Electronic product production line order intelligent scheduling management system based on deep learning

Abstract

This invention discloses a deep learning-based intelligent scheduling and management system for electronic product production line orders, comprising: a segment sequence construction module, which collects multi-source operation data of the production line and splits orders to form process segment sequences; a potential energy weight generation module, which inputs segment sequences and temporal perturbation data and outputs multi-dimensional scheduling potential energy weights via a JANET network; a curved segment identification module, which constructs a three-dimensional grid and potential field to generate streamlines and analyzes gradients and curvatures to determine key segments; a bridging segment determination module, which constructs a hierarchical metastable mirror body and calculates metastable energy barriers based on perturbation costs to identify bridging segments; a locking kernel generation module, which constructs a chain complex, performs continuous homology analysis, maps conformal manifolds to generate locking kernel structures; and a perturbation rearrangement scheduling module, which freezes the locking kernel and performs minimum perturbation rearrangement on segments outside the kernel to generate a scheduling scheme. This invention achieves minimum perturbation stable scheduling of production line orders through JANET potential energy modeling and locking kernel control.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial artificial intelligence technology, and in particular to an intelligent scheduling and management system for electronic product production line orders based on deep learning. Background Technology

[0002] Electronic product production lines typically include continuous processes such as SMT placement, component insertion, testing, and assembly. Orders must simultaneously meet multiple constraints during production, including equipment availability, material availability, tooling reuse, testing resource availability, and delivery deadlines. Existing order scheduling methods largely rely on rule-based scheduling, priority ranking, or heuristic optimization algorithms within MES and APS systems. These methods generate production plans based on order delivery dates, equipment load, material availability, and processing time, and can complete basic production scheduling tasks under normal, stable production conditions.

[0003] However, in multi-order parallel production scenarios, order insertion, material shortage, equipment failure, test equipment malfunction, and process cycle fluctuations occur frequently. Existing methods usually reorder orders as a whole, lacking the identification of the continuous connection between patch segments, plug-in segments, test segments, and assembly segments within an order. It is difficult to determine whether moving, repositioning, or removing a segment of an order will disrupt the process continuity, material continuity, tooling reuse, and test resource connection of subsequent processes.

[0004] Existing technologies cannot identify and protect critical order segments that, while not having the most urgent delivery dates, play a crucial role in continuous cross-process operations in real time. After dynamic disturbances occur, global or coarse-grained local rearrangements are easily adopted, leading to increased line changeovers, frequent switching of test programs, semi-finished product retention, queuing at subsequent workstations, and local congestion transfers, which in turn affect the stability of order delivery and the overall operating efficiency of the production line.

[0005] Therefore, how to provide an intelligent scheduling and management system for electronic product production line orders based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent scheduling and management system for electronic product production line orders based on deep learning. This invention fully utilizes multi-dimensional potential energy gating JANET networks, time-process-equipment three-dimensional potential field modeling, hierarchical metastable-bridging mirror analysis, and continuous homology-conformal locking kernel technology. It describes in detail the intelligent scheduling process for electronic product production line orders, including process segmentation, key acceptance segment identification, locking protection under dynamic disturbances, and minimum disturbance rearrangement. It has the advantages of high stability of cross-process acceptance, fast order insertion response speed, low loss during line changeover and test switching, less semi-finished product retention, and high order delivery reliability.

[0007] The intelligent scheduling and management system for electronic product production line orders based on deep learning according to an embodiment of the present invention includes:

[0008] The segment sequence construction module collects multi-source operational data from the electronic product production line, breaks down each order into multiple order process segments, and constructs an order process segment sequence.

[0009] The potential energy weight generation module inputs the order process segment sequence and the timing disturbance data within the continuous time window into the multidimensional potential energy gating JANET network. The multidimensional potential energy gating JANET network adopts a single forget gate memory unit and sets a potential energy allocation gate at the output end. By adaptively fusing the historical state decay factor and the current pulse gain, it outputs multidimensional scheduling potential energy weights.

[0010] The bending segment identification module maps order process segments to a three-dimensional grid space of time-process-equipment. Based on the multi-dimensional scheduling potential energy weight, it calculates the local potential energy corresponding to each order process segment, constructs a multi-potential-well potential field, generates adaptive bending streamlines along the gradient descent direction of the potential field, and performs gradient change analysis and curvature change analysis to determine key bending segments.

[0011] The bridging segment determination module constructs a hierarchical metastable-bridging mirror body for each key bending segment. Based on the mirror disturbance cost data corresponding to line switching, tooling, testing, waiting, queuing, delay and streamline detour in each mirror state, it calculates the metastable energy barrier and determines the metastable bridging segment.

[0012] The locking core generation module constructs a time-process-equipment chain complex based on key bending segments and non-stable bridging segments, performs continuous homology analysis, maps the time-process-equipment chain complex to a two-dimensional conformal manifold, and generates a continuous homology-conformal locking core.

[0013] When a dynamic disturbance event is detected, the disturbance rearrangement scheduling module freezes the order process segments and their corresponding acceptance relationships in the continuous coherence-conformal locking core, performs minimum disturbance rearrangement on the order process segments outside the continuous coherence-conformal locking core, and generates the final intelligent order scheduling scheme.

[0014] Optionally, the multi-source operational data includes order number, customer level, product model, order quantity, delivery deadline, placement program, stencil number, feeder configuration, insertion tooling, test program, test fixture, assembly version, equipment status, equipment load, equipment changeover time, material arrival status, material batch status, test equipment queuing status, semi-finished product waiting status, work-in-process queue depth, actual processing time, and order delay results.

[0015] Optionally, the step of splitting each order into multiple order process segments and constructing an order process segment sequence includes:

[0016] Read the product model, process route, equipment candidate set, material requirements, tooling requirements, test resource requirements, and delivery deadline for each order. Divide each order into surface mount (SMT) segment, component insertion segment, test segment, and assembly segment according to the surface mount (SMT) process, component insertion (Component Insertion) process, test segment, and assembly segment. Label each order's process segment with the order number, process type, preceding segment number, following segment number, candidate equipment number, material batch number, tooling number, test program number, estimated start time, and estimated end time. Sort each order's process segment according to process sequence, equipment occupancy, and time continuity to generate an order process segment sequence.

[0017] Optionally, the output multidimensional scheduling potential weights include:

[0018] Read the order process segment sequence and the timing disturbance data within the continuous time window corresponding to the current scheduling time. Encode the segment status of the process type, equipment occupancy status, material arrival status, test resource occupancy status, sequential status of preceding and following processes, and remaining delivery time of the order in the order process segment sequence. Encode the disturbance status of the line change change, tooling change change, test change change, semi-finished product waiting change, equipment load change, material arrival change, and equipment abnormality change in the timing disturbance data. Combine the segment status encoding results and the disturbance status encoding results into the JANET input status.

[0019] The JANET input state is input into the input mapping layer of the multidimensional potential energy gated JANET network to uniformly map the current acceptance state of the order process segment and the production line disturbance state, and generate candidate memory states.

[0020] The JANET input state and the memory state of the previous scheduling time are input into the single forget gate memory unit. The single forget gate memory unit generates the historical state decay factor based on the current disturbance intensity, the changes in the succession of the preceding and following processes, and the potential energy state of the previous scheduling time.

[0021] The memory state of the previous scheduling moment is attenuated and retained based on the historical state attenuation factor. The current pulse gain is generated based on the current line switching change, tooling change, test change, waiting change, delivery pressure change and abnormal trigger change. The attenuated and retained historical memory state is fused with the candidate memory state corresponding to the current pulse gain to generate the potential energy memory state of the current scheduling moment.

[0022] The potential energy memory state at the current scheduling moment is input into the potential energy allocation gate. The potential energy allocation gate allocates the potential energy memory state according to the impact of line switching, tooling reuse, test resource acceptance, semi-finished product waiting, delivery pressure, and equipment abnormality, and outputs the multi-dimensional scheduling potential energy weights corresponding to line switching, tooling, testing, waiting, delivery, and abnormality.

[0023] Optionally, determining the key bending segment includes:

[0024] Read the scheduling start time, scheduling end time, process type set and equipment candidate set of the order process segment sequence, divide the time range from the scheduling start time to the scheduling end time into continuous time grids according to the preset scheduling granularity, set the chip placement process, plug-in process, testing process and assembly process as process grids respectively, set the available equipment or workstations of each process as equipment grids, and generate a time-process-equipment three-dimensional grid space by combining the continuous time grids, process grids and equipment grids;

[0025] Read the executable time window, the process to which it belongs, the candidate equipment, the processing time, the preceding segment and the following segment corresponding to each order process segment, map each order process segment to the grid position that matches the executable time window, the process to which it belongs and the candidate equipment, and establish the process succession relationship between adjacent grids based on the succession relationship between the preceding segment and the following segment;

[0026] Read the multi-dimensional scheduling potential energy weights, generate segment local potential energy based on the line change cost, tooling cost, testing cost, waiting cost, delivery cost and exception cost corresponding to each order process segment, write the segment local potential energy into the corresponding grid position, and transfer the segment local potential energy to adjacent grids according to time succession relationship, process succession relationship and equipment occupancy relationship to construct a multi-potential well potential field;

[0027] In a multi-potential well field, starting from the preceding grid position of each order process segment, grid positions that satisfy the process continuity relationship, equipment occupation relationship, material arrival relationship and test resource acceptance relationship are selected sequentially along the local potential energy decrease direction of the segment, forming an adaptive curved streamline that runs through the chip placement process, insertion process, testing process and assembly process.

