Clothing production line intelligent scheduling control system and method

Abstract

The invention relates to the technical field of production line scheduling, and discloses an intelligent scheduling control system and method for a garment production line, and the system comprises a deviation calculation module, a topological optimization module, a skill correction module and a strategy curved surface generation module. According to the system, dynamic process parameters and station working condition parameters of an operator are collected, a related vector field and a frequency domain spectrum are constructed, and process-load deviation is calculated; on the basis, the topological structure of the production line is dynamically reconstructed, and an optimal sub-station segmentation scheme is obtained; correcting an operator skill entropy weight by using an augmented reality technology; a third-order feature tensor is constructed, production takt stability is modeled, and a multi-scale strategy curved surface is output; the method can improve the matching degree of personnel skills and processes, reduces the oscillation frequency of bottleneck processes, shortens the training period of green hands, and effectively optimizes the garment production line scheduling.

Description

Technical Field

[0001] The present invention relates to the technical field of production line scheduling, and more particularly to an intelligent scheduling control system and method for a clothing production line. Background Art

[0002] In the modern apparel industry, with the increasing diversification and personalization of market demand, efficient scheduling and collaborative operations on clothing production lines have become increasingly critical. This is especially true in hanging system workshops, where a mix of experienced and novice technicians operate. Addressing issues such as process load oscillation and bottleneck process drift caused by skill disparity has become a major challenge in improving production efficiency and product quality.

[0003] In the prior art, a Chinese patent with authorization announcement number CN105045236B discloses a production scheduling method and system for an assembly line, covering the production scheduling, tracking, feedback, and adjustment stages. This method performs initial allocation by obtaining site configuration and process information, tracks production data in real time, and analyzes site processing capabilities, thereby predicting delivery status and adjusting task allocation and resource allocation. However, this method has limitations when dealing with the problem of differences in personnel skills in clothing production. In the clothing production scenario, the skill levels of different workers vary, which significantly affects the completion efficiency and quality of the process. This prior art does not fully consider the impact of individual worker skill differences on production scheduling and is unable to accurately allocate tasks based on worker skills. For example, when senior technicians and novices are working in a mixed environment, it is difficult to ensure that each process is assigned to a worker with the most appropriate skill level, which can easily lead to uneven process loads, causing bottleneck processes to drift, affecting overall production efficiency.

[0004] The Chinese patent application with publication number CN117035328A discloses a production scheduling method, device, equipment and medium for a clothing production line. It proposes to obtain process beats and equipment information from the waiting pool, and determine the recommended stations in combination with the station beats and the information of the equipment that has been scheduled, so as to improve the rationality of the process arrangement. However, this method has shortcomings in dealing with process load oscillations caused by differences in personnel skills. In the clothing production process, differences in workers' skill levels will cause the completion time and quality of the same process to be different when operated by different workers. This existing technology does not conduct in-depth analysis and optimization of this key factor, and cannot dynamically adjust the process allocation according to the actual skills of the workers. As a result, in actual production, when encountering a combination of workers with a large skill gap, the process load cannot be effectively balanced. It is easy for some processes to have excessively high or low loads due to mismatched workers' skills, which in turn causes process load oscillations, affecting product quality and production stability.

[0005] Existing technologies are insufficient in solving problems such as process load oscillation and bottleneck process drift caused by differences in personnel skills in clothing production. Summary of the Invention

[0006] The intelligent scheduling control system and method for garment production lines presented in this invention are primarily used in the production workshops of garment manufacturers, particularly in hanging system workshops where a mix of senior and novice technicians operate. In such workshops, workers of varying skill levels participate in production, and the processes are complex and varied, placing extremely high demands on the flexibility and precision of production scheduling. For example, when producing a wide variety of fashion styles, the skill requirements for each process vary significantly, making traditional scheduling methods difficult to adapt. However, this system can effectively address this complex scenario, improving production efficiency and product quality.

[0007] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent scheduling and control system for a clothing production line, which aims to collect the operator's dynamic process parameters and workstation working condition parameters, construct relevant vector fields and frequency domain spectra, calculate the process-load deviation, and then optimize the production line topology. It uses augmented reality technology to correct the operator's skill entropy weight and construct a multi-scale strategy surface to achieve precise matching of personnel skills and processes, reduce the oscillation frequency of bottleneck processes, shorten the novice training cycle, and effectively improve the scheduling level and production efficiency of the clothing production line.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] Intelligent scheduling and control system for clothing production lines, including:

[0010] Deviation calculation module: used to collect the operator's dynamic process parameters, construct the process characteristic vector field based on the operator's dynamic process parameters, and calculate the skill entropy weight of each operator; collect the working parameters of each workstation in real time, and construct the dynamic oscillation frequency domain spectrum of the process load based on the working parameters of each workstation; calculate the process-load deviation based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load;

[0011] Topology optimization module: used to construct the initial adjacency matrix PTC0 of the production line's physical topology chain and the adjacency matrix STC of the operator skill topology chain, and to modify the initial adjacency matrix PTC0 of the production line's physical topology chain to generate the PTC-STC bimodal workstation chain adjacency matrix. When the process-load deviation CAD exceeds the preset deviation threshold ε, the dynamic reconstruction condition of the production line topology structure is triggered. When the dynamic reconstruction condition of the production line topology structure is triggered, the optimal sub-workstation segmentation scheme is obtained based on the PTC-STC bimodal workstation chain adjacency matrix, and the newly split sub-workstations are obtained according to the optimal sub-workstation segmentation scheme.

[0012] Skill correction module: For the newly split sub-stations, the operator's skill entropy weight is corrected using augmented reality technology;

[0013] Strategy surface generation module: Based on the process feature vector field and the dynamic oscillation frequency domain spectrum of the process load, it constructs a third-order feature tensor in the three dimensions of operator, workstation, and time. It applies a three-dimensional convolutional gated recurrent unit network to model the production rhythm stability and output a multi-scale strategy surface.

[0014] Furthermore, the dynamic process parameters include curve trajectory dwell time ratio STDR and seam track quantization deviation QSD;

[0015] Calculating the skill entropy weight of each operator includes:

[0016] Performing Min-Max normalization processing on the dynamic process parameters of each operator and mapping them to the interval [0, 1] to obtain normalized dynamic process parameters, wherein the normalized dynamic process parameters include the normalized curve trajectory dwell time ratio STDR' and the seam quantization deviation QSD';

[0017] According to the normalized dynamic process parameters and the dimension weight λ of the dynamic process parameters in the preset process feature vector field i” , calculate the skill feature vector TFV of each operator i ;where i' is the index of the dynamic process parameter dimension and i is the index of the operator;

[0018] For each operator's skill feature vector TFV i Normalize the modulus length to obtain the unit skill feature vector UTFV i ;

[0019] Based on the unit skill feature vector UTFV i , calculate the skill entropy weight SEW of each operator i .

[0020] Furthermore, the working condition parameters include the station dwell time-varying gradient TDG, the material rheological curvature MRC and the load value; one station corresponds to one operator;

[0021] The dynamic oscillation frequency domain spectrum of the construction process load includes:

[0022] Set the station dwell time-varying gradient of station j to TDG j and the material rheological curvature is MRC j , based on TDG j and MRC j Construct the load-logistics pair (TDG) of station j j ,MRC j ), the load-logistics binary (TDG j ,MRC j) is projected into the amplitude-frequency domain to obtain the dynamic oscillation frequency domain spectrum DOFS of the process load of station j j , where 1≤j≤n, and n is the total number of workstations.

[0023] Furthermore, calculating the process-load deviation CAD includes:

[0024] The skill entropy weight of each operator in the process feature vector field is time-series aligned with the dynamic oscillation frequency domain spectrum of the process load of its corresponding workstation to construct the interaction matrix between operator skill and process load.

[0025] Extract the entropy weight time series of each operator's operation skill and the oscillation frequency domain time series of the process load of each workstation from the interaction matrix of operation skill and process load;

[0026] The cumulative error metric between the operation skill entropy weight time series and the process load oscillation frequency domain time series is calculated, and the cumulative error metric is defined as the process-load deviation CAD.

[0027] Furthermore, the construction of the initial adjacency matrix PTC0 of the physical topology chain of the production line includes:

[0028] The physical location relationship of the equipment on the clothing production line is obtained. Based on the physical location relationship of the equipment, the initial adjacency matrix PTC0 of the physical topology chain of the production line is constructed. The matrix element PTC0(i1,j1) represents the material direct arrival time between workstation i1 and workstation j1. The initial adjacency matrix PTC0 of the physical topology chain of the production line contains physical edges, and the material direct arrival time is used as the physical edge weight.

[0029] Furthermore, constructing the operator skill topology chain adjacency matrix STC includes:

[0030] The skill entropy weights of each operator are sorted across workstations, and the operator skill topology chain adjacency matrix STC is generated according to the principle of skill complementarity. The matrix element STC(p,q) represents the skill substitution difficulty between operator p and operator q. The operator skill topology chain adjacency matrix STC contains virtual edges, and the skill substitution difficulty is used as the virtual edge weight.

[0031] Furthermore, the modification of the initial adjacency matrix PTC0 of the physical topology chain of the production line includes:

[0032] Based on the principle of material transfer efficiency balance, the initial adjacency matrix PTC0 of the physical topology chain of the production line is modified to obtain the first physical topology matrix PTC1; based on the principle of skill complementary optimization, the first physical topology matrix PTC1 is modified to obtain the second physical topology matrix PTC2.

[0033] Furthermore, obtaining the first physical topology matrix PTC1 includes:

[0034] Calculate the average material transfer time t1 between any two workstations in the initial adjacency matrix PTC0 of the physical topology chain of the production line, and define the physical edges with a transfer time greater than 1.5t1 as physical edges with excessive transfer time; for each physical edge (i2, j2) with excessive transfer time, i2 and j2 represent the workstations corresponding to the endpoints of the physical edge with excessive transfer time; assuming that workstation i2 corresponds to operator i3, and workstation j2 corresponds to operator j3, find the virtual edge with operator i3 or operator j3 as the endpoint and the smallest virtual edge weight in the operator skill topology chain adjacency matrix STC, and define it as the virtual edge with the lowest skill substitution difficulty; multiply the physical edge weight in the initial adjacency matrix PTC0 of the physical topology chain of the production line corresponding to the virtual edge with the lowest skill substitution difficulty by the attenuation factor α1 to obtain the first physical topology matrix PTC1; where 0<α1<1.

[0035] Furthermore, obtaining the optimal sub-station segmentation solution includes:

[0036] When the dynamic reconstruction condition of the production line topology is triggered, the dynamic balance factor μ between production efficiency and operational fairness is set, and the trade-off objective function of the bimodal workstation chain reconstruction is established based on μ.

[0037] The Monte Carlo method is used to randomly sample in the high-dimensional solution space formed by the adjacency matrix M of the PTC-STC dual-modal workstation chain. Each sampling point corresponds to a potential sub-workstation reconstruction solution.

[0038] A genetic algorithm is used to combinatorially optimize the potential sub-station reconstruction schemes obtained by Monte Carlo sampling to obtain candidate sub-station reconstruction schemes. The candidate sub-station reconstruction schemes are substituted into the trade-off objective function to calculate the J value. The top 10% schemes with the highest J value are selected as dominant species. After crossover mutation, the next generation population is generated and multiple iterations are carried out until convergence.

[0039] Output the optimal solution after convergence as the optimal sub-station division plan under the current conditions.

