A Quaternary Knowledge Graph-Driven Library Approach and System for Home Appliance Warehouse Scenarios

Abstract

This invention discloses a warehouse management method and system based on a four-element knowledge graph driven by a home appliance warehouse scenario. It constructs a four-element knowledge graph with goods, storage locations, AGVs, and equipment as core entities; builds a small-scale vertical model for warehouse management, integrating storage location allocation algorithms, AGV path optimization algorithms, and anomaly closed-loop handling algorithms; storage location allocation employs a fusion method of graph embedding and neural factor machine; AGV path planning uses a hybrid algorithm combining global exploration with bidirectional A* local optimization; and anomaly handling employs a two-stage prediction process using isolated forest and rule engine, along with a full-link closed-loop process. Finally, an execution application layer is constructed through an intelligent inbound planning module, a dynamic warehouse management execution module, and a real-time anomaly handling module, deployed on a cloud-edge-device collaborative architecture to achieve intelligent perception, accurate decision-making, and automatic execution of warehouse management. This invention solves the problems of data silos, coarse decision-making, and delayed response in traditional warehousing.

Description

Technical Field

[0001] This invention belongs to the field of intelligent logistics and warehousing technology. Specifically, it relates to a database management method and system based on quaternary knowledge graphs in the context of home appliance warehouses. Background Technology

[0002] In the field of intelligent logistics and warehousing, home appliances, due to their large size, heavy weight, and the sensitivity of some high-end products to environmental factors (temperature and humidity), require extremely high levels of precision in warehouse storage and scheduling. Currently, the industry generally uses Warehouse Management Systems (WMS) for inventory management and is gradually introducing automated equipment such as AGVs for material handling. However, in the warehouse management process (i.e., closed-loop operations such as inbound storage, dynamic transfer, and anomaly handling), the following core challenges remain: matching goods with storage locations relies on manual experience or simple rules, resulting in low storage space utilization and a high risk of damage; AGV scheduling schemes are static and rigid, leading to path conflicts and high empty-run rates; and responses to emergencies such as abnormal temperature and humidity and equipment failures are delayed, lacking a closed-loop handling mechanism. Summary of the Invention

[0003] This invention addresses the technical problems of low warehouse location adaptability, inefficient AGV scheduling, and delayed anomaly response and handling in existing warehouse management systems. It proposes a warehouse management method and system based on a quaternary knowledge graph in the context of home appliance warehouses. This method achieves dynamic and optimal warehouse location allocation and AGV scheduling, intelligent warehouse management decisions tailored to industry characteristics, and full-link automated and closed-loop management of anomalies from discovery to resolution.

[0004] The present invention is implemented using the following technical solutions:

[0005] A quaternion knowledge graph-driven knowledge base approach is proposed for the home appliance warehouse scenario, including:

[0006] A four-element knowledge graph is constructed with goods, storage locations, AGVs, and equipment as core entities;

[0007] Construct and train a small-scale model of warehouse management vertical domain; the small-scale model of warehouse management vertical domain adopts a three-layer architecture of basic model, vertical domain adaptation layer and application plugin layer, wherein the application plugin layer includes warehouse location allocation plugin, AGV movement optimization plugin and exception handling plugin.

[0008] Intelligent database management tasks are performed based on a quaternion knowledge graph and a small-scale database management model within the database, including:

[0009] Intelligent inbound planning: The storage location allocation plugin is invoked to generate target storage location recommendations;

[0010] Dynamic inventory management execution: Based on order forecasts and inventory changes, the AGV movement optimization plugin is invoked to generate inventory transfer tasks and AGV scheduling instructions;

[0011] Real-time anomaly handling: Real-time detection of abnormal events, invoking the anomaly handling plugin to issue anomaly warnings and triggering the anomaly handling process.

[0012] In some embodiments of the present invention, constructing a quaternary knowledge graph data foundation includes multi-source data acquisition, data standardization processing, hierarchical extraction of entities and relations, and graph storage and updating; wherein, hierarchical extraction of entities and relations includes:

[0013] For unstructured data, the LSTM-CNN-CRF joint relation extraction model is used, which takes text word vectors as input and completes named entity recognition and relation extraction end-to-end.

[0014] For structured data, the R2RML mapping language is used to map relational data into resource description framework triples, including URI rule definition, attribute mapping and relation mapping;

[0015] For this structured data, the wrapper induction method is adopted, and the core fields and relationships are extracted through web page cleaning, sample labeling, rule learning and performance verification.

[0016] In some embodiments of the present invention, the storage location allocation plugin employs a storage location allocation algorithm that fuses graph embedding representation with neural factor machine, including:

[0017] Call the entity / relation vector of the four-element knowledge graph, input the entity base vector, graph association vector and dynamic scene vector, and concatenate them to form a fused embedding vector;

[0018] The fused vector is reduced in dimensionality, and Min-Max normalization is used to retain key information;

[0019] By integrating low-order and high-order features through feature cross-pooling and using the ReLU activation function to achieve non-linear transformation, the implicit adaptation relationship of entities is captured.

[0020] A three-layer feedforward neural network is constructed to learn the intrinsic correlation between goods and storage locations, and output a matching probability of 0-1.

