A smart error correction scheduling method and system for chip packaging processes
By employing intelligent deviation correction scheduling methods and utilizing data acquisition and model prediction technologies, the problem of linking equipment status with production queues after positioning deviation alarms in chip packaging equipment has been solved, thereby improving equipment utilization and production efficiency.
Patent Information
- Application Number
- CN202611375880.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
AI Technical Summary
When a positioning offset alarm is triggered on existing chip packaging equipment, the equipment status and production queue cannot be linked in real time, resulting in delayed abnormal response and reduced utilization of available equipment.
By collecting positioning offset data to identify alarm status, screening non-calibrated standby capacity, using multilayer perceptron models and autoregressive time series prediction models to predict available processing time intervals, constructing a conflict avoidance path population and performing genetic iteration, and finally realizing the flow scheduling of wafer batches.
It shortened the response time for abnormal scheduling, improved equipment utilization and the feasibility of path scheduling, reduced resource conflicts, and enhanced equipment availability and production efficiency.
Smart Images

Figure CN123117682A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of chip packaging technology, and in particular to an intelligent error correction scheduling method and system for chip packaging processes. Background Technology
[0002] Currently, the die bonding and wire bonding processes in chip packaging require equipment to maintain high-precision positioning. Mechanical wear, temperature rise, and environmental changes can cause a decrease in the actual positioning accuracy or a change in zero-point offset. When an offset alarm occurs, the alarming equipment needs to stop accepting new tasks and enter calibration mode. The original queued wafer batch needs to be transferred to a replacement machine. This transfer task will compete with the original queue on the replacement machine for processing time, and the calibration window will conflict with the original production schedule of the alarming equipment. Therefore, an industrial control system is needed to coordinate equipment status, in-process batches, calibration time periods, and production scheduling information to achieve automatic anomaly handling and task scheduling after an alarm is triggered.
[0003] Existing packaging production lines mostly employ a loosely coupled operation of local equipment-level compensation and manufacturing execution system (MES)-level rule-based task assignment. Some packaging equipment has embedded real-time dynamic visual compensation, while other equipment uses periodic calibration or single-point visual alignment, with the equipment controller performing coordinate compensation based on its own detection values. The MES forms equipment queues based on first-come, first-served, delivery-priority, and process recipe matching rules, and schedules preventative calibrations according to the calibration calendar. Equipment compensation records, calibration status, and MES task assignment queues are updated separately, without establishing real-time linkage based on the same offset alarm event. After a machine alarms, the operator still needs to read the equipment status and queued tasks, select process-compatible machines, modify the task assignment list, and then the industrial control system issues processing tasks. This loosely coupled approach creates a time lag between equipment status changes and scheduling adjustments, and causes resource conflicts between calibration periods, replacement machines' original tasks, and transferred batches.
[0004] Therefore, existing technologies have the problem that the equipment status and production queue cannot be linked in real time after the offset alarm is triggered, which can easily cause abnormal response delays and a decrease in the utilization rate of available equipment. Summary of the Invention
[0005] This invention provides an intelligent error correction scheduling method and system for chip packaging processes, which shortens the abnormal scheduling response time after an offset alarm and improves the effective utilization rate of available equipment.
[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent error correction scheduling method for chip packaging processes, comprising: Collect positioning offset data of chip packaging equipment and identify offset alarm status to obtain offset alarm status; When the offset alarm state is triggered, the alarm queue wafer set, initial process requirements and equipment operation data are obtained, and non-calibration standby capacity is screened according to the initial process requirements and the equipment operation data to obtain the standby capacity set and equipment load data. Obtain historical deviation characteristics and current production schedule; compare load time-series deviations based on the spare capacity set, equipment load data and historical deviation characteristics to obtain a deviation distribution matrix; and predict machine availability windows based on the deviation distribution matrix and current production schedule to obtain available processing time intervals. Obtain real-time production instructions, schedule wafer transfer paths based on the available processing time interval, the initial process requirements, and the real-time production instructions, and obtain an initial transfer path set and conflict detection results; When the conflict detection result indicates that a conflict exists, a conflict avoidance path population is constructed based on the conflict detection result for the initial transfer path set to obtain a candidate path population. Then, a candidate path genetic iteration is performed based on the candidate path population to obtain a converged path population. Finally, an optimal path decoding and reconstruction is performed based on the converged path population to obtain a set of corrective transfer paths. Based on the set of correction and transfer paths, the alarm queued wafer set is subjected to wafer batch transfer scheduling, and processing instructions are issued to the target processing machine according to the final transfer path.
[0007] Secondly, the present invention provides an intelligent correction and scheduling system for chip packaging processes, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.
[0008] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.
[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention collects equipment positioning offset data and identifies offset alarm status. When an alarm is triggered, it obtains the affected queued wafers, initial process requirements, and equipment operation data. Based on the process requirements and equipment operation status, it filters non-calibrated standby capacity to obtain a set of standby capacity and its equipment load data. Since the offset alarm status can limit the affected wafer batches, the process requirements can exclude incompatible machines, and the equipment operation data can reflect the current availability of the machines, this invention solves the problem in the prior art where operators have to check the machine and batch information one by one after an offset alarm. This shortens the response time for abnormal capacity switching and improves the accuracy of standby processing resource screening.
[0010] (2) This invention forms a deviation distribution matrix by comparing the equipment load data of standby machines with historical deviation characteristics over time. It then uses a multilayer perceptron model to fuse the load deviation characteristics and uses an autoregressive time series prediction model combined with the current production schedule to predict continuous low-load periods, thus obtaining the available processing time interval. Subsequently, wafer transfer path scheduling and conflict detection are performed based on the available processing time interval, initial process requirements, and real-time production instructions. Since the deviation distribution matrix retains the time difference between the current load and the historical state, the available processing time interval can reflect the future acceptance period of the machine, and the real-time production instructions can verify the new resource occupancy. This solves the problem that fixed rule dispatching is difficult to adapt to changes in machine load and production schedule in a synchronous manner, improves the executability of transfer path scheduling, and reduces the situation of new resource conflicts caused by batch reassignment.
[0011] (3) When a path conflict is detected, this invention constructs a candidate path population based on the conflict node, the alternative machine node, the real-time location of the guide vehicle, and the workshop road network topology. The path is encoded according to the double-layer integer encoding rule of the process and the machine. Fitness evaluation, roulette wheel selection, and crossover mutation iteration are implemented based on the maximum completion time and transportation distance. The optimal path in the converged path population is then decoded to replace the conflict path segment in the initial transfer path. Since the candidate paths are generated around the conflict node and the alternative resources, the genetic iteration can continuously screen the path individuals with better completion time and transportation distance from multiple path combinations. The optimal path decoding and reconstruction can convert the calculation results into an executable flow path, which solves the problem of slow manual reassignment and resource competition that still occurs after path adjustment. It reduces the waiting time and transportation time of wafer batches, and improves the abnormal scheduling response capability and available equipment utilization after the offset alarm. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of the intelligent error correction scheduling method for chip packaging process provided in the first embodiment of the present invention. Detailed Implementation
[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0014] Reference Figure 1 The first embodiment of the present invention provides an intelligent error correction scheduling method for chip packaging processes, comprising the following steps: S11, Collect positioning offset data of chip packaging equipment and identify offset alarm status to obtain offset alarm status; S12, when the offset alarm state is triggered, the alarm queue wafer set, initial process requirements and equipment operation data are obtained, and non-calibration standby capacity is screened according to the initial process requirements and the equipment operation data to obtain the standby capacity set and equipment load data. S13, obtain historical deviation characteristics and current production schedule, compare load time-series deviations based on the spare capacity set, the equipment load data and the historical deviation characteristics to obtain a deviation distribution matrix, and predict machine availability time windows based on the deviation distribution matrix and the current production schedule to obtain available processing time intervals; S14, Obtain real-time production instructions, schedule wafer transfer paths according to the available processing time interval, the initial process requirements and the real-time production instructions, and obtain the initial transfer path set and conflict detection results; S15, when the conflict detection result indicates that a conflict exists, a conflict avoidance path population is constructed for the initial transfer path set based on the conflict detection result to obtain a candidate path population, and a candidate path genetic iteration is performed based on the candidate path population to obtain a converged path population, and an optimal path decoding and reconstruction is performed based on the converged path population to obtain a correction transfer path set. S16, perform wafer batch transfer scheduling on the alarm queue wafer set according to the set of correction transfer paths, and issue processing instructions to the target processing machine according to the final transfer path.