[0040] An intelligent scheduling and control method for a clothing production line is provided, which is based on the above-mentioned intelligent scheduling and control system for a clothing production line. The method comprises:

[0041] The operator's dynamic process parameters are collected, and based on these parameters, a process characteristic vector field is constructed to calculate the skill entropy weight of each operator. The working parameters of each workstation are collected in real time, and based on these parameters, a dynamic oscillation frequency domain spectrum of the process load is constructed. The process-load deviation is calculated based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load.

[0042] The initial adjacency matrix PTC0 of the production line's physical topology chain and the adjacency matrix STC of the operator's skill topology chain are constructed, and the initial adjacency matrix PTC0 of the production line's physical topology chain is modified to generate a PTC-STC dual-modal workstation chain adjacency matrix. When the process-load deviation CAD exceeds a preset deviation threshold ε, the dynamic reconstruction condition of the production line topology structure is triggered. When the dynamic reconstruction condition of the production line topology structure is triggered, the optimal sub-workstation segmentation scheme is obtained based on the PTC-STC dual-modal workstation chain adjacency matrix, and the newly split sub-workstations are obtained according to the optimal sub-workstation segmentation scheme.

[0043] For the newly split sub-stations, the operator's skill entropy weight is corrected using augmented reality technology;

[0044] Based on the dynamic oscillation frequency domain spectrum of the process characteristic vector field and the process load, a third-order feature tensor in the three dimensions of operator, workstation, and time is constructed. A three-dimensional convolutional gated recurrent unit network is applied to model the production rhythm stability and output a multi-scale strategy surface.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The intelligent scheduling control system and method for a clothing production line provided by the present invention comprehensively improves the scheduling level of a clothing production line through the coordinated operation of multiple modules. The deviation calculation module collects and analyzes data, quantifies process-load deviations, and provides a basis for subsequent optimization. The topology optimization module reconstructs the production line topology based on the deviations, obtains a reasonable sub-station segmentation scheme, and optimizes the production layout. The skill correction module uses augmented reality technology to correct the operator's skill entropy weight in real time, improve the skill level of novices, and shorten the training cycle. The strategy surface generation module constructs a third-order feature tensor and outputs a multi-scale strategy surface to assist in production decision-making. These functions work together to significantly improve the matching degree between personnel skills and process steps, reduce process load oscillations caused by skill gaps, and reduce the oscillation frequency of bottleneck processes, ultimately achieving efficient and stable operation of the clothing production line, improving production efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 This is a functional module diagram of the intelligent scheduling control system for a clothing production line in the present invention;

[0049] Figure 2Flowchart of a method for calculating process-load deviation in the intelligent scheduling and control system of a clothing production line of the present invention;

[0050] Figure 3 A flow chart of a method for obtaining an optimal sub-station segmentation scheme in the intelligent scheduling and control system for a garment production line of the present invention;

[0051] Figure 4 This is a flow chart of a method for correcting an operator's skill entropy weight using augmented reality technology in the intelligent scheduling and control system for a clothing production line of the present invention;

[0052] Figure 5 A flow chart of a method for constructing a third-order feature tensor in the intelligent scheduling and control system for a clothing production line of the present invention;

[0053] Figure 6 Schematic diagram of the principle of the intelligent scheduling control method for a clothing production line in the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] Example 1

[0056] See also Figure 1 As shown, this embodiment provides an intelligent scheduling and control system for a clothing production line, including:

[0057] Deviation calculation module: used to collect the operator's dynamic process parameters, construct the process characteristic vector field based on the operator's dynamic process parameters, and calculate the skill entropy weight of each operator; collect the working parameters of each workstation in real time, and construct the dynamic oscillation frequency domain spectrum of the process load based on the working parameters of each workstation; calculate the process-load deviation based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load;

[0058] Topology optimization module: used to construct the initial adjacency matrix PTC0 of the production line's physical topology chain and the adjacency matrix STC of the operator skill topology chain, and to modify the initial adjacency matrix PTC0 of the production line's physical topology chain to generate the PTC-STC bimodal workstation chain adjacency matrix. When the process-load deviation CAD exceeds the preset deviation threshold ε, the dynamic reconstruction condition of the production line topology structure is triggered. When the dynamic reconstruction condition of the production line topology structure is triggered, the optimal sub-workstation segmentation scheme is obtained based on the PTC-STC bimodal workstation chain adjacency matrix, and the newly split sub-workstations are obtained according to the optimal sub-workstation segmentation scheme.

[0059] Skill correction module: For the newly split sub-stations, the operator's skill entropy weight is corrected using augmented reality technology;

[0060] Strategy surface generation module: Based on the process feature vector field and the dynamic oscillation frequency domain spectrum of the process load, it constructs a third-order feature tensor in the three dimensions of operator, workstation, and time. It applies a three-dimensional convolutional gated recurrent unit network to model the production rhythm stability and output a multi-scale strategy surface.

[0061] In the deviation calculation module, the acquisition of the operator's dynamic process parameters, the construction of the process characteristic vector field based on the operator's dynamic process parameters, and the calculation of the skill entropy weight of each operator include:

[0062] Step S1110 , collecting the operator's dynamic process parameters, wherein the dynamic process parameters include the curve trajectory dwell time ratio STDR and the seam quantization deviation QSD;

[0063] Specifically, the operator's dynamic process parameters such as the curve trajectory dwell time ratio (STDR) and the seam quantization deviation (QSD) are first collected through a high-dimensional motion capture system, with a sampling frequency of not less than 120Hz. This is because a higher sampling frequency can ensure that sufficiently detailed motion trajectory information is collected, thereby more accurately reflecting the operator's behavioral characteristics on the critical process path. STDR is used to quantitatively reflect the operator's dwell time distribution characteristics on the critical process path. For example, in the process of sewing clothes, if a certain part requires more sewing time to ensure quality, STDR can reflect the proportion of the operator's dwell time in this part to the total sewing time. QSD is used to measure the seam quantization deviation, that is, the level of sewing accuracy. For example, when sewing straight seams, QSD can indicate the degree of deviation between the actual seam and the standard straight line. After collecting these parameters, they are subsequently used to construct the process feature vector field.

[0064] The beneficial effect of this step is that by collecting these two key parameters that can quantify operational skills, a data foundation is provided for the subsequent accurate assessment of the operator's skill level. In clothing production, different sewing tasks require different operational skills. Accurately assessing skill levels can help to rationally allocate tasks and improve overall production efficiency and product quality. Taking the production of a pair of jeans as an example, the sewing of the back pocket requires high precision. If the operator's skill level in this area can be accurately assessed, this task can be assigned to an operator with a smaller QSD value and higher sewing precision, thereby reducing the defective rate. At the same time, collecting detailed motion trajectory information can promptly identify the operator's bad operating habits or skill shortcomings, providing a basis for targeted training and further improving the operator's overall skill level.

[0065] Step S1120 , reducing the dimension of the operator's dynamic process parameters and mapping them to the process feature vector field, and calculating the skill entropy weight of each operator.

[0066] Furthermore, step S1120 includes:

[0067] Step S1121, performing Min-Max normalization processing on the dynamic process parameters of each operator, mapping them to the interval [0, 1], and obtaining normalized dynamic process parameters, wherein the normalized dynamic process parameters include the normalized curve trajectory dwell time ratio STDR' and the seam quantization deviation QSD';

[0068] Step S1122: according to the normalized dynamic process parameters and the dimension weight λ of the dynamic process parameters in the preset process feature vector field i” , calculate the skill feature vector TFV of each operator i ;where i' is the index of the dynamic process parameter dimension and i is the index of the operator;

[0069] The skill feature vector TFV of each operator i The calculation formula includes:

[0070] TFV i =(λ1×STDR' i ,λ2×QSD' i );

[0071] Where λ1 is the weight of STDR' and λ2 is the weight of QSD'.

[0072] Step S1123: skill feature vector TFV of each operator i Normalize the modulus length to obtain the unit skill feature vector UTFV i ;

[0073] Step S1124: Based on the unit skill feature vector UTFV i , calculate the skill entropy weight SEW of each operator i .

[0074] Specifically, STDR and QSD are normalized because the numerical ranges of STDR and QSD for different operators in different process tasks may vary greatly, making it difficult to compare and analyze directly using the original data. Through normalization, all data are unified into the interval [0,1], making the parameters of different operators comparable. For example, the original STDR value of operator A1 is 5 seconds, and the original STDR value of operator B1 is 10 seconds. Without normalization, it is difficult to intuitively judge the impact of the difference between the two on skill evaluation. After normalization, the skill performance of the two can be compared on the same scale. The beneficial effect of this step is that it provides standardized data for the subsequent accurate calculation of skill feature vectors and skill entropy weights, avoiding calculation errors caused by differences in data scales, and improving the accuracy and reliability of skill evaluation.

[0075] The skill feature vector TFV of each operator is calculated based on the normalized dynamic process parameters and the dimension weight λi of the dynamic process parameters in the preset process feature vector field. i , the calculation formula is TFV i =(λ1×STDR' i ,λ2×QSD' i ), where i'' is the index of the dynamic process parameter dimension (in this embodiment, for STDR' and QSD', i'' takes the values of 1 and 2), i is the operator's index, λ1 is the weight of STDR', and λ2 is the weight of QSD'. These weights are pre-set based on the importance of the process and the degree of impact on the overall skill level. For example, if sewing accuracy has a greater impact on product quality in a certain garment production process, then the weight λ2 of QSD' may be set relatively high. Through this formula, the normalized dynamic process parameters are combined with the weights to comprehensively consider the contribution of different parameters to the skill feature vector. Assuming λ1 = 0.4, λ2 = 0.6, and operator C1's STDR' is 0.7 and QSD' is 0.8, then the first component of its skill feature vector TFVC is 0.4×0.7=0.28, and the second component is 0.6×0.8=0.48. The beneficial effect of this step is that, by introducing weights, it can highlight the differences in the importance of different process parameters in skill evaluation, make the calculated skill feature vector more in line with actual process requirements, more accurately reflect the operator's skill characteristics, and provide a more targeted basis for subsequent skill entropy weight calculation and task allocation.

[0076] The purpose of modulus normalization is to normalize the vector length to 1, giving it a uniform scale and facilitating subsequent calculations and comparisons. This eliminates the differences in skill feature vector lengths, allowing for the comparison and analysis of skill features across different operators using a unified standard. This improves the consistency and comparability of skill assessment results, facilitates a more accurate assessment of skill differences between operators, and provides a more reliable basis for the rational scheduling of production tasks.

[0077] The calculation of the skill entropy weight SEW of each operator i The calculation formula is:

[0078] SEW i =γ(UTFVi)=A×(UTFV i ) T ×H×UTFV i ;

[0079] in:

[0080] γ(·) is the skill entropy weight mapping function, which is used to map the unit skill feature vector to the skill entropy weight;

[0081] A is the skill entropy gain coefficient, ranging from [1,10];

[0082] H is the skill dimension interaction matrix, and the matrix element H k”l (k”,l=1,2) represents the interaction intensity between the k”th skill dimension and the lth skill dimension, and its value range is [0,1].

[0083] Skill entropy weight SEW i It quantitatively represents the proficiency and stability of the i-th operator's mastery of process skills. Among them, the coefficient A reflects the contribution of the operator's overall skill level to the entropy weight. The larger the A value, the higher the operator's comprehensive skill level and the greater the skill entropy weight. The matrix H describes the mutual influence and constraint relationship between different skill dimensions. k”l The closer it is to 1, the stronger the positive correlation between the k-th dimension skill and the l-th dimension skill, the more obvious the symbiotic effect, and thus the greater the incremental effect on the entropy weight.

