Intelligent warehouse logistics line material feeding control method and system based on visual feedback

Abstract

This invention discloses a material supply control method and system for intelligent warehousing and logistics lines based on visual feedback, specifically relating to the field of automation control. The method includes constructing a dynamic perception field integrating multi-view images and 3D point clouds to extract material supply opportunity factors; establishing a dynamic demand potential energy field for downstream workstations, converting opportunity factors into response potential fields, and automatically identifying the optimal supply target through field coupling resonance analysis; constructing an energy state model of the actuator, adjusting virtual potential well parameters based on real-time perception, and generating a time-varying stiffness capture trajectory to complete picking and delivery; calculating the change in system orderliness after supplying material, and dynamically adjusting perception sensitivity and resonance threshold to achieve adaptive optimization; and encoding the successful supply process as a context memory unit, directly activating the memory sequence to execute operations when matching perception patterns. This invention integrates real-time perception, intelligent decision-making, and adaptive execution, improving supply accuracy, efficiency, and system adaptability.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and more specifically, to a method and system for controlling material supply in intelligent warehousing and logistics lines based on visual feedback. Background Technology

[0002] With the rapid development of intelligent manufacturing and logistics automation, warehousing and logistics systems have placed higher demands on the real-time, accuracy and adaptability of material supply. Traditional material supply methods rely on fixed procedures or manual intervention, which are difficult to cope with the dynamic and ever-changing material status and downstream task requirements. There is an urgent need for an integrated control solution that can integrate environmental perception, intelligent decision-making and precise execution to achieve efficient and flexible automated material supply.

[0003] The existing solution is a robot grasping system based on machine vision. It deploys industrial cameras in the feeding area to acquire material location information, and then controls the robotic arm to pick up and place materials according to preset grasping rules. Specifically, it uses image recognition technology to locate the material coordinates, generates a grasping trajectory with fixed stiffness through a path planning algorithm, and finally executes the grasping actions sequentially.

[0004] However, in practical use, it still has some shortcomings, such as a single perception dimension, relying only on two-dimensional images and lacking a comprehensive judgment of the material's three-dimensional posture and surface properties, resulting in insufficient grasping adaptability; a rigid decision-making mechanism, unable to autonomously couple and optimize based on real-time task status and material dynamic characteristics, and the selection of feeding targets is not intelligent enough; a lack of system-level adaptive adjustment capability, unable to optimize control parameters based on historical operation feedback, making it difficult to guarantee efficiency and robustness during long-term operation; and a lack of memory and association capabilities in the execution process, requiring recalculation in similar scenarios, resulting in low response efficiency. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a material supply control method and system for intelligent warehousing and logistics lines based on visual feedback, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a material supply control method for intelligent warehousing logistics lines based on visual feedback, comprising S1: deploying a visual sensor group in the supply area to collect a continuous multi-view image stream, fusing it to generate a real-time dynamic perception field of logistics, embedding an unsupervised feature extraction module, and mapping the material state to opportunity factors, wherein the opportunity factors include operability factors, path compliance factors, and time urgency factors.

[0007] S2: Establish the demand potential energy field of downstream workstations, which dynamically evolves with task priority and real-time status of workstations. Convert the opportunity factors into the response potential field of materials, and calculate the spatial coupling coefficient and temporal correlation coefficient between the response potential field of each material and the demand potential energy field of its location. When a material satisfies that both the spatial coupling coefficient and the temporal correlation coefficient exceed the set threshold, and the field coupling resonance exceeds the dynamically adjusted adaptive threshold, determine that the current material is the optimal material supply target and trigger the material supply decision.

[0008] S3: For the identified material supply target, construct the energy state model of the actuator. The energy state model includes the energy storage state and the energy release triggering condition. Based on the real-time spatial relationship of the sensing field feedback, dynamically adjust the virtual potential well parameters of the actuator. When the energy release condition is met, automatically generate a time-varying stiffness efficient capture trajectory to complete the picking up and directional delivery of the target material.

[0009] S4: Collect system status data after the feeding action is completed, calculate the change in system order caused by the current operation, and dynamically adjust the sensitivity of the sensing field feature extraction and the resonance threshold of the field coupling process based on the change in order, so that the system can achieve continuous adaptive control performance.

[0010] S5: The complete material supply decision and execution process is encoded into a scenario memory unit that includes sensing field feature segments, field coupling parameters and execution energy state sequences. When the system detects a sensing field pattern or physical vibration spectrum feature that highly matches the historical memory unit, it actively activates the associated memory unit and performs real-time fine-tuning based on the stored control sequence to directly drive the material supply operation.