[0028] Perform streamline gradient change analysis and curvature change analysis on each order process segment in each adaptive curved streamline, and identify the order process segments that are at the positions of abrupt change in the direction of potential energy decrease, abrupt change in the direction of streamline curvature, the position of confluence of cross-process, or the starting position of streamline bypass as critical curved segments.

[0029] Optionally, the determination of the stable bridging segment includes:

[0030] Read the order number, process type, current grid position, previous segment, subsequent segment, candidate equipment, material batch, tooling number, test program number and adaptive bending streamline assignment result corresponding to each key bending segment, and generate the baseline state of the key bending segment;

[0031] A local mirror layer is constructed based on the baseline state of the key bending segment. The local mirror layer includes a retained in-situ mirror, a forward-moving mirror, a backward-moving mirror, a removed mirror, and a replaced device mirror. The execution time, device usage, tooling usage, test resource usage, preceding continuation state, and subsequent continuation state are recorded for each local mirror.

[0032] Centered on the critical bending segment, read the preceding segment, the following segment, the segment adjacent to the same equipment, and the segment adjacent to the same test resource to build a neighborhood mirror layer. Within the same scheduling time window, read the order process segment that has a material continuity relationship, process succession relationship, or test resource acceptance relationship with the critical bending segment to build a global constraint mirror layer. The hierarchical meta-stable-bridging mirror body is composed of the local mirror layer, the neighborhood mirror layer, and the global constraint mirror layer.

[0033] The image disturbance cost data corresponding to line switching, tooling, testing, waiting, queuing, delay and flow detour in each image state are collected respectively. The image disturbance cost data are weighted and aggregated according to the multidimensional scheduling potential energy weight. The comprehensive cost increment formed by each disturbed image relative to the original image is determined as the meta-stability energy barrier of the corresponding image state.

[0034] The meta-stability energy barrier formed by each key bending segment under forward mirror, backward mirror, removal mirror, equipment replacement mirror, neighbor mirror layer or global constraint mirror layer is compared with the pre-set meta-stability judgment threshold. It is checked whether there are interruptions in process continuity, equipment occupation conflicts, tooling reuse interruptions, test resource conflicts, semi-finished products waiting to be expanded or adaptive bending streamline detours in the corresponding mirror state. The key bending segment whose meta-stability energy barrier reaches the meta-stability judgment threshold and exhibits any of the above states is determined as a meta-stability bridging segment.

[0035] Optionally, the generation of the persistent homology-conformal latching kernel includes:

[0036] The key bending segments and non-stable bridging segments are collected to generate a candidate locking segment set. Each order process segment in the candidate locking segment set is used as the vertex of the chain complex, and the order number, process type, grid position, equipment number, material batch number, tooling number, test program number and adaptive bending streamline assignment result are written for each chain complex vertex.

[0037] Based on the process sequence relationship, material batch continuity relationship, equipment reuse relationship, tooling reuse relationship, test procedure acceptance relationship, test fixture acceptance relationship, semi-finished product flow relationship and delivery constraint relationship between the vertices of each chain complex, chain complex edges are established. Chain complex edges that can sequentially connect the chip placement process, insertion process, testing process and assembly process are combined into two-dimensional receiving cavities. Two-dimensional receiving cavities that simultaneously close time continuity, process continuity and equipment occupation are combined into time-process-equipment chain complexes.

[0038] The filtering order is set according to the local potential energy of the segment, the stability barrier, the image perturbation cost data, the process continuity strength, the test resource acceptance strength and the waiting status of the semi-finished product. The chain complex vertex, chain complex edge and two-dimensional acceptance cell in the time-process-equipment chain complex are added step by step to identify the segment connection bridge connecting two cross-process acceptance loops and the potential congestion voids formed by waiting accumulation, equipment queuing or test resource conflict.

[0039] The vertices of the chain complex in the time-process-equipment chain complex are taken as manifold nodes, and the edges of the chain complex are taken as manifold connecting edges. The mapping distance between nodes is determined based on the bearing strength, element stability energy barrier, mirror perturbation cost data and semi-finished product waiting status corresponding to each manifold connecting edge. Under the condition of keeping the node adjacency relationship and process bearing direction unchanged, a two-dimensional unfolding mapping is performed on the time-process-equipment chain complex to generate a two-dimensional conformal manifold.

[0040] Based on the segment association distance, local curvature change, cross-process acceptance loop coverage, segment connection bridge location, and potential congestion cavity coverage in the two-dimensional conformal manifold, order process segments covering cross-process acceptance loops, order process segments connecting two cross-process acceptance loops, and order process segments that expand potential congestion cavities after release are retained. The retained order process segments and their corresponding chain complex edges are combined into a continuous homology-conformal locking core.

[0041] Optionally, the intelligent scheduling scheme for generating the final order includes:

[0042] Read the disturbance type, disturbance occurrence time, affected processes, affected equipment, affected materials, affected test resources, and affected order process segments corresponding to the order insertion event, material shortage event, equipment failure event, test equipment abnormality event, or process cycle abnormality event, and generate the disturbance impact range;

[0043] Read the order process segments, relative order of segments, process succession relationship, material continuity relationship, test resource acceptance relationship, equipment occupation relationship and semi-finished product flow relationship in the continuous homology-conformal lockout kernel, and set freeze flags for the order process segments and corresponding acceptance relationships in the continuous homology-conformal lockout kernel;

[0044] Based on the scope of the disturbance, rearrangeable order process segments are read from the outside of the continuous homology-conformal lockout core. Combined with available equipment time windows, material arrival time windows, test resource idle time windows, and process continuity time windows, a set of rearrangeable segments outside the core and a set of insertable time windows are generated.

[0045] When the disturbance type is an insertion event, the insertion order is split into patch acceptance segment, plug-in acceptance segment, test acceptance segment and assembly acceptance segment. The process segment of the insertion order is combined with the set of reconfigurable segments outside the core. Under the condition that the relative order of segments within the continuous homology-conformal locking core and the corresponding acceptance relationship are not changed, candidate reconfiguration schemes are generated.

[0046] Perform feasibility verification and perturbation sorting on the candidate rearrangement schemes, and determine the candidate rearrangement scheme that passes the feasibility verification and whose perturbation sorting results meet the minimum perturbation condition as the final intelligent order scheduling scheme.

[0047] The beneficial effects of this invention are:

[0048] This invention breaks down orders in an electronic product production line into multiple order process segments and combines a multi-dimensional potential energy gating JANET network to process temporal disturbance data within a continuous time window. It can dynamically generate multi-dimensional scheduling potential energy weights based on factors such as line changeover, tooling, testing, waiting, delivery time, and anomalies. This allows order scheduling to no longer rely solely on fixed priorities or static rules, but to adjust scheduling judgment criteria in real time according to order insertions, material shortages, equipment failures, and changes in process cycle time, thereby improving the matching degree between order scheduling results and actual production status.

[0049] This invention constructs a three-dimensional grid space of time, process, and equipment, and a multi-potential well potential field. It generates adaptive curved streamlines along the gradient descent direction of the potential field and combines hierarchical metastable-bridging mirror bodies to identify key curved segments and metastable bridging segments. This enables the discovery of critical order segments that are difficult to identify by traditional scheduling methods, thus avoiding increased chip-mount line changes, repeated tooling assembly, frequent switching of test programs, expansion of semi-finished products, and congestion and transfer of subsequent workstations due to the accidental movement of such segments. This improves the stability of continuous operation of electronic product production lines across processes.

[0050] This invention generates a continuous homology-conformal locking kernel through continuous homology analysis and two-dimensional conformal manifold mapping. When a dynamic disturbance event is detected, the order process segments and corresponding succession relationships in the locking kernel are frozen. Minimal disturbance rearrangement is performed only on the order process segments outside the locking kernel. This can reduce the damage to the original stable production scheduling structure while ensuring the efficiency of order insertion response and anomaly handling, reduce the rearrangement range, reduce production cycle fluctuations, and improve equipment utilization, order on-time delivery rate, and overall production line scheduling reliability. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1The flowchart shows the intelligent scheduling and management system for electronic product production line orders based on deep learning proposed in this invention.

[0053] Figure 2 This is a schematic diagram of the structure of the multidimensional potential energy gating JANET network that generates multidimensional scheduling potential energy weights for the intelligent scheduling management system for electronic product production line orders based on deep learning proposed in this invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figure 1 and Figure 2 A deep learning-based intelligent scheduling and management system for electronic product production line orders includes:

[0056] The segment sequence construction module collects multi-source operational data from the electronic product production line, breaks down each order into multiple order process segments, and constructs an order process segment sequence.

[0057] The potential energy weight generation module inputs the order process segment sequence and the timing disturbance data within the continuous time window into the multidimensional potential energy gating JANET network. The multidimensional potential energy gating JANET network adopts a single forget gate memory unit and sets a potential energy allocation gate at the output end. By adaptively fusing the historical state decay factor and the current pulse gain, it outputs multidimensional scheduling potential energy weights.

[0058] The bending segment identification module maps order process segments to a three-dimensional grid space of time-process-equipment. Based on the multi-dimensional scheduling potential energy weight, it calculates the local potential energy corresponding to each order process segment, constructs a multi-potential-well potential field, generates adaptive bending streamlines along the gradient descent direction of the potential field, and performs gradient change analysis and curvature change analysis to determine key bending segments.