[0084] The coefficient A and matrix H can be obtained by fitting the historical process data of a large number of operators through multivariate regression analysis. Specifically, the UTFV of each sample operator is calculated according to steps S1121-S1123. i , then in UTFV iAs the independent variable, the actual process yield of the corresponding operator is the dependent variable, and the least squares method is used for surface fitting to obtain the regression coefficients A and H. In actual production, the values of A and H can be dynamically adjusted according to the level of product quality requirements. When the quality requirements are high, the A value and H value should be appropriately increased. k”l value to enhance the influence of the operator's skill level on the entropy weight.

[0085] In general, the skill entropy weight SEW i The inherent relationship between the skill status of the i-th operator and the process quality is revealed, which is an important basis for subsequent human-machine collaborative optimization scheduling. i The higher the value, the higher the operator's proficiency and stability, and the operator can be given priority to undertake more critical process tasks. In addition, by tracking and analyzing the SEW of each operator i The dynamic change trend of production processes can be monitored and the risk of skill degradation can be discovered in a timely manner, and targeted personnel deployment and training measures can be taken to achieve predictive control of process quality.

[0086] In the deviation calculation module, the real-time acquisition of the working parameters of each workstation and the construction of the dynamic oscillation frequency domain spectrum of the process load based on the working parameters of each workstation include:

[0087] Step S1210: Real-time acquisition of working parameters of each workstation, including the workstation dwell time-varying gradient TDG, material rheological curvature MRC, and load value; one workstation corresponds to one operator;

[0088] Step S1220, constructing a dynamic oscillation frequency domain spectrum of the process load of each workstation based on the working condition parameters;

[0089] Set the station dwell time-varying gradient of station j to TDG j and the material rheological curvature is MRC j , based on TDG j and MRC j Construct the load-logistics pair (TDG) of station j j ,MRC j ), the load-logistics binary (TDG j ,MRC j ) is projected into the amplitude-frequency domain to obtain the dynamic oscillation frequency domain spectrum DOFS of the process load of station j j .

[0090] Specifically, TDG reflects the rate of change in the dwell time of a workpiece at each workstation. For example, on a clothing production line, a high TDG value at a particular workstation indicates that the dwell time at that workstation varies significantly. This may be due to unstable worker speeds or complex process flows at that workstation, resulting in large fluctuations in operation time. MRC indicates the smooth flow of materials along the production line. A low MRC value indicates smooth material flow, without significant blockages or stagnation. Conversely, a high MRC value suggests potential obstructions to material flow, such as poor material transport tracks or untimely material handovers between workstations. The load value reflects the current workload at each workstation. By collecting these parameters in real time, comprehensive information on the current status of each workstation on the production line can be obtained. This provides an accurate data foundation for subsequent analysis. For example, in a clothing production workshop, continuous monitoring of TDG, MRC, and load values revealed that the TDG of a particular workstation was gradually increasing, the MRC was also outside the normal range, and the load value was excessively high, indicating a high probability of a production bottleneck at that workstation. Based on this data, managers can intervene in a timely manner and adjust production arrangements, such as deploying other workers to assist, optimizing process flows, etc., to avoid production delays and ensure the efficient operation of the production line.

[0091] The 2D-FFT algorithm is a mathematical tool that can convert time domain signals into frequency domain signals. In this embodiment, the algorithm can be used to convert TDG and MRC that reflect the real-time status of the production line from time domain features to frequency domain features. For example, suppose that the TDG and MRC of a certain workstation show complex fluctuations over time, and it is difficult to intuitively judge its patterns and potential problems in the time domain. After being processed by the 2D-FFT algorithm, the amplitude of different frequency components can be clearly seen in the frequency domain. If the amplitude is large at a certain frequency, it means that the workstation has a more prominent load fluctuation problem on the corresponding time scale, which may be a potential production bottleneck. The beneficial effect of this step is that by analyzing the process load from a frequency domain perspective, potential problems that are not easy to find in the time domain analysis can be discovered, and a more accurate perception and prediction of the production rhythm can be achieved. In this way, the bottleneck links that may appear in the production line can be discovered in advance, providing a strong basis for timely adjustment of production strategies. For example, in actual production, according to DOFS j It was discovered that a certain workstation had abnormal amplitude in the high-frequency band. Further investigation determined that the unstable operation time was due to aging equipment. Equipment maintenance or replacement could then be arranged in advance to avoid production stagnation caused by equipment problems and improve the stability and efficiency of the production line.

[0092] In the deviation calculation module, such as Figure 2 As shown, the process-load deviation is calculated based on the dynamic oscillation frequency domain spectrum of the process characteristic vector field and the process load, including:

[0093] Step S1310 , aligning the skill entropy weight of each operator in the process feature vector field with the dynamic oscillation frequency domain spectrum of the process load of the corresponding workstation to construct an interaction matrix between operator skill and process load;

[0094] Step S1320 , extracting the entropy weight time series of each operator's operation skill and the oscillation frequency domain time series of the process load of each workstation from the operation skill and process load interaction matrix;

[0095] Step S1330 , calculating the cumulative error metric between the operation skill entropy weight time series and the process load oscillation frequency domain time series, and defining the cumulative error metric as the process-load deviation CAD.

[0096] Specifically, the skill entropy weight, calculated in the previous steps, quantifies the operator's proficiency and stability in mastering process skills. Higher values indicate greater proficiency and stability. The dynamic oscillation frequency domain spectrum of the process load reflects the load conditions at each workstation at different times. Temporal alignment involves matching the two in the temporal dimension to ensure comparison and analysis occur on the same time scale. For example, the skill entropy weight of operator A2 at 10:00 AM is aligned with the dynamic oscillation frequency domain spectrum of the process load at his or her workstation at that time. Each element in the constructed interaction matrix of operator skill and process load represents the degree of match between the operator's skill level and the intensity of the process load at that moment. Suppose, at a certain moment, operator B2 has a high skill entropy weight, indicating that he or she is able to complete the work efficiently, while the process load at his or her workstation is relatively low. In this case, the corresponding element in the interaction matrix will have a large value, indicating a high degree of match between the two and potentially high production efficiency. Conversely, a small element value indicates a low degree of match, potentially indicating problems with human-machine collaboration. The beneficial effect of this step is that by constructing an interaction matrix, it can intuitively present the matching relationship between operator skills and process load at different times, providing a clear data basis for analyzing the human-machine collaboration status during the production process. For example, on a clothing production line, if the interaction matrix element values of multiple workstations are generally low during a certain period of time, it means that there is a problem with human-machine collaboration during this period. This may be due to a mismatch between operator skills and work tasks, or an unreasonable process arrangement. Further analysis and adjustments are needed to improve production efficiency.

[0097] Since each workstation corresponds to one operator, the time series of the entropy weights of the operator's skills over time, reflecting the changes in the operator's skills over time, and the time series of the process load oscillation frequency domain, reflecting the changes in the workstation load over time, can be extracted from the interaction matrix in chronological order. For example, the skill entropy weights of operator C2 at different times such as 9:00, 10:00, and 11:00 a.m. are sequentially extracted from the interaction matrix to form the time series of the entropy weights of the operator's skills; at the same time, the key characteristic data (such as the bottleneck drift characteristic factor) in the process load oscillation frequency domain spectrum of the corresponding workstation at the corresponding time are extracted to form the time series of the process load oscillation frequency domain. The purpose of this step is to further refine and organize the data in the interaction matrix for subsequent more accurate quantitative analysis. The beneficial effect is that by extracting these two time series data, the dynamic changes in operator skills and process load over time can be clearly observed. For example, if we observe a gradual decrease in the skill entropy weight of an operator by observing its skill entropy weight over time, this could be due to fatigue from long hours, which has impacted their skill. Meanwhile, if the process load oscillation frequency domain time series at their workstation shows a gradual increase in load, this could indicate a potential process-load imbalance. By identifying these trends in advance, timely measures can be taken, such as adjusting personnel work hours and reallocating tasks, to avoid decreased production efficiency and product quality issues.

[0098] Step S1330 uses the dynamic time warping (DTW) algorithm to calculate the cumulative error metric. The DTW algorithm is an algorithm for calculating the similarity between two time series. It can find the optimal matching path between the two when the time series have different lengths and changing speeds, thereby calculating the cumulative error metric. In this embodiment, the operating skill entropy weight time series and the process load oscillation frequency domain time series are used as inputs to the DTW algorithm, and the cumulative error metric obtained by calculation reflects the degree of difference between the two, namely the process-load deviation (CAD). For example, assuming that the operating skill entropy weight time series indicates that the operator's skill level has always been high and stable, while the process load oscillation frequency domain time series shows that the workstation load fluctuates greatly, the CAD value calculated by the DTW algorithm will be large, indicating that the process-load imbalance is more serious. The calculation of the CAD deviation fully considers the dynamic change characteristics of the operating skill and process load in the time dimension, and can timely quantify the imbalance of the human-machine collaboration level in the production cycle. When the imbalance exceeds the limit, reverse gradient compensation is activated in time to prevent problems before they occur. This step has the beneficial effect of accurately calculating process-load deviations, enabling real-time monitoring of the state of human-machine collaboration during production. If an imbalance is detected, such as a CAD value exceeding a preset threshold, indicating that the current production schedule may be irrational, a reverse gradient compensation mechanism can be activated, such as adjusting personnel allocation and optimizing process flows, to restore production to an efficient and stable state, ensuring the normal operation of the production line and avoiding problems such as reduced production efficiency and product quality caused by imbalances.

[0099] In the topology optimization module, the construction of the production line's physical topology chain initial adjacency matrix PTC0 and the operator skill topology chain adjacency matrix STC, the production line's physical topology chain initial adjacency matrix PTC0 is modified to generate the PTC-STC dual-modal workstation chain adjacency matrix M, including:

[0100] Step S2110: Obtain the physical location relationship of the equipment on the clothing production line. Based on the physical location relationship of the equipment, construct the initial adjacency matrix PTC0 of the physical topology chain of the production line. The matrix element PTC0(i1,j1) represents the direct material arrival time between workstation i1 and workstation j1. The initial adjacency matrix PTC0 of the physical topology chain of the production line contains physical edges, and the direct material arrival time is used as the physical edge weight.

[0101] Specifically, in the clothing production scenario, the physical location of equipment determines the material transfer path and time between workstations. The Floyd algorithm is used to construct PTC0. The matrix element PTC0(i1,j1) represents the direct material transfer time between workstation i1 and workstation j1, and this time serves as the physical edge weight. For example, suppose a clothing production line has workstations A3, B3, and C3. Through measurement and calculation, it is determined that the direct material transfer time from workstation A3 to workstation B3 is 5 minutes. Therefore, PTC0(A,B) = 5. This step, by clarifying the material transfer time between workstations, provides basic data for subsequent material flow analysis. Its beneficial effect is that it provides an intuitive data basis for modeling the material flow of the entire production line. By analyzing PTC0, we can clearly understand the speed and efficiency of material flow between workstations, providing a reference for optimizing material distribution routes and rationalizing production sequencing, thereby helping to improve overall production efficiency. For example, if you find that the material delivery time between two workstations is too long, you can consider adjusting their relative positions or adding an intermediate buffer area to reduce the material waiting time.

[0102] Step S2120: sort the skill entropy weights of each operator across workstations and generate an operator skill topology chain adjacency matrix STC based on the principle of skill complementarity. The matrix element STC(p,q) represents the skill substitution difficulty between operator p and operator q. The operator skill topology chain adjacency matrix STC includes virtual edges, and the skill substitution difficulty is used as the virtual edge weight.