[0011] The intelligent warehousing and logistics line material supply control system based on visual feedback includes a dynamic perception field construction and feature extraction module: including a visual sensor group and edge computing unit deployed in the feeding area, which integrates multi-view images and three-dimensional point cloud data to generate a real-time dynamic perception field, and maps the material state into opportunity factors through unsupervised feature extraction.

[0012] Field Coupling Decision and Target Identification Module: Includes a workstation status monitoring unit and a central decision unit, used to construct a dynamic demand potential energy field based on the real-time status of downstream workstations, convert the opportunity factor into a material response potential field, calculate the spatial and temporal coupling degree between the two fields, and autonomously identify the optimal material supply target and trigger decision instructions when the field coupling resonance exceeds the adaptive threshold.

[0013] The energy state model-driven execution control module includes an actuator driver and a trajectory planner, which are used to construct the energy state model of the actuator, dynamically adjust the virtual potential well parameters according to the feedback of the sensing field, and generate a time-varying stiffness efficient capture trajectory when the energy storage-release conditions are met, so as to complete the precise picking and directional delivery of materials.

[0014] System entropy feedback adaptive module: Collects performance index data after material feeding is completed, calculates the change in system orderliness, and dynamically adjusts feature extraction sensitivity and field coupling resonance threshold to achieve closed-loop optimization of control parameters;

[0015] Contextual memory and associative execution module: The successful feeding process is encoded as a contextual memory unit. When a real-time sensing pattern or vibration spectrum that highly matches the historical memory is detected, the corresponding memory unit is activated and the feeding operation is directly driven after fine-tuning based on its stored sequence.

[0016] The technical effects and advantages of this invention are as follows:

[0017] 1. This invention constructs a high-resolution dynamic perception field by integrating multi-view vision and three-dimensional point cloud data, and combines an opportunity factor and demand potential energy field coupled resonance decision-making model. It integrates multi-dimensional information such as the three-dimensional posture of the material, surface properties and the urgency of downstream tasks, and intelligently filters and identifies the optimal material supply target. This overcomes the problems of poor grasping adaptability and rigid decision-making caused by the reliance on two-dimensional positioning and fixed rules in traditional methods.

[0018] 2. This invention introduces an orderliness evaluation mechanism based on entropy weight method. Based on the feedback of multiple performance indicators after each material feeding operation, the core parameters of sensing sensitivity and field coupling threshold are dynamically adjusted to form a complete closed loop of "sensing-decision-execution-evaluation". This enables the system to continuously learn and optimize its own performance and effectively cope with environmental changes and task disturbances.

[0019] 3. By encoding the successful feeding process into a storable and recallable context memory unit, this invention directly activates the historical optimal control sequence and fine-tunes its execution when a similar perception pattern or physical scene is detected. This avoids the repetitive calculation of complex models, significantly shortens the system's response time in repetitive or similar tasks, and improves the overall feeding efficiency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0021] Figure 2 This is a schematic diagram of the dynamic sensing field construction of the present invention;

[0022] Figure 3 This is a schematic diagram of the field coupling resonance decision-making process of the present invention;

[0023] Figure 4 This is a schematic diagram of the energy state model and trajectory execution of the present invention;

[0024] Figure 5 This is a schematic diagram illustrating the execution of the contextual memory and association mechanism of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] As attached Figure 1 Appendix Figure 2 The material supply control method for intelligent warehousing logistics lines based on visual feedback shown includes S1: deploying a visual sensor group in the supply area to collect a continuous multi-view image stream, fusing them to generate a real-time dynamic perception field of logistics, embedding an unsupervised feature extraction module, and mapping the material state to opportunity factors, wherein the opportunity factors include operability factors, path compliance factors, and time urgency factors.

[0027] It should be specifically noted that the opportunity factor is a dimensionless comprehensive evaluation index that comprehensively characterizes whether materials are easily captured and transported by the current system and meet downstream timeliness requirements. Its value range is [0,1]. The opportunity factor is composed of three sub-factors: operability factor F1, path compliance factor F2, and timeliness urgency factor F3, with linear weighting. The calculation formula is as follows: The weights are 0.4, 0.3, and 0.3, which correspond to F1, F2, and F3, respectively. The grid search method is based on 1,000 sets of experimental data on the supply of warehouse materials. The samples cover fragile items, rigid parts, and irregularly shaped parts. The weights are determined by minimizing the objective function of the supply error. The delivery accuracy is ±0.1mm and the response time is ≤0.5s.

