An object identifier-based smart laundry sorting system and device

Abstract

This invention discloses an intelligent clothing sorting system and device based on object identifiers, comprising: an initial feature construction module, which collects unique identifiers and multi-dimensional feature vectors to generate an initial feature set; a hierarchical heterogeneous modeling module, which performs node mapping to construct an initial hierarchical heterogeneous graph model; a cross-hierarchical feature alignment module, which uses a hierarchical contrast loss function to process node features and generate an initial hierarchical alignment model; a diffusion logic reconstruction module, which introduces a logical diffusion guidance mechanism and uses logical gradients to reconstruct the feature distribution structure; an iterative denoising decision module, which combines flow logic to perform inverse denoising iteration and generate sorting decision instructions; and a feedback scheduling optimization module, which performs sorting scheduling and adjusts feature weights based on feedback data. This invention significantly improves sorting accuracy and scheduling flexibility in complex environments through deep coupling of logical diffusion and hierarchical graph modeling.

Description

Technical Field

[0001] This invention relates to the field of logistics automation and intelligent sorting data processing technology, and in particular to an intelligent clothing sorting system and device based on object identifiers. Background Technology

[0002] With the widespread application of RFID and computer vision technologies in the logistics field, automated garment sorting systems have become a core component of the modern laundry industry and apparel management. Existing sorting technologies primarily rely on object identifiers for task matching. Sensors collect unique identification codes and basic image features of garments, and a pre-defined logical rule base drives sorting arms or conveyor belts to perform physical displacement, thereby classifying the garments.

[0003] In practical applications, garment sorting involves complex hierarchical attributes and variable flow constraints. Existing sorting systems mostly employ shallow feature matching mechanisms, determining scheduling paths by establishing simple lookup table mapping relationships. This approach struggles to uncover the deep relationships between garment static attributes, geospatial locations, and business logic tasks when processing massive and heterogeneous sorting data. This leads to incomplete feature representation or logical redundancy when facing high-concurrency tasks or complex process constraints. Furthermore, existing technologies rely heavily on hard-coded rules in the logic verification stage, lacking the ability to guide feature distribution within a high-dimensional solution space. This means that sorting decisions, when handling conflicting scheduling instructions, often fail to balance path cost minimization and logical compliance, limiting the reliability and intelligence level of automated production lines.

[0004] Specifically, when sorting tasks involve dynamic feedback, existing data processing models often treat each attribute as an independent dimension, ignoring the semantic alignment requirements of heterogeneous data across different levels. This not only reduces the accuracy of feature extraction but also makes the system susceptible to noise interference when performing reverse decision-making reasoning, making it difficult to produce highly consistent execution sequences. Due to the lack of an effective logical gradient guidance mechanism for structured reconstruction of model weights, feedback data during sorting scheduling is also difficult to translate into effective corrections to the model's feature distribution in real time and accurately, thus limiting the system's adaptive adjustment capabilities under complex operating conditions.

[0005] Therefore, how to provide a smart clothing sorting system and device based on object identifiers is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to propose an intelligent clothing sorting system and device based on object identifiers. Addressing the technical problems of weak feature correlation and lack of high-dimensional guidance capability in logic verification when handling heterogeneous attributes in existing sorting systems, this invention proposes a technical solution that constructs a multi-dimensional attribute mapping relationship through hierarchical heterogeneous modeling and cross-level feature alignment. It also introduces a logic diffusion guidance mechanism to reconstruct the feature distribution using the logical gradient generated by business rules, combined with inverse denoising and iterative generation of sorting decision instructions. This invention possesses the ability to deeply explore the deep coupling relationship between static attributes of clothing and flow logic, significantly improving the logical consistency of sorting instructions and the reliability of scheduling paths under complex working conditions, achieving the beneficial effects of accurate decision-making and adaptive feedback updates for massive heterogeneous sorting tasks.

[0007] According to an embodiment of the present invention, a smart clothing sorting system and apparatus based on object identifiers includes the following modules: Original feature construction module: It obtains the unique identifier code, multi-dimensional feature vector and preset clothing flow logic of the clothing to be processed through the data acquisition interface, and retrieves the associated business attribute set according to the unique identifier code to obtain the original feature set containing the static attributes of clothing and flow constraints. Hierarchical heterogeneous modeling module: Performs logical node mapping and associated edge construction processing on the original feature set, converting the unique identifier, multi-dimensional feature vector and business attribute set into interrelated object nodes, location nodes and task nodes, to obtain an initial hierarchical heterogeneous graph model arranged in topology; Cross-level feature alignment module: Based on the initial hierarchical heterogeneous graph model, the module uses the hierarchical contrast loss function to perform association constraint processing on the node features of different levels, thereby obtaining an initial hierarchical alignment model with multi-dimensional attribute mapping relationships; Diffusion logic reconstruction module: In the initial hierarchical alignment model, a logic diffusion guidance mechanism is introduced. The logic gradient generated in the high-dimensional solution space by the preset classification rules is used to reconstruct the feature distribution structure in the initial hierarchical alignment model, so as to obtain the improved hierarchical heterogeneous graph model. Iterative denoising decision module: Based on the improved hierarchical heterogeneous graph model, it extracts the reconstructed feature vector of the object node after feature distribution reconstruction, and performs reverse denoising iterative processing in combination with the preset clothing flow logic to obtain sorting decision instructions; Feedback scheduling optimization module: Based on the sorting decision instructions, it executes the corresponding sorting scheduling and distribution processing, and collects the status feedback data after the transfer and returns it to the improved hierarchical heterogeneous graph model to obtain the adjusted feature weights.