[0103] Specifically, the principle of skill complementarity involves connecting operators with complementary skill levels to improve overall production efficiency. The matrix element STC(p,q) represents the skill substitution difficulty between operator p and operator q, with skill substitution difficulty serving as the virtual edge weight. For example, if operator A excels at sewing straight lines, while operator B excels at sewing curved lines, their skills are complementary. If operator A is temporarily absent, it would be easier for operator B to replace A, and the weight of STC(A,B) would be relatively low. This step, by constructing STC, quantifies and visualizes the skill relationships between operators, facilitating subsequent personnel deployment and task allocation based on skill complementarity. This beneficial effect is that it optimizes staffing and increases the flexibility and adaptability of the production line. If an operator is unable to work due to special circumstances, a suitable replacement can be quickly found, minimizing the impact of personnel changes on production. Furthermore, collaboration among workers with complementary skills can fully leverage their respective strengths, improving product quality and production efficiency. For example, when producing a complex garment, having workers with different sewing skills collaborate can enhance the sewing process and reduce the defective rate.

[0104] Step S2130: Based on the principle of material transfer efficiency balance, the initial adjacency matrix PTC0 of the physical topology chain of the production line is modified to obtain a first physical topology matrix PTC1;

[0105] Calculate the average material transfer time t1 between any two workstations in the initial adjacency matrix PTC0 of the physical topology chain of the production line, and define the physical edges with a transfer time greater than 1.5t1 as physical edges with excessive transfer time; for each physical edge (i2, j2) with excessive transfer time, i2 and j2 represent the workstations corresponding to the endpoints of the physical edge with excessive transfer time; assuming that workstation i2 corresponds to operator i3, and workstation j2 corresponds to operator j3, find the virtual edge with operator i3 or operator j3 as the endpoint and the smallest virtual edge weight in the operator skill topology chain adjacency matrix STC, and define it as the virtual edge with the lowest skill substitution difficulty; multiply the physical edge weight in the initial adjacency matrix PTC0 of the physical topology chain of the production line corresponding to the virtual edge with the lowest skill substitution difficulty by the attenuation factor α1 to obtain the first physical topology matrix PTC1; where 0<α1<1.

[0106] Specifically, based on the principle of balanced material transfer efficiency, the goal is to make the material transfer time distribution on the production line more uniform, avoiding situations where some paths have excessively long material transfer times, thereby improving overall material transfer efficiency. For example, if the calculated average material transfer time t1 is 3 minutes, and the material transfer time of a physical edge is 5 minutes (greater than 1.5t1 = 4.5 minutes), this edge connects workstations X and Y, corresponding to operators X1 and Y1. If the virtual edge weight between operators X1 and Z1 is found to be the smallest in STC, then the physical edge weight between workstations X and Y in PTC0 is multiplied by the attenuation factor α1 (e.g., α1 = 0.8) to obtain the corresponding edge weight in PTC1. This step aims to implement negative feedback regulation of material flow through a "people-carrying-materials" strategy, reducing the weight of physical edges with excessively long transfer times and guiding material flow toward paths with higher transfer efficiency. The beneficial effect is to optimize material transfer paths, improve material transfer efficiency, and reduce the time materials spend on the production line, thereby improving overall production efficiency. For example, if a material transmission path originally took a long time due to a long distance or equipment problem, by adjusting the weight, the material can choose other more efficient paths more often, avoiding material backlogs and ensuring the smooth operation of the production line.

[0107] Step S2140: Based on the skill complement optimization principle, the first physical topology matrix PTC1 is modified to obtain the second physical topology matrix PTC2;

[0108] Calculate the mean of each row of the operator skill topology chain adjacency matrix STC, mark it as the row mean, and define the workstations corresponding to elements less than 50% of the row mean as critical workstations with insufficient skill complementarity; find the row corresponding to each critical workstation k in PTC1, and define the workstation corresponding to the non-zero element with the smallest physical edge weight as the alternative workstation with the shortest direct material access time for the critical workstation k; multiply the physical edge weight between the critical workstation and the alternative workstation in PTC1 by the gain factor β1 to obtain the second physical topology matrix PTC2; where β1>1.

[0109] Specifically, based on the skill complementarity optimization principle, it means that by analyzing the skill complementarity relationship between workstations, the connection between workstations with insufficient skill complementarity is enhanced to optimize the skill synergy effect of the entire production line. For example, assuming that the mean value of a certain row of STC is 0.6, the workstation with an element value less than 0.3 (50%×0.6) is the key workstation k2. In PTC1, in the row corresponding to the key workstation k2, the workstation corresponding to the non-zero element with the smallest physical edge weight is N2, then the physical edge weight between workstation k2 and workstation N2 in PTC1 is multiplied by the gain factor β1 (such as β1=1.2) to obtain the corresponding edge weight in PTC2. The purpose of this step is to optimize the skill matching between workstations by enhancing the physical edge weights of workstations with insufficient skill complementarity. Its beneficial effect is that it further improves the overall skill synergy of the production line and makes the production process smoother. For example, workstations that originally had insufficient skill complementarity may have problems with poor coordination during production. By adjusting the weights, the connection between these workstations is strengthened, making the collaboration between workers closer, reducing production delays caused by skill mismatch, and improving production efficiency and product quality.

[0110] Step S2150: Multiply the second physical topology matrix PTC2 and the operator skill topology chain adjacency matrix STC bit by bit to obtain the PTC-STC dual-modal workstation chain adjacency matrix M.

[0111] Specifically, bitwise multiplication refers to the multiplication of the elements at corresponding positions in two matrices. The element M(i1, j1) of matrix M represents the comprehensive matching degree between workstations i1 and j1, taking into account multiple factors such as material transfer time (derived from PTC2) and the difficulty of operator skill substitution (derived from STC). For example, if the physical edge weight between workstations A3 and B3 in PTC2 is 4, and the skill substitution difficulty weight between operators A1 and B1 in STC is 0.5, then M(A3, B3) = 4 × 0.5 = 2. This matrix M provides an important data foundation for the subsequent dynamic reconstruction of the production line topology based on process-load deviations. Its beneficial effect lies in that, by comprehensively considering multiple factors in the human-machine operation process, it provides key data support for achieving dynamic coordination between production tact and operational rhythm. For example, during production line scheduling, the values of each element in matrix M can be used to quickly determine which workstations have a high degree of matching, thereby rationally arranging production tasks and personnel deployment, thereby improving the overall operational efficiency of the production line.

[0112] In the topology optimization module, such as Figure 3 As shown, when the process-load deviation CAD exceeds the preset deviation threshold ε, the dynamic reconstruction condition of the production line topology is triggered; when the dynamic reconstruction condition of the production line topology is triggered, the optimal sub-station segmentation scheme is obtained based on the PTC-STC dual-modal station chain adjacency matrix, including:

[0113] Step S2210: When the dynamic reconstruction condition of the production line topology is triggered, a dynamic balance factor μ between production efficiency and operational fairness is set, and a trade-off objective function for the reconstruction of the bimodal workstation chain is established based on μ;

[0114] Specifically, in the garment production process, when the process-load deviation (CAD) exceeds a preset deviation threshold ε, it indicates a serious imbalance in the human-machine collaboration of the current production line. For example, a mismatch between operator skills and process loads can lead to reduced production efficiency and unstable product quality. In this case, the production line topology needs to be dynamically reconfigured to optimize the production process.

[0115]

[0116] Where μ is a dynamic balancing factor between production efficiency and operational fairness; its value ranges from 0 to 1 and is used to weigh the relative importance of these two objectives during the bimodal workstation chain reconstruction process. When μ approaches 1, the optimization process prioritizes improving production efficiency, such as prioritizing reducing production time and increasing output. When μ approaches 0, greater emphasis is placed on operational fairness, ensuring that each operator is appropriately assigned tasks that match their skills, avoiding uneven task distribution.

[0117] Tr(·) is the trace of the matrix; it is the sum of the main diagonal elements of the matrix. In this embodiment, it is used to quantify the Kronecker product of PTC2 (the modified production line physical topology chain matrix) and STC (the operator skill topology chain adjacency matrix). Matrix characteristics after operation. The Kronecker product is a special matrix product that combines two matrices according to specific rules. In this embodiment, this operation comprehensively considers the information of the physical topology of the production line and the topology of the operator's skills.

[0118] is the Kronecker product;

[0119] n is the total number of workstations;

[0120] J is a trade-off objective function, a quantitative metric that comprehensively considers production efficiency and operational fairness. It is used to evaluate the advantages and disadvantages of different bimodal workstation chain reconfiguration schemes. A larger J value indicates that, under the current μ setting, the corresponding scheme performs better in terms of increasing output, balancing loads, and suppressing dynamic mismatch between process and load. For example, in a clothing production line, if a reconfiguration scheme significantly improves production efficiency while also ensuring operational fairness, then this scheme will have a relatively high J value when substituted into the trade-off objective function. By maximizing J, the bimodal workstation chain reconfiguration scheme that best meets production requirements under current production conditions can be found. The trade-off objective function provides a clear quantitative optimization direction for the dynamic reconfiguration of production line topology. By properly adjusting μ, production efficiency and operational fairness can be flexibly balanced according to actual production needs, avoiding the pursuit of one objective while neglecting others. This allows for optimal allocation of production resources and improves overall production efficiency.

[0121] is the sum of the skill entropy weights of the operators, where m is the total number of operators, m=n.

[0122] The dynamic balance factor μ, which ranges from 0 to 1, balances the relative importance of these two objectives during the bimodal workstation chain reconstruction process. When μ approaches 1, the optimization process prioritizes improving production efficiency, such as reducing production time and increasing output. When μ approaches 0, greater emphasis is placed on operational fairness, ensuring that each operator is appropriately assigned a task that matches their skills, avoiding uneven task distribution.

[0123] The principle behind this trade-off objective function is to balance production efficiency-related indicators (such as the relationship between a certain measure of the PTC2 and STC calculation results and the CAD) and operational fairness-related indicators (i.e., the sum of the operator skill entropy weights) by adjusting the value of μ. When the process-load deviation (CAD) is large, it indicates an imbalance in human-machine collaboration during production. In this case, this objective function can guide the optimization process, maximizing load balance across workstations while improving output and suppressing dynamic detuning between process and load. For example, on a clothing production line, if current production efficiency is low and workstation loads are uneven (large CAD values), while operational fairness is acceptable (relatively uniform distribution of operator skill entropy weights), the value of μ can be appropriately increased to shift the optimization process toward improving production efficiency. This can be achieved by adjusting workstation allocation and optimizing material flow to reduce the CAD value and achieve efficient and stable production. The beneficial effect is that, by introducing the dynamic balancing factor μ, it provides an adjustable optimization direction for the dynamic reconfiguration of the production line topology. The value of μ can be flexibly adjusted to meet diverse production goals in different production scenarios and requirements. When the market demand for products is high, μ can be increased to improve production efficiency; when the requirements for product quality and operational fairness are high, μ can be appropriately reduced to ensure that each operator can be reasonably assigned tasks that match their skills, thereby improving overall production quality and employee satisfaction.

[0124] When the process-load deviation CAD exceeds the preset threshold ε, the production line topology is dynamically reconstructed by establishing a trade-off objective function and performing an optimization solution to improve the overall performance of the production line; if it does not exceed the threshold, the current production status is maintained and continuously monitored.

[0125] Step S2220: Randomly sample in the high-dimensional solution space formed by the adjacency matrix M of the PTC-STC dual-modal workstation chain using the Monte Carlo method, where each sampling point corresponds to a potential sub-workstation reconstruction solution;

[0126] Specifically, because the solution space described by the matrix M is extremely complex, directly finding the optimal sub-station reconfiguration is extremely difficult. Therefore, the Monte Carlo method becomes an effective exploration tool. The Monte Carlo method randomly generates a large number of sampling points in the solution space, each corresponding to a potential sub-station reconfiguration solution. For example, in the clothing production scenario, one potential solution might be to reorder certain processes and reallocate workers; another might be to split a complex process and assign the sub-processes to workers with different skill sets. Through such a large number of random samplings, a wide range of possibilities in the solution space can be fully explored. Leveraging the random nature of the Monte Carlo method, a large number of potential sub-station reconfiguration solutions can be rapidly explored in the complex solution space, providing a rich and diverse initial solution for the subsequent genetic algorithm. This avoids the problem of being trapped in local optimal solutions due to local search and increases the likelihood of finding the global optimal solution.

