A fast response scheduling method and system for semiconductor packaging and testing area based on experience learning

Abstract

The present invention relates to a semiconductor packaging and testing area fast response scheduling method and system based on experience learning. The method constructs a semiconductor packaging and testing area fast response scheduling system, transmits the bonding process lot data and machine or unit data to the bonding process scheduling module through the semiconductor packaging and testing area data module, and is used for the scheduling of the bonding process in the semiconductor packaging and testing area; uses a machine learning method to estimate the initial scheduling plan of the bonding process; uses an adaptive overlapping decomposition method to decompose it into lot scheduling sub-problems; uses BAHA to solve the scheduling sub-problems, thereby obtaining a complete bonding process scheduling plan; completes the global scheduling plan of semiconductor packaging and testing, and uses a GUI module to view the relevant operating data in the scheduling process. The method solves the problem of multiple constraints, diverse product types, and a large number of scheduling unit lots in the scheduling of semiconductor packaging and testing areas, which affect the delivery penalty time, reduces the delivery penalty of semiconductor packaging and testing production workshops, and realizes efficient workshop scheduling control.

Description

Technical Field

[0001] The present invention relates to a workshop scheduling control technology, and in particular to a semiconductor packaging and testing area fast response scheduling method and system based on experience learning. Background Art

[0002] As a key link in my country's semiconductor industry chain, semiconductor packaging and testing has significant economic value and strategic significance. In the field of semiconductor packaging and testing, there are many constraints such as flexible constraints on delivery penalty time and rigid constraints on machine modification time. At the same time, it faces the challenges of diversity of product types and thousands of lots to be scheduled. Some existing technical solutions, such as Chinese patents CN104576441A and CN102393687A, use simple heuristic rules for scheduling, but the scheduling effect is not good; while other technical solutions, such as Chinese patents CN117434902A and CN116151572A, use intelligent optimization algorithms, which have higher scheduling quality but affect practical applications due to long calculation time. The above technical defects limit the ability of semiconductor packaging and testing workshops to achieve efficient and high-quality scheduling. Therefore, exploring a method that can both ensure scheduling quality and effectively shorten the solution time is of great practical significance and academic value for improving the efficiency of semiconductor packaging and testing. Semiconductor packaging and testing involves multiple complex steps, including but not limited to wafer inspection, backside thinning, wafer dicing, chip mounting, wire bonding, plastic encapsulation, laser printing, rib cutting, electroplating, and final testing. It is worth noting that the wire bonding process is often considered to be the key bottleneck in packaging and testing because its processing speed has a decisive impact on the efficiency of the entire production process. In view of the current multiple constraints, the variety of product types, and the large number of scheduling lots, there is an urgent need to develop a new scheduling method that can quickly generate high-quality scheduling solutions to optimize the delivery penalty time of the semiconductor packaging and testing area. Summary of the invention

[0003] Aiming at the problem of multiple constraints, diverse product types and large number of scheduling units in the semiconductor packaging and testing area scheduling, which affect the delivery penalty time, a fast response scheduling method and system for the semiconductor packaging and testing area based on experience learning is proposed to reduce the delivery penalty time. This method uses machine learning methods to learn historical scheduling experience, estimate the scheduling plan of the bonding process, and decompose the estimated scheduling plan into multiple overlapping sub-scheduling problems. Then, the bidirectional artificial hummingbird algorithm (BAHA) is used to solve each sub-problem to complete the scheduling of the bonding process. Finally, global scheduling is achieved through the global scheduling module of the semiconductor packaging and testing area to ensure that high-quality scheduling plans can be quickly obtained. Therefore, it is of great theoretical and practical significance to conduct a holistic study on the scheduling of semiconductor packaging and testing production workshops.

[0004] The technical solution of the present invention is:

[0005] A semiconductor packaging and testing area fast response scheduling method based on experience learning, constructing a semiconductor packaging and testing area fast response scheduling system, the system includes a semiconductor packaging and testing area data module, a bonding process scheduling module, a semiconductor packaging and testing global scheduling module and a GUI module. The semiconductor packaging and testing area data module is used to store and transmit lot and machine or unit data; the bonding process scheduling module includes a learning submodule, a decomposition submodule and a scheduling submodule; the semiconductor packaging and testing area global scheduling module includes a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule; the GUI module is used to display a Gantt chart and operation data; the method transmits the bonding process lot data and machine or unit data to the bonding process scheduling module through the semiconductor packaging and testing area data module for scheduling the bonding process in the semiconductor packaging and testing area; in the bonding process scheduling The learning submodule of the module uses machine learning methods to estimate the initial scheduling plan of the bonding process; in the decomposition submodule of the bonding process scheduling module, based on the initial scheduling plan, the bonding process scheduling problem is decomposed into multiple lot scheduling subproblems with overlapping parts using an adaptive overlapping decomposition method; in the scheduling submodule of the bonding process scheduling module, BAHA is used to solve the scheduling subproblems and complete the solution of all subproblems, thereby completing the scheduling problem of the bonding process and obtaining a complete bonding process scheduling plan; then, the semiconductor packaging and testing area global scheduling module completes the semiconductor packaging and testing global scheduling plan through the scheduling plan of the bonding process and the packaging and testing area lot data and machine or unit data in the semiconductor packaging and testing area data module, and uses the GUI module to view the Gantt chart during the scheduling process and the operation data related to the lot and machine or unit.