[0028] Further explanation is needed regarding the deployment of the vision sensor group and image acquisition, specifically including: the feeding area is divided into 3 sensing sub-regions (feed inlet area, buffer area, and grasping area), with each sub-region deploying 1 set of industrial cameras (resolution 2592×1944, frame rate 30fps) and 1 laser contour sensor (measurement accuracy 0.1μm); the industrial cameras acquire RGB image streams to obtain the material's color, texture, and 2D position information; the laser contour sensor acquires the material's 3D contour data to obtain the material's height, volume, and orientation information; during the acquisition process, a timestamp synchronization mechanism ensures the temporal consistency of multi-view images, with a synchronization error ≤1ms.

[0029] The edge computing unit preprocesses multi-view RGB images, including using adaptive histogram equalization to eliminate the influence of illumination changes, removing noise points from the image through an improved RANSAC algorithm, and introducing spatial neighborhood weights to improve the accuracy of noise removal. Specifically, it constructs a 3×3 (stable illumination scene) or 5×5 (fluctuating illumination scene) spatial neighborhood window centered on any pixel P(i,j) in the image, and calculates the gray-level gradient consistency G within the window. 灰度 and location clustering L 聚集 When G 灰度 With L 聚集 If the preset threshold is met, pixel P(i,j) is included in the set of valid points; otherwise, it is determined to be a noise point.

[0030] G 灰度 calculate: ,in Let be the Sobel gradient magnitude of pixel P(i,j). Let be the gradient magnitude of any pixel in the neighborhood, and n be the total number of pixels in the window. G is the global maximum gradient value of the image. 灰度 ∈[0,1], G 灰度 The closer the value is to 1, the more consistent the grayscale change trend of the pixel with its neighbors. A threshold is set accordingly. The images of stored materials were statistically calibrated.

[0031] L 聚集 calculate: ,in The grayscale value is within the neighborhood, and the range is [value]. The number of pixels, the interval is obtained by adaptive thresholding of the image, L 聚集 ∈[0,1],L 聚集 The closer the value is to 1, the denser the effective material points around the pixel. A threshold is set accordingly. The images of stored materials were statistically calibrated.

[0032] In the RANSAC interior point determination phase, the initial valid point set G is selected through the aforementioned dual threshold filtering. 灰度 ≥ And L 聚集 ≥ Then, the model is fitted based on the initial effective point set, and the subsequent residual judgment is only for pixels in the initial set, avoiding the problem of misjudging isolated noise points and material edges.

[0033] A combined algorithm of direct-pass filtering and statistical filtering is used to perform point cloud filtering on the laser contour data, removing distance anomalies and isolated points. An attention-based image fusion algorithm is employed, using material contour sharpness as the attention weight to avoid missing information from a single viewpoint. The preprocessed multi-view RGB images are fused with the 3D point cloud data. The specific fusion process involves: calibrating multi-source sensors using a checkerboard calibration board to obtain the intrinsic and extrinsic parameters of the industrial camera, as well as the relative extrinsic parameters between the laser contour sensor and the camera; and using a coordinate transformation matrix to uniformly map the pixel coordinates of all viewpoint RGB images and the 3D point cloud world coordinates to the material feeding area world. A coordinate system is used to ensure that the positions of the same material points are accurately aligned in the two types of data, with the alignment error controlled within 0.1mm. Material edge texture features are extracted from RGB images, and surface normal vectors and curvature features are extracted from 3D point clouds to construct a 0.5mm×0.5mm×0.5mm 3D grid field. Each grid cell synchronously stores multi-dimensional information such as color, texture, spatial coordinates, and normal vectors, forming a real-time dynamic perception field that includes the material's two-dimensional position, three-dimensional pose, volume, and texture features. The spatial resolution of the perception field is set to 0.5mm×0.5mm, and the temporal resolution is consistent with the camera frame rate (30Hz).

[0034] An unsupervised feature extraction module based on a variational autoencoder (VAE) is embedded in the sensing field. Prior constraints based on material physical properties are introduced to improve the targeting of feature extraction. The extracted features include: material surface roughness, contour regularity, spatial dispersion, and motion trend. The reasons for this selection are: surface roughness directly affects grasping stability (a core factor influencing operability), contour regularity determines the adaptability of the grasping posture (a core factor influencing path compliance), and spatial dispersion and motion trend determine the timeliness of material supply (a core factor influencing timeliness urgency).