[0008] Optionally, the original feature construction module specifically includes: Collect raw label data, multispectral image data, production scheduling data, and material attribute data of the clothing to be processed; The collected raw tag data, multispectral image data, production scheduling data, and material attribute data are preprocessed. The preprocessing includes unique identifier parsing, pixel feature extraction, priority mapping, and data standardization to obtain preprocessed multidimensional feature data. Based on the preprocessed multidimensional feature data, field aggregation processing is performed to construct an original feature set containing the static attributes and circulation constraints of clothing.

[0009] Optionally, the hierarchical heterogeneous modeling module specifically includes: Based on the unique identifier and multidimensional feature vector in the original feature set, entities with unique attribute identifiers are generated using preset object instantiation rules, thus obtaining object nodes; Extract the clothing flow logic from the original feature set, and perform spatial coordinate mapping based on the preset physical sorting points and logical workstation information in the clothing flow logic to obtain the location nodes; The business attribute set in the original feature set is analyzed, the washing requirements and delivery time limit parameters are extracted, and the washing requirements and delivery time limit parameters are converted into action constraints in the sorting process of the clothes to be processed, thus obtaining task nodes; Using preset hierarchical partitioning rules, object nodes, location nodes, and task nodes are deployed to the entity layer, spatial layer, and logical layer of the heterogeneous graph structure, respectively, to obtain a hierarchical node set; Based on the clothing flow logic, establish the affiliation association between object nodes and location nodes, establish the trigger association between task nodes and location nodes, and establish logical associations according to the execution order between object nodes and task nodes to obtain a set of associated edges; By using the set of associated edges to perform topological connection processing on the set of hierarchical nodes, an initial hierarchical heterogeneous graph model arranged according to the topological structure is constructed.

[0010] Optionally, the cross-level feature alignment module specifically includes: Based on the original feature vectors of object nodes, location nodes and task nodes in the initial hierarchical heterogeneous graph model, the original feature vectors of object nodes, location nodes and task nodes are transformed into a hidden feature space of a unified dimension using a preset linear mapping function, thus obtaining a cross-level basic feature set. Calculate the spatial distribution consistency between object nodes and associated location nodes in the cross-level basic feature set, and calculate the logical association strength between object nodes and corresponding task nodes to obtain the inter-level association weight coefficient. The hierarchical contrastive loss function is used to perform contrastive enhancement processing on the cross-level basic feature set. By maximizing the mutual information between positive sample pairs and minimizing the semantic association between negative sample pairs, the contrastive loss value is obtained. The initial hierarchical heterogeneous graph model is iteratively optimized using the contrastive loss value. The coordinate distribution of object nodes, position nodes, and task nodes in the hidden feature space is adjusted according to the inter-level correlation weight coefficients to obtain the hierarchical heterogeneous feature matrix after feature alignment. Semantic features of position nodes and task nodes that have topological connections with object nodes in the initial hierarchical heterogeneous graph model are retrieved. Attention weighting is used to aggregate the semantic features of position nodes and task nodes into the empty attribute bits of object nodes. Value pairs of logical attributes of object nodes are filled to obtain attribute-enhanced feature vectors. The attribute-enhanced feature vectors are mapped to the node index positions corresponding to the hierarchical heterogeneous feature matrix to construct an initial hierarchical alignment model with multi-dimensional attribute mapping relationships.

[0011] Optionally, the diffusion logic reconstruction module specifically includes: The business logic and scheduling priority constraints in the preset classification rules are analyzed, and a logical gradient representing the compliance trend of features is constructed in the feature space defined by the initial hierarchical alignment model. The node feature distribution in the initial hierarchical alignment model is regarded as the initial probability distribution through the logical diffusion guidance mechanism, and the logical gradient is set as the external guiding term in the feature evolution process. Based on the logical gradient, the abnormal feature components that violate the preset classification rules in the initial probability distribution are offset, and the compliant feature components that conform to the preset classification rules are clustered, resulting in the feature distribution space after the execution logic converges. Projection alignment is performed on the feature distribution space after the execution logic converges, and the evolved feature components are remapped to the topology node index positions of the initial hierarchical alignment model to obtain the reconstructed feature distribution structure. The reconstructed feature distribution structure is injected into the initial hierarchical alignment model, and the constraint force generated by the logical gradient is solidified by updating the model weight parameters to obtain the improved hierarchical heterogeneous graph model.

