A Semiconductor Intelligent Scheduling Method Based on Multi-Objective Optimization

Abstract

This invention discloses a semiconductor intelligent scheduling method based on multi-objective optimization, relating to the field of semiconductor packaging and testing production planning and scheduling technology. The method mainly includes: splitting each production order into multiple sub-work orders; traversing each process record from back to front according to capacity records; training a UPH model and a preference model using historical production data; updating the real-time UPH using the median absolute deviation method and merging it with the historical UPH; classifying orders into out-of-line orders and partially produced orders; generating scheduling units for out-of-line orders and partially produced orders using an order processing function; passing the partially produced order scheduling units into an exclusion set; deleting scheduling units of all sub-work orders already in the MES; and using a scheduler to perform production scheduling to obtain the scheduling details. Implementing this invention can improve the accuracy, robustness, and configurability of semiconductor production scheduling, reduce changeover time, increase the utilization rate of bottleneck equipment, and shorten order delivery cycles.

Description

Technical Field

[0001] This invention relates to the field of semiconductor packaging and testing production planning and scheduling technology, and more specifically, to a semiconductor intelligent scheduling method based on multi-objective optimization. Background Technology

[0002] Semiconductor packaging production lines typically include core processes such as die bonding, wire bonding, molding, electroplating, lead trimming, and testing, as well as auxiliary processes such as baking and inspection. They are characterized by strict process sequences, a large number of machines, complex constraints between equipment and packaging, large production volumes, and tight order deadlines. Existing production scheduling systems mostly rely on manual experience or simple First-In-First-Out (FIFO) rules, which are insufficient to address the following technical challenges: First, the capacity data recorded by Manufacturing Execution Systems (MES) often suffers from missing end times, discontinuous process records, and inconsistent hourly output (Units Per Hour). First, inconsistent unit (Hour, UPH) results in misjudgments of work-in-process status and inaccurate predictions of end times. Second, significant differences in hourly output across different products, machines, and batches mean that simple historical averages or real-time values ​​cannot guarantee both accuracy and robustness. Third, the wire bonding process is a bottleneck, and the number of parallel devices must match the output cycle of the upstream die bonding process; otherwise, it can lead to wasted capacity or exacerbate the bottleneck. Fourth, the physical layout of equipment results in high changeover time costs, and traditional scheduling focuses only on time and efficiency, ignoring the proximity of equipment locations. Fifth, when urgent orders require securing resources, interrupted work-in-process orders must be correctly resumed to avoid data loss or duplicate scheduling. Sixth, different sub-work orders (some in-process, some not yet started) under the same master work order tend to use different equipment due to independent scheduling paths, increasing the frequency of changeovers.

[0003] Therefore, improving the accuracy, robustness, and configurability of semiconductor production scheduling, reducing changeover time, increasing the utilization rate of bottleneck equipment, and shortening order delivery cycles are urgent problems to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide a semiconductor intelligent scheduling method based on multi-objective optimization, which can improve the accuracy, robustness and configurability of semiconductor scheduling, reduce changeover time, increase the utilization rate of bottleneck equipment, and shorten the order delivery cycle.

[0005] This invention provides a semiconductor intelligent scheduling method based on multi-objective optimization, comprising the following steps: S1: Obtain production orders, split each production order into multiple sub-work orders according to the packaging type and fixed sub-batch, and construct the total order mapping and sub-work order plan mapping; S2: Obtain the capacity records within the specified time window from the manufacturing execution system. Based on the capacity records, total order mapping, and sub-work order plan mapping, group them by sub-work orders, sort them by start time, and traverse each process record from back to front to obtain the work-in-process tasks, completed processes, and end time. S3: Train the UPH model and preference model using historical production data to obtain the trained UPH model and preference model; the UPH model is used to construct a median lookup table with process, machine, product, packaging, and batch range as keys, and backtracks from fine to coarse granularity during prediction to obtain the output per unit hour; the preference model is used to calculate the output weighted probability of each machine according to process, product, and packaging to obtain the preference score of each machine. S4: Obtain the real-time unit hour output from the manufacturing execution system, remove outliers using the median absolute deviation method, and take the median to obtain the new real-time unit hour output. Then, perform weighted fusion of the real-time unit hour output with the historical unit hour output to obtain the fused real-time unit hour output, which is used as the input for actual scheduling. S5: Based on the set of sub-work orders in the manufacturing execution system and the parent work order number of the partial production, the orders are divided into out-of-line orders and partial production orders. The order processing function is used to generate out-of-line order scheduling units and partial production order scheduling units. S6: Pass some production order scheduling units into the exclusion set to skip sub-work orders already in MES, delete all scheduling units of sub-work orders already in the manufacturing execution system, and obtain the filtered scheduling units. S7: Based on the filtered scheduling units, the fusion of real-time hourly output, the trained UPH model and preference model, the work-in-process tasks, the completed processes and end times, the scheduler is used to schedule production and obtain the scheduling details.

[0006] Furthermore, the semiconductor intelligent scheduling method based on multi-objective optimization also includes: confirming the existence of urgent orders; finding currently occupied equipment with an end time later than the current time based on the equipment required for key processes; sorting the equipment by its end time from latest to earliest; releasing the orders of the equipment at the top of the sort; adding the released order numbers to the preempted set; regenerating scheduling units for the unstarted sub-work orders in the preempted set; setting the end time of the in-process tasks of the preempted equipment to the current time; obtaining the scheduling result based on the regenerated scheduling unit, the scheduling result including scheduling details, equipment view, order delivery prediction, and generating a wire bonding timeout alarm file.