[0021] Output the top storage locations in descending order of matching probability;

[0022] We receive feedback from manual adjustments, use the adjustment records as training samples, and update the graph embedding model parameters and NFM model weights monthly.

[0023] In some embodiments of the present invention, the AGV movement optimization plugin employs a hybrid algorithm of RRT* global exploration and bidirectional A* local optimization, including:

[0024] The RRT* algorithm is used to generate feasible paths from the starting point to the target point. ;

[0025] In feasible paths In the local area, the bidirectional A* algorithm is run with the local starting point and local ending point as the starting points for bidirectional search, the loss function is calculated, and the path with the minimum loss is searched iteratively. The starting point and ending point of each segment are searched for paths that are close to each other until the two paths converge.

[0026] The optimized local path segments replace the corresponding segments in the original RRT*, and are finally integrated to form a globally optimized path.

[0027] In some embodiments of the present invention, the anomaly handling plugin employs anomaly handling algorithms to achieve the perception, early warning, and closed-loop handling of anomalies within the warehouse, including:

[0028] Integrate three data sources—sensor data, vertical domain small model prediction, and manual reporting—and standardize their formats;

[0029] The alert level is determined and pushed out based on hierarchical rules;

[0030] Generate exception handling work orders and allocate work orders based on load balancing and skill matching;

[0031] Receive work orders through the terminal and execute operations according to the instructions;

[0032] Set a period to summarize anomaly handling data, analyze the distribution of anomaly types, the time taken to handle and the reasons for failure, and generate anomaly review reports.

[0033] In some embodiments of the present invention, the vertical domain small model prediction is implemented based on an anomaly prediction algorithm. The anomaly detection algorithm employs a two-stage model combining isolated forest and rule engine, inputting multi-source data within the data warehouse and outputting the anomaly type, warning level, and impact range.

[0034] The first stage of the isolated forest model uses historical data of anomalies within the warehouse as the training set to construct an isolated forest model. By isolating abnormal samples, it initially identifies candidate abnormal samples.

[0035] The second-stage rule engine model performs secondary screening of candidate samples based on the home appliance warehousing business rules, and outputs the final abnormal results.

[0036] This paper proposes a knowledge graph-driven knowledge base system for home appliance warehouse scenarios, including:

[0037] The data infrastructure layer includes a quaternary knowledge graph construction module, which uses goods, storage locations, AGVs, and equipment as core entities to construct a quaternary knowledge graph.

[0038] The algorithm engine layer deploys a warehouse management vertical domain small model module to build and train the warehouse management vertical domain small model. The warehouse management vertical domain small model adopts a three-layer architecture of basic model, vertical domain adaptation layer and application plugin layer. The application plugin layer includes warehouse location allocation plugin, AGV movement optimization plugin and exception handling plugin.

[0039] The application layer is used to execute intelligent warehouse management operations based on a quaternion knowledge graph and a small-scale warehouse management model. It includes an intelligent inbound planning module, a dynamic warehouse management execution module, and a real-time anomaly handling module.

[0040] The intelligent inbound planning module is used to call the storage location allocation plugin to generate target storage location recommendations;

[0041] The dynamic inventory management execution module is used to generate inventory transfer tasks and AGV scheduling instructions by calling the AGV movement optimization plugin based on order forecasts and inventory changes.

[0042] The real-time anomaly handling module is used to detect abnormal events in real time, call the anomaly handling plugin to issue anomaly warnings and trigger the anomaly handling process.

[0043] In some embodiments of the present invention, the system adopts a cloud-edge-device collaborative architecture; wherein,

[0044] The terminal layer includes in-warehouse AGVs, temperature and humidity sensors, positioning tags, and PDA operation terminals, which are responsible for data collection and operation execution.

[0045] In the edge layer, each warehouse deploys several edge nodes, which are responsible for high-frequency data preprocessing, real-time task execution, and breakpoint resume.

[0046] The cloud layer adopts a hybrid cloud architecture, with the private cloud storage warehouse containing all data, training small models for vertical domains, and formulating global strategies; the public cloud provides elastic computing power to support split views by region / category.

[0047] The terminal layer collects data and uploads it to the edge layer for preprocessing and real-time task execution. The edge layer synchronizes the processed data to the cloud layer for model training and global policy optimization. The cloud layer distributes the optimized policy to the edge layer to update the local decision model. The edge layer drives the terminal layer to execute the job, and the terminal layer feeds back the execution data to the cloud layer to form a closed loop.

[0048] Compared with existing technologies, the advantages and positive effects of this invention are as follows: By constructing a quaternary knowledge graph, this invention breaks down data silos within the warehouse, enabling decisions to be made based on the global relationships between goods, storage locations, and equipment. This significantly improves the accuracy of storage location allocation and the overall efficiency of AGV scheduling. Because it uses a vertically integrated small model that deeply integrates knowledge from the home appliance field for decision-making, rather than a general algorithm, the model can accurately respond to business constraints such as heavy loads and environmental sensitivity in tasks like storage location allocation and path planning, thereby greatly improving the scenario adaptability and business fit of the decision results. By designing a closed-loop handling model that integrates anomaly perception, graph-based location association, automatic work order dispatch, and equipment linkage, the system changes the traditional passive alarm mode of monitoring solutions, achieving proactive, rapid, and accurate responses to anomalies, thus effectively reducing the risk of damage to high-end goods. Furthermore, the deployment of the "cloud-edge-device" collaborative architecture places tasks with extremely high real-time requirements on the edge side for processing, reducing the pressure on the cloud. Complex model training and global optimization are placed in the cloud, balancing the stringent requirements of warehouse operations for instantaneous response with the global optimality of system decision-making, ensuring the high reliability and high availability of the entire system. The terminal receives and executes instructions and feeds back data, ensuring efficient operation of each layer.

