Warehouse intelligent agent self-evolution scheduling method and device, equipment and medium

Abstract

The application provides a warehouse intelligent agent self-evolution scheduling method and device, equipment and medium. The method comprises the following steps: obtaining running data of a current warehouse scheduling system, and performing abnormality detection on the running data to obtain an evolution trigger event set; performing directional variation based on the evolution trigger event set to generate a variation strategy candidate set; performing simulation calculation on the variation strategy candidate set to obtain an adaptability sorting result; selecting the updated local strategy gene library according to the adaptability sorting result; publishing and recording the updated local strategy gene library to construct an evolution knowledge graph; and performing self-evolution scheduling based on a scheduling cycle strategy and the evolution knowledge graph. Through the implementation of the embodiment of the application, the system has self-evolution capability, continuously performs detection, generation, calculation, selection, construction and self-evolution scheduling in operation, realizes long-term and continuous improvement of scheduling efficiency, and realizes full-link traceability of strategy evolution through the evolution knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of warehouse scheduling technology, and in particular to a self-evolving scheduling method, apparatus, equipment and medium for warehouse intelligent agents. Background Technology

[0002] Intelligent warehouse scheduling systems face challenges from continuously changing operating environments during long-term operation, including order structure shifts (such as changes in SKU popularity and order size fluctuations), equipment status changes (such as performance degradation due to equipment aging and the introduction of new equipment), and adjustments to business rules (such as new temperature control requirements and changes in delivery time requirements). Traditional scheduling systems, once deployed, tend to have fixed scheduling strategies, requiring periodic adjustments through manual intervention. This not only results in delayed responses but also makes it difficult to identify and utilize potential optimization opportunities within operational data.

[0003] In addition, in recent years, research has emerged on automatic algorithm design that combines Large Language Models (LLMs) with evolutionary computation, leveraging the natural language understanding and code generation capabilities of LLMs to assist in the automatic evolution of heuristic algorithms. However, these works mainly focus on general combinatorial optimization problems (such as the packing problem, the traveling salesman problem, and the shop floor scheduling problem), and their evolutionary objects are algorithm code fragments or scheduling rule functions, rather than structured strategy units specific to warehousing scenarios. Therefore, they cannot be directly applied to practical warehousing scheduling environments with high requirements for security and traceability.

[0004] Some existing solutions employ online learning or adaptive adjustment mechanisms, but they suffer from the following shortcomings: **Strategy solidification and lag in response:** Traditional systems tend to solidify their strategies after deployment, relying on periodic manual tuning. This makes them unable to respond in real-time to environmental dynamics such as order shifts and equipment changes, and makes it difficult to uncover potential optimization opportunities within the data. **Lack of strategy-level interpretability and traceability:** Existing online learning mechanisms primarily adjust model parameters (such as neural network weights) rather than structured strategy units, resulting in an uninterpretable, unauditable, and unmanageable decision-making process. Furthermore, the lack of a genealogical record of strategy evolution makes it impossible to trace the optimization process. **Lack of security verification and protection mechanisms:** The absence of sandbox simulation verification, gray-scale deployment, and effective rollback mechanisms makes online updates of new strategies prone to introducing instability, failing to guarantee the security and reliability of the actual operating environment. **Lack of proactive environmental awareness:** The system cannot proactively perceive environmental shifts such as order structure or equipment performance, and cannot automatically detect strategy degradation and trigger re-optimization. **Limitations of evolutionary mechanisms:** Existing evolutionary algorithms are mostly used as offline optimization tools for one-time applications. The evolutionary object is the algorithm parameters rather than interpretable strategy units, making online dynamic self-evolution impossible. Cross-warehouse strategy migration is also only static inheritance, rather than dynamic evolution based on actual performance.

[0005] Therefore, existing warehouse scheduling methods suffer from rigid strategies and delayed responses, lack of strategy-level interpretability and traceability, lack of genealogical records of strategy evolution, lack of security verification and guarantee mechanisms, lack of proactive environmental perception capabilities, and limitations of evolutionary mechanisms, resulting in low scheduling efficiency. Summary of the Invention

[0006] This invention provides a self-evolving scheduling method, apparatus, device, and medium for warehouse intelligent agents, aiming to solve the problems of low scheduling efficiency caused by existing warehouse scheduling methods, such as rigid strategies and delayed responses, lack of strategy-level interpretability and traceability, lack of genealogical records of strategy evolution, lack of security verification and guarantee mechanisms, lack of proactive perception of the environment, and limitations of evolutionary mechanisms.

[0007] To address the aforementioned problems, in a first aspect, embodiments of the present invention provide a self-evolving scheduling method for warehouse intelligent agents, applied to a warehouse scheduling system, the method comprising:

[0008] Obtain the current operational data of the warehouse scheduling system, and perform anomaly detection on the operational data to obtain an evolution trigger event set;

[0009] Based on the set of evolutionary triggering events, targeted mutations are performed to generate a candidate set of mutation strategies;

[0010] The fitness ranking results are obtained by performing simulation calculations on the candidate set of mutation strategies.

[0011] The updated local strategy gene library is obtained by selecting from the fitness ranking results.

[0012] The updated local strategy gene library is published and recorded to construct an evolutionary knowledge graph;

[0013] Self-evolutionary scheduling is performed based on the scheduling loop strategy and the evolutionary knowledge graph.

[0014] Secondly, embodiments of this application provide a self-evolving scheduling device for warehouse intelligent agents, applied to a warehouse scheduling system, the device comprising:

[0015] An anomaly detection unit is used to acquire the current operational data of the warehouse scheduling system and perform anomaly detection on the operational data to obtain an evolution trigger event set;

[0016] The mutation generation unit is used to perform targeted mutations based on the set of evolution triggering events to generate a candidate set of mutation strategies.

[0017] The simulation calculation unit is used to perform simulation calculations on the mutation strategy candidate set to obtain fitness ranking results.

[0018] The selection unit is used to select from the fitness ranking results to obtain the updated local strategy gene library.

[0019] The publishing and recording unit is used to publish and record the updated local strategy gene library in order to construct an evolutionary knowledge graph;

[0020] The scheduling loop unit is used to perform self-evolutionary scheduling processing based on the scheduling loop strategy and the evolutionary knowledge graph.

[0021] Thirdly, embodiments of this application provide a computer device, the computer device including a memory and a processor connected to the memory; the memory is used to store a computer program, and the processor is used to run the computer program stored in the memory to perform the method described in the first aspect above.

[0022] Fourthly, embodiments of this application provide a storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, implement the method described in the first aspect above.