[0059] The bridging segment determination module constructs a hierarchical metastable-bridging mirror body for each key bending segment. Based on the mirror disturbance cost data corresponding to line switching, tooling, testing, waiting, queuing, delay and streamline detour in each mirror state, it calculates the metastable energy barrier and determines the metastable bridging segment.

[0060] The locking core generation module constructs a time-process-equipment chain complex based on key bending segments and non-stable bridging segments, performs continuous homology analysis, maps the time-process-equipment chain complex to a two-dimensional conformal manifold, and generates a continuous homology-conformal locking core.

[0061] When a dynamic disturbance event is detected, the disturbance rearrangement scheduling module freezes the order process segments and their corresponding acceptance relationships in the continuous coherence-conformal locking core, performs minimum disturbance rearrangement on the order process segments outside the continuous coherence-conformal locking core, and generates the final intelligent order scheduling scheme.

[0062] In this embodiment, the multi-source operational data includes order number, customer level, product model, order quantity, delivery deadline, placement program, stencil number, feeder configuration, insertion tooling, test program, test fixture, assembly version, equipment status, equipment load, equipment changeover time, material arrival status, material batch status, test equipment queuing status, semi-finished product waiting status, work-in-process queue depth, actual processing time, and order delay results.

[0063] In this embodiment, the step of splitting each order into multiple order process segments and constructing an order process segment sequence includes:

[0064] Read the product model, process route, equipment candidate set, material requirements, tooling requirements, test resource requirements, and delivery deadline for each order. Divide each order into surface mount (SMT) segment, component insertion segment, test segment, and assembly segment according to the surface mount (SMT) process, component insertion (Component Insertion) process, test segment, and assembly segment. Label each order's process segment with the order number, process type, preceding segment number, following segment number, candidate equipment number, material batch number, tooling number, test program number, estimated start time, and estimated end time. Sort each order's process segment according to process sequence, equipment occupancy, and time continuity to generate an order process segment sequence.

[0065] In this embodiment, the output multidimensional scheduling potential weight includes:

[0066] Read the order process segment sequence and timing disturbance data within the continuous time window corresponding to the current scheduling time. Encode the segment status of the process segment sequence for the operation type, equipment occupancy status, material arrival status, test resource occupancy status, sequential status of preceding and following operations, and remaining delivery time of the order. Encode the disturbance status of the timing disturbance data for changes in line switching, tooling switching, test switching, semi-finished product waiting, equipment load, material arrival, and equipment anomaly. Combine the segment status encoding results with the disturbance status encoding results to form the JANET input status.

[0067] First, at scheduling time t, all order process segments in the current queue are read, totaling 128 segments. Simultaneously, disturbance monitoring values ​​within a 40-minute rolling window are extracted. A 36-dimensional segment state vector is generated for each segment: the first 4 bits use one-hot encoding to represent four types of processes: surface mount, component insertion, testing, and assembly; bits 5-9 record the segment's occupancy status on the corresponding equipment set, written as 1 for enabled equipment and 0 for unused equipment; bits 10-14 record the arrival status of key material batches, set to 1 when the material availability rate is ≥95%; bits 15-19 record... The test resource occupancy status is checked. When a segment occupies test program T-05, fixture J-02, and test slot C-03, the corresponding bit is written as 1. Bits 20-25 indicate the continuation relationship with the preceding and following segments. The current continuation is marked as 1 if it is completed and 0 if it is to be completed. Bits 26-36 store the remaining time for order delivery, using 11 bits of binary encoding for the interval of 0-2047 minutes. The example encoding for the remaining 256 minutes is 0001000000000. At the same time, the state vectors of 128 segments are organized into a segment feature matrix P(t) of size 128×36.

[0068] The disturbance data within the same window are aggregated into a 14-dimensional disturbance vector D(t). The first 6 bits record the incremental duration and number of changes for three types of switching: line switching, tooling switching, and testing. For example, if the total duration of line switching is 32 minutes and the number of changes is 4, it is mapped to the line switching increment bit 1; if the tooling switching increment is 20 minutes and the number of changes is 2, it is mapped to the tooling increment bit 1; if the testing switching increment is 15 minutes and the number of changes is 3, it is mapped to the testing increment bit 1. The 7th to 9th bits record the average waiting time of the semi-finished product queue, the average load rate of the equipment, and the instantaneous arrival rate of materials. If the waiting time of the semi-finished product increases by 18 minutes, the equipment load rate rises to 87%, and the material arrival rate drops to 88%, the three bits are written as 1, 1, and 1 respectively. The last 5 bits identify the type of equipment abnormality and the duration marker. For example, if the test equipment M-13 experiences a 25-minute short stop, the bit sequence 01001 indicates that the test segment is abnormal and the duration is in level 2. Thus, a disturbance feature vector D(t) of size 1×14 is obtained.

[0069] Each row of the fragment feature matrix P(t) is concatenated with the perturbation feature vector D(t) in sequence to form a 128×50 JANET input matrix X(t), where the first 36 columns are ordered fragment state codes and the last 14 columns are unified perturbation state codes.

[0070] The JANET input state is input into the input mapping layer of the multidimensional potential energy gated JANET network to uniformly map the current acceptance state of the order process segment and the production line disturbance state, generating candidate memory states. Specifically, the generation of candidate memory states involves:

[0071] Each row of the JANET input matrix X(t) is input into the input mapping layer of the multidimensional potential energy-gated JANET network. First, normalization and compression are performed on the 36-dimensional segment state bits and the 14-dimensional perturbation state bits respectively, converting the binary bits and hierarchical bits into continuous features with a 0–1 scale. The 50-dimensional input vector is projected onto the 64-dimensional feature space through a set of fully connected mapping weights. The hyperbolic activation function is used to enhance the nonlinear expressive power. After projection, a candidate memory matrix of size 128×64 is obtained. For example, in the 64-dimensional candidate memory vector corresponding to sample segment P-021, the mean value of the line-related feature channel is 0.42, and the mean value of the test-related feature channel is 0.58; while in segment P-047, due to the superposition of test queuing pressure and equipment abnormality marking, the mean value of the test-related channel rises to 0.76, and the mean value of the abnormality-related channel rises to 0.63. The candidate memory matrix serves as the candidate memory state of each segment at this moment under the combined effect of the current acceptance state and the production line perturbation state.

[0072] The JANET input state and the memory state from the previous scheduling time are input into a single forget gate memory unit. The single forget gate memory unit generates a historical state decay factor based on the current disturbance strength, changes in the sequence of preceding and following processes, and the potential energy state from the previous scheduling time. Specifically, the generation of the historical state decay factor by the single forget gate memory unit based on the current disturbance strength, changes in the sequence of preceding and following processes, and the potential energy state from the previous scheduling time is as follows:

[0073] The input state matrix X(t) at the current moment and the potential energy memory matrix H(t-1) retained from the previous scheduling moment are synchronously sent to the single forget gate memory unit. The single forget gate memory unit first calculates three indicators for each order process segment: First, the disturbance intensity score within the current window, which is obtained by weighting the line change increment, test switch increment, semi-finished product waiting increment, and equipment abnormality duration, with a value range of 0 to 1; Second, the continuity change score between preceding and following processes, which is obtained by normalizing the number of continuity completion marker changes between the segment and adjacent segments, with a value range of 0 to 1; Third, the potential energy state score from the previous moment, which is taken from H(t-1). The average channel potential of the corresponding segment is normalized, with a value ranging from 0 to 1. The single forget gate memory unit weights the three scores according to preset weights to generate a history state decay factor between 0.2 and 0.9. Among them, the disturbance intensity score has the highest weight. When the test equipment queue suddenly increases and the equipment abnormality is marked as true, the history state decay factor can reach 0.85, indicating that the memory of the previous moment needs to be discarded in large quantities. When the production line is stable and the successive changes are small, the history state decay factor can be reduced to 0.25, indicating that more historical memory is retained. The history state decay factor is then applied to H(t-1) to decay and retain the old memory.

[0074] The memory state from the previous scheduling moment is attenuated and preserved based on the historical state attenuation factor. The current pulse gain is generated based on current changes such as line switching mutations, tooling mutations, testing mutations, waiting mutations, delivery pressure changes, and abnormal triggering changes. The attenuated and preserved historical memory state is then fused with the candidate memory state corresponding to the current pulse gain to generate the potential energy memory state for the current scheduling moment. Specifically, the generation of the current pulse gain is as follows:

[0075] The amplitude of the six types of disturbance indicators collected most recently within the continuous time window is detected, and the detection results are mapped into a six-dimensional gain vector according to a preset three-level gain table, specifically:

[0076] Calculate the increments of line changeover time, tooling changeover times, test changeover times, semi-finished product waiting time, delivery pressure level changes, and equipment anomaly duration. Compare each increment with its corresponding threshold: assign high gain when the line changeover time increment is ≥20 min, medium gain when it is between 10 min and 20 min, and low gain when it is less than 10 min. Use 2 and 1 changeover times as the dividing lines for tooling and test changeover increments; use 15 min and 5 min as the dividing lines for waiting time; and assign delivery pressure by a 2-level jump. Level 1 is the dividing line; equipment malfunctions are divided by durations of 20 min and 10 min. High, medium, and low gains are internally represented as 0.8, 0.5, and 0.2, respectively, corresponding to the positions of each dimension of the six-dimensional gain vector. Taking the current window as an example, if the line change increment is 24 min, tooling change increment is 3 times, test change increment is 1 time, waiting increment is 18 min, delivery level increases by 1 level, and equipment malfunction lasts for 12 min, then the six-dimensional pulse gain vector is <0.8,0.8,0.5,0.8,0.5,0.5>.