[0049] Other features and advantages of the present invention will become clearer after reading the detailed description of the embodiments of the present invention in conjunction with the accompanying drawings. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is the execution flow of the quaternary knowledge graph-driven library method proposed in the home appliance warehouse scenario of this invention;

[0052] Figure 2 This is the overall architecture diagram of the knowledge base system driven by a quaternary knowledge graph in the home appliance warehouse scenario proposed in this invention;

[0053] Figure 3 This is a schematic diagram of the algorithm engine layer in the library system of this invention. Detailed Implementation

[0054] 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 embodiments of the present invention, and not all embodiments. 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.

[0055] This invention focuses on the core business processes of warehouse management in home appliance warehousing, taking "data-driven, algorithm-based decision-making, and scenario-based implementation" as its fundamental logical thread, and combining... Figure 1 and Figure 2 As shown, a three-layer collaborative technical architecture consisting of a "data foundation layer, algorithm engine layer, and execution application layer" is constructed. This architecture aims to achieve intelligent and refined operation of the entire warehouse management process, with the overall goal of systematically improving the adaptability of warehouse location resources, the scheduling efficiency of handling equipment, and effectively reducing the damage rate of goods.

[0056] The data foundation layer includes a warehouse-based four-element knowledge graph construction module. This module serves as the data foundation for the entire system, dedicated to integrating multi-source, heterogeneous data from four core elements within the warehouse: "goods, storage locations, AGVs, and equipment." By constructing a knowledge graph specific to the home appliance warehousing scenario, it achieves data standardization, correlation, and integrated governance. Its core value lies in providing unified, real-time, and high-quality data support for upper-level algorithmic decision-making, fundamentally solving the problem of data silos within the warehouse and ensuring the comprehensiveness and accuracy of decision-making basis.

[0057] The algorithm engine layer includes a warehouse management vertical model module. This module serves as the intelligent decision-making hub, focusing on developing vertical models deeply adapted to the home appliance warehousing scenario. This vertical model integrates three core sub-models: intelligent warehouse location allocation, AGV flow optimization, and anomaly prediction and closed-loop handling. During the decision-making process, it fully incorporates typical business characteristics of home appliances, such as "large, non-standard items, high-end products being environmentally sensitive, and drastic order fluctuations during peak seasons," thereby outputting management strategies that are both accurate and scenario-appropriate.

[0058] The application execution layer comprises three core business operation modules. This layer translates the algorithm engine's decisions into specific warehouse operation instructions, encompassing three core modules: intelligent inbound planning, dynamic inventory management execution, and real-time anomaly handling. The intelligent inbound planning module accurately matches goods characteristics with warehouse capacity and generates the optimal inbound plan. The dynamic inventory management execution module dynamically generates transfer tasks and AGV scheduling instructions based on real-time order and inventory status changes. The real-time anomaly handling module tracks and monitors all types of anomalies within the warehouse throughout the entire process, ensuring rapid response and a closed-loop management system.

[0059] First, the data infrastructure layer.

[0060] The quaternary knowledge graph construction module uses a quaternary knowledge graph construction algorithm to integrate, standardize, and correlate multi-source data from "cargo location-warehouse location-AGV-equipment" within the warehouse, breaking down data silos and building a high-quality graph data foundation to provide real-time and accurate data support for subsequent warehouse management decisions.

[0061] Quadruple knowledge graph algorithms include:

[0062] (1) Collect data from multiple sources.

[0063] It connects to IO data sources such as the warehouse WMS, AGV scheduling system, IoT platform, and manual data entry terminal, and collects the following data types:

[0064] Structured data: cargo attribute table, storage location capacity table, AGV operation log (obtained directly from the database, supporting incremental synchronization).

[0065] Semi-structured data: device alarm logs and manually reported anomaly information (key fields are extracted through JSON parsing and regular expression matching).

[0066] Unstructured data: warehouse layout diagram, AGV movement design diagram (coordinate information is extracted through OCR recognition).

[0067] The dynamic data (AGV location, equipment status) is updated once per second, and the static data (basic cargo attributes) is updated once per day.

[0068] (2) Data standardization processing.

[0069] Establish dedicated data standards within the warehouse to resolve data format conflicts across multiple systems:

[0070] Unified Coding Rule 1: Storage locations use the "Area-Type-Number" coding (e.g., "A01-High-001", where A01 is the area, High is the high-end product area, and 001 is the storage location number), and goods use the "Category-Value-Number" coding (e.g., "Refrigerator-High-001").

[0071] Unified data format: Date format is "YYYY-MM-DDHH:MM:SS", numerical values ​​are retained to 2 decimal places (e.g., weight 100.50kg), and Boolean values ​​are represented by 0 / 1.