[0023] This invention provides a self-evolutionary scheduling method, apparatus, device, and medium for warehouse intelligent agents, applied to a warehouse scheduling system. The method includes: acquiring the current operating data of the warehouse scheduling system and performing anomaly detection on the operating data to obtain an evolutionary trigger event set; performing targeted mutation based on the evolutionary trigger event set to generate a mutation strategy candidate set; performing simulation calculation on the mutation strategy candidate set to obtain a fitness ranking result; selecting from the fitness ranking result to obtain an updated local strategy gene library; publishing and recording the updated local strategy gene library to construct an evolutionary knowledge graph; and performing self-evolutionary scheduling processing based on a scheduling loop strategy and the evolutionary knowledge graph. Therefore, this invention draws on the core mechanism of "mutation-selection-heredity" in biological evolution, enabling the system to possess self-optimization capabilities similar to biological evolution. During operation, it continuously performs detection, generation, calculation, selection, construction, and self-evolutionary scheduling processing, achieving long-term and continuous improvement in scheduling efficiency, and realizing full-link traceability of strategy evolution through the evolutionary knowledge graph. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1A flowchart illustrating the self-evolutionary scheduling method for warehouse intelligent agents provided in an embodiment of the present invention;

[0026] Figure 2 A schematic block diagram of a warehouse intelligent agent self-evolution scheduling device provided in an embodiment of the present invention;

[0027] Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0030] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0031] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0032] To facilitate understanding of this invention, the technical terms involved in this invention are explained as follows:

[0033] A strategy gene is the smallest independently executable strategy unit in a warehouse scheduling system. Its internal structure includes the strategy intent (the scheduling goal the strategy aims to achieve), preconditions (the environmental conditions required for the strategy to be triggered), execution actions (a series of scheduling operations to be performed after the strategy is triggered), and constraints (security and business restrictions that cannot be violated during strategy execution). Strategy genes are encoded, stored, and propagated as structured data objects, analogous to the functional units of biological genes.

[0034] Fitness: A quantitative indicator that measures the performance of a strategy in a specific operating environment. Its value is calculated by a fitness function, which typically considers multiple dimensions such as throughput, latency, congestion rate, and equipment utilization. The higher the fitness, the more effectively the strategy can cope with the current operating environment.

[0035] Evolutionary knowledge graph: A knowledge network using a graph database to record the version evolution relationships of strategy genes, the lineage of mutation operations, the trajectory of fitness changes, and environmental context relationships. This graph not only describes "which strategies are effective," but also comprehensively records the entire process of "how strategies evolve from nothing to something."

[0036] Sandbox simulation environment: refers to an isolated simulation instance built based on real-time data snapshots of the current warehouse operating environment, used to conduct security assessments of candidate strategies without affecting actual operation.

[0037] Gray-scale verification: refers to a gradual verification mechanism in the actual operating environment, in which a small amount of traffic is switched to run under the new strategy and its performance is monitored, in order to reduce the risk of the new strategy being fully launched.

[0038] Shannon Diversity Index: A diversity metric derived from information theory, used to assess the richness and evenness of strategy types in a strategy gene pool. A higher value indicates greater diversity of strategy types.

[0039] Vector Database: A database system specifically designed for storing and retrieving high-dimensional vector data. It supports approximate nearest neighbor retrieval based on metrics such as cosine similarity and Euclidean distance. In this invention, it is used to store feature vectors of abnormal patterns and feature codes of historical repair cases to achieve fast similar case retrieval.

[0040] Gene distribution protocol: refers to the standardized communication specifications for cross-node propagation of strategy genes as defined in this invention, including the serialization encoding format of gene content, transmission message structure, version conflict resolution rules, and the legality verification process of the receiver, to ensure the secure and complete propagation of genes between different repository nodes.

[0041] Please see Figure 1 , Figure 1 This is a flowchart illustrating the self-evolutionary scheduling method for warehouse intelligent agents provided in an embodiment of the present invention. Figure 1 As shown, this embodiment of the invention provides a self-evolutionary scheduling method for warehouse intelligent agents, which is applied to a warehouse scheduling system. Specifically, this invention can be used to carry out end-to-end strategy self-evolutionary scheduling optimization for complex operational topologies formed by AGVs, stacker cranes, conveyor lines, sorting equipment, etc. in large automated warehouses and intelligent distribution centers. The method includes the following steps S110-S160.

[0042] S110. Obtain the current operating data of the warehouse scheduling system, and perform anomaly detection on the operating data to obtain an evolution trigger event set.

[0043] In this embodiment, the operational data refers to the continuously collected multi-dimensional operational data of the current warehouse scheduling system, which may include equipment status data, cargo flow data, and scheduling execution feedback data. After acquiring the operational data, anomaly detection can be performed on the operational data to obtain an evolution trigger event set.

[0044] In one embodiment, the step of acquiring the current operational data of the warehouse scheduling system and performing anomaly detection on the operational data to obtain an evolution trigger event set includes:

[0045] The system collects real-time data on equipment status, cargo flow, and scheduling execution feedback from the current warehouse scheduling system to obtain the operational data.

[0046] Sliding window statistics and baseline deviation detection are performed on the operational data to identify abnormal fluctuation events of key performance indicators, thereby obtaining an abnormal signal set;

[0047] Spatiotemporal correlation analysis is performed on the abnormal signal set to identify the topological location, temporal pattern, and association with the equipment of the anomaly, thereby obtaining an abnormal pattern feature set;

[0048] The set of evolution trigger events is obtained by judging the abnormal pattern feature set based on the judgment rules.

[0049] In this embodiment, the equipment status data includes conveyor line speed, AGV (Automated Guided Vehicle) position and power, stacker crane load, etc.; the cargo flow data includes cargo transit time, throughput at each node, queue length, etc.; the scheduling execution feedback data includes strategy triggering frequency, execution success rate, actual effect and expected deviation, etc.

[0050] The process involves performing sliding window statistics and baseline deviation detection on the operational data to identify abnormal fluctuations in key performance indicators (KPIs) and obtain an abnormal signal set. Specifically, the KPIs include throughput, average on-time, congestion rate, and equipment utilization rate. These KPIs can be obtained from the operational data, and sliding window statistics and baseline deviation detection are then performed on them. The baseline deviation detection uses an exponentially weighted moving average (EWMA, a time series smoothing method that assigns higher weights to recent data), and its calculation formula is as follows: Among them, the initial value of the dynamic baseline. Take the arithmetic mean of the observations within the first complete statistical period (e.g., the first 1-hour window) after the system starts up, to ensure that the recursive process of baseline calculation has a reasonable starting state; The baseline estimate at time t. The current observation value is λ; λ is the smoothing coefficient, which usually ranges from 0.1 to 0.3. The larger the λ is, the more sensitive the response to the most recent data is. When the deviation between the actual value and the baseline estimate exceeds k times the standard deviation (k can be configured according to the actual situation, usually 2 to 3), an anomalous signal is generated, and several anomalous signals are used as the set of anomalous signals.

[0051] The process involves performing spatiotemporal correlation analysis on the abnormal signal set to identify the topological location, temporal pattern, and equipment association of the anomalies, thereby obtaining an abnormal pattern feature set. Specifically, spatiotemporal correlation analysis is performed on the abnormal signal set to identify the topological location, temporal pattern, and equipment association of the anomalies. The topological location includes specific merging points, a section of conveyor line, a storage tunnel, etc.; the temporal pattern includes fixed time periods, occurrence under specific order patterns, etc.; and the equipment association refers to a specific AGV or conveyor line segment. The spatiotemporal correlation analysis employs a co-occurrence statistical method within a sliding time window: the system iterates through the timestamps and spatial identifiers of the abnormal signals, counts the abnormal signal pairs that co-occur within a preset time window (e.g., 30 minutes) and spatial neighborhood (e.g., a topological distance of no more than 3 nodes), calculates their co-occurrence frequency, and constructs an abnormal correlation matrix based on the co-occurrence frequency. When the co-occurrence frequency of certain signal pairs in the correlation matrix is ​​significantly higher than the random baseline, abnormal pattern features are formed, and several abnormal pattern features are used as the abnormal pattern feature set.

