AI-assisted garment production routing system

Abstract

The present application relates to the technical field of intelligent manufacturing and industrial engineering, and particularly relates to an AI-assisted clothing production scheduling system, comprising: manufacturing network state perception, man-machine physical feature extraction, non-stationary service rate solution, mold change feature equivalent mapping, collaborative Lyapunov optimization and flexible closed-loop execution module. The system obtains the true in-process product queue length, semi-finished product arrival rate and target and current tooling features of the candidate workstations, extracts the number of processing pieces and continuous operation time of the operators; based on the learning curve and fatigue model, the non-stationary service rate expectation value is solved; through feature comparison, the mold change penalty coefficient is obtained and converted into an equivalent virtual queue; the virtual queue and the real queue are superimposed to construct a collaborative Lyapunov function, and a control equation is constructed by introducing the non-stationary service rate to solve the target optimal workstation, and material and material preparation orders are issued and state feedback is performed. The present application takes the mold change time and personnel efficiency into the queuing optimization decision.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial engineering technology, and in particular to an AI-assisted garment production line layout system. Background Technology

[0002] As the garment manufacturing industry shifts towards a small-batch, fast-response model, flexible mixed-flow production lines face extremely high-frequency material and style changes. Existing dynamic line scheduling systems typically allocate materials based on rules such as the shortest queue priority in queuing theory. However, existing line control logic has serious inherent flaws: it abstracts the underlying operational units into standard black boxes with constant throughput, completely stripping away the constraints of the physical changeover time of the underlying equipment and the non-stationary learning curve constraints of the operators.

[0003] When dealing with mixed-style production changes, conventional scheduling systems often blindly assign new style materials to the workstation with the shortest current network queue, ignoring the fact that the candidate workstation receiving the material is likely to lack matching physical sewing tools, requiring lengthy downtime for physical mold changes. They also overlook the learning curve and non-linear ramp-up period that operators must go through when faced with a new sewing process. This severe disconnect between the algorithm's pre-set constant ideal capacity and the non-linear output of the machines on site causes the so-called optimal network routing instructions issued by the system to induce a highly concealed mirage of capacity during the physical execution phase. This directly leads to a precipitous drop in the effective utilization rate of sewing equipment, rapidly turning previously idle nodes in the short queue into secondary congestion sources that cause overall line rhythm loss and work-in-process accumulation. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an AI-assisted garment production line layout system, which aims to improve the problem that traditional garment line layout mostly adopts static short queue rules, which ignores the time spent on equipment mold changes and the fluctuation of human efficiency, resulting in serious congestion of the production line during frequent production changes.

[0005] This invention provides the following technical solution: an AI-assisted garment production line layout system, comprising the following modules:

[0006] The manufacturing network status awareness module is used to obtain the actual work-in-process queue length, semi-finished product arrival rate, and target tooling feature set of the material to be allocated at the candidate workstation.

[0007] The human-machine physical feature extraction module is used to obtain the current physical tooling feature set of the candidate workstation, as well as the cumulative number of processed parts and continuous working time of the operator;

[0008] The non-stationary service rate calculation module is used to calculate the expected value of the non-stationary service rate based on the cumulative number of processed pieces, the continuous operation time, the learning curve and the fatigue decay model.

[0009] The mold-changing feature equivalent mapping module is used to calculate the physical mold-changing penalty coefficient by calculating the current physical tooling feature set and the target tooling feature set, and to convert the physical mold-changing penalty coefficient into an equivalent virtual queue by combining the semi-finished product arrival rate and the benchmark mold-changing time constant.

[0010] The collaborative Lyapunov optimization module is used to construct a collaborative Lyapunov function by superimposing the lengths of the equivalent virtual queue and the real work-in-process queue, and to construct a control equation by introducing the expected value of the non-stationary service rate to solve for the target optimal workstation.

[0011] The flexible closed-loop execution module is used to issue the material token and preparation list to be allocated according to the target optimal workstation, and to feed back to the manufacturing network status perception module and the human-machine physical feature extraction module.

[0012] By adopting the above technical solution, the virtual queue for mold changing is superimposed with the actual backlog and combined with the non-stationary service rate to solve for the optimal workstation, and then the flexible scheduling is executed in a closed loop. This improves the problem that traditional garment production lines mostly use static short queue rules, which ignore the time spent on mold changing and the fluctuation of human efficiency, resulting in serious congestion of the production line when frequently changing production.

[0013] The present invention has the following beneficial effects:

[0014] 1. In this invention, by superimposing the virtual queue for mold changing with the actual backlog and combining it with the non-stationary service rate to solve for the optimal workstation, flexible scheduling is executed in a closed loop. This improves the problem that traditional garment production lines mostly use static short queue rules, which ignore the time spent on mold changing and fluctuations in human efficiency, resulting in severe congestion of the production line during frequent production changes.

[0015] 2. In this invention, by superimposing equivalent virtual queues with real backlogs and combining them with non-stationary service rates to solve for the optimal workstation, and then performing closed-loop work assignment, the problem of scheduling being out of sync with actual production capacity is improved by using a single short queue rule in traditional garment production, which ignores mold change time and human error.