[0035] The extracted features are normalized using min-max normalization, mapping the feature values ​​to the [0,1] interval. The random factor is then calculated. Among them, the operability factor F1 is obtained by weighting surface roughness features and contour regularity features (weight ratio 1:1), the path compliance factor F2 is obtained by weighting contour regularity features and spatial location dispersion features (weight ratio 2:1), and the timeliness factor F3 is obtained by weighting spatial location dispersion features and movement trend features (weight ratio 1:2). The weight ratios are initially calibrated using grid search.

[0036] As attached Figure 3As shown, S2: Establish the demand potential energy field of the downstream workstation, which dynamically evolves with the task priority and the real-time status of the workstation. Convert the opportunity factor into the response potential field of the material, and calculate the spatial coupling coefficient and temporal correlation coefficient between the response potential field of each material and the demand potential energy field of its location. When there is a material that satisfies that both the spatial coupling coefficient and the temporal correlation coefficient exceed the set threshold, and the field coupling resonance exceeds the adaptive threshold of dynamic adjustment, determine that the current material is the optimal material supply target and trigger the material supply decision.

[0037] It should be specifically noted that the demand potential energy field is a scalar field defined in the warehouse space coordinate system, used to quantify the urgency of the demand for materials at any point in space at a certain moment. It takes each downstream workstation as the potential energy source, the intensity of which is determined by the real-time status of the workstation and decays with distance in space.

[0038] The response potential field is a scalar defined on the material to quantify the material's "willingness" or "fit" to respond to downstream demands. It is derived from the material's opportunity factor combined with the current environmental context.

[0039] The field coupling resonance refers to a state in which the response potential field of a certain material and the demand potential field of a certain point in space exhibit a high degree of matching in terms of spatial location and change sequence.

[0040] Further explanation is needed regarding the collection of status data from downstream workstations, including the quantity of materials awaiting completion at each workstation (Q); task priority (P); current workstation operating efficiency (A, units: pieces / minute); and workstation equipment health status (H), calculated using vibration sensor data, ranging from 0 to 1, with 1 being the optimal value. Task priority (P) is divided into 1-5 levels, with level 1 being the highest. The specific grading standards are as follows: Level 1 is an urgent material-awaiting task at a bottleneck workstation; a material shortage will directly lead to a complete line shutdown. Level 2 is a high-priority task at a core process workstation; a material shortage will cause three or more downstream workstations to wait. Level 3 is a standard task at a regular process workstation; a material shortage only affects the operation of this workstation. Level 4 is a low-priority task at an auxiliary process workstation, which can switch to backup materials to maintain operation. Level 5 is a material replenishment task at a buffer workstation, where there is currently no immediate production demand. The rationale for this selection is that the quantity of materials awaiting completion directly reflects the urgency of the demand, task priority determines the material supply sequence, operating efficiency determines the matching degree of the material supply rhythm, and equipment health status avoids supplying materials to faulty workstations, preventing material accumulation.

[0041] The above data is mapped to the workstation demand potential value U using a fuzzy logic algorithm. The fuzzy rules are set as follows: when Q≥5 and P≥4, U≥0.8 (high demand); when 3≤Q<5 and 3≤P<4, 0.5≤U<0.8 (medium-high demand); when 1≤Q<3 and 2≤P<3, 0.2≤U<0.5 (medium-low demand); when Q=0 or P=1, U<0.2 (low demand). The threshold is calibrated based on the physical properties of the material and the historical safety margin.

[0042] The demand potential field is centered on each workstation, and a Gaussian decay model is adopted according to the distance decay law: Where r is the distance from the material to the workstation, and σ is the attenuation coefficient, which is dynamically adjusted according to the warehouse layout, with a value range of 0.5-2m. It forms a spatial distribution in the material supply area and dynamically evolves with the real-time update of workstation status data (update frequency 30Hz).

[0043] The opportunity factor F is converted into the material's response potential field V, and the conversion formula is as follows: Where k is a conversion coefficient, which is dynamically adjusted by the material density in the current feeding area. When the material density ρ≥5 pieces / m², k=1.2; when 3≤ρ<5 pieces / m², k=1.0; when ρ<3 pieces / m², k=0.8. This ensures that materials with high chance factors are selected first when the materials are dense, and the selection range is expanded when the materials are sparse. The threshold is calibrated according to the physical properties of the materials and the historical safety margin.