[0007] Furthermore, the key processes include die bonding and wire bonding; the equipment view includes the current task, the next task, and the future task; the alarm file is obtained by comparing the estimated end time of the initial lock with the current time.

[0008] Furthermore, the process of traversing each process record from back to front is as follows: if the end time of the current process exists and is later than the current time, the current process is treated as a work-in-process task; if the end time of the current process does not exist, if the current process has a subsequent process and the subsequent process has an end time, the current process is treated as a completed process with missing data; if the current process does not have a subsequent process, the current process is treated as a work-in-process task; and the end time of the work-in-process task is predicted based on the process record.

[0009] Furthermore, the method for predicting the end time of work-in-process tasks is as follows: If the current process of the work-in-process task is determined to be electroplating, a 3-7 day cycle is randomly generated based on the parent work order to obtain the end time of the work-in-process task; if the current process of the work-in-process task is determined not to be electroplating, the remaining working hours are calculated based on the total order volume mapping and real-time hourly output to obtain the end time of the work-in-process task; if it is confirmed that the data source for calculating the remaining working hours is insufficient, the end time of the work-in-process task is calculated using historical typical durations; if it is confirmed that the data for historical typical durations is insufficient, the end time of the work-in-process task is calculated using fixed machine occupancy time.

[0010] Furthermore, the real-time unit hour output refers to the real-time unit hour output of the manufacturing execution system for the most recent 3 hours.

[0011] Furthermore, the method for scheduling production using a scheduler includes: for a current work-in-process task, obtaining the next core process based on the current process of the work-in-process task, and generating a first continuation scheduling unit using the predicted end time of the current task as the readiness time; for a sub-work order that has completed the core process but still has subsequent processes, finding the last core process with an end time from the historical data of the manufacturing execution system, and generating a second continuation scheduling unit starting from the next process of the core process; merging the first and second continuation scheduling units and placing them at the beginning of the scheduling task list to obtain a new scheduling task list; using the scheduler, employing a greedy strategy to process each scheduling unit sequentially, traversing according to the process order to obtain the scheduling details.

[0012] Furthermore, the process of traversing according to the process sequence includes: for each process, firstly, candidate machines are screened based on the equipment master data and packaging constraints. If it is a wire bonding process, the parallel upper limit and the target total output per unit hour are dynamically calculated. Based on the parallel upper limit and the target total output per unit hour, the start time, processing time, and end time of each candidate machine are calculated, and the comprehensive cost is evaluated using a cost function. The cost function includes: the difference between the completion time and the current time as the main term, the logarithm of the output per unit hour as the efficiency term, the preference score of the trained preference model as the historical habit term, the difference between the sequence number of the equipment and the anchor point and the cross-prefix penalty as the distance term, the wire bonding cycle time matching gap as the penalty term, and the historical equipment of the parent work order as the reward term. The machine with the minimum comprehensive cost is selected, the available time of the machine is updated, the equipment anchor point is recorded, and the set of equipment used in the parent work order is updated to obtain the scheduling details.

[0013] Furthermore, the method for dynamically calculating the parallel upper limit and the target total unit hour output includes: collecting the fused real-time unit hour output of all candidate devices, sorting them by sequence number difference based on the anchor device, selecting a local pool, calculating the discrete coefficient of high-efficiency devices from the local pool, and determining the target total unit hour output based on the discrete coefficient.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described semiconductor intelligent scheduling method based on multi-objective optimization.

[0015] Implementing the semiconductor intelligent scheduling method based on multi-objective optimization provided by this invention has the following beneficial effects: This invention addresses the shortcomings of existing MES technologies, such as poor data quality and lack of timing verification and completion logic. It employs a multi-level prediction method for in-process task identification and end-time prediction based on process timing analysis. The method groups tasks by sub-work orders (Lots), sorts them by process start time, and iterates through subsequent processes from back to front to determine in-process status. For in-process tasks lacking end times, it uses a multi-level prediction priority based on "total order quantity / real-time UPH calculated remaining working hours → historical typical duration → fixed machine occupation time." The electroplating process is separately generated with a fixed, random 3-7 day cycle based on the parent work order. All predictions undergo rationality verification to avoid overflow. This solves the problems of incomplete MES data (missing end times, discontinuous process records) leading to misjudgments of in-process task status and inaccurate end-time predictions. It accurately identifies the true status of in-process tasks, improves end-time prediction accuracy, provides accurate in-process task data for subsequent scheduling, and avoids status misjudgments.

[0016] This invention addresses the shortcomings of existing UPH (Universal Power Requirement) prediction technologies, which fail to consider multi-dimensional differences and lack robust rollback and outlier handling. It utilizes a hierarchical rollback and batch binning UPH prediction method to construct a UPH median table across five levels: "process-machine-product," "process-machine-packaging," "process-machine," "process-packaging," and "process." Batch intervals are divided according to the number of sub-lots to achieve fine-grained mapping. The Median Absolute Deviation (MAD) method is used to remove outliers from the most recent 3 hours of real-time UPH, and the median is then fused with historical UPH values ​​using weighted methods as scheduling input. This solves the problem that significant differences in UPH exist across different products, machines, and batches, and that simple historical averages or real-time values ​​cannot simultaneously guarantee prediction accuracy and robustness. It achieves precise, fine-grained UPH prediction, adapts to different data availability scenarios through multi-level rollback, and improves real-time data robustness through outlier removal, providing accurate input for scheduling time and efficiency evaluation.