[0016] 3. In this invention, by deriving the physical mold change penalty and combining it with the arrival rate to convert it into an equivalent virtual queue, the cross-dimensional mapping of machine changeover time to the number of backlogged parts is realized. This improves the problem that traditional production scheduling mostly uses independent and fragmented time parameters for evaluation. Due to the lack of a unified evaluation space, this leads to long downtime caused by material dispatching to workstations without matching auxiliary tools. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of an AI-assisted garment production line layout system proposed in this invention;

[0018] Figure 2 This is a flowchart of the state perception and data extraction process of an AI-assisted garment production line layout system proposed in this invention;

[0019] Figure 3 This is a flowchart of the optimization control solution for an AI-assisted garment production line layout system proposed in this invention;

[0020] Figure 4 This is a flowchart illustrating the flexible closed-loop execution and feedback process of an AI-assisted garment production line layout system proposed in this invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1:

[0023] In a first embodiment of the present invention, the present invention provides an AI-assisted garment production line layout system, such as... Figures 1-4 As shown, it includes the following steps:

[0024] The manufacturing network status awareness module is used to obtain the actual work-in-process queue length, semi-finished product arrival rate, and target tooling feature set of the material to be allocated at the candidate workstation.

[0025] Furthermore, a network state awareness module is manufactured to perform:

[0026] By scanning the tokens of the materials to be allocated using RFID sensors deployed at the scheduling decision points of the production line, the set of target tooling features bound to the tokens of the materials to be allocated is extracted.

[0027] Call the underlying data interface of the manufacturing execution system to extract the actual length of the work-in-process queue in front of the candidate workstation;

[0028] Based on the flow log, extract the number of semi-finished products flowing into the candidate workstation within the set time window, and output the semi-finished product arrival rate by performing a ratio calculation on the number of semi-finished products and the set time window.

[0029] Specifically, within the operational logic of the manufacturing network status perception module, the input end directly receives the underlying physical electrical signals transmitted by RFID sensors on the production line and simultaneously reads the underlying flow logs with system timestamps from the manufacturing execution system server, while simultaneously setting the time window benchmark for observation and calculation. After data parsing and ratio calculation within the module, the output end sends the target tooling feature set of the material to be allocated, the actual work-in-process queue length in front of the candidate workstations, and the semi-finished product arrival rate to the network control center. The target tooling feature set, the actual work-in-process queue length, and the semi-finished product arrival rate together constitute the underlying environmental status discrete parameters required for subsequent nonlinear algorithms.

[0030] In the process of parsing and outputting the semi-finished product arrival rate, the manufacturing network status awareness module performs mathematical analytical calculations based on the discrete-time queuing model. The specific calculation model for the semi-finished product arrival rate is set as follows:

[0031] ;

[0032] in This represents a specific candidate workstation on the assembly line that is either idle or in the process of processing. This represents the current absolute system time when the manufacturing network state awareness module triggers the scheduling decision. This represents a pre-set time window for mobile observation, with a value range of 15 to 30 minutes. This represents the total number of semi-finished products that physically flow into front of a specific candidate workstation within the flow observation time window. This represents the semi-finished product arrival rate of a specific candidate workstation in the current scheduling cycle, with the physical dimension uniformly converted to pieces per minute. To eliminate instantaneous data jitter caused by crosstalk from underlying radio frequency signals or network communication delays, an exponential moving average filter is connected in series at the system's underlying layer. This filter performs weighted filtering on the semi-finished product arrival rate of the current scheduling cycle and the historical smoothed value before inputting it into the subsequent calculation link.

[0033] By using RFID sensors deployed at the production line scheduling decision points to scan the materials to be allocated and extract the target tooling feature set, combined with calling the underlying data interface of the manufacturing execution system to extract the actual work-in-process queue length, and calculating the ratio of the quantity of semi-finished products to the set time window based on the flow log to output the semi-finished product arrival rate, this improves upon the traditional technical defects of garment production scheduling software that relies solely on static scheduling plans for unidirectional open-loop dispatching and lacks a data feedback mechanism, resulting in an inability to perceive the instantaneous physical congestion level and tooling matching degree of each underlying operation node. This avoids the serious accumulation of work-in-process at local workstations on the production line and the downtime due to tooling mismatch caused by the lack of underlying real-state feature support in mixed production scenarios of small-batch, multi-batch garment orders.

[0034] The manufacturing network status perception module accurately collects the actual work-in-process queue length and semi-finished product arrival rate, completing the substantial transformation of the physical workshop congestion status into a discrete digital matrix. This provides a strictly dimensionally consistent data basis for the Lyapunov optimization module to perform equivalent virtual queue transformation.

[0035] The human-machine physical feature extraction module is used to obtain the current physical tooling feature set of the candidate workstation, as well as the cumulative number of processed parts and continuous operation time of the operator;

[0036] Furthermore, the human-machine physical feature extraction module is used to perform:

[0037] The electronic tags of physical fixtures assembled at candidate workstations are read through the on-site equipment ports, and the electronic tags of physical fixtures are integrated to generate the current set of physical fixture features;

[0038] Collect the identification information of the operators and initiate a data retrieval request to the human resources database based on the identification information;

[0039] Extract the cumulative number of pieces processed by the operator for the same material within the current shift cycle, and extract the continuous working time calculated from the initial clock-in time.

