A sorting path adaptive switching method and system based on fabric overlapping state

Abstract

The application discloses a sorting path adaptive switching method and system based on fabric overlapping state, and belongs to the field of intelligent sorting of waste textiles. The application obtains images and height data through multi-modal perception, and constructs a topological graph taking fabrics as nodes and overlapping relationships as directed edges. Scene-level overlapping features such as the number of nodes of the maximum connected subgraph and the proportion of the overlapping area of the whole graph are extracted through a graph attention network. After normalization, the comprehensive overlapping risk score is calculated through a weighted summation formula, and the first to fourth risk levels are divided. Accordingly, the direct sorting, speed reduction re-inspection, scattering and expansion or pause backflow paths are adaptively selected from the path strategy library, the corresponding execution mechanism is controlled to act, and the risk threshold and strategy are updated by the downstream state sensor feedback. The application solves the problem that the prior art cannot perceive the overlapping state, significantly reduces the material blocking rate and the mis-sorting rate, and improves the continuous operation capacity of the production line.

Description

Technical Field

[0001] This invention relates to the field of intelligent recycling technology for waste textiles, and in particular to an adaptive switching method and system for sorting paths based on the overlapping state of fabrics. Background Technology

[0002] With the increasing demand for circular economy and recycled resource utilization, large-scale and automated sorting of waste textiles has become one of the key technologies in the recycling process. Existing fabric sorting production lines typically use vision inspection modules to identify fabric type, color, material, or defect status, and control actuators such as spray valves, robotic arms, material feeding mechanisms, and diversion baffles to complete path diversion.

[0003] However, fabrics are typical flexible materials, lacking a fixed shape and stable posture, and are prone to edge overlap, partial obstruction, layering, curling, and entanglement. In this case, there is a significant difference between the test results of a single piece of fabric and its actual sortable state.

[0004] Existing technology discloses an automated control system for a fabric sorting production line (publication number CN122076712A, hereinafter referred to as D1). This system identifies fabric categories and predicts estimated times through a central processing unit, and coordinates multiple sorting actuators through a collaborative control module, enabling multi-component collaboration. However, the technical solution of D1 has at least the following drawbacks: First, the perception dimension is limited. The D1 only uses ordinary industrial cameras to collect two-dimensional image data for analysis, and cannot perceive the distribution information of fabrics in the height dimension. As a result, it is completely unable to judge the spatial relationships such as the overlapping and occlusion of fabrics, and is prone to misjudging multiple overlapping fabrics as a single target.

[0005] Second, it lacks a functional module for handling overlapping states. In the D1 system architecture, the software functions of the central processing unit and the collaborative control module are limited to single-objective classification and task cost-based execution mechanism scheduling. It does not construct an overlapping topology model to express the occlusion and overlap relationships between multiple fabrics, nor can it extract quantitative features that reflect the overall scene overlap complexity.

[0006] Third, the path decision-making basis is simple and cannot adaptively handle complex overlapping situations. D1 only adjusts the conveyor belt speed based on the material accumulation height in the downstream recycling bin, or reduces speed and alarms when the actuator cannot arrive on time. Its processing logic is limited to simple "speed adjustment" or "alarm," and it cannot dynamically select intervention paths such as slowing down for re-inspection, unpacking and unfolding, or pausing the return flow based on the degree of fabric overlap itself. This makes it prone to jamming and mis-sorting when facing clumps of fabric or severe entanglement, resulting in low processing efficiency and frequent reliance on manual intervention.

[0007] Fourth, the control loop is open-loop and lacks self-optimization capabilities. D1 lacks an automatic confirmation and feedback mechanism for the results of sorting operations, making it impossible to know whether sorting was successful or whether residue occurred. System parameters (such as thresholds and strategies) cannot be automatically updated based on actual operating results, causing the system to remain at its initial settings level after long-term operation, making it difficult to adapt to changes in production line conditions.

[0008] Therefore, there is an urgent need for a method and system that can structurally express the overlapping relationship between fabrics, quantitatively assess the risk of overlap, and adaptively switch sorting paths according to the risk level, in order to solve the problems of functional deficiencies and insufficient performance in the existing technology. Summary of the Invention

[0009] This invention addresses the problems of existing technologies, such as limited perception dimensions, inability to quantify and process fabric overlap states, lack of adaptability in path decision-making, and open-loop control links. It proposes a sorting path adaptive switching method and system based on fabric overlap states, which can comprehensively perceive the three-dimensional morphology of the fabric, construct a structured overlap relationship diagram, quantify the risk of scene overlap, and adaptively switch to the optimal processing path according to the risk level. This achieves the technical effects of significantly reducing material jamming and missorting rates, improving the continuous operation capability of the production line in complex overlap scenarios, and enhancing the system's adaptive optimization level.

