Data preset learning-based dynamic material allocation system for multiple stations

Abstract

The invention relates to the technical field of resource allocation, in particular to a data preset learning-based dynamic material allocation system for multiple stations, which comprises a topology modeling module, a demand clustering module, a strategy generation module and a decision optimization module. According to the method, laser ranging and dynamic time warping are combined, the work station layout is converted into a moving time consumption parameter, a space correlation coefficient is adjusted according to a difference value proportion, a topological relation matrix is updated in real time, DTW quantifies the mode similarity of a demand interval standard deviation and a usage variable coefficient, a multi-dimensional demand feature vector is constructed, and the prediction and actual consumption matching degree is improved; the method comprises the steps of establishing a response capability curve based on OEE data, dynamically associating residual capacity with demand characteristics through linear programming, realizing strategy pre-generation and capacity early warning, constructing a cooperation unit through a Hungary algorithm, matching material and equipment parameters through cosine similarity, triggering strategy reconstruction, synchronously updating topology, forming a closed-loop optimization system and enhancing the cross-station dynamic response capability.

Description

Technical Field

[0001] The present invention relates to the technical field of resource allocation, and particularly to a material dynamic allocation system based on data pre-set learning for multi-workstations. Background Art

[0002] Resource allocation technology is a systematic technology for spatio-temporal allocation and task scheduling of limited resources such as personnel, equipment, and materials through scientific methods. Its system covers core modules such as resource demand forecasting, supply planning, scheduling model construction and optimization, and real-time monitoring and adjustment, and is widely applied in fields such as manufacturing, logistics, and project management. Among them, as a typical application, the material dynamic allocation system relies on algorithm models and data system support, and through technologies such as real-time inventory monitoring, demand intelligent identification, supply-demand matching decision-making, and path optimization selection, constructs a resource priority evaluation mechanism based on historical data and production plans to achieve the efficient transfer of materials from storage to usage units. Its core is to dynamically optimize the configuration of resource allocation strategies and transportation paths through a parametric model.

[0003] The prior art relies on fixed historical data and rules to evaluate resource priorities and cannot respond to changes in production plans in real time, resulting in delays in supply-demand matching. The static parameter model ignores the dynamic changes in equipment efficiency, and there is no warning mechanism when the workstation load approaches the critical value, which is likely to cause imbalance in resource allocation. The path selection is based on a fixed topological relationship and does not consider the spatial correlation changes after layout adjustment, resulting in deviations in transportation time consumption and a decrease in efficiency. The demand identification does not quantify the pattern similarity of intervals and usage fluctuations, and single-dimensional prediction is difficult to match complex consumption characteristics, resulting in inventory redundancy. The collaborative scheduling lacks the multi-dimensional adaptation ability of equipment parameters and material attributes, and compatibility problems are likely to occur during cross-workstation collaboration, increasing the cost of secondary allocation. For example, in a flexible manufacturing scenario, the fixed topological relationship causes the material cross-region transportation time consumption to exceed expectations, and uneven equipment loads cause line blockages. Summary of the Invention

[0004] The purpose of the present invention is to solve the disadvantages existing in the prior art, and to propose a material dynamic allocation system based on data pre-set learning for multi-workstations.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A material dynamic allocation system based on data pre-set learning for multi-workstations includes: A topology modeling module, configured to obtain a set of workstation layout coordinates through a laser ranging device, input the coordinate set into a dynamic time warping algorithm to calculate moving time consumption parameters, adjust the spatial correlation coefficient based on the product of the difference ratio of the time consumption parameters and the equipment moving acceleration, generate a topological relationship matrix, and transmit it to the demand clustering module; A demand clustering module, which is used to call the material demand time series record of the production line MES system, input the demand interval standard deviation and the single - use coefficient of variation into the dynamic time warping algorithm to calculate the pattern similarity, generate a demand feature vector, and perform a Cartesian product operation with the topological relationship matrix and then transfer it to the policy generation module; A policy generation module, which is used to establish a response - ability curve equation based on the equipment OEE data, activate the constraint condition when the number of allocated material types at the workstation reaches the inflection point of the curve equation, input the ratio of the demand feature vector to the remaining capacity of the workstation into the linear programming model to generate a set of preset strategies, and transfer it to the decision - making optimization module.

[0006] As a further solution of the present invention, the topological relationship matrix includes a spatial correlation coefficient, a workstation layout structure, and a coordinate space position relationship. The demand feature vector includes a demand pattern similarity, a demand interval regularity, and a material usage feature. The set of preset strategies is specifically a material configuration strategy, a capacity matching plan, and a workstation load balancing plan.

[0007] As a further solution of the present invention, the dynamic time warping algorithm dynamically corrects the spatial correlation coefficient by matching the morphological difference degree of the moving trajectory and the deviation coefficient of the time reference value; The dynamic time warping algorithm performs sequence alignment with a constraint condition that the window width is twice the demand interval standard deviation; The response - ability curve equation fits the corresponding relationship between the overall equipment efficiency and the material handling volume through cubic polynomial regression, and the regression fitting goodness - of - fit threshold is set to 0.85.

[0008] As a further solution of the present invention, the topological modeling module includes: a coordinate acquisition sub - module scans the workstation layout through a laser ranging device, acquires the three - dimensional coordinate data of multiple equipment nodes, uses the Gaussian filtering algorithm to eliminate environmental noise interference, sets the coordinate offset threshold to 10% of the average distance between adjacent nodes as the reference value, performs interpolation compensation on abnormal coordinate points, and generates a workstation coordinate set; A moving time - consuming calculation sub - module calls the workstation coordinate set, extracts the Euclidean distance between adjacent coordinate points, takes the ratio of the equipment moving speed to the trajectory length as the time reference value, aligns the time series of the moving trajectory through the dynamic time warping algorithm, and when calculating the deviation coefficient of the trajectory morphological difference degree and the time reference value, uses the morphological difference degree = the trajectory curvature change rate × the square of the time reference value to obtain the moving time - consuming parameter; A topological relationship construction sub - module calculates the time difference ratio between different moving paths based on the moving time - consuming parameter, performs row - vector normalization processing on the spatial correlation matrix with the reciprocal of the path length as the weight coefficient, and performs a Hadamard product operation on the normalization result and the time difference ratio to generate a topological relationship matrix; The row vector normalization process is achieved by calculating the Euclidean length of each row element and performing a division operation.

[0009] As a further solution of the present invention, the demand clustering module includes: the demand feature extraction sub-module calls the material demand time series record of the production line MES system, extracts the time interval sequence of adjacent demand events, calculates the sequence standard deviation as the time fluctuation amount, collects the single material usage data, calculates the coefficient of variation as the usage discreteness amount, performs Z-score normalization processing on the time fluctuation amount and the usage discreteness amount, and generates a demand feature vector; The pattern similarity calculation sub-module, based on the demand feature vector, uses the dynamic time warping algorithm to stretch and compress the time axis of the differential demand sequence, calculates the cumulative distance value of the minimum bending path between sequences, and maps the distance value to a preset similarity conversion function to obtain a pattern similarity coefficient; The matrix operation sub-module calls the pattern similarity coefficient and the topological relationship matrix, constructs the Cartesian product space of the coefficient matrix and the topological matrix, performs the product operation of the corresponding positions of the matrix elements, sums the product results by row vector, and generates a demand-topology association matrix; The Z-score normalization process uses the data of the last 30 production cycles as the calculation benchmark; The similarity conversion function is an S-shaped curve function, and its slope parameter is inversely proportional to the material allocation response time threshold; The Cartesian product space improves the material distribution efficiency by increasing the weight factor of the moving time between workstations.