[0012] Optionally, the iterative denoising decision module specifically includes: Based on the improved hierarchical heterogeneous graph model, the reconstructed feature vectors of object nodes in the entity layer are extracted, and the reconstructed feature vectors are mapped to the initial state distribution of the discrete scheduling space to obtain the initial latent variable sequence. Based on the preset clothing flow logic, the spatiotemporal topological constraints and path cost functions of the physical sorting points are extracted, and the spatiotemporal topological constraints and path cost functions of the physical sorting points are converted into logical guiding terms in the reverse denoising process. Based on the initial latent variable sequence, the weight parameters in the improved hierarchical heterogeneous graph model are called to perform single-step denoising. During the single-step denoising process, a logical guiding term is introduced to perform distribution correction, and intermediate state feature vectors are obtained. The intermediate state feature vector is used to perform multiple rounds of inverse denoising iteration. The intermediate state feature vector is continuously subjected to manifold constraint projection through the logical guiding term until the intermediate state feature vector meets the preset convergence threshold, thus obtaining the logical consistency representation vector. Perform feature decoding and symbolic mapping on the logical consistency representation vector to generate and reconstruct the discrete sorting execution sequence corresponding to the feature vector; Align the discrete sorting execution sequence with the real-time status of the physical sorting points to obtain sorting decision instructions.

[0013] Optionally, the feedback scheduling optimization module specifically includes: Based on sorting decision instructions, the sorting and processing of the clothing to be processed is completed; The system collects execution result data in real time during the sorting process, compares the execution result data with the preset clothing flow logic, and extracts status feedback parameters including path deviation and time consumption. Based on the state feedback parameters, the hierarchical heterogeneous graph model is updated to obtain the updated hierarchical heterogeneous graph model.

[0014] Optionally, the object identifier-based intelligent clothing sorting device stores a program that, when executed, enables the object identifier-based intelligent clothing sorting device to perform the object identifier-based intelligent clothing sorting system according to any one of claims 1 to 7.

[0015] The beneficial effects of this invention are: This invention introduces a logical diffusion guidance mechanism into the initial hierarchical alignment model and utilizes preset classification rules to generate logical gradients representing the compliance trend of features. This achieves deep reconstruction of the feature distribution structure in the high-dimensional solution space, effectively improving the coupling dimension between feature representation and business logic, and enhancing the logical consistency and accurate positioning capability of sorting decision instructions under complex constraints. Through hierarchical heterogeneous modeling and cross-level feature alignment, the invention achieves mapping and attribute enhancement of object nodes, location nodes, and task nodes in the hidden feature space, significantly improving the correlation accuracy of heterogeneous data at different semantic levels. It also demonstrates better scheduling reliability and path cost balance in high-concurrency, multi-station intelligent washing and sorting scenarios. In terms of model adaptive adjustment and robustness, this invention effectively solves the problems of incomplete feature representation and sorting conflicts caused by hard-coded logical rules in existing technologies through an inverse denoising iterative processing mechanism and a global parameter fine-tuning method based on state feedback. It breaks through the technical bottleneck of traditional sorting systems that are difficult to correct model weights in real time when dealing with large-scale heterogeneous attributes, achieving a significant advancement from static mapping to dynamic logical evolution generation, and effectively improving the intelligent processing level in the field of industrial automation sorting. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of a clothing intelligent sorting system based on object identifiers proposed in this invention; Figure 2 This is a flowchart of the logical diffusion reconstruction and reverse denoising iteration process of an intelligent clothing sorting system based on object identifiers proposed in this invention. Detailed Implementation

[0017] Combination Figures 1-2 The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Figure 1 and Figure 2 The present invention provides an intelligent clothing sorting system based on object identifiers, comprising the following modules: Original feature construction module: It obtains the unique identifier code, multi-dimensional feature vector and preset clothing flow logic of the clothing to be processed through the data acquisition interface, and retrieves the associated business attribute set according to the unique identifier code to obtain the original feature set containing the static attributes of clothing and flow constraints. Hierarchical heterogeneous modeling module: Performs logical node mapping and associated edge construction processing on the original feature set, converting the unique identifier, multi-dimensional feature vector and business attribute set into interrelated object nodes, location nodes and task nodes, to obtain an initial hierarchical heterogeneous graph model arranged in topology; Cross-level feature alignment module: Based on the initial hierarchical heterogeneous graph model, the module uses the hierarchical contrast loss function to perform association constraint processing on the node features of different levels, thereby obtaining an initial hierarchical alignment model with multi-dimensional attribute mapping relationships; Diffusion logic reconstruction module: In the initial hierarchical alignment model, a logic diffusion guidance mechanism is introduced. The logic gradient generated in the high-dimensional solution space by the preset classification rules is used to reconstruct the feature distribution structure in the initial hierarchical alignment model, so as to obtain the improved hierarchical heterogeneous graph model. Iterative denoising decision module: Based on the improved hierarchical heterogeneous graph model, it extracts the reconstructed feature vector of the object node after feature distribution reconstruction, and performs reverse denoising iterative processing in combination with the preset clothing flow logic to obtain sorting decision instructions; Feedback scheduling optimization module: Based on the sorting decision instructions, it executes the corresponding sorting scheduling and distribution processing, and collects the status feedback data after the transfer and returns it to the improved hierarchical heterogeneous graph model to obtain the adjusted feature weights.

[0018] In this embodiment, the original feature construction module specifically includes: Collect raw label data, multispectral image data, production scheduling data, and material attribute data of the clothing to be processed; Specifically, the data collection is performed using an ultra-high frequency RFID read / write array and a multispectral imaging unit deployed at the feed end of the sorting line.

[0019] The collected raw tag data, multispectral image data, production scheduling data, and material attribute data are preprocessed. The preprocessing includes unique identifier parsing, pixel feature extraction, priority mapping, and data standardization to obtain preprocessed multidimensional feature data. Based on the preprocessed multidimensional feature data, field aggregation processing is performed to construct an original feature set containing the static attributes and circulation constraints of clothing.