[0006] Further, the method specifically includes the following steps:

[0007] Step 1: Extract the bonding process lot data and unit data from the semiconductor packaging and testing area data module and pass them to the bonding process scheduling module;

[0008] Step 2: In the bonding process scheduling module, the bonding process scheduling scheme is solved through the learning submodule, the decomposition submodule and the scheduling submodule, and output to the semiconductor packaging and testing area global scheduling module;

[0009] Step 2.1: In the learning submodule of the bonding process scheduling module, a machine learning method is used to extract the relationship between the lot's own attributes and global attributes and the scheduling priority from the bonding process experience library, and an estimated scheduling plan based on the lot priority order is generated according to the lot attribute data provided by the data module;

[0010] Step 2.2: In the decomposition submodule of the bonding process scheduling module, based on the estimated scheduling scheme generated by the learning submodule, an adaptive overlapping decomposition method is used to decompose the lot scheduling task in the complete bonding process scheduling problem into N w overlapping scheduling subproblems;

[0011] Step 2.3: In the scheduling submodule of the bonding process scheduling module, the bidirectional artificial hummingbird algorithm is applied to solve each scheduling subproblem decomposed in the decomposition submodule one by one; then, these scheduling subproblems are merged in a forward-to-backward order, and the scheduling results of the subsequent subproblems are used to update the overlapping parts with the previous subproblems, and finally a complete bonding process scheduling solution is obtained;

[0012] Step 3: In the semiconductor packaging and testing area global scheduling module, the semiconductor packaging and testing area global scheduling plan is completed using the bonding process scheduling plan obtained in the bonding process scheduling module and the semiconductor packaging and testing area lot data and machine or unit data in the semiconductor packaging and testing area data module, and output to the GUI module;

[0013] Step 4: In the GUI module, you can view the Gantt chart of the schedule and the operation data related to the lot and machine or unit during the schedule process.

[0014] Furthermore, the specific operations of step 2 are:

[0015] Step 2.1: In the learning submodule of the bonding process scheduling module, a machine learning method is used to extract the relationship between the lot's own attributes and global attributes and the scheduling priority from the bonding process experience library, and an estimated scheduling plan based on the lot priority order is generated according to the lot attribute data provided by the data module;

[0016] The specific operations of step 2.1 are:

[0017] Step 2.1.1: Extract the lot's own attributes and global attributes, and perform data preprocessing on the attribute values; input information into the learning submodule, the information includes multiple types of key data, and use Z-score standardization to preprocess the input data, as shown in formula (1);

[0018] Step 2.1.2: Use machine learning methods to learn the relationship between the lot's own attributes and global attributes and its scheduling priority;

[0019] The machine learning network uses machine learning methods to extract the features of each lot and its global attributes, and outputs the priority coefficient of a single lot; then the priority coefficients are arranged in descending order to form a priority list of the lot, which serves as the estimated optimal solution to assist the subsequent decomposition method; specifically, the structure of the machine learning model is: the input layer has multiple nodes; one or more hidden layers; the output layer has one node; this method shows the relationship between the lot-related attributes and the priority coefficient, as shown in formula (2); formula (3) represents the ReLU activation function;

[0020]

[0021] In formula (1), x’ , x , μ , σ Respectively represent the standardized data, original data, feature mean and feature standard deviation; in formula (2), , w 1. w 2. b 1. b 2 respectively represent the priority coefficient of lot, input layer weight matrix, hidden layer weight matrix, input layer bias vector and hidden layer bias vector, T represents the transpose of the matrix; in formula (3), z Represents the input data of the ReLU activation function;

[0022] Step 2.1.3: Output an estimated scheduling plan based on the relationship between the priority coefficients of the lots;