[0010] As a further solution of the present invention, the strategy generation module includes: the response ability modeling sub-module collects the equipment OEE data, extracts the equipment running time, theoretical cycle time and the number of qualified products, calculates the equipment overall efficiency benchmark value as the product of the running time utilization rate and performance, uses cubic polynomial regression to fit the corresponding relationship between the overall efficiency and the material handling volume, and generates a response ability curve equation; The constraint condition activation sub-module calls the response ability curve equation, calculates the zero coordinate of the second derivative function, monitors the real-time value of the current number of material types at the workstation, and when the value reaches the abscissa value corresponding to the zero coordinate, triggers the capacity constraint flag and generates a workstation constraint flag set; The strategy generation sub-module, based on the workstation constraint flag set, calculates the ratio matrix of the remaining capacity of the workstation and the demand feature vector, constructs the objective function of the linear programming model as the maximization of the sum of the product of matrix elements, sets the material type constraint as an inequality equation system, and uses the simplex method to solve the feasible solution of the model to generate a preset strategy set; The goodness-of-fit threshold of the cubic polynomial regression is set to 0.85; The calculation accuracy of the zero point coordinates is reserved to two decimal places; The inequality equation set includes the upper limit constraint of the material type and the non - negative constraint of the station capacity.

[0011] As a further solution of the present invention, the system further includes: A decision - making optimization module, which is used to construct a collaborative unit from the work - station groups with correlation coefficients higher than the critical value in the topological relationship matrix through the Hungarian algorithm, perform cosine similarity matching between the material types in the preset policy set and the equipment parameters of the collaborative unit, trigger a policy reconstruction mechanism to generate a dynamic allocation instruction and synchronously update it to the topological modeling module.

[0012] As a further solution of the present invention, the dynamic allocation instruction is specifically a material distribution path, a work - station cooperation method, and a material - equipment combination plan.

[0013] As a further solution of the present invention, the critical value is the triple weighted sum of the mean and standard deviation of the elements of the topological relationship matrix, where the mean weight is 0.6 and the standard deviation weight is 0.4; The policy reconstruction mechanism adjusts the policy weight coefficient through the gradient descent method, and its learning rate is set as the reciprocal of the difference between the real - time load of the work - station and the maximum capacity of the collaborative unit.

[0014] As a further solution of the present invention, the decision - making optimization module includes: a collaborative unit construction sub - module calls the topological relationship matrix, calculates the mean and standard deviation of the matrix elements, sets the critical value of the correlation coefficient as the triple weighted sum of the mean and standard deviation, constructs a bipartite graph of work - station nodes, traverses the adjacency matrix using the Hungarian algorithm to capture the maximum weight matching, and generates a set of collaborative units; A policy matching sub - module, based on the set of collaborative units, extracts the equipment processing speed and capacity parameters to form a feature vector, calculates the cosine angle value between the material type features in the preset policy set and the equipment parameters, sets the matching threshold as the reciprocal of the ratio of the feature vector norm lengths, filters the policy items with the angle value lower than the threshold, and generates an optimized policy candidate set; A dynamic allocation sub - module calls the optimized policy candidate set, detects the difference between the real - time load of the work - station and the maximum capacity of the collaborative unit, adjusts the policy weight coefficient using the gradient descent method, encodes the updated policy instruction into a JSON - format data packet, and generates a dynamic allocation instruction; The coefficient selection of the triple weighted sum is based on the correlation distribution corresponding to the peak value of the work - station cooperation efficiency in historical data; The calculation formula for the ratio of the feature vector norm lengths is the square root of the sum of the squares of the equipment processing speed and capacity parameters; The iteration termination condition of the gradient descent method is that the change amount of the weight coefficient is less than 0.01 in three consecutive iterations.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In the present invention, through the laser ranging device combined with the dynamic time warping algorithm, the layout coordinates of the workstations are converted into moving time-consuming parameters, and the spatial correlation coefficient is adjusted by the difference ratio, so that the topological relationship matrix can reflect the changes of the physical environment in real time. The dynamic time warping algorithm quantifies the pattern similarity of the demand interval standard deviation and the usage variation coefficient, constructs a multi-dimensional demand feature vector, and improves the matching accuracy between demand prediction and actual consumption of workstations. Based on the OEE data of the equipment, a response capacity curve equation is established, and combined with the linear programming model, the remaining capacity of the workstations is dynamically associated with the demand characteristics, realizing strategy pre-generation and capacity inflection point warning, and optimizing the resource allocation efficiency. The Hungarian algorithm constructs a workstation collaboration unit with a high correlation coefficient, and the cosine similarity matches the material types and equipment parameters, triggers strategy reconstruction and synchronously updates the topological relationship, forming a closed-loop optimization system, and enhancing the dynamic response ability of cross-workstation collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a system flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0018] Please refer to Figure 1 , a material dynamic allocation system based on data pre-set learning for multi-workstations includes: A topology modeling module, a demand clustering module, a strategy generation module, and a decision optimization module.

[0019] The topological modeling module includes: The coordinate acquisition sub-module refers to the laser ranging device. First, the operator deploys a 3D laser scanner with the model number Leica ScanStation P50 at the center of the predetermined scanning area of the workstation, sets the scanning density to 10,000 points per square meter, and the scanning angle range covers 360 degrees horizontally and 270 degrees vertically. Then, the scanning program is started. The laser head emits pulsed laser and receives the reflected signal to obtain the original 3D coordinate data cloud of the surface points of each device within the workpiece processing area, such as robot A, numerically controlled machine tool B, conveyor belt C, etc. For example, for a feature point PA_raw on the base of robot A, the collected original coordinates may be (1005.3, 2010.1, 503.7) millimeters. At the same time, vibrations and temperature changes in the environment may introduce noise, causing the original data points to have a random fluctuation of ±5 millimeters around their true positions. Next, the system calls the Gaussian filtering algorithm to process the collected original coordinate data cloud. Specifically, for each data point, all neighboring points within a radius of 50 millimeters around it are selected. The weights are calculated using the Gaussian function according to the spatial distances between the neighboring points and the central point. The closer the distance, the greater the weight, and vice versa. Then, the coordinates of these neighboring points are weighted and averaged to obtain the filtered coordinates of the central point. For example, after Gaussian filtering of PA_raw and its neighboring points, its coordinates are updated to PA_filtered(1004.9, 2009.8, 503.5) millimeters. This process effectively reduces the environmental noise interference of ±5 millimeters, reducing the standard deviation of the coordinate data from the original 3.5 millimeters to 0.8 millimeters. Subsequently, in order to identify and process abnormal coordinate points, a coordinate offset threshold is set, which is calculated based on the average distance between adjacent nodes. First, the pairwise Euclidean distances between all the identified device nodes within the workstation (such as the centroid points of robot A, machine tool B, and conveyor belt C) are calculated. For example, node A(1000, 2000, 500) millimeters, node B(4000, 2000, 500) millimeters, node C(1000, 5000, 500) millimeters. Then the distance between A and B is 3000 millimeters, the distance between A and C is 3000 millimeters, and the distance between B and C is approximately 4242.6 millimeters. Assuming there are mainly these three nodes in the workstation, adjacent nodes are defined as node pairs with a direct material flow or cooperation relationship. For example, A - B and A - C are adjacent, and their average distance is (3000 + 3000) / 2 = 3000 millimeters (simplified calculation here, and all adjacent pairs will be considered in actual applications). Then the coordinate offset threshold is set to 10% of this average distance, that is , the setting of this threshold refers to the maximum deviation range allowed for the equipment layout accuracy in industry practice. Usually, 10% of the average spacing can ensure the recognition accuracy while avoiding misjudging normal large-span equipment as abnormal. Experimental verification shows that when the threshold is set in the range of 8% - 12% of the average spacing, the accuracy of abnormal point recognition is the highest. 10% is selected as the balance point. Finally, the system traverses the filtered coordinate points. If the deviation of a certain coordinate point from its theoretical position predicted based on the model (or the interpolation position of the surrounding credible points) exceeds 300 millimeters, it is determined as an abnormal coordinate point. For example, a certain collection point Perror(1200, 2300, 850) millimeters, and the compensated coordinate obtained by linear interpolation of its surrounding valid points (such as using its two adjacent valid path points Pprev(1000, 2000, 500) and Pnext(1100, 2100, 550). If Perror should be located at the midpoint of the line connecting the two, the theoretical position is (1050, 2050, 525), and its deviation is much greater than 300 millimeters) is Pcompensated(1050, 2050, 525) millimeters. The system replaces Perror with this compensated coordinate, and this process is iterated until all abnormal points are compensated. Finally, a workstation coordinate set that accurately describes the spatial positions of each equipment node in the workstation is output. For example, this set contains robot A_final(1004.9, 2009.8, 503.5) millimeters, machine tool B_final(3995.2, 2005.1, 501.0) millimeters, etc.