[0052] The evolution trigger event set is obtained by judging the abnormal pattern feature set based on the judgment rules. Specifically, the judgment rules include multiple criteria, such as criterion 1: the frequency of a certain abnormal pattern in the most recent M time periods exceeds a preset threshold T_freq; criterion 2: the moving average of a certain performance indicator shows a downward trend for K consecutive periods and the cumulative decrease exceeds a preset percentage D_pct; and criterion 3: a completely new abnormal pattern that the system has never encountered before (i.e., no matching record in the historical abnormal pattern library) has appeared. When any of the above criteria are met, the system determines that it needs to trigger strategy evolution and generates evolution trigger events, with several evolution trigger events forming the evolution trigger event set. The content of the evolution trigger events may include the triggering reason, associated abnormal patterns, the affected topological region, and the identifier of the currently used strategy, etc.

[0053] S120. Based on the set of evolution triggering events, perform targeted mutation to generate a candidate set of mutation strategies.

[0054] In this embodiment, after obtaining the set of evolution triggering events, targeted mutations can be performed based on the set of evolution triggering events to generate a candidate set of mutation strategies.

[0055] In one embodiment, the step of performing targeted mutation based on the set of evolutionary triggering events to generate a candidate set of mutation strategies includes:

[0056] The local strategy gene library is located based on the anomaly type and associated topological location of each event in the evolution trigger event set to obtain the set of genes to be mutated;

[0057] The parameter perturbation mutation is performed on each gene in the gene set to be mutated to obtain the parameter mutated gene set.

[0058] Structural variations are performed on each gene in the set of genes to be mutated to obtain a set of structurally mutated genes.

[0059] The innovative mutant gene set is obtained by performing pre-processing on the set of genes to be mutated and the acquired historical repair cases using a pre-set model and verification rules.

[0060] The parameter variant gene set, the structural variant gene set, and the innovative variant gene set are merged to obtain a candidate set of variant strategies.

[0061] In this embodiment, the local strategy gene library is located based on the anomaly type and associated topological location of each event in the evolution trigger event set to obtain a set of genes to be mutated. Specifically, the anomaly type, topological region identifier, and time window of each event in the evolution trigger event set are obtained. In the local strategy gene library, the specific genes to be mutated are retrieved using the anomaly type, the topological region identifier, and the time window as search conditions, and then combined to form the set of genes to be mutated. The data structure of each strategy gene in the local strategy gene library includes at least the following fields: gene identifier (gene_id, a globally unique identifier, in the format "GN-{region}-{function}-{serial number}"), strategy intent (natural language text describing the scheduling goal to be achieved by the strategy), preconditions (an array of condition type-threshold key-value pairs), action sequence (an array of action objects arranged in execution order), constraints (including maximum speed limit, minimum spacing requirement, device load limit, etc.), version number (an integer incrementing), parent gene identifier (parent_gene_id, pointing to the previous generation gene that produced this gene), and fitness history (a fitness score array arranged in time). For example, if the evolution trigger event is "merging point A3 is continuously congested from 3 to 5 pm", the system will search the local strategy gene library for all strategy genes whose prerequisites include the topological region and time window "15:00-17:00" based on the topological region identifier "merging point A3" and the time window "15:00-17:00" in the event, and locate the diversion scheduling strategy gene that controls the merging point.

[0062] The parameter perturbation mutation of each gene in the gene set to be mutated is used to obtain a parameter-mutated gene set. Specifically, parameter perturbation mutation involves applying Gaussian noise perturbation to the execution parameters of the genes to generate fine-tuned versions of the parameters. For example, the shunt opening threshold is perturbed from the original value of 15 to multiple candidate values ​​between 12 and 18, and the velocity setting is perturbed from 0.8 m / s to candidate values ​​between 0.6 m / s and 1.0 m / s. The perturbation formula is as follows: Where θ is the original parameter value, and θ' is the mutated parameter value. This represents a Gaussian distributed random number with a mean of 0 and a standard deviation of σ; σ is the standard deviation of the disturbance, which can be adaptively adjusted according to the severity of the anomaly, and its calculation formula is as follows: ;in, The system uses the standard deviation of the basic perturbation (configurable, e.g., set to 5% of the parameter value), η as the sensitivity coefficient (configurable, e.g., set to 2.0), and severity as the normalized score of the anomaly severity (ranging from 0 to 1). For each gene to be mutated, the system generates P candidate values ​​(P is configurable, e.g., set to 5 to 10), thus forming the parameter-mutated gene set.

[0063] The structural mutation of each gene in the set of genes to be mutated is performed to obtain a structurally mutated gene set. Specifically, structural mutation involves modifying the execution sequence of the genes to be mutated, including changing the action order (e.g., changing from deceleration followed by diversion to diversion followed by deceleration), adding actions (e.g., adding AGV avoidance instructions, adding temporary buffer activation actions), deleting actions (e.g., removing unnecessary waiting steps, deleting redundant detection instructions), and adjusting precondition boundaries (e.g., adjusting the trigger time window from "15:00-17:00" to "14:30-17:30", adjusting the congestion threshold from "queue length > 10" to "queue length > 8"). Structural mutation uses a random selection of mutation operators: the system maintains a structural mutation operator library, including operators such as sequence exchange, action insertion, action deletion, condition boundary expansion, and condition boundary contraction. Each time mutation occurs, 1 to 3 operators are randomly selected and applied sequentially to the target gene to generate the structurally mutated gene set.

[0064] The system utilizes a preset model and validation rules to process the set of genes to be mutated and the acquired historical repair cases to obtain an innovative set of mutated genes. Specifically, the preset model is a Large Language Model (LLM), which provides structured prompts containing the following information: (a) the type, location, temporal pattern, and severity of the current anomaly pattern; (b) a topological description of the affected area, including the types of adjacent devices, connectivity relationships, and capacity parameters; and (c) the top-K most similar historical successful repair cases (K is typically set to 3 to 5) retrieved from the vector database based on the cosine similarity of the anomaly pattern feature vectors. Each case includes an anomaly description, the repair strategy adopted, and data on the improvement in effectiveness after repair. The LLM generates a new strategy scheme based on the above information and requires the LLM to output the results in a predefined JSON Schema format, which includes four required fields: intent, preconditions, action_sequence, and constraints. The system performs schema validation on the LLM output, including field integrity validation, data type validation, and constraint rationality validation (such as checking whether the speed parameter is within the rated range of the device and checking whether the referenced device identifier exists in the current warehouse device list). If the schema validation passes, it is mapped to a standard gene format data object as an innovative variant gene; if the validation fails, the system retryes the LLM up to R times (R is configurable, for example, set to 3 times). Each retry appends the error message from the previous validation to guide the LLM to correct its output. If the validation still fails, the innovative variant gene is discarded.

[0065] The three sets of variant genes—the parameter variant gene set, the structural variant gene set, and the innovative variant gene set—are combined into a candidate set for the variant strategy.