[0015] In step S11, the positioning offset data of the chip packaging equipment is collected and the offset alarm status is identified to obtain the offset alarm status.
[0016] Specifically, the positioning offset data is synchronously generated by the existing vision alignment unit and motion controller of the chip packaging equipment after each positioning action. The acquisition device includes an equipment camera, an optical mark recognition module, a workbench position encoder, and an equipment controller. Each record of the positioning offset data is arranged sequentially as follows: equipment number, workstation number, wafer batch number, positioning action number, reference horizontal coordinate, reference vertical coordinate, measured horizontal coordinate, measured vertical coordinate, positioning quality mark, and positioning action timestamp. The reference horizontal and vertical coordinates are read from the currently effective packaging process recipe, while the measured horizontal and vertical coordinates are read from the output of the vision alignment unit and converted to micrometers according to the coordinate ratio in the workbench position encoder calibration certificate. The equipment controller uses the positioning action completion signal as the acquisition trigger signal and writes the positioning action timestamp using the equipment clock synchronized by the manufacturing execution system, so that the positioning offset data and the wafer batch history use the same time base. The positioning sampling period is taken as the single positioning cycle period specified in the equipment product technical document, and is verified by the difference between adjacent positioning action timestamps. If the deviation exceeds the equipment clock resolution, the positioning offset data is disconnected at that point. The effective range of the measured coordinates is the intersection of the calibration range of the vision alignment unit and the calibration effective range of the worktable position encoder. When the positioning quality flag is invalid, the image recognition confidence level is lower than the recognition lower limit in the vision control file, or the measured coordinates exceed the intersection, the corresponding positioning action record is deleted. When the same positioning action is recorded repeatedly, the record with the latest completion time of the equipment controller and a valid positioning quality flag is retained. The determination of duplicate records is based on the fact that the four fields of equipment number, workstation number, wafer batch number, and positioning action number are completely identical. Records that lack equipment number, wafer batch number, or positioning action timestamp will not enter the offset alarm state identification.
[0017] For each valid positioning offset data, the lateral positioning offset is first obtained by subtracting the reference lateral coordinate from the measured lateral coordinate, and then the longitudinal positioning offset is obtained by subtracting the reference lateral coordinate from the measured ordinate. The lateral offset alarm threshold and the longitudinal offset alarm threshold are respectively taken as the smaller of the maximum permissible positioning error in the corresponding direction specified in the equipment calibration certificate and the upper limit of the corresponding direction offset specified in the packaging process control plan. When the absolute value of the lateral positioning offset is not greater than the lateral offset alarm threshold and the absolute value of the longitudinal positioning offset is not greater than the longitudinal offset alarm threshold, the identification value of the current positioning action is recorded as no alarm. When the absolute value of the positioning offset in any direction exceeds the corresponding offset alarm threshold, the identification value of the current positioning action is recorded as exceeding the limit, and the number of times the same device exceeds the limit is continuously accumulated according to the positioning action timestamp from small to large. The number of consecutive exceedances is read from the equipment alarm interlock design file. When the accumulated number reaches the number of consecutive exceedances, the offset alarm status is set to triggered. When the number of consecutive exceedances is not reached and the subsequent positioning action recovers to within the two threshold ranges, the accumulated number is cleared to zero. The offset alarm status records the alarm event number, device number, workstation number, first over-limit timestamp, trigger timestamp, lateral positioning offset, longitudinal positioning offset, and trigger flag sequentially, arranged in ascending order of trigger timestamp. If the number of consecutive over-limits recorded for positioning actions is insufficient, if there are timestamp breaks in the records, or if the device enters calibration mode during the accumulation period, the current offset alarm status will not be triggered and accumulation will restart.
[0018] In step S12, when the offset alarm state is triggered, the alarm queue wafer set, initial process requirements and equipment operation data are acquired, and non-calibrated standby capacity is screened according to the initial process requirements and the equipment operation data to obtain the standby capacity set and equipment load data.
[0019] Specifically, after an offset alarm is triggered, the alarm event number and device number in the offset alarm status are used as the association key to read the wafer batches that have not yet started processing from the work-in-process queue of the manufacturing execution system, forming an alarm queue wafer set. Each record in the alarm queue wafer set is arranged sequentially with the wafer batch number, wafer box number, current process number, queue position, queue entry timestamp, delivery timestamp, and remaining maximum waiting time. The initial process requirements are read from the currently effective process route card and packaging process recipe for each wafer batch in the alarm queue wafer set. The fields include process sequence, packaging type, product model, recipe number, qualified machine group, standard processing time, changeover requirements, and maximum queuing time. Equipment operation data is read from the Manufacturing Execution System (MES) equipment interface and the equipment automation control system at the trigger timestamp of the offset alarm state. Fields include machine number, equipment type, operating status, current recipe number, calibration status, calibration validity deadline, maintenance lock flag, current processing batch, queue length, and scheduled processing time. The scheduled processing time is derived from the allowed processing start and end points in the currently effective processing instructions stored in the equipment automation control system, and is synchronously updated by the MES equipment interface when each machine completes its current processing task or issues a new processing instruction. Equipment operation data uses the same factory clock as the MES and is sampled based on the equipment status reporting cycle, which is obtained by reading the equipment interface design file and verified by the difference between adjacent equipment status timestamps. The valid range of equipment operation data is limited to the latest record with complete equipment number, operating status, and calibration status, no more than one equipment status reporting cycle prior to the trigger timestamp. Machines exceeding this time range are not included in the standby capacity screening. Alarm queue wafer sets, initial process requirements, and equipment operation data are all arranged in ascending order by wafer batch number or machine number, and their original timestamps are retained for subsequent association.
[0020] For each wafer batch in the alarm queue wafer set, first read the qualified machine group from the initial process requirements, then extract the machine records belonging to the qualified machine group from the equipment operation data to form a process-compatible machine set. Remove the alarm devices corresponding to the offset alarm state, machines with a fault or shutdown status, machines with a locked maintenance lock flag, and machines with a calibration status of "under calibration" from the process-compatible machine set. For the remaining machines, add the standard processing time, changeover time, and loading preparation time from the initial process requirements sequentially to obtain the estimated occupancy time of the wafer batch on the corresponding machine. The estimated occupancy period starts at the later of the offset alarm state trigger timestamp and the machine's existing scheduled end time, and ends at the starting point plus the estimated occupancy time. If the effective deadline for machine calibration is not later than the end of the estimated occupancy period, the current formulation is incompatible with the target formulation, or the equipment product technical documents do not record the processing capability for the corresponding package type, delete the corresponding machine record. Machines corresponding to at least one alarm queued wafer batch and passing the above screening are arranged in ascending order of machine number to obtain a spare capacity set. Each record in the set includes the machine number, compatible wafer batch number, available start time, and calibration effective end time. If no machine in the same wafer batch passes the screening, a "no available capacity" flag for that wafer batch is written into the spare capacity set. The wafer batch number, original queue position, reason for no available capacity, and writing timestamp are combined to form a pending wafer batch record. These records are then collected to obtain a pending wafer batch set. Pending wafer batches are retained in the original queue of the manufacturing execution system without adding machines that do not meet the initial process requirements.