[0020] Table 1 Example of Coordinate Data of Workstation Equipment Nodes

[0021] As shown in Table 1, the data of some equipment nodes after coordinate acquisition and preliminary processing are listed, showing the correction effect of Gaussian filtering on the original coordinate data.

[0022] The moving time-consuming calculation sub-module calls the workstation coordinate set generated in the previous step, for example, the coordinate data containing robot A_final(1004.9, 2009.8, 503.5) millimeters and machine tool B_final(3995.2, 2005.1, 501.0) millimeters. First, the Euclidean distance between adjacent coordinate points is extracted as the theoretical shortest path length for the movement of the equipment (such as an automated guided vehicle AGV). For the movement from A to B, its Euclidean distance is calculated as millimeters, that is millimeters, approximately 2.99 meters. Next, a time reference value is set, which is defined as the ratio of the trajectory length to the average moving speed of the equipment. Assume the average straight-line moving speed of the AGV is 0.5 m / s, then the time reference value , this time reference value represents the time taken under an ideal straight-line path without any detours or speed changes. Subsequently, the time series of the actual or planned movement trajectory (represented in the form of a time series, for example, recording the position coordinates of the AGV every 0.1 seconds) is aligned with the time series of the ideal straight-line trajectory through the dynamic time warping (DTW) algorithm. DTW constructs a cost matrix, where each element represents the -th time point of the actual trajectory and the -th time point of the ideal trajectory represents the position deviation (Euclidean distance) between them, and then finds a minimum cumulative cost path from the lower left corner to the upper right corner of the matrix. This path represents the best alignment of the two time series and can handle the time axis deviation from the ideal uniform straight-line motion caused by possible accelerations, decelerations, or short stops in the actual trajectory. In the process of calculating the deviation coefficient of the trajectory shape difference degree from the time reference value, a specific formula is used to calculate the shape difference degree, that is, the shape difference degree is a dimensionless parameter that quantifies the degree of trajectory bending. For example, by discretizing the actual trajectory into a series of small line segments, calculating the instantaneous curvature at the endpoints of each line segment (approximated by the reciprocal of the circumradius of three consecutive points), then calculating the cumulative change of these curvature values along the path length, and finally normalizing it. Assume that for the actual movement trajectory from A to B, its is calculated (for example, the actual path is 5% longer than the straight-line path and contains two main turns, comprehensively evaluated) to be 0.08 (dimensionless), and the time reference value , then the shape difference degree . This shape difference degree reflects the non-ideal degree of the trajectory. Finally, based on this shape difference degree and the time reference value, the final movement time parameter is obtained. This parameter adds the extra time generated by the trajectory bending to the basic time. The calculation formula is . Substituting the values, . This value is used as an estimated movement time for one time from node A to node B.

[0023] The descriptions of the parameters in formula are as follows: : The finally calculated movement time parameter, in seconds (s), representing the estimated movement time of the device considering non-ideal factors such as path bending. : The time reference value, in seconds (s), defined as the ratio of the theoretical trajectory length to the average movement speed of the device, that is , where is the Euclidean distance (meters), is the equipment speed (m / s). It is the moving time under ideal conditions. : The rate of change of trajectory curvature, a dimensionless parameter that quantifies the bending or complexity of the actual moving trajectory relative to the ideal straight path. The larger the value of this parameter, the more the trajectory deviates from a straight line and the more it bends. It is usually obtained by analyzing the geometric shape of the actual trajectory. For example, the coordinates of a series of points on the path are collected , and the curvature is calculated in segments , and then the change in curvature is calculated, such as and normalized. In this example, by analyzing the historical movement data or planned path of the AGV from point A to point B, the sensor is set to record the (x, y) coordinates of the AGV every 100 milliseconds, obtaining a sequence of path points . For every three consecutive points , the curvature (the reciprocal of the radius) of the circle passing through these three points can be calculated, thus obtaining a series of curvature values. can be the variance of these curvature values, or the average deviation from the straight path (curvature is 0), obtained after normalization. For example, if the average curvature of a path increases by 8% compared to the straight path, then takes 0.08.

[0024] Parameter assignment and calculation process: 1. The coordinates of point A (1004.9, 2009.8, 503.5) mm and point B (3995.2, 2005.1, 501.0) mm are extracted from the workstation coordinate set. 2. Calculate the Euclidean distance between A and B mm m. 3. The average moving speed of the AGV is set to 0.5 m / s (obtained from the equipment manual or actual calibration. For example, through multiple round-trip tests on a 10-meter straight track, the average speed is stable between 0.48 - 0.52 m / s, and 0.5 m / s is taken). 4. Calculate the time reference value . 5. The rate of change of trajectory curvature is set to 0.08 according to the actual path analysis (this value is obtained by analyzing historical trajectory data. It is found that the average turning radius of this path segment is about 5 meters, while the turning radius of the straight path is infinite. This difference is quantified through a specific model. For example, by comprehensively quantifying factors such as the average number of turns and turning angles and comparing with the standard non-turning path, the deviation is obtained as 8%). 6. Substitute into the formula to calculate the moving time-consuming parameter: , , , .

[0025] The topological relationship construction sub-module calculates the time difference ratio between different differential movement paths based on the movement time-consuming parameter, performs row vector normalization on the spatial correlation matrix with the reciprocal of the path length as the weight coefficient, and performs Hadamard product operation on the normalization result and the time difference ratio to generate a topological relationship matrix; The movement time-consuming calculation sub-module calls the set of workstation coordinates generated in the previous step, such as coordinate data including robot A_final(1004.9, 2009.8, 503.5) millimeters and machine tool B_final(3995.2, 2005.1, 501.0) millimeters. First, it extracts the Euclidean distance between adjacent coordinate points as the theoretical shortest path length for the movement of equipment (such as an automatic guided vehicle AGV). For the movement from A to B, its Euclidean distance is calculated as millimeters, that is, millimeters, approximately 2.99 meters. Next, a time reference value is set, which is defined as the ratio of the trajectory length to the average movement speed of the equipment. Assuming the average straight-line movement speed of the AGV is 0.5 m / s, then the time reference value , this time reference value represents the time-consuming under the ideal straight-line path without any detours or speed changes. Subsequently, through the dynamic time warping algorithm (DTW), the time series of the actual or planned movement trajectory (represented in the form of a time series, for example, recording the position coordinates of the AGV every 0.1 second) is aligned with the time series of the ideal straight-line trajectory. DTW constructs a cost matrix, where each element represents the position deviation (Euclidean distance) between the th time point of the actual trajectory and the th time point of the ideal trajectory. Then, a minimum cumulative cost path from the lower left corner to the upper right corner of the matrix is found, and this path represents the best alignment of the two time series, which can handle the time axis deviation from the ideal uniform straight-line motion caused by acceleration, deceleration, or short stops in the actual trajectory. In the process of calculating the deviation coefficient between the trajectory shape difference degree and the time reference value, a specific formula is used to calculate the shape difference degree, that is, the shape difference degree , where the trajectory curvature change rate is a dimensionless parameter that quantifies the degree of trajectory bending. For example, by discretizing the actual trajectory into a series of small line segments, calculating the instantaneous curvature at the endpoints of each line segment (approximated by the reciprocal of the circumradius of three consecutive points), then calculating the cumulative change of these curvature values along the path length, and then performing normalization processing. Assuming for the actual movement trajectory from A to B, its is calculated (for example, the actual path is 5% longer than the straight-line path and contains two main turns, comprehensively evaluated) to be 0.08 (dimensionless), and the time reference value , then the morphological difference degree , this morphological difference degree reflects the non-ideal degree of the trajectory. Finally, based on this morphological difference degree and the time reference value, the final moving time-consuming parameter is obtained , this parameter adds extra time caused by trajectory bending on the basis of the basic time-consuming, and the calculation formula is , substitute the values , this value is used as an estimated moving time-consuming between node A and node B.