[0072] Data cleaning: The data is cleaned in three steps: outlier removal (e.g., weight > 1000 kg is considered outlier) → missing value completion (filling with mean / median) → duplicate value removal. After cleaning, the accuracy rate is ≥ 98%.

[0073] (3) Extract entities and relations.

[0074] For the three types of unstructured, structured, and semi-structured data in home appliance warehousing, a hierarchical extraction method is adopted to ensure the accuracy of entities and relationships:

[0075] For the extraction of unstructured data (equipment logs, manual records), the "LSTM-CNN-CRF joint relation extraction model" is used. Taking text word vectors as input, it completes named entity recognition and relation extraction end-to-end. For example, from "AGV-001 stopped due to overloading when carrying a 50kg refrigerator", extract entities "AGV-001" and "50kg refrigerator", and the relation "AGV-001 - overloaded - 50kg refrigerator"; from "The temperature and humidity in storage location A01 exceeded the standard, and the storage of high-end air conditioners was abnormal", extract entities "storage location A01" and "high-end air conditioner", and the relation "storage location A01 - storage abnormal - high-end air conditioner". The overall extraction accuracy is ≥97%.

[0076] For the extraction of structured data (WMS / AGV scheduling system tables), the "R2RML mapping language" is used to formulate exclusive mapping rules for home appliance warehousing to achieve standardized data conversion. Among them, the URI rule: the goods URI is uniformly "http: / / rrs.com / hw / {encoding}" (such as "http: / / rrs.com / hw / HW-refrigerator 50kg"), and the storage location URI is uniformly "http: / / rrs.com / kw / {encoding}" (such as "http: / / rrs.com / kw / KW-A01"); property mapping: map the "weight in the goods table" field to the "<home appliance weight>" property, map "the maximum bearing capacity in the storage location table" to the "<maximum bearing capacity>" property, and map "the load capacity in the AGV table" to the "<maximum load>" property; relation mapping: associate through "the storage location ID in the goods table" and "the storage location ID in the storage location table" to automatically generate the "<HW-refrigerator 50kg> - <stored in> - " triple; associate through "the compatible goods type in the AGV table" and "the category in the goods table" to generate the " - <compatible with> - <HW-refrigerator 50kg>" triple, and the mapping has no error.

[0077] For semi-structured data extraction (warehouse location webpage, scheduling report), the wrapper induction method is adopted, and the extraction is completed in four steps. (1) Webpage cleaning: remove advertisements and redundant code in the warehouse location layout webpage, and retain the core fields "warehouse location number-channel width-load limit-temperature and humidity range"; (2) Sample labeling: manually label 100 home appliance scene samples (such as "KW-A01-3m-500kg-18-25℃" and "KW-B02-2.5m-300kg-10-30℃") to train the wrapper; rule learning: the wrapper automatically learns the field matching rules; performance evaluation: verify the performance of the wrapper by "accuracy ≥ 95% + recall ≥ 93%", and finally extract attributes and relationships such as "warehouse location-channel width" and "warehouse location-adjacent AGV" to support the subsequent map construction.

[0078] (4) Map storage and updating.

[0079] A dual-storage architecture of graph database + relational database is adopted: the graph database (Neo4j) stores entity nodes and relationships, supports efficient relational queries, and the single query response time is ≤50ms; the relational database (MySQL) backs up static data (such as basic cargo attributes and storage location capacity parameters) to ensure data security.

[0080] Update mechanism: Dynamic data is synchronized in real time via WebSocket, static data is updated incrementally every day at midnight, and full data is verified every month. A retry mechanism is triggered when an update fails (the number of retries is ≤3) to ensure the integrity of the map data.

[0081] Second, the algorithm engine layer.

[0082] Warehouse management vertical domain small model modules such as Figure 3 As shown, a three-layer architecture of basic model + vertical domain adaptation layer + application plugins is adopted. For the specific vertical domain of home appliance warehousing, a small-scale artificial intelligence module is trained by integrating business rules and domain characteristics to adapt to the specific needs of the home appliance warehousing scenario, thereby achieving higher scenario decision-making accuracy.

[0083] The base model layer uses LightGBM as the base model, which provides general feature extraction and classification capabilities. The initial model parameters are set as follows: learning rate 0.01, tree depth 8, and number of leaf nodes 32.

[0084] The vertical adaptation layer incorporates exclusive features of home appliance warehousing, including the load-bearing constraints of large appliances, the temperature and humidity sensitivity of high-end products, and the frequency of outbound shipments during peak seasons. Through feature cross-referencing (such as cargo weight × warehouse load-bearing capacity) and feature selection (using mutual information to screen the top 20 key features), the model's scenario adaptability is improved.

[0085] The application plugin layer contains three core functional plugins: a warehouse location allocation plugin (outputs the top 3 recommended warehouse locations), an AGV movement optimization plugin (generates the optimal driving path), and an anomaly prediction and closed-loop handling plugin (outputs the anomaly type and warning level). The plugins are linked through API interfaces and can be configured to be called individually or in combination.