[0017] This invention addresses the shortcomings of existing technologies that lack dynamic calculation and cycle time matching logic for wire bonding parallelism. It utilizes a dynamic parallelism and cycle time matching method for the wire bonding process. A local equipment pool is constructed using the wire bonding equipment initially selected in the master work order as the anchor point. The local pool is sorted by equipment number difference, and the dispersion coefficient of high-efficiency equipment within the pool is calculated. The target total UPH is adaptively calculated, and the fused UPH of equipment within the local pool is accumulated from high to low until the target total UPH is reached. The accumulated number of equipment is the upper limit of parallel equipment (cap), which is also constrained by the global parallelism upper limit. This solves the problem of mismatch between the number of parallel equipment in the wire bonding process and the output cycle time of the upstream die bonding process, leading to wasted capacity or exacerbated bottlenecks. It dynamically adapts to the actual parallelism capability of the wire bonding process, matches the output cycle time of the upstream die bonding process, improves the utilization rate of bottleneck equipment, and avoids wasted capacity or bottleneck overload.

[0018] This invention addresses the shortcomings of existing scheduling technologies that do not consider the physical proximity of equipment. It utilizes a scheduling cost function that combines equipment anchors and distance penalties. It records the first equipment selected for the same product as the anchor and applies a penalty based on the difference between the prefix and sequence number of the equipment number. This encourages tasks in the same group to select equipment with similar numbers, solving the problems of high line changeover time costs caused by physical equipment layout and the neglect of equipment location proximity in traditional scheduling. It reduces the physical distance of equipment allocation for tasks in the same group, reduces line running time during equipment switching, and lowers line changeover time costs.

[0019] This invention addresses the shortcomings of existing technologies that lack logic for expedited preemption and interrupted task continuation. It utilizes a preemption and task continuation method for expedited orders, focusing only on equipment for critical processes like die bonding and wire bonding. The first N devices are released in ascending order of their occupied end time, and the released orders are added to the preempted set. A second order retrieval process regenerates the unstarted sub-lots of preempted orders into Jobs with a priority of P1 (P0 for expedited orders). The end time of the in-process tasks on the preempted equipment is forcibly set to the current time for immediate release. This solves the problem of data loss or duplicate scheduling for interrupted in-process orders when expedited orders require resource preemption, ensuring rapid response for expedited orders while ensuring correct continuation of interrupted in-process orders, avoiding scheduling data corruption or duplicate scheduling.

[0020] This invention addresses the deficiency in existing technologies that lack equipment consistency constraints at the master work order level by constructing a soft constraint for equipment consistency at the master work order level. It extracts the set of equipment used in each process of each master work order from historical MES data and applies a negative reward to these used equipment in the cost function, prioritizing their selection without forcing them to be constrained. This solves the problem that different sub-work orders under the same master work order may use different equipment due to independent scheduling, increasing the frequency of line changes. It improves the consistency of equipment allocation among sub-work orders under the same master work order, reduces equipment switching frequency, and lowers line changeover costs.

[0021] This invention addresses the shortcomings of existing technologies that lack a two-level filtering mechanism (order-work order) and fallback check. It proposes a batch scheduling method for offline orders based on this two-level filtering. First, it extracts the parent work order number for partial production from the set of sub-work orders (Lots) already present in the MES (Manufacturing Execution System). Orders are then categorized into "completely offline" and "partially in production." For the "partially in production" category, an exclusion set is passed to skip sub-work orders (Lots) already in the MES. Finally, at the scheduler entry point, all Jobs corresponding to sub-work orders (Lots) already in the MES are forcibly deleted as a fallback. This solves the problem of unfiltered sub-work orders already in production or partially in production during offline order scheduling, which easily leads to duplicate scheduling. It ensures that sub-work orders already in production or partially in production are not repeatedly included in the schedule, avoiding scheduling conflicts and resource waste.

[0022] This invention addresses the shortcomings of existing technologies where wire bonding timeout alarms lack a locking mechanism and are prone to jitter. It proposes a time-locked monitoring method for wire bonding tasks that fails to complete within timeout. Once the estimated end time of the wire bonding task is locked for the first time, it remains unchanged. The timeout time is compared with the current time to determine whether an alarm has been triggered. This solves the problem of frequent fluctuations in wire bonding process timeout alarms and poor monitoring reliability. It also avoids repeated jitter in alarm time and improves the stability and reliability of timeout monitoring.

[0023] In summary, this invention, as a multi-objective intelligent scheduling method that integrates historical data learning, real-time capacity perception, equipment dynamic constraints, rush order grabbing, and work-in-process time-series continuation, improves the accuracy, robustness, and configurability of semiconductor scheduling, reduces changeover time, increases bottleneck equipment utilization, and shortens order delivery cycles. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the semiconductor intelligent scheduling method based on multi-objective optimization provided by the present invention; Figure 2 This is a schematic diagram of the overall framework of the semiconductor intelligent scheduling method based on multi-objective optimization provided by the present invention; Figure 3 This is a diagram of the production scheduling results provided by the present invention. Detailed Implementation

[0025] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] Figure 1 A schematic diagram of the semiconductor intelligent scheduling method based on multi-objective optimization according to this embodiment is shown. In this embodiment, the semiconductor intelligent scheduling method based on multi-objective optimization includes the following steps: S1: Obtain production orders, split each production order into multiple sub-work orders according to the packaging type and fixed sub-batch, and construct the total order mapping and sub-work order plan mapping.

[0027] S2: Obtain the capacity records within the specified time window from the manufacturing execution system. Based on the capacity records, total order mapping, and sub-work order plan mapping, group them by sub-work orders, sort them by start time, and traverse each process record from back to front to obtain the work-in-process tasks, completed processes, and end times.