[0066] S130. The candidate set of mutation strategies is processed by simulation to obtain the fitness ranking result.

[0067] In this embodiment, after obtaining the mutation strategy candidate set, the mutation strategy candidate set can be processed by simulation calculation to obtain the fitness ranking result.

[0068] In one embodiment, the step of performing simulation calculations on the mutation strategy candidate set to obtain fitness ranking results includes:

[0069] The current warehouse scheduling system's environmental data is acquired, and a sandbox simulation environment instance is constructed based on the environmental data.

[0070] Each candidate gene in the mutation strategy candidate set is loaded into the sandbox environment instance in sequence for independent simulation to obtain the simulation execution trajectory of each candidate gene.

[0071] Multidimensional performance indicators are extracted from the simulation execution trajectory, and the first fitness score of each candidate gene is calculated by calculating the multidimensional performance indicators according to the preset fitness function.

[0072] The candidate genes are sorted in descending order according to their first fitness scores to obtain the fitness ranking results.

[0073] In this embodiment, the environmental data of the current warehouse scheduling system is acquired, and a sandbox environment instance is constructed based on the environmental data. Specifically, a global consistency snapshot mechanism is adopted for the acquisition of real-time data snapshots: at a certain moment, all data sources of the system are synchronously sampled to obtain the environmental data. The environmental data includes the current location and status of each device, the location and destination of all goods in transit, the queue length of each node, the current valid order pool, and the speed settings of conveyor lines and AGVs, etc., in real-time snapshot data. To ensure the time consistency of snapshots, the system pauses accepting new scheduling decisions at the snapshot acquisition time (the pause time is usually on the order of milliseconds), and resumes normal scheduling after the snapshot data is written. In the sandbox simulation engine, the environmental data is used to construct the sandbox environment instance using discrete event simulation (DES, a simulation method that infers system state changes by simulating the sequence of events). Each device, each conveyor line segment, each merging point and splitting point is modeled as an independent simulation entity, and the entities interact with each other through event messages. The behavioral model parameters of the simulated entity are extracted from the operating parameters of the actual equipment, including the equipment's acceleration curve, failure rate distribution, maintenance interval, etc.

[0074] The process involves sequentially loading each candidate gene from the mutation strategy candidate set into the sandbox environment instance for independent simulation to obtain the simulation execution trajectory of each candidate gene. Specifically, each candidate gene is sequentially loaded into the sandbox environment instance for independent simulation, and each candidate gene runs under the same initial conditions for a preset duration (e.g., simulating 1 hour of actual operation with a simulation time step of 100 milliseconds) to obtain its simulation execution trajectory. To ensure fairness in the evaluation, each candidate gene uses the same data snapshot as its initial state and the same random seed sequence to simulate random events (e.g., randomly arriving new orders, randomly occurring brief equipment failures, etc.).

[0075] The process involves extracting multidimensional performance indicators from the simulation execution trajectory and calculating the first fitness score for each candidate gene using a preset fitness function. Specifically, the process involves extracting multidimensional performance indicators from the simulation trajectory and calculating the first fitness score based on the fitness function. The fitness function is as follows: Wherein, ΔT(g) represents the throughput increment of candidate gene g relative to the currently used strategy (the difference in the amount of goods passing through per unit time), ΔL(g) represents the average latency increment of candidate gene g (positive values ​​indicate increased latency, negative values ​​indicate decreased latency), ΔC(g) represents the congestion increment of candidate gene g (measured by the difference in the proportion of congestion duration), and ΔU(g) represents the equipment utilization increment of candidate gene g (measured by the difference in the proportion of effective equipment operation time). α, β, γ, and δ are non-negative weighting coefficients. In the formula, a positive sign before ΔT(g) and ΔU(g) indicates that higher throughput and equipment utilization are better, while a negative sign before ΔL(g) and ΔC(g) indicates that lower latency and congestion are better. The values ​​of each weighting coefficient can be configured according to the enterprise's business priorities. For example, when the enterprise prioritizes throughput, α=0.4, β=0.3, γ=0.2, and δ=0.1 can be set. The multidimensional performance indicators include throughput increment, average latency increment, congestion increment, and equipment utilization increment.

[0076] In actual calculations, all performance indicators are normalized and mapped to the same dimension before being weighted and summed to ensure comparability of indicators across different dimensions. Specifically, the normalization of each incremental indicator uses the min-max method, with the normalization formula: ΔX_norm=(ΔX-ΔX_min) / (ΔX_max-ΔX_min); where ΔX_min and ΔX_max are the minimum and maximum values ​​of the increment of this indicator in the current batch of candidate genes, respectively; when ΔX_max equals ΔX_min, the normalization value is set to 0.

[0077] The candidate genes are sorted in descending order according to the first fitness score to obtain the fitness ranking result; specifically, the candidate genes are sorted in descending order according to the first fitness score to obtain the fitness ranking result.

[0078] S140. Based on the fitness ranking results, select to obtain the updated local strategy gene library.

[0079] In this embodiment, after determining the fitness ranking result, an updated local strategy gene library can be obtained by selecting from the fitness ranking result.

[0080] In one embodiment, the step of selecting the updated local strategy gene library based on the fitness ranking results includes:

[0081] A predetermined number of candidate genes are obtained from the fitness ranking results as the winning gene set;

[0082] Obtain the parent genes of each winning gene in the winning gene set, and compare the differences between each winning gene and its corresponding parent gene to obtain evolutionary event data;

[0083] Obtain the currently used strategies and their corresponding second fitness scores from the local strategy gene library;

[0084] If the first fitness score of each gene in the set of winning genes is greater than or equal to the second fitness score, grayscale verification is performed on the winning genes corresponding to the first fitness score to obtain the verification result.

[0085] If the verification result is successful, the corresponding winning gene replaces the corresponding old version gene in the local strategy gene library to obtain the updated local strategy gene library.

[0086] In this embodiment, obtaining a preset number of candidate genes from the fitness ranking results as the winning gene set specifically involves selecting the top N (N is configurable, for example, 3 to 5) candidate genes from the fitness ranking results as the winning gene set.

[0087] The process involves obtaining the parent genes of each winning gene in the winning gene set and comparing the differences between each winning gene and its corresponding parent gene to obtain evolutionary event data. Specifically, for each winning gene, the system records its mutation operation type and mutation parameters (and fitness increase magnitude) relative to its parent gene, generating structured evolutionary event data. The mutation operation type includes parameter perturbation, structural mutation, or LLM innovation, and the mutation parameters include perturbation magnitude and newly added action descriptions. The data structure of the evolutionary event data includes at least: event identifier (event_id), timestamp, parent gene identifier (parent_gene_id), child gene identifier (child_gene_id), mutation operator type (mutation_type), mutation operator parameters (mutation_params), parent fitness (parent_fitness), child fitness (child_fitness), fitness increment (fitness_delta), and trigger pattern identifier (trigger_pattern_id).

[0088] The step of obtaining the currently used strategies and their corresponding second fitness scores in the local strategy gene library can specifically be achieved by calculating the second fitness score of the currently used strategies using the fitness function.