[0020] In this embodiment, the hierarchical heterogeneous modeling module specifically includes: Based on the unique identifier and multidimensional feature vector in the original feature set, entities with unique attribute identifiers are generated using preset object instantiation rules, thus obtaining object nodes; Specifically, the unique identifier code is extracted from the original feature set as the global index of the node. The fabric composition, color features, and washing frequency parsed from the multi-dimensional feature vector are then standardized and mapped into attribute feature vectors of a preset dimension. Using preset object instantiation rules, the attribute feature vectors are bound to the unique identifier code, generating a digital twin entity with a unique attribute identifier. This allows for the dynamic creation of object nodes within the entity layer of the heterogeneous graph structure. Extract the clothing flow logic from the original feature set, and perform spatial coordinate mapping based on the preset physical sorting points and logical workstation information in the clothing flow logic to obtain the location nodes; The business attribute set in the original feature set is analyzed, the washing requirements and delivery time limit parameters are extracted, and the washing requirements and delivery time limit parameters are converted into action constraints in the sorting process of the clothes to be processed, thus obtaining task nodes; Using preset hierarchical partitioning rules, object nodes, location nodes, and task nodes are deployed to the entity layer, spatial layer, and logical layer of the heterogeneous graph structure, respectively, to obtain a hierarchical node set; Based on the clothing flow logic, establish the affiliation association between object nodes and location nodes, establish the trigger association between task nodes and location nodes, and establish logical associations according to the execution order between object nodes and task nodes to obtain a set of associated edges; By using the set of associated edges to perform topological connection processing on the set of hierarchical nodes, an initial hierarchical heterogeneous graph model arranged according to the topological structure is constructed.

[0021] In this embodiment, the cross-level feature alignment module specifically includes: Based on the original feature vectors of object nodes, location nodes and task nodes in the initial hierarchical heterogeneous graph model, the original feature vectors of object nodes, location nodes and task nodes are transformed into a hidden feature space of a unified dimension using a preset linear mapping function, thus obtaining a cross-level basic feature set. Calculate the spatial distribution consistency between object nodes and associated location nodes in the cross-level basic feature set, and calculate the logical association strength between object nodes and corresponding task nodes to obtain the inter-level association weight coefficient. Specifically, the hidden feature vectors of object nodes and the spatial coordinate vectors of associated location nodes are extracted from the cross-level basic feature set. The cosine similarity function is used to measure the overlap between the predicted path distribution of object nodes and the topological attributes of physical sorting points, representing the consistency of spatial distribution. The washing process attributes of object nodes and the execution constraint parameters of corresponding task nodes are extracted to quantify the responsiveness of object nodes to expedited processing or special drying requirements, representing the strength of logical association. The spatial distribution consistency and logical association strength are weighted and fused to produce a numerical index of the coupling degree between mapping nodes, obtaining the inter-level association weight coefficient, which is used to adjust the coordinate distribution of object nodes, location nodes, and task nodes in the hidden feature space.

[0022] The hierarchical contrastive loss function is used to perform contrastive enhancement processing on the cross-level basic feature set. By maximizing the mutual information between positive sample pairs and minimizing the semantic association between negative sample pairs, the contrastive loss value is obtained. The initial hierarchical heterogeneous graph model is iteratively optimized using the contrastive loss value. The coordinate distribution of object nodes, position nodes, and task nodes in the hidden feature space is adjusted according to the inter-level correlation weight coefficients to obtain the hierarchical heterogeneous feature matrix after feature alignment. Semantic features of position nodes and task nodes that have topological connections with object nodes in the initial hierarchical heterogeneous graph model are retrieved. Attention weighting is used to aggregate the semantic features of position nodes and task nodes into the empty attribute bits of object nodes. Value pairs of logical attributes of object nodes are filled to obtain attribute-enhanced feature vectors. The attribute-enhanced feature vectors are mapped to the node index positions corresponding to the hierarchical heterogeneous feature matrix to construct an initial hierarchical alignment model with multi-dimensional attribute mapping relationships.

[0023] In this embodiment, the diffusion logic reconstruction module specifically includes: The business logic and scheduling priority constraints in the preset classification rules are analyzed, and a logical gradient representing the compliance trend of features is constructed in the feature space defined by the initial hierarchical alignment model. The node feature distribution in the initial hierarchical alignment model is regarded as the initial probability distribution through the logical diffusion guidance mechanism, and the logical gradient is set as the external guiding term in the feature evolution process. Specifically, the execution logic of the logical diffusion guidance mechanism is as follows: the feature vectors of each node in the initial hierarchical alignment model are mapped to an initial probability distribution in a high-dimensional space, and preset business logic rules and sorting constraint parameters are extracted to construct a logical gradient representing the compliance trend of the feature distribution; the logical gradient is used as an external guiding term to perform nonlinear correction on the diffusion path of node features during the reverse denoising process of feature evolution, forcibly inducing feature components that do not conform to logical constraints to gather in the compliant region in the high-dimensional solution space; through the nonlinear attraction generated by the logical gradient, the topological reconstruction of the initial probability distribution is realized, producing an improved feature vector with logical self-consistency, which is used to solidify the business association constraints between object nodes, location nodes and task nodes.