[0021] The load observation start point is set at the trigger timestamp of the offset alarm state. The load observation duration is taken as a rolling production scheduling cycle specified in the manufacturing execution system scheduling design document, and continuous load sampling time is divided according to the equipment status reporting cycle. For each machine in the standby capacity set, the remaining processing time of the current processing batch, the standard processing time of the queued wafer batch, the changeover time, and the reserved processing time are collected at each load sampling time to obtain the planned occupancy time. The planned occupancy time is first divided by the load observation duration to obtain the machine load rate. Then, the machine load rate of machines with a value less than zero is set to zero, and the machine load rate of machines with a value greater than one is set to one. Each record of equipment load data is arranged in the following order: machine number, load sampling timestamp, machine load rate, queue length, current operating status, and most recent reserved end time, and arranged in ascending order of machine number and ascending order of timestamp.
[0022] If any machine lacks two or more consecutive load sampling moments within the load observation period, the equipment load data before and after the gap is divided into independent sequence segments, and no interpolation is performed on the missing load rate. The standby capacity set and equipment load data are directly used as inputs for load timing deviation comparison in step S13. If no standby capacity set is formed, all corresponding alarm queued wafers are added to the wafer batch set to be processed, retaining their original queue positions. An alarm indicating no available standby capacity is written to the manufacturing execution system's abnormal scheduling queue, and step S12 is re-executed after the next equipment status reporting cycle arrives. Each re-execution accumulates the number of abnormal retries. The maximum number of abnormal retries is calculated by dividing the abnormal scheduling response time limit specified in the manufacturing execution system scheduling design document by the equipment status reporting cycle, rounding down, and being at least one. When the maximum number of abnormal retries is reached, automatic retries are stopped, and a processing notification is sent to the operation terminal.
[0023] In step S13, historical deviation characteristics and the current production schedule are obtained. Based on the spare capacity set, the equipment load data, and the historical deviation characteristics, a load time-series deviation comparison is performed to obtain a deviation distribution matrix. Then, based on the deviation distribution matrix and the current production schedule, machine availability time windows are predicted to obtain the available processing time interval, including: The continuous load sampling values of each machine in the standby capacity set are extracted from the equipment load data to obtain the machine load sequence; The load sequence of the machine is segmented by a sliding window to obtain the load fluctuation sequence; The load fluctuation sequence and the historical deviation characteristics are dynamically time-normalized according to time nodes to obtain the deviation distribution matrix.
[0024] The bias distribution matrix is input into a pre-trained multilayer perceptron model for feature fusion to obtain weighted load features; When the weighted load characteristics are lower than the preset load boundary threshold, the weighted load characteristics and the current production schedule are input into the pre-built autoregressive time series prediction model to obtain the load prediction sequence. The available processing time interval is obtained by extracting consecutive low-load periods from the load prediction sequence.
[0025] Specifically, historical deviation features are derived from offset alarm correction records completed within the manufacturing execution system's retention period for the same packaging production line. Historical records consistent with the current standby capacity set are filtered by equipment type, packaging type, and process number. Only historical records with complete alarm timestamps, continuous equipment load data, completed final wafer batch flow scheduling, and no processing scrapping or secondary equipment failure are retained. Each historical deviation feature is arranged sequentially with historical event number, equipment type, process number, load sampling timestamp, historical machine load rate, and load rate difference between adjacent sampling times, arranged from earliest to latest historical event occurrence time. The number of historical templates is the smaller of the number of timing deviation values that the load prediction model input area can accommodate and the number of historical records meeting the filtering criteria. The number that the model input area can accommodate is calculated by dividing the number of bytes in the input area by the number of bytes stored for a single timing deviation value, rounded down. Historical deviation features closest to the current trigger timestamp and equal in number to the number of historical templates are selected from the valid historical records. Machine load rates with consecutive timestamps are extracted from the equipment load data by machine number and arranged in ascending order of timestamp to obtain the machine load sequence for each machine.
[0026] The sliding window length is the nearest positive integer obtained by dividing the load observation duration in step S12 by the equipment status reporting cycle. The window movement step size is one equipment status reporting cycle. Load sequences of machines with fewer than the sliding window length of load samples do not form load fluctuation sequences. Within each sliding window, the load rate of the previous machine is subtracted from the load rate of the next machine in sequence, and the resulting differences are arranged in the original timestamp order to obtain the load fluctuation sequence. The machine load rates are already within a closed interval of zero to one, and the historical machine load rates in the historical deviation characteristics use the same load rate caliber, so the two can be directly compared in time series.
[0027] For each load fluctuation sequence of each machine and each historical deviation feature with a quantity equal to the number of historical templates, dynamic time warping is performed. The local distance is the absolute value of the difference between the current load rate difference and the historical load rate difference. Dynamic time warping starts from the first difference between the two sequences, adding the current local distance to the previous load fluctuation position, the previous historical position, and the three cumulative distances corresponding to the simultaneous forward movement of both positions, and taking the minimum of the three sums as the current cumulative distance. The cumulative distance of the first position is taken as the first local distance, and the cumulative distance outside the boundary of any sequence is set to an unreachable value, so that the warping path must extend from the beginning to the end of the two sequences. The warping bandwidth is the maximum status reporting delay specified in the equipment interface design document divided by the equipment status reporting cycle and rounded up. Sequence positions exceeding the warping bandwidth are not included in the matching. The cumulative distance at the end is divided by the number of positions included in the warping path to obtain the time series deviation value of the current load fluctuation sequence relative to the corresponding historical deviation feature. All time series deviation values are arranged according to the machine number in the standby capacity set as the row and the historical event occurrence time as the column to obtain the deviation distribution matrix. The number of rows in the deviation distribution matrix represents the number of standby generators, and the number of columns represents the number of historical templates. When any load fluctuation sequence or historical deviation feature contains non-finite load rates, non-existent regular paths, or non-finite cumulative distances at the ends, no deviation distribution matrix rows are generated for the corresponding generators.
[0028] The pre-trained multilayer perceptron model adopts a per-machine shared structure, performing feature fusion on each row of the deviation distribution matrix. The model input consists of time-series deviation values arranged chronologically by historical event occurrence time, with a dimension equal to the number of historical templates. The model output is the weighted load feature for the next load observation period for the corresponding machine. The per-machine shared structure means that all standby machines share the same set of model weights and bias parameters. The model inputs the row vector of the deviation distribution matrix for each machine, outputting the corresponding weighted load feature row by row. Different machines do not have independent model parameters. The fitted samples of the multilayer perceptron model are derived from the same batch of valid historical records that formed the historical deviation features. Each fitted sample uses the row of the deviation distribution matrix before the trigger time of the historical event as the independent variable and the arithmetic mean of the measured machine load rate within one load observation period after the trigger time as the dependent variable. The fitted samples are arranged from earliest to latest according to the occurrence time of historical events, with the first 80% used as training samples and the last 20% as validation samples. This division ratio is written into the load prediction model design file.