[0040] Specifically, in the execution control link of the human-machine physical feature extraction module, the data input end receives the radio frequency signal of the physical auxiliary tool electronic tag transmitted from the field equipment port through the physical communication bus, and receives the operator's identity information collected by the workstation check-in terminal. The human-machine physical feature extraction module uses the identity information to send a log query command to the human resources database to obtain personnel scheduling and attendance records. After completing the parsing and timestamp alignment of the underlying heterogeneous data, the data output end sends the current physical tooling feature set, the cumulative number of processed parts, and the continuous operation time to the non-stationary service rate calculation module and the mold change feature equivalent mapping module. The current physical tooling feature set, the cumulative number of processed parts, and the continuous operation time establish the dual constraint boundary of physical machine hardware and human factors engineering status in the system algorithm.

[0041] In the process of extracting continuous operation time and the current physical tooling feature set, the human-machine physical feature extraction module performs variable analysis and calculation based on the underlying physical state. To eliminate invalid time accumulation caused by manual absence or sewing equipment idling and to ensure the statistical accuracy of the underlying data, the mathematical calculation model for continuous operation time is replaced with:

[0042] ;

[0043] in It represents the actual continuous working time of the operator, and the physical dimension is set to minutes. This represents a valid work time segment. The human-machine physical feature extraction module collects transient operating current in real time through a current transformer deployed on the main spindle motor of the sewing equipment. Only when the transient operating current continuously exceeds a preset standby idling threshold is the human-machine physical feature extraction module included in and accumulated to obtain a valid work time segment. .

[0044] The discrete space representation model of the current physical tooling feature set is defined as follows:

[0045] ;

[0046] in The set of physical tooling features representing the current assembly of candidate workstations. The representative obtained by analyzing the electronic tags of physical assistive devices Each physical assist hardware code points to a specific sewing presser foot or pull tube device.

[0047] This solution reads the electronic tags of physical fixtures assembled at candidate workstations through on-site equipment ports to generate a set of current physical tooling features. Combined with the collected operator identification information, it extracts the cumulative number of processed pieces and continuous operation time for the same material within the current shift cycle from the human resources database. This improves upon the traditional technical shortcomings of flexible production scheduling control systems, which abstract manual work units into standard black boxes with constant throughput and completely isolate the differences in underlying machine physical components. Traditional solutions lack real-time measurement methods for tracking the physical exertion of assembly line operators and the proficiency ramp-up period for new process actions. This solution avoids large-scale downtime caused by the system blindly assigning new process orders to nodes lacking corresponding hardware fixtures, and eliminates the production operation risk of forcibly assigning complex cutting pieces to personnel at the point of fatigue, leading to localized process bottlenecks and paralysis.

[0048] The human-machine physical feature extraction module accurately extracts the current physical tooling feature set, cumulative number of processed parts, and continuous operation time, and performs high-precision digital mirror reconstruction of the manual operation efficiency features and the physical equipment assembly status, establishing a real and reliable engineering reference benchmark for subsequent algorithm execution of non-stationary capacity extrapolation and mold change penalty evaluation.

[0049] The non-stationary service rate calculation module is used to calculate the expected value of the non-stationary service rate based on the cumulative number of processed pieces, continuous operation time, learning curve and fatigue decay model.

[0050] Furthermore, the non-stationary service rate calculation module is used to perform:

[0051] Extract the standard rated capacity corresponding to the candidate workstation and the initial extremely low capacity at the beginning of the model change.

[0052] The cumulative number of processed pieces is substituted into the preset learning curve as the independent variable to perform nonlinear evolution calculation and extract the cognitive learning rate corresponding to the operator.

[0053] The continuous working time is substituted into the preset fatigue decay model as the independent variable to perform exponential decay calculation and extract the physical fatigue decay index corresponding to the operator.

[0054] The non-stationary service rate calculation module is also used to perform:

[0055] Perform a difference calculation between the standard calibration capacity and the initial extremely low capacity to generate a capacity boundary benchmark value;

[0056] The dynamic capacity loss is obtained by multiplying the capacity boundary benchmark, cognitive learning rate and physical fatigue decay index.

[0057] Perform a correlation calculation between the standard rated capacity and the dynamic capacity loss, and output the expected value of the non-stationary service rate.

[0058] Specifically, in the computational execution chain of the non-stationary service rate calculation module, the input end receives the cumulative number of processed parts and continuous working time of the operator from the human-machine physical feature extraction module, and retrieves the standard calibrated capacity and initial minimum capacity of candidate workstations for the current process style from the system's underlying database based on the work order attributes. After internal nonlinear evolution calculations, difference calculations, and correlation product processing, the output end sends the expected value of the non-stationary service rate when the operator accepts and allocates material tokens to the collaborative Lyapunov optimization module. The expected value of the non-stationary service rate constitutes the underlying human-physical boundary constraints when the optimization algorithm executes routing decisions.