[0077] Generate the potential energy memory state at the current scheduling moment, specifically as follows:

[0078] First, the memory channel value from the previous moment is multiplied by the historical state attenuation factor to retain it. Then, the weighted sum of the candidate memory value and the pulse gain value is superimposed on the same channel: the weight of the historical part is consistent with the attenuation factor, and the combined weight of the candidate memory and the pulse gain is supplemented by the attenuation factor. Taking the test channel as an example, the historical memory value is 0.62, the attenuation factor is 0.35, the candidate memory value is 0.71, and the pulse gain is 0.50. Then, the historical contribution of 0.22 is retained. The candidate memory and the pulse gain are superimposed in the same direction to form 0.46. The two parts are added together to obtain a new test channel potential value of 0.68. If the result is greater than 1, it is truncated to 1; if it is less than 0, it is truncated to 0. The same operation is performed on each of the six channels: line switching, tooling, testing, waiting, delivery, and abnormality, to complete the full channel update of 128 segments. The result is then merged to form a new potential memory matrix, which is recorded as the potential memory state at the current scheduling moment.

[0079] The potential energy memory state at the current scheduling moment is input into the potential energy allocation gate. The potential energy allocation gate allocates the potential energy memory state according to the impact of line switching, tooling reuse, test resource acceptance, semi-finished product waiting, delivery pressure, and equipment anomaly. It outputs the multi-dimensional scheduling potential energy weights corresponding to line switching, tooling, testing, waiting, delivery, and anomaly. Specifically, the allocation of the potential energy memory state and the output of the multi-dimensional scheduling potential energy weights corresponding to line switching, tooling, testing, waiting, delivery, and anomaly are as follows:

[0080] The potential energy memory state generated at the current scheduling moment is fed into the potential energy allocation gate. The mean values ​​of the six sets of channel features corresponding to line switching, tooling, testing, waiting, delivery date, and anomaly in the memory state are calculated respectively. The six means are proportionalized by summation to obtain a set of six-dimensional normalized weight vectors. When the proportion of the line switching mean, testing mean, and waiting mean is high, the corresponding weights are automatically increased. When the proportion of the delivery date mean and anomaly mean decreases, the corresponding weights are automatically decreased. Finally, the potential energy weights for line switching, tooling, testing, waiting, delivery date, and anomaly are output.

[0081] In this embodiment, determining the key bending segment includes:

[0082] The system reads the scheduling start time, scheduling end time, process type set, and equipment candidate set of the order process segment sequence. It divides the time range from the scheduling start time to the scheduling end time into continuous time grids according to a preset scheduling granularity. It sets the placement process, insertion process, testing process, and assembly process as process grids, and sets the available equipment or workstations for each process as equipment grids. The continuous time grids, process grids, and equipment grids are combined to generate a three-dimensional time-process-equipment raster space. Specifically, the division of the time range from the scheduling start time to the scheduling end time into continuous time grids according to a preset scheduling granularity is as follows:

[0083] The scheduling engine reads the start and end times of the current round of scheduling. Assuming the start time is the 1st minute of the day and the end time is the 480th minute, the total scheduling duration is 479 minutes. The time period is discretized according to the preset scheduling granularity: if the granularity is set to 10 minutes, a time grid is divided every 10 minutes from the start time. When the remaining time is less than 10 minutes, a tail grid is still opened, so that the time axis is divided into 48 equal-width main time grids; if the granularity is changed to 5 minutes, the time axis is discretized into 96 time grids.

[0084] A three-dimensional time-process-equipment raster space is generated by combining continuous time grids, process grids, and equipment grids, specifically as follows:

[0085] First, four fixed process grids are established for the process dimension: SMT1-SMT4, Component Insertion, Test, and Assembly, identified by R-SMD, R-AI, R-ATE, and R-ASM respectively. Then, the equipment resource table is searched, and the SMT1-SMT4 of the SMT1 segment, AI1-AI3 of the Component Insertion segment, ATE1-ATE6 of the Test segment, and ASM1-ASM4 of the Assembly segment are registered as equipment grids in sequence, for a total of 17 equipment grids. Each equipment grid records the equipment number and its current availability status. A triple loop is used to traverse the time grid, process grid, and corresponding equipment grid: within the time grid T1, four processes are generated first. First, lay out four device units (SMT1-SMT4) in the surface mount layer, three device units (AI1-AI3) in the through-hole layer, six device units (ATE1-ATE6) in the test layer, and four device units (ASM1-ASM4) in the assembly layer. After completion, switch to time grid T2 and repeat the same steps. Iterate in this way until time grid T48, finally forming 48×4×17=3264 independent cubic grids. Each grid has a unique index (time grid number, process grid identifier, device grid number) and an idle / occupied flag, forming a three-dimensional grid space of time-process-device.

[0086] Read the executable time window, the process to which it belongs, the candidate equipment, the processing time, the preceding segment and the following segment corresponding to each order process segment, map each order process segment to the grid position that matches the executable time window, the process to which it belongs and the candidate equipment, and establish the process succession relationship between adjacent grids based on the succession relationship between the preceding segment and the following segment;

[0087] Read the multi-dimensional scheduling potential energy weights, generate segment local potential energy based on the changeover cost, tooling cost, testing cost, waiting cost, delivery cost, and exception cost corresponding to each order's process segment, write the segment local potential energy into the corresponding grid position, and transfer the segment local potential energy to adjacent grids according to time sequence relationship, process sequence relationship, and equipment occupancy relationship to construct a multi-potential-well potential field, where:

[0088] The local potential energy of the generated segment is as follows:

[0089] Read the output six-dimensional scheduling potential energy weights, and then extract the corresponding line change time increment, tooling change number increment, test change number increment, semi-finished product waiting time increment, delivery pressure level increment, and equipment abnormality duration increment for each order process segment. Normalize each increment to the 0-1 range according to its own baseline threshold; multiply each item by the weight and the normalization cost and sum them to obtain a 0-1 floating-point value as the segment's local potential energy.

[0090] Constructing a multi-potential-well potential field, specifically as follows:

[0091] The local potential energy of each segment is written into the starting grid in the 3D grid. The grid is filled along the time axis according to the number of time grids continuously occupied by the processing time. The potential energy value is decayed by 0.1 as the time diffusion decay coefficient. According to the process sequence relationship, a connection edge is established between the same equipment grid at the end time grid of the current segment and the starting time grid of the next segment. The potential energy is transferred to the starting grid of the subsequent segment with a decay coefficient of 0.6. If there are segments with shared stencils, shared test programs or shared plug-in tooling in the same time grid, a process adjacency edge is established between the corresponding equipment grids, and the potential energy is diffused laterally with a decay coefficient of 0.3. After 128 segments are written and diffused, an uneven potential energy distribution is formed in the 48×4×17 3D grid. Significant potential wells are formed in the SMT2-T15 and SMT3-T16 positions in the chip area, and a secondary potential well is formed near ATE4-T27 in the test area.

[0092] In a multi-potential well field, starting from the preceding grid position of each order's process segment, grid positions that satisfy the relationships of process continuity, equipment occupancy, material arrival, and test resource acceptance are selected sequentially along the direction of local potential energy decrease in the segment. This forms an adaptive curved streamline that runs through the placement, insertion, testing, and assembly processes. Specifically, the adaptive curved streamline that runs through the placement, insertion, testing, and assembly processes is as follows:

[0093] Using the grid where the placement segment of an order is located as the starting point of the streamline, the system searches the time grid and candidate equipment grid for grids that meet the conditions of increasing time, progressive process, available equipment, material availability, and test resource acceptance. From these grids, the grid with the largest decrease in local potential energy is selected as the next node. Taking order O-06 as an example, placement segment P-021 is located in <T8, R-SMD, SMT2>. <T13, R-AI, AI2>, where the potential energy decreases from 0.62 to 0.54, is selected to connect to insertion segment P-022. Then, <T18, R-ATE, ATE2>, where the potential energy decreases to 0.49 and the test procedure is consistent, is selected to connect to the test segment. Finally, <T23, R-ASM, ASM3> is selected to connect to the assembly segment. If the path sequentially passes through four types of process grids—placement, insertion, testing, and assembly—and meets the adjacent node connection conditions, it is determined to be an adaptive curved streamline. If multiple feasible paths exist, the path with the minimum cumulative local potential energy, the fewest equipment switching times, and the highest test procedure continuity is retained.

[0094] For each adaptive curved streamline, streamline gradient change analysis and curvature change analysis are performed on the order process segments. Order process segments located at abrupt changes in potential energy decrease direction, abrupt changes in streamline curvature direction, inter-process confluence points, or streamline detour initiation points are identified as critical curved segments. Specifically, the streamline gradient change analysis and curvature change analysis performed on each adaptive curved streamline are as follows:

[0095] Calculate the local potential energy difference between adjacent nodes on the streamline. When the potential energy difference between a segment and the preceding or following node is more than 1.5 times the average difference of the entire streamline, the segment is determined to be located at a position where the potential energy decreases abruptly.

[0096] The streamline turning angle at the statistical segment is used. When the turning angle exceeds 45° and the arc length of the two segments before and after is not less than 2 nodes, the segment is determined to be located at the position where the streamline bending direction changes abruptly.