[0026] Formula The descriptions of each parameter in the formula are as follows: : The finally calculated moving time-consuming parameter, with the unit of second (s), representing the estimated moving time of the device after considering non-ideal factors such as path bending. : The time reference value, with the unit of second (s), defined as the ratio of the theoretical trajectory length to the average moving speed of the device, that is , where is the Euclidean distance (meter), is the device speed (meter / second). It is the moving time under ideal conditions. : The trajectory curvature change rate, a dimensionless parameter, which quantifies the bending or complexity of the actual moving trajectory relative to the ideal straight path. The larger the value of this parameter, the more the trajectory deviates from the straight line and the more bending. The way to obtain it is usually to analyze the geometric shape of the actual trajectory. For example, collect the coordinates of a series of points on the path , calculate the curvature in segments , and then calculate the change of the curvature, such as and perform normalization processing. In this example, by analyzing the historical moving data or planned path of the AGV from point A to point B, set the sensor to record the (x, y) coordinates of the AGV every 100 milliseconds, and obtain the path point sequence . For every three consecutive points , the curvature (the reciprocal of the radius) of the circle passing through these three points can be calculated, so as to obtain a series of curvature values. can be the variance of these curvature values, or the average deviation from the straight path (curvature is 0), and is obtained after normalization processing. For example, if the average curvature of a path increases by 8% compared with the straight path, then takes 0.08.

[0027] Parameter assignment and calculation process: 1. Extract the coordinates of point A (1004.9, 2009.8, 503.5) mm and the coordinates of point B (3995.2, 2005.1, 501.0) mm from the workstation coordinate set. 2. Calculate the Euclidean distance between A and B mm m. 3. The average moving speed of the AGV Set to 0.5 m / s (obtained from the equipment manual or through actual calibration. For example, during multiple round-trip tests on a 10-meter straight track, the average speed is stable between 0.48 - 0.52 m / s, and 0.5 m / s is taken). 4. Calculate the time reference value . 5. Track curvature change rate Set to 0.08 according to the actual path analysis (this value is obtained by analyzing historical track data. It is found that the average turning radius of this path segment is about 5 meters, while the turning radius of a straight path is infinite. This difference is quantified through a specific model. For example, after comprehensively quantifying factors such as the average number of turns and turning angles and comparing with the standard non-turning path, the deviation is found to be 8%). 6. Substitute into the formula to calculate the moving time-consuming parameter: , , , ; Row vector normalization is achieved by calculating the Euclidean length of each row element and performing a division operation.

[0028] The demand clustering module includes: The demand feature extraction sub-module calls the demand time-series data recorded in the production line manufacturing execution system (MES) for the cutting material at Station A (for example, steel plate SPCC, thickness 2 mm). This data details the timestamp and demand quantity of each demand initiation in the past several production cycles. First, the system extracts the time intervals between adjacent demand events from these time-series records to form a time interval sequence , for example, within the most recent shift (8 hours), the demand timestamps for this steel plate at Station A are 08:05, 08:32, 09:10, 09:40, 10:25 respectively, and the corresponding time intervals are 27 minutes, 38 minutes, 30 minutes, 45 minutes, forming the sequence minutes. Next, calculate the standard deviation of this time interval sequence as the "time fluctuation amount" to measure the uncertainty of demand time points , for the sequence , its mean minutes, and the standard deviation minutes. Then, the system collects the specific usage data of this material each time it is demanded at Station A within the same time period to form a usage sequence , for example, the corresponding four demand usages are 10 pieces, 12 pieces, 8 pieces, 15 pieces, that is pieces. Subsequently, calculate the coefficient of variation (CV) of this usage sequence as the "usage dispersion amount" to measure the uncertainty of demand quantity , the coefficient of variation is defined as the standard deviation divided by the mean. For the sequence , its mean pieces, and its standard deviation pieces. Therefore, the usage dispersion amount (Dimensionless). Finally, the calculated "time fluctuation amount" in minutes and "usage discreteness amount" are subjected to Z-score standardization. The calculation benchmark for standardization is derived from the mean and standard deviation obtained by statistically analyzing the and values calculated daily within the most recent 30 production cycles (e.g., the most recent 30 days). Assume that the mean of the time fluctuation amount obtained from the statistical analysis of the data in the past 30 cycles is minutes, and the standard deviation is minutes; the mean of the usage discreteness amount is , and the standard deviation is . The setting of these benchmark values is based on long-term tracking and statistical analysis of historical data, ensuring their representativeness and stability. Selecting 30 production cycles takes into account the seasonal fluctuations in production and monthly plan adjustments, and can better reflect the recent normal level. Then, the standardized time fluctuation amount for the current cycle is , and the standardized usage discreteness amount is . These two standardized values form a two-dimensional demand feature vector , which concisely represents the volatility and discreteness characteristics of the current material demand.

[0029] Table 2 Example data for calculating demand characteristics

[0030] As shown in Table 2, it demonstrates the calculation process and intermediate data of the demand characteristics of a certain material at Station A, and finally obtains the standardized demand feature vector.

[0031] The pattern similarity calculation sub-module is based on a specific demand generated in the previous step (e.g., the steel plate demand at Station A, whose feature vector is ) or its original demand sequence, and calls other historical or different types of demand patterns (e.g., the aluminum material demand sequence at Station B, or the high-volatility demand pattern sequence in the history of Station A) for comparison. First, the dynamic time warping (DTW) algorithm is used to process the two differential demand sequences to be compared. For example, the daily steel plate demand sequence at the current Station A is (where is the demand at the t-th time unit, such as the demand per hour), and a historical typical demand pattern sequence is . The DTW algorithm non-linearly stretches or compresses the time axes of these two sequences to find the best alignment path between them. Specifically, when implementing, a distance matrix is constructed, where the element has a value of (usually p = 1 or 2), and then the dynamic programming idea is used to find a path from to The path with the smallest sum of elements on this path is the minimum bending path. Next, calculate the cumulative distance value on this minimum bending path, denoted as , which quantifies the morphological difference degree between two demand sequences. Even if they are misaligned in time or have different rates, assume that the cumulative distance value between the current demand sequence at Station A and the historical high-fluctuation pattern P1 is obtained through DTW calculation (The unit is related to the demand unit or has been normalized to dimensionless). Subsequently, map this calculated cumulative distance value through a preset similarity conversion function to obtain the final pattern similarity coefficient , and this conversion function is set as an S-shaped curve function, and the specific form can be , where is the input DTW distance value, is the slope parameter, is the distance offset (or called the center point), which determines the conversion center position of the S-shaped curve. According to the description, the slope parameter is inversely proportional to the material allocation response time threshold , that is , where is a positive proportional constant. The material allocation response time threshold refers to the maximum time allowed from the generation of demand to the delivery of materials. For example, it is set to 30 minutes. The setting of this threshold is based on a comprehensive consideration of the production rhythm and the material waiting cost. Through experimental analysis, when the response time exceeds 30 minutes, the average waiting time loss at the station increases significantly, so it is set to 30 minutes. The value of the proportional constant can be calibrated through historical data so that the similarity coefficient can effectively distinguish different patterns. For example, take , then (if the unit of D is related to minutes), if D is a dimensionless DTW distance, the unit of k is adjusted accordingly, or D is first normalized. Assume (dimensionless, obtained through calibration so that the DTW distance can produce better similarity discrimination in the range of 20 - 80), (the center point of the S-shaped curve, representing the DTW distance of general differences) is set to 50. This value is selected near the mean of the pairwise DTW distances of a large number of historical demand sequence pairs. For example, the average DTW distance is statistically obtained as 52, then can be taken as 50. Then for , the pattern similarity coefficient , and this 0.6035 is the similarity coefficient between the current demand pattern at Station A and the historical high-fluctuation pattern P1.

