A flexible fabric sortation execution window generation and dynamic control system and method

Abstract

The application discloses a flexible fabric sorting execution window generation and dynamic control system and method, and the system comprises a multi-modal perception module, a fabric state vector generation module, an execution mechanism constraint modeling module, an execution window generation module and a dynamic control module. The multi-modal perception module collects image data containing depth information; the fabric state vector generation module extracts morphological parameters such as flattening index and curling degree to construct a state vector; the execution window generation module generates a multi-dimensional execution window containing a spatially executable area and an executable time interval in combination with mechanism constraints; and the dynamic control module drives the execution mechanism according to the window. The application unifies real-time fabric morphology and mechanism capability modeling, generates a spatiotemporally coordinated execution window, and introduces an online feedback optimization mechanism, thereby improving the success rate and operation continuity of flexible fabric sorting, and having multi-type execution mechanism adaptation and self-evolution capabilities.

Description

Technical Field

[0001] This invention relates to the field of intelligent sorting technology for waste fabrics, and in particular to a flexible fabric sorting execution window generation and dynamic control system and method. Background Technology

[0002] With the development of the recycling industry, the demand for automated sorting of waste textiles is constantly increasing. Existing waste textile sorting systems typically use visible light image recognition, near-infrared spectral recognition, or instance segmentation to identify the material, color, and type of the textiles, and then trigger mechanisms such as spray valves, robotic arms, or push rods to complete the sorting.

[0003] Prior art document D1 (CN114082669B) discloses a robot sorting system and method with a flexible feeding mechanism. This system utilizes the flexible feeding mechanism to rearrange stacked parts, converting their 3D posture to a finite number of 2D / 2.5D postures. A vision system acquires 2D images and extracts visual information such as the position and posture of the parts, determining whether each part meets the picking requirements of "not stacked, separated, and a specific face facing a specified direction." If the requirements are met, the type of end effector is determined, and the robot is controlled to perform grasping and sorting. For materials that do not meet the picking requirements, the flexible feeding mechanism continues to vibrate or shake to rearrange them before re-sorting. However, the technical solution disclosed by D1 has significant shortcomings in circuit performance, processing efficiency, and functional coverage: its perception system relies solely on ordinary 2D cameras (or optional 3D cameras used only for pose acquisition), and cannot simultaneously obtain the depth contour and material information of materials. This results in its image processing module only being able to extract simple geometric features of rigid parts, and failing to characterize key morphological parameters such as the flattening degree, curling state, folding height, and edge exposure degree of flexible fabrics. The circuit processing capability cannot support multi-dimensional state estimation for flexible objects. Its execution control logic only drives the robotic arm to grasp with fixed-rule "compliant / non-compliant" binary judgments, lacking a dynamic execution window generation mechanism based on the real-time state of the material and the capability constraints of the actuator. This results in low processing efficiency, and when faced with the complex deformation, drift, and overlap of flexible fabrics, it is prone to grasping failures, misclassification, and omissions. In addition, D1's grating detection only provides open-loop prompts for placement errors, and there is no closed-loop feedback circuit and algorithm that automatically corrects control parameters or optimizes the recognition model based on the execution results. The system cannot learn from failures, and its long-term operating performance cannot be improved. Furthermore, it cannot be compatible with the unified control of multiple types of actuators such as spray valve arrays and push rod baffles.

[0004] Therefore, there is an urgent need to propose a sorting control system that can analyze the fabric morphology and mechanism response delay online, dynamically generate an adaptive execution window, and achieve closed-loop optimization, in order to solve the problems of low execution success rate, frequent equipment malfunctions, and inability of the system to adaptively evolve due to material deformation and motion uncertainty in the automated sorting of flexible fabrics. Summary of the Invention

[0005] This invention solves the problems of high failure rate and poor adaptability, and proposes a flexible fabric sorting execution window generation and dynamic control system and method, which achieves the technical effects of significantly improving sorting success rate, reducing missorting and omission, enhancing system operation continuity and having online self-learning capability.