[0127] Step S2230: Using a genetic algorithm to perform combinatorial optimization on the potential sub-station reconstruction schemes obtained by Monte Carlo sampling to obtain candidate sub-station reconstruction schemes. Substituting the candidate sub-station reconstruction schemes into the trade-off objective function to calculate the J value, the top 10% schemes with the highest J value are selected as the dominant species, and the next generation population is generated after crossover mutation. Multiple iterations are performed until convergence.

[0128] Specifically, the potential sub-station reconstruction plans obtained through Monte Carlo sampling are first used as the initial population. These plans are then substituted into the trade-off objective function to calculate the J value, which measures the performance of the plans in terms of the optimization objective. The population is then ranked according to the J value, and the top 10% of plans with the highest J values are selected as the parents (dominant species). This is because plans with high J values are more effective in balancing production efficiency and operational fairness, and are more likely to guide the evolutionary process towards a more optimal outcome. For example, among a batch of sampled plans, one plan significantly increases production line output while also balancing load across the stations and ensuring good operational fairness. Its high J value will be selected as the parent. The parent plans are then subjected to crossover recombination and local mutation operations. Crossover recombination involves swapping elements, such as the station allocation order, between different parent plans, combining the strengths of each plan to generate new combinations. For example, swapping some elements of the station allocation order 1-2-3 in parent plan A4 with the station allocation order 3-1-2 in parent plan B4 yields new plans C4 and D4. Local variation involves adjusting certain parts of a solution, such as changing the number of sub-stations. This introduces a degree of randomness to prevent the algorithm from prematurely converging to a local optimal solution. For example, by fine-tuning the number of sub-stations in solution C4, we obtain solution C'.

[0129] Afterwards, the J values of the newly generated solutions (such as solution C4, solution D4 and solution C') are recalculated, and these new solutions are merged with the parent generation to form an extended population. Then, the process of selection, crossover and mutation is repeated, and it is continuously iterated. In each iteration, the solutions with higher J values are retained, and the poorer solutions are eliminated, so that the population gradually evolves in a better direction. When the variance of the population J value is less than the preset threshold or the maximum number of iterations is reached, the algorithm stops iterating. The purpose of this step is to gradually improve the sub-station reconstruction scheme through iterative optimization of the genetic algorithm, so that it is continuously optimized in meeting the goals of production efficiency and operational fairness. Its beneficial effect is that the genetic algorithm utilizes the diversity of individuals in the population and searches for a better sub-station reconstruction scheme through continuous evolution. Compared with traditional optimization algorithms, it does not require complex mathematical operations such as derivation of the objective function, and is suitable for handling complex multi-objective optimization problems, such as the dynamic reconstruction problem of the production line topology structure in this patent. Through multiple iterations, it is possible to improve operational fairness while ensuring production efficiency, make the operation of the production line more efficient and stable, and reduce production delays and quality problems caused by unreasonable station allocation.

[0130] Step S2240: output the optimal solution after convergence as the optimal sub-station division plan under the current conditions, and obtain the newly divided sub-stations according to the optimal sub-station division plan.

[0131] Specifically, step S2240 outputs the optimal solution after the genetic algorithm converges, and uses it as the optimal sub-station segmentation scheme under the current conditions, and obtains the newly split sub-stations based on this scheme. When the genetic algorithm meets the convergence conditions (the variance of the population J value is less than the preset threshold or the maximum number of iterations is reached), the optimal solution obtained at this time is an ideal sub-station reconstruction scheme that balances multiple objectives such as production efficiency and operational fairness. For example, after multiple iterations, the genetic algorithm determines that the buttonhole process should be split into three parallel sub-stations and reasonably allocated to operators with matching skills, which greatly improves production efficiency while ensuring operational fairness and achieving a high J value. This splitting and allocation method is the optimal sub-station segmentation scheme under the current conditions.

[0132] Based on this optimal sub-station partitioning scheme, corresponding adjustments are made on the actual production line to create new sub-stations. For example, in a garment production workshop, the optimal sub-station partitioning scheme can be used to rearrange equipment layout, adjust material transfer routes, and assign operators, thereby organizing the production process into the new sub-stations. This step aims to apply the theoretical results obtained by the optimization algorithm to actual production, achieving dynamic reconstruction of the production line topology. The beneficial effect is that by implementing the optimal sub-station partitioning scheme, the production line's operational status can be effectively improved. By rationally dividing complex processes and assigning them to appropriate operators, production efficiency can be improved, the workload at each workstation can be reduced, and production bottlenecks can be avoided. Furthermore, by considering operational fairness, each operator's skills can be fully utilized, improving product quality and employee motivation. For example, in garment production, by rationally dividing the complex sewing process, operators at each sub-station can focus on their respective strengths, which not only speeds up production but also ensures sewing quality, thereby enhancing the economic benefits and competitiveness of the entire production line.

[0133] In the skill correction module, such as Figure 4 As shown, the correction of the operator's skill entropy weight using augmented reality technology for the newly split sub-station includes:

[0134] Step S2310: equip each sub-station with AR smart glasses to collect the pupil focus heat map PFHM and gesture trajectory residual of the novice operator in real time;

[0135] In step S2320, the collected gesture trajectory residual is normalized to obtain the normalized gesture trajectory residual GTR, and the GTR is substituted into the calculation formula of the skill entropy weight. The skill entropy weight of the operator is corrected in real time by introducing the attenuation factor SDF.

[0136] The calculation formula of the SDF is:

[0137] SDF=1-e -k'×GTR ;

[0138] Where k' is the decay rate, GTR is the normalized gesture trajectory residual, and e is the natural logarithm base.

[0139] Specifically, the main purpose of step S2300 is to use augmented reality technology to construct a dynamic skill correction loop for the newly split sub-stations, correcting the operator's skill entropy weights. This improves the operator's skill level, ensures efficient and stable production processes, and enhances product quality. This step is a key component of optimizing operator skills within the intelligent scheduling and control system for the entire garment production line. It is closely related to the previously mentioned steps of constructing the bimodal workstation chain adjacency matrix and determining the optimal sub-station segmentation scheme, and together they contribute to improving the overall performance of the production line.

[0140] During the garment production process, novice operators, due to lack of experience, may have shortcomings in their operational skills, impacting production efficiency and product quality. AR smart glasses, as advanced equipment, can capture critical information related to operator performance in real time. The pupil focus heat map (PFHM) reflects the operator's level of focus on the process path. For example, when sewing a garment with a complex pattern, if the operator's PFHM shows a short focus time on certain key areas, it may indicate insufficient attention to these areas, leading to sewing errors. The hand gesture trajectory residual, represented by the GTR, reflects the degree to which the operator's hand movement deviates from the standard paradigm. For example, when sewing a straight seam, the standard hand gesture trajectory should be smooth and straight. A high GTR value indicates that the operator's actual hand movement deviates from the standard paradigm, potentially resulting in an uneven seam and affecting product quality. By collecting these two data sets, a comprehensive and real-time understanding of the novice operator's actual state during operation is achieved, providing an accurate data foundation for subsequent skill assessment and correction. The beneficial effect of this step is that by monitoring operators' operating behaviors in real time, potential problems can be promptly identified, providing a basis for precisely improving operator skills. For example, in a garment production workshop, managers can use PFHM and GTR data to quickly identify which operators have operational problems and at which stages. This allows them to provide targeted guidance and training, preventing the large number of defective products caused by improper operation and improving production efficiency and product quality.

[0141] Normalizing gesture trajectory residuals ensures comparability between different operators, eliminating data discrepancies caused by factors such as measurement units and operating environment. In actual production, different operators have varying operating habits and proficiency levels, leading to significant variations in raw gesture trajectory residual data. Normalization maps these data to a unified scale, facilitating subsequent calculations and analysis. The normalized GTR is substituted into the skill entropy weight calculation formula and corrected in real time using the decay factor (SDF). The decay factor (SDF) adjusts the skill entropy weight based on the magnitude of the gesture trajectory residual. When the GTR value increases, meaning the operator's hand movements deviate more from the standard paradigm, the SDF value approaches 0, resulting in a larger correction to the skill entropy weight, reducing the operator's current skill entropy weight. Conversely, when the GTR value decreases, the SDF value approaches 1, reducing the correction to the skill entropy weight. For example, if a novice operator has a large residual of the hand gesture trajectory during sewing, and the normalized GTR is 0.8, and the decay rate k′ is 1, the SDF is calculated according to the formula: -1×0.8 ≈0.55, which indicates that the operator's skill entropy weight needs to be lowered and corrected accordingly based on this SDF value to more accurately reflect his current actual skill level.

[0142] The dynamic correction result of the skill entropy weight will be fed back to step S1100, forming a closed-loop iteration of skill feature extraction. The beneficial effect of this closed-loop iterative process is that it can continuously optimize the assessment of the operator's skill level, enabling the production scheduling system to make more reasonable task allocations based on the operator's real-time skill status. For example, after a period of operation, when an operator's skill entropy weight is corrected due to gesture deviation, the production scheduling system can re-evaluate the process tasks that the operator is suitable for based on the new skill entropy weight, and assign more complex and demanding tasks to operators with higher and more stable skill entropy weights, thereby improving the production efficiency and product quality of the entire production line. At the same time, this dynamic correction mechanism also helps to timely discover the changing trend of the operator's skills. If it is found that the skill entropy weight of an operator continues to decline, it may mean that the operator has encountered difficulties or fatigue during the operation. At this time, appropriate measures can be taken in a timely manner, such as adjusting work arrangements or conducting targeted training, to ensure the stability and continuity of the production process.

[0143] In the strategy surface generation module, the dynamic oscillation frequency domain spectrum based on the process feature vector field and process load is used to construct a third-order feature tensor in the three dimensions of operator, workstation, and time. A three-dimensional convolutional gated recurrent unit network is used to model the production cycle stability. The output multi-scale strategy surface includes:

[0144] Step S3100 , fusing the process feature vector field, the dynamic oscillation frequency domain spectrum of the process load, and the pupil focus heat map to construct a third-order feature tensor;

[0145] The core objective of step S3100 is to fuse the process characteristic vector field, the dynamic oscillation frequency domain spectrum of the process load, and the pupil focus heat map to construct a third-order characteristic tensor. This provides the data foundation for subsequent accurate modeling of production cycle stability and comprehensive analysis and optimization of the production process. This step integrates multiple key information from the production process, contributing to a deeper understanding and optimization of the operating mechanisms of the garment production line.