[0029] The multilayer perceptron model consists of an input layer with the same number of nodes as the historical templates, a hidden layer with twice the number of nodes as the historical templates, and an output layer with one node. The weight matrix between the input and hidden layers has twice the number of rows and columns as the historical templates, and the hidden layer bias vector has twice the number of components. The hidden layer first multiplies the input column vector by the weight matrix, then adds the hidden layer bias vector component by component, and sets the negative components to zero to obtain the hidden layer output vector. The output layer weight vector has twice the number of components as the historical templates. The output layer weight vector is multiplied component by component with the hidden layer output vector, summed, and then the output layer bias is added. Finally, a function obtained by dividing one by one and the sum of the negative net input power of the natural constant is used to restrict the weighted load feature to an open interval between zero and one. The dependent variable in the training and validation samples lies in a closed interval between zero and one; historical records outside this interval are deleted as invalid fitting samples.
[0030] For each input location, the time-series deviation values in the training samples are arranged in ascending order. The first percentile and the 99th percentile are read. The normalized time-series deviation value is obtained by subtracting the first percentile value from the current time-series deviation value and then dividing the result by the difference between the 99th and 100th percentile values. If the difference between the 99th and 100th percentile values is less than the equipment load rate numerical resolution, the equipment load rate numerical resolution is used as the divisor, and the normalized time-series deviation value is restricted to a closed interval between zero and one. The equipment load rate numerical resolution is read from the decimal places of the load rate defined in the equipment interface field. Each weight and bias uses the software version number in the load prediction model design file to generate a fixed pseudo-random seed. The initial weight values are uniformly generated between the square root and the reciprocal of the number of positive and negative corresponding input nodes, and the initial bias value is set to zero.
[0031] During training, the model output is subtracted from the corresponding dependent variable, the squared differences are summed, and divided by the number of training samples to obtain the training error. Then, the error gradients of each weight and bias are calculated in reverse order from the output layer to the hidden layers. In each update round, the current weight or bias is subtracted from the product of the update step size and the corresponding error gradient. The update step size is the reciprocal of the square of the largest singular value of the training input matrix. Offline training terminates when the largest singular value is zero. Updates stop when the training error is no greater than the square of the numerical resolution of the device load rate. If this condition is not met, the update stops at the maximum number of iterations. The maximum number of iterations is the allowed offline training time specified in the load prediction model design document divided by the longest single-round execution time in the software test report, rounded down. The upper limit of the training error is the square of the numerical resolution of the device load rate. After each update round, validation samples are used to calculate the validation error. Only the weights and biases with the smallest validation error that is not greater than the upper limit of the training error are saved. If offline training fails to generate weights and biases that meet the conditions, the most recently validated multilayer perceptron model within its effective version period will continue to be called. If there is no effective model version, a record of wafer batches to be disposed of will be formed with the corresponding wafer batch number, original queuing position, reason for model unavailability, and write timestamp. The record will be added to the set of wafer batches to be disposed of and will not be included in subsequent predictions. If there is an effective model version, each effective row of the deviation distribution matrix will be input sequentially and the model output will be arranged according to the machine number to obtain the weighted load features that correspond one-to-one with the machines in the standby capacity set.
[0032] The load boundary threshold is the smaller of the upper limit of the low load rate specified in the manufacturing execution system scheduling design document and the 25th percentile of the machine load rate of the successfully received alarm queued wafers in the effective historical correction records. The 25th percentile of the historical machine load rate is determined by arranging the historical machine load rates that meet the screening conditions in step S13 from smallest to largest and reading the 25th percentile position. When the weighted load characteristic is not lower than the load boundary threshold, the corresponding machine is marked as not entering the load forecast; when it is lower than the load boundary threshold, the weighted load characteristic and the current production schedule are input into the pre-built autoregressive time series forecast model.
[0033] It should be noted that the pre-construction refers to solving the model coefficients through singular value decomposition based on historical machine load sequences in an offline state, and saving the obtained autoregressive coefficients, characteristic coefficients, planned occupancy rate coefficients, and constant terms as a model parameter file; during online prediction, the model parameter file is directly loaded, and the current input data is input in the order of model order, weighted load characteristics, and planned occupancy rate, multiplied by the corresponding coefficients, summed, and the constant term is added to obtain the prediction result for the next time step; the model does not rely on online retraining; the fitted sequence of the autoregressive time series prediction model comes from the same batch of valid historical machine load sequences that generated the multilayer perceptron fitted samples, and is established separately according to machine type.
[0034] The model input consists of machine load rates, current weighted load characteristics, and planned occupancy rates of the current production schedule at the next sampling time, arranged sequentially up to the current sampling time and in quantities equal to the model order. The model output is the machine load rate at the next sampling time. The model order is calculated by dividing the load observation duration in step S12 by the equipment status reporting cycle and rounding down, ensuring that the result is at least one and not greater than the shortest consecutive sampling quantity of any valid fitted sequence minus one. The planned occupancy rate of the current production schedule is calculated sequentially according to the sampling interval. When each scheduled task spans multiple sampling intervals, the overlap duration between the task's occupied time period and each crossed sampling interval is calculated. The overlap durations are accumulated only to the corresponding sampling interval, and then the cumulative overlap duration of the corresponding sampling interval is divided by the equipment status reporting cycle and restricted to a closed interval between zero and one.
[0035] Arrange all valid fitted sequences according to the historical event occurrence time. For each output machine load rate, along with its corresponding model order, previous machine load rates, weighted load characteristics, and planned occupancy rates, form a fitted record. The input objects of each fitted record are used to form a fitted matrix, and the corresponding output machine load rates are used to form a load rate column vector. Singular value decomposition is then performed on the fitted matrix. Singular values not greater than the product of the largest singular value and the minimum resolution floating-point value of the scheduling server are set to zero. The reciprocals of the remaining singular values are multiplied by the left and right singular vectors to form the pseudo-inverse of the fitted matrix. This pseudo-inverse is then multiplied by the load rate column vector to obtain the autoregressive coefficients, characteristic coefficients, planned occupancy rate coefficients, and constant term. The minimum resolution floating-point value of the scheduling server is read from the server processor's technical file. If the fitted matrix does not contain any singular values greater than the product, or if the number of valid fitted records is not greater than the number of coefficients, no new autoregressive time series prediction model is generated, and the most recent model that has passed fitting verification and is within its valid version period continues to be used. When no valid model version exists, a record of wafer batches to be processed is formed using the corresponding wafer batch number, original queuing position, reason for unavailability of the load prediction model, and write timestamp. This record is added to the set of wafer batches to be processed without subsequent time window extraction. When a valid model version exists, the machine load rate with the most recent quantity equal to the model order, the current weighted load characteristic, and the planned occupancy rate of the current production schedule are multiplied by the corresponding coefficients and added to a constant term to obtain the predicted load rate at the next sampling time. The newly generated predicted load rate is moved into the most recent machine load rate sequence and the earliest load rate is deleted. The predicted timestamp is incremented according to the equipment status reporting cycle and repeated until it covers the end timestamp of the current production schedule. The predicted load rate is set to zero when it is below zero and set to one when it is above one. If any input has a non-finite value or a timestamp gap, the subsequent prediction for the corresponding machine is stopped. The valid predicted load rates are arranged in ascending order of machine number and in ascending order of predicted timestamp to obtain the load prediction sequence.