[0010] To achieve the above objectives, the following technical solution is proposed: An adaptive switching method for sorting paths based on fabric overlap states includes: Acquire image data and height distribution data of the fabrics to be sorted on the conveyor belt; Based on the image data and height distribution data, instances of each fabric target are identified, and an overlapping topology graph is constructed. The overlapping topology graph represents each fabric target as a node, and directed edges are established based on the spatial overlap, height difference, and / or occlusion relationship between any two fabric targets to express the overlapping state between fabrics. The overlapping topology graph is input into a trained graph neural network to extract scene-level overlapping features that reflect the complexity of multi-fabric overlap. A comprehensive overlap risk score is calculated based on the scene-level overlap features, and the overlap risk level of the current scene is determined based on the comprehensive overlap risk score. Based on the overlap risk level, a corresponding sorting processing path is selected from a preset path strategy library. The sorting processing path includes at least a direct sorting path and at least one intervention path for handling different degrees of overlap. Based on the selected sorting process path, the corresponding sorting execution mechanism is controlled to perform sorting actions on the fabric to be sorted.

[0011] By employing the above technical solutions, and by acquiring multimodal information including height distribution data, the system can accurately perceive the overlapping and occlusion relationships between fabrics in three-dimensional space, overcoming the limitations of D1 which relies solely on two-dimensional images. By constructing an overlapping topology graph that represents individual fabrics and their spatial relationships using nodes and directed edges, this solution transforms the originally complex and disordered physical overlap state into a structured, machine-processable data model, laying the data foundation for subsequent quantitative analysis. Furthermore, by using a trained graph neural network to process the overlapping topology graph, it can automatically and effectively aggregate local adjacency relationships and global graph structure information, thereby extracting abstract features reflecting the overall complexity of the scene's overlap. These features comprehensively characterize the difficulty of the sorting task compared to a single indicator.

[0012] Preferably, the image data includes color images, and the height distribution data includes depth maps and / or line laser profile data; the acquisition of image data and height distribution data of the fabric to be sorted on the conveyor belt is achieved by at least one of an industrial color camera, a depth sensor, and a line laser profile sensor.

[0013] By using the above technical solutions, the specific sensor type for acquiring height distribution data is limited, thus ensuring that the hardware can reliably capture the three-dimensional morphological information of the fabric surface.

[0014] Preferably, in the overlapping topology graph: node features include at least one of fabric projection area, equivalent height, contour compactness, edge exposure ratio, curl index, target motion stability, and perceived confidence; the edge attributes of the directed edges include at least one of the following: the percentage of overlapping area between two fabrics, the average depth difference, the occlusion direction, the contact boundary length, and the relative motion direction.

[0015] By using the above technical solutions, a fine set of attributes is defined for the nodes and edges of the topological graph, so that the overlapping relationship is not just "present" or "absent", but is quantified in a multi-dimensional and high-precision way. For example, the curl index can be used to identify easily entangled targets, and the occlusion direction can be used to determine the relationship between upper and lower layers. This provides rich raw information for subsequent graph neural network inference.

[0016] Preferably, the graph neural network is a graph attention network, and the scene-level overlap features include at least one of the following: the maximum number of nodes in the connected subgraph, the proportion of overlapping area in the whole graph, the maximum number of stacked layers, the degree of dispersion of edge weights, the number of severely curled boundaries, and the proportion of abnormal height regions.

[0017] By employing the above technical solutions and utilizing graph attention networks, different attention weights can be adaptively assigned to different neighboring nodes, allowing the model to focus more on key occlusion relationships while ignoring weak contacts. Furthermore, by defining a set of scene-level feature metrics, fabric stacks of arbitrary size and complexity can be uniformly mapped to a set of fixed-length feature vectors, achieving standardization and comparable quantification of overlapping complexity across different scenes.