[0028] In one exemplary embodiment, the process of traversing the process records from back to front is as follows: If the end time of the current process exists and is later than the current time, then the current process is treated as a work-in-process task. If it is confirmed that the end time of the current process does not exist, or if it is confirmed that there is a subsequent process for the current process and the subsequent process has an end time, then the current process is treated as a completed process with missing data. If it is confirmed that there is no subsequent process for the current process, then the current process is treated as a work-in-process task. Based on the process records, predict the end time of the work-in-process tasks.

[0029] In one exemplary embodiment, the method for predicting the end time of in-process tasks is as follows: The current process of the work-in-process task is determined to be electroplating. Based on the fixed and random cycle of 3-7 days generated from the master work order, the end time of the work-in-process task is obtained. If the current process of the work-in-process task is determined not to be the electroplating process, the remaining working hours are calculated based on the total order volume mapping and the real-time unit hourly output to obtain the end time of the work-in-process task. It was confirmed that the data source for calculating the remaining working hours was insufficient, so the end time of the in-process task was calculated using historical typical durations; Since the data on typical historical durations is insufficient, the end time of in-process tasks is calculated using fixed machine occupancy durations.

[0030] S3: Train the UPH model and preference model using historical production data to obtain the trained UPH model and preference model; the UPH model is used to construct a median lookup table with process, machine, product, packaging, and batch range as keys, and backtracks from fine to coarse granularity during prediction to obtain the output per unit hour; the preference model is used to calculate the output weighted probability of each machine according to process, product, and packaging to obtain the preference score of each machine.

[0031] S4: Obtain the real-time unit hour output from the manufacturing execution system, remove outliers using the median absolute deviation method, and take the median to obtain the new real-time unit hour output. Then, perform weighted fusion of the real-time unit hour output with the historical unit hour output to obtain the fused real-time unit hour output, which is used as the input for actual scheduling.

[0032] In one exemplary embodiment, the real-time unit hour output is the real-time unit hour output of the manufacturing execution system for the most recent 3 hours.

[0033] S5: Based on the set of sub-work orders in the manufacturing execution system and the parent work order number of the partial production, the orders are divided into out-of-line orders and partial production orders. The order processing function is used to generate out-of-line order scheduling units and partial production order scheduling units.

[0034] S6: Pass some production order scheduling units into the exclusion set to skip sub-work orders already in MES, delete scheduling units of all sub-work orders already in the manufacturing execution system, and obtain the filtered scheduling units.

[0035] S7: Based on the filtered scheduling units, the fusion of real-time hourly output, the trained UPH model and preference model, the work-in-process tasks, the completed processes and end times, the scheduler is used to schedule production and obtain the scheduling details.

[0036] In one exemplary embodiment, the method of scheduling production using a scheduler includes: For the current in-process task, the next core process is obtained according to the current process of the in-process task, and the first continuation scheduling unit is generated with the predicted end time of the current task as the preparation and readiness time. For sub-work orders that have completed the core process but still have subsequent processes, find the last core process with an end time from the historical data of the manufacturing execution system, and generate a second continuation scheduling unit starting from the next core process. The first and second continuation scheduling units are merged and placed at the beginning of the scheduling task list to obtain a new scheduling task list. The scheduler uses a greedy strategy to process each scheduling unit sequentially, traversing according to the process order to obtain the scheduling details.

[0037] In one exemplary embodiment, the process of traversing according to the process sequence includes: For each process, candidate machines are first screened based on the equipment master data and packaging constraints; if it is a wire bonding process, the parallel upper limit and the target total output per unit hour are dynamically calculated. Based on the parallel upper limit and the target total output per unit hour, calculate the start time, processing time, and end time of each candidate machine, and evaluate the comprehensive cost using a cost function. The cost function includes: the difference between the completion time and the current time as the main term, the logarithm of the output per unit hour as the efficiency term, the preference score of the trained preference model as the historical habit term, the difference between the sequence number of the equipment and the anchor point and the cross-prefix penalty as the distance term, the wire bonding cycle matching gap as the penalty term, and the historical equipment of the parent work order as the reward term. Select the machine with the lowest overall cost, update the available time of the machine, record the equipment anchor point, update the set of equipment used in the master work order, and obtain the scheduling details.

[0038] In one exemplary embodiment, the method for dynamically calculating the parallel upper limit and the target total unit hour output includes: collecting the fused real-time unit hour output of all candidate devices, sorting them by sequence number difference based on the anchor device, selecting a local pool, calculating the discrete coefficient of high-efficiency devices from the local pool, and determining the target total unit hour output based on the discrete coefficient.

[0039] In one exemplary embodiment, the semiconductor intelligent scheduling method based on multi-objective optimization further includes: Once an urgent order is confirmed, based on the equipment required for the critical process, find the equipment that is currently occupied and whose end time is later than the current time, sort them from latest to earliest by occupation end time, release the orders of the equipment at the top of the sort, and add the order numbers of the released orders to the preempted set; Regenerate scheduling units for unstarted sub-work orders in the preempted set, and set the end time of the in-process tasks of the preempted equipment to the current time. Based on the regenerated scheduling unit, the scheduling results are obtained, including scheduling details, equipment view, order delivery forecast, and a wire bonding timeout alarm file is generated.

[0040] In one exemplary embodiment, the key processes include die bonding and wire bonding; The device view includes the current task, the next task, and future tasks; The alarm file is obtained by comparing the expected end time of the initial lock with the current time.

[0041] In some embodiments, the above-described semiconductor intelligent scheduling method based on multi-objective optimization can also be implemented in the following ways.