[0032] The matrix operation sub-module calls the pattern similarity coefficient calculated in the previous step (for example, the similarity coefficient vectors obtained by calculating the current demand pattern of station X with K historical typical demand patterns P1, P2, …, PK respectively , or, if calculating the similarity of demand patterns between stations, a pattern similarity matrix between stations is obtained , where the element is the similarity coefficient of the demand patterns of station i and station j), and the topological relationship matrix generated in the topological modeling module (whose element reflects the topological connection strength between station i and station j). First, the system needs to construct a "coefficient matrix" and a topological matrix for operation. Here, the "coefficient matrix" is closely related to the "pattern similarity coefficient". Suppose we consider the characteristics of the demand patterns of each station itself relative to a global benchmark, or the demand collaboration between stations. If the "pattern similarity coefficient" is a single value (such as the similarity between the overall demand pattern of the current production batch and a certain standard pattern), then this coefficient may be used to scale the entire topological relationship matrix or some of its elements. However, more commonly, we need a "coefficient matrix" with the same dimension as . For example, can represent the demand pattern synchronization or complementarity score between station i and station j, which is jointly determined by their respective "pattern similarity coefficients" (similarity relative to a certain benchmark pattern) and their interaction relationship. In the description of "constructing the Cartesian product space of the coefficient matrix and the topological matrix and performing the product operation of matrix elements at corresponding positions", the "Cartesian product space" may refer to a conceptually combined space, and the actual operation of "performing the product operation of matrix elements at corresponding positions" clearly refers to the Hadamard product, provided that and have the same dimension (N is the number of stations). Let be the interaction coefficient calculated based on the pattern similarity coefficients of stations i and j. For example, if the demand pattern of station i has a high similarity with a high-fluctuation pattern, while the demand pattern of station j has a high similarity with a stable pattern, their interaction coefficient may be low. On the contrary, if both are similar to a high-collaboration demand pattern, the interaction coefficient is high. Let's take a specific example. Suppose is directly the demand pattern similarity between station i and station j , and is the topological connection strength between them, then the product matrix is obtained by multiplying the elements at corresponding positions , and , its elements , for example, if the demand pattern similarity between station 1 and station 2 , the topological relationship value between them (this value comes from the calculation result in paragraph 3, for example, it is a representative value after a certain scaling or selection), then , this product more comprehensively reflects the comprehensive relevance between stations i and j in terms of demand pattern and physical topology. Regarding "the Cartesian product space improves the material distribution efficiency by increasing the weight factor of the moving time between stations", this usually means itself already contains the weight factor related to the moving time (as described in paragraph 3, its construction process considers the moving time), so naturally inherits this as well. Subsequently, sum the result of this product matrix "by row vector", which means for each row of , calculate the sum of all its elements , this will generate a scalar value for each station , representing the overall correlation strength of this station with all other stations at the demand-topology level. For example, if the first row of is , then the row sum of station 1 is . Finally, according to the description "generate the demand-topology correlation matrix", if this "demand-topology correlation matrix" is a matrix, then it may be itself, and the result of "summing by row vector" is a further analysis or aggregation step based on . If refers to , it directly quantifies the coupling effect of the demand pattern similarity and topological connectivity between each pair of stations.

[0033] The Z-score normalization process uses the data of the most recent 30 production cycles as the calculation benchmark; The similarity conversion function is an S-shaped curve function, and its slope parameter is inversely proportional to the material allocation response time threshold; The Cartesian product space improves the material distribution efficiency by increasing the weight factor of the moving time between stations.

[0034] The strategy generation module includes: The response ability modeling sub-module collects the historical Overall Equipment Effectiveness (OEE) data of a specified workstation (e.g., the welding workstation W1). These data are usually automatically recorded by the MES system or the equipment monitoring system and contain the operating performance of the equipment over a period of time. First, key parameters are extracted from these OEE data records, including the actual operating time during the planned operating time of the equipment (e.g., in an 8-hour shift, i.e., 480 minutes, due to reasons such as waiting for materials and changeovers, the actual operating time of the equipment is 400 minutes), the theoretical cycle time per unit product (e.g., the theoretical fastest time to weld a standard part is 0.8 minutes per piece, and this data is provided by the process design or equipment manufacturer), and the number of qualified products produced during this actual operating time (e.g., 450 qualified products are produced within 400 minutes). Next, an "Overall Equipment Effectiveness benchmark value" is calculated , which is defined as the product of the "operating time utilization rate" and "performance". Here, the "operating time utilization rate" is defined as the actual operating time divided by the total planned production time (e.g., the shift time), that is , and "performance" is defined here as (number of qualified products × theoretical cycle time) / actual operating time, that is . Therefore, the Overall Equipment Effectiveness benchmark value . This benchmark value is a snapshot of the efficiency performance under specific periods and specific material handling volumes. Subsequently, to more comprehensively describe the relationship between efficiency and output (material handling volume), the system uses the cubic polynomial regression method to fit multiple groups of historical data points collected (each group of data contains the "material handling volume" and the corresponding "overall efficiency" . Here, the overall efficiency is calculated as described above ). The goal is to establish a mathematical model in the form of , where represents the material handling volume (e.g., the number of workpieces processed per hour), represents the overall efficiency (a value between 0 and 1), and the coefficients are estimated by the least squares method. For example, by analyzing the data of the past 20 shifts, 20 data points are obtained, and the fitted equation may be , where is in units of pieces per hour, is the efficiency value. The goodness of fit is evaluated by the coefficient of determination R2, and the threshold of R2 is set to 0.85, that is, it is required that , the setting of this threshold is based on industry experience and multiple simulation experiments. Experiments show that when R2 is lower than 0.85, the prediction error of the model is large and it cannot accurately guide production, while the model with R2 higher than 0.85 can better reflect the actual trend and ensure the reliability of the model. For example, the R2 obtained from this fitting is 0.89, which meets the requirements. Finally, this cubic polynomial equation is the response ability curve equation of this station generated.

[0035] Formula (simplified to ) The descriptions of each parameter are as follows: : The benchmark value of overall equipment effectiveness, dimensionless, usually between 0 and 1, indicating the overall performance of the equipment under specific conditions. : The actual operating time, in minutes or hours, refers to the time that the equipment is actually used for production during the planned production time. : The total planned production time, with the same unit as , such as the total duration of a shift. : The number of qualified products, in pieces or other counting units, refers to the number of products that meet the quality requirements produced within time. : The theoretical cycle time per unit product, in minutes / piece or hours / piece, refers to the shortest time required to produce a qualified product under ideal conditions.

[0036] Parameter assignment and calculation process: 1. Collect data from the MES system: The total planned production time minutes. 2. The actual operating time minutes. 3. The number of qualified products pieces. 4. The theoretical cycle time per unit product minutes / piece (this data is provided by the equipment supplier and verified through tests under small-batch ideal conditions. For example, when continuously producing 10 pieces, the shortest total time is 7.9 minutes, with an average of 0.79 minutes per piece, rounded up to 0.8 minutes). 5. Substitute into the formula to calculate the benchmark value of overall equipment effectiveness: , , , .