[0018] Preferably, the calculation of the comprehensive overlap risk score based on the scene-level overlap features is specifically as follows: The maximum number of connected subgraph nodes N_connected, the total overlapping area ratio P_overlap, the maximum number of overlapping layers L_maximum, and the edge weight dispersion σ_edge in the scene-level overlap features are normalized to obtain normalized feature values ​​N_connected_norm, P_overlap_norm, L_maximum_norm, and σ_edge_norm. The comprehensive overlap risk score is calculated using the weighted summation formula R_overlap = a1·N_connected_norm + a2·P_overlap_norm + a3·L_maximum_norm + a4·σ_edge_norm, where a1 to a4 are weighting coefficients. The step of determining the overlap risk level of the current scenario based on the comprehensive overlap risk score includes: when the R overlap is lower than the first threshold, it is determined as a first-level risk; when the R overlap is between the first threshold and the second threshold, it is determined as a second-level risk; when the R overlap is between the second threshold and the third threshold, it is determined as a third-level risk; and when the R overlap is higher than the third threshold, it is determined as a fourth-level risk.

[0019] The above technical solution, by providing a clear, configurable weighted summation formula that includes normalization processing and a hierarchical logic based on multiple thresholds, transforms complex overlapping scenario risks into an objective, calculable, and numerically consistent normalized index, overcoming the limitations of D1's reliance on a single threshold or human experience. This formula has low computational cost, is suitable for real-time industrial calculations, and its weighting coefficients can be flexibly adjusted to adapt to different production lines.

[0020] Preferably, the step of selecting a corresponding sorting processing path from a preset path strategy library specifically involves: For the first level of risk, select the direct sorting route; For the second level of risk, choose the reduced-speed re-inspection path; Corresponding to the third level of risk, a dispersed deployment path is selected; For the fourth level of risk, choose either to pause the return process or conduct a manual review. The action parameters in the path strategy library include at least one of the following: conveying speed, number of re-inspection frames, action intensity of the dispersing mechanism, angle of the diversion baffle, selection of buffer channel, and alarm strategy.

[0021] By employing the above technical solutions, and by establishing a complete mapping relationship from risk level to specific intervention paths, the system can adopt tiered response strategies, ranging from "mild intervention" to "forced isolation," based on the severity of overlap. For example, mild overlap can be addressed by slowing down re-inspection to maintain efficiency, while severe entanglement can be actively isolated to avoid material jams, significantly improving the success rate and automation level in complex scenarios.

[0022] Preferably, the present invention further includes: The sorting action execution results are obtained by deploying status sensors downstream of the sorting mechanism. The execution results include whether the fabric passes through smoothly and whether there is any residue or jamming. Based on the execution results, the risk threshold, weight coefficients, and / or the path strategy library used in calculating the comprehensive overlap risk score are updated.

[0023] Through the above technical solutions, by constructing a closed-loop feedback chain from the sorting execution end to the risk decision-making front end, the system possesses the ability to adaptively learn and continuously optimize. It can automatically discover the shortcomings of the initial model or strategy and iteratively adjust it based on actual operating data, so that the system's sorting performance continuously improves as the operating time increases, overcoming the performance solidification defect under D1 open-loop control.

[0024] An adaptive switching system for sorting paths based on fabric overlap states includes: The multimodal sensing module is configured to acquire image data and height distribution data of the fabric to be sorted on the conveyor belt; An overlapping topology modeling module, coupled to the multimodal perception module, is configured to identify instances of each fabric target based on the image data and height distribution data, and construct an overlapping topology graph. The overlapping topology graph represents each fabric target as a node, and establishes directed edges based on the spatial overlap, height difference, and / or occlusion relationship between any two fabric targets. The graph neural network inference module is coupled to the overlapping topology modeling module and is configured to receive the overlapping topology graph and extract scene-level overlapping features that reflect the complexity of multi-fabric overlap. The risk quantification and assessment module, coupled with the graph neural network inference module, is configured to calculate a comprehensive overlap risk score based on the scene-level overlap features and determine the overlap risk level. The path switching decision module, coupled with the risk quantification assessment module, is configured to select a corresponding sorting and processing path from a preset path strategy library based on the overlapping risk level. The sorting and processing path includes at least a direct sorting path and at least one intervention path for handling different degrees of overlap. The execution control module, coupled with the path switching decision module, is configured to control the corresponding sorting execution mechanism to perform sorting actions on the fabric to be sorted according to the selected sorting processing path.

[0025] Preferably, the sorting execution mechanism coupled to the execution control module includes at least one of the following: a spray valve array, a material feeding roller, a vibration dispersing mechanism, a diversion baffle, a buffer conveyor, and an alarm device.

[0026] Preferably, the system also includes a status confirmation module, coupled to the execution control module, the risk quantification assessment module, and the path switching decision module. The status confirmation module is configured as follows: The sorting action results are obtained through downstream sensors, including whether the fabric passes through smoothly and whether there is any residue or jamming. The execution results are fed back to the risk quantification assessment module and / or the path switching decision module to trigger an update of at least one of the risk threshold, weight coefficient, and path strategy library.