[0059] In calculating the expected value of the non-stationary service rate, the non-stationary service rate calculation module performs dynamic deduction based on the underlying human factors engineering operating rules. The comprehensive calculation model for the expected value of the non-stationary service rate is set as follows:

[0060] ;

[0061] in This represents the expected value of the non-stationary service rate calculated when operators receive specific types of materials, with the physical dimension uniformly set to pieces per minute. This represents the standard rated capacity of the candidate workstation under fully proficient conditions. This represents the initial, extremely low production capacity when operators are just beginning to understand the new materials. This represents the capacity boundary benchmark value generated by performing a difference calculation between the standard calibrated capacity and the initial extremely low capacity. This represents the cumulative number of pieces processed by the operator for the same type of material. The cognitive learning rate index represents the value between zero and one. The cognitive learning rate index is obtained by the system retrieving the actual production capacity log of the operator's historical process switching cycle in the manufacturing execution system and calibrating the nonlinear output discrete points by curve fitting using the least squares method. The cognitive learning rate is generated by substituting the cumulative number of processed parts into the learning curve and performing nonlinear evolution calculations. This represents the continuous working time of the operator. The physical fitness attenuation coefficient represents the value between zero and one. The physical fitness attenuation coefficient is dynamically derived by the system based on the historical shift fatigue data of the operator in the human resources database using a maximum likelihood estimation model. The physical fatigue decay index is generated by substituting continuous working time into the fatigue decay model and performing exponential decay calculation. This represents the dynamic capacity loss obtained by multiplying the capacity boundary benchmark, cognitive learning rate, and physical fatigue decay index.

[0062] By extracting the standard calibration capacity corresponding to the candidate workstations and the initial extremely low capacity at the beginning of the product change, the cumulative number of processed parts and continuous operation time are substituted into the preset learning curve and fatigue decay model to extract the cognitive learning rate and physical fatigue decay index. Then, the dynamic capacity loss is obtained by performing a product operation on the capacity boundary benchmark value, cognitive learning rate and physical fatigue decay index. Finally, the standard calibration capacity and dynamic capacity loss are correlated to output the non-stationary service rate expectation value. This improves the traditional production scheduling control algorithm that treats manual operation units as absolute static constants. It eliminates the traditional technical defects of the action exploration stage that operators must go through when facing new production processes and the physical decline phenomenon caused by the increase in working hours. This avoids the scheduling system from excessively tilting the distribution of materials to workstations that have not passed the learning ramp-up period based on false high capacity prediction instructions, and completely cuts off the physical causes that cause serious blockage of work-in-process in the workshop and imbalance of the overall production line rhythm.

[0063] The mold-changing feature equivalent mapping module is used to calculate the physical mold-changing penalty coefficient by comparing the current physical tooling feature set with the target tooling feature set, and then convert the physical mold-changing penalty coefficient into an equivalent virtual queue by combining the semi-finished product arrival rate and the benchmark mold-changing time constant.

[0064] Furthermore, the model-changing feature equivalent mapping module is used to perform:

[0065] The current physical tooling feature set and the target tooling feature set are analyzed separately to extract the types of physical tools currently in use and the types of physical tools required.

[0066] Perform an intersection operation on the types of physical assistive devices currently in use and the types of physical assistive devices required, and extract the number of shared assistive devices;

[0067] Perform a union operation on the types of physical assistive devices currently in use and the types of physical assistive devices required to extract the total number of assistive devices;

[0068] Calculate the ratio of the number of shared assistive devices to the total number of assistive devices, and output the physical mold change penalty coefficient based on the ratio.

[0069] The model-changing feature equivalent mapping module is also used to perform:

[0070] Retrieve the system's preset benchmark mold change time constant and determine the dimensional matching status between the benchmark mold change time constant and the semi-finished product arrival rate;

[0071] After confirming the dimensional matching status, the physical mold change penalty coefficient, the semi-finished product arrival rate, and the benchmark mold change time constant are subjected to multiplication mapping operation;

[0072] Based on the result of the multiplication mapping operation, an equivalent virtual queue with the dimension of number of items is generated.

[0073] Specifically, in the data interaction link of the mold-changing feature equivalent mapping module, the input end receives the current physical tooling feature set, the target tooling feature set, and the semi-finished product arrival rate transmitted by the front-end module. Internally, this module sequentially executes set operations and dimension conversion mechanisms to complete the parsing and reconstruction of the underlying physical feature data. The output end sends an equivalent virtual queue with component dimensions to the subsequent collaborative optimization control hub. This equivalent virtual queue, as a core environmental variable, directly participates in the mathematical deduction process of the optimal solution for subsequent network node transitions.

[0074] In the process of analyzing physical tooling differences and outputting the physical mold-changing penalty coefficient, the mold-changing feature equivalent mapping module performs deduction based on the set-theoretic metric space rule. The mathematical model for calculating the physical mold-changing penalty coefficient is set as follows:

[0075] ;

[0076] in This represents the physical mold change penalty coefficient calculated for a specific candidate workstation; it is a pure numerical value without physical dimensions. The current set of physical tooling features consists of the types of physical aids currently in use at the candidate workstation. The target tooling feature set consists of the types of physical tools required for the material to be allocated. This represents the number of shared aids obtained by performing an intersection operation on the two sets mentioned above. This represents the total number of assistive devices obtained by performing a union operation on the two sets mentioned above.

[0077] After determining the dimensional matching status, the mathematical mapping model for generating the equivalent virtual queue is set as follows:

[0078] ;

[0079] in The equivalent virtual queue generated by the transformation and deduction is strictly defined by the number of pieces. The arrival rate of semi-finished products flowing into candidate workstations is represented by the physical unit of pieces per minute. Representing the dynamic reference mold-changing time, the system has a two-dimensional mold-changing time calibration matrix pre-built at the bottom layer that maps different types of physical assistive devices. Based on the specific assistive device codes in the feature difference set, the system addresses, matches and accumulates the corresponding dynamic reference mold-changing time in the two-dimensional mold-changing time calibration matrix, thereby replacing the fixed time constant.