[0097] If a test segment intersects with other streamlines in the same process, equipment, or testing resource, or if it branches off and forms a new path after the segment, these segments are marked as cross-process confluence positions and streamline detour starting positions, respectively, to identify key bending segments.

[0098] In this embodiment, determining the stable bridging segment includes:

[0099] Read the order number, process type, current grid position, previous segment, subsequent segment, candidate equipment, material batch, tooling number, test program number and adaptive bending streamline assignment result corresponding to each key bending segment, and generate the baseline state of the key bending segment;

[0100] A local mirror layer is constructed based on the baseline state of the key bending segment. This local mirror layer includes in-situ mirrors, forward-moving mirrors, backward-moving mirrors, removed mirrors, and replaced device mirrors. The execution time, device usage, tooling usage, test resource usage, preceding continuation state, and following continuation state are recorded for each local mirror. Specifically, the construction of the local mirror layer based on the baseline state of the key bending segment is as follows:

[0101] The current execution time, device number, tooling number, and test resource number of the key bending segment are copied to generate an in-situ image, and the status of its preceding sequence being completed and its subsequent sequence being pending is recorded synchronously.

[0102] Search forward for the idle time window of the process where the segment is located. If the idle window is longer than the segment processing time and the materials, tooling, and testing resources are all available, then advance the segment execution time index to the start of the idle window, keep the equipment unchanged, generate a forward-moving image, and update the previous sequence continuation status to incomplete.

[0103] Search backwards for available time windows for the same process. If an idle window is found that meets the processing time and the subsequent process time does not conflict, the execution time of the segment is postponed to the beginning of the idle window, a backward mirror is generated, and the subsequent continuation status is marked as delayed.

[0104] The baseline segment is temporarily removed from the current scheduling sequence, the execution time is recorded as empty, the usage of equipment, tooling, and test resources is set to zero, a removal image is generated, and the connection status before and after is marked as disconnected.

[0105] Within the same time window, find backup equipment or workstations. If the backup equipment meets the process compatibility requirements and the cost of changing the line is lower than that of the original equipment, keep the execution time unchanged, replace the equipment number, maintain the original occupancy relationship between tooling and test resources, generate a replacement equipment image, and record the equipment change mark as replaced.

[0106] Centered on the critical bending segment, preceding segments, subsequent segments, segments adjacent to the same equipment, and segments adjacent to the same test resources are read to construct a neighborhood mirror layer. Within the same scheduling time window, order process segments that have material continuity, process succession, or test resource acceptance relationships with the critical bending segment are read to construct a global constraint mirror layer. The hierarchical meta-stable-bridging mirror body is composed of the local mirror layer, the neighborhood mirror layer, and the global constraint mirror layer, where:

[0107] The neighborhood mirror layer is constructed as follows: taking the key bending segment as the center, the directly preceding segment, the directly following segment, the adjacent segments on the same device within the preceding and following 20-minute time windows, and the segments queued adjacently on the same test program or test fixture are retrieved in sequence. The retrieved segments are packaged together with the key bending segment itself to generate a neighborhood mirror set. For each segment in the set, five types of local mirrors are copied and retained, moved forward, moved backward, removed, and replaced to evaluate the chain effect of local disturbances in a narrow range.

[0108] Construct a global constraint mirror layer, specifically: within the same scheduling time window as the key bending segment, read all order process segments that have continuous material batches, process succession dependencies, or test resource dependencies, fix the segments as the baseline state without changing their position or equipment, and record the process sequence, material binding, and test resource occupancy as hard constraints. When there are local or neighboring mirror disturbances, keep the segments and dependencies within the global constraint mirror layer constant, and limit the maximum impact boundary of the mirror disturbance.

[0109] Mirror disturbance cost data for each mirror state, including line switching, tooling, testing, waiting, queuing, delay, and streamline detour, are collected. The mirror disturbance cost data are then weighted and aggregated according to multi-dimensional scheduling potential energy weights. The comprehensive cost increment of each disturbed mirror state relative to the original mirror state is determined as the meta-stability energy barrier for that mirror state.

[0110] The image perturbation cost data is weighted and aggregated according to the multidimensional scheduling potential energy weights, specifically as follows:

[0111] The seven increments—line changeover time increment, tooling changeover count increment, test changeover count increment, semi-finished product waiting time increment, test equipment queuing time increment, order delivery offset time increment, and streamline detour marker—are read separately. These increments are then normalized and allocated to six potential energy dimensions: line changeover increment is assigned to the line changeover dimension, tooling changeover increment to the tooling dimension, test changeover and queuing increments are both assigned to the test dimension, waiting increment to the waiting dimension, delivery offset increment to the delivery date dimension, and streamline detour marker to the anomaly dimension. The six-dimensional scheduling potential energy weights of the current scheduling cycle are then called, and the increments of each dimension are multiplied bitwise to obtain six weighted cost values. Finally, the six weighted values ​​are summed to form the comprehensive cost of a single mirror disturbance.

[0112] The incremental cost of each disturbed mirror image relative to the preserved in-situ mirror image is defined as the meta-stability energy barrier of the corresponding mirror image state, specifically:

[0113] Using the overall cost of preserving the original image as a benchmark, the overall costs of moving forward, moving backward, removing and replacing device images, as well as neighboring and global images generated for the same segment are compared with the benchmark value one by one. A positive difference indicates an increase in cost, and the magnitude of the difference is the meta-stability barrier of the image; if the difference is negative, it is considered that the meta-stability barrier is zero and the cost is recorded as reduced.

[0114] The meta-stability energy barriers formed by each key bending segment under forward mirroring, backward mirroring, removal mirroring, equipment replacement mirroring, neighbor mirroring layer, or global constraint mirroring layer are compared with a pre-set meta-stability judgment threshold. It is checked whether process continuity interruption, equipment occupation conflict, tooling reuse interruption, test resource conflict, semi-finished product waiting for expansion, or adaptive bending streamline detour occurs under the corresponding mirroring state. Key bending segments whose meta-stability energy barriers reach the meta-stability judgment threshold and exhibit any of the above states are identified as meta-stability bridging segments, wherein:

[0115] The elemental stability energy barrier formed by each key bending segment under forward mirroring, backward mirroring, removal mirroring, device replacement mirroring, neighbor mirroring layer, or global constraint mirroring layer is compared with a pre-set elemental stability judgment threshold. Specifically:

[0116] For each critical bending segment, an energy barrier comparison table containing six types of mirror images is established. The comprehensive cost increment value of the corresponding mirror image is read, and the increment value is compared with the current production line's calibrated non-stability judgment threshold of 0.6 one by one. If the increment value is greater than or equal to the threshold, the mirror image is marked as having met the energy barrier in the comparison table; otherwise, it is marked as having insufficient energy barrier. Then, the number of mirror images that meet the energy barrier for the same segment is counted according to the comparison table. If at least one type of mirror image meets the energy barrier and any one of the following is detected at the same time: process continuity interruption, equipment occupation conflict, tooling reuse interruption, test resource conflict, semi-finished product waiting for expansion, or bending flow line detour, the critical bending segment is identified as a non-stability bridging segment, and the triggered mirror image type and corresponding instability cause are recorded.

[0117] In this embodiment, generating a persistent homology-conformal latching kernel includes:

[0118] The key bending segments and non-stable bridging segments are collected to generate a candidate locking segment set. Each order process segment in the candidate locking segment set is used as the vertex of the chain complex, and the order number, process type, grid position, equipment number, material batch number, tooling number, test program number and adaptive bending streamline assignment result are written for each chain complex vertex.

[0119] Based on the process sequence relationships, material batch continuity relationships, equipment reuse relationships, tooling reuse relationships, test procedure acceptance relationships, test fixture acceptance relationships, semi-finished product flow relationships, and delivery constraint relationships among the vertices of each chain complex, chain complex edges are established. Chain complex edges that can sequentially connect the surface mount process, insertion process, testing process, and assembly process are combined into two-dimensional receiving cavities. Two-dimensional receiving cavities that simultaneously close due to time continuity, process continuity, and equipment occupancy are combined into time-process-equipment chain complexes. Specifically, the combination of chain complex edges that can sequentially connect the surface mount process, insertion process, testing process, and assembly process into two-dimensional receiving cavities is as follows:

[0120] Using the same order number as an index, retrieve the chain complex vertices of the order in the four processes of surface mount, insertion, testing and assembly. If the four vertices are strictly increasing on the time axis and adjacent vertices are connected by chain complex edges through at least one relationship such as continuous material batch, test program acceptance or tooling reuse, then they are closed into a ring edge set in the order of <surface mount → insertion>, <insertion → testing>, and <testing → assembly>. The closed edge set is defined as a two-dimensional receiving cell. The order number, cross-process sequence and associated equipment number are recorded in the cell metadata.

[0121] Combining the two-dimensional receiving cavities that simultaneously close time continuity, process continuity, and equipment occupancy into a time-process-equipment chain complex is as follows:

[0122] Within the same time frame, all formed two-dimensional receiving cavities are retrieved and aggregated. If the two-dimensional receiving cavities are continuous in the time frame, have a complete process sequence, and occupy equipment without conflict, then adjacent cavities are spliced ​​into a volume unit according to three-dimensional coordinates, using the time frame number as the vertical index, the process layer number as the horizontal index, and the equipment number as the depth index. When the four layers of cavities of patching, plugging, testing, and assembly in any three adjacent time frames are spliced, it is considered as a three-dimensional closure of time-process-equipment. The volume unit is uniformly identified as a time-process-equipment chain complex. The order flow, equipment occupation, and material flow relationship are simultaneously marked on the vertices, edges, and cavities of the complex to complete the construction of the time-process-equipment chain complex.