[0006] To achieve the above objectives, the following technical solution is proposed: A flexible fabric sorting execution window generation and dynamic control system includes: A multimodal sensing module is configured to acquire image data of fabric on a conveyor belt, wherein the image data includes at least two-dimensional texture information and depth information; The fabric state vector generation module is communicatively coupled with the multimodal perception module and is configured to extract the morphological state parameters of the fabric based on the image data and construct a fabric state vector representing the real-time state of the fabric. The actuator constraint modeling module stores constraint parameters for at least one actuator. The execution window generation module is communicatively coupled to the fabric state vector generation module and the actuator constraint modeling module, and is configured to generate an execution window based on the fabric state vector and the constraint parameters. The execution window includes at least a spatially executable region and an executable time interval on the fabric. The dynamic control module is communicatively coupled to the execution window generation module and is configured to drive the corresponding execution mechanism to perform sorting operations on the spatially executable area within the executable time interval according to the control instructions of the execution window generation mechanism.

[0007] Through the above technical solutions, by setting up a multimodal perception module to synchronously acquire image data containing depth information, compared to the D1 solution which only uses a two-dimensional camera, this system can acquire the three-dimensional contour, height distribution, and local texture of the fabric, thus providing a high-dimensional data foundation for subsequent morphological analysis. The fabric state vector generation module uses this data to extract features specific to flexible materials such as flattening, curling, and folding, and constructs a unified state vector. This allows the system to characterize the fabric state beyond simple geometric selection, delving into the fine-grained physical properties that affect the success or failure of execution, effectively compensating for the D1's inability to determine whether the fabric is stable. The module addresses the shortcomings of "fixed execution" by jointly modeling the real-time fabric status with constraints such as the response delay and reachability of the actuator. This generates a multi-dimensional execution window that includes both spatial and temporal regions, expanding the controlled object from a traditional single point to a spatiotemporally coordinated window region. This significantly improves the accuracy and robustness of the execution actions. The dynamic control module calculates specific mechanism action parameters based on this window, ensuring that actions such as blowing, gripping, or pushing are accurately applied to the most suitable stress area of ​​the fabric. This avoids execution failures caused by curled edges or overlapping areas, thus achieving a leap from "being able to recognize" to "being able to execute stably."

[0008] Preferably, the multimodal sensing module includes an RGB-D camera and a near-infrared spectral sensor, configured to perform time synchronization and spatial registration of the acquired RGB images, depth images and near-infrared spectral data, and generate fused data that integrates texture, depth and material information.

[0009] Through the above technical solutions, by introducing the combination of RGB-D camera and near-infrared spectral sensor and performing spatiotemporal alignment, the system not only obtains the appearance and three-dimensional shape of the fabric, but also simultaneously acquires material composition information. This enables accurate differentiation of fabrics with similar colors but different materials. At the same time, the depth information makes the estimation of fold height and number of layers more reliable. The fused data provides rich raw signals for the high-precision calculation of fabric state vectors, overcoming the perception bottleneck of the D1 single vision sensor in the sorting of complex fabrics.

[0010] Preferably, the fabric state vector generation module is configured to run a fabric state analysis model, and the fabric state vector includes at least one or more of the following parameters: fabric projected area, flattening index, fold or bulge height, edge exposure length, curling degree, number of overlapping layers, percentage of overlapping area, movement speed, posture angle and perceived confidence.

[0011] Through the above technical solutions, by defining a multi-dimensional set of state parameters for flexible fabrics, the system can comprehensively quantify the fabric's flatness, the size of the stress-bearing parts, motion stability, and visual uncertainty. These parameters directly determine the quality of the execution window generation. For example, when the flattening index is too low or the curling degree is too high, the system can predict the execution risk in advance and select an alternative strategy, unlike D1 which only mechanically judges whether the target box overlaps. This significantly reduces problems such as false spraying and grasping drop caused by forced execution.

[0012] Preferably, the execution window generation module is configured to generate a multi-dimensional execution window that includes a spatially executable region R, an executable time interval T, an execution mode M, control parameters P, and a risk level U. The module then performs a comprehensive evaluation based on the flattening score, edge exposure score, motion stability score, institutional accessibility score, overlap risk, curling risk, and uncertainty of the candidate execution windows to select the optimal execution window.