[0027] Therefore, the present invention has at least the following beneficial effects: First, by introducing a multimodal perception module composed of depth sensors and other components, and constructing an overlapping topology map and a graph neural network inference module, the system has achieved a leap from "two-dimensional image classification" to "three-dimensional spatial relationship deep understanding" in terms of perception dimension and data processing logic, completely solving the functional deficiency problem that existing technologies cannot perceive and process the overlapping state of fabrics.

[0028] Secondly, by comprehensively considering overlapping risk scoring, multi-level risk classification, and the mapping mechanism of the path strategy library, the system constructs a computable, configurable, and tiered adaptive decision-making system. Compared with the single speed adjustment or alarm logic of existing technologies, this significantly improves the ability to handle complex scenarios such as severe entanglement and clumping, greatly reduces material jamming and missorting rates, and enhances the continuous operation capability of the production line. Actual tests show that the number of machine stoppages due to material jams can be reduced from 5.8 times per 100 pieces to 0.9 times, and the target channel misentry rate can be reduced from 12% to 2%.

[0029] Third, by adding a status confirmation module and establishing its feedback coupling relationship with the risk assessment and path decision modules, the system is upgraded from open-loop control to closed-loop adaptive optimization. It can automatically adjust the model parameters according to the execution results, so that the system performance can be continuously improved over time, avoiding the drawback of performance solidification after long-term operation. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overlapping topology construction and risk quantification mechanism based on the overlapping state of fabrics in this invention.

[0031] Figure 2 This is a schematic diagram of the sorting path adaptive switching method based on fabric overlap state according to the present invention.

[0032] Figure 3 This is a schematic diagram of the hardware module connection and signal flow of an embodiment of the system of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. However, it should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0034] Figure 3 A schematic diagram of the hardware module connections and signal flow of an embodiment of the system of the present invention is shown. In this embodiment, the system hardware mainly includes a multimodal sensing module, an edge computing unit, an execution control module, a sorting execution mechanism group, a downstream status confirmation module, and a display operation screen.

[0035] The multimodal sensing module is installed approximately 1.5 meters directly above the conveyor belt detection area, covering a complete 1.2-meter-wide passageway. This module specifically includes a 5-megapixel industrial color camera, a structured light-based depth sensor, a line laser profile sensor, and an incremental rotary encoder coaxially mounted with the conveyor belt drive shaft. The color camera and depth sensor connect to the edge computing unit via a USB 3.0 interface to acquire 1920×1080 resolution color images and depth maps. The line laser profile sensor connects to the edge computing unit via Gigabit Ethernet (GigE Vision), outputting a profile point cloud of the conveyor belt's transverse cross-section at a frame rate of 500Hz for accurate local height calculation. The encoder transmits pulse signals to the edge computing unit via an RS-485 serial bus to provide real-time conveyor belt speed (typically 0.5 m / s to 1.0 m / s) and position compensation.

[0036] The edge computing unit is an industrial computer (IPC) deployed within a production line control cabinet. Its core configuration includes an Intel Core i7 processor, 32GB of memory, and a single NVIDIA RTX A5000 graphics processing unit (GPU) with a computing power of 21 TOPS. Internally, this IPC runs instances of four core software modules: an overlapping topology modeling module, a graph neural network inference module, a risk quantification and assessment module, and a path switching decision module. These modules execute sequentially in a pipelined manner, with data flowing to the next module and signals flowing in the following order: Figure 3 As shown.

[0037] The execution control module is a programmable logic controller (PLC) independent of the IPC, specifically a Siemens S7-1500 series model. The IPC and PLC are coupled via an industrial Ethernet bus based on the EtherCAT protocol. This connection ensures that path switching commands are delivered to the actuator with a microsecond delay. The PLC's digital output module is hardwired to the solenoid valves of the spray valve array and the relays of the alarm device for direct sorting and alarm functions. The analog output module is connected to the servo motor driver of the diversion baffle, controlling the baffle angle by outputting a 0-10V voltage signal. Another digital output controls the start and stop of the vibration dispersing mechanism and the feeding rollers via an intermediate relay. The sorting execution mechanism group in this embodiment specifically includes: a spray valve array controlled by eight high-speed solenoid valves, a vibration dispersing mechanism with an adjustable frequency motor, a three-way diversion baffle driven by a servo motor, a buffer conveyor, a set of feeding rollers, and an audible and visual alarm device.