[0089] If the first fitness score of each winning gene in the set of winning genes is greater than or equal to the second fitness score, the winning gene corresponding to the first fitness score is subjected to gray-scale verification to obtain the verification result; specifically, for winning genes whose fitness score exceeds the current strategy, the system sets them to the ready state and enters the gray-scale verification process. Gray-scale verification adopts a progressive flow switching mechanism, which is divided into multiple incremental stages. Gray-scale verification is divided into the following stages: (1) Small flow verification stage - a small number of scheduling tasks (e.g., 10% of the flow) are switched to the new strategy and the actual performance indicators are continuously monitored for at least 1 hour; (2) Medium flow verification stage - if the actual performance of the new strategy during the small flow verification period is not lower than the preset ratio expected by the simulation (e.g., 90%) and no safety alarm is triggered, the flow ratio is increased (e.g., to 30%) and the monitoring is continuously carried out for at least 2 hours; (3) Large flow verification stage - the flow ratio is further increased (e.g., to 50%) and the monitoring is continuously carried out for at least 4 hours; (4) Full switch - after confirming that each stage has passed, 100% flow switch is performed. The specific traffic ratio, duration, and performance threshold for each of the above stages can be configured according to the actual business characteristics and risk appetite of the warehouse. In any stage, if the actual performance of the winning gene is significantly lower than or equal to the security alarm data, the verification result is a failure. The system immediately performs a rollback operation, switching all traffic back to the original strategy, and marking the reason and stage information for the gray-scale verification failure in the evolution event data. The security alarm data includes collision risks, equipment overload, etc.

[0090] If the verification result is successful, the corresponding winning gene replaces the corresponding old version gene in the local strategy gene library to obtain the updated local strategy gene library. Specifically, if the actual performance of the winning gene is significantly higher than the security alarm data, the verification result is successful. After the gray-scale verification is successful, the new gene is officially fixed to the local strategy gene library, that is, the corresponding winning gene replaces the corresponding old version gene in the local strategy gene library to obtain the updated local strategy gene library. At the same time, the old version gene is retained as a rollback backup, and the retention time is not less than the preset gene archiving period.

[0091] S150. Publish and record the updated local strategy gene library to construct an evolutionary knowledge graph.

[0092] In this embodiment, after determining the updated local strategy gene library, the updated local strategy gene library can be published and recorded to construct an evolutionary knowledge graph.

[0093] In one embodiment, the step of publishing and recording the updated local strategy gene library to construct an evolutionary knowledge graph includes:

[0094] The updated local strategy gene library is used to obtain the published data, and the published data is published to the gene sharing network;

[0095] Obtain the version evolution relationship of the updated local strategy gene library, and use the version evolution relationship to construct a gene phylogenetic map;

[0096] Evolutionary statistical indicators were calculated based on the aforementioned gene pedigree chart;

[0097] The evolutionary knowledge graph is constructed by associating the evolutionary statistical indicators with warehouse environment fingerprints and business scenario tags.

[0098] In this embodiment, the updated local strategy gene library is used to obtain the release data, and the release data is then published to the gene sharing network. Specifically, the gene release protocol is used to publish the release data through the entire network gene sharing network. The entire network gene sharing network adopts a distributed publish-subscribe architecture, where each repository node is both a publisher and a subscriber of genes, and asynchronous propagation of genes is achieved through message middleware. First, the winning genes in the updated local strategy gene library are serialized into gene messages in standard JSON format as the release data. The release data includes gene content fields (including a complete structured description of strategy intent, preconditions, execution actions, constraints, etc.), source environment fingerprint fields (such as the warehouse environment feature vector that generated the gene, including numerical encoding of dimensions such as warehouse area, equipment type and quantity, daily order volume range, main SKU category, and conveyor topology), fitness score fields, complete evolutionary event record fields (including parent gene ID, mutation operation description, fitness change trajectory), and version lineage information fields. Simultaneously, the published data is digitally signed to ensure its integrity and source credibility. Finally, the signed published data is published to the gene topic channel of the message middleware to complete the publication. After receiving the published data, the subscriber first performs signature verification and format verification. If the verification passes, the new gene is written into the local global gene cache library as a candidate heterologous gene pool in the subsequent evolutionary cycle. It should be noted that the gene publication in this step focuses on broadcasting and sharing the winning genes generated by the self-evolution of this repository as evolutionary experience. Its purpose is to enrich the diversity of the global gene library and provide greater variation space and reference materials for the subsequent self-evolution process of each repository, rather than directly performing cross-repository strategy deployment migration.

[0099] The system obtains the version evolution relationship of the updated local strategy gene library and constructs a gene phylogenetic graph using this relationship. Specifically, the system records the version evolution relationship of genes in the graph database (a database system that stores data in a node-edge-attribute structure, such as Neo4j) to construct the gene phylogenetic graph. Each node in the phylogenetic graph represents a gene version, and node attributes include gene identifier, version number, fitness score, creation timestamp, and source repository identifier. Edges represent mutation operations, and edge attributes include mutation type (parameter perturbation, structural mutation, or LLM innovation), mutation parameter summary, and fitness change.

[0100] The evolutionary statistical indicators are calculated based on the gene pedigree. Specifically, the evolutionary statistical indicators are calculated based on the pedigree. The statistical indicators include evolutionary depth (the number of variations from the initial version to the current version), branch width (the number of mutated branches generated by the same parent generation), fitness growth curve (the trend of fitness as the number of generations changes), and evolutionary success rate (the proportion of positive fitness improvement in each generation of mutations).

[0101] The evolutionary knowledge graph is constructed by associating the evolutionary statistical indicators with the warehouse environment fingerprint and business scenario tags. Specifically, the evolutionary knowledge graph with a multi-dimensional index is constructed by associating these indicators with the warehouse environment fingerprint and the business scenario tags. This knowledge graph not only records "which strategies are effective," but also comprehensively records the entire process of "how strategies evolve from low fitness to high fitness through self-evolution," providing historical decision-making basis for subsequent strategy degradation detection and re-evolution in this warehouse.

[0102] S160. Perform self-evolutionary scheduling processing based on the scheduling loop strategy and the evolutionary knowledge graph.

[0103] In this embodiment, after determining the evolutionary knowledge graph, self-evolutionary scheduling processing can be performed based on the scheduling loop strategy and the evolutionary knowledge graph.

[0104] In one embodiment, self-evolutionary scheduling processing based on a scheduling loop strategy and the evolutionary knowledge graph includes:

[0105] The fitness of all used genes in the updated local strategy gene library is re-evaluated according to a preset period to obtain a third fitness score.

[0106] Obtain historical fitness records, compare the trend of the third fitness score with the historical fitness records, and identify the genes whose fitness is continuously declining to obtain a set of degenerate genes;

[0107] Based on the evolutionary knowledge graph and the set of degenerate genes, a set of candidate replacement genes is obtained.

[0108] When the set of degenerate genes is a non-empty set, the set of candidate replacement genes is used as the set of evolutionary triggering events;

[0109] To continue performing targeted mutations based on the set of evolutionary triggering events to generate a candidate set of mutation strategies, the candidate set of mutation strategies is processed by simulation calculation to obtain fitness ranking results, and an updated local strategy gene library is obtained by selecting from the fitness ranking results.

[0110] The updated local strategy gene library is evaluated based on the diversity index assessment strategy to obtain the diversity index.