[0086] The warehouse location allocation algorithm employs a multimodal integration method combining Knowledge Graph Embedding Representation (KG-BERT) and Neural Factor Machine (NFM) to construct a warehouse location allocation algorithm based on feature embedding, cross-fusion, and matching decision. It captures implicit relationships between entities through knowledge graphs and combines multimodal feature fusion to output matching probabilities, achieving dynamic priority allocation of warehouse locations. The objective function is to maximize the matching probability between goods and warehouse locations, including:

[0087] (1) Call the entity / relation vector generated by the four-element knowledge graph model, input the following three types of core features and convert them into low-dimensional fusion vectors to achieve deep representation of entity associations:

[0088] Entity base vector: 128-dimensional base vectors for goods (weight, storage requirements, high-end product labels) and storage locations (load-bearing capacity, temperature and humidity, location), among which the weight of the "high-end product goods" vector is increased by 20%, and the weight of the "load-bearing capacity adaptation" dimension in the "large goods (weight > 50kg)" vector is increased by 30%.

[0089] Graph Association Vector: Extracts 64-dimensional vectors from the quaternary knowledge graph to represent relationships such as "goods-historical storage location" and "storage location-adjacent AGV". For example, if "goods were once stored in a storage location near the export location", the "efficiency dimension" score in the association vector will be improved.

[0090] Dynamic Scenario Vector: A 32-dimensional vector containing dynamic data such as real-time AGV load and current warehouse congestion. For example, if AGVs around the warehouse are congested, the score for the "Dynamic Adaptation" dimension decreases.

[0091] The three types of vectors are concatenated to form a 224-dimensional fused embedding vector, which provides a basic representation for subsequent matching.

[0092] (2) The NFM model achieves accurate matching of goods and storage locations through an embedding layer, a dual interaction layer, and a feedforward neural network (FNN) layer.

[0093] The embedding layer reduces the dimensionality of the 224-dimensional fused vector to 64-dimensional vectors and uses Min-Max normalization to retain more than 95% of the key information, thus solving the problem of data sparsity (such as the problem of missing attributes for low-frequency goods).

[0094] The dual-interaction layer integrates low-order and high-order features through feature cross-pooling. Low-order features include "cargo weight × warehouse load-bearing capacity" and "cargo temperature and humidity requirements × warehouse temperature and humidity". High-order features include "high-end product label × warehouse exclusive area attribute" and "cargo outbound frequency × warehouse to exit distance". The ReLU activation function is used to achieve non-linear transformation and capture the implicit adaptation relationship of entities.

[0095] The FNN layer (probability output) constructs a 3-layer feedforward neural network (the number of hidden layer nodes is 64→32→16), learns the inherent correlation between goods and storage locations, and outputs a matching probability of 0-1 (the probability value corresponds to the original "fitness score", such as 0.88 corresponding to the original 88 points). The probability calculation incorporates the business dimensions of "fitness (weight 0.5) + efficiency (weight 0.3) + security (weight 0.2)", which is consistent with the original scoring logic.

[0096] (3) Output results and iterative optimization.

[0097] Recommended output: Output the top 3 storage locations in descending order of matching probability, along with "matching dimension breakdown" (e.g., "Storage location KW-A01: basic adaptation 0.85 (load-bearing capacity + temperature and humidity) + associated adaptation 0.92 (historical storage records) + dynamic adaptation 0.88 (AGV load), total score 0.88"). Supports "one-click adoption" or manual adjustment. Adjustment records are automatically marked as "manual feedback samples".

[0098] Model iteration: Every month, "human feedback samples" (such as reasons for rejecting recommended locations and adjusted matching results) are added to the training set to update the graph embedding model parameters and NFM model weights, ensuring that the matching accuracy continues to improve, with a target accuracy of ≥94%, which is 4 percentage points higher than the original solution.

[0099] The model constraints include load constraints, temperature and humidity constraints, and region constraints.

[0100] The AGV path optimization algorithm adopts a hybrid algorithm of two-stage fusion fast random exploration tree (RRT*) + bidirectional A*. It generates feasible paths through global exploration and improves path accuracy through local fine optimization. It explores areas in unknown environments and calculates the optimal path scheme.

[0101] The RRT* algorithm is used to quickly find a feasible, but not necessarily optimal, path from the starting point to the target point, including:

[0102] (1) Construct the RRT* tree: including nodes This represents the position of the AGV in three-dimensional space, initially including only the starting point. Tree structure , For a set of nodes, Let edge represent the set of feasible path segments between adjacent nodes; starting point Target location .

[0103] (2) Define the path cost as the cumulative Euclidean distance from the starting point to the node:

[0104] ;

[0105] Tree Initialization, containing only the starting point : .

[0106] (3) Start the loop iteration: Each iteration includes sampling the target point and finding the nearest neighbor node:

[0107] In three-dimensional space With probability Towards Perform random sampling:

[0108] ;

[0109] In the current tree Find the distance The nearest node In the formula To calculate the Euclidean distance.

[0110] Expand new nodes, from Towards Expand a new node in the direction :

[0111] ;

[0112] And for each Constraints are imposed: safe distance, main passage width, and avoidance of collisions with static obstacles and dynamically sensed obstacles in real time.

[0113] definition for Nearby node set, select node This makes the connection arrive Path cost Minimum:

[0114] :

[0115] Node join in and with It is the parent node.

[0116] Continuously optimize the tree Until finally getting a way from arrive Low-cost feasible initial path .