[0036] Sampling times with predicted load rates not exceeding the load boundary threshold are extracted from the load forecast sequence by machine number. The time continuity tolerance is the sum of the maximum permissible jitter of the timestamp specified in the equipment interface design document and the maximum synchronization error stated in the factory time synchronization service test report. When the absolute value of the difference between adjacent timestamps and the difference between the equipment status reporting cycle is not greater than the time continuity tolerance, the corresponding sampling times are merged into a continuous low-load period. The start point of each continuous low-load period is the first sampling timestamp, and the end point is the last sampling timestamp plus the equipment status reporting cycle. The period length is the end point minus the start point. The minimum wafer batch acceptance time is obtained by sequentially adding the standard processing time, changeover time, and loading preparation time in the initial process requirements, and then matching them according to the corresponding machine and wafer batch. The continuous low-load period is retained when the length is not less than the minimum wafer batch acceptance time and does not overlap with the maintenance period, calibration period, and locked task period in the current production schedule. The retained periods are arranged according to machine number, compatible wafer batch number, start timestamp, and end timestamp to obtain the available processing time interval. When adjacent available processing time intervals of the same machine are within the time continuity tolerance and apply to the same wafer batch, they are merged into one interval. If an interval endpoint is missing or the end point is no later than the start point, the corresponding interval is deleted. When any alarm queued wafer does not form an available processing time interval, its no-available-time-window flag is retained. A wafer batch record to be processed is formed with the corresponding wafer batch number, original queue position, reason for no available-time-window, and write timestamp. The record is added to the wafer batch set to be processed, its original queue position in the manufacturing execution system is retained, and the wafer batch is stopped from entering step S14.
[0037] In step S14, real-time production instructions are obtained, and wafer transfer paths are scheduled according to the available processing time interval, the initial process requirements, and the real-time production instructions to obtain an initial transfer path set and conflict detection results.
[0038] Specifically, real-time production instructions are read from the manufacturing execution system's dispatch interface. The reading time is determined by the end-of-processing timestamp of the available processing time interval generated in step S13, using the unified clock of the manufacturing execution system. Each record of a real-time production instruction is arranged sequentially as follows: production instruction number, wafer batch number, current process number, target quantity to be completed, production priority, instruction issuance timestamp, required completion timestamp, and instruction version number. Only real-time production instructions whose instruction version number is the currently effective version, whose wafer batch number belongs to the alarm queue wafer set, and whose current process number is consistent with the initial process requirements are retained. If the required completion timestamp of a real-time production instruction is earlier than the instruction issuance timestamp, the target quantity to be completed is greater than the corresponding batch quantity in the alarm queue wafer set, or the production instruction number is duplicated, the corresponding real-time production instruction is marked as invalid. Valid real-time production instructions are associated with the available processing time interval according to the wafer batch number, and their consistency is verified with the qualified machine group, recipe number, and standard processing time in the initial process requirements. The associated records are arranged in ascending order according to production priority from high to low, remaining maximum waiting time from short to long, instruction issuance timestamp from early to late, and wafer batch number, forming a path scheduling order.
[0039] Wafer batches are processed sequentially according to the path scheduling. From the available processing time intervals, machine time slots belonging to qualified machine groups with an interval length not less than the sum of the standard processing time and the changeover time are extracted. For each machine time slot, the later of the interval start and the corresponding machine's occupied time slot end is used as the expected processing start. The standard processing time and changeover time are then added to the expected processing start to obtain the expected processing end. If the expected processing end is not later than the interval end and not later than the completion timestamp of the real-time production instruction, the corresponding machine time slot is retained and sorted in ascending order of expected processing end and machine number. The machine time slot at the top of the list is selected as the target machine time slot for the current wafer batch. The standard material flow node sequence from the current storage node to the target machine loading node is read from the initial process requirements to form an initial transfer path. Each record of the initial transfer path is arranged sequentially with the path number, production instruction number, wafer batch number, source node, standard flow node sequence, target machine node, expected handling start, expected processing start, and expected processing end. All records are then combined to obtain the initial transfer path set. According to the expected processing time period in the initial transfer path set, when the expected processing time periods of two paths overlap or the same wafer batch corresponds to two target machine nodes, the conflict detection result is set to "conflict exists". The conflict detection result records the conflict flag, conflict type, conflicting machine node, associated path number, overlap start time stamp, and overlap end time stamp in sequence, and is arranged according to the conflicting machine node and overlap start time stamp. Wafer batches that have not formed a target machine time period are written into the unscheduled flag of the initial transfer path set, and a record of wafer batches to be processed is formed with the corresponding wafer batch number, original queue position, reason for not forming a target machine time period, and writing timestamp. The record is added to the set of wafer batches to be processed, retains its original queue position in the manufacturing execution system, and is not included in the processing instruction issuance; when the conflict detection result does not show any conflict, the initial transfer path set containing the valid path record is directly determined as the correction transfer path set and proceeds to step S16.
[0040] In step S15, when the conflict detection result indicates the existence of a conflict, a conflict avoidance path population is constructed based on the conflict detection result of the initial transfer path set to obtain a candidate path population. Then, a candidate path genetic iteration is performed based on the candidate path population to obtain a converged path population. Finally, an optimal path decoding and reconstruction is performed based on the converged path population to obtain a correction transfer path set, including: The conflict nodes are obtained by extracting the machine nodes that are competing for resources from the conflict detection results. Obtain the status of redundant equipment with the same process capability in the workshop, and filter out idle machine nodes based on the status of the redundant equipment to obtain alternative machine nodes; Obtain the real-time location of the guide vehicle and the workshop road network topology, and perform a depth-first path search based on the conflict node, the alternative machine node, the real-time location of the guide vehicle, and the workshop road network topology to obtain candidate transfer paths; The candidate transfer paths are aggregated to obtain the candidate path population.
[0041] Each candidate transfer path in the candidate path population is encoded according to the process and machine tool double-layer integer encoding rule to obtain the path encoding sequence; The fitness of the path coding sequence is evaluated based on the estimated path duration and transport distance corresponding to the path coding sequence to obtain an individual fitness value. The path encoding sequence is selected by roulette wheel selection based on the individual fitness value; The selected path encoding sequence is updated by crossover mutation to obtain the updated path population; The updated path population is used as the candidate path population for the next round. The continuous steps from fitness evaluation to crossover and mutation update are repeated until the optimal fitness value remains unchanged or the number of iterations reaches a preset number, thus obtaining the converged path population.
[0042] Based on the maximum completion time and total transport distance of each path individual in the convergence path population, the fitness of each path individual is evaluated to obtain a convergence fitness sequence. Based on the convergence fitness sequence, the path encoding sequence with the highest fitness value is selected from the convergence path population to obtain the optimal path encoding sequence; The optimal path encoding sequence is decoded to obtain the optimal transfer path; Extract the path segments corresponding to the conflict detection results from the initial set of transfer paths to obtain conflict path segments; The optimal transfer path is used to replace and reconstruct the conflicting path segments to obtain the set of corrective transfer paths.
[0043] The step of evaluating the fitness of the path coding sequence based on the estimated path duration and transport distance corresponding to the path coding sequence to obtain an individual fitness value includes: Extract the machine processing nodes and guide vehicle transport segments of each candidate transfer path from the path coding sequence to obtain the processing node sequence and the transport segment sequence. The estimated path duration is obtained by accumulating the machine processing time based on the processing node sequence; The transport distance is obtained by accumulating the distances of the transport segments according to the transport segment sequence; Using minimizing the maximum completion time and minimizing the total transport distance as evaluation objectives, a fitness mapping is performed on the estimated path duration and the transport distance to obtain the individual fitness value.
[0044] The step of performing crossover mutation update on the selected path encoding sequence to obtain the updated path population includes: The selected path encoding sequences are combined according to the pairing relationship to obtain path encoding sequence pairs; Partial mapping and cross-mapping are performed on the path encoding sequence pairs to obtain cross-path encoding sequences; Single-point mutation is performed on the machine allocation gene in the cross-path coding sequence to obtain the updated path coding sequence; The updated path encoding sequences are aggregated to obtain the updated path population.