[0123] Based on the local potential energy of segments, the stability barrier, the cost of mirror perturbation data, the strength of process continuity, the strength of test resource acceptance, and the waiting status of semi-finished products, a filtering order is set. A step-by-step addition analysis is performed on the chain complex vertices, chain complex edges, and two-dimensional acceptance cavities in the time-process-equipment chain complex. This identifies segment connection bridges connecting two cross-process acceptance loops and potential congestion voids formed by waiting accumulation, equipment queuing, or test resource conflicts. Among these:

[0124] A stepwise addition analysis was performed on the chain complex vertices, chain complex edges, and two-dimensional receiving cavities in the time-process-equipment chain complex, specifically as follows:

[0125] Sort the fragments by local potential energy from low to high, and add vertices with potential energy below 0.40 to the complex filter view in batches. Based on the view, examine the associated edges in sequence. If the meta-stability barrier of the associated edge is not higher than 0.55 and the mirror perturbation cost is lower than the current window average, add the edge to the view and update the process continuity strength. When the test resource acceptance strength is higher than 0.70 and the semi-finished product waiting status is in the normal range, the corresponding two-dimensional acceptance cavity is activated and added to the view. Then, increase the threshold and expand the upper limit of potential energy to 0.60, and repeat the above vertex-edge-cavity addition process. When the threshold is increased to 0.80, the complex filter view has covered all acceptance loops. Add the remaining elements that have not yet entered the view at once, and count the test queuing, waiting accumulation and conflict node positions caused by the new elements. Identify the fragment connection bridge connecting two cross-process acceptance loops, as well as potential congestion holes formed by semi-finished product waiting accumulation, equipment queuing or test resource conflicts.

[0126] Using the vertices of the time-process-equipment chain complex as manifold nodes and the edges as manifold connecting edges, the mapping distance between nodes is determined based on the bearing strength, elemental stability barrier, mirror perturbation cost data, and semi-finished product waiting state corresponding to each manifold connecting edge. While maintaining the node adjacency relationship and process bearing direction, a two-dimensional unfolding mapping is performed on the time-process-equipment chain complex to generate a two-dimensional conformal manifold, where:

[0127] Determine the mapping distance between nodes, specifically as follows:

[0128] For each chain complex edge, a comprehensive continuity weight is calculated. The continuity weight is composed of four factors: the strength of the connection is weighted positively, with a higher value indicating a tighter connection; the energy barrier of the immature stability is weighted negatively, with a higher barrier indicating a weaker connection; the cost of mirror perturbation is weighted negatively, with a higher cost indicating a greater risk; and the waiting state of the semi-finished product is weighted negatively, with a longer waiting time indicating poorer liquidity. After proportional normalization, the four factors are summarized into a single weight in the 0-1 range. The reciprocal of this weight is taken and linearly scaled to obtain the mapping distance of the corresponding manifold connection edge. The larger the weight, the shorter the distance, and the smaller the weight, the longer the distance.

[0129] Generate a two-dimensional conformal manifold, specifically:

[0130] Using the vertices of the chain complex as initial coordinate points, a force-oriented algorithm is first used to iteratively distribute points on a two-dimensional plane, gradually bringing the distance between nodes closer to the target distance. Then, a local angle preservation correction is performed on the iteration results to ensure that the angle between adjacent nodes changes within 5°. After completion, the entire plane is stretched and normalized to eliminate the boundary contraction effect, while retaining the node adjacency matrix. The final output node distribution is the two-dimensional conformal manifold, where the straight-line distance between nodes in the manifold reflects the comprehensive continuity risk in the three-dimensional chain complex, and the concentrated curvature area of ​​the manifold corresponds to the location of potential congestion cavities.

[0131] Based on the segment association distance, local curvature change, cross-process acceptance loop coverage, segment connection bridge location, and potential congestion cavity coverage in the two-dimensional conformal manifold, order process segments covering cross-process acceptance loops, order process segments connecting two cross-process acceptance loops, and order process segments that expand potential congestion cavities after release are retained. The retained order process segments and their corresponding chain complex edges are combined into a continuous homology-conformal locking core, where:

[0132] The retained order process segments and corresponding chain complex edges are combined into a continuous homology-conformal locking core, specifically:

[0133] A segment labeling table is established on a two-dimensional conformal manifold. Segments located inside cross-process receiving loops are labeled as "inside loops," segments located at both ends of segment connecting bridges and whose distance from the bridge center does not exceed a set threshold are labeled as "bridge ends," and segments located at the boundaries of potential congestion cavities and whose removal would expand the cavity area are labeled as "boundaries." Subsequently, all chain complex edges are traversed. Edges connecting nodes in two loops are assigned a closure label, edges connecting nodes inside loops and bridge end nodes are assigned a bridging label, and edges connecting boundary nodes and whose removal would cause the cavity area to increase beyond a threshold are labeled. Assign locking edge identifiers; gather all nodes and edges of the three types of identifiers to form the first version of the fragment-edge set; then check the integrity of the set according to the process sequence. If a continuous path is missing in any process layer such as patching, plugging, testing, or assembly, backtrack and add the candidate edge with the second lowest weight that was eliminated in the previous step until the path is closed; finally, perform a redundancy removal process on the set, removing edges that can be bypassed by two or more alternative paths in the 3D chain complex without affecting the closure of the loop, to obtain the smallest node-edge closed block, i.e., the continuous homology-conformal locking kernel.

[0134] In this embodiment, the intelligent scheduling scheme for generating the final order includes:

[0135] Read the disturbance type, disturbance occurrence time, affected processes, affected equipment, affected materials, affected test resources, and affected order process segments corresponding to the order insertion event, material shortage event, equipment failure event, test equipment abnormality event, or process cycle abnormality event, and generate the disturbance impact range;

[0136] Read the order process segments, relative order of segments, process succession relationship, material continuity relationship, test resource acceptance relationship, equipment occupation relationship and semi-finished product flow relationship in the continuous homology-conformal lockout kernel, and set freeze flags for the order process segments and corresponding acceptance relationships in the continuous homology-conformal lockout kernel;

[0137] Based on the scope of the disturbance, rearrangeable order process segments are read from the outside of the continuous homology-conformal lockout core. Combined with available equipment time windows, material arrival time windows, test resource idle time windows, and process continuity time windows, a set of rearrangeable segments outside the core and a set of insertable time windows are generated.

[0138] When the disturbance type is an insertion event, the insertion order is split into patch acceptance segment, plug-in acceptance segment, test acceptance segment and assembly acceptance segment. The process segment of the insertion order is combined with the set of reconfigurable segments outside the core. Under the condition that the relative order of segments within the continuous homology-conformal locking core and the corresponding acceptance relationship are not changed, candidate reconfiguration schemes are generated.

[0139] Feasibility verification and perturbation sorting are performed on candidate rearrangement schemes. The candidate rearrangement scheme that passes the feasibility verification and whose perturbation sorting results meet the minimum perturbation condition is determined as the final intelligent order scheduling scheme. The feasibility verification includes equipment occupancy verification, material arrival verification, test resource occupancy verification, process continuity verification, delivery deadline verification, and continuous coherence-conformal interlocking core integrity verification. The perturbation sorting is determined based on the number of rearranged segments, segment start time offset, number of equipment changes, number of line changes, number of test switching, changes in semi-finished product waiting time, and order delay marking.

[0140] Example 1: In a continuous production shift of electronic products, the system receives a batch of multi-source operation data from the production execution system, scheduling system, equipment acquisition system, and test management system. This data includes 32 orders, 7 product models, and a planned output of 18,600 units. The production line comprises 4 surface mount lines, 3 component insertion stations, 6 functional testing devices, and 4 assembly stations. Of the original orders, 11 orders share the same type of stencil, 9 orders share the same test program, and 6 orders use the same component insertion fixtures. The system reads an average equipment load rate of 76.4%, a material arrival rate of 91.8%, an average queue length of 5.7 segments for testing devices, and a historical average changeover time of 17.6 minutes.

[0141] After the data enters the processing flow, the system breaks down the 32 orders into 128 order process segments according to placement, insertion, testing, and assembly. Each segment is recorded with process type, candidate equipment, material batch, tooling number, test program number, estimated processing time, and the successive relationship between preceding and following segments. For example, order O-06 is broken down into P-021, P-022, P-023, and P-024, where P-021 is the placement segment with a processing time of 42 minutes, using stencil S-03 and feeder configuration F-07; P-023 is the testing segment, using test program T-05 and fixture J-02, with an estimated testing time of 31 minutes. After the splitting is completed, the system obtains 128 segment nodes and 384 successive relationship records.

[0142] During the training phase, the system used 8600 historical samples, including 6200 order process segment samples and 2400 disturbance samples. In sample A, three consecutive patch segments shared stencil S-02, with 0 actual line changes and a semi-finished product waiting time of 14 minutes. In sample B, after test program T-04 was inserted into other programs, the test switching increased by 2 times, and the test queuing time increased from 18 minutes to 53 minutes. In sample C, a certain plug-in tooling was released prematurely, increasing the waiting time for subsequent segments with the same tooling by 27 minutes. In sample D, equipment malfunction lasted for 22 minutes, resulting in three delayed segments after the original production schedule was shifted. The system used these samples to train a multi-dimensional potential energy gating JANET network, enabling it to learn the impact of line changes, tooling, testing, waiting, delivery time, and abnormal factors on scheduling under different disturbance states.