[0013] By expanding the execution window into a five-dimensional parameter space and applying multi-factor weighted scoring to each candidate window, the system can quantitatively balance execution effectiveness and failure risk, and automatically select the execution scheme with the highest overall benefit. For example, when the fabric edge is well exposed and the movement is stable, a high-efficiency blowing mode can be selected. If there is partial obstruction, it will automatically switch to low-speed grasping or secondary flattening before processing. This intelligent decision-making capability is not available in D1's fixed threshold logic, which greatly improves the level of adaptability in complex situations.

[0014] Preferably, the system also includes an anomaly handling module, which is communicatively coupled to the execution window generation module. The module is configured to trigger alternative execution strategies, including delayed execution, secondary flattening, reflow processing, or temporary storage for re-inspection, when the risk level U exceeds a preset threshold or the fabric state vector indicates that the fabric is in a state that does not meet the stable execution conditions.

[0015] Through the above technical solutions, by adding an abnormal handling module and setting a risk gating mechanism, when the fabric is severely curled, overlapped, or the perceived quality is too low, the system can actively suppress high-risk execution actions and instead guide the fabric into the flattening device or return channel. This avoids material jamming, missorting, and equipment damage caused by indiscriminate shaking or direct grabbing in D1, and significantly improves the continuous operation time and safety of the production line.

[0016] Preferably, the system also includes an online feedback optimization module, which is communicatively coupled to the dynamic control module and the execution window generation module. This module is configured to collect actual result data after sorting execution and dynamically correct the spatial offset, timing compensation, and executability evaluation threshold of the execution window based on the deviation between the actual result data and the target result.

[0017] Through the above technical solutions, by establishing a closed-loop feedback path from execution results to control parameters, the system can automatically learn and compensate for long-term drift factors, such as small changes in conveyor belt speed or attenuation of spray valve pressure, and maintain high execution accuracy without manual intervention. In contrast, D1's grating detection can only detect and prompt errors, lacking a closed-loop circuit and algorithm to convert the results into control quantity correction, and cannot avoid the recurrence of similar errors.

[0018] Preferably, the online feedback optimization module is further configured to feed back execution failure samples, window prediction deviation samples, and manual review results to the training sample pool to optimize the fabric state analysis model in the fabric state vector generation module and the executability prediction model in the execution window generation module.

[0019] Through the above technical solutions, by feeding back failed cases as high-value training data into the model training process, the deep learning model in the system can be continuously iterated and strengthened as the production line runs, and its generalization ability for new fabrics or rare patterns is gradually enhanced, achieving true online evolution. In contrast, D1 does not contain any model training and update mechanism, and its performance cannot be improved autonomously under long-term operation.

[0020] Preferably, the actuator is one or more of the following: a spray valve array, a robotic arm, a push rod, a lever, a diverter baffle, or a conveying branch mechanism; the dynamic control module is configured to calculate the corresponding spray valve combination, opening time, and pulse width, or calculate the gripping point and path of the robotic arm, or calculate the action time and stroke of the push rod and the baffle, based on the spatial executable area and executable time interval in the execution window.

[0021] Through the above technical solutions, by using a unified execution window as an abstract interface, the dynamic control module can adapt the same set of decision logic to actuators with different physical principles. For example, for a spray valve array, the spatial area can be mapped to a combination of nozzles and a spray pulse width, while for a robotic arm, it is transformed into a grasping posture and trajectory. This multi-mechanism compatibility enables the system to flexibly integrate multiple sorting methods in a production line, which is far superior to the single control architecture of D1, which is only for robot grasping.

[0022] A method for generating and dynamically controlling the execution window for flexible fabric sorting includes: Collect multimodal image data of the fabric on the conveyor belt, wherein the multimodal image data includes at least depth information; Based on the multimodal image data, the morphological state parameters of the fabric are extracted, and a fabric state vector representing the real-time state of the fabric is constructed. Obtain constraint parameters for at least one actuator; An execution window is generated based on the fabric state vector and the constraint parameters. The execution window includes at least a spatially executable region and an executable time interval on the fabric. According to the control instructions of the execution window generation mechanism, the corresponding execution mechanism is driven to perform sorting operations on the spatially executable area within the executable time interval.