[0044] A field coupling coefficient C is introduced to characterize the degree of coupling between the response potential field and the demand potential energy field. The field coupling coefficient is calculated. The value range is [-1, 1]. When C ≥ 0.7, it is determined to be a spatial coupling match. Simultaneously, the temporal correlation coefficient R between the two is calculated using the Pearson correlation coefficient, with a value range of [-1, 1]. When R ≥ 0.6, it is determined to be a temporal coupling match. The adaptive threshold B is dynamically adjusted based on the system's historical material supply success rate. S represents the success rate of the last 100 feeding attempts, where S = number of successful feeding attempts / total number of feeding attempts, and its value ranges from [0,1]. When C ≥ 0.7 and R ≥ 0.6, and... When the field coupling resonance exceeds the adaptive threshold, it is automatically identified as the current optimal feeding target, triggering a feeding decision. Based on 500 historical feeding data, the threshold is determined using the "ROC curve analysis method".

[0045] As attached Figure 4 As shown, S3: For the identified material supply target, construct the energy state model of the actuator. The energy state model includes the energy storage state and the energy release triggering condition. Based on the real-time spatial relationship of the sensing field feedback, dynamically adjust the virtual potential well parameters of the actuator. When the energy release condition is met, automatically generate a time-varying stiffness efficient capture trajectory to complete the picking up and directional delivery of the target material.

[0046] It should be specifically noted that the energy state model is an abstract description of the internal state and triggering mechanism of an actuator (such as a robotic arm), comprising two core parts:

[0047] Energy storage state E: A dimensionless comprehensive evaluation index in the range of [0,1], used to quantify whether the actuator has sufficient motion capacity to perform efficient and stable grasping at the current moment. It is obtained by weighted fusion of characteristic quantities reflecting the thermal load of the motor and the motion state of the mechanism after normalization.

[0048] Release trigger condition D: Logical judgment condition. When all conditions are met, the trigger state machine changes from the "energy storage" state to the "energy release execution" state, and initiates trajectory planning and execution. The core conditions include the distance threshold between the end effector and the target material and the expected clamping force threshold.

[0049] The virtual potential well is an abstract potential field model constructed for trajectory planning, used in control algorithms to guide the end effector of the actuator to approach and stably grasp the target. The mathematical model is typically expressed as follows: ,in The current position of the end. For the target location, K 势阱 To determine the potential well stiffness, K is adjusted. 势阱 Sum of potential well depth H 势阱 This changes the stiffness and convergence characteristics of the trajectory.

[0050] It should be further explained that, using the SCARA robot and the electric gripper as actuators, an energy state model is constructed. The model includes the energy storage state E and the energy release triggering condition D. The core data selected include: the motor current I of each joint of the robot and the equivalent resistance of the motor windings. Current duration within the sampling period Equivalent rotational inertia of joints , Joint angular velocity ω, gripper clamping force N, distance d from robot end effector to material.

[0051] The energy storage state E is calculated as follows: Where 0.6 and 0.4 are weighting coefficients calibrated through experiments; the energy release triggering condition D is set as follows: d≤5mm and N reaches the minimum clamping force threshold of the material, where N is calculated from the material weight and friction coefficient. m is the mass of the material, g is the gravitational acceleration, and μ is the static friction coefficient between the gripper and the material.

[0052] Based on the real-time 3D attitude of the material (position coordinates (x,y,z), attitude angle (α,β,γ)) and the real-time spatial position of the robot's end effector, a virtual potential well model is constructed. The core parameters of the virtual potential well include the potential well depth H. 势阱 Sum of potential well stiffness K 势阱 H is dynamically adjusted according to the energy storage state E and the material state. 势阱 and K 势阱When E≥0.8 (high energy storage) and the material is fragile (judged by the hardness characteristics of the material in the sensing field), H is set. 势阱 =0.3, K 势阱 =5N / m (low stiffness, avoid impact); when E≤0.5 (low energy storage) and the material is rigid, set H. 势阱 =0.8, K 势阱 =20N / m (high rigidity, improving gripping efficiency), the threshold is calibrated based on the material's physical properties and historical safety margin.

[0053] Potential well parameter K 势阱 H 势阱 The system adjusts the energy storage state E and material hardness online to achieve the approximate behavior of "time-varying stiffness". For example, when grasping fragile materials with high energy storage (large E), smaller K and H are used to achieve compliant contact; when grasping rigid materials with low energy storage (small E), larger K and H are used to achieve fast and rigid positioning.

[0054] When the energy release trigger condition is met, a time-varying stiffness capture trajectory is generated based on the Model Predictive Control (MPC) algorithm. The core design is as follows: the prediction model adopts a linearized discrete-time dynamics model to balance computational complexity and prediction accuracy; the state vector is selected as the position and velocity of the robot's end effector in Cartesian space, and the control input is the end effector acceleration. The objective function solves an optimization problem in the finite time domain at each control time, designed to balance trajectory tracking accuracy, control smoothness, and energy characteristics; the constraint optimization solution satisfies the following physical and task constraints, where the time-varying stiffness is provided in real time by the task stage (approach / grasp / deployment) and material hardness online decision module, and is input as a parameter to the MPC.