[0023] Step 2.2: In the decomposition submodule of the bonding process scheduling module, based on the estimated scheduling scheme generated by the learning submodule, an adaptive overlapping decomposition method is used to decompose the lot scheduling task in the complete bonding process scheduling problem into N w overlapping scheduling subproblems;

[0024] The specific operations of step 2.2 are:

[0025] Step 2.2.1: Calculate the parameters required for adaptive overlapping decomposition based on the scale of the scheduling problem, including the number of scheduling sub-problems after decomposition N w , calculate the sequence number of the left and right lots of each scheduling subproblem NL i and NR i ;

[0026] According to formula (4), the lot scheduling problem is decomposed into N w The number of overlapping lots is the same as the number of bonding machine units. According to formula (5) and formula (6), the serial numbers of the left and right lots of each scheduling subproblem are calculated. NL i and NR i ;

[0027]

[0028] In formula (4), formula (5), and formula (6), N w represents the number of scheduling subproblems decomposed, B is a fixed multiple, n , K are the number of lots and the number of bonding machine units respectively; N w -1 The number of slots in the scheduling subproblem is n / N w + K ; NL i and NR i and i The overlapping parts of adjacent sub-scheduling problems are K lot;

[0029] Step 2.2.2: Based on the estimated scheduling solution and the parameters required for adaptive overlap decomposition, the bonding process scheduling problem is decomposed into N w have K The scheduling subproblem of overlapping lots;

[0030] Step 2.3: In the scheduling submodule of the bonding process scheduling module, the bidirectional artificial hummingbird algorithm is applied to solve each scheduling subproblem decomposed in the decomposition submodule one by one; then, the optimal scheduling solutions of these scheduling subproblems are merged in a forward-to-backward order, and the scheduling solutions of the subsequent subproblems are used to update the overlapping parts with the previous subproblems, and finally a complete bonding process scheduling solution is obtained;

[0031] The specific operations of step 2.3 are:

[0032] Step 2.3.1: The overall number of iterations and population size are I total and P total , the number of iterations and the number of populations of the scheduling subproblem are calculated by formula (7) and formula (8) respectively: I S and P S ;

[0033] The number of iterations to solve the scheduling subproblem I S and population P S As shown in formula (7) and formula (8), where: I min and P min is the minimum value of the number of iterations and population of the scheduling subproblem; the overlap between the solved scheduling subproblem and the next problem, i.e. the final K The lot will be used as the lot in the new scheduling problem; in addition, the processing state of the unit of the solved part of the previous scheduling sub-problem is used as the initial state of the unit of the next scheduling problem; in the whole solution process, the main factors of the running time of different solution algorithms are deduced through the decoding process; the ratio of the running time with decomposition submodule and without decomposition submodule is preliminarily calculated, as shown in formula (9);

[0034]

[0035] In formula (7) and formula (8), I S and P S are the number of iterations and the number of iteration populations for solving the scheduling subproblem respectively; in formula (9), t and t no are the theoretical time estimates required to solve the schedule with and without decomposed submodules, respectively;

[0036] Step 2.3.2: Use BAHA to perform optimization search on the decomposed scheduling subproblems in the decomposed submodule;

[0037] The specific operations of step 2.3.2 are:

[0038] Step 2.3.2.1: Initialize the initial population through bidirectional initialization;

[0039] Bidirectional initialization refers to generating initial solutions through two methods: directional generation based on machine learning methods and heuristic rules, and random generation; heuristic rules include the shortest delivery time priority rule and the shortest delivery time slack time priority rule; using the above two methods to generate the same number of initial solutions;

[0040] Step 2.3.2.2: Determine the number of iterations. If the current number of iterations is less than I S , then two-way guided foraging or two-way territorial foraging will be performed randomly with equal probability, otherwise proceed to step 2.3.6;

[0041] In the process of two-way guided foraging, each hummingbird will select a new guide individual with the same probability in two ways: one is random selection, and the other is selection based on fitness value weight. There are three ways of two-way guided foraging: selecting one, multiple, K Perform POX crossover on lots to generate two new hummingbird individuals, and keep the hummingbird individual with the minimum delivery penalty.