[0024] Based on the logical gradient, the abnormal feature components that violate the preset classification rules in the initial probability distribution are offset, and the compliant feature components that conform to the preset classification rules are clustered, resulting in the feature distribution space after the execution logic converges. Projection alignment is performed on the feature distribution space after the execution logic converges, and the evolved feature components are remapped to the topology node index positions of the initial hierarchical alignment model to obtain the reconstructed feature distribution structure. The reconstructed feature distribution structure is injected into the initial hierarchical alignment model, and the constraint force generated by the logical gradient is solidified by updating the model weight parameters to obtain the improved hierarchical heterogeneous graph model.

[0025] In this embodiment, the iterative denoising decision module specifically includes: Based on the improved hierarchical heterogeneous graph model, the reconstructed feature vectors of object nodes in the entity layer are extracted, and the reconstructed feature vectors are mapped to the initial state distribution of the discrete scheduling space to obtain the initial latent variable sequence. Based on the preset clothing flow logic, the spatiotemporal topological constraints and path cost functions of the physical sorting points are extracted, and the spatiotemporal topological constraints and path cost functions of the physical sorting points are converted into logical guiding terms in the reverse denoising process. Specifically, the logic for generating and transforming the logical guiding term is as follows: extracting the time constraints regarding washing delivery time limits and sorting priorities from the preset garment flow logic, and simultaneously obtaining the coordinate spacing of physical sorting points on the automated track and the response frequency of the robotic arm's movements to construct a spatiotemporal topological constraint characterizing the sorting execution efficiency; using a preset Euclidean distance and energy consumption model to calculate the movement cost of garments from the current workstation to the target location, generating a path cost function reflecting scheduling costs; and using a mapping unit to project the spatiotemporal topological constraints and path cost function onto a high-dimensional latent variable space, transforming them into nonlinear partial derivatives that induce the feature distribution to shift towards the global optimal solution, thus obtaining the logical guiding term, which is used to correct the evolution trajectory of the feature vector in real time during the inverse denoising iteration.

[0026] Based on the initial latent variable sequence, the weight parameters in the improved hierarchical heterogeneous graph model are called to perform single-step denoising. During the single-step denoising process, a logical guiding term is introduced to perform distribution correction, and intermediate state feature vectors are obtained. The intermediate state feature vector is used to perform multiple rounds of inverse denoising iteration. The intermediate state feature vector is continuously subjected to manifold constraint projection through the logical guiding term until the intermediate state feature vector meets the preset convergence threshold, thus obtaining the logical consistency representation vector. Perform feature decoding and symbolic mapping on the logical consistency representation vector to generate and reconstruct the discrete sorting execution sequence corresponding to the feature vector; Align the discrete sorting execution sequence with the real-time status of the physical sorting points to obtain sorting decision instructions.

[0027] In this embodiment, the feedback scheduling optimization module specifically includes: Based on sorting decision instructions, the sorting and processing of the clothing to be processed is completed; The system collects execution result data in real time during the sorting process, compares the execution result data with the preset clothing flow logic, and extracts status feedback parameters including path deviation and time consumption. Based on the state feedback parameters, the hierarchical heterogeneous graph model is updated to obtain the updated hierarchical heterogeneous graph model.

[0028] In this embodiment, the intelligent clothing sorting device based on object identifiers stores a program that, when executed, enables the intelligent clothing sorting device based on object identifiers to perform the intelligent clothing sorting system based on object identifiers as described in any one of claims 1 to 7.

[0029] Example 1: To verify the feasibility of this invention in practice, this example applies it to the auxiliary transportation and logistics automation center at the mine entrance of a large coal mine. In the actual application scenario, this coal mine is a high-yield and high-efficiency mine, requiring the daily processing of tool washing and sorting for over 8,000 underground workers. Due to the complex underground working environment, these tools are generally covered with high concentrations of coal dust, rock dust, and hydraulic oil stains. The fabrics include antistatic coated cloth, heavy-duty canvas, and special flame-retardant fibers. Each piece of tool has a high-temperature and high-pressure resistant UHF RFID electronic tag sewn inside the collar. The mine entrance sorting center is equipped with a 300-meter-long heavy-duty chain-type automated conveyor track with 60 sorting stations, each corresponding to different mining teams, shifts, and special degreasing process requirements.

[0030] In this application scenario, traditional technical solutions typically rely on manual sorting or simple static database comparison. Existing methods often fail to capture the deep logical relationship between tooling ownership, stain type, and real-time mine-head downhole plans in real time and accurately when faced with visual features obscured by coal dust and extremely complex shift scheduling instructions. This leads to a significant drop in sorting accuracy during peak shift change periods. Because traditional methods lack dynamic modeling of mine logistics topology constraints and washing path costs, technical problems such as sorting lever action conflicts, hook path deadlocks, and tools being misassigned to the wrong teams frequently occur, severely impacting the efficiency of mine logistics support. When handling temporarily added work clothes for emergency repairs at mining faces, existing technologies struggle to achieve smooth embedding of logical priorities while ensuring optimal overall paths, resulting in a surge in sorting line energy consumption and increased wear on robotic arms. To address these technical problems, this invention introduces hierarchical heterogeneous graph modeling and a logical diffusion guidance mechanism, achieving a deep coupling and reconstruction of the sorting logic and physical space of mine tools, significantly improving decision-making accuracy in extreme environments.