[0038] The status confirmation module includes a pair of through-beam photoelectric sensors installed downstream of each sorting channel, approximately 0.5 meters from the diversion point. These sensors are connected to the PLC's digital input module to capture the light-blocking signal sequence as the fabric passes through. If no correctly timed light-blocking signal is detected within a specific time window after the instruction is executed, the system determines it as "residual" or "mistakenly entered the channel"; if the signal is continuously blocked for more than a preset time, it is determined as "material jam." This status information, collected by the PLC, is transmitted back to the IPC via the EtherCAT bus for closed-loop updates by its internal risk quantification assessment module and path switching decision module.

[0039] The following is combined Figure 1 and Figure 2 This section details the complete sorting method and process based on the aforementioned hardware.

[0040] After the system is powered on and initialized, the conveyor belt runs at a preset initial speed (e.g., 0.8 m / s). When the encoder signal indicates that fabric has entered the sensing area, the multimodal sensing module is triggered, simultaneously acquiring a color image and a depth map, and obtaining the current real-time speed V_belt of the conveyor belt.

[0041] Step 1: Overlapping Topology Map Construction. Color image and depth map data are fed into the overlapping topology modeling module in the IPC. This module first uses a pre-trained instance segmentation model based on the Mask R-CNN architecture to process the color image, obtaining accurate masks, class labels, and bounding boxes for each individual fabric target in the image. Simultaneously, using pre-calibrated camera intrinsic and extrinsic parameters, the depth map point cloud is registered with the color image pixels, ensuring that each pixel within each mask region has a height value Z.

[0042] Suppose the current detection frame detects 3 fabric targets (target A: jeans, target B: T-shirt, target C: towel). The module calculates node features for each target, and the specific calculation process is as follows: Projected area: Directly count the number of pixels within the mask and multiply by the physical area corresponding to a single pixel. The measured areas of A, B, and C are 0.42m², 0.35m², and 0.28m², respectively.

[0043] Equivalent height: The average of all effective height values ​​within the mask area. The measured equivalent heights of A, B, and C are 0.05m, 0.04m, and 0.03m, respectively.

[0044] Contour compactness: calculated as (4π × area) / (perimeter²). The compactnesses of A, B, and C are 0.78, 0.65, and 0.81, respectively.

[0045] Edge Exposure Ratio: Calculates the ratio of the length of the boundary within the target bounding box that is not adjacent to any other target mask to its total boundary length. A is 0.50, B is 0.35, and C is 0.70.

[0046] Curl index: Calculated as 1 minus the ratio of the mask area to the area of ​​its smallest bounding rectangle. A is 0.22, B is 0.45, and C is 0.12.

[0047] Target motion stability: calculated by the standard deviation of the change vector of the centroid position of the same target in two consecutive frames, which is initially set to 0.9.

[0048] Perceived confidence score: The classification confidence score output by the instance segmentation model. A is 0.98, B is 0.96, and C is 0.99.

[0049] Thus, we obtain the 7-dimensional feature vectors of the 3 nodes. For example, the node feature vector of target A is [0.42, 0.05, 0.78, 0.50, 0.22, 0.9, 0.98].

[0050] Next, the module iterates through all pairwise target combinations to construct directed edges. If the intersection-over-union (IoU) ratio of the masks of two targets on the 2D plane is greater than 0.05, it is determined that there is spatial contact, and the edge attributes are calculated. Assume that A and B overlap, and B and C overlap.

[0051] Regarding A and B: Overlap area ratio: The ratio of the intersection area of ​​the two masks to the area of ​​mask A is calculated to be 0.2, and the ratio to the area of ​​mask B is 0.24.

[0052] Mean depth difference: Comparing the depth values ​​of A and B in the overlapping area, it was found that the average depth of A (distance from the camera) is less than that of B, meaning that A is on the upper layer. The mean depth difference is 0.015m.

[0053] Occlusion direction: Since A is on top, a directed edge is created from A to B, indicating that A occludes B. The edge properties in the other direction can represent the occlusion relationship.

[0054] Contact boundary length: Calculate the pixel length of the overlapping area edge and convert it to physical length, which is 0.18m.

[0055] Relative motion direction: Calculate the angle of the velocity vector difference between the centers of mass A and B, which is 15 degrees.

[0056] Therefore, a row in the edge attribute matrix is ​​in the form of [0.20, 0.015, 1.0 (meaning A occludes B), 0.18, 15].