[0111] When the diversity index is lower than the preset index threshold, the updated local strategy gene library is published and recorded to complete the self-evolutionary scheduling process.

[0112] In this embodiment, the fitness re-evaluation of all used genes in the updated local strategy gene library is performed according to a preset period to obtain a third fitness score; specifically, the fitness re-evaluation of all used genes in the local strategy gene library is performed according to a preset period (such as daily or weekly) to calculate the third fitness score for all used genes using the fitness function.

[0113] The process involves acquiring historical fitness records, comparing the third fitness score with these historical fitness records to identify trends, and identifying genes whose fitness is continuously declining, thus obtaining a set of degenerate genes. Specifically, due to shifts in the operating environment (such as seasonal changes in order structure or equipment aging), strategies that were originally highly fit may experience performance decline, i.e., strategy degradation. The system uses a degradation detection formula to compare the current fitness (i.e., the third fitness score) with historical fitness records to identify trends, and uses linear regression to fit the fitness sequence for the most recent W evaluation periods. The degradation detection formula is as follows: ;in, Let W be the fitness score for the i-th evaluation period, and W be the backtracking window length. When the slope value is less than negative ε (ε is the degradation threshold, which is configurable, for example, set to 0.01), the gene is determined to be a degenerate gene, thus obtaining the set of degenerate genes.

[0114] The system retrieves a candidate replacement gene set based on the evolutionary knowledge graph and the degenerate gene set. Specifically, for each degenerate gene in the degenerate gene set, the system first performs an environmental similarity search on the evolutionary knowledge graph using a search formula to determine if there are any replacement genes with similar environments and higher fitness. The environmental similarity search uses the cosine similarity of the source environment fingerprint as a metric, and the search formula is as follows: ;in and These are the environmental fingerprint vectors of two repositories. When the cosine similarity exceeds a set threshold (e.g., 0.85), the two environments are considered sufficiently similar, and policy transfer can proceed. When the degenerate gene set is non-empty, the candidate replacement gene set is used as the evolutionary trigger event set, so that the degenerate genes and candidate replacement genes jointly trigger a new round of evolutionary cycles.

[0115] To continue performing targeted mutations based on the evolution trigger event set to generate a mutation strategy candidate set, the mutation strategy candidate set is processed by simulation calculation to obtain fitness ranking results, and selection is made according to the fitness ranking results to obtain an updated local strategy gene library; specifically, the above steps S120-S140 are repeated to obtain the updated local strategy gene library.

[0116] The updated local strategy gene pool is evaluated using a diversity index-based assessment strategy to obtain a diversity index. When the diversity index is lower than a preset threshold, the updated local strategy gene pool is published and recorded to complete the self-evolutionary scheduling process. Specifically, the strategy diversity of the updated local strategy gene pool is evaluated based on the Shannon diversity index, and the calculation formula is as follows: Where S represents the number of types of strategy genes (classified by strategy intent). This represents the percentage of usage frequency of strategy type i among the currently active strategies. It should be noted that when the percentage of usage frequency of a certain strategy... When = 0, it is agreed that =0, which is consistent with the function in The extreme values ​​approaching 0 are consistent. When the diversity index H is lower than the preset threshold H_min (H_min can be dynamically set according to the number of strategy types, for example, when the number of strategy types S=5, H_min can be set to 1.2, and when S=10, H_min can be set to 1.8), it indicates that the gene pool strategies tend to be homogeneous. The system triggers exploratory evolution, actively searching the entire network gene pool for heterogeneous genes (i.e., strategies from different warehouses or different business scenarios) whose source environment fingerprints are more similar to the local library than the preset threshold (e.g., 0.7) and whose fitness ranks among the top. These heterogeneous genes are introduced into the local strategy gene pool to participate in subsequent evolutionary cycles to increase strategy diversity and avoid local optimum traps.

[0117] The following is a specific numerical example to illustrate the implementation process of the present invention. Assume an e-commerce automated warehouse contains 120 AGVs, eight main conveyor lines, 16 merging points, and 12 branching points, processing an average of 50,000 orders per day. The warehouse scheduling system, during continuous monitoring, detects abnormal congestion at merging point A3 between 15:00 and 17:00 daily.

[0118] Specifically, in the anomaly detection operation of step S110, the system uses the EWMA method with a smoothing coefficient λ=0.2 to calculate the throughput baseline of the merging port A3. Operational data from the past week shows that the actual throughput of this merging port between 15:00 and 17:00 is 2800 pieces per hour, while the EWMA baseline is 3500 pieces per hour, a deviation of -20%, exceeding the alarm threshold of twice the standard deviation (standard deviation is 280 pieces / hour, twice the standard deviation corresponds to a deviation tolerance of 560 pieces / hour), thus generating an anomaly signal. Spatiotemporal correlation analysis reveals that this anomaly is associated with the high-load operation of the three branch conveyor lines during the same period (co-occurrence frequency reaches 85%), forming an anomaly pattern. Since this anomaly pattern has occurred 80% of the last 10 evaluation periods, exceeding the preset threshold T_freq=50%, the system determines that evolution needs to be triggered, generating an evolution trigger event.

[0119] In the mutation generation operation of step S120, the system locates the diversion scheduling strategy gene (gene_id="GN-A3-MERGE-001", version=3) of the merging point A3. Three types of mutation operations are performed on this gene: parameter perturbation mutation generates 8 candidate genes (the diversion point opening threshold is perturbed from 15 to 10, 11, 12, 13, 14, 16, 17, and 18 respectively); structural mutation generates 3 candidate genes (adding AGV avoidance instructions, swapping the order of deceleration and diversion actions, and adding temporary buffer activation actions); LLM innovation mutation, by calling a large language model and providing anomaly descriptions and historical repair cases, generates 2 new strategy candidate genes (an active diversion strategy based on time-period prediction and a dynamic rate-limiting strategy based on real-time queue length). After merging, 13 candidate genes are obtained.

[0120] In the simulation calculation operation of step S130, fitness was calculated using weighting coefficients α=0.4, β=0.3, γ=0.2, and δ=0.1. The 13 candidate genes were each simulated for 1 hour under the same initial conditions. The fitness scores of the top 3 candidate genes were: the active diversion strategy based on time-period prediction (LLM innovative variant, F=0.87), the parameter variant version with a diversion port opening threshold of 12 (F=0.82), and the structural variant version with added AGV avoidance instructions (F=0.78). The fitness score of the currently used strategy was F=0.65.

[0121] In step S140, during the selection and solidification process, all three winning genes entered the gray-scale verification process. Ultimately, the top-ranked proactive diversion strategy based on time-period prediction, after verification under low-flow (10% flow, 1 hour), medium-flow (30% flow, 2 hours), and high-flow (50% flow, 4 hours) conditions, achieved over 92% of the simulation's expected performance in all cases, successfully completing the full switchover and solidifying it into the local strategy gene library. This strategy increased the throughput of merging point A3 from 15:00 to 17:00 from 2800 pieces per hour to 3650 pieces per hour, and reduced the congestion duration from 35% to 8%.

[0122] In the release and recording operation of step S150, the winning gene is released to the whole network gene sharing network and its complete lineage is recorded in the graph database: from the initial version GN-A3-MERGE-001 (v1, fitness 0.55) through two parameter mutations to v3 (fitness 0.65), and then through LLM innovation mutation to v4 (fitness 0.87), with an evolutionary depth of 3 and a cumulative fitness increase of 58%.