[0146] Further, if Figure 5 As shown, step S3100 includes:

[0147] Step S3110: extracting focus data from the pupil focus heat map and extracting bottleneck drift characteristic factors from the dynamic oscillation frequency domain spectrum of the process load;

[0148] Specifically, focus data is an indicator used to quantify the operator's attention to the process path during the operation. In the clothing production scenario, it reflects the degree of concentration of the operator's eyes on different process links when performing operations such as sewing. For example, when sewing complex patterns or key parts (such as collars, cuffs, etc.), high focus means that the operator can concentrate and is more likely to ensure the accuracy and high quality of the operation; conversely, low focus may lead to operational errors and affect product quality. Focus data is extracted from the pupil focus heat map, mainly using image processing and data analysis methods. First, the collected pupil focus heat map is preprocessed. Since the original image may have noise, uneven lighting and other problems, which will affect the accuracy of data extraction, it is necessary to perform grayscale processing to convert the color image into a grayscale image to simplify subsequent calculations. For example, through a specific grayscale conversion formula, the image in the RGB color mode is converted into a grayscale image with only brightness values, making the image data easier to process. At the same time, filtering algorithms are used to remove noise. A common method is Gaussian filtering, which smoothes the image and reduces noise interference by performing a weighted average of each pixel and its neighboring pixels. Next, a region of interest (ROI) is determined. In garment production, the process path area relevant to the current operation is defined as the ROI based on the specific process flow. For example, if the primary operation in a garment sewing process is the hem, the heat map area corresponding to the hem is designated as the ROI. This eliminates interference from other irrelevant areas and allows for more accurate analysis of the focus related to the current process. Focus data is then calculated based on the identified ROI. A common calculation method is based on pixel value statistics. In pupil focus heat maps, different brightness levels typically represent different levels of focus. Higher brightness levels often indicate a longer pupil focus time or a higher degree of focus. Focus data is obtained by counting the number of pixels in different brightness ranges within the ROI and performing a weighted calculation based on preset weights. Focus data calculated in this way can more accurately quantify the operator's attention to the process path.

[0149] The method for extracting the bottleneck drift characteristic factor from the dynamic oscillation frequency domain spectrum of the process load is:

[0150] DOFS j Perform wavelet packet decomposition and extract the energy value as the bottleneck drift characteristic factor BDFF j .

[0151] Wavelet packet decomposition is a signal processing technology that can decompose the signal into different frequency bands, so as to analyze the signal characteristics in more detail. In this embodiment, the energy value of the 3-5Hz frequency band is selected as the bottleneck drift feature factor BDFF. jIn garment production, process loads change as the production process progresses. By analyzing the energy values at different frequencies, we can capture the characteristic information of load changes. The 3-5 Hz frequency band was chosen because the energy value changes within this frequency band are closely related to the emergence of production bottlenecks. Bottleneck drift characteristic factor BDFF j It reflects the possibility of the workstation becoming a bottleneck. The larger the amplitude, the more out of sync the load level of the workstation is with the production line rhythm, and the higher the probability of a production bottleneck. For example, when the BDFF of a workstation is j A sudden increase in amplitude may indicate that a workstation is about to become, or has become, a production bottleneck, impacting the production rhythm of the entire production line. This step is beneficial because, by extracting focus data and bottleneck drift characteristic factors, it is possible to obtain important information about the production process from two key aspects: operator status and process load. This provides critical data support for subsequent analysis of production rhythm stability. By monitoring and analyzing this data, potential production issues, such as operator inattention or potential bottlenecks at a workstation, can be promptly identified, allowing proactive adjustments to ensure smooth production line operation.

[0152] Step S3120 , performing three-dimensional high-order interpolation on three types of heterogeneous process data: the skill entropy weight of the process feature vector field, the bottleneck drift characteristic factor of the dynamic oscillation frequency domain spectrum of the process load, and the focus data of the pupil focus heat map;

[0153] Specifically, the skill entropy weights of the process feature vector field, the bottleneck drift characteristic factors of the dynamic oscillation frequency domain spectrum of the process load, and the focus data of the pupil focus heatmap differ in their collection and representation formats, making them heterogeneous data. For example, the skill entropy weights are numerical values reflecting the operator's skill level, obtained through a series of calculations; the dynamic oscillation frequency domain spectrum of the process load is frequency-domain data; and the focus data of the pupil focus heatmap is a measurement based on visual perception. The sampling density and resolution of these data may also differ across time and space, making it difficult to directly use them for comprehensive analysis. Three-dimensional high-order interpolation is a mathematical method that expands the spatial resolution of tensors, aligning the sampling densities of these three heterogeneous data types across the operator, workstation, and time dimensions. This approach aims to integrate these different types of data into a unified framework, facilitating subsequent analysis and processing. For example, assuming the skill entropy weights have a larger sampling interval in the time dimension, while the dynamic oscillation frequency domain spectrum of the process load has a smaller sampling interval in this dimension, three-dimensional high-order interpolation can be used to adjust their sampling densities in the time dimension to match each other. This step has the beneficial effect of enabling unified processing of heterogeneous data, laying the foundation for the subsequent construction of more accurate production process models. By ensuring consistency across time and space, the interrelationships between various factors in the production process can be more comprehensively and accurately reflected, improving the accuracy and reliability of analysis, and contributing to a deeper understanding of complex phenomena in the production process, providing strong support for optimizing production scheduling and control.

[0154] Step S3130: Use Tucker decomposition algorithm to align and reduce the dimension of the three types of heterogeneous process data after three-dimensional high-order interpolation, project them into a shared low-dimensional latent space, and obtain the compressed third-order feature tensor CFT∈R N×T×D , where N, T, and D represent the data lengths of the three dimensions of operator, workstation, and time, respectively.

[0155] Specifically, the Tucker decomposition algorithm is a multilinear algebra method that can reduce the dimensionality of high-dimensional data while retaining the main features of the data. In this step, although the data after three-dimensional high-order interpolation tends to be consistent in sampling density, it is still high-dimensional data. Directly processing these high-dimensional data will face problems such as high computational complexity and data redundancy. The Tucker decomposition algorithm finds a suitable low-dimensional subspace and projects the heterogeneous process data after three-dimensional high-order interpolation into this shared low-dimensional latent space to obtain a compressed third-order characteristic tensor CFT. This not only reduces the dimension of the data and the amount of calculation, but also removes redundant information in the data and highlights the key features of the data. For example, in actual production, there may be some data features that have little effect on the stability of the production cycle. Through Tucker decomposition, these unimportant information can be removed, and only the parts that are important for analysis and modeling can be retained. The beneficial effects of this step are, on the one hand, reducing the complexity of data processing and improving the efficiency of subsequent modeling and analysis; on the other hand, by removing redundant information, it more centrally reflects the interactive modulation mechanism of human-machine-time in the production process, making multi-scale slices more suitable for modeling needs at different decision-making levels. For example, at the production planning level, the compressed tensor can be used to analyze the overall trend of the production process from a macro perspective; at the production site control level, the production status of specific operators, workstations, and time points can be monitored from a micro perspective, providing more targeted and effective support for production decision-making, thereby optimizing the entire production process.

[0156] Step S3200: Building a three-dimensional convolutional gated recurrent unit network to extract the production cycle stability PSE';

[0157] Design a three-dimensional convolutional gated recurrent unit as a basic component, build a three-dimensional convolutional gated recurrent unit network, input the third-order feature tensor CFT into the built three-dimensional convolutional gated recurrent unit network, extract the process stability entropy value PSE, and use the process stability entropy value as the production cycle stability PSE';

[0158] The method for extracting the process stability entropy value PSE is:

[0159]

[0160] Of which: PSE t is the process stability entropy value at the current moment, t is the current moment, τ is the time window length, σ(·) is the Sigmoid activation function, CFT t-τ:t represents a slice on the time dimension from time t-τ to time t, represents the element-by-element addition of tensors; Conv3D(·) is a three-dimensional convolution operation, which is used to extract the local correlation features of CFT in the three dimensions of operator-station-time; GRU(·) is a gated recurrent unit, which is used to explore the long-range dependency of CFT in the time dimension.

[0161] Specifically, step S3200 builds a three-dimensional convolutional gated recurrent unit network to extract the production cycle stability (PSE') to quantitatively assess the stability of the production process and provide a basis for subsequent strategy implementation. This step is a key component of the intelligent scheduling and control system for monitoring and optimizing production cycle time. By calculating and analyzing the process stability entropy, dynamic control of the production process is achieved.

[0162] When constructing a three-dimensional convolutional gated recurrent unit network, a three-dimensional convolutional gated recurrent unit (3D-ConvGRU) is designed as a basic component. The three-dimensional convolution operation (Conv3D) is used to extract local correlation features of the third-order feature tensor (CFT) in the three dimensions of operator, workstation, and time. For example, in garment production, the changes in the operating patterns of different operators at specific workstations over time (such as the speed and stitching changes of different workers when sewing the same part) and the impact of different workstations on the production rhythm at different time points (such as the impact of fluctuations in the processing time of a certain workstation on the rhythm of the entire production line) are all contained in the CFT tensor and can be effectively extracted through the three-dimensional convolution operation.

[0163] The Gated Recurrent Unit (GRU) is used to exploit the long-range dependencies of the CFT over time. In the garment production process, the stability of the production cycle is often affected by operations and load conditions at multiple previous time points. The GRU can remember this historical information, avoiding gradient vanishing or gradient exploding problems, and thus better capture long-term trends in the production process. For example, if a workstation has been under high load for several time periods, the GRU can remember this information and consider this long-term impact in subsequent calculations, thereby more accurately assessing the stability of the current production cycle.

[0164] The three-dimensional convolutional gated recurrent unit network built by inputting the third-order feature tensor CFT, Extract the process stability entropy PSE and use it as the production cycle stability PSE'. t represents the process stability entropy at the current moment, t is the current moment, and τ is the time window length, which determines the length of the historical period considered when calculating PSE. The Sigmoid activation function σ(·) is used to map the calculation result to the range of 0-1, which facilitates the quantitative evaluation of production cycle stability. CFT t-τ:tRepresents a slice along the time dimension from time t-τ to time t. By processing the CFT data within this time period, recent production conditions can be comprehensively considered. ⊕ represents element-by-element addition of tensors, fusing the results of the three-dimensional convolution operation and the gated recurrent unit calculations to more comprehensively reflect the dynamic detuning of the production cycle.

[0165] For example, on a clothing production line, by processing CFT data over a period of time (e.g., from time t-10 to time t), a three-dimensional convolution operation extracts the local operating patterns and interrelationship features of each workstation during this period. The GRU then memorizes the tempo of these operating patterns over time. The two results are summed and mapped using a sigmoid function to obtain the production cycle stability (PSE') at the current time t. A PSE' value close to 1 indicates a relatively stable production cycle and good coordination between various links. A lower PSE' value, such as one below a preset stability threshold η, indicates instability in the production process and may require appropriate adjustments.

[0166] This step has the beneficial effect of quantifying production cycle stability, providing a clear indicator for production line management and optimization. Based on the PSE' value, production managers can promptly identify unstable links in the production process, such as irregular operations at a certain workstation leading to fluctuations in production cycle. They can then take targeted measures, such as adjusting personnel arrangements and optimizing process flows, to ensure stable production line operation and improve production efficiency and product quality. Furthermore, this indicator helps predict potential production problems, enabling preventive measures and adjustments to mitigate losses caused by production instability.

[0167] Step S3300: output a multi-scale strategy surface.

[0168] Furthermore, step S3300 includes:

[0169] Step S3310: Extract the load value of each workstation from the dynamic oscillation frequency domain spectrum of the process load. Calculate the workstation load balance degree (LBD) and operator skill balance degree (SBD) of the entire production line based on the load value of each workstation and the operator skill entropy weight corresponding to the newly split sub-workstations.

[0170]

[0171] Where: std(·) is the standard deviation, L i is the load value of the i-th workstation.