[0057] Finally, the overlapping topology modeling module packages and transmits the following data to the graph neural network inference module: a node feature matrix with dimensions [3, 7], an edge index tensor with dimensions [2, 2] (recording the connection relationships between A→B and B→C), and an edge attribute matrix with dimensions [2, 5].

[0058] Step Two: Graph Neural Network Inference and Risk Quantization. After receiving the graph data, the graph neural network inference module inputs it into a pre-trained, offline-trained Graph Attention Network (GAT) with a 3-layer structure, 256 hidden units per layer, and 4 attention heads. GAT dynamically calculates weights for each node's neighbors using its attention mechanism. For example, in this embodiment, since occlusion of B by A is the primary relationship, GAT assigns a significantly higher attention weight to the edge A→B than to other secondary connections. After message passing and node updates by GAT, all node features are finally globally averaged and pooled, then concatenated with the overall graph attributes, and directly mapped by its output layer to the following scene-level overlapping feature indicators: The maximum number of nodes in a connected subgraph, N_connected, is 3.

[0059] The overlap area ratio of the entire image, P_overlap, is 0.25.

[0060] The maximum number of stacked layers, Lmax, is 2.

[0061] The edge weight dispersion σedge = 0.32.

[0062] The number of severely curled boundaries is 0.

[0063] The proportion of areas with abnormal height is 0.06.

[0064] In this frame, the curl index of all fabrics did not exceed 0.8, therefore the number of severely curled boundaries was 0. The proportion of abnormal height areas was obtained by statistically analyzing the percentage of pixels whose height values ​​exceeded twice the standard deviation of the average height among all fabric pixels.

[0065] Next, the risk quantification and assessment module normalizes the above features by dividing N-connectivity, P-overlap, L-maximum, and σ-edge by their respective preset baseline constants (the baseline constant for N-connectivity is 10, for P-overlap is 0.8, for L-maximum is 5, and for σ-edge is 0.6). After normalization, the feature values ​​are: N-connectivity_norm = 0.30, P-overlap_norm = 0.31, L-maximum_norm = 0.40, and σ-edge_norm = 0.53.

[0066] The risk quantification assessment module calls the comprehensive overlapping risk scoring formula for calculation. The weighting coefficients at this time are the current optimal values ​​loaded from the parameter server: a1=0.30, a2=0.20, a3=0.30, a4=0.20.

[0067] Substituting into the normalized formula R_overlap = a1·N_connected_norm + a2·P_overlap_norm + a3·L_maximum_norm + a4·σ_edge_norm, we get: R_overlap = 0.30×0.30 + 0.20×0.31 + 0.30×0.40 + 0.20×0.53 = 0.090 +0.062 + 0.120 + 0.106 = 0.378.

[0068] The system's current three thresholds are: first threshold T1 = 0.25, second threshold T2 = 0.45, and third threshold T3 = 0.65. Since 0.25 < 0.378 < 0.45, the overlap risk level of the current scenario is determined to be Level 2 risk.

[0069] Step 3: Adaptive Path Switching and Execution Control. The risk quantification assessment module sends the "Level 2 Risk" level signal to the path switching decision module. This module queries its internal path strategy library to find the action strategy corresponding to Level 2 risk—"Speed ​​Re-inspection Path". The specific action parameter set of this strategy is as follows: reduce the conveyor belt speed from 0.8m / s to 0.4m / s; trigger the industrial camera to continuously acquire 3 supplementary inspection images after a 200ms delay; send the supplementary images to the overlap topology modeling module to re-segment instances and construct the topology; if the recalculated R overlap after re-inspection drops below the Level 1 risk threshold, the direct sorting path is restored.

[0070] The path switching decision module encapsulates these action parameters into a control command message, which is sent to the execution control module (PLC) via the EtherCAT bus. After parsing the command, the PLC controls the frequency converter through the analog output channel to reduce the frequency of the main drive motor of the conveyor belt, thereby achieving speed reduction. At the same time, the PLC sends a synchronization signal to the IPC, triggering the camera to complete the acquisition of 3 frames of re-inspection images according to the set timing sequence.