[0023] By integrating multimodal depth perception, state vector construction, and execution window generation into a coherent process, the output of each step becomes the input of the next step. The entire control chain forms a complete information loop from "seeing" to "execution." Compared to the fragmented control logic of D1, which relies solely on two-dimensional image judgment for direct capture, this method has higher orderliness and traceability when processing flexible fabrics.

[0024] As a preferred option, it also includes: Predict the future trajectory of the fabric after the actuator response delay; The generation of the execution window includes generating a multi-dimensional execution window based on the fabric state vector, the future trajectory, and the constraint parameters, which includes a spatially executable region, an executable time interval, an execution mode, control parameters, and a risk level. The method further includes: collecting execution results and dynamically adjusting the timing compensation amount, executability evaluation threshold, and execution window scoring weight based on the deviation between the execution results and the target results.

[0025] By incorporating the future trajectory prediction step, the system can compensate for the time delay between perception and execution, ensuring that the generated execution window accurately matches the actual position and posture of the fabric when it arrives at the execution station. This effectively solves the motion deviation caused by response lag under high-speed conveying. Furthermore, the system integrates post-execution deviation feedback to correct timing and thresholds online, making the entire method a fully closed-loop adaptive control process of "perception-prediction-decision-execution-correction," which has a fundamental advantage over the open-loop control of D1.

[0026] Therefore, the present invention has at least the following beneficial effects: First, this invention introduces the real-time morphological state of flexible fabric and the capability constraints of the actuator into the execution control process, generating a multi-dimensional execution window that includes spatial regions, time intervals and control parameters. This overcomes the shortcomings of D1, which only relies on fixed target positions and binary judgments, and significantly improves the sorting success rate. Experiments have shown that the success rate can be increased from about 72% to over 93%.

[0027] Second, this invention deeply characterizes the fine-grained features of flexible fabrics, such as flattening, curling, and overlapping, through multimodal perception fusion and fabric state vector construction. This enables the system to have the ability to predict the feasibility of execution before execution, avoiding misclassification, omission, and material jamming caused by D1 forcibly executing when the material is in poor condition. The effective running time ratio is increased to 96%.

[0028] Third, this invention uses an online feedback optimization mechanism to dynamically correct control parameters and update the deep learning model based on the actual execution results, enabling the system to self-evolve during long-term operation and solving the problems of D1's lack of closed-loop learning ability and the recurrence of similar errors.

[0029] Fourth, this invention uses the execution window as a unified interface, which can be adapted to various execution mechanisms such as spray valve arrays, robotic arms, push rods, and baffles. The control architecture has strong versatility and scalability, making it easy to deploy in various flexible material sorting production lines. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall architecture of a flexible fabric sorting execution window generation and dynamic control system according to the present invention.

[0031] Figure 2 This is a schematic diagram of the execution window generation and dynamic control process in this invention. Detailed Implementation

[0032] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the described embodiments are for illustrative purposes only and are not intended to limit the scope of protection. All equivalent substitutions or simple modifications made based on the inventive concept should be included within the scope of protection of the present invention.

[0033] This embodiment provides a waste clothing air-blowing sorting system based on multimodal perception. Its core lies in the generation and dynamic control of the flexible fabric sorting execution window. (Reference) Figure 1 , Figure 1 This demonstrates the data flow and control relationships between the field perception layer, edge computing and window generation layer, execution control layer, and online feedback optimization. It mainly includes the field perception layer, edge computing and window generation layer, execution control layer, and online feedback optimization chain. The main body of the production line includes an adjustable speed conveyor belt, a flattening device, a multimodal sensing module, edge computing equipment, a spray valve array, flow dividers, a return channel, and a visual feedback unit in the material discharge area.