[0031] When the workpiece to be processed is conveyed into the sensing station of the sorting line at the mine entrance by the conveyor belt, the UHF RFID reader array deployed on the side of the track immediately senses the workpiece tag, reads and parses the unique EPC code as the object identifier. The data acquisition is carried out by the UHF RFID reader array and multispectral imaging unit deployed at the feeding end of the sorting line, which synchronously senses and acquires its unique identification code and multispectral image data containing coal dust distribution and fabric texture at the moment the workpiece is mounted; production scheduling data and material attribute data in the mine MES production management system are retrieved in real time through the industrial communication interface. The above multi-source heterogeneous data is subjected to denoising and dimension alignment processing using a preset field aggregation algorithm to construct an original feature set with spatiotemporal consistency.

[0032] After entering the hierarchical heterogeneous modeling stage, based on the unique identifier and multi-dimensional feature vector in the original feature set, entities with unique attribute identifiers are generated using preset object instantiation rules, resulting in object nodes. The object nodes are instantiated as follows: the unique identifier in the original feature set is extracted as the global index of the node, and the stain components, reflective strip features, and washing frequency parsed from the multi-dimensional feature vector are standardized and mapped to attribute feature vectors of preset dimensions. Using preset object instantiation rules, the attribute feature vectors are bound and encapsulated with the unique identifier to generate digital twin entities with unique attribute identifiers, thereby completing the dynamic creation of object nodes in the entity layer of the heterogeneous graph structure. Simultaneously, the physical layout parameters of the wellhead sorting center are retrieved to instantiate spatial layer location nodes representing sorting stations and unloading positions, and logical layer task nodes representing coal mining teams, tunneling teams, and auxiliary teams are generated according to the mine area scheduling instruction sequence.

[0033] When performing cross-level feature alignment, a cross-level basic feature set is constructed, and the association weights between nodes at each level are calculated. The generation logic of the inter-level association weight coefficients is as follows: extract the hidden feature vectors of object nodes and the spatial coordinate vectors of associated position nodes from the cross-level basic feature set; use a cosine similarity function to measure the overlap between the predicted path distribution of object nodes and the topological attributes of physical sorting points, representing spatial distribution consistency; extract the washing process attributes of object nodes and the execution constraint parameters of corresponding task nodes; use an attention scoring mechanism to quantify the response of object nodes to expedited drying or special oil stain cleaning requirements, representing logical association strength. The execution logic of the weighted fusion is as follows: extract a preset weight allocation ratio and perform linear weighted superposition on spatial distribution consistency and logical association strength, mapping multi-source features to a feature subspace of a unified scale; use a nonlinear activation function to perform nonlinear mapping on the superimposed feature results, inducing a nonlinear shift in feature distribution driven by mining area business logic, producing a numerical index that dynamically represents the degree of coupling between nodes, and obtaining the inter-level association weight coefficients. On this basis, feature alignment is achieved by maximizing the mutual information between positive sample pairs and minimizing the semantic association between negative sample pairs. The execution logic of maximization and minimization is as follows: extract the feature vectors of positive sample pairs with alignment relationship in the hierarchical heterogeneous graph, calculate their cosine similarity in the hidden feature space using the contrastive loss function, and induce the model weights to shift in the direction of increasing this similarity through the gradient descent algorithm; at the same time, extract the feature vectors of negative sample pairs without alignment relationship, use the inter-class exclusion mechanism to widen their spatial distance, and produce node representation features with high recognizability.

[0034] In the process of diffusion logic reconstruction and iterative denoising decision-making, the node feature distribution in the initial hierarchical alignment model is regarded as the initial probability distribution through a logic diffusion guidance mechanism, and the logic gradient is set as the external guiding term in the feature evolution process. The execution logic of the logic diffusion guidance mechanism is to map the feature vectors of each node in the initial hierarchical alignment model to the initial probability distribution in a high-dimensional space, and extract the preset mining area scheduling logic rules and physical constraint parameters to construct a logic gradient representing the compliance trend of the feature distribution; using the logic gradient as an external guiding term, nonlinear correction is performed on the diffusion path of node features in the inverse denoising process of feature evolution, forcibly inducing feature components that do not conform to the logic constraints to gather towards the compliant region in the high-dimensional solution space; through the nonlinear attraction generated by the logic gradient, the topological reconstruction of the initial probability distribution is realized.

[0035] When generating decision instructions, based on the preset tooling flow logic, the spatiotemporal topological constraints and path cost functions of the physical sorting points are extracted. The generation and transformation logic of the logical guidance term involves extracting the time constraints regarding wellhead delivery time limits and sorting priorities of each team from the preset logic, and simultaneously acquiring the coordinate spacing of the sorting points on the automated track and the action response frequency of the actuator to construct spatiotemporal topological constraints characterizing sorting execution efficiency; using a preset Euclidean distance and energy consumption model, the movement cost of the tooling from the current attachment point to the target workstation is calculated to generate a path cost function reflecting scheduling costs; using a mapping unit, the spatiotemporal topological constraints and path cost function are projected onto a high-dimensional latent variable space, transforming them into nonlinear partial derivatives that induce the feature distribution to shift towards the global optimal solution, thus obtaining the logical guidance term. Through multiple rounds of reverse denoising iteration, the final sorting decision instruction is generated and drives the actuator.