[0045] Specifically, when a conflict is detected, the machine number involved in the resource contention is extracted from the conflict machine node field of the conflict detection result, and duplicate numbers are removed to obtain the conflict node. The redundant equipment status is read from the timestamp generated by the equipment automation control system at the time of the conflict detection result generation. Fields include machine number, equipment type, process capability code, current operating status, current recipe, calibration validity deadline, maintenance lock flag, and loading node number. For each conflict node, a replacement machine node is obtained by retaining the machine whose process capability code covers the initial process requirements, whose current operating status is idle, whose maintenance lock flag is unlocked, and whose calibration validity deadline is later than the corresponding expected processing endpoint. The real-time location of the guided vehicle is read from the guided vehicle scheduling system. Fields include guided vehicle number, current location node, operating status, remaining power, and location timestamp. The difference between the location timestamp and the conflict detection result generation timestamp must not exceed the location reporting cycle specified in the guided vehicle interface design document. The workshop road network topology is read from the currently active digital map of the guided vehicle scheduling system, arranged according to node number, adjacent node number, road segment direction, road segment distance, traffic status, and occupancy status. For each alarm-queued wafer, the current location node of the corresponding guide vehicle is used as the search starting point, and the loading nodes of each alternative machine are used as the search ending point. A depth-first path search is performed in ascending order of adjacent node numbers. During the search, visited nodes are written into the current access sequence, and the estimated handling time of the next segment is obtained by dividing the current cumulative segment distance by the rated operating speed. If the next node is already in the current access sequence and the segment's passage status is closed or occupied, continuously covering the estimated handling time, the next node is not entered. The upper limit of the search depth is taken as the total number of nodes in the workshop road network topology. When the alternative machine loading node is reached, the current access sequence, the alternative machine node, and the corresponding wafer batch are combined into a candidate transfer path, and the candidate transfer paths are aggregated to obtain a candidate path population.
[0046] The estimated transport time of candidate transfer routes is obtained by dividing the distance of each segment by the rated operating speed in the technical document of the guided vehicle product and then summing the results. If the remaining power is lower than the sum of the power required to complete the candidate transfer route and the minimum return power of the guided vehicle, the corresponding candidate transfer route is deleted. The power required for a candidate transfer route is obtained by multiplying the distance of each segment by the power consumption per unit distance in the technical document of the guided vehicle product, summing the results, and then adding the power consumption for loading and unloading operations. The minimum return power is read from the guided vehicle scheduling safety file. The candidate route population size is the smaller of the number of candidate transfer routes and the number of routes that the scheduling software can accommodate. The number of routes that the scheduling software can accommodate is calculated by dividing the number of bytes in the path population memory area by the number of bytes in the longest single path code and then rounding down. When the number of candidate transfer routes is greater than the candidate route population size, they are first arranged in order of estimated transport time from shortest to longest, the estimated processing endpoint of the substitute machine from earliest to latest, and the number of path nodes from fewest to most. Then, candidate transfer routes with a number equal to the candidate route population size are retained.
[0047] Each candidate transfer path is encoded according to a two-layer integer coding rule for processes and machines. The first layer writes the process numbers according to the process order required by the initial process requirements. The second layer writes the machine number that undertakes the process at the corresponding gene position. The network node sequence of the candidate transfer path is associated with the corresponding machine gene by the path index, without adding process genes that do not belong to the initial process requirements. The two-layer integer coding rule for processes and machines adopts fixed-length coding, and the total coding length is equal to the total number of remaining processes corresponding to the alarm queue wafers. The process numbers in the first layer are fixedly arranged according to the process execution order in the initial process requirements. The machine numbers in the second layer correspond one-to-one with the first layer according to the gene position. The process number and machine number at the same gene position together constitute the processing allocation scheme of that process on the corresponding machine. The number of gene positions in all path coding sequences is taken as the number of remaining processes in the initial process requirements corresponding to the alarm queue wafers. Wafer batches with inconsistent process numbers form independent candidate path populations. If any path encoding sequence contains a machine number that is not in the qualified machine group, a duplicate process number, or a missing process number, the corresponding path encoding sequence is deleted. If the candidate path population is less than two, genetic iteration is not performed, and the only valid candidate transfer path is taken as the optimal transfer path.
[0048] From each path coding sequence, machine processing nodes and associated guide vehicle transport segments are extracted according to gene location, resulting in a processing node sequence and a transport segment sequence. For each machine in the processing node sequence, the latest time among the machine's existing task end time, guide vehicle arrival time, and the start of the available processing time interval is taken as the processing start point. Then, the corresponding standard processing time and changeover time are added to the processing start point to obtain the machine processing end point. The maximum difference between the processing end points of all machines in the same path coding sequence and the timestamp of the real-time production instruction is taken to obtain the estimated path duration. Finally, the distances of each segment in the transport segment sequence are added together to obtain the transport travel distance.
[0049] Within the current candidate path population, the minimum and maximum values of the estimated path duration and transport distance are read respectively. The normalized estimated duration is obtained by subtracting the corresponding minimum value from the estimated duration of each path and then dividing by the difference between the maximum and minimum values. The normalized transport distance is obtained using the same subtraction and division order. If any maximum and minimum value are the same, the corresponding normalized value is set to zero. Both normalized values are restricted to a closed interval between zero and one. The maximum completion time and total transport distance are used as two parallel evaluation objectives. The normalized estimated duration and normalized transport distance are each multiplied by half and then added together to obtain the comprehensive cost of the path. The weights of the two objectives are fixed, and no adjustable objective weights are set. Within the current candidate path population, the minimum and maximum values of the overall path cost are read. If they differ, the individual fitness value within the closed interval of zero to one is obtained by subtracting the current path cost value from the maximum path cost value and dividing by the difference between the maximum and minimum path cost values. If they are the same, all individual fitness values are set to one to amplify the difference between superior and inferior individuals and avoid fitness concentration. When the variance of the population's individual fitness values is lower than the current generation's fitness value range (i.e., 1% of the difference between the maximum and minimum values), the selection pressure is deemed insufficient, and the fitness value is squared to amplify the differences between individuals. If the range is zero, the original value remains unchanged. Here, 1% is the default threshold, and the specific value can be adjusted linearly between 1% and 5% according to the population size of the genetic algorithm; the larger the population size, the lower the threshold. The current generation fitness sequence is obtained by arranging the fields in the order of path encoding sequence first, path estimated duration, transport distance, and individual fitness value last.
[0050] It should be noted that the roulette wheel selection maps the proportion of each individual's fitness value to the total fitness value of all individuals to the length of the roulette wheel interval. The higher the fitness value of an individual, the longer its roulette wheel interval. Random numbers are generated using a fixed pseudo-random seed, and individuals are selected based on the cumulative roulette wheel interval in which the random numbers fall, so that individuals with high fitness values have a higher probability of being selected.