[0080] The physical mold-changing penalty coefficient is calculated by extracting the number of shared auxiliary tools and the total number of auxiliary tools from the tooling feature set. Then, a multiplicative mapping operation is performed based on the semi-finished product arrival rate and the baseline mold-changing time constant to transform it into an equivalent virtual queue. This execution logic overcomes the inherent defect of conventional flexible scheduling software that calculates the data network queuing status and the underlying machine physical changeover time separately. Conventional control systems tend to execute the shortest queue dispatch principle when dealing with multi-style mixed-flow production, ignoring the long-cycle downtime and debugging caused by mismatched underlying hardware auxiliary tools at the receiving station.

[0081] By using the semi-finished product arrival rate as a mathematical bridge, the abstract physical tool disassembly and installation time, under strict dimensional consistency constraints, is equivalently transformed into the number of semi-finished products that should have been processed but were forced to accumulate in front of the workstation during the downtime period. This cross-physical dimension homomorphic transformation mechanism forcibly pulls mold change losses into the mathematical game space of queue backlog, eliminating the control risk of the scheduling system blindly dispatching new process orders to short queue nodes lacking corresponding hardware tools, and cutting off the physical cause of the precipitous drop in the effective utilization rate of sewing equipment due to frequent changes of presser feet or pull tubes at the source.

[0082] The Collaborative Lyapunov Optimization Module is used to construct a Collaborative Lyapunov Function by superimposing the lengths of the equivalent virtual queue and the real work-in-process queue, and to construct the control equation by introducing the expected value of the non-stationary service rate to solve for the target optimal workstation.

[0083] Furthermore, the Lyapunov optimization module is used to execute:

[0084] Extract the length of the actual work-in-progress queue in terms of the number of available parts and the equivalent virtual queue in terms of the number of available parts;

[0085] Add the actual work-in-process queue length to the equivalent virtual queue using the same units, and output the sum of the queue lengths.

[0086] The queue stacking and quadratic form expansion processes are performed to generate a single-node quadratic form function that produces nonlinear cross-coupling terms;

[0087] The single-node quadratic functions of all candidate workstations on the production line are summed to output a cooperative Lyapunov function.

[0088] The collaborative Lyapunov optimization module is also used to execute:

[0089] Extract the single-step conditional drift between adjacent decision cycles based on the collaborative Lyapunov function;

[0090] The expected value of non-stationary service rate is converted into a negative reward term and introduced into the control logic;

[0091] Extract dynamic weight adjustment parameters, and use these parameters to perform a weighted correlation between the single-step conditional drift and the negative reward term to construct the control equation;

[0092] Find the candidate workstations corresponding to the minimum value of the control equations, and set the network transition nodes as the target optimal workstations.

[0093] Specifically, in the data interaction link of the collaborative Lyapunov optimization module, the input end receives in parallel the actual work-in-process queue length, the equivalent virtual queue with the same quantity of parts, and the expected value of the non-stationary service rate transmitted by the front-end module. The collaborative Lyapunov optimization module uses these as environmental variables and substitutes them into the underlying mathematical model to perform calculations, completing the optimization solution of nonlinear superposition of the same quantity and drift penalty. The output end generates the target optimal workstation network address for the current material to be allocated and sends this network address to the flexible closed-loop execution module to trigger the underlying physical transmission action.

[0094] In the process of deriving the network topology and outputting the target optimal workstation, the cooperative Lyapunov optimization module executes control law solutions based on stochastic network optimization theory. The construction model of the cooperative Lyapunov function is defined as follows:

[0095] ;

[0096] The optimization model for the single-step conditional drift and the joint control equations is defined as follows:

[0097] ;

[0098] in This represents the discrete decision time slot triggered by the current scheduling instruction. It represents a finite set of all candidate workstations currently in a connected and running state on the production line. This represents a specific candidate workstation within the set of candidate workstations. This represents the actual work-in-process queue length in front of the specific candidate workstation extracted. This represents the equivalent virtual queue obtained after pre-mapping processing. The cooperative Lyapunov function, representing the sum of quadratic functions of all single nodes in the global production line, characterizes the virtual congestion potential energy of the entire workshop. Before calculating the single-step conditional drift, the discrete state evolution equation of the real work-in-process queue is established at the system's underlying level as follows:

[0099] ;

[0100] in The actual work-in-process queue length represents the next discrete decision time interval for a specific candidate workstation; To obtain the non-negative boundary operator, which is used to ensure that the physical queue length does not have a negative value; This represents the arrival rate of semi-finished products flowing into a specific candidate workstation during the current scheduling cycle. The single-step conditional drift between adjacent decision cycles is calculated based on the co-Lyapunov function. The derivation of the single-step conditional drift is strictly controlled by the boundary constraints of the discrete state evolution equation mentioned above, and it represents the upper bound of the deterioration rate of the queue backlog state. This represents the expected non-stationary service rate for a specific candidate workstation to undertake new styles of materials. This represents a dynamic weight adjustment parameter. The system adaptively adjusts the total length of the global actual work-in-process queue based on the current total length of the global actual work-in-process queue. When the total length of the global actual work-in-process queue approaches the extreme value of the workshop's physical capacity, the system lowers the weight using an exponential decay function. The value is increased to force priority in suppressing queue backlog; conversely, it is increased if the value is decreased. The value focuses on the benefits of production capacity.