[0036] To quantify the technical advantages of this invention, this embodiment sets rigorous experimental conditions. The model training uses the Adam optimizer with a learning rate of 0.0003 and a batch size of 64. The total number of steps in the diffusion process is set to 1000, with 50 sampling steps. During a continuous 48-hour high-intensity shift test, approximately 300,000 iterations of optimization were performed, and the global average error was controlled at an extremely low level.

[0037] The table below presents the measured performance data of the present invention under different samples: Table 1 Sample number Sorting destination prediction accuracy (%) Logical constraint violation rate (%) Measured values ​​of path cost function Response latency (ms) Composite Reconstruction Error (MSE) S_Coal_01 (Comprehensive Mining Team 1 / Urgent) 99.88 0.01 11.23 40.2 0.0011 S_Coal_02 (Tunneling Team 2 / Standard) 99.94 0.01 14.56 37.8 0.0007 S_Coal_03 (Airlift Team / Bulk) 99.82 0.03 17.89 44.5 0.0019 S_Coal_04 (Emergency Repair Team / Fast Track) 99.85 0.02 12.05 39.1 0.0014 S_Coal_05 (Civil Defense Team / Special) 99.91 0.01 13.44 41.6 0.0009 Analysis of the data in Table 1 shows that the sorting accuracy is extremely high in high-dust, multi-logic-constraint environments like coal mines, with an average accuracy of 99.88%, significantly better than existing technologies in similar environments. Especially in the case of the emergency repair team (S_Coal_04), where timeliness is extremely critical, the response latency is only 39.1ms, and the logic constraint violation rate is extremely low, demonstrating the robustness of the logic diffusion guidance mechanism in handling dynamic and sudden tasks. The measured path cost remains low, indicating that the logic guidance term effectively achieves an optimal balance between energy consumption and path optimization. The overall reconstruction error (MSE) is consistently in the extremely low range, reflecting the precise control of high-dimensional feature reconstruction during the reverse denoising iteration process.

[0038] This invention solves the technical bottleneck of chaotic sorting logic and low efficiency in the auxiliary transportation of coal mine shafts by combining hierarchical heterogeneous graphs with a logic diffusion mechanism.