[0051] The fitness values of all individuals in the current generation fitness sequence are summed, and each individual fitness value is divided by the sum to obtain the roulette interval length of the corresponding path encoding sequence. When the sum of individual fitness values is zero, the roulette interval length of each path encoding sequence is the reciprocal of the candidate path population size. A fixed pseudo-random seed is formed by combining the alarm event number and the current iteration round, and selection values between zero and one are generated sequentially. The path encoding sequence is selected according to the cumulative roulette interval in which the selection value falls, resulting in the selected path encoding sequence. Crossover and mutation updates are performed in the order of crossover first and mutation second, sequentially exchanging gene fragments and randomly adjusting machine numbers on the parent path encoding sequence, so that each iteration produces a child path encoding sequence different from the parent, so as to continuously explore new path combinations in the solution space. Crossover is used to exchange superior gene fragments between parents, and mutation is used to prevent the population from getting trapped in local optima. The two work together to achieve a balance between global search and local search. Every two paths are paired according to the order of selection. When the population size is odd, the path encoding sequence with the highest individual fitness value is paired with the last selected path encoding sequence. For each path coding sequence pair, partial mapping crossover is performed at the process coding layer. The two crossover positions are the rounded-down positions corresponding to one-third and two-thirds of the number of gene positions, respectively, and the machine codes at the corresponding positions are exchanged simultaneously. If the number of gene positions is less than three, partial mapping crossover is not performed, and the parent path coding sequence is directly copied. For each crossover path coding sequence, a machine is selected to assign a gene, and the original machine number is replaced with the next qualified replacement machine number arranged in ascending order. If the original machine number is the last in ascending order, it wraps back to the first. The selected gene position for mutation is determined by reading the combination value of the current iteration round and the path coding sequence length, taking the modulo of the path coding sequence length, and adding one to the modulo value as the mutation position to ensure that each gene position is evenly covered during the iteration process. When there is only one qualified replacement machine in the current process, the original machine number is retained. The path coding sequence with the highest fitness value of the current generation is retained as is, and the crossover path coding sequences and mutated path coding sequences that have been verified by time interval, process order, and road network connectivity are collected to obtain the updated path population.
[0052] It should be noted that after partial mapping crossover, if a certain process number is repeated or missing in the path coding sequence, the repeated process will be replaced with the missing process in the order of the complement of the existing process numbers in the uncrossed part, so that each process number appears exactly once in the sequence, and the machine number of the corresponding machine coding layer will be adjusted synchronously to form a valid path coding sequence.
[0053] When the number of updated path populations is less than the size of the candidate path population, non-repeating parent path coding sequences are added according to the parent individual fitness values from high to low. When the number exceeds the size of the candidate path population, path coding sequences equal to the size of the candidate path population are retained according to the newly calculated individual fitness values from high to low. The updated path population is used as the candidate path population for the next round, and the current generation fitness sequence is regenerated, recording the path coding sequence with the highest individual fitness value. The consecutive optimal maintenance rounds are taken as the number of gene positions in the path coding sequence. Iteration stops when the highest individual fitness value and the optimal path coding sequence remain unchanged for consecutive rounds reaching the specified number. The maximum number of iterations is calculated by dividing the abnormal scheduling response time limit specified in the manufacturing execution system scheduling design document by the longest execution time of a single round of genetic iteration in the scheduling software test report, rounded down, and at least to 1. Iteration stops when the maximum number of iterations is reached, without continuing to occupy scheduling response time due to not reaching the consecutive optimal maintenance rounds. The candidate path population at the time of stopping iteration is arranged according to the individual fitness values from high to low to obtain the converged path population. When no feasible path encoding sequence exists in the candidate path population during the iteration, no converged path population is generated. A non-feasible correction path flag is written into the conflict detection result to form a wafer batch record to be processed by associating the wafer batch number, original queuing position, reason for non-feasible correction path, and writing timestamp. The record is then added to the set of wafer batches to be processed.
[0054] For each path individual in the converged path population, the maximum completion time, total transport distance, and individual fitness value are recalculated in the same order as described above, resulting in a converged fitness sequence. The path coding sequence with the highest individual fitness value is selected from the converged fitness sequence. If individual fitness values are the same, the path coding sequences with shorter maximum completion time, shorter total transport distance, and smaller lexicographical order are selected in that order to obtain the optimal path coding sequence. The process execution order is restored according to the process coding layer, the target machine for each process is restored according to the machine coding layer, and the guide vehicle transport segment is restored according to the associated road network node sequence to obtain the optimal transfer path. Corresponding path segments are extracted from the initial transfer path set according to the associated path number and overlapping timestamp of the conflict detection results to obtain conflict path segments. The optimal transfer path is used to replace the conflict path segments, keeping the source node before the conflict path segment and the non-conflicting target node after the conflict path segment unchanged, to obtain the reconstructed path. When adjacent nodes of the reconstructed path are not connected, the target machine does not belong to the qualified machine group, the expected processing time exceeds the available processing time interval, or duplicate machine allocation occurs in the same wafer batch, the next undecoded path encoding sequence is used according to the convergence fitness sequence order until the path is reviewed or all path encoding sequences are reviewed. All reconstructed paths that have passed the review are collected and arranged according to the wafer batch number and the expected handling start point to obtain the correction and transfer path set. When all path encoding sequences fail the review, a "no solution" flag for the reconstructed path review is written into the conflict detection result. This flag is associated with the wafer batch number, the original queue position, the reason for the no solution in the reconstructed path review, and the timestamp to form a wafer batch record to be processed. This record is added to the wafer batch set to be processed, retains its original queue position in the manufacturing execution system, and is not included in the processing instruction issuance. This step does not call external classification models, regression models, or interpolation models. The candidate path population, fitness mapping, iteration boundary, and abnormal branches are all formed by the input objects of this step and the determination process.
[0055] In step S16, the alarm queued wafer set is batch-based for wafer transfer scheduling according to the set of correction transfer paths, and processing instructions are issued to the target processing machine according to the final transfer path.
[0056] Specifically, wafer batches with valid path records in the correction transfer path set are associated one by one with the alarm queue wafer set according to their wafer batch numbers, forming a transfer scheduling record containing wafer batch number, wafer box number, source node, final transfer path, target machine node, expected handling start point, expected processing start point, and expected processing end point. Wafer batches in the pending wafer batch set without associated valid path records are formed into anomaly handling records, with fields including alarm event number, wafer batch number, original queue position, handling reason, number of anomaly retries, and most recent processing timestamp. These wafer batches remain in the original queue of the manufacturing execution system and do not generate handling tasks or processing instructions. The transfer scheduling records are arranged in ascending order according to the expected handling start point (earliest to latest), production priority (highest to lowest), remaining maximum waiting time (shortest to longest), and wafer batch number. For each transfer scheduling record, it is first verified that the wafer batch is still at the source node and not locked by other scheduling tasks, then verified that the target machine's operating status is idle or waiting for materials, the effective calibration deadline is later than the expected processing end point, and the target machine's formula is consistent with the initial process requirements. If any verification condition is not met, the current workflow scheduling record will not be issued. The abnormal retry count for the alarm event will be incremented by one. If the maximum abnormal retry count is not reached, the latest equipment operation data of the target machine will be returned to step S12 to re-execute the backup capacity screening. When the maximum abnormal retry count mentioned in step S12 is reached, the corresponding wafer batch will be added to the wafer batch set to be processed, automatic return will be stopped, and a processing notification will be sent to the operation terminal. The verified workflow scheduling record uses a manufacturing execution system transaction lock to lock the wafer batch number and the target machine node. The effective duration of the transaction lock is the length of time from the expected handling start point to the expected processing end point. When there are multiple verified workflow scheduling records for the same wafer batch, only the first one in the corrective transfer path set will be retained, and the transaction locks corresponding to the other records will be released. The final workflow path is formed by sequentially connecting the source node, the guide car handling node sequence, and the target machine node in the retained workflow scheduling records.