[0055] The trajectory planning is divided into three stages: the approach stage, from the current position to 5mm above the material, uses S-curve acceleration and deceleration, with stiffness gradually decreasing as the distance decreases; the grasping stage, from 5mm above the material to the clamping position, uses constant force control, with stiffness dynamically matching the material hardness; and the delivery stage, from the clamping position to the designated position at the downstream workstation, uses adaptive speed control, with stiffness adjusted according to the delivery accuracy requirements. During trajectory generation, real-time data from the robot's end-effector displacement sensor (sampling frequency 1kHz) is collected, and trajectory deviations are corrected using a PID algorithm to ensure positioning accuracy ≤ ±0.1mm.

[0056] S4: Collect system state data after the feeding action is completed, calculate the change in system order caused by the current operation, and dynamically adjust the sensitivity of the sensing field feature extraction and the resonance threshold of the field coupling process based on the change in order, so that the system can achieve continuous adaptive control performance.

[0057] It should be specifically noted that the system orderliness is a comprehensive evaluation index that assesses the degree of improvement in the operational order of the entire warehousing and logistics system after a single material supply operation. The value range is [0,1]. It is obtained by normalizing and weighting multiple performance indicators (such as completion time, accuracy, energy consumption, etc.). The higher the orderliness, the more efficient and stable the system operation.

[0058] It should be further noted that after the material feeding action is completed, the collected system status data includes the material feeding completion time T. 完成 Material delivery accuracy Robot energy consumption Eh, downstream workstation waiting time T 等待 Material breakage rate G.

[0059] The system's orderliness O is defined as the comprehensive evaluation result of the above indicators. The entropy weight method is used to determine the weight of each indicator, T. 完成 , Eh, T 等待 The weights of G are 0.25, 0.2, 0.2, 0.2, and 0.15 respectively. The degree of order O is calculated as follows: ,in Let i be the weight of the i-th indicator. Given the entropy value of the i-th index (the larger the entropy value, the higher the disorder of the index), calculate the change in the system's orderliness caused by the current operation. , For the current degree of order, This refers to the orderliness of the material supply since the last feeding.

[0060] according to The sensitivity 'a' for extracting the features of the sensing field and the resonance threshold 'b' for the field coupling process are dynamically adjusted; the adjustment rule is: when... When ≥0.1 (significantly improved orderliness), based on the current value, a increases by 10% and b decreases by 5%, enhancing perception sensitivity and expanding the target selection range for feed materials; when -0.05≤ When <0.1 (order degree is basically stable), a and b remain unchanged; when When the value is less than -0.05 (orderliness decreases), a decreases by 15% and b increases by 8%, thus reducing the sensitivity of perception and strictly screening the material supply targets.

[0061] As attached Figure 5 As shown, S5: The complete material feeding decision and execution process is encoded into a scenario memory unit that includes sensing field feature segments, field coupling parameters and execution energy state sequences. When the system detects a sensing field pattern or physical vibration spectrum feature that highly matches the historical memory unit, it actively activates the associated memory unit and performs real-time fine-tuning based on the stored control sequence to directly drive the material feeding operation.

[0062] It should be specifically noted that the scenario memory unit is a compressed coded representation of a successful feeding process. It is a data structure that includes a snapshot of the perception features of the triggering scenario; key parameters of the decision-making process; control sequence of the execution process; and a strategy for quickly retrieving near-optimal strategies when encountering similar scenarios.

[0063] It should be further explained that the complete material supply decision and execution process is encoded into a contextual memory unit. Each memory unit includes three core parts: a sensing field feature segment, field coupling parameters, and an execution energy state sequence. The sensing field feature segment selects the core features of the material supply target, specifically including surface roughness, contour regularity, and three-dimensional position. Principal component analysis (PCA) is used for dimensionality reduction, retaining 95% of the feature information and reducing storage requirements. The field coupling parameters include the coupling coefficient, temporal correlation coefficient, and adaptive threshold. The execution energy state sequence is a temporal sequence (sampling interval 10ms) of the energy storage state, virtual potential well parameters, and coordinates of key trajectory points during the material supply process.