Claims

1. A smart laundry sorting system based on object identifier, characterized in that, Specifically, it includes the following modules: Original feature construction module: It obtains the unique identifier code, multi-dimensional feature vector and preset clothing flow logic of the clothing to be processed through the data acquisition interface, and retrieves the associated business attribute set according to the unique identifier code to obtain the original feature set containing the static attributes of clothing and flow constraints. Hierarchical heterogeneous modeling module: Performs logical node mapping and associated edge construction processing on the original feature set, converting the unique identifier, multi-dimensional feature vector and business attribute set into interrelated object nodes, location nodes and task nodes, to obtain an initial hierarchical heterogeneous graph model arranged in topology; Cross-level feature alignment module: Based on the initial hierarchical heterogeneous graph model, the module uses the hierarchical contrast loss function to perform association constraint processing on the node features of different levels, thereby obtaining an initial hierarchical alignment model with multi-dimensional attribute mapping relationships; Diffusion logic reconstruction module: In the initial hierarchical alignment model, a logic diffusion guidance mechanism is introduced. The logic gradient generated in the high-dimensional solution space by the preset classification rules is used to reconstruct the feature distribution structure in the initial hierarchical alignment model, so as to obtain the improved hierarchical heterogeneous graph model. Iterative denoising decision module: Based on the improved hierarchical heterogeneous graph model, it extracts the reconstructed feature vector of the object node after feature distribution reconstruction, and performs reverse denoising iterative processing in combination with the preset clothing flow logic to obtain sorting decision instructions; Feedback scheduling optimization module: Based on the sorting decision instructions, it executes the corresponding sorting scheduling and distribution processing, and collects the status feedback data after the transfer and returns it to the improved hierarchical heterogeneous graph model to obtain the adjusted feature weights. 2.The smart laundry sorting system based on object identifier according to claim 1, wherein, The original feature construction module specifically includes: Collect raw label data, multispectral image data, production scheduling data, and material attribute data of the clothing to be processed; The collected raw tag data, multispectral image data, production scheduling data, and material attribute data are preprocessed. The preprocessing includes unique identifier parsing, pixel feature extraction, priority mapping, and data standardization to obtain preprocessed multidimensional feature data. Based on the preprocessed multidimensional feature data, field aggregation processing is performed to construct an original feature set containing the static attributes and circulation constraints of clothing. 3.The smart laundry sorting system based on object identifier according to claim 1, wherein, The hierarchical heterogeneous modeling module specifically includes: Based on the unique identifier and multidimensional feature vector in the original feature set, entities with unique attribute identifiers are generated using preset object instantiation rules, thus obtaining object nodes; Extract the clothing flow logic from the original feature set, and perform spatial coordinate mapping based on the preset physical sorting points and logical workstation information in the clothing flow logic to obtain the location nodes; The business attribute set in the original feature set is analyzed, the washing requirements and delivery time limit parameters are extracted, and the washing requirements and delivery time limit parameters are converted into action constraints in the sorting process of the clothes to be processed, thus obtaining task nodes; Using preset hierarchical partitioning rules, object nodes, location nodes, and task nodes are deployed to the entity layer, spatial layer, and logical layer of the heterogeneous graph structure, respectively, to obtain a hierarchical node set; Based on the clothing flow logic, establish the affiliation association between object nodes and location nodes, establish the trigger association between task nodes and location nodes, and establish logical associations according to the execution order between object nodes and task nodes to obtain a set of associated edges; By using the set of associated edges to perform topological connection processing on the set of hierarchical nodes, an initial hierarchical heterogeneous graph model arranged according to the topological structure is constructed. 4.The smart laundry sorting system based on object identifier according to claim 1, wherein, The cross-level feature alignment module specifically includes: Based on the original feature vectors of object nodes, location nodes and task nodes in the initial hierarchical heterogeneous graph model, the original feature vectors of object nodes, location nodes and task nodes are transformed into a hidden feature space of a unified dimension using a preset linear mapping function, thus obtaining a cross-level basic feature set. Calculate the spatial distribution consistency between object nodes and associated location nodes in the cross-level basic feature set, and calculate the logical association strength between object nodes and corresponding task nodes to obtain the inter-level association weight coefficient. The hierarchical contrastive loss function is used to perform contrastive enhancement processing on the cross-level basic feature set. By maximizing the mutual information between positive sample pairs and minimizing the semantic association between negative sample pairs, the contrastive loss value is obtained. The initial hierarchical heterogeneous graph model is iteratively optimized using the contrastive loss value. The coordinate distribution of object nodes, position nodes, and task nodes in the hidden feature space is adjusted according to the inter-level correlation weight coefficients to obtain the hierarchical heterogeneous feature matrix after feature alignment. Semantic features of position nodes and task nodes that have topological connections with object nodes in the initial hierarchical heterogeneous graph model are retrieved. Attention weighting is used to aggregate the semantic features of position nodes and task nodes into the empty attribute bits of object nodes. Value pairs of logical attributes of object nodes are filled to obtain attribute-enhanced feature vectors. The attribute-enhanced feature vectors are mapped to the node index positions corresponding to the hierarchical heterogeneous feature matrix to construct an initial hierarchical alignment model with multi-dimensional attribute mapping relationships. 5.The smart laundry sorting system based on object identifier according to claim 1, wherein, The diffusion logic reconstruction module specifically includes: The business logic and scheduling priority constraints in the preset classification rules are analyzed, and a logical gradient representing the compliance trend of features is constructed in the feature space defined by the initial hierarchical alignment model. The node feature distribution in the initial hierarchical alignment model is regarded as the initial probability distribution through the logical diffusion guidance mechanism, and the logical gradient is set as the external guiding term in the feature evolution process. Based on the logical gradient, the abnormal feature components that violate the preset classification rules in the initial probability distribution are offset, and the compliant feature components that conform to the preset classification rules are clustered, resulting in the feature distribution space after the execution logic converges. Projection alignment is performed on the feature distribution space after the execution logic converges, and the evolved feature components are remapped to the topology node index positions of the initial hierarchical alignment model to obtain the reconstructed feature distribution structure. The reconstructed feature distribution structure is injected into the initial hierarchical alignment model, and the constraint force generated by the logical gradient is solidified by updating the model weight parameters to obtain the improved hierarchical heterogeneous graph model. 6.The smart laundry sorting system based on object identifier according to claim 1, wherein, The iterative denoising decision module specifically includes: Based on the improved hierarchical heterogeneous graph model, the reconstructed feature vectors of object nodes in the entity layer are extracted, and the reconstructed feature vectors are mapped to the initial state distribution of the discrete scheduling space to obtain the initial latent variable sequence. Based on the preset clothing flow logic, the spatiotemporal topological constraints and path cost functions of the physical sorting points are extracted, and the spatiotemporal topological constraints and path cost functions of the physical sorting points are converted into logical guiding terms in the reverse denoising process. Based on the initial latent variable sequence, the weight parameters in the improved hierarchical heterogeneous graph model are called to perform single-step denoising. During the single-step denoising process, a logical guiding term is introduced to perform distribution correction, and intermediate state feature vectors are obtained. The intermediate state feature vector is used to perform multiple rounds of inverse denoising iteration. The intermediate state feature vector is continuously subjected to manifold constraint projection through the logical guiding term until the intermediate state feature vector meets the preset convergence threshold, thus obtaining the logical consistency representation vector. Perform feature decoding and symbolic mapping on the logical consistency representation vector to generate and reconstruct the discrete sorting execution sequence corresponding to the feature vector; Align the discrete sorting execution sequence with the real-time status of the physical sorting points to obtain sorting decision instructions. 7.The smart laundry sorting system based on object identifier according to claim 1, wherein, The feedback scheduling optimization module specifically includes: Based on sorting decision instructions, the sorting and processing of the clothing to be processed is completed; The system collects execution result data in real time during the sorting process, compares the execution result data with the preset clothing flow logic, and extracts status feedback parameters including path deviation and time consumption. Based on the state feedback parameters, the hierarchical heterogeneous graph model is updated to obtain the updated hierarchical heterogeneous graph model. 8.A laundry smart sorting apparatus based on an object identifier, characterized by, The object identifier-based intelligent clothing sorting device stores a program that, when executed, enables the object identifier-based intelligent clothing sorting device to perform the object identifier-based intelligent clothing sorting system according to any one of claims 1 to 7.

Images