[0057] The system issues a handling task to the guided vehicle scheduling system according to the final flow path. The handling task fields are arranged in the following order: task number, wafer batch number, wafer box number, source node, guided vehicle handling node sequence, target machine loading node, earliest start timestamp, and task version number. After the guided vehicle scheduling system returns a task acceptance flag and a receipt timestamp, it issues a processing instruction to the equipment automation control interface of the target processing machine. The processing instruction fields are arranged in the following order: production instruction number, wafer batch number, process number, recipe number, target quantity, allowed processing start point, allowed processing end point, instruction version number, and checksum. The checksum is generated by concatenating the fields in the processing instruction (excluding the checksum) in the field order and calling the cyclic redundancy check rule specified in the manufacturing execution system interface design document. The target processing machine only accepts the processing instruction if the checksum matches and the instruction version number is higher than the received version number. After the target processing machine returns an instruction acceptance flag, it updates the wafer batch status to "pending handling," updates the target machine's reserved time slot to "locked," and records the issuance timestamp and receipt timestamp. If a handling task or processing instruction does not return a receipt flag within the response time limit specified in the manufacturing execution system interface design document, a new task number is not generated; instead, the same task number and an incrementing version number are used for resending. When the number of resends reaches the upper limit specified in the interface design document, the transaction lock is released and the number of abnormal retries is incremented. If the maximum number of abnormal retries is not reached, the process returns to step S14 for rescheduling. When the maximum number of abnormal retries is reached, the corresponding wafer batch is added to the set of wafer batches to be processed, its original queue position is retained, and a processing notification is sent to the operation terminal. Processing begins after the guided vehicle arrives at the loading node of the target machine and the target processing machine verifies that the wafer box number, process number, and recipe number are consistent. After completion, the actual processing start and end times are written into the wafer batch history.
[0058] The second embodiment of the present invention provides an intelligent correction and scheduling system for a chip packaging process, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.
[0059] It should be noted that the intelligent deviation correction scheduling system for chip packaging process provided in this embodiment of the invention is used to execute all process steps of the intelligent deviation correction scheduling method for chip packaging process in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.
[0060] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.
Claims
1. An intelligent deviation correction scheduling method for a chip packaging process, characterized in that, include: Collect positioning offset data of chip packaging equipment and identify offset alarm status to obtain offset alarm status; When the offset alarm state is triggered, the alarm queue wafer set, initial process requirements and equipment operation data are obtained, and non-calibration standby capacity is screened according to the initial process requirements and the equipment operation data to obtain the standby capacity set and equipment load data. Obtain historical deviation characteristics and current production schedule; compare load time-series deviations based on the spare capacity set, equipment load data and historical deviation characteristics to obtain a deviation distribution matrix; and predict machine availability windows based on the deviation distribution matrix and current production schedule to obtain available processing time intervals. Obtain real-time production instructions, schedule wafer transfer paths based on the available processing time interval, the initial process requirements, and the real-time production instructions, and obtain an initial transfer path set and conflict detection results; When the conflict detection result indicates that a conflict exists, a conflict avoidance path population is constructed based on the conflict detection result for the initial transfer path set to obtain a candidate path population. Then, a candidate path genetic iteration is performed based on the candidate path population to obtain a converged path population. Finally, an optimal path decoding and reconstruction is performed based on the converged path population to obtain a set of corrective transfer paths. Based on the set of correction and transfer paths, the alarm queued wafer set is subjected to wafer batch transfer scheduling, and processing instructions are issued to the target processing machine according to the final transfer path.
2. The intelligent deviation correction scheduling method for chip packaging process according to claim 1, characterized in that, The step of comparing load time-series deviations based on the set of reserve capacity, the equipment load data, and the historical deviation characteristics to obtain a deviation distribution matrix includes: The continuous load sampling values of each machine in the standby capacity set are extracted from the equipment load data to obtain the machine load sequence; The load sequence of the machine is segmented by a sliding window to obtain the load fluctuation sequence; The load fluctuation sequence and the historical deviation characteristics are dynamically time-normalized according to time nodes to obtain the deviation distribution matrix.
3. The intelligent deviation correction scheduling method for chip packaging process according to claim 1, characterized in that, The step of predicting machine availability windows based on the deviation distribution matrix and the current production schedule to obtain the available processing time interval includes: The bias distribution matrix is input into a pre-trained multilayer perceptron model for feature fusion to obtain weighted load features; When the weighted load characteristics are lower than the preset load boundary threshold, the weighted load characteristics and the current production schedule are input into the pre-built autoregressive time series prediction model to obtain the load prediction sequence. The available processing time interval is obtained by extracting consecutive low-load periods from the load prediction sequence.
4. The intelligent deviation correction scheduling method for chip packaging processes according to claim 1, characterized in that, The step of constructing a conflict avoidance path population based on the conflict detection results of the initial transfer path set to obtain a candidate path population includes: The conflict nodes are obtained by extracting the machine nodes that are competing for resources from the conflict detection results. Obtain the status of redundant equipment with the same process capability in the workshop, and filter out idle machine nodes based on the status of the redundant equipment to obtain alternative machine nodes; Obtain the real-time location of the guide vehicle and the workshop road network topology, and perform a depth-first path search based on the conflict node, the alternative machine node, the real-time location of the guide vehicle, and the workshop road network topology to obtain candidate transfer paths; The candidate transfer paths are aggregated to obtain the candidate path population.
5. The intelligent deviation correction scheduling method for chip packaging processes according to claim 1, characterized in that, The step of performing genetic iteration on candidate paths based on the candidate path population to obtain a convergent path population includes: Each candidate transfer path in the candidate path population is encoded according to the process and machine tool double-layer integer encoding rule to obtain the path encoding sequence; The fitness of the path coding sequence is evaluated based on the estimated path duration and transport distance corresponding to the path coding sequence to obtain an individual fitness value. The path encoding sequence is selected by roulette wheel selection based on the individual fitness value; The selected path encoding sequence is updated by crossover mutation to obtain the updated path population; The updated path population is used as the candidate path population for the next round. The continuous steps from fitness evaluation to crossover and mutation update are repeated until the optimal fitness value remains unchanged or the number of iterations reaches a preset number, thus obtaining the converged path population.
6. The intelligent deviation correction scheduling method for chip packaging process according to claim 1, characterized in that, The step of performing optimal path decoding and reconstruction based on the converged path population to obtain a set of correction and transfer paths includes: Based on the maximum completion time and total transport distance of each path individual in the convergence path population, the fitness of each path individual is evaluated to obtain a convergence fitness sequence. Based on the convergence fitness sequence, the path encoding sequence with the highest fitness value is selected from the convergence path population to obtain the optimal path encoding sequence; The optimal path encoding sequence is decoded to obtain the optimal transfer path; Extract the path segments corresponding to the conflict detection results from the initial set of transfer paths to obtain conflict path segments; The optimal transfer path is used to replace and reconstruct the conflicting path segments to obtain the set of corrective transfer paths.
7. The intelligent deviation correction scheduling method for chip packaging processes according to claim 5, characterized in that, The step of evaluating the fitness of the path encoding sequence based on the estimated path duration and transport distance corresponding to the path encoding sequence to obtain an individual fitness value includes: Extract the machine processing nodes and guide vehicle transport segments of each candidate transfer path from the path coding sequence to obtain the processing node sequence and the transport segment sequence. The estimated path duration is obtained by accumulating the machine processing time based on the processing node sequence; The transport distance is obtained by accumulating the distances of the transport segments according to the transport segment sequence; Using minimizing the maximum completion time and minimizing the total transport distance as evaluation objectives, a fitness mapping is performed on the estimated path duration and the transport distance to obtain the individual fitness value.
8. The intelligent deviation correction scheduling method for chip packaging processes according to claim 5, characterized in that, The step of performing crossover mutation update on the selected path encoding sequence to obtain the updated path population includes: The selected path encoding sequences are combined according to the pairing relationship to obtain path encoding sequence pairs; Partial mapping and cross-mapping are performed on the path encoding sequence pairs to obtain cross-path encoding sequences; Single-point mutation is performed on the machine allocation gene in the cross-path coding sequence to obtain the updated path coding sequence; The updated path encoding sequences are aggregated to obtain the updated path population.
9. An intelligent deviation correction scheduling system for a chip packaging process, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 8.