[0101] The calculation step of adding the actual work-in-process queue length and the equivalent virtual queue length to each other with the same dimensions and then performing a quadratic expansion generates a nonlinear cross-coupling term at the system's underlying layer, which is the product of the actual queue length and the virtual queue length. This coupling term breaks the linear independence of physical variables in conventional algorithms, forcing the scheduling logic to perform strict avoidance actions and cutting off the path for distributing new materials to nodes that have already accumulated semi-finished products and whose physical mold changing time is extremely long.

[0102] Based on this, the single-step conditional drift is calculated, and the expected value of the non-stationary service rate is converted into a negative reward term. A control equation is constructed using dynamically weighted parameters. To meet the millisecond-level control cycle requirements of industrial sites, when the number of candidate workstation nodes exceeds the system's computing power threshold, the underlying controller employs a local topology sampling strategy to extract a subset of candidate nodes. By traversing the candidate node subsets, the discrete cost of the control equation is calculated for each subset, and the specific network address that minimizes the discrete cost is output as the target optimal workstation. This improves upon the technical shortcomings of conventional production scheduling control algorithms, which treat network queuing time, machine mold change and debugging cycles, and the operator's new process proficiency ramp-up period as separate calculations. Conventional solutions lack a unified measurement and evaluation space, easily falling into the trap of forcibly assigning new materials to completely idle workstations that lack corresponding molds and have no experience with new process processing, resulting in local optima. Within a rigorous Lyapunov drift and penalty control domain, this solution simultaneously weighs the risks of queue backlog outbreaks, physical downtime penalties, and the potential efficiency of human operations. It avoids the risk of production loss of control in scenarios with small-batch, high-frequency production changes, where the system blindly pursues absolute physical load balance, leading to frequent downtime for repacking and inefficient cognitive ramp-up.

[0103] The flexible closed-loop execution module is used to issue the material tokens and preparation lists to be allocated based on the target optimal workstation, and feed them back to the manufacturing network status perception module and the human-machine physical feature extraction module.

[0104] Furthermore, a flexible closed-loop execution module is used to execute:

[0105] Based on the target optimal workstation, generate a network routing message, send the network routing message to the underlying transmission network, and control the underlying transmission network to physically guide the materials to be allocated to the target optimal workstation.

[0106] Extract the target tooling feature set of the material to be allocated and the current physical tooling feature set corresponding to the target optimal workstation;

[0107] Perform a difference comparison between the target tooling feature set and the current physical tooling feature set to extract the feature difference set, parse the feature difference set to generate a material preparation list, and push the material preparation list to the corresponding on-site operation terminal;

[0108] The system monitors the first piece completion status signal of the target optimal workstation. Upon receiving the first piece completion status signal, the system extracts the target tooling feature set and the previous work order historical feature set to calculate the cosine similarity. Only when the cosine similarity is lower than the preset experience inheritance threshold will the cumulative number of processed pieces and continuous operation time be reset. The reset cumulative number of processed pieces and continuous operation time will be fed back to the manufacturing network status perception module and the human-machine physical feature extraction module, thus avoiding unnecessary learning curve resets caused by minor process adjustments.

[0109] Specifically, within the data interaction link of the flexible closed-loop execution module, the input end receives the target optimal workstation network address issued by the collaborative Lyapunov optimization module, and simultaneously retrieves the target tooling feature set of the material to be allocated and the current physical tooling feature set corresponding to the target optimal workstation. Internal processing nodes complete routing message encapsulation and difference set parsing. The output end sends network routing messages to the underlying transmission network, pushes the material preparation list to the field operation terminal, and sends the reset status parameters to the manufacturing network status perception module and the human-machine physical feature extraction module after the physical changeover action is completed.

[0110] During the extraction of feature difference sets and the resetting of underlying parameters, the flexible closed-loop execution module performs underlying operations based on set theory and discrete state machine logic. The feature difference set extraction model is defined as follows:

[0111] ;

[0112] in This represents the feature difference set output by the difference comparison calculation, and its elements precisely indicate the physical assistive device hardware codes that must be prepared by the on-site auxiliary workers. This represents the set of target tooling features that the material to be allocated must possess for the token processing. This represents the current set of physical tooling features that are currently mounted at the target optimal workstation. The relative complement operator for a set.

[0113] The state reset model is set as follows after the first piece is processed:

[0114] ;

[0115] ;

[0116] in This represents the cumulative number of processed parts after the reset. In this control node system, its initial value is forcibly reset back to 1. This represents the continuous working time after the reset. The initial value of this value is forcibly reset to 0 in the control node system.

[0117] Based on the target optimal workstation, a network routing message is generated and sent to the underlying transmission network to physically guide the material to be allocated. A difference comparison is performed between the target tooling feature set and the current physical tooling feature set to extract the feature difference set, and a material preparation list is generated and pushed to the field operation terminal. Then, after detecting the first piece processing completion status signal, the cumulative number of processed pieces and continuous operation time are reset, and data feedback is implemented. The above execution steps establish a physical control link from the upper-level mathematical decision-making center to the lower-level manufacturing execution hardware.