[0117] Feasible paths generated by the bidirectional A* algorithm in RRT* Above, for complex, narrow, and obstacle-ridden areas (such as densely packed shelving areas and corners), a local bidirectional A* fine-grained search is performed to optimize the accuracy, length, and safety of the path, including:

[0118] (1) For the path Divided into several key segments according to geometric features, each segment starting from... ,end .

[0119] (2) For each path segment, construct a 3D space and normalize it to a 3D model. A grid, and marking the grid, including , And whether it is permissible to pass.

[0120] (3) Select a new node ,against Construct a loss function as a path guide, including the distance and cost from the current node to the target. (Including energy consumption, time, and congestion):

[0121] .

[0122] The bidirectional A* algorithm is run on a local map using the above method to calculate the loss function, continuously optimize the planning method, and iteratively find the path with the minimum loss. For each segment, a path search is performed to find the closest possible points to the starting and ending points until the two paths converge.

[0123] The optimized local path segments replace the corresponding segments in the original RRT*, and are finally integrated to form a globally optimized path. This meets the actual needs of AGV handling tasks in home appliance warehousing.

[0124] The anomaly handling algorithm enables rapid detection, accurate early warning, efficient handling, and closed-loop review of anomalies within the warehouse, shortening the anomaly response time (level 3 early warning response time ≤ 10 min), reducing cargo loss caused by anomalies (high-end product damage rate ≤ 1.2%), and accumulating anomaly handling experience to optimize the handling process.

[0125] The algorithm is executed in five steps, forming a closed-loop process, as detailed below:

[0126] (1) Integrate three data sources: sensors, vertical domain small model prediction, and manual reporting. Access abnormal information in real time through data interface. The unified format is "Abnormal ID-Type-Occurrence Time-Location-Description", such as "ERR20240601001-Temperature and Humidity Exceed Standard-2024-06-0109:30:00-A01 Storage Location-Temperature 28℃ (Requirement 18-25℃)", with a collection delay of ≤1s.

[0127] Among them, the anomaly prediction algorithm of the vertical domain small model adopts a two-stage model of isolated forest + rule engine. It takes multi-source data in the warehouse as input and outputs anomaly type, warning level and impact range, including:

[0128] Phase 1 (Isolation Forest): Using 12 months of abnormal historical data (temperature and humidity, equipment voltage, AGV operating parameters, etc.) in the warehouse as the training set, an isolation forest model is constructed. By isolating abnormal samples (abnormal data is more easily isolated in the feature space), abnormal candidate samples are initially identified, with an identification accuracy of ≥85%.

[0129] The second stage (rule engine) involves secondary screening of candidate samples based on the home appliance warehousing business rules, such as "temperature and humidity exceeding the standard for more than 5 minutes and involving high-end product storage locations → Level 3 warning" and "AGV voltage < 36V and current load > 50% → Level 2 warning". The final abnormal results are output (abnormality type, warning level, list of affected storage locations / equipment).

[0130] Model iteration: Add new abnormal data (after labeling) to the training set every month, update the model parameters, and ensure that the prediction accuracy continues to improve (accuracy improvement ≥2% per quarter).

[0131] (2) Based on the hierarchical rules of the anomaly handling model, the warning level is automatically determined and pushed to the responsible person through multiple channels:

[0132] Level 1 alert: A system message is pushed to the warehouse administrator;

[0133] Level 2 alert: System message + SMS push to warehouse administrators and maintenance personnel;

[0134] Level 3 alert: The system message, SMS and telephone notification will be sent to the warehouse supervisor and operations supervisor, and the responsible person will be required to confirm receipt within 10 minutes.

[0135] (3) Automatically generate anomaly handling work orders, including "anomaly details, handling requirements, time limits, and associated resources (such as backup AGV number, temperature and humidity control equipment ID)", and allocate work orders using the "load balancing + skill matching" principle:

[0136] Load balancing: Prioritize assigning tasks to the person with the least current workload;

[0137] Skill matching: Equipment failure work orders are assigned to maintenance personnel, and temperature and humidity work orders are assigned to warehouse managers;

[0138] Work order allocation time is ≤2 minutes, and manual reassignment is supported (the reason must be stated when reassigning).

[0139] (4) Receive work orders through the terminal, perform operations according to the handling instructions (e.g., if the temperature and humidity exceed the standard, start the air conditioning dehumidification equipment), and submit the handling results (including on-site photos and data change curves) after the operation is completed.

[0140] The system automatically verifies the handling effect: such as whether the temperature and humidity have returned to the normal range and whether the equipment has resumed operation. If the verification passes, the work order is closed; if the verification fails, a second warning is triggered (upgraded to level one). The handling execution time is: Level 3 warning ≤ 10 min, Level 2 warning ≤ 30 min, and Level 1 warning ≤ 2 h.

[0141] (5) Summarize the abnormal handling data weekly, analyze the "abnormal type distribution, handling time and failure reasons", and output the "Abnormal Review Report"; optimize preventive measures (such as increasing sensor density) for high-frequency abnormalities (such as exceeding temperature and humidity limits more than 5 times per month); optimize resource allocation (such as increasing backup equipment) for links with long handling time (such as equipment maintenance more than 1 hour) to continuously improve the efficiency of abnormal handling.

[0142] Third, execute the application layer.