[0034] The multimodal sensing module, located above the conveyor belt, consists of an RGB-D camera and a near-infrared (NIR) spectral sensor. The RGB-D camera outputs a 1920×1080 resolution color image and a synchronized depth image at a frame rate of 30fps. The NIR sensor acquires the fabric's reflectance spectrum in the 900-1700nm band for material classification. After the data from both sensors enters the edge computing device, it is aligned using hardware timestamps and spatially registered using pre-calibrated intrinsic and extrinsic parameters to generate aligned RGB-D-NIR four-channel fused data for each frame. This fused data simultaneously includes texture, 3D geometry, and spectral material information, and is sent to the target localization and trajectory estimation module via a gigabit Ethernet interface.

[0035] The target localization and trajectory estimation module is deployed on an edge computing device, running a deep learning-based instance segmentation network (such as YOLACT++) to perform pixel-level segmentation on each piece of fabric in the fused data, obtaining instance masks, bounding boxes, and contours. Combining the displacement increments provided by the conveyor belt encoder, this module uses a Kalman filter to estimate the centroid position and velocity of each fabric, and predicts its position after a future time Δt, to compensate for the total delay from command issuance to the arrival of the blown airflow on the fabric (approximately 120ms, of which valve opening delay accounts for approximately 30ms, and the remainder is communication and airflow transmission time). The predicted trajectory is passed to the fabric state vector generation module via shared memory.

[0036] The fabric state vector generation module extracts a series of state features from the target mask, depth contour, and continuous frame motion information to construct the fabric state vector X_i(t). This module runs a lightweight fabric state analytical model based on a multi-task learning network. Its inputs are aligned RGB-D-NIR local regions and the mask, and its outputs include various morphological parameters: fabric projected area S, flattening index F (calculated using depth variance and gradient histogram), fold or bulge height H (local depth peak), edge exposure length L (arc length at the mask edge that does not overlap with other fabric and is not curled), curling degree C (based on contour curvature and depth edge abrupt change detection), number of overlapping layers N and overlapping area percentage R, motion velocity V, acceleration fluctuation A, attitude angle θ, and perception confidence Q. These parameters together constitute the state vector X_i(t). This module interacts with the actuator constraint modeling module and the execution window generation module via the PCIe bus.

[0037] The actuator constraint modeling module is a parameter storage and query unit, internally storing the constraint vectors G_m of various actuators in the form of a structure array. For the spray valve array in this embodiment, G includes the installation spacing of the spray valves, the effective spray diameter of a single valve, the valve opening delay τ (approximately 30ms), the maximum switching frequency (bandwidth B), the spray pressure-force mapping table (capability constraint K), and safety boundary conditions (such as prohibiting two consecutive valves from opening simultaneously to avoid a sudden drop in air pressure). These parameters can be configured offline through the human-machine interface and loaded into memory during system initialization for use by the executability assessment module and the execution window generation module.

[0038] The executability assessment module receives X_i(t) from the fabric state vector generation module and G_m from the actuator constraint modeling module. Internally, it runs an executability prediction model (using a lightweight gradient boosting tree or a small multilayer perceptron). This model outputs the success probability of blowing and a comprehensive risk value U (a continuous value ranging from 0 to 1, with higher values ​​indicating greater risk). U can be linearly mapped to a discrete risk level: U < 0.3 is low, 0.3 ≤ U < 0.6 is medium, and U ≥ 0.6 is high. When the success probability is higher than 0.85 and U < 0.3 (low risk), the fabric is marked as "direct execution"; if the flattening index F is lower than the threshold of 0.6 but the edge exposure length is acceptable, it is marked as "flattening treatment"; if the number of overlapping layers N ≥ 2 or the perception confidence Q is lower than 0.5, it is marked as "reflow treatment". This assessment result serves as a prerequisite for the execution window generation module.