[0037] The constraint condition activation sub-module calls the response ability curve equation generated in the previous step. For example, the response ability curve of the welding station W1 is , where is the overall efficiency, is the material handling capacity (pieces per hour). First, the system calculates the second derivative function of this response capacity curve equation and solves for the points where the second derivative is zero (i.e., the inflection points of the curve), which typically correspond to the turning points where the efficiency growth rate changes significantly. Mathematically, , then the second derivative , let , that is , and the solution is pieces per hour. The calculation result is rounded to two decimal places as required, that is pieces per hour. This value is interpreted in the current context as a critical point of the material type handling capacity of the workstation or a boundary point of the optimal efficiency range (although usually the direct correlation between the inflection point and the optimal efficiency point requires more context, here it is carried out according to the logic described in the patent). Next, the system monitors in real time the number of material types currently being processed at workstation W1 , for example, by reading the production order or bill of materials information associated with this workstation, it is known that workstation W1 currently needs to process 3 different specifications but similar process parts, that is . Then, the system compares the number of material types monitored in real time with the horizontal axis value corresponding to the zero point of the previously calculated second derivative (here the patent description compares the material handling capacity with the number of material types, which may mean that does not represent the inflection point of the processing rate, but a certain quantity related to the processing capacity, or there is a certain correspondence or conversion relationship between the "number of material types" here and the "material handling capacity" in the response capacity curve in a specific scenario. Assume that is a threshold related to the "carrying capacity of the number of material types"), and when the real-time number of material types reaches or exceeds this horizontal axis value , the system triggers a capacity constraint flag, indicating that the workstation has reached or is approaching a critical point in terms of handling material diversity in its design or experience, which may lead to a decrease in efficiency or an increase in management complexity. In this example, , while , because , the capacity constraint flag is not triggered. If it is assumed that refers to a threshold related to the upper limit of the number of material types that the workstation can handle efficiently, such as represents an "equivalent maximum material type processing index" after calibration, and the current is also converted into a comparable index, or simply, if is directly the maximum allowable number of material types, for example , and the current , because , the capacity constraint flag of station W1 is set to "activated" or "True". Finally, the system aggregates the constraint flag status of all stations to generate a set of station constraint flags. For example, for a station containing three stations W1, W2, and W3, its set of constraint flags may be {W1: True, W2: False, W3: True}, indicating that the constraints have been triggered for stations W1 and W3 in terms of material type processing.

[0038] Based on the set of station constraint flags obtained in the previous step (e.g., {W1: True, W2: False, W3: True}, where the material type processing of W1 and W3 has reached the constraints), and the specific operating parameters of each station, the strategy generation sub-module first calculates the ratio matrix between the "remaining capacity of the station" and the "demand feature vector". Here, the "remaining capacity of the station" For station , it can refer to the margin after subtracting the assigned tasks from its effective production capacity under the current constraint conditions. For example, if station W2 has not triggered the constraint, its theoretical maximum production capacity is 100 units / hour, and the assigned tasks occupy 60 units / hour, then the remaining capacity units / hour. For station W1 that has triggered the constraint, the calculation of its remaining capacity may need to consider the efficiency loss caused by the constraint or be directly set to a lower value. For example units / hour, the "demand feature vector" comes from the demand clustering module (such as in paragraph 4), which describes the characteristics of the demand to be assigned , such as the volatility and discreteness of the demand quantity. Suppose we have 3 stations W1, W2, W3 and 2 demands D1, D2 to be processed, and their demand feature vectors are respectively and . We need to convert the demand feature vector into a "demand quantity" or "load index" that can be compared with the capacity. For example, we can take its modulus length or weighted sum. Let be the equivalent load of demand , then the element of the ratio matrix (or some form of matching degree score, such as ). For example, if units, units, then , . This ratio matrix reflects the relative abundance or suitability of each station to process each demand. Next, the system constructs a linear programming model, and its objective function is set to maximize the sum of the products of these ratios (or matching degree scores) and the decision variables (indicating whether to assign the demand to the station), that is , where is a binary decision variable. If demand is assigned to work station , then , otherwise 0. Then, the constraints of the model are set. These constraints are given in the form of a system of inequality equations, mainly including "material type upper limit constraint" and "work station capacity non - negative constraint". The material type upper limit constraint means that the total number of different material types (or demand types) assigned to each work station cannot exceed the upper limit of types preset for this work station or determined by the constraint - condition activation sub - module , that is (here judges whether demand j is a new type on work station i), or if represents the allocation quantity, then , where is an indicator function. The work station capacity non - negative constraint ensures that the total load assigned to each work station does not exceed its remaining effective capacity , and all allocation quantities (if representing specific quantities) need to be non - negative. Finally, the simplex method (Simplex Method) is used to solve the constructed linear programming model. The simplex method finds the optimal solution by iteratively moving on the vertices of the feasible solution domain, and each step moves in the direction of improving the objective function value until a combination that satisfies all constraints and maximizes the objective function is found. For example, after solving, it may obtain , indicating that allocating both demand D1 and D2 to work station W2 is the current optimal (or feasible) strategy. Finally, all calculated feasible solutions (which may include the optimal solution and some sub - optimal solutions, or the Pareto solution set under multiple objectives) are sorted out and output to form a preset strategy set. For example, Strategy A: {D1 -> W2, D2 -> W2}, Strategy B: {D1 -> W2, D2 -> W3} (if W3 also has a certain capacity).

[0039] The decision - making optimization module includes: The collaborative unit construction sub - module calls the topological relationship matrix generated in the previous steps (for example, a matrix, where the element represents the topological association strength from work station i to work station j, and the larger the value, the stronger the association. Assume the matrix is ). First, calculate the mean and standard deviation of all non - diagonal elements in this matrix. For the above , the non - diagonal elements are 0.8, 0.3, 0.8, 0.6, 0.3, 0.6, and its mean , standard deviation ; Next, set a "correlation coefficient threshold" , which is defined as the triple weighted sum of the mean and the standard deviation, i.e., , where the weight coefficient , this coefficient is selected based on the analysis of historical data. It is found that when the collaborative efficiency between workstations reaches its peak, the corresponding topological correlation values mostly distribute in and above regions. For example, collect the tasks involving the collaborative operation of two workstations in the past year, count their completion efficiency, and perform a correlation analysis with the value between these two workstations. It is found that when , the probability of the collaborative efficiency increasing by more than 15% significantly increases. Take as a conservative and effective threshold, then . Since the maximum value in is 0.8, this value is too high, indicating that there may be no strong correlation in the current sample data. To continue the example, we adjust the assumption. For example, assume that calculated by another method, then . Subsequently, based on this threshold, construct a workstation node graph (not specifically a bipartite graph, but an ordinary graph, where nodes represent workstations and edges represent the association strength between workstations). If , then form an edge between workstation i and workstation j, and the weight of the edge can be itself. For example, for and , then there is an edge with a weight of 0.8 between workstation 1 and workstation 2, and other direct connections do not meet this condition (both 0.3 and 0.6 are less than 0.7). Therefore, a graph containing only the 1-2 connection is obtained. Then, the system uses the Hungarian algorithm (Hungarian algorithm) to "traverse the adjacency matrix to capture the maximum weight matching". The Hungarian algorithm is usually used to solve the maximum weight matching problem of a weighted bipartite graph. If the goal here is to form pairwise collaborative units (for example, pair N workstations into N / 2 pairs), a bipartite graph can be constructed by duplicating the workstation set, where the nodes in one part are the original workstations, and the nodes in the other part are also the original workstations. The weight of the edge is (if and ) Then, use the Hungarian algorithm to find the matching that maximizes the total weight. Or, if the "collaborative units" are not limited to pairs but small-scale collaborative clusters, other community discovery algorithms may be adopted. If strictly following the "Hungarian algorithm" and "maximum weight matching", we assume that here it is to find the optimal workstation pairing combination. Even if the number of workstations is odd, it can be handled by introducing virtual nodes. For example, for workstations 1, 2, and 3, if we want to form an optimal pairing and a single independent workstation, assume we construct a weighted adjacency matrix , where if else 0, and apply the Hungarian algorithm (or its variants such as the auction algorithm for non-bipartite graph matching) to find several disjoint edge sets that maximize the sum of the weights of the matching edges. In this example, only the 1-2 connection meets the condition with a weight of 0.8. Then, a collaborative unit is identified as {W1, W2}, while W3 is alone. Finally, the set of collaborative units is output, for example: {{W1, W2}, {W3}}.

[0040] Formula The explanations of each parameter in it are as follows: : The critical value of the correlation coefficient. This is a threshold used to determine whether the element value in the topological relationship matrix is high enough to indicate that there is significant collaborative potential between two workstations. : The topological relationship matrix The mean of the (usually non-diagonal) elements in it. : The topological relationship matrix The standard deviation of the (usually non-diagonal) elements in it. : The weight coefficient, set to 3 here. It determines the magnification factor of the standard deviation when calculating the critical value, which usually corresponds to the "three-sigma" criterion in statistics and can cover most data.