[0064] During the encoding process, a lossless compression algorithm based on LZW is used to compress the memory units and store them in the context memory storage unit (SSD solid-state drive). At the same time, the most recent (nearly 50 times) memory units are loaded into the memory-mapped cache to improve the reading speed.

[0065] The system collects the current sensing field pattern in real time and extracts the core features, physical vibration spectrum features of the actuator (collected by vibration sensors and converted into spectrum features by Fast Fourier Transform (FFT)); calculates the cosine similarity between the current sensing field features and those in the historical memory unit, and the Euclidean distance between the current vibration spectrum features and the spectrum features of the corresponding execution stage in the historical memory unit.

[0066] The activation conditions are set as follows: cosine similarity ≥ 0.85 and Euclidean distance ≤ 0.2. When the activation conditions are met, the associated memory unit is actively activated, and the stored field coupling parameters and execution energy state sequence are extracted as the substrate. Based on the difference between the current sensing field and the historical sensing field, the incremental PID algorithm is used to fine-tune the field coupling parameters and execution energy state sequence in real time. The fine-tuning object is the deviation between the target position of the current cycle and the expected target position of the historical trajectory starting point. The fine-tuned parameters directly drive the material feeding operation, which shortens the response time of the system in similar scenarios.

[0067] The intelligent warehousing and logistics line material supply control system based on visual feedback includes a dynamic perception field construction and feature extraction module: including a visual sensor group and edge computing unit deployed in the feeding area, which integrates multi-view images and 3D point cloud data to generate a real-time dynamic perception field, and maps the material state into opportunity factors through unsupervised feature extraction.

[0068] Field Coupling Decision and Target Identification Module: Includes a workstation status monitoring unit and a central decision unit, used to construct a dynamic demand potential energy field based on the real-time status of downstream workstations, convert the opportunity factor into a material response potential field, calculate the spatial and temporal coupling degree between the two fields, and autonomously identify the optimal material supply target and trigger decision instructions when the field coupling resonance exceeds the adaptive threshold.

[0069] The energy state model-driven execution control module includes an actuator driver and a trajectory planner, which are used to construct the energy state model of the actuator, dynamically adjust the virtual potential well parameters according to the feedback of the sensing field, and generate a time-varying stiffness efficient capture trajectory when the energy storage-release conditions are met, so as to complete the precise picking and directional delivery of materials.

[0070] System entropy feedback adaptive module: Collects performance index data after material feeding is completed, calculates the change in system orderliness, and dynamically adjusts feature extraction sensitivity and field coupling resonance threshold to achieve closed-loop optimization of control parameters.

[0071] Contextual memory and associative execution module: The successful feeding process is encoded as a contextual memory unit. When a real-time sensing pattern or vibration spectrum that highly matches the historical memory is detected, the corresponding memory unit is activated and the feeding operation is directly driven after fine-tuning based on its stored sequence.