[0143] The intelligent inbound planning module, as the entry point for warehouse management, achieves precise matching of "goods and storage locations," avoiding mismatched storage locations from the source and supporting the standardization of home appliance inbound operations. It analyzes goods characteristics, calls the storage location allocation plugin in the algorithm engine layer, and performs global matching calculations based on the real-time status of the current quaternary knowledge graph (occupancy status of each storage location, AGV congestion, equipment operating status, etc.), outputting the top 3 recommended storage locations and their matching probabilities.

[0144] The dynamic inventory management module dynamically adjusts the warehouse layout based on order forecasts and inventory changes to improve the efficiency of appliance outbound operations and adapt to order fluctuation scenarios (such as peak sales seasons). Based on order forecasts and warehouse status analysis, it calls the AGV movement optimization algorithm in the algorithm engine layer to generate a global optimized path for AGVs and matches the AGV model. It collects the location nodes of AGVs in real time and monitors whether AGVs follow path constraints (minimum turning radius, maximum lifting speed, maximum horizontal speed). If a sudden dynamic obstacle is encountered (such as temporary goods stacking), it triggers local bidirectional AGV re-optimization to generate a new local path segment.

[0145] The real-time anomaly handling module detects anomalies in the warehouse in real time, calls the anomaly handling algorithm to automatically trigger the anomaly handling process, and tracks the progress to reduce damage to home appliances and ensure the continuity of warehouse operations.

[0146] To address the above, this invention constructs a cloud-edge-device collaborative architecture, connecting the entire chain from data collection, processing, decision-making to execution within the warehouse, adapting to the real-time response and dynamic adjustment needs of home appliance warehousing, and ensuring the stable operation and efficient collaboration of each module.

[0147] The terminal layer includes equipment such as AGVs in the warehouse, temperature and humidity sensors, positioning tags, and PDA operation terminals, which are responsible for data collection (sensors collect temperature and humidity, and AGVs collect position) and work execution (AGVs perform handling and PDAs receive work orders).

[0148] In the edge layer, each warehouse deploys 1-2 edge nodes, which are responsible for high-frequency data preprocessing (such as AGV position filtering and sensor data noise reduction), real-time task execution (AGV dynamic obstacle avoidance and abnormal early warning push), and breakpoint resume (caching data when the network is interrupted and synchronizing it after recovery).

[0149] The cloud layer adopts a hybrid cloud architecture, with the private cloud storing all data (12 months of historical data + real-time data), training small models for vertical domains, and formulating global strategies (such as cross-warehouse management rules); the public cloud provides elastic computing power (expanding capacity during peak seasons) and supports splitting views by region / category (such as home appliance warehouse and high-end product warehouse in East China).

[0150] The process involves: terminal data collection → edge layer preprocessing and real-time task execution → cloud layer model training and global strategy formulation → edge layer optimization strategy → terminal job execution → terminal feedback of execution data to the cloud, forming a closed loop.

[0151] It should be noted that, in the specific implementation process, the above-mentioned control part can be implemented by a hardware processor executing computer-executable instructions in software form stored in memory, which will not be elaborated here. The programs corresponding to the actions performed by the above control circuit can all be stored in the computer-readable storage medium of the system in software form, so that the processor can call and execute the operations corresponding to the above modules.

[0152] The computer-readable storage media mentioned above may include volatile memory, such as random access memory; may also include non-volatile memory, such as read-only memory, flash memory, hard disk or solid-state drive; and may also include combinations of the above types of memory.

[0153] The term "processor" as mentioned above can also refer to a collective of multiple processing elements. For example, a processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors, 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. A general-purpose processor can be a microprocessor or any conventional processor, and it can also be a special-purpose processor.

[0154] It should be noted that the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A knowledge graph-driven library management method for home appliance warehouse scenarios, characterized in that, include: A four-element knowledge graph is constructed with goods, storage locations, AGVs, and equipment as core entities; Construct and train a small-scale model of warehouse management vertical domain; the small-scale model of warehouse management vertical domain adopts a three-layer architecture of basic model, vertical domain adaptation layer and application plugin layer, wherein the application plugin layer includes warehouse location allocation plugin, AGV movement optimization plugin and exception handling plugin. Intelligent database management tasks are performed based on a quaternion knowledge graph and a small-scale database management model within the database, including: Intelligent inbound planning: The storage location allocation plugin is invoked to generate target storage location recommendations; Dynamic inventory management execution: Based on order forecasts and inventory changes, the AGV movement optimization plugin is invoked to generate inventory transfer tasks and AGV scheduling instructions; Real-time anomaly handling: Real-time detection of abnormal events, invoking the anomaly handling plugin to issue anomaly warnings and triggering the anomaly handling process.

2. The method for refining a knowledge base based on a quaternary knowledge graph in the home appliance warehouse scenario according to claim 1, characterized in that, Constructing a quaternary knowledge graph data foundation includes multi-source data acquisition, data standardization processing, hierarchical extraction of entities and relations, and graph storage and updating; among which, hierarchical extraction of entities and relations includes: For unstructured data, the LSTM-CNN-CRF joint relation extraction model is used, which takes text word vectors as input and completes named entity recognition and relation extraction end-to-end. For structured data, the R2RML mapping language is used to map relational data into resource description framework triples, including URI rule definition, attribute mapping and relation mapping; For this structured data, the wrapper induction method is adopted, and the core fields and relationships are extracted through web page cleaning, sample labeling, rule learning and performance verification.

