set; align and splice the multi-modal state sub-vector set to obtain a time window statistical feature sequence; perform missing value interpolation on the time window statistical feature sequence, and then perform continuous feature splicing to obtain a fusion feature tensor; organize the fusion feature tensor into a step input in chronological order, and perform dimension rearrangement and mask construction to output an environment state vector sequence; The message generation unit is specifically configured to parse a control target based on the plan execution data, and perform instruction granularity segmentation to obtain an instruction unit sequence; match the instruction unit sequence with a device identifier and map the instruction unit sequence to an address label based on a PLC control channel to obtain a device mapping instruction table; perform theme mapping on the device mapping instruction table to obtain a protocol mapping table; construct a control payload based on the protocol mapping table, and organize the control payload into a load structure set according to an interface specification; The load structure set is serialized and encoded to obtain a control message frame set; the control message frame set is merged and grouped based on the PLC control channel to output a to-be-transmitted scheduling control message for a conveying line controller; The scheduling output unit is specifically configured to establish a transmission adaptation object to identify and configure an association key for message metadata in the to-be-transmitted scheduling control message, generate a delivery record for result binding; obtain actual operation data based on the delivery record, and aggregate to obtain an execution feedback original set; extract required interaction parameters in the execution feedback original set to obtain a feedback element set; write the feedback element set into a vector database and establish a similarity retrieval index to obtain a feedback vector index set; Based on the feedback vector index set, a cargo parameter index is calculated to obtain a structured result item; the structured result item is associated with the delivery record to obtain a scheduling control result set.

6. A computer device, comprising: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the automatic logistics conveying line scheduling optimization method according to any one of claims 1 to 4.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the automatic logistics conveying line scheduling optimization method according to any one of claims 1 to 4.

Images