[0039] After receiving the executability assessment result for "direct execution," the execution window generation module begins generating candidate execution windows. It first generates the spatially executable region R of the fabric: within the instance mask, after removing curled edges, folded protrusions, and overlapping boundary areas, a set of connected regions suitable for blowing is obtained. For each connected region, combined with the trajectory prediction results, the executable time interval T in the conveyor belt's movement direction is calculated, and the execution mode M (e.g., conventional blowing, high-pressure blowing) and control parameters P are set according to the fabric material type. P includes the valve number combination, blowing pressure, pulse width, and advance compensation. Finally, each candidate window is represented as W_i,m={R,T,M,P,U}, where U is the risk value corresponding to that candidate window. The module scores each candidate window using the comprehensive scoring formula F_w = a1·A_flat + a2·E_edge + a3·S_motion + a4·R_each + a5·V_task - a6·R_overlap - a7·R_curl - a8·U - a9·C_action. The weighting coefficients are configured based on the characteristics of the jetting mechanism as follows: a1=0.2, a2=0.25, a3=0.15, a4=0.1, a5=0.1, a6=0.08, a7=0.07, a8=0.03, a9=0.02. These weighting coefficients are empirically assigned based on the jetting mechanism characteristics, and all scoring items have been normalized to the [0,1] interval. Therefore, the scoring result is a dimensionless weighted index of merit. The window with the highest score and a risk level not exceeding "medium" (i.e., U<0.6) is selected as the optimal execution window and sent to the dynamic control module. If the scores of all candidate windows are lower than the preset safety threshold or U≥0.6, the window generation module will not output a valid window and will instead trigger the exception handling module.

[0040] The system's decision-making hierarchy is as follows: After the feasibility assessment module completes the initial screening, for "directly executable" targets, the execution window generation module performs refined window generation and scoring; if the U-value of the optimal window is still higher than the safety threshold, the exception handling module takes over and implements alternative strategies. The exception handling module also handles the initially screened non-directly executable targets.

[0041] After receiving the selected execution window, the dynamic control module maps the spatial region R to a specific valve number in the spray valve array. Based on the time interval T and the advance compensation amount, it calculates the opening time and converts the pulse width and pressure level in the control parameter P into digital signals. These signals are then sent to the drive controller of the spray valve array via Industrial Ethernet (EtherCAT) to precisely control the spatiotemporal distribution of the blowing airflow. Simultaneously, when the fabric is marked as "flattening" or "returning," the dynamic control module coordinates the movement of the flattening device's pressure rollers or the angle switching of the diversion baffle to guide the fabric into the flattening station or return channel.

[0042] The anomaly handling module continuously monitors the output status of the execution window generation module and the risk level U of the executability assessment module. When the risk level U exceeds a preset threshold, or the fabric state vector indicates that the fabric is in a state that does not meet the stable execution conditions (such as curl degree C greater than 0.8, number of overlapping layers N ≥ 3, edge exposure length L less than the minimum graspable length, etc.), the module immediately suppresses the direct execution command output of the dynamic control module and triggers alternative actions according to preset strategies: for example, for fabrics with local curl that can be flattened, a command is sent to reduce the conveyor belt speed, and the flattening device is started for secondary flattening; for fabrics with severe overlap or lumps, the diversion baffle is controlled to guide the material into the return channel and back to the feeding port; if sensor data is intermittently missing, a "re-inspection and temporary storage" command is issued to temporarily store the material in the buffer area for re-inspection. The execution status of the anomaly handling strategy is fed back to the online feedback optimization module in real time.

[0043] The online feedback optimization module acquires images of the actual landing point of the fabric after it is blown by the air jet through the visual feedback unit (industrial camera) in the material drop area, compares them with the target diversion port position, and calculates the landing point offset vector. The module statistically analyzes the mean and variance of the offset over a fixed time window. When the offset systematically increases, it automatically corrects the spatial bias (e.g., moving the target point in the opposite direction of the offset) and timing compensation (e.g., advancing or delaying by 1-3ms) in the dynamic control module. Simultaneously, based on long-term success rate changes, it adjusts the success probability threshold in the executability evaluation module. Furthermore, the module automatically labels and stores execution failure samples (e.g., landing point deviation exceeding 10cm, fabric not entering the target port), window prediction deviation samples, and manually reviewed classification error samples. When the accumulated samples reach a preset number, the background training process is initiated to incrementally fine-tune the fabric state analysis model (in the fabric state vector generation module) and the executability prediction model (in the executability evaluation module). The updated model is deployed online through the model version management system, achieving closed-loop self-evolution.