[0042] In this embodiment, the following are achieved: First, a multi-level prediction of in-process task identification and completion time based on process timing analysis: By grouping by sub-work orders, sorting by start time, and traversing the completion status of subsequent processes from back to front, it is determined whether the current process is truly in-process. For tasks with missing completion times, prediction is performed sequentially using total order volume, real-time UPH, historical typical duration, and fixed machine occupation time. Random periodic processing of electroplating outsourcing is also introduced. Second, a hierarchical rollback and batch binning UPH prediction: A median table of UPH is constructed for five levels: process-machine-product, process-machine-packaging, process-machine, process-packaging, and process. Batch intervals are divided according to the number of sub-work orders to achieve fine-grained prediction and robust rollback. Third, dynamic parallel capability and cycle time matching for wire bonding processes: Using the wire bonding equipment initially selected in the parent work order as the anchor point, a local equipment pool is constructed. The target total UPH is adaptively calculated using the discrete coefficients of high-efficiency equipment within the pool, and dynamically... The system includes: 1) Determining the upper limit of parallel devices; 2) Implementing a scheduling cost function for device anchor points and distance penalties, recording the first selected devices for the same product, and applying penalties based on the difference between the prefix and sequence number of the device number to encourage tasks in the same group to select similar devices; 3) Implementing soft constraints on device consistency at the parent work order level, extracting the set of devices used in each process of each parent work order from historical MES data, and giving negative rewards to these devices in the cost function to prioritize their selection without forcing them; 4) Implementing batch scheduling of offline orders based on two-level filtering of orders and work orders: using the set of child work orders that have appeared in MES to achieve precise filtering and mandatory fallback; 5) Implementing device preemption and task continuation for urgent orders: only performing preemption judgment on key processes, dynamically releasing occupied devices, and ensuring the correct continuation of interrupted orders through secondary order pulling and priority adjustment; 6) Implementing a time-sequence locking-based monitoring method for wire bonding timeouts, locking the estimated end time initially to avoid alarm jitter. The overall flowchart is as follows: Figure 2 As shown.

[0043] In this embodiment, the semiconductor intelligent scheduling method based on multi-objective optimization includes the following steps: First, the system retrieves production orders from the order API, splits each order into several sub-loans based on packaging type and fixed sub-batch size, and constructs an order total mapping and sub-loan plan mapping. Simultaneously, it retrieves capacity records within a specified time window from the MES interface, groups them by sub-loans, sorts them by start time, and iterates through each process record from back to front: if a process's end time exists and is later than the current time, it is considered in-process; if the end time is missing, it checks if there are subsequent processes with end times. If so, the current process is considered completed (data missing); otherwise, it is considered in-process. For tasks determined to be in-process, the system predicts their end time: if it's an electroplating process, it generates a fixed cycle of 3 to 7 days based on the parent work order; otherwise, it prioritizes calculating the remaining working hours using the order total and real-time UPH. If the data is insufficient, it reverts to historical typical durations; if still insufficient, it uses fixed machine occupancy time. All predictions undergo a reasonableness check to avoid overflow.

[0044] Then, the system uses historical production data to train the UPH model and the preference model. The UPH model constructs a median lookup table using process, machine, product, packaging, and batch range as keys, and backtracks from fine to coarse granularity during prediction. The preference model calculates the output weighted probability of each machine at the process, product, and packaging levels. Simultaneously, the system calculates the real-time UPH for the most recent 3 hours from MES data, uses the Median Absolute Deviation (MAD) method to remove outliers, takes the median, and merges it with historical UPH data according to weights as input for actual scheduling.

[0045] Next, the system processes the orders that are not yet online in batches: it extracts the parent work order number of the partially produced order based on the set of sub-work orders that have appeared in MES, divides the order lines into two categories: "not yet online" and "partially produced", and calls the order processing function to generate Jobs for each category. The latter needs to pass in the exclusion set to skip the sub-work orders that are already in MES. Finally, at the scheduler entry point, all Jobs (i.e., scheduling units) of the sub-work orders that are already in MES are forcibly deleted again as a fallback filter.

[0046] The system then generates ongoing tasks: For a current ongoing task, the next core process is obtained based on its current process, and a follow-up job is generated using the predicted end time of the current task as the ReadyTime; for a sub-work order that has completed a core process but still has subsequent processes (e.g., wire bonding is completed in MES but there is no encapsulation record), the last core process with an end time is found from the MES historical data, and a follow-up job starting from the next process is generated. These two types of follow-up jobs are merged and placed at the top of the scheduled task list.

[0047] The scheduler uses a greedy strategy to process each job sequentially, traversing the process order. For each process, candidate machines are first selected based on the equipment master data and packaging constraints. If it is a wire bonding process, the parallel upper limit and the target total UPH are dynamically calculated: the fused UPH of all candidate devices is collected, and a local pool is selected based on the difference in sequence number of the anchor device (if the total number of devices is not large, all devices are used). The Top N efficient devices in the local pool are selected to calculate the coefficient of variation α, and then the target total UPH is determined. The UPH is then accumulated from high to low until the target is reached. The number of accumulated devices is cap, which is subject to the global upper limit constraint.

[0048] The system then calculates the start time, processing time, and end time for each candidate machine, and evaluates the overall cost using a cost function. The cost function includes: the difference between the completion time and the current time as the main term; the logarithm of the UPH (Uptime Per Hour) as the efficiency term; the preference model score as the historical habit term; the difference in the sequence number between the machine and the anchor point and the cross-prefix penalty as the distance term; the wire bonding cycle time matching gap as the penalty term; and the historical equipment of the parent work order as the reward term. Each weight is configurable. The system selects the machine with the lowest cost, updates its available time, records the machine anchor point, updates the set of used machines in the parent work order, and outputs the scheduling details.