[0071] To demonstrate the system's comprehensive handling capabilities for different risk levels, three more examples are given. When the R overlap is lower than T1, indicating a Level 1 risk, the path switching decision module activates the direct sorting path. The conveyor belt maintains a speed of 0.8 m / s, and the PLC controls the corresponding solenoid valve of the spray valve array to open based solely on the fabric category identification result, directly blowing the fabric into the corresponding collection channel. When the risk is determined to be Level 3, the path switching decision module activates the dispersing and unfolding path. The specific action parameters are: reducing the conveyor belt speed to 0.3 m / s; controlling the vibration dispersing mechanism to operate at a frequency of 40 Hz and an amplitude of 5 mm for 1.5 seconds, while simultaneously starting the material-feeding roller to perform auxiliary mechanical turning of the fabric; after dispersing, guiding the fabric into the return channel, allowing it to return to the feeding area for a second sensing. When the risk is determined to be Level 4, such as when the R overlap reaches 0.78 due to clump accumulation, the module activates the pause return or manual review path, controls the diversion baffle to guide the fabric into the buffer conveyor, and triggers the audible and visual alarm device to prompt the operator for manual intervention. The signal flow throughout the entire execution process is clear, achieving a precise transformation from high-level decision-making to physical actions.

[0072] Step 4: Status Confirmation and Closed-Loop Feedback. Approximately 5 seconds after the deceleration re-inspection and subsequent sorting actions, the downstream through-beam photoelectric sensor of the status confirmation module detects a fabric signal. If the signal sequence matches expectations, the sorting is marked as "successful." If the fabric re-enters the sensing area after entering the return channel and is successfully sorted, the entire process is still recorded as a "successful after scattering." All execution results are stored in the FIFO buffer of the edge computing unit.

[0073] The system automatically triggers a parameter update every 1000 sorting actions' worth of buffer data. Assuming that the slow-down re-inspection path was called 150 times out of these 1000 actions, with 120 ultimately succeeding in sorting after re-inspection (80% success rate), and 5 triggering of Level 4 risk manual review due to severe entanglement, the optimization algorithm detects a manual review trigger rate of 0.5%, below the 5% adjustment threshold, and therefore does not adjust the third threshold. However, it finds that the 80% re-inspection success rate is lower than the target value of 90%, so it automatically increases the re-inspection frame count parameter from 3 frames to 4 frames and records this update. All updates have amplitude limits (no more than 5% per update) and can be manually reviewed and confirmed to ensure the stability of the production line operation.

[0074] Through the above specific implementation methods, the present invention realizes a complete technical closed loop from multimodal perception to closed-loop optimization, effectively solving the functional and performance defects caused by the inability of existing technologies to handle the overlapping state of fabrics.

[0075] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the technical concept and spirit of the present invention should be included within the scope of protection of the present invention.

Claims

1. A sorting path adaptive switching method based on fabric overlap state, characterized in that, include: Acquire image data and height distribution data of the fabrics to be sorted on the conveyor belt; Based on the image data and height distribution data, instances of each fabric target are identified, and an overlapping topology graph is constructed. The overlapping topology graph represents each fabric target as a node, and directed edges are established based on the spatial overlap, height difference, and / or occlusion relationship between any two fabric targets to express the overlapping state between fabrics. The overlapping topology graph is input into a trained graph neural network to extract scene-level overlapping features that reflect the complexity of multi-fabric overlap. A comprehensive overlap risk score is calculated based on the scene-level overlap features, and the overlap risk level of the current scene is determined based on the comprehensive overlap risk score. Based on the overlap risk level, a corresponding sorting processing path is selected from a preset path strategy library. The sorting processing path includes at least a direct sorting path and at least one intervention path for handling different degrees of overlap. Based on the selected sorting process path, the corresponding sorting execution mechanism is controlled to perform sorting actions on the fabric to be sorted.

2. The sorting path adaptive switching method based on fabric overlap state according to claim 1, characterized in that, The image data includes color images, and the height distribution data includes depth maps and / or line laser profile data; the acquisition of image data and height distribution data of the fabric to be sorted on the conveyor belt is achieved by at least one of an industrial color camera, a depth sensor, and a line laser profile sensor.

3. The sorting path adaptive switching method based on fabric overlap state according to claim 1, characterized in that, In the overlapping topology graph: The node features include at least one of the following: fabric projected area, equivalent height, contour compactness, edge exposure ratio, curl index, target motion stability, and perceived confidence. The edge properties of a directed edge include at least one of the following: the percentage of overlapping area between two fabrics, the average depth difference, the occlusion direction, the contact boundary length, and the relative motion direction.

4. The sorting path adaptive switching method based on fabric overlap state according to claim 1, characterized in that, The graph neural network is a graph attention network, and the scene-level overlap features include at least one of the following: the maximum number of nodes in the connected subgraph, the proportion of overlapping area in the whole graph, the maximum number of stacked layers, the degree of dispersion of edge weights, the number of severely curled boundaries, and the proportion of abnormal height regions.