[0044] A method for generating and dynamically controlling a flexible fabric sorting execution window, applicable to the aforementioned flexible fabric sorting execution window generation and dynamic control system, comprising: Collect multimodal image data of the fabric on the conveyor belt, wherein the multimodal image data includes at least depth information; Based on the multimodal image data, the morphological state parameters of the fabric are extracted, and a fabric state vector representing the real-time state of the fabric is constructed. Obtain constraint parameters for at least one actuator; An execution window is generated based on the fabric state vector and the constraint parameters. The execution window includes at least a spatially executable region and an executable time interval on the fabric. According to the control instructions of the execution window generation mechanism, the corresponding execution mechanism is driven to perform sorting operations on the spatially executable area within the executable time interval.

[0045] It also includes: predicting the future trajectory of the fabric after the actuator response delay; The generation of the execution window includes generating a multi-dimensional execution window based on the fabric state vector, the future trajectory, and the constraint parameters, which includes a spatially executable region, an executable time interval, an execution mode, control parameters, and a risk level. The method further includes: collecting execution results and dynamically adjusting the timing compensation amount, executability evaluation threshold, and execution window scoring weight based on the deviation between the execution results and the target results.

[0046] refer to Figure 2 , Figure 2 This demonstrates the complete process of the system, from multimodal synchronous acquisition, target detection and tracking, state vector and trajectory prediction, executability assessment, candidate window generation, window scoring and risk gating, dynamic execution control to feedback optimization. After the system is powered on, the multimodal perception module continuously and synchronously acquires RGB-D-NIR data; the target localization and trajectory estimation module performs instance segmentation and Kalman filter tracking on each frame of image, outputting a list of fabric targets with IDs and predicted trajectories; the fabric state vector generation module calculates the state vector for each target; subsequently, the executability assessment module performs executability assessment in conjunction with actuator constraints. If the fabric does not meet the direct execution conditions, the anomaly handling module is triggered to execute flattening or recirculation strategies; if executable, the execution window generation module generates multiple candidate windows and scores them to select the best one; the optimal window is sent to the dynamic control module, which converts it into the specific control timing of the spray valve array; after the action is executed, the online feedback optimization module collects the landing point results and performs parameter correction and model optimization, completing one closed-loop cycle.

[0047] The technical advantages of this embodiment compared to the prior art are as follows: This system was applied to a mixed waste clothing sorting production line and a 72-hour comparative test was conducted. Under the same operating conditions, the system using the D1 architecture (2D vision + fixed-delay grasping / blowing) had an average execution success rate of approximately 72%, an effective running time ratio of 82%, and an average of 1.2 material jams per 8 hours. In contrast, the system using this invention achieved an execution success rate of 93%, an effective running time ratio of 96%, and reduced material jam frequency to 0.2 times per 8 hours. The performance improvement stems primarily from three aspects: First, the depth and material information provided by multimodal perception fusion allows for accurate estimation of key states such as flattening index and curling degree, avoiding invalid executions due to state misjudgment; second, the execution window mechanism dynamically adjusts the blowing area and timing, ensuring that the airflow precisely acts on the flat areas of the fabric, reducing tumbling and landing point deviation; and third, online feedback optimization continuously corrects biases and thresholds, compensating for time-varying factors such as conveyor belt wear and air pressure fluctuations. This comparison fully verifies the advanced nature of this invention in terms of circuit architecture and processing logic.

[0048] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural modifications made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A flexible fabric sorting execution window generation and dynamic control system, characterized in that, include: A multimodal sensing module is configured to acquire image data of fabric on a conveyor belt, wherein the image data includes at least two-dimensional texture information and depth information; The fabric state vector generation module is communicatively coupled with the multimodal perception module and is configured to extract the morphological state parameters of the fabric based on the image data and construct a fabric state vector representing the real-time state of the fabric. The actuator constraint modeling module stores constraint parameters for at least one actuator. The execution window generation module is communicatively coupled to the fabric state vector generation module and the actuator constraint modeling module, and is configured to generate an execution window based on the fabric state vector and the constraint parameters. The execution window includes at least a spatially executable region and an executable time interval on the fabric. The dynamic control module is communicatively coupled to the execution window generation module and is configured to drive the corresponding execution mechanism to perform sorting operations on the spatially executable area within the executable time interval according to the control instructions of the execution window generation mechanism.