[0042] In the two-way field foraging, all hummingbird individuals are selected and selected by fitness weights. P S hummingbirds, mixed and randomly selected P S hummingbird individuals are selected as individuals participating in the territory foraging process; during the territory foraging process, two, multiple, K The lots are randomly exchanged to generate a new hummingbird individual, and the hummingbird individual with the minimum delivery penalty is retained;

[0043] Step 2.3.2.3: Determine the number of iterations. If the current number of iterations is 0.1× I S multiples of, conduct enhanced two-way elite foraging, otherwise proceed to step 2.3.5;

[0044] In the bidirectional elite foraging, 2× N e hummingbird individuals conduct elite taboo searches, N e The values ​​are shown in formula (10); each selected hummingbird individual undergoes 2× PS / N e sub-territorial foraging;

[0045]

[0046] Step 2.3.2.4: Perform bidirectional migration for food and proceed to step 2.3.3;

[0047] In bidirectional migration foraging, 0.1× P S The number of temporary hummingbird individuals is shuffled and replaced with the 0.1× with the worst target value in the population P S Number of hummingbird individuals;

[0048] Step 2.3.2.5: Output the individual with the best fitness value, that is, the optimal solution for the scheduling subproblem;

[0049] Step 2.3.3: Merge the optimal scheduling solutions of these scheduling sub-problems in order from front to back, and use the scheduling results of subsequent sub-problems to update the overlapping parts with the previous sub-problems, and finally obtain a complete bonding process scheduling solution.

[0050] Furthermore, the specific operations of step 3 are:

[0051] Step 3.1: In the heuristic rule automatic scheduling submodule, the bonding process scheduling scheme obtained in the bonding process scheduling module and the semiconductor packaging and testing area lot data and machine or unit data in the semiconductor packaging and testing area data module are extracted, and the shortest delivery period priority rule and the delivery period slack time rule are used to complete the generation of the preliminary global scheduling scheme, and the scheme is transmitted to the expert experience fine-tuning scheduling submodule;

[0052] Step 3.2: In the expert experience fine-tuning scheduling submodule, there are two heuristic rule bases: lot selection rule base and equipment selection rule base. The scheduling specialist selects appropriate simple rules or designs compound rules from these two rule bases according to the current specific needs of the closed-loop testing area to adjust the preliminary global scheduling plan and finally generate an optimized global scheduling plan.

[0053] Furthermore, the machine learning methods include multi-submodule perceptron network MLP, enhanced multi-submodule perceptron network EMLP, convolutional neural network CNN, long short-term memory network LSTM, and Transformer.

[0054] Furthermore, in the learning submodule in step 2.1.1, the input information includes the processing time, delivery time, slack value, slack percentage of the lot and the statistical values ​​of these four input information, the weight of early completion, the weight of delayed completion, the serial number of the bonding wire type, the total number of this type of bonding wire type, the bonding frame type, the total number of this type of bonding frame type, the number of available machines and the total number of lots in the case, totaling 32 types of data; among which, the statistical values ​​include the maximum value, the minimum value, the average value, the variance, and the rank in the case.

[0055] Furthermore, the lot selection rule library includes the rule that the shorter the processing time, the higher the priority, the longer the processing time, the first-come-first-served rule, the earlier the delivery date, the higher the priority, the higher the equipment precision required by the lot, the lower the equipment precision required by the lot, and the higher the importance of the lot user; the equipment selection rule library includes the rule that the shortest processing time is the priority, the shortest machine change time is the priority, the highest equipment precision is the priority, the lowest equipment precision is the priority, the most order types that can be processed are the priority, and the least order types that can be processed are the priority; the design of composite rules is achieved by selecting multiple rules in the same library at the same time and assigning different priorities to each rule.

[0056] A semiconductor packaging and testing area fast response scheduling system based on experience learning, used to implement the semiconductor packaging and testing area fast response scheduling method based on experience learning as described above, the system includes four parts: semiconductor packaging and testing area data module, bonding process scheduling module, semiconductor packaging and testing global scheduling module and GUI module;

[0057] The semiconductor packaging and testing area data module is used to store and transmit lot and machine or group data;

[0058] The bonding process scheduling module is used to generate the scheduling plan for the bonding process. It consists of three parts: the learning submodule, the decomposition submodule, and the scheduling submodule. In the learning submodule, machine learning is used to estimate the scheduling plan. In the decomposition submodule, the large-scale bonding process problem is decomposed into multiple scheduling subproblems through the adaptive overlapping decomposition method. In the scheduling submodule, BAHA is used to solve a single scheduling problem.

[0059] The semiconductor packaging and testing area global scheduling module includes a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule, which are used to automatically generate a preliminary global scheduling plan and fine-tune the final global scheduling plan respectively;

[0060] The GUI module includes a Gantt chart display submodule and an operation data submodule.

[0061] Furthermore, the system also includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it implements a semiconductor packaging and testing area data module, a bonding process scheduling module including a learning submodule, a decomposition submodule and a scheduling submodule, a semiconductor packaging and testing area global scheduling module including a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule, and a GUI module.