[0118] Conventional scheduling systems only issue instructions at the pure data flow level. On-site operators can only check the process details and find matching accessories offline after the materials actually arrive at the physical workstation. This serial operation mode directly induces long-term downtime of sewing equipment. This solution decomposes the pure network control optimization result into material space transfer instructions and accessory replacement details. The underlying transmission network transfers physical cut pieces according to routing messages, and the on-site operation terminal guides workers to prepare accessories in advance based on feature difference sets before the materials arrive. This parallel collaborative execution logic reduces the absolute time of physical mold change. In addition, the status parameter reset and feedback action after the first piece is processed remaps the worker's learning and recovery status in the early stage of the new process into the system state decision machine, providing a real-time updated underlying data source for the non-stationary service rate calculation of the next scheduling cycle.

[0119] Example 2:

[0120] The garment manufacturing workshop operates a multi-style mixed-flow production operation. Multiple serial and parallel sewing stations are arranged on the assembly line, with automated guided vehicles or intelligent hanging systems at the bottom responsible for distributing cut materials to each station. Within the same shift cycle, the production line frequently receives micro-orders with different process characteristics, each with varying sewing requirements. In addition to performing regular sewing actions, operators must physically disassemble, assemble, and adjust hardware accessories such as presser feet, pull tubes, or needle plates to accommodate changes in order styles.

[0121] In mixed-flow production scenarios, when facing frequent changes in multiple product styles, conventional line control logic primarily relies on the length of the work-in-process queue at each workstation to perform a coarse, single-task dispatching approach, tending to directly assign newly arriving materials to the shortest idle node in the current queue. This network topology scheduling strategy completely strips away the dual constraints of the underlying physical execution environment. The first constraint lies in the time-consuming physical hardware replacement; when new materials do not match the existing physical aids at the receiving workstation, it inevitably leads to a long downtime for disassembly and assembly. The second constraint lies in the non-stationary operational efficiency of human factors; when operators face new processes, they inevitably experience a cognitive adaptation phase, resulting in a non-linear lag in actual capacity ramp-up. The scheduling center treats manual work units as static machines with a constant throughput rate, mistakenly forcing new style materials onto workstations with short queues but lacking corresponding aids and relevant process experience. A severe disconnect exists between the algorithm's preset ideal production capacity and the actual nonlinear output on-site, inducing a highly concealed production capacity mirage phenomenon. This leads to a precipitous drop in the effective utilization rate of sewing equipment, causing nodes that were originally in short queues to quickly deteriorate into secondary congestion sources that trigger overall line rhythm loss and work-in-process accumulation. To solve the above problems, this invention provides an AI-assisted garment production line layout system, the structure of which is as follows: Figure 1 As shown. The specific implementation process of this system is as follows:

[0122] The system collects data on the actual backlog of work-in-process, tooling configuration features, and continuous worker operation data through a manufacturing network status perception module and a human-machine physical feature extraction module. A non-stationary service rate calculation module introduces learning curves and fatigue models to quantify the worker's skill progression and physical exertion when facing new processes. A mold-changing feature equivalent mapping module converts the disassembly and assembly time of physical aids into an equivalent virtual queue with a quantity dimension. A collaborative Lyapunov optimization module superimposes the actual backlog and virtual mold-changing penalties on the same dimension, and combines this with dynamic capacity constraints to solve for the globally optimal work assignment node. Finally, a flexible closed-loop execution module issues physical scheduling instructions and material preparation lists. The entire execution logic forcibly incorporates the dynamic operational efficiency of personnel and the cost of equipment hardware switching into the topology decision, eliminating the control risks of blindly assigning new orders to short-queue workstations lacking matching hardware and relevant processing experience. This cuts off the source of equipment utilization rate decline and secondary production line congestion caused by frequent downtime and debugging in multi-style mixed-flow production scenarios.

[0123] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-assisted garment production line layout system, characterized in that, Includes the following modules: The manufacturing network status awareness module is used to obtain the actual work-in-process queue length, semi-finished product arrival rate, and target tooling feature set of the material to be allocated at the candidate workstation. The human-machine physical feature extraction module is used to obtain the current physical tooling feature set of the candidate workstation, as well as the cumulative number of processed parts and continuous working time of the operator; The non-stationary service rate calculation module is used to calculate the expected value of the non-stationary service rate based on the cumulative number of processed pieces, the continuous operation time, the learning curve and the fatigue decay model. The mold-changing feature equivalent mapping module is used to calculate the physical mold-changing penalty coefficient by calculating the current physical tooling feature set and the target tooling feature set, and to convert the physical mold-changing penalty coefficient into an equivalent virtual queue by combining the semi-finished product arrival rate and the benchmark mold-changing time constant. The collaborative Lyapunov optimization module is used to construct a collaborative Lyapunov function by superimposing the lengths of the equivalent virtual queue and the real work-in-process queue, and to construct a control equation by introducing the expected value of the non-stationary service rate to solve for the target optimal workstation. The flexible closed-loop execution module is used to issue the material token and preparation list to be allocated according to the target optimal workstation, and to feed back to the manufacturing network status perception module and the human-machine physical feature extraction module.