[0123] In the scheduling cycle operation of step S106, the system confirmed that the fitness of the policy remained stable (slope=0.002, no degradation threshold ε=0.01) during continuous monitoring over the next four weeks, verifying the long-term effectiveness of the evolution results and completing the self-evolutionary scheduling process.

[0124] In summary, this invention discloses a process for acquiring current warehouse scheduling system operation data, performing anomaly detection on the operation data to obtain an evolutionary trigger event set, performing targeted mutation based on the evolutionary trigger event set to generate a mutation strategy candidate set, performing simulation calculations on the mutation strategy candidate set to obtain fitness ranking results, selecting from the fitness ranking results to obtain an updated local strategy gene library, publishing and recording the updated local strategy gene library to construct an evolutionary knowledge graph, and performing self-evolutionary scheduling processing based on a scheduling loop strategy and the evolutionary knowledge graph. Therefore, this invention draws on the core mechanism of "mutation-selection-heredity" in biological evolution, enabling the warehouse scheduling system to possess self-optimization capabilities similar to biological evolution. During operation, it continuously performs detection, generation, calculation, selection, construction, and self-evolutionary scheduling processing, achieving long-term and continuous improvement in scheduling efficiency, and realizing full-link traceability of the strategy evolution process through the evolutionary knowledge graph. Compared with existing offline evolutionary algorithm optimization schemes, this invention achieves online self-evolution, and the evolution object is a structured strategy gene unit rather than algorithm parameters, which has interpretability and traceability. Compared with existing cross-warehouse strategy migration schemes, this invention focuses on the dynamic self-evolution of the strategy based on actual performance during operation, rather than the static copying and adaptation of existing strategies. The two are complementary.

[0125] Figure 2 This is a schematic block diagram of a warehouse intelligent agent self-evolution scheduling device provided in an embodiment of the present invention. Figure 2 As shown, this embodiment of the invention provides a warehouse intelligent agent self-evolution scheduling device 700 that implements the method described above, applied to a warehouse scheduling system. Specifically, this invention can perform end-to-end strategy self-evolution scheduling optimization for complex operational topologies formed by AGVs, stacker cranes, conveyor lines, sorting equipment, etc., in large automated warehouses and intelligent distribution centers. For details, please refer to... Figure 2 The warehouse intelligent agent self-evolution scheduling device 700 includes:

[0126] Anomaly detection unit 701 is used to acquire the current operating data of the warehouse scheduling system and perform anomaly detection on the operating data to obtain an evolution trigger event set;

[0127] The mutation generation unit 702 is used to perform targeted mutations based on the set of evolution triggering events to generate a candidate set of mutation strategies.

[0128] The simulation calculation unit 703 is used to perform simulation calculations on the mutation strategy candidate set to obtain fitness ranking results.

[0129] Selection unit 704 is used to select the updated local strategy gene library based on the fitness ranking results.

[0130] The publishing and recording unit 705 is used to publish and record the updated local strategy gene library in order to construct an evolutionary knowledge graph;

[0131] The scheduling loop unit 706 is used to perform self-evolutionary scheduling processing based on the scheduling loop strategy and the evolutionary knowledge graph.

[0132] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned device can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0133] The above-described device can be implemented as a computer program, and the computer program can be implemented in, for example... Figure 3 It runs on the computer device shown.

[0134] Please see Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of the present invention. The electronic device 800 can be a terminal or a server. The terminal can be an electronic device with communication functions. The server can be a standalone server or a server cluster composed of multiple servers.

[0135] See Figure 3The electronic device 800 includes a processor 802, a memory, and a network interface 805 connected via a system bus 801. The memory may include a non-volatile storage medium 803 and internal memory 804.

[0136] The non-volatile storage medium 803 may store an operating system 8031 ​​and a computer program 8032. The computer program 8032 includes program instructions that, when executed, cause the processor 802 to perform a self-evolving scheduling method for warehouse agents.

[0137] The processor 802 provides computing and control capabilities to support the operation of the entire electronic device 800.

[0138] The internal memory 804 provides an environment for the operation of the computer program 8032 in the non-volatile storage medium 803. When the computer program 8032 is executed by the processor 802, the processor 802 can execute a self-evolution scheduling method for warehouse intelligence agents.

[0139] This network interface 805 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device 800 to which the present invention is applied. The specific electronic device 800 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0140] The processor 802 is used to run a computer program 8032 stored in the memory to perform the following steps:

[0141] Obtain the current operational data of the warehouse scheduling system, and perform anomaly detection on the operational data to obtain an evolution trigger event set;

[0142] Based on the set of evolutionary triggering events, targeted mutations are performed to generate a candidate set of mutation strategies;

[0143] The fitness ranking results are obtained by performing simulation calculations on the candidate set of mutation strategies.

[0144] The updated local strategy gene library is obtained by selecting from the fitness ranking results.

[0145] The updated local strategy gene library is published and recorded to construct an evolutionary knowledge graph;

[0146] Self-evolutionary scheduling is performed based on the scheduling loop strategy and the evolutionary knowledge graph.

[0147] It should be understood that, in this embodiment of the invention, the processor 802 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0148] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0149] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the following steps:

[0150] Obtain the current operational data of the warehouse scheduling system, and perform anomaly detection on the operational data to obtain an evolution trigger event set;

[0151] Based on the set of evolutionary triggering events, targeted mutations are performed to generate a candidate set of mutation strategies;

[0152] The fitness ranking results are obtained by performing simulation calculations on the candidate set of mutation strategies.

[0153] The updated local strategy gene library is obtained by selecting from the fitness ranking results.

[0154] The updated local strategy gene library is published and recorded to construct an evolutionary knowledge graph;

[0155] Self-evolutionary scheduling is performed based on the scheduling loop strategy and the evolutionary knowledge graph.

[0156] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0157] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0158] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0159] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0160] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0161] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The software tools, models, or components appearing in the embodiments of the present invention are merely illustrative examples and do not represent actual use.