[0049] When an urgent order exists, the system identifies the equipment required for its critical processes (die bonding, wire bonding), locates currently occupied equipment with an end time later than the current time, sorts them by end time from latest to earliest, and releases the first N units (N is configurable). The order numbers of the released units are added to the preempted set. Then, the system re-enters the order, regenerates the unstarted sub-work orders of the preempted orders into a Job, and assigns them priority P1 (P0 for urgent orders). Simultaneously, the end time of the in-process tasks on the preempted equipment is forcibly set to the current time, causing it to be released immediately.

[0050] Finally, the system outputs a scheduling result Excel file, including scheduling details, equipment views (current / next / future tasks), order delivery forecasts, and generates a wire bonding timeout alarm file. The alarm is based on a comparison between the initially locked estimated end time and the current time to avoid repeated fluctuations. The scheduling result interface diagram is shown below. Figure 3 As shown.

[0051] It should be noted that the key points of this invention are: using time series analysis to solve the work-in-process misjudgment caused by incomplete MES data; adopting a multi-level rollback and batch binning UPH prediction model to improve accuracy; achieving equipment consistency through anchor point distance penalty and soft constraints of historical equipment in the parent work order; dynamically calculating the parallel capability of welding lines to achieve cycle time matching; two-level filtering and forced fallback to ensure that work-in-process is not re-scheduled; and expedited order preemption and continuation scheduling mechanisms to ensure response speed.

[0052] Compared with existing technologies, this invention can improve scheduling accuracy, reduce changeover time, increase bottleneck equipment utilization, shorten order delivery cycle, and has good robustness and configurability, making it suitable for semiconductor packaging manufacturing scenarios.

[0053] It should be noted that MAD stands for Median Absolute Deviation, a robust statistic in statistics that measures the dispersion of data. The calculation steps are as follows: ① For the input dataset to be processed (real-time UPH values ​​of each machine / process in the last 3 hours), first calculate the median of all data, denoted as M; ② Calculate the absolute difference between each data point and the median M in turn to obtain the set of absolute differences; ③ Calculate the median again for the set of absolute differences, and the result is the MAD value.

[0054] The outlier determination rule for MAD (Median Displacement) is as follows: The outlier threshold is typically "median ± k times MAD", where k is an adjustment coefficient (usually 2-3; the smaller k is, the higher the sensitivity to outliers). If data points... <M-k MAD, or data point > M+k If a value is MAD, it is considered an outlier and is removed. The median of the remaining non-outliers is then taken, which is the robust UPH median after removing outliers.

[0055] The role of the MAD method in this invention is as follows: ① The input is the raw statistical data of real-time UPH for the most recent 3 hours. This data may contain extreme outliers due to sudden equipment failures, data entry errors, temporary line changes, etc. ② The robust median obtained after removing the above outliers using the MAD method is then fused with the historical hierarchical UPH according to weights to obtain the actual UPH used for scheduling. ③ Compared with the traditional standard deviation method, MAD does not require the assumption that the data follows a normal distribution, nor is it affected by the skewed statistical results caused by extreme outliers. It can more accurately reflect the normal production capacity level of the equipment, improve the robustness of UPH prediction, and provide reliable input for subsequent scheduling time and efficiency assessment.

[0056] Production capacity data in production scenarios often exhibits occasional anomalies (such as equipment failure causing UPH to be 0 in a certain hour, or statistical errors causing UPH to far exceed the normal value). If the standard deviation is used to calculate the degree of dispersion, extreme values ​​will directly increase the standard deviation, causing normal data to be misjudged as anomalies. The robustness of MAD perfectly matches the data robustness requirements of production scheduling scenarios, which is in line with the design goal of this invention to balance the accuracy and robustness of UPH prediction.

[0057] It should be noted that the core objective of the UPH model is to output the predicted hourly output value under different feature combinations. It features multi-level backoff, batch binning, real-time UPH fusion, and high robustness requirements, and can adopt the following structure: 1. Shallow Multilayer Perceptron (MLP) Based on Feature Embedding Discrete classification features such as process, machine, product, packaging, and batch range are mapped to low-dimensional dense embedding vectors. All embedding vectors, real-time UPH after MAD processing, historical UPH statistics, and other continuous features are concatenated and input into a 2-3 layer fully connected network to finally output the predicted UPH value.

[0058] ① Matching multi-level rollback requirements: A hierarchical sharing mechanism for feature embeddings can be set up, with all levels sharing the basic embeddings of processes, machines, products, and packages. When there are insufficient samples of fine-grained combinations (such as "process-machine-product"), the model automatically degenerates to using only the embeddings of coarse-grained features (such as "process-machine") for prediction, thus achieving soft multi-level rollback and satisfying hard rollback logic; ② Matching batch binning design: The batch intervals divided according to the number of sub-work orders are used as category features input into the embedding, and the model can automatically learn the UPH differences under different batches; ③ Matching real-time UPH fusion requirements: Real-time UPH can be used as an additional input feature, or a lightweight gating unit can be added. The fusion weight of historical prediction values ​​and real-time UPH can be dynamically adjusted according to task characteristics (such as whether the line has been changed or the recent fault status of the equipment), which is more flexible than fixed weight fusion.

[0059] 2. Gradient Boosting Tree Model (LightGBM / XGBoost) It takes discrete / continuous features such as process, machine, product, packaging, and batch range as input, and outputs UPH prediction values ​​through ensemble learning of multiple decision trees.