2. The flexible fabric sorting execution window generation and dynamic control system according to claim 1, characterized in that, The multimodal perception module includes an RGB-D camera and a near-infrared spectral sensor, configured to perform time synchronization and spatial registration of the acquired RGB images, depth images and near-infrared spectral data, and generate fused data that integrates texture, depth and material information.

3. The flexible fabric sorting execution window generation and dynamic control system according to claim 1, characterized in that, The fabric state vector generation module is configured to run a fabric state analysis model. The fabric state vector includes at least one or more of the following parameters: fabric projected area, flattening index, fold or bulge height, edge exposure length, curling degree, number of overlapping layers, percentage of overlapping area, movement speed, attitude angle, and perception confidence.

4. The flexible fabric sorting execution window generation and dynamic control system according to claim 1, characterized in that, The execution window generation module is configured to generate a multi-dimensional execution window that includes a spatially executable region R, an executable time interval T, an execution mode M, control parameters P, and a risk level U. It then performs a comprehensive evaluation based on the flattening score, edge exposure score, motion stability score, institutional accessibility score, overlap risk, curling risk, and uncertainty of the candidate execution windows to select the optimal execution window.

5. The flexible fabric sorting execution window generation and dynamic control system according to claim 4, characterized in that, It also includes an exception handling module, which is communicatively coupled to the execution window generation module. It is configured to trigger alternative execution strategies, including delayed execution, secondary flattening, reflow processing, or temporary storage for re-inspection, when the risk level U exceeds a preset threshold or the fabric state vector indicates that the fabric is in a state that does not meet the stable execution conditions.

6. The flexible fabric sorting execution window generation and dynamic control system according to claim 1, characterized in that, It also includes an online feedback optimization module, which is communicatively coupled with the dynamic control module and the execution window generation module. It is configured to collect the actual result data after sorting execution and dynamically correct the spatial offset, timing compensation, and executability evaluation threshold of the execution window based on the deviation between the actual result data and the target result.

7. The flexible fabric sorting execution window generation and dynamic control system according to claim 6, characterized in that, The online feedback optimization module is also configured to feed back execution failure samples, window prediction deviation samples, and manual review results to the training sample pool to optimize the fabric state analysis model in the fabric state vector generation module and the executability prediction model in the execution window generation module.

8. The flexible fabric sorting execution window generation and dynamic control system according to claim 1, characterized in that, The actuator is one or more of the following: a spray valve array, a robotic arm, a push rod, a lever, a diverter baffle, or a conveying branch mechanism; the dynamic control module is configured to calculate the corresponding spray valve combination, opening time, and pulse width, or calculate the gripping point and path of the robotic arm, or calculate the action time and stroke of the push rod and the baffle, based on the spatial executable area and executable time interval in the execution window.

9. A method for generating and dynamically controlling the execution window for flexible fabric sorting, applicable to the flexible fabric sorting execution window generation and dynamic control system described in any one of claims 1-8, characterized in that, include: Collect multimodal image data of the fabric on the conveyor belt, wherein the multimodal image data includes at least depth information; Based on the multimodal image data, the morphological state parameters of the fabric are extracted, and a fabric state vector representing the real-time state of the fabric is constructed. Obtain constraint parameters for at least one actuator; An execution window is generated based on the fabric state vector and the constraint parameters. The execution window includes at least a spatially executable region and an executable time interval on the fabric. According to the control instructions of the execution window generation mechanism, the corresponding execution mechanism is driven to perform sorting operations on the spatially executable area within the executable time interval.

10. The method for generating and dynamically controlling the flexible fabric sorting execution window according to claim 9, characterized in that, Also includes: Predict the future trajectory of the fabric after the actuator response delay; The generation of the execution window includes generating a multi-dimensional execution window based on the fabric state vector, the future trajectory, and the constraint parameters, which includes a spatially executable region, an executable time interval, an execution mode, control parameters, and a risk level. The method further includes: collecting execution results and dynamically adjusting the timing compensation amount, executability evaluation threshold, and execution window scoring weight based on the deviation between the execution results and the target results.

Images