Claims

1. A self-evolutionary scheduling method for warehouse intelligent agents, characterized in that, The method, applied to a warehouse scheduling system, includes: Obtain the current operational data of the warehouse scheduling system, and perform anomaly detection on the operational data to obtain an evolution trigger event set; Based on the set of evolutionary triggering events, targeted mutations are performed to generate a candidate set of mutation strategies; The fitness ranking results are obtained by performing simulation calculations on the candidate set of mutation strategies. The updated local strategy gene library is obtained by selecting from the fitness ranking results. The updated local strategy gene library is published and recorded to construct an evolutionary knowledge graph; Self-evolutionary scheduling processing is performed based on the scheduling loop strategy and the evolutionary knowledge graph. The process of acquiring the current operational data of the warehouse scheduling system and performing anomaly detection on the operational data to obtain an evolution trigger event set includes: The system collects real-time equipment status data, cargo flow data, and scheduling execution feedback data from the current warehouse scheduling system to obtain operational data. It then performs sliding window statistics and baseline deviation detection on the operational data to identify abnormal fluctuations in key performance indicators, resulting in an abnormal signal set. Finally, it performs spatiotemporal correlation analysis on the abnormal signal set to identify the topological location, temporal patterns, and equipment associations of the anomalies, obtaining an abnormal pattern feature set. Based on judgment rules, it determines the abnormal pattern feature set to obtain an evolution trigger event set. The simulation calculation of the mutation strategy candidate set to obtain the fitness ranking result includes: The system acquires environmental data from the current warehouse scheduling system and constructs a sandbox simulation environment instance based on the acquired data. It then loads each candidate gene from the mutation strategy candidate set into the sandbox environment instance for independent simulation, obtaining the simulation execution trajectory of each candidate gene. Multidimensional performance indicators are extracted from the simulation execution trajectories, and the first fitness score of each candidate gene is calculated using a preset fitness function. Finally, the candidate genes are sorted in descending order according to their first fitness scores to obtain the fitness ranking result. The self-evolutionary scheduling process based on the scheduling loop strategy and the evolutionary knowledge graph includes: According to a preset cycle, the fitness of all used genes in the updated local strategy gene library is re-evaluated to obtain a third fitness score; historical fitness records are obtained, and the third fitness score is compared with the historical fitness records to identify used genes with continuously declining fitness, thus obtaining a degenerate gene set; a search is performed based on the evolutionary knowledge graph and the degenerate gene set to obtain a candidate replacement gene set; when the degenerate gene set is not empty, the candidate replacement gene set is used as the evolution trigger event set; targeted mutation is then performed based on the evolution trigger event set to generate a mutation strategy candidate set, and the mutation strategy candidate set is processed by simulation to obtain a fitness ranking result; an updated local strategy gene library is obtained by selecting from the fitness ranking results; the updated local strategy gene library is evaluated based on a diversity index evaluation strategy to obtain a diversity index; when the diversity index is lower than a preset index threshold, the updated local strategy gene library is published and recorded to complete the self-evolutionary scheduling process.

2. The method according to claim 1, characterized in that, The step of performing targeted mutation based on the set of evolutionary triggering events to generate a candidate set of mutation strategies includes: The local strategy gene library is located based on the anomaly type and associated topological location of each event in the evolution trigger event set to obtain the set of genes to be mutated; The parameter perturbation mutation is performed on each gene in the gene set to be mutated to obtain the parameter mutated gene set. Structural variations are performed on each gene in the set of genes to be mutated to obtain a set of structurally mutated genes. The innovative mutant gene set is obtained by performing pre-processing on the set of genes to be mutated and the acquired historical repair cases using a pre-set model and verification rules. The parameter variant gene set, the structural variant gene set, and the innovative variant gene set are merged to obtain a candidate set of variant strategies.

3. The method according to claim 1, characterized in that, The step of selecting the updated local strategy gene library based on the fitness ranking results includes: A predetermined number of candidate genes are obtained from the fitness ranking results as the winning gene set; Obtain the parent genes of each winning gene in the winning gene set, and compare the differences between each winning gene and its corresponding parent gene to obtain evolutionary event data; Obtain the currently used strategies and their corresponding second fitness scores from the local strategy gene library; If the first fitness score of each gene in the set of winning genes is greater than or equal to the second fitness score, grayscale verification is performed on the winning genes corresponding to the first fitness score to obtain the verification result. If the verification result is successful, the corresponding winning gene replaces the corresponding old version gene in the local strategy gene library to obtain the updated local strategy gene library.

4. The method according to claim 1, characterized in that, The process of publishing and recording the updated local strategy gene library to construct an evolutionary knowledge graph includes: The updated local strategy gene library is used to obtain the published data, and the published data is published to the gene sharing network; Obtain the version evolution relationship of the updated local strategy gene library, and use the version evolution relationship to construct a gene phylogenetic map; Evolutionary statistical indicators were calculated based on the aforementioned gene pedigree chart; The evolutionary knowledge graph is constructed by associating the evolutionary statistical indicators with warehouse environment fingerprints and business scenario tags.

5. A self-evolving scheduling device for warehouse intelligent agents, characterized in that, The device is applied to a warehouse scheduling system; the device also includes: An anomaly detection unit is used to acquire the current operational data of the warehouse scheduling system and perform anomaly detection on the operational data to obtain an evolution trigger event set; The mutation generation unit is used to perform targeted mutations based on the set of evolution triggering events to generate a candidate set of mutation strategies. The simulation calculation unit is used to perform simulation calculations on the mutation strategy candidate set to obtain fitness ranking results. The selection unit is used to select from the fitness ranking results to obtain the updated local strategy gene library. The publishing and recording unit is used to publish and record the updated local strategy gene library in order to construct an evolutionary knowledge graph; The scheduling loop unit is used to perform self-evolutionary scheduling processing based on the scheduling loop strategy and the evolutionary knowledge graph. The process of acquiring the current operational data of the warehouse scheduling system and performing anomaly detection on the operational data to obtain an evolution trigger event set includes: The system collects real-time equipment status data, cargo flow data, and scheduling execution feedback data from the current warehouse scheduling system to obtain operational data. It then performs sliding window statistics and baseline deviation detection on the operational data to identify abnormal fluctuations in key performance indicators, resulting in an abnormal signal set. Finally, it performs spatiotemporal correlation analysis on the abnormal signal set to identify the topological location, temporal patterns, and equipment associations of the anomalies, obtaining an abnormal pattern feature set. Based on judgment rules, it determines the abnormal pattern feature set to obtain an evolution trigger event set. The simulation calculation of the mutation strategy candidate set to obtain the fitness ranking result includes: The system acquires environmental data from the current warehouse scheduling system and constructs a sandbox simulation environment instance based on the acquired data. It then loads each candidate gene from the mutation strategy candidate set into the sandbox environment instance for independent simulation, obtaining the simulation execution trajectory of each candidate gene. Multidimensional performance indicators are extracted from the simulation execution trajectories, and the first fitness score of each candidate gene is calculated using a preset fitness function. Finally, the candidate genes are sorted in descending order according to their first fitness scores to obtain the fitness ranking result. The self-evolutionary scheduling process based on the scheduling loop strategy and the evolutionary knowledge graph includes: According to a preset cycle, the fitness of all used genes in the updated local strategy gene library is re-evaluated to obtain a third fitness score; historical fitness records are obtained, and the third fitness score is compared with the historical fitness records to identify used genes with continuously declining fitness, thus obtaining a degenerate gene set; a search is performed based on the evolutionary knowledge graph and the degenerate gene set to obtain a candidate replacement gene set; when the degenerate gene set is not empty, the candidate replacement gene set is used as the evolution trigger event set; targeted mutation is then performed based on the evolution trigger event set to generate a mutation strategy candidate set, and the mutation strategy candidate set is processed by simulation to obtain a fitness ranking result; an updated local strategy gene library is obtained by selecting from the fitness ranking results; the updated local strategy gene library is evaluated based on a diversity index evaluation strategy to obtain a diversity index; when the diversity index is lower than a preset index threshold, the updated local strategy gene library is published and recorded to complete the self-evolutionary scheduling process.

6. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-4.

7. A storage medium, characterized in that, The storage medium stores a computer program, which includes program instructions that, when executed by a processor, can implement the method as described in any one of claims 1-4.

Images