[0172] Specifically, the calculation formula of the workstation load balance degree LBD is std(L1, L2, ..., L n ) represents the load value L of each workstation iThe degree of discreteness (i ranges from 1 to n, where n is the total number of workstations). The greater the degree of discreteness, the greater the difference between the load values of each workstation. It represents the average value of all workstation load values, which reflects the average load level of the entire production line. This formula measures the degree of balance of workstation loads by subtracting the ratio of the standard deviation to the average value from 1. When the difference between the load values of each workstation is small, the standard deviation is small, and the value of LBD will be close to 1, indicating that the workstation load balance is high; conversely, when the difference between the load values of each workstation is large, the standard deviation is large, and the value of LBD will decrease, indicating that the workstation load balance is low. In the calculation formula of the operating skill balance SBD, std(SEW1, SEW2,…, SEW m ) represents the skill entropy weight SEW of each operator i The degree of discreteness (i ranges from 1 to m, where m is the total number of operators and m=n, because one workstation corresponds to one operator) reflects the difference between the skill entropy weights of different operators. This represents the average of all operator skill entropy weights, reflecting the overall skill level. Similar to workstation load balance, this formula measures the balance of operator skills by subtracting the ratio of the standard deviation to the mean (1). When the differences in skill entropy weights are small, the SBD value is close to 1, indicating a high degree of skill balance. When the differences are large, the SBD value decreases, indicating a low degree of skill balance.

[0173] The beneficial effect of calculating these two balance measures is that they can quantify the balance between workstation load and operator skill within a production line. In garment production, uneven workstation load can lead to overcrowding at some stations, creating production bottlenecks and impacting overall production efficiency. Uneven operator skill levels can also lead to inconsistent product quality due to significant differences in operator skill levels in some processes. By calculating LBD and SBD, we can intuitively understand the balance between these two aspects of the production line, providing data for subsequent optimized scheduling. For example, a low LBD indicates significant disparity in workstation load. In this case, targeted adjustments can be made to production task allocation, shifting some overloaded tasks to less-loaded stations to improve overall production efficiency. A low SBD indicates significant disparity in operator skill levels. Therefore, skills training or staffing adjustments can be implemented to achieve a more balanced operator skill level across stations, thereby ensuring consistent product quality.

[0174] Step S3320: Use the production cycle stability PSE', workstation load balance LBD, and operation skill balance SBD as three coordinate axes to construct a three-dimensional decision space. Project the strategy parameters into the three-dimensional decision space to form a multi-scale strategy surface. The strategy parameters include the dimensional weight λ of the dynamic process parameters in the process feature vector field. i”, the dynamic balance factor μ of production efficiency and operational fairness in the trade-off objective function of the bimodal workstation chain reconstruction, the attenuation factor α1, the gain factor β1, and the attenuation rate k' in the calculation formula of the attenuation factor SDF.

[0175] Specifically, when constructing a three-dimensional decision space, the production cycle stability (PSE') quantifies the degree of dynamic detuning of the production cycle, reflecting the stability of the coordinated working of various factors in the production process. The workstation load balance (LBD) and operator skill balance (SBD) further describe the status of the production line from the perspectives of load distribution and skill balance. Using these indicators as coordinate axes to construct a three-dimensional decision space provides a more comprehensive picture of the production line's operating status. For example, in this three-dimensional space, a point (PSE', LBD, SBD) represents the comprehensive status of the production line at a given moment in terms of production cycle stability, workstation load balance, and operator skill balance.

[0176] Projecting the policy parameters onto the three-dimensional decision space forms a multi-scale policy surface. Surface slices at different scales correspond to different production rhythm patterns. This is because different combinations of policy parameters have varying impacts on production line operations, leading to different production rhythms. For example, the dynamic balance factor μ is used to balance production efficiency and operational fairness. When μ is large, the trade-off objective function for bimodal workstation chain reconstruction favors improved production efficiency; when μ is small, operational fairness is prioritized. The attenuation factor α1 is used to modify the physical topology chain adjacency matrix and adjust the weights of workstations with excessively long material transfer times. When α1 is small, the weight adjustment for physical edges with excessively long transfer times is larger, which is more conducive to optimizing material flow; conversely, the adjustment is smaller. The gain factor β1 is used to increase the weights of physical edges at workstations with insufficient skill complementarity, optimizing skill matching between workstations. A larger β1 value results in a more pronounced weight increase for workstations with insufficient skill complementarity. The decay rate k' dynamically adjusts the decay speed of the skill entropy weight in AR correction, and associates the real-time correction of skill entropy with the gesture trajectory residual (GTR). The larger the k', the faster the skill entropy weight decays with the gesture trajectory residual, and the faster the feedback adjustment of the operator's skill shortcomings.

[0177] By applying a clustering algorithm to perform multi-scale segmentation of the strategy surface, strategy parameter combinations for different production scenarios can be classified. Surface slices at different scales correspond to different production rhythm patterns, enabling them to meet the diverse demands of high-, medium-, and low-end production orders. For example, for high-end production orders, a balance between product quality and operational skill may be more important. In this case, a surface slice corresponding to a strategy parameter combination with a high operational skill balance (SBD) and a high production rhythm stability (PSE') can be selected. For low-end production orders, production efficiency may be more important. The surface slice corresponding to a strategy parameter combination with a dominant production efficiency-related factor (such as a large dynamic balance factor μ) can be selected. The construction and application of this multi-scale strategy surface allows for the rapid identification of appropriate strategy parameter combinations based on the demands of different production orders, enabling optimized production line scheduling, improving production flexibility and adaptability, and mitigating problems such as low production efficiency or unstable product quality caused by changes in production orders.

[0178] Example 2

[0179] This embodiment provides an intelligent scheduling control method for a clothing production line based on the embodiment 1, such as Figure 6 Shown, including:

[0180] Step S1000: Collect the operator's dynamic process parameters, construct a process characteristic vector field based on the operator's dynamic process parameters, and calculate the skill entropy weight of each operator; collect the working parameters of each workstation in real time, and construct the dynamic oscillation frequency domain spectrum of the process load based on the working parameters of each workstation; calculate the process-load deviation based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load;

[0181] Furthermore, step S1000 includes:

[0182] Step S1100: collecting the dynamic process parameters of the operators, constructing a process feature vector field based on the dynamic process parameters of the operators, and calculating the skill entropy weight of each operator;

[0183] Furthermore, step S1100 includes:

[0184] Step S1110 , collecting the operator's dynamic process parameters, wherein the dynamic process parameters include the curve trajectory dwell time ratio STDR and the seam quantization deviation QSD;

[0185] Step S1120 , reducing the dimension of the operator's dynamic process parameters and mapping them to the process feature vector field, and calculating the skill entropy weight of each operator.

[0186] Furthermore, step S1120 includes:

[0187] Step S1121, performing Min-Max normalization processing on the dynamic process parameters of each operator, mapping them to the interval [0, 1], and obtaining normalized dynamic process parameters, wherein the normalized dynamic process parameters include the normalized curve trajectory dwell time ratio STDR' and the seam quantization deviation QSD';

[0188] Step S1122: according to the normalized dynamic process parameters and the dimension weight λ of the dynamic process parameters in the preset process feature vector field i” , calculate the skill feature vector TFV of each operator i ;where i' is the index of the dynamic process parameter dimension and i is the index of the operator;

[0189] Step S1123: skill feature vector TFV of each operator i Normalize the modulus length to obtain the unit skill feature vector UTFV i ;

[0190] Step S1124: Based on the unit skill feature vector UTFV i , calculate the skill entropy weight SEW of each operator i .

[0191] Step S1200: collecting the working parameters of each workstation in real time, and constructing a dynamic oscillation frequency domain spectrum of the process load based on the working parameters of each workstation;

[0192] Furthermore, step S1200 includes:

[0193] Step S1210: Real-time acquisition of working parameters of each workstation, including the workstation dwell time-varying gradient TDG, material rheological curvature MRC, and load value; one workstation corresponds to one operator;

[0194] Step S1220, constructing a dynamic oscillation frequency domain spectrum of the process load of each workstation based on the working condition parameters;

[0195] Set the station dwell time-varying gradient of station j to TDG j and the material rheological curvature is MRC j , based on TDG j and MRC j Construct the load-logistics pair (TDG) of station j j ,MRC j ), the load-logistics binary (TDG j ,MRC j ) is projected into the amplitude-frequency domain to obtain the dynamic oscillation frequency domain spectrum DOFS of the process load of station j j .

[0196] Step S1300 : Calculate the process-load deviation CAD based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load.

[0197] Furthermore, step S1300 includes:

[0198] Step S1310 , aligning the skill entropy weight of each operator in the process feature vector field with the dynamic oscillation frequency domain spectrum of the process load of the corresponding workstation to construct an interaction matrix between operator skill and process load;

[0199] Step S1320 , extracting the entropy weight time series of each operator's operation skill and the oscillation frequency domain time series of the process load of each workstation from the operation skill and process load interaction matrix;

[0200] Step S1330 , calculating the cumulative error metric between the operation skill entropy weight time series and the process load oscillation frequency domain time series, and defining the cumulative error metric as the process-load deviation CAD.

[0201] Step S2000: construct the initial adjacency matrix PTC0 of the physical topology chain of the production line and the adjacency matrix STC of the operator skill topology chain, modify the initial adjacency matrix PTC0 of the physical topology chain of the production line, and generate the PTC-STC bimodal workstation chain adjacency matrix; when the process-load deviation CAD exceeds the preset deviation threshold ε, trigger the dynamic reconstruction condition of the production line topology structure; when the dynamic reconstruction condition of the production line topology structure is triggered, based on the PTC-STC bimodal workstation chain adjacency matrix, obtain the optimal sub-workstation segmentation scheme, and obtain the newly split sub-workstation according to the optimal sub-workstation segmentation scheme; for the newly split sub-workstation, use augmented reality technology to correct the skill entropy weight of the operator;

[0202] Furthermore, step S2000 includes:

[0203] Step S2100: constructing the production line's physical topology chain initial adjacency matrix PTC0 and the operator skill topology chain adjacency matrix STC, and modifying the production line's physical topology chain initial adjacency matrix PTC0 to generate a PTC-STC dual-modal workstation chain adjacency matrix M;

[0204] Furthermore, step S2100 includes:

[0205] Step S2110: Obtain the physical location relationship of the equipment on the clothing production line. Based on the physical location relationship of the equipment, construct the initial adjacency matrix PTC0 of the physical topology chain of the production line. The matrix element PTC0(i1,j1) represents the direct material arrival time between workstation i1 and workstation j1. The initial adjacency matrix PTC0 of the physical topology chain of the production line contains physical edges, and the direct material arrival time is used as the physical edge weight.

[0206] Step S2120: sort the skill entropy weights of each operator across workstations and generate an operator skill topology chain adjacency matrix STC based on the principle of skill complementarity. The matrix element STC(p,q) represents the skill substitution difficulty between operator p and operator q. The operator skill topology chain adjacency matrix STC includes virtual edges, and the skill substitution difficulty is used as the virtual edge weight.

[0207] Step S2130: Based on the principle of material transfer efficiency balance, the initial adjacency matrix PTC0 of the physical topology chain of the production line is modified to obtain a first physical topology matrix PTC1;

[0208] Step S2140: Based on the skill complement optimization principle, the first physical topology matrix PTC1 is modified to obtain the second physical topology matrix PTC2;

[0209] Step S2150: Multiply the second physical topology matrix PTC2 and the operator skill topology chain adjacency matrix STC bit by bit to obtain the PTC-STC dual-modal workstation chain adjacency matrix M.

[0210] Step S2200: When the process-load deviation CAD exceeds a preset deviation threshold ε, a dynamic reconstruction condition of the production line topology is triggered. When the dynamic reconstruction condition of the production line topology is triggered, a trade-off objective function for bimodal station chain reconstruction is established. Based on the trade-off objective function and the PTC-STC bimodal station chain adjacency matrix, an optimal sub-station segmentation scheme is obtained, and newly split sub-stations are obtained according to the optimal sub-station segmentation scheme.