5. The sorting path adaptive switching method based on fabric overlap state according to claim 1, characterized in that, The calculation of the comprehensive overlap risk score based on the scene-level overlap features is specifically as follows: The maximum number of connected subgraph nodes N_connected, the total overlapping area ratio P_overlap, the maximum number of overlapping layers L_maximum, and the edge weight dispersion σ_edge in the scene-level overlap features are normalized to obtain normalized feature values ​​N_connected_norm, P_overlap_norm, L_maximum_norm, and σ_edge_norm. The comprehensive overlap risk score is calculated using the weighted summation formula R_overlap = a1·N_connected_norm + a2·P_overlap_norm + a3·L_maximum_norm + a4·σ_edge_norm, where a1 to a4 are weighting coefficients. The step of determining the overlap risk level of the current scenario based on the comprehensive overlap risk score includes: when the R overlap is lower than the first threshold, it is determined as a first-level risk; when the R overlap is between the first threshold and the second threshold, it is determined as a second-level risk; when the R overlap is between the second threshold and the third threshold, it is determined as a third-level risk; and when the R overlap is higher than the third threshold, it is determined as a fourth-level risk.

6. The sorting path adaptive switching method based on fabric overlap state according to claim 5, characterized in that, The step of selecting the corresponding sorting processing path from the preset path strategy library specifically involves: For the first level of risk, select the direct sorting route; For the second level of risk, choose the reduced-speed re-inspection path; Corresponding to the third level of risk, a dispersed deployment path is selected; For the fourth level of risk, choose either to pause the return process or conduct a manual review. The action parameters in the path strategy library include at least one of the following: conveying speed, number of re-inspection frames, action intensity of the dispersing mechanism, angle of the diversion baffle, selection of buffer channel, and alarm strategy.

7. A sorting path adaptive switching method based on fabric overlap state according to any one of claims 1 to 6, characterized in that, Also includes: The sorting action execution results are obtained by deploying status sensors downstream of the sorting mechanism. The execution results include whether the fabric passes through smoothly and whether there is any residue or jamming. Based on the execution results, the risk threshold, weight coefficients, and / or the path strategy library used in calculating the comprehensive overlap risk score are updated.

8. A sorting path adaptive switching system based on fabric overlap state, applicable to the sorting path adaptive switching method based on fabric overlap state as described in any one of claims 1-6, characterized in that, include: The multimodal sensing module is configured to acquire image data and height distribution data of the fabric to be sorted on the conveyor belt; An overlapping topology modeling module, coupled to the multimodal perception module, is configured to identify instances of each fabric target based on the image data and height distribution data, and construct an overlapping topology graph. The overlapping topology graph represents each fabric target as a node, and establishes directed edges based on the spatial overlap, height difference, and / or occlusion relationship between any two fabric targets. The graph neural network inference module is coupled to the overlapping topology modeling module and is configured to receive the overlapping topology graph and extract scene-level overlapping features that reflect the complexity of multi-fabric overlap. The risk quantification and assessment module, coupled with the graph neural network inference module, is configured to calculate a comprehensive overlap risk score based on the scene-level overlap features and determine the overlap risk level. The path switching decision module, coupled with the risk quantification assessment module, is configured to select a corresponding sorting and processing path from a preset path strategy library based on the overlapping risk level. The sorting and processing path includes at least a direct sorting path and at least one intervention path for handling different degrees of overlap. The execution control module, coupled with the path switching decision module, is configured to control the corresponding sorting execution mechanism to perform sorting actions on the fabric to be sorted according to the selected sorting processing path.

9. The sorting path adaptive switching system based on fabric overlap state according to claim 8, characterized in that, The sorting execution mechanism coupled to the execution control module includes at least one of the following: a spray valve array, a material feeding roller, a vibration dispersing mechanism, a diversion baffle, a buffer conveyor, and an alarm device.

10. The sorting path adaptive switching system based on fabric overlap state according to claim 8, characterized in that, It also includes a status confirmation module, coupled to the execution control module, the risk quantification assessment module, and the path switching decision module. The status confirmation module is configured as follows: The sorting action results are obtained through downstream sensors, including whether the fabric passes through smoothly and whether there is any residue or jamming. The execution results are fed back to the risk quantification assessment module and / or the path switching decision module to trigger an update of at least one of the risk threshold, weight coefficient, and path strategy library.

Images