2. The AI-assisted garment production line layout system according to claim 1, characterized in that, The manufacturing network state awareness module is used to perform: By scanning the token of the material to be allocated using RFID sensors deployed at the scheduling decision point of the production line, the set of target tooling features bound to the token of the material to be allocated is extracted. Call the underlying data interface of the manufacturing execution system to extract the actual length of the work-in-process queue in front of the candidate workstation; The number of semi-finished products flowing into the candidate workstation within a set time window is extracted based on the flow log. The semi-finished product arrival rate is output by performing a ratio calculation between the number of semi-finished products and the set time window.

3. The AI-assisted garment production line layout system according to claim 1, characterized in that, The human-machine physical feature extraction module is used to perform: The electronic tags of the physical aids assembled at the candidate workstation are read through the field equipment port, and the physical aid electronic tags are integrated to generate the current physical tooling feature set; Collect the identification information of the operators and initiate a data retrieval request to the human resources database based on the identification information; Extract the cumulative number of pieces processed by the operator for the same material within the current shift cycle, and extract the continuous working time calculated from the initial clock-in time.

4. The AI-assisted garment production line layout system according to claim 1, characterized in that, The non-stationary service rate calculation module is used to perform: Extract the standard rated capacity and the initial extremely low capacity at the beginning of the product changeover for the candidate workstations; The cumulative number of processed parts is substituted into a preset learning curve as an independent variable to perform nonlinear evolution calculation, and the cognitive learning rate corresponding to the operator is extracted. The continuous working time is used as an independent variable and substituted into a preset fatigue decay model to perform exponential decay calculation, thereby extracting the physical fatigue decay index corresponding to the operator.

5. The AI-assisted garment production line layout system according to claim 4, characterized in that, The non-stationary service rate calculation module is also used to perform: Perform a difference calculation between the standard calibrated capacity and the initial extremely low capacity to generate a capacity boundary benchmark value; The dynamic capacity loss is obtained by multiplying the capacity boundary benchmark value, the cognitive learning rate, and the physical fatigue decay index. Perform a correlation calculation between the standard rated capacity and the dynamic capacity loss, and output the expected value of the non-stationary service rate.

6. The AI-assisted garment production line layout system according to claim 1, characterized in that, The mold-changing feature equivalent mapping module is used to perform: The current physical tooling feature set and the target tooling feature set are analyzed respectively to extract the types of physical aids currently in use and the types of physical aids required. Perform an intersection operation on the types of physical assistive devices currently in use and the types of physical assistive devices required, and extract the number of shared assistive devices; Perform a union operation on the types of physical assistive devices currently in use and the types of physical assistive devices required to extract the total number of assistive devices; Calculate the ratio of the number of shared assistive devices to the total number of assistive devices, and output the physical mold change penalty coefficient based on the ratio.

7. The AI-assisted garment production line layout system according to claim 6, characterized in that, The mold-changing feature equivalent mapping module is also used to perform: The system retrieves the preset benchmark mold change time constant and determines the dimensional matching status between the benchmark mold change time constant and the semi-finished product arrival rate. After the dimensional matching status is confirmed, the physical mold change penalty coefficient, the semi-finished product arrival rate, and the benchmark mold change time constant are subjected to a multiplication mapping operation. Based on the result of the multiplication mapping operation, an equivalent virtual queue with the dimension of number of items is generated.

8. The AI-assisted garment production line layout system according to claim 1, characterized in that, The collaborative Lyapunov optimization module is used to perform: Extract the actual work-in-process queue length in terms of the number of components and the equivalent virtual queue in terms of the number of components; The actual work-in-process queue length is added to the equivalent virtual queue using the same unit of measurement, and the sum of the queue lengths is output. The queues are stacked and subjected to quadratic expansion to generate a single-node quadratic function that produces nonlinear cross-coupling terms; The single-node quadratic functions of all candidate workstations on the production line are accumulated to output the cooperative Lyapunov function.

9. The AI-assisted garment production line layout system according to claim 8, characterized in that, The collaborative Lyapunov optimization module is also used to perform: The single-step conditional drift between adjacent decision cycles is extracted based on the aforementioned collaborative Lyapunov function; The non-stationary service rate expectation is converted into a negative reward term and introduced into the control logic; Extract dynamic weight adjustment parameters, and use these parameters to perform a weighted association between the single-step conditional drift and the negative reward term to construct a control equation; Find the candidate workstations corresponding to the minimum value of the control equation, and set the network transition nodes as the target optimal workstations.

10. The AI-assisted garment production line layout system according to claim 1, characterized in that, The flexible closed-loop execution module is used to execute: Based on the target optimal workstation, a network routing message is generated and sent to the underlying transmission network to control the underlying transmission network to physically guide the material to be allocated to the target optimal workstation. Extract the target tooling feature set of the material to be allocated and the current physical tooling feature set corresponding to the target optimal workstation; Perform a difference comparison between the target tooling feature set and the current physical tooling feature set to extract the feature difference set, parse the feature difference set to generate a material preparation list, and push the material preparation list to the corresponding field operation terminal; The system monitors the first piece completion status signal of the target optimal workstation. Upon receiving the first piece completion status signal, it resets the cumulative number of processed pieces and the continuous operation time, and feeds back the reset cumulative number of processed pieces and the continuous operation time to the manufacturing network status perception module and the human-machine physical feature extraction module.

Images