[0060] ① Natural matching multi-level backoff logic: When splitting features, the tree model will automatically select the feature granularity with the highest information gain. When there are insufficient samples or low information gain of fine-grained combinations (such as "process-machine-product"), it will automatically prioritize the use of coarse-grained features (such as "process-machine" and "process") for splitting and prediction. There is no need to manually set backoff rules, which satisfies the design logic of the hierarchical UPH median table. ② It natively supports discrete feature input without the need for additional embedding mapping. You can directly input the interval features after batch binning, the device number and other category features, and automatically learn the UPH differences of different batches and different device combinations. ③ It is naturally robust to outliers. With MAD outlier preprocessing, the prediction accuracy can be further improved without the need for additional outlier robustness design.

[0061] 3. Fusion prediction model of static features + temporal features (Embedding + TCN / LSTM) Static features such as process, machine, product, packaging, and batch range are input into the Embedding layer to extract low-dimensional representations. The real-time UPH time series sequence of the most recent N hours is input into the temporal convolutional network (TCN) or lightweight LSTM to extract dynamic temporal features. The two types of features are concatenated and input into the fully connected layer to output the UPH prediction value.

[0062] It meets the real-time UPH fusion requirements and can capture the temporal change patterns of UPH (such as capacity ramp-up after equipment replacement and recovery trends after failure), further improving the accuracy of UPH prediction. It is suitable for production line scenarios with high requirements for prediction accuracy and sufficient historical time data.

[0063] It should be noted that the core objective of the preference model is to output the preference scores of each machine under the combination of (process, product, packaging, and parent work order), which are used for the historical habitual term of the cost function and the soft constraint of the consistency of the parent work order equipment. It has the characteristics of hierarchical statistics, output weighting, and support for the preferences of the parent work order equipment, and can adopt the following structure: 1. Feature-interaction shallow MLP Discrete features such as processes, products, packaging, machines, and master work orders are mapped to embedding vectors. The cross weights of different features are learned through inner product or shallow fully connected layers, and the preference score of each machine under the current task features is output. During training, the historical output of the machine under the corresponding feature combination is used as the sample weight to achieve the learning objective of "output weighting".

[0064] ① Matching hierarchical statistical needs: The model automatically learns the cross weights of features at different levels (such as process-machine, product-machine, process-product-machine), and the output feature importance can correspond to the statistical logic at different levels, matching the hierarchical preference design; ② Match the consistency requirements of the parent work order equipment: The similarity calculation item between the embedding of the equipment used in the parent work order in the past and the embedding of the current candidate machine can be added as an extra bonus to the preference score, so that the machine used in the past of the parent work order scores higher, matching the design of "giving negative rewards (i.e. higher preference scores) to the equipment used in the parent work order". ③ Output weighting is achieved through sample weights. The model will give higher scores to machine-feature combinations with higher output, matching the logic of "output weighted probability".

[0065] 2. Hierarchical matrix factorization model The combination of process + product + packaging + mother work order is used as the "user-side" feature and mapped to a low-dimensional user vector; the machine is used as the "item-side" feature and mapped to a low-dimensional item vector; the inner product of the user vector and the item vector is the preference score of the machine under the corresponding combination; during training, the logarithm or normalized value of historical output is used as the weight of the interaction matrix to achieve output weighting.

[0066] ① Naturally adaptable to hierarchical statistical logic: By setting shared user embeddings, preference prediction for different hierarchical features (such as only process + machine, process + product + machine) can be achieved. When fine-grained combination samples are insufficient, coarse-grained preference fallback can be achieved through shared embeddings to match hierarchical preference design. ② It has a strong ability to generalize to sparse feature combinations, making it suitable for scenarios with a large number of rare process-product-machine combinations in semiconductor production scheduling. Even if there are very few historical samples for some combinations, it can still give reasonable preference scores through the generalization of Embedding. ③ The equipment preference of the parent work order can be realized by adding the feature vector of the parent work order separately, using the machines used in the history of the parent work order as positive samples, improving the preference score of the corresponding machine, and matching the equipment consistency requirements of the parent work order.

[0067] 3. Lightweight Gradient Boosting Tree Model Using process, product, packaging, machine, parent work order, historical output, and historical usage tags of parent work order as input features, a preference score is output through a tree model.

[0068] It can automatically learn the preference weights of different feature combinations and directly output the importance of features, which can be mapped to the logic of "statistics by process, product, packaging and other levels". The historical usage mark of the parent work order can be directly input as a feature, and the model will automatically give higher scores to the machines used by the parent work order, without the need for additional design constraints.

[0069] It should be noted that semiconductor production scheduling scenarios have high requirements for model interpretability and small sample adaptability. The following models are recommended: UPH model adopts LightGBM gradient boosting tree, which matches the multi-level backoff requirements and has extremely strong interpretability; the preference model adopts feature interaction shallow MLP, which can flexibly add the consistency constraint of the mother work order equipment, with low overall implementation cost, strong interpretability, and close to the technical design logic.

[0070] This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described semiconductor intelligent scheduling method based on multi-objective optimization.