[0041] Parameter assignment and calculation process (using adjusted hypothetical data to make the process meaningful): 1. The topological relationship matrix . 2. Calculate the mean of the non-diagonal elements . 3. Calculate the standard deviation of the non-diagonal elements . 4. Set the weight coefficient . The basis for choosing this coefficient: Historical data shows that when the topological correlation coefficient between workstations exceeds , the average efficiency of the collaborative units composed of them in performing complex tasks is 20% - 25% higher than that of the units below this threshold, and the failure rate is reduced by 10%. For example, 100 groups of historical collaborative task data are selected, corresponding to the values between workstations. It is found that in the top 5% of the values (roughly corresponding to The pairing within the above intervals) has significantly improved both the collaborative success rate and efficiency metrics. 5. Substitute into the formula to calculate the critical value of the correlation coefficient: , , , , (as mentioned before, this result is too high for the example matrix. If you want to continue, assume is adjusted to 0.7).

[0042] The policy matching sub-module is based on the set of collaborative units generated in the previous step, such as {{W1, W2}, {W3}}, and conducts matching evaluations for each unit (or the devices within the unit) and each policy in the "preset policy set". First, extract the key "device processing speed" and "capacity parameter" from the collaborative unit or individual device to form a device feature vector , for example, for the collaborative unit {W1, W2}, its combined processing speed may be the sum of the two or a certain weighted average, such as 15 pieces per minute, and the combined capacity is 80 pieces (batch capacity or cache capacity), then its feature vector , for the individual work station W3, its feature vector may be . Next, the system analyzes each policy in the "preset policy set". A policy usually involves allocating a certain material (or task) to a certain device or collaborative unit, and a "material type feature" vector needs to be extracted for this material (or task) , and this vector should be able to reflect the requirements of the material for the device. For example, a certain material M1 requires a relatively high processing speed and medium batch processing capacity, and its features can be quantified as . Then, calculate the cosine angle value (i.e., cosine similarity) between the material feature vector and the device feature vector , , for example, for the material M1 ( ) and the collaborative unit {W1, W2} ( ): , , , , and this value is very close to 1, indicating that the directions of the two feature vectors are highly consistent. Subsequently, set a matching threshold , and this threshold is described as "the reciprocal of the ratio of the feature vector norms". According to "the formula for calculating the ratio of the feature vector norms is the square root of the sum of the squares of the device processing speed and the capacity parameter", this refers to the calculation of a single norm, that is . For "the reciprocal of the ratio of the norms", we interpret it as to ensure that the threshold is within the [0, 1] interval so that it can be compared with the cosine similarity. For example, , , then , the logic for setting this threshold is that when the magnitudes of two vectors (representing the overall "scale" or "ability") differ significantly, even if their directions are the same, the matching degree should be affected. The closer the magnitude ratio is to 1, the higher the threshold, and a more consistent direction is required. Conversely, when the magnitude difference is large, the threshold decreases, and the requirement for direction consistency is slightly relaxed. Finally, filter out those "policy items with an included angle value lower than the threshold" in the preset policy. If the "included angle value" refers to itself, and the "threshold" refers to an angular threshold (for example, obtained by inverse cosine), then filter the policies, which is equivalent to , if the "included angle value" refers to , and "lower than the threshold" is taken literally, that is, filter the policies, which will select items with a poor matching degree. This may be meaningful in some scenarios (such as finding differential supplementary policies), but usually optimization aims to find high-matching ones. Here, we follow the literal "included angle value ( ) lower than the threshold ( )", then for the above example, , , because , so this policy item will not be filtered out (if following this literal logic). If it is understood as filtering for high-quality matches, then , this policy will be selected. Considering the goal of "optimized policy candidate set", the latter understanding is more common, but strictly following the literal meaning of the question, assuming it is filtering , then only those policies with a low enough cosine similarity (lower than the threshold determined by the magnitude ratio) will be retained. For example, if another policy calculates and its corresponding , then because , this policy item will be selected into the optimized policy candidate set. This set contains those policies that are preliminarily filtered according to the current matching criteria (which may be to find specific differences or meet specific non-matching conditions).

[0043] The dynamic allocation sub-module calls the optimized policy candidate set generated by the previous step of filtering (for example, containing policies S1’, S2’, S3’, which may be selected because the cosine similarity of their device and material characteristics is "lower than" a dynamic threshold determined by the reciprocal of the magnitude ratio, suggesting that they may represent some options that require special attention, are not highly matched but still valuable, or, if the filtering logic is to take "higher than" the threshold, then they are high-quality matches). First, the system detects the real-time load of the current workstation (or collaborative unit) and its maximum capacity determined during the previous collaborative unit construction stage The difference (e.g., the maximum collaborative capacity of the collaborative unit {W1, W2} is 100 units per hour) , this difference reflects the current idle capacity or overloading situation of the collaborative unit. For example, if the current load of the unit {W1, W2} is 90 units per hour, then units per hour (idle), if the load is 110, then (overloaded). Next, the system uses the gradient descent method to adjust the weight coefficients of each strategy in the optimization strategy candidate set (each strategy has an initial weight, such as evenly distributed or an initial value based on previous evaluations, e.g., ), the goal of gradient descent is to optimize a certain performance metric function , this function may be related to improving overall output, reducing waiting time, or balancing the loads of each unit, and will be affected by (e.g., preferentially reducing the strategy weight of overloaded units, or increasing the strategy weight of idle units). Specifically, calculate the partial derivative of the performance metric with respect to each strategy weight , and then update the weight along the negative gradient direction: , where is the learning rate (e.g., , this value is set through experiments, tested in the range of 0.01 to 0.1, and 0.05 can converge well within 100 iterations and is not prone to oscillation). The iterative process of gradient descent will continue until the termination condition is met. The termination condition is set as: in three consecutive iterations, the change amount (absolute value) of all strategy weight coefficients is less than 0.01, that is, for any strategy , , , and . This 0.01 threshold is set based on the sensitivity analysis of the impact of weight changes on the final scheduling instructions. When the change is less than this value, it is considered that the actual benefit improvement brought by further iteration is not obvious. For example, after several iterations, the weight is updated to And in the subsequent three iterations, if the change in each weight is less than 0.01, the iteration stops. Finally, these optimized and updated strategies (now with a new weight distribution, where some strategy weights may have increased significantly and become dominant strategies) and their related execution parameters (such as which workstation to allocate to, the specific quantity, priority, etc.) are encoded into a data packet in JSON (JavaScript Object Notation) format. JSON format is chosen because of its lightweight nature and ease of machine parsing and human reading. For example, a deployment instruction may be encoded as: {"command_id":"CMD001","timestamp":"2025-05-16T14:35:00Z","target_unit":"Unit_W1_W2","strategy_selected":"S1'_optimal","tasks":[{"material_id":"M005","quantity":50,"priority":1},{"material_id":"M008","quantity":30,"priority":2}],"estimated_completion_time":"2025-05-16T15:05:00Z"}. This JSON data packet is the finally generated dynamic deployment instruction and will be sent to the execution system (such as the AGV scheduling system, workstation control system) to guide the actual material deployment and production activities.

[0044] The coefficient selection of the triple weighted sum is based on the correlation distribution corresponding to the peak of the cooperation efficiency of workstations in historical data; The calculation formula for the ratio of the feature vector norm lengths is the square root of the sum of the squares of the equipment processing speed and capacity parameters; The iteration termination condition of the gradient descent method is that the change in the weight coefficient is less than 0.01 in three consecutive iterations.

[0045] The above are only the preferred embodiments of the present invention and do not limit the present invention in other forms. Any person skilled in the relevant art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution content of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A material dynamic allocation system based on data preset learning for a multi-station, characterized in that, The system includes: A topology modeling module, which is used to obtain the workstation layout coordinate set through a laser ranging device, input the coordinate set into the dynamic time warping algorithm to calculate the moving time-consuming parameter, adjust the spatial correlation coefficient based on the product of the difference ratio of the time-consuming parameter and the equipment moving acceleration, generate a topology relationship matrix and transmit it to the demand clustering module; A demand clustering module, which is used to call the material demand time series record of the production line MES system, input the demand interval standard deviation and the single usage coefficient of variation into the dynamic time warping algorithm to calculate the pattern similarity, generate a demand feature vector, perform a Cartesian product operation with the topology relationship matrix, and then transmit it to the policy generation module; A policy generation module, which is used to establish a response ability curve equation based on the equipment OEE data, activate the constraint condition when the number of material types allocated to the workstation reaches the inflection point of the curve equation, input the ratio of the demand feature vector to the remaining capacity of the workstation into the linear programming model to generate a preset policy set, and transmit it to the decision optimization module.