[0143] In the current production shift, the system uses 12 consecutive scheduling windows as input, each window lasting 10 minutes. Input timing disturbance data shows an average line changeover fluctuation of 16.9 minutes, tooling reinstallation of 11.4 minutes, test switching of 8.6 minutes, and semi-finished product waiting time of 33.2 minutes. Equipment anomaly flags appeared twice. The multi-dimensional potential energy gating JANET network first retains the effective line changeover and test impacts from the previous state through a single forget gate memory unit, and then generates the current pulse gain based on the current increase in test queue, increased order insertion pressure, and equipment anomaly occurrences. The multi-dimensional scheduling potential energy weights given by the output potential energy allocation gate are: line changeover 0.24, tooling 0.16, testing 0.22, waiting 0.18, delivery time 0.13, and anomaly 0.07. Compared to the static weights, the testing and waiting weights increased by 0.08 and 0.05 respectively, indicating that the current production bottleneck has shifted from simple delivery time pressure to pressure related to test acceptance and semi-finished product flow.

[0144] The system then constructs a three-dimensional grid space encompassing time, process, and equipment. The scheduling scope covers 360 minutes, divided into 36 time grids with a 10-minute granularity. The process dimension is set into four grids: patching, insertion, testing, and assembly. The equipment dimension is set into 17 equipment or workstation grids, resulting in 2448 mappable grids. The system maps 128 order process segments to corresponding executable time windows, their respective processes, and candidate equipment grids, and calculates the local potential energy of each segment based on multi-dimensional scheduling potential energy weights. After calculation, the local potential energy of segment P-021 is 0.62, segment P-023 has a local potential energy of 0.81 due to a long queue in test program T-05, and segment P-037 has a local potential energy of 0.88 due to steel mesh switching and delivery time pressure.

[0145] After generating the multi-potential well potential field, the system tracks the low-dissipation path of the order segment along the direction of potential energy decrease. The system generates a total of 19 adaptive curved streamlines, 12 of which can completely penetrate the patch, insertion, testing, and assembly stages. The 5th streamline includes P-017, P-018, P-019, P-020, P-031, and P-032, with the average potential energy decreasing from 0.79 to 0.46. The 8th streamline shows a significant inflection point at P-047, with the potential energy of the preceding segment being 0.51 and the subsequent segment 0.84, and the bending direction changing from test equipment M-11 to M-14. Based on the changes in streamline gradient and curvature, the system identifies 18 key curved segments, including 7 patch segments, 3 insertion segments, 6 testing segments, and 2 assembly segments.

[0146] For 18 key bending segments, the system constructs hierarchical metastable-bridging mirror images. Each segment generates five types of local mirror images: in-situ retention, forward movement, backward movement, removal, and equipment replacement. These are combined with preceding segments, subsequent segments, segments adjacent to the same equipment, and segments adjacent to the same test resource to generate neighborhood mirror images, resulting in a total of 126 mirror states. Taking segment P-047 as an example, in the in-situ retention state, test program T-05 executes continuously, with a test queue time of 21 minutes; in the forward movement mirror image, fixture J-02 conflicts with P-044, increasing the queue time to 58 minutes; in the backward movement mirror image, the assembly section wait time increases by 34 minutes; and in the removal mirror image, the 8th bending streamline detours, adding one line change. After aggregating these mirror disturbance costs according to multi-dimensional scheduling potential energy weights, the system obtains a metastable energy barrier of 0.76 for P-047, which is higher than the preset judgment threshold of 0.60, thus identifying it as a metastable bridging segment. After calculating all mirror images, the system identifies 12 metastable bridging segments from the 18 key bending segments.

[0147] The system merged and deduplicated 18 key bending segments and 12 non-stable bridging segments, resulting in 23 candidate locking segments. Using these 23 segments as chain complex vertices, the system established 68 chain complex edges based on process continuity, material continuity, equipment reuse, tooling reuse, test acceptance, and semi-finished product flow relationships, forming 9 two-dimensional acceptance cavities. Continuous homology analysis showed that 4 cross-process acceptance loops remained for more than 80 minutes, 3 segment connecting bridges served as transition connections between two acceptance loops, and 2 potential congestion cavities were located at the junction of the test and assembly sections. After the system unfolded the chain complex into a conformal manifold, it found that 7 segments were covered by cavities in the test section, with local curvature concentrated near P-047, P-052, and P-068. The system ultimately retained 15 order process segments and 41 acceptance edges, generating a continuous homology-conformal locking core.

[0148] At the 140-minute mark of production execution, the system detected a dynamic disturbance event: a new insert order of 1200 units was added, reducing the remaining delivery time to 260 minutes; simultaneously, material batch B-06 was delayed by 40 minutes, and test equipment M-13 experienced a 25-minute abnormal shutdown. Traditional scheduling methods directly insert the insert order and globally adjust 58 segments, increasing the number of patch line changes from 8 to 14, test program switches from 9 to 17, and the average waiting time for semi-finished products from 31.6 minutes to 49.8 minutes. This invention freezes 15 segments and 41 receiving edges in the continuous coherence-conformal locking core, performing minimal disturbance rearrangement only on the 37 segments outside the core. After the insert order is split into 4 segments, the system inserts its patch segment into the outside patch window at the 170-minute mark, inserts the test segment into the gap window after test program T-05 ends, and postpones the execution of the 2 outside segments affected by material shortages until after material arrival. Ultimately, only 24 segments experienced time adjustments, and 6 segments experienced device adjustments, with the continuous cohomology-conformal locking core remaining intact.

[0149] In the same batch of 32 orders and under the same set of disturbance events, the system compares the method of this invention with the traditional rule scheduling method. The training samples are 8600 sets and the test simulation segments are 128. The traditional method involves 58 rearranged segments, while this invention involves 24; the traditional method requires 14 SMT line changes, while this invention requires 9; the traditional method requires 19 feeder reconfigurations, while this invention requires 11; the traditional method requires 17 test program switches, while this invention requires 10; the traditional method requires 15 test fixture switches, while this invention requires 8; the traditional method has an average waiting time of 49.8 minutes for semi-finished products, while this invention requires 26.4 minutes; the traditional method has an average queuing time of 42.7 minutes for test equipment, while this invention requires 23.5 minutes; the traditional method has a cumulative waiting time of 88 minutes for assembly section idling, while this invention requires 36 minutes; the traditional method has 5 delayed orders, while this invention requires 2; the traditional method has an on-time order delivery rate of 84.4%, while this invention achieves 93.8%; and the traditional method has an equipment utilization rate of 79.6%, while this invention achieves 86.7%.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent scheduling and management system for electronic product production line orders, characterized in that: include: The segment sequence construction module collects multi-source operational data from the electronic product production line, breaks down each order into multiple order process segments, and constructs an order process segment sequence. The potential energy weight generation module inputs the order process segment sequence and the timing disturbance data within the continuous time window into the multidimensional potential energy gating JANET network. The multidimensional potential energy gating JANET network adopts a single forget gate memory unit and sets a potential energy allocation gate at the output end. By adaptively fusing the historical state decay factor and the current pulse gain, it outputs multidimensional scheduling potential energy weights. The bending segment identification module maps order process segments to a three-dimensional grid space of time-process-equipment. Based on the multi-dimensional scheduling potential energy weight, it calculates the local potential energy corresponding to each order process segment, constructs a multi-potential-well potential field, generates adaptive bending streamlines along the gradient descent direction of the potential field, and performs gradient change analysis and curvature change analysis to determine key bending segments. The bridging segment determination module constructs a hierarchical metastable-bridging mirror body for each key bending segment. Based on the mirror disturbance cost data corresponding to line switching, tooling, testing, waiting, queuing, delay and streamline detour in each mirror state, it calculates the metastable energy barrier and determines the metastable bridging segment. The locking core generation module constructs a time-process-equipment chain complex based on key bending segments and non-stable bridging segments, performs continuous homology analysis, maps the time-process-equipment chain complex to a two-dimensional conformal manifold, and generates a continuous homology-conformal locking core. When a dynamic disturbance event is detected, the disturbance rearrangement scheduling module freezes the order process segments and their corresponding acceptance relationships in the continuous coherence-conformal locking core, performs minimum disturbance rearrangement on the order process segments outside the continuous coherence-conformal locking core, and generates the final intelligent order scheduling scheme.

2. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The multi-source operational data includes order number, customer level, product model, order quantity, delivery deadline, placement program, stencil number, feeder configuration, insertion tooling, test program, test fixture, assembly version, equipment status, equipment load, equipment changeover time, material arrival status, material batch status, test equipment queuing status, semi-finished product waiting status, work-in-process queue depth, actual processing time, and order delay results.

3. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The step of splitting each order into multiple order process segments and constructing an order process segment sequence includes: Read the product model, process route, equipment candidate set, material requirements, tooling requirements, test resource requirements, and delivery deadline for each order. Divide each order into surface mount (SMT) segment, component insertion segment, test segment, and assembly segment according to the surface mount (SMT) process, component insertion (Component Insertion) process, test segment, and assembly segment. Label each order's process segment with the order number, process type, preceding segment number, following segment number, candidate equipment number, material batch number, tooling number, test program number, estimated start time, and estimated end time. Sort each order's process segment according to process sequence, equipment occupancy, and time continuity to generate an order process segment sequence.

4. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The output multidimensional scheduling potential weights include: Read the order process segment sequence and the timing disturbance data within the continuous time window corresponding to the current scheduling time. Encode the segment status of the process type, equipment occupancy status, material arrival status, test resource occupancy status, sequential status of preceding and following processes, and remaining delivery time of the order in the order process segment sequence. Encode the disturbance status of the line change change, tooling change change, test change change, semi-finished product waiting change, equipment load change, material arrival change, and equipment abnormality change in the timing disturbance data. Combine the segment status encoding results and the disturbance status encoding results into the JANET input status. The JANET input state is input into the input mapping layer of the multidimensional potential energy gated JANET network to uniformly map the current acceptance state of the order process segment and the production line disturbance state, and generate candidate memory states. The JANET input state and the memory state of the previous scheduling time are input into the single forget gate memory unit. The single forget gate memory unit generates the historical state decay factor based on the current disturbance intensity, the changes in the succession of the preceding and following processes, and the potential energy state of the previous scheduling time. The memory state of the previous scheduling moment is attenuated and retained based on the historical state attenuation factor. The current pulse gain is generated based on the current line switching change, tooling change, test change, waiting change, delivery pressure change and abnormal trigger change. The attenuated and retained historical memory state is fused with the candidate memory state corresponding to the current pulse gain to generate the potential energy memory state of the current scheduling moment. The potential energy memory state at the current scheduling moment is input into the potential energy allocation gate. The potential energy allocation gate allocates the potential energy memory state according to the impact of line switching, tooling reuse, test resource acceptance, semi-finished product waiting, delivery pressure, and equipment abnormality, and outputs the multi-dimensional scheduling potential energy weights corresponding to line switching, tooling, testing, waiting, delivery, and abnormality.

5. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The determination of key bending segments includes: Read the scheduling start time, scheduling end time, process type set and equipment candidate set of the order process segment sequence, divide the time range from the scheduling start time to the scheduling end time into continuous time grids according to the preset scheduling granularity, set the chip placement process, plug-in process, testing process and assembly process as process grids respectively, set the available equipment or workstations of each process as equipment grids, and generate a time-process-equipment three-dimensional grid space by combining the continuous time grids, process grids and equipment grids; Read the executable time window, the process to which it belongs, the candidate equipment, the processing time, the preceding segment and the following segment corresponding to each order process segment, map each order process segment to the grid position that matches the executable time window, the process to which it belongs and the candidate equipment, and establish the process succession relationship between adjacent grids based on the succession relationship between the preceding segment and the following segment; Read the multi-dimensional scheduling potential energy weights, generate segment local potential energy based on the line change cost, tooling cost, testing cost, waiting cost, delivery cost and exception cost corresponding to each order process segment, write the segment local potential energy into the corresponding grid position, and transfer the segment local potential energy to adjacent grids according to time succession relationship, process succession relationship and equipment occupancy relationship to construct a multi-potential well potential field; In a multi-potential well field, starting from the preceding grid position of each order process segment, grid positions that satisfy the process continuity relationship, equipment occupation relationship, material arrival relationship and test resource acceptance relationship are selected sequentially along the local potential energy decrease direction of the segment, forming an adaptive curved streamline that runs through the chip placement process, insertion process, testing process and assembly process. Perform streamline gradient change analysis and curvature change analysis on each order process segment in each adaptive curved streamline, and identify the order process segments that are at the positions of abrupt change in the direction of potential energy decrease, abrupt change in the direction of streamline curvature, the position of confluence of cross-process, or the starting position of streamline bypass as critical curved segments.

6. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The determined stable bridging segment includes: Read the order number, process type, current grid position, previous segment, subsequent segment, candidate equipment, material batch, tooling number, test program number and adaptive bending streamline assignment result corresponding to each key bending segment, and generate the baseline state of the key bending segment; A local mirror layer is constructed based on the baseline state of the key bending segment. The local mirror layer includes a retained in-situ mirror, a forward-moving mirror, a backward-moving mirror, a removed mirror, and a replaced device mirror. The execution time, device usage, tooling usage, test resource usage, preceding continuation state, and subsequent continuation state are recorded for each local mirror. Centered on the critical bending segment, read the preceding segment, the following segment, the segment adjacent to the same equipment, and the segment adjacent to the same test resource to build a neighborhood mirror layer. Within the same scheduling time window, read the order process segment that has a material continuity relationship, process succession relationship, or test resource acceptance relationship with the critical bending segment to build a global constraint mirror layer. The hierarchical meta-stable-bridging mirror body is composed of the local mirror layer, the neighborhood mirror layer, and the global constraint mirror layer. The image disturbance cost data corresponding to line switching, tooling, testing, waiting, queuing, delay and flow detour in each image state are collected respectively. The image disturbance cost data are weighted and aggregated according to the multidimensional scheduling potential energy weight. The comprehensive cost increment formed by each disturbed image relative to the original image is determined as the meta-stability energy barrier of the corresponding image state. The meta-stability energy barrier formed by each key bending segment under forward mirror, backward mirror, removal mirror, equipment replacement mirror, neighbor mirror layer or global constraint mirror layer is compared with the pre-set meta-stability judgment threshold. It is checked whether there are interruptions in process continuity, equipment occupation conflicts, tooling reuse interruptions, test resource conflicts, semi-finished products waiting to be expanded or adaptive bending streamline detours in the corresponding mirror state. The key bending segment whose meta-stability energy barrier reaches the meta-stability judgment threshold and exhibits any of the above states is determined as a meta-stability bridging segment.

7. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The generation of the persistent homology-conformal latching kernel includes: The key bending segments and non-stable bridging segments are collected to generate a candidate locking segment set. Each order process segment in the candidate locking segment set is used as the vertex of the chain complex, and the order number, process type, grid position, equipment number, material batch number, tooling number, test program number and adaptive bending streamline assignment result are written for each chain complex vertex. Based on the process sequence relationship, material batch continuity relationship, equipment reuse relationship, tooling reuse relationship, test procedure acceptance relationship, test fixture acceptance relationship, semi-finished product flow relationship and delivery constraint relationship between the vertices of each chain complex, chain complex edges are established. Chain complex edges that can sequentially connect the chip placement process, insertion process, testing process and assembly process are combined into two-dimensional receiving cavities. Two-dimensional receiving cavities that simultaneously close time continuity, process continuity and equipment occupation are combined into time-process-equipment chain complexes. The filtering order is set according to the local potential energy of the segment, the stability barrier, the image perturbation cost data, the process continuity strength, the test resource acceptance strength and the waiting status of the semi-finished product. The chain complex vertex, chain complex edge and two-dimensional acceptance cell in the time-process-equipment chain complex are added step by step to identify the segment connection bridge connecting two cross-process acceptance loops and the potential congestion voids formed by waiting accumulation, equipment queuing or test resource conflict. The vertices of the chain complex in the time-process-equipment chain complex are taken as manifold nodes, and the edges of the chain complex are taken as manifold connecting edges. The mapping distance between nodes is determined based on the bearing strength, element stability energy barrier, mirror perturbation cost data and semi-finished product waiting status corresponding to each manifold connecting edge. Under the condition of keeping the node adjacency relationship and process bearing direction unchanged, a two-dimensional unfolding mapping is performed on the time-process-equipment chain complex to generate a two-dimensional conformal manifold. Based on the segment association distance, local curvature change, cross-process acceptance loop coverage, segment connection bridge location, and potential congestion cavity coverage in the two-dimensional conformal manifold, order process segments covering cross-process acceptance loops, order process segments connecting two cross-process acceptance loops, and order process segments that expand potential congestion cavities after release are retained. The retained order process segments and their corresponding chain complex edges are combined into a continuous homology-conformal locking core.

8. The intelligent scheduling and management system for electronic product production line orders based on deep learning according to claim 1, characterized in that, The intelligent scheduling scheme for generating the final order includes: Read the disturbance type, disturbance occurrence time, affected processes, affected equipment, affected materials, affected test resources, and affected order process segments corresponding to the order insertion event, material shortage event, equipment failure event, test equipment abnormality event, or process cycle abnormality event, and generate the disturbance impact range; Read the order process segments, relative order of segments, process succession relationship, material continuity relationship, test resource acceptance relationship, equipment occupation relationship and semi-finished product flow relationship in the continuous homology-conformal lockout kernel, and set freeze flags for the order process segments and corresponding acceptance relationships in the continuous homology-conformal lockout kernel; Based on the scope of the disturbance, rearrangeable order process segments are read from the outside of the continuous homology-conformal lockout core. Combined with available equipment time windows, material arrival time windows, test resource idle time windows, and process continuity time windows, a set of rearrangeable segments outside the core and a set of insertable time windows are generated. When the disturbance type is an insertion event, the insertion order is split into patch acceptance segment, plug-in acceptance segment, test acceptance segment and assembly acceptance segment. The process segment of the insertion order is combined with the set of reconfigurable segments outside the core. Under the condition that the relative order of segments within the continuous homology-conformal locking core and the corresponding acceptance relationship are not changed, candidate reconfiguration schemes are generated. Perform feasibility verification and perturbation sorting on the candidate rearrangement schemes, and determine the candidate rearrangement scheme that passes the feasibility verification and whose perturbation sorting results meet the minimum perturbation condition as the final intelligent order scheduling scheme.

Images