[0211] Furthermore, step S2200 includes:

[0212] Step S2210: When the dynamic reconstruction condition of the production line topology is triggered, a dynamic balance factor μ between production efficiency and operational fairness is set, and a trade-off objective function for the reconstruction of the bimodal workstation chain is established based on μ;

[0213] Step S2220: Randomly sample in the high-dimensional solution space formed by the adjacency matrix M of the PTC-STC dual-modal workstation chain using the Monte Carlo method, where each sampling point corresponds to a potential sub-workstation reconstruction solution;

[0214] Step S2230: Using a genetic algorithm to perform combinatorial optimization on the potential sub-station reconstruction schemes obtained by Monte Carlo sampling to obtain candidate sub-station reconstruction schemes. Substituting the candidate sub-station reconstruction schemes into the trade-off objective function to calculate the J value, the top 10% schemes with the highest J value are selected as the dominant species, and the next generation population is generated after crossover mutation. Multiple iterations are performed until convergence.

[0215] Step S2240: output the optimal solution after convergence as the optimal sub-station division plan under the current conditions, and obtain the newly divided sub-stations according to the optimal sub-station division plan.

[0216] Step S2300: For the newly split sub-station, the operator's skill entropy weight is corrected using augmented reality technology;

[0217] Furthermore, step S2300 includes:

[0218] Step S2310: equip each sub-station with AR smart glasses to collect the pupil focus heat map PFHM and gesture trajectory residual of the novice operator in real time;

[0219] In step S2320, the collected gesture trajectory residual is normalized to obtain the normalized gesture trajectory residual GTR, and the GTR is substituted into the calculation formula of the skill entropy weight. The skill entropy weight of the operator is corrected in real time by introducing the attenuation factor SDF.

[0220] In step S3000, based on the process feature vector field and the dynamic oscillation frequency domain spectrum of the process load, a third-order feature tensor in the three dimensions of operator, workstation, and time is constructed, and a three-dimensional convolutional gated recurrent unit network is applied to model the production rhythm stability and output a multi-scale strategy surface.

[0221] Furthermore, step S3000 includes:

[0222] Step S3100 , fusing the process feature vector field, the dynamic oscillation frequency domain spectrum of the process load, and the pupil focus heat map to construct a third-order feature tensor;

[0223] Step S3200: Building a three-dimensional convolutional gated recurrent unit network to extract the production cycle stability PSE';

[0224] Step S3300: output a multi-scale strategy surface.

[0225] The methods and systems of the present application may be implemented in many ways. For example, the methods and systems of the present application may be implemented using software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps used in the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified.

[0226] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0227] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific 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 shall be included within the scope of protection of the present invention.

Claims

1. Intelligent scheduling and control system for clothing production line, characterized by: The system comprises: Deviation calculation module: used to collect the operator's dynamic process parameters, construct the process characteristic vector field based on the operator's dynamic process parameters, and calculate the skill entropy weight of each operator; collect the working parameters of each workstation in real time, and construct the dynamic oscillation frequency domain spectrum of the process load based on the working parameters of each workstation; calculate the process-load deviation based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load; Topology optimization module: used to construct the initial adjacency matrix PTC0 of the production line's physical topology chain and the adjacency matrix STC of the operator skill topology chain, and to modify the initial adjacency matrix PTC0 of the production line's physical topology chain to generate the PTC-STC bimodal workstation chain adjacency matrix. When the process-load deviation CAD exceeds the preset deviation threshold ε, the dynamic reconstruction condition of the production line topology structure is triggered. When the dynamic reconstruction condition of the production line topology structure is triggered, the optimal sub-workstation segmentation scheme is obtained based on the PTC-STC bimodal workstation chain adjacency matrix, and the newly split sub-workstations are obtained according to the optimal sub-workstation segmentation scheme. Skill correction module: For the newly split sub-stations, the operator's skill entropy weight is corrected using augmented reality technology; Strategy surface generation module: Based on the process feature vector field and the dynamic oscillation frequency domain spectrum of the process load, it constructs a third-order feature tensor in the three dimensions of operator, workstation, and time. It applies a three-dimensional convolutional gated recurrent unit network to model the production rhythm stability and output a multi-scale strategy surface.

2. The intelligent scheduling and control system for a clothing production line according to claim 1 is characterized in that: The dynamic process parameters include curve trajectory dwell time ratio STDR and seam track quantization deviation QSD; Calculating the skill entropy weight of each operator includes: Performing Min-Max normalization processing on the dynamic process parameters of each operator and mapping them to the interval [0, 1] to obtain normalized dynamic process parameters, wherein the normalized dynamic process parameters include the normalized curve trajectory dwell time ratio STDR' and the seam quantization deviation QSD'; According to the normalized dynamic process parameters and the dimension weight λ of the dynamic process parameters in the preset process feature vector field i” , calculate the skill feature vector TFV of each operator i ;where i' is the index of the dynamic process parameter dimension and i is the index of the operator; For each operator's skill feature vector TFV i Normalize the modulus length to obtain the unit skill feature vector UTFV i ; Based on the unit skill feature vector UTFV i , calculate the skill entropy weight SEW of each operator i .

3. The intelligent scheduling and control system for a clothing production line according to claim 2 is characterized in that: The working condition parameters include the station dwell time-varying gradient TDG, the material rheological curvature MRC and the load value; one station corresponds to one operator; The dynamic oscillation frequency domain spectrum of the construction process load includes: Set the station dwell time-varying gradient of station j to TDG j and the material rheological curvature is MRC j , based on TDG j and MRC j Construct the load-logistics pair (TDG) of station j j ,MRC j ), the load-logistics binary (TDG j ,MRC j ) is projected into the amplitude-frequency domain to obtain the dynamic oscillation frequency domain spectrum DOFS of the process load of station j j , where 1≤j≤n, and n is the total number of workstations.

4. The intelligent scheduling and control system for a clothing production line according to claim 3 is characterized in that: Calculation process-load deviation CAD includes: The skill entropy weight of each operator in the process feature vector field is time-series aligned with the dynamic oscillation frequency domain spectrum of the process load of its corresponding workstation to construct the interaction matrix between operator skill and process load. Extract the entropy weight time series of each operator's operation skill and the oscillation frequency domain time series of the process load of each workstation from the interaction matrix of operation skill and process load; The cumulative error metric between the operation skill entropy weight time series and the process load oscillation frequency domain time series is calculated, and the cumulative error metric is defined as the process-load deviation CAD.

5. The intelligent scheduling and control system for a clothing production line according to claim 4 is characterized in that: The physical topology chain initial adjacency matrix PTC0 of the production line is constructed as follows: The physical location relationship of the equipment on the clothing production line is obtained. Based on the physical location relationship of the equipment, the initial adjacency matrix PTC0 of the physical topology chain of the production line is constructed. The matrix element PTC0(i1,j1) represents the material direct arrival time between workstation i1 and workstation j1. The initial adjacency matrix PTC0 of the physical topology chain of the production line contains physical edges, and the material direct arrival time is used as the physical edge weight.

6. The intelligent scheduling and control system for a clothing production line according to claim 5, characterized in that: Constructing the operator skill topology chain adjacency matrix STC includes: The skill entropy weights of each operator are sorted across workstations, and the operator skill topology chain adjacency matrix STC is generated according to the principle of skill complementarity. The matrix element STC(p,q) represents the skill substitution difficulty between operator p and operator q. The operator skill topology chain adjacency matrix STC contains virtual edges, and the skill substitution difficulty is used as the virtual edge weight.

7. The intelligent scheduling and control system for a clothing production line according to claim 6, characterized in that: The modification of the initial adjacency matrix PTC0 of the physical topology chain of the production line includes: Based on the principle of material transfer efficiency balance, the initial adjacency matrix PTC0 of the physical topology chain of the production line is modified to obtain the first physical topology matrix PTC1; based on the principle of skill complementary optimization, the first physical topology matrix PTC1 is modified to obtain the second physical topology matrix PTC2.

8. The intelligent dispatching and control system for a clothing production line according to claim 7, characterized in that: Obtaining the first physical topology matrix PTC1 includes: Calculate the average material transfer time t1 between any two workstations in the initial adjacency matrix PTC0 of the physical topology chain of the production line, and define the physical edges with a transfer time greater than 1.5t1 as physical edges with excessive transfer time; for each physical edge (i2, j2) with excessive transfer time, i2 and j2 represent the workstations corresponding to the endpoints of the physical edge with excessive transfer time; assuming that workstation i2 corresponds to operator i3, and workstation j2 corresponds to operator j3, find the virtual edge with operator i3 or operator j3 as the endpoint and the smallest virtual edge weight in the operator skill topology chain adjacency matrix STC, and define it as the virtual edge with the lowest skill substitution difficulty; multiply the physical edge weight in the initial adjacency matrix PTC0 of the physical topology chain of the production line corresponding to the virtual edge with the lowest skill substitution difficulty by the attenuation factor α1 to obtain the first physical topology matrix PTC1; where 0<α1<1.

9. The intelligent dispatching and control system for a clothing production line according to claim 8, characterized in that: The optimal sub-station segmentation scheme includes: When the dynamic reconstruction condition of the production line topology is triggered, the dynamic balance factor μ between production efficiency and operational fairness is set, and the trade-off objective function of the bimodal workstation chain reconstruction is established based on μ. The Monte Carlo method is used to randomly sample in the high-dimensional solution space formed by the adjacency matrix M of the PTC-STC dual-modal workstation chain. Each sampling point corresponds to a potential sub-workstation reconstruction solution. A genetic algorithm is used to combinatorially optimize the potential sub-station reconstruction schemes obtained by Monte Carlo sampling to obtain candidate sub-station reconstruction schemes. The candidate sub-station reconstruction schemes are substituted into the trade-off objective function to calculate the J value. The top 10% schemes with the highest J value are selected as dominant species. After crossover mutation, the next generation population is generated and multiple iterations are carried out until convergence. Output the optimal solution after convergence as the optimal sub-station division plan under the current conditions.

10. An intelligent scheduling and control method for a clothing production line, based on the intelligent scheduling and control system for a clothing production line according to any one of claims 1 to 9, characterized in that: The method comprises: The operator's dynamic process parameters are collected, and based on these parameters, a process characteristic vector field is constructed to calculate the skill entropy weight of each operator. The working parameters of each workstation are collected in real time, and based on these parameters, a dynamic oscillation frequency domain spectrum of the process load is constructed. The process-load deviation is calculated based on the process characteristic vector field and the dynamic oscillation frequency domain spectrum of the process load. The initial adjacency matrix PTC0 of the production line's physical topology chain and the adjacency matrix STC of the operator's skill topology chain are constructed, and the initial adjacency matrix PTC0 of the production line's physical topology chain is modified to generate a PTC-STC dual-modal workstation chain adjacency matrix. When the process-load deviation CAD exceeds a preset deviation threshold ε, the dynamic reconstruction condition of the production line topology structure is triggered. When the dynamic reconstruction condition of the production line topology structure is triggered, the optimal sub-workstation segmentation scheme is obtained based on the PTC-STC dual-modal workstation chain adjacency matrix, and the newly split sub-workstations are obtained according to the optimal sub-workstation segmentation scheme. For the newly split sub-stations, the operator's skill entropy weight is corrected using augmented reality technology; Based on the dynamic oscillation frequency domain spectrum of the process characteristic vector field and the process load, a third-order feature tensor in the three dimensions of operator, workstation, and time is constructed. A three-dimensional convolutional gated recurrent unit network is applied to model the production rhythm stability and output a multi-scale strategy surface.

Images