3. The method for refining a database based on a quaternary knowledge graph in the home appliance warehouse scenario according to claim 1, characterized in that, The storage location allocation plugin employs a storage location allocation algorithm that fuses graph embedding representations and neural factor machines, including: Call the entity / relation vector of the four-element knowledge graph, input the entity base vector, graph association vector and dynamic scene vector, and concatenate them to form a fused embedding vector; The fused vector is reduced in dimensionality, and Min-Max normalization is used to retain key information; By integrating low-order and high-order features through feature cross-pooling and using the ReLU activation function to achieve non-linear transformation, the implicit adaptation relationship of entities is captured. A three-layer feedforward neural network is constructed to learn the intrinsic correlation between goods and storage locations, and output a matching probability of 0-1. Output the top storage locations in descending order of matching probability; We receive feedback from manual adjustments, use the adjustment records as training samples, and update the graph embedding model parameters and NFM model weights monthly.

4. The method for refining a database based on a quaternary knowledge graph in the home appliance warehouse scenario according to claim 1, characterized in that, The AGV movement optimization plugin employs a hybrid algorithm combining RRT* global exploration and bidirectional A* local optimization, including: The RRT* algorithm is used to generate feasible paths from the starting point to the target point. ; In feasible paths In the local area, the bidirectional A* algorithm is run with the local starting point and local ending point as the starting points for bidirectional search, the loss function is calculated, and the path with the minimum loss is searched iteratively. The starting point and ending point of each segment are searched for paths that are close to each other until the two paths converge. The optimized local path segments replace the corresponding segments in the original RRT*, and are finally integrated to form a globally optimized path.

5. The method for refining a database based on a quaternary knowledge graph in the home appliance warehouse scenario according to claim 1, characterized in that, The anomaly handling plugin employs anomaly handling algorithms to detect, warn of, and handle anomalies within the warehouse, including: Integrate three data sources—sensor data, vertical domain small model prediction, and manual reporting—and standardize their formats; The alert level is determined and pushed out based on hierarchical rules; Generate exception handling work orders and allocate work orders based on load balancing and skill matching; Receive work orders through the terminal and execute operations according to the instructions; Set a period to summarize anomaly handling data, analyze the distribution of anomaly types, the time taken to handle and the reasons for failure, and generate anomaly review reports.

6. The method for resolving data in a home appliance warehouse scenario based on a quaternary knowledge graph as described in claim 5, characterized in that, The vertical domain small-model prediction is based on an anomaly prediction algorithm. This anomaly detection algorithm employs a two-stage model combining isolated forest and rule engine, inputting multi-source data from within the data repository and outputting the anomaly type, warning level, and impact range. The first stage of the isolated forest model uses historical data of anomalies within the warehouse as the training set to construct an isolated forest model. By isolating abnormal samples, it initially identifies candidate abnormal samples. The second-stage rule engine model performs secondary screening of candidate samples based on the home appliance warehousing business rules, and outputs the final abnormal results.

7. A knowledge graph-driven database system for home appliance warehouse scenarios, characterized in that, include: The data infrastructure layer includes a quaternary knowledge graph construction module, which uses goods, storage locations, AGVs, and equipment as core entities to construct a quaternary knowledge graph. The algorithm engine layer deploys a warehouse management vertical domain small model module to build and train the warehouse management vertical domain small model. The warehouse management vertical domain small model adopts a three-layer architecture of basic model, vertical domain adaptation layer and application plugin layer. The application plugin layer includes warehouse location allocation plugin, AGV movement optimization plugin and exception handling plugin. The application layer is used to execute intelligent warehouse management operations based on a quaternion knowledge graph and a small-scale warehouse management model. It includes an intelligent inbound planning module, a dynamic warehouse management execution module, and a real-time anomaly handling module. The intelligent inbound planning module is used to call the storage location allocation plugin to generate target storage location recommendations; The dynamic inventory management execution module is used to generate inventory transfer tasks and AGV scheduling instructions by calling the AGV movement optimization plugin based on order forecasts and inventory changes. The real-time anomaly handling module is used to detect abnormal events in real time, call the anomaly handling plugin to issue anomaly warnings and trigger the anomaly handling process.

8. The knowledge graph-driven system for appliance storage scenarios according to claim 7, characterized in that, The system adopts a cloud-edge-device collaborative architecture; wherein, The terminal layer includes in-warehouse AGVs, temperature and humidity sensors, positioning tags, and PDA operation terminals, which are responsible for data collection and operation execution. In the edge layer, each warehouse deploys several edge nodes, which are responsible for high-frequency data preprocessing, real-time task execution, and breakpoint resume. The cloud layer adopts a hybrid cloud architecture, with the private cloud storage warehouse containing all data, training small models for vertical domains, and formulating global strategies; the public cloud provides elastic computing power to support split views by region / category. The terminal layer collects data and uploads it to the edge layer for preprocessing and real-time task execution. The edge layer synchronizes the processed data to the cloud layer for model training and global policy optimization. The cloud layer distributes the optimized policy to the edge layer to update the local decision model. The edge layer drives the terminal layer to execute the job, and the terminal layer feeds back the execution data to the cloud layer to form a closed loop.

Images