[0072] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments of this disclosure. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0073] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for material feeding control of intelligent warehouse logistics line based on visual feedback, characterized in that, include: S1: Deploy a visual sensor group in the material supply area to collect a continuous multi-view image stream, fuse it to generate a real-time dynamic perception field of logistics, embed an unsupervised feature extraction module, and map the material status into opportunity factors, including operability factors, path compliance factors and time urgency factors. S2: Establish the demand potential energy field of downstream workstations, which dynamically evolves with task priority and real-time status of workstations. Convert the opportunity factors into the response potential field of materials, and calculate the spatial coupling coefficient and temporal correlation coefficient between the response potential field of each material and the demand potential energy field of its location. When a material satisfies that both the spatial coupling coefficient and the temporal correlation coefficient exceed the set threshold, and the field coupling resonance exceeds the dynamically adjusted adaptive threshold, determine that the current material is the optimal material supply target and trigger the material supply decision. S3: For the identified material supply target, construct the energy state model of the actuator. The energy state model includes the energy storage state and the energy release triggering condition. Based on the real-time spatial relationship of the sensing field feedback, dynamically adjust the virtual potential well parameters of the actuator. When the energy release condition is met, automatically generate a time-varying stiffness efficient capture trajectory to complete the picking up and directional delivery of the target material. S4: Collect system status data after the feeding action is completed, calculate the change in system order caused by the current operation, and dynamically adjust the sensitivity of the sensing field feature extraction and the resonance threshold of the field coupling process based on the change in order, so that the system can achieve continuous adaptive control performance. S5: The complete material supply decision and execution process is encoded into a scenario memory unit that includes sensing field feature segments, field coupling parameters and execution energy state sequences. When the system detects a sensing field pattern or physical vibration spectrum feature that highly matches the historical memory unit, it actively activates the associated memory unit and performs real-time fine-tuning based on the stored control sequence to directly drive the material supply operation. 2.The visual feedback based intelligent warehouse logistics line material feeding control method according to claim 1, characterized in that: The real-time dynamic sensing field of logistics is generated through a fusion process that includes: aligning the RGB image stream acquired by the industrial camera with the three-dimensional contour data acquired by the laser contour sensor at the pixel level through multi-sensor spatial registration and timestamp synchronization, thereby constructing a high-resolution rasterized sensing field that combines two-dimensional texture features and three-dimensional spatial attributes. 3.The visual feedback based intelligent warehouse logistics line material feeding control method of claim 1, wherein: The dynamic evolution process of the demand potential energy field is based on the multi-dimensional real-time status data of downstream workstations. The quantity of materials to be delivered, task priority, work efficiency and equipment health status are mapped to demand potential energy values ​​through fuzzy logic rules, and a Gaussian decay model is used to form a potential energy distribution field that decays with distance in space. 4.The visual feedback based intelligent warehouse logistics line material feeding control method of claim 1, wherein: The field coupling resonance is determined based on the following criteria: calculating the spatial coupling coefficient and temporal correlation coefficient between the response potential field and the demand potential field; dynamically calculating and comparing the adaptive threshold based on the historical material supply success rate; and triggering a material supply decision when the spatial, temporal, and intensity conditions are met simultaneously. 5.The visual feedback based intelligent warehouse logistics line material feeding control method of claim 1, wherein: The energy state model is constructed based on: the real-time dynamic parameters of the actuator; the energy storage state is calculated by fusing the current and angular velocity characteristics of the joint motor; and the energy release triggering condition is jointly determined by the distance between the end effector and the material and the clamping force threshold. 6.The visual feedback based intelligent warehouse logistics line material feeding control method according to claim 1, characterized in that: The efficient capture trajectory is generated by a model predictive control algorithm. The stiffness parameter is adjusted online according to the material hardness and energy storage state. The trajectory planning includes a variable stiffness control strategy in three stages: approach, grasping, and release. 7.The visual feedback based intelligent warehouse logistics line material feeding control method of claim 1, wherein: The change in the system's orderliness is calculated by using the entropy weight method to fuse multiple indicators, including material supply completion time, delivery accuracy, energy consumption, downstream waiting time, and material breakage rate. Based on the sign and magnitude of the change, a differentiated adjustment strategy is implemented for the sensitivity of the sensing field feature extraction and the field coupling resonance threshold. 8.The visual feedback based intelligent warehouse logistics line material feeding control method of claim 1, wherein: The context memory unit is encoded using principal component analysis to reduce the dimensionality of the perceptual field feature fragments and a lossless compression algorithm for storage; the conditions for actively activating the associated memory unit include the cosine similarity threshold between the current perceptual field features and historical memories, and the Euclidean distance threshold between the current physical vibration spectrum and the historical spectrum.

9. The system for the method of any one of claims 1-8, wherein the system comprises: a visual feedback system; a control system; and a material supply system. include: Dynamic sensing field construction and feature extraction module: including a visual sensor group and edge computing unit deployed in the feeding area, which integrates multi-view images and 3D point cloud data to generate a real-time dynamic sensing field, and maps the material state into opportunity factors through unsupervised feature extraction; Field Coupling Decision and Target Identification Module: Includes a workstation status monitoring unit and a central decision unit, used to construct a dynamic demand potential energy field based on the real-time status of downstream workstations, convert the opportunity factor into a material response potential field, calculate the spatial and temporal coupling degree between the two fields, and autonomously identify the optimal material supply target and trigger decision instructions when the field coupling resonance exceeds the adaptive threshold. The energy state model-driven execution control module includes an actuator driver and a trajectory planner, which are used to construct the energy state model of the actuator, dynamically adjust the virtual potential well parameters according to the feedback of the sensing field, and generate a time-varying stiffness efficient capture trajectory when the energy storage-release conditions are met, so as to complete the precise picking and directional delivery of materials. System entropy feedback adaptive module: Collects performance index data after material feeding is completed, calculates the change in system orderliness, and dynamically adjusts feature extraction sensitivity and field coupling resonance threshold to achieve closed-loop optimization of control parameters; Contextual memory and associative execution module: The successful feeding process is encoded as a contextual memory unit. When a real-time sensing pattern or vibration spectrum that highly matches the historical memory is detected, the corresponding memory unit is activated and the feeding operation is directly driven after fine-tuning based on its stored sequence.

Images