2. The material dynamic allocation system based on data preset learning for a multi-station according to claim 1, wherein The topology relationship matrix includes the spatial correlation coefficient, the workstation layout structure, and the coordinate spatial position relationship. The demand feature vector includes the demand pattern similarity, the demand interval regularity, and the material usage characteristics. The preset policy set is specifically a material configuration policy, a capacity matching plan, and a workstation load balancing plan.

3. The material dynamic allocation system based on data preset learning for a multi-station according to claim 2, wherein The dynamic time warping algorithm realizes the dynamic correction of the spatial correlation coefficient by matching the morphological difference degree of the moving trajectory and the deviation coefficient of the time reference value; The dynamic time warping algorithm uses a constraint condition with a window width twice the demand interval standard deviation for sequence alignment; The response ability curve equation fits the corresponding relationship between the equipment comprehensive efficiency and the material handling volume through cubic polynomial regression, and the regression fitting goodness-of-fit threshold is set to 0.

85.

4. The material dynamic allocation system based on data preset learning for a multi-station according to claim 3, characterized in that, The topology modeling module includes: A coordinate acquisition sub-module scans the workstation layout through a laser ranging device, acquires the three-dimensional coordinate data of multiple equipment nodes, uses the Gaussian filtering algorithm to eliminate environmental noise interference, sets the coordinate offset threshold as 10% of the average distance between adjacent nodes as the reference value, performs interpolation compensation on abnormal coordinate points, and generates a workstation coordinate set; The moving time-consuming calculation sub-module calls the workstation coordinate set, extracts the Euclidean distance between adjacent coordinate points, uses the ratio of the equipment moving speed to the trajectory length as the time reference value, aligns the time series of the moving trajectory through the dynamic time warping algorithm, and when calculating the deviation coefficient of the trajectory morphological difference degree and the time reference value, uses the morphological difference degree = the trajectory curvature change rate × the square of the time reference value to obtain the moving time-consuming parameter; The topology relationship construction sub-module calculates the time difference ratio between different moving paths based on the moving time-consuming parameter, uses the reciprocal of the path length as the weight coefficient to perform row vector normalization processing on the spatial correlation matrix, and performs a Hadamard product operation on the normalization result and the time difference ratio to generate a topology relationship matrix; The row vector normalization processing is realized by calculating the Euclidean length of each row element and performing a division operation.

5. The material dynamic allocation system based on data preset learning for a multi-station according to claim 4, characterized in that, The demand clustering module includes: a demand feature extraction sub-module that calls the material demand time series record of the production line MES system, extracts the time interval sequence of adjacent demand events, calculates the standard deviation of the sequence as the time fluctuation amount, collects the single material usage data, calculates the coefficient of variation as the usage dispersion amount, performs Z-score standardization processing on the time fluctuation amount and the usage dispersion amount, and generates a demand feature vector; A pattern similarity calculation sub-module, based on the demand feature vector, uses the dynamic time warping algorithm to stretch and compress the time axis of the differentiated demand sequences, calculates the cumulative distance value of the minimum bending path between the sequences, and maps the distance value to a preset similarity conversion function to obtain a pattern similarity coefficient; A matrix operation sub-module calls the pattern similarity coefficient and the topological relationship matrix, constructs the Cartesian product space of the coefficient matrix and the topological matrix, performs the product operation of the corresponding positions of the matrix elements, sums the product results by row vectors, and generates a demand-topology association matrix; The Z-score standardization processing uses the data of the last 30 production cycles as the calculation benchmark; The similarity conversion function is an S-shaped curve function, and its slope parameter is inversely proportional to the material allocation response time threshold; The Cartesian product space improves the material distribution efficiency by increasing the weight factor of the moving time between workstations.

6. The material dynamic allocation system based on data preset learning for a multi-station according to claim 5, characterized in that The strategy generation module includes: a response ability modeling sub-module that collects equipment OEE data, extracts the equipment running time, theoretical cycle time, and the number of qualified products, calculates the equipment comprehensive efficiency benchmark value as the product of the running time utilization rate and performance, uses cubic polynomial regression to fit the corresponding relationship between the comprehensive efficiency and the material handling volume, and generates a response ability curve equation; A constraint condition activation sub-module calls the response ability curve equation, calculates the zero coordinate of the second derivative function, monitors the real-time value of the current number of material types at the workstation, and when the value reaches the abscissa value corresponding to the zero coordinate, triggers a capacity constraint flag and generates a workstation constraint flag set; A strategy generation sub-module, based on the workstation constraint flag set, calculates the ratio matrix of the remaining capacity of the workstation and the demand feature vector, constructs the objective function of the linear programming model as the maximization of the sum of the product of the matrix elements, sets the material type constraint as an inequality equation system, and uses the simplex method to solve the feasible solution of the model to generate a preset strategy set; The goodness-of-fit threshold of the cubic polynomial regression is set to 0.85; The calculation accuracy of the zero coordinate is reserved to two decimal places; The inequality equation system includes the upper limit constraint of the material type and the non-negativity constraint of the workstation capacity.

7. The material dynamic allocation system based on data preset learning for a multi-station according to claim 6, wherein The system further includes: A decision optimization module, which uses the Hungarian algorithm to construct the workstation group with the correlation coefficient higher than the critical value in the topological relationship matrix into a collaborative unit, performs cosine similarity matching between the material types in the preset strategy set and the equipment parameters of the collaborative unit, triggers a strategy reconstruction mechanism to generate a dynamic allocation instruction and synchronously update it to the topological modeling module.

8. The material dynamic allocation system based on data preset learning for a multi-station according to claim 7, characterized in that, The dynamic allocation instruction is specifically a material distribution path, a workstation cooperation method, and a material-equipment combination plan.

9. The material dynamic allocation system based on data preset learning for a multi-station according to claim 8, characterized in that, The critical value is the triple weighted sum of the mean and standard deviation of the elements of the topological relation matrix, where the mean weight is 0.6 and the standard deviation weight is 0.4; The policy reconstruction mechanism adjusts the policy weight coefficient through the gradient descent method, and its learning rate is set to the reciprocal of the difference between the real-time load of the workstation and the maximum capacity of the collaborative unit.

10. The material dynamic allocation system based on data preset learning for a multi-station according to claim 9, characterized in that, The decision optimization module includes: the collaborative unit construction sub-module calls the topological relation matrix, calculates the mean and standard deviation of the matrix elements, sets the critical value of the correlation coefficient as the triple weighted sum of the mean and standard deviation, constructs the bipartite graph of the workstation nodes, and uses the Hungarian algorithm to traverse the adjacency matrix to capture the maximum weight matching and generate the collaborative unit set; The policy matching sub-module, based on the collaborative unit set, extracts the device processing speed and capacity parameters to form a feature vector, calculates the cosine angle value between the material type features in the preset policy set and the device parameters, sets the matching threshold as the reciprocal of the ratio of the feature vector modulus lengths, filters the policy items with the angle value lower than the threshold, and generates an optimized policy candidate set; The dynamic allocation sub-module calls the optimized policy candidate set, detects the difference between the real-time load of the workstation and the maximum capacity of the collaborative unit, uses the gradient descent method to adjust the policy weight coefficient, encodes the updated policy instruction into a JSON format data packet, and generates a dynamic allocation instruction; The coefficient selection of the triple weighted sum is based on the correlation distribution corresponding to the peak value of the workstation collaboration efficiency in the historical data; The calculation formula for the ratio of the feature vector modulus lengths is the square root of the sum of the squares of the device processing speed and capacity parameters; The iteration termination condition of the gradient descent method is that the change amount of the weight coefficient is less than 0.01 in three consecutive iterations.

Images