[0071] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A semiconductor intelligent scheduling method based on multi-objective optimization, characterized in that, Includes the following steps: S1: Obtain production orders, split each production order into multiple sub-work orders according to the packaging type and fixed sub-batch, and construct the total order mapping and sub-work order plan mapping; S2: Obtain the capacity records within the specified time window from the manufacturing execution system. Based on the capacity records, total order mapping, and sub-work order plan mapping, group them by sub-work orders, sort them by start time, and traverse each process record from back to front to obtain the work-in-process tasks, completed processes, and end time. S3: Train the UPH model and preference model using historical production data to obtain the trained UPH model and preference model; the UPH model is used to construct a median lookup table with process, machine, product, packaging, and batch range as keys, and backtracks from fine to coarse granularity during prediction to obtain the output per unit hour; the preference model is used to calculate the output weighted probability of each machine according to process, product, and packaging to obtain the preference score of each machine. S4: Obtain the real-time unit hour output from the manufacturing execution system, remove outliers using the median absolute deviation method, and take the median to obtain the new real-time unit hour output. Then, perform weighted fusion of the real-time unit hour output with the historical unit hour output to obtain the fused real-time unit hour output, which is used as the input for actual scheduling. S5: Based on the set of sub-work orders in the manufacturing execution system and the parent work order number of the partial production, the orders are divided into out-of-line orders and partial production orders. The order processing function is used to generate out-of-line order scheduling units and partial production order scheduling units. S6: Pass some production order scheduling units into the exclusion set to skip sub-work orders already in MES, delete all scheduling units of sub-work orders already in the manufacturing execution system, and obtain the filtered scheduling units. S7: Based on the filtered scheduling units, the fusion of real-time hourly output, the trained UPH model and preference model, the work-in-process tasks, the completed processes and end times, the scheduler is used to schedule production and obtain the scheduling details.

2. The method according to claim 1, characterized in that, Semiconductor intelligent scheduling methods based on multi-objective optimization also include: Once an urgent order is confirmed, based on the equipment required for the critical process, find the equipment that is currently occupied and whose end time is later than the current time, sort them from latest to earliest by occupation end time, release the orders of the equipment at the top of the sort, and add the order numbers of the released orders to the preempted set; Regenerate scheduling units for unstarted sub-work orders in the preempted set, and set the end time of the in-process tasks of the preempted equipment to the current time. Based on the regenerated scheduling unit, the scheduling results are obtained, including scheduling details, equipment view, order delivery forecast, and a wire bonding timeout alarm file is generated.

3. The method according to claim 2, characterized in that, The key processes include die bonding and wire bonding; the equipment view includes the current task, the next task, and the future task; the alarm file is obtained by comparing the estimated end time of the initial lock with the current time.

4. The method according to claim 1, characterized in that, The process of traversing the records of each process step from back to front is as follows: If the end time of the current process exists and is later than the current time, then the current process is treated as a work-in-process task. If it is confirmed that the end time of the current process does not exist, or if it is confirmed that there is a subsequent process for the current process and the subsequent process has an end time, then the current process is treated as a completed process with missing data. If it is confirmed that there is no subsequent process for the current process, then the current process is treated as a work-in-process task. Based on the process records, predict the end time of the work-in-process tasks.

5. The method according to claim 4, characterized in that, The method for predicting the end time of in-process tasks is as follows: The current process of the work-in-process task is determined to be electroplating. Based on the fixed and random cycle of 3-7 days generated from the master work order, the end time of the work-in-process task is obtained. If the current process of the work-in-process task is determined not to be the electroplating process, the remaining working hours are calculated based on the total order volume mapping and the real-time unit hourly output to obtain the end time of the work-in-process task. It was confirmed that the data source for calculating the remaining working hours was insufficient, so the end time of the in-process task was calculated using historical typical durations; Since the data on typical historical durations is insufficient, the end time of in-process tasks is calculated using fixed machine occupancy durations.

6. The method according to claim 1, characterized in that, The real-time hourly output refers to the real-time hourly output of the manufacturing execution system for the most recent 3 hours.

7. The method according to claim 1, characterized in that, The method of scheduling production using a scheduler includes: For the current in-process task, the next core process is obtained according to the current process of the in-process task, and the first continuation scheduling unit is generated with the predicted end time of the current task as the preparation and readiness time. For sub-work orders that have completed the core process but still have subsequent processes, find the last core process with an end time from the historical data of the manufacturing execution system, and generate a second continuation scheduling unit starting from the next core process. The first and second continuation scheduling units are merged and placed at the beginning of the scheduling task list to obtain a new scheduling task list. The scheduler uses a greedy strategy to process each scheduling unit sequentially, traversing according to the process order to obtain the scheduling details.

8. The method according to claim 7, characterized in that, The process of traversing according to the process sequence includes: For each process, candidate machines are first screened based on the equipment master data and packaging constraints. If it is a wire bonding process, the parallel upper limit and the target total output per unit hour are dynamically calculated. Based on the parallel upper limit and the target total output per unit hour, calculate the start time, processing time, and end time of each candidate machine, and evaluate the comprehensive cost using a cost function. The cost function includes: the difference between the completion time and the current time as the main term, the logarithm of the output per unit hour as the efficiency term, the preference score of the trained preference model as the historical habit term, the difference between the sequence number of the equipment and the anchor point and the cross-prefix penalty as the distance term, the wire bonding cycle matching gap as the penalty term, and the historical equipment of the parent work order as the reward term. Select the machine with the lowest overall cost, update the available time of the machine, record the equipment anchor point, update the set of equipment used in the master work order, and obtain the scheduling details.

9. The method according to claim 8, characterized in that, The method for dynamically calculating the parallel upper limit and the target total unit hour output includes: collecting the fusion real-time unit hour output of all candidate devices, sorting them by sequence number difference based on the anchor device, selecting local pools, calculating the discrete coefficients of high-efficiency devices from the local pools, and determining the target total unit hour output based on the discrete coefficients.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the semiconductor intelligent scheduling method based on multi-objective optimization as described in any one of claims 1-9.

Images