[0062] The beneficial effects of the present invention are:

[0063] The present invention discloses a fast response scheduling method and system for a semiconductor packaging and testing area based on experience learning, which solves the problems of low solution quality and long solution time in the scheduling optimization of the semiconductor packaging and testing area, reduces the delivery penalty time of the bonding process, thereby reducing the delivery penalty of the semiconductor packaging and testing production workshop, and realizes efficient scheduling control of the semiconductor packaging and testing area, reducing the cost of the enterprise and improving the enterprise efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] FIG1 is a framework diagram of a semiconductor packaging and testing area fast response scheduling system based on experience learning according to the present invention;

[0065] FIG2 is a diagram of the complete scheduling process of the bonding process scheduling module of the present invention;

[0066] FIG3 is a method diagram of a process of estimating a scheduling solution based on experience learning in a learning submodule of the present invention;

[0067] FIG4 is a method diagram of overlapping scheduling windows based on scheduling environment in the decomposition submodule of the present invention;

[0068] FIG5 is a flow chart of a bidirectional artificial hummingbird algorithm in a scheduling submodule of the present invention. DETAILED DESCRIPTION

[0069] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0070] The present invention provides a semiconductor packaging and testing area fast response scheduling method and system based on experience learning, the system includes a semiconductor packaging and testing area data module, a bonding process scheduling module, a semiconductor packaging and testing global scheduling module and a GUI module. Figure 1 The semiconductor packaging and testing area data module is used to store and transmit lot and machine (unit) data; the bonding process scheduling module includes three parts: learning submodule, decomposition submodule and scheduling submodule; the semiconductor packaging and testing area global scheduling module includes two parts: heuristic rule automatic scheduling submodule and expert experience fine-tuning scheduling submodule; the GUI module is used to display the Gantt chart and operation data.

[0071] Among them, machine (unit): machine is a single machine, and unit is a collection of multiple machines. Machines exist in different processes, but in step 2, i.e. bonding process, the scheduling unit is the unit.

[0072] The present invention transmits the bonding process lot data and machine (unit) data to the bonding process scheduling module through the semiconductor packaging and testing area data module, which is used for the scheduling of the bonding process in the semiconductor packaging and testing area. In the learning submodule of the bonding process scheduling module, the machine learning method is used to estimate the initial scheduling plan of the bonding process; in the decomposition submodule of the bonding process scheduling module, the initial scheduling plan is decomposed based on the decomposition, and the bonding process scheduling problem is decomposed into multiple lot scheduling sub-problems with overlapping parts using the adaptive overlapping decomposition method; in the scheduling submodule of the bonding process scheduling module, BAHA is used to solve the scheduling sub-problems, and all sub-problems are solved, thereby completing the scheduling problem of the bonding process and obtaining a complete bonding process scheduling plan. Afterwards, the semiconductor packaging and testing area global scheduling module completes the semiconductor packaging and testing global scheduling plan through the scheduling plan of the bonding process and the packaging and testing area lot data and machine (unit) data in the semiconductor packaging and testing area data module, and uses the GUI module to view the Gantt chart and the operation data related to the lot and machine (unit) during the scheduling process. Compared with the scheduling system that uses conventional intelligent optimization algorithms for scheduling, this system generates scheduling plans for the packaging and testing area faster and can generate higher-quality scheduling plans, that is, scheduling plans with smaller delivery penalty time, thereby improving the production efficiency of the semiconductor packaging and testing area.

[0073] The method is implemented based on the system for scheduling process, and specifically includes the following steps:

[0074] Step 1: Extract the bonding process lot data and unit data from the semiconductor packaging and testing area data module and pass them to the bonding process scheduling module;

[0075] Step 2: In the bonding process scheduling module, the bonding process scheduling scheme is solved through the learning submodule, decomposition submodule and scheduling submodule, and output to the semiconductor packaging and testing area global scheduling module. The complete scheduling process of the bonding process scheduling module is as follows: Figure 2 As shown;

[0076] The specific operations of step 2 are:

[0077] Step 2.1: In the learning submodule of the bonding process scheduling module, a machine learning method is used to extract the relationship between the lot's own attributes and global attributes and the scheduling priority from the bonding process experience library, and based on the lot attribute data provided by the data module, an estimated scheduling plan based on the lot priority order is generated. The method of estimating the scheduling plan based on experience learning in the learning submodule is shown in the figure. Figure 3As shown. Figure 3 In it, x1-x32, Seq1-Seqn and n are the input data, processing order and the total number of lots in the scheduling problem, respectively.

[0078] The specific operations of step 2.1 are:

[0079] Step 2.1.1: Extract the lot's own attributes and global attributes, and perform data preprocessing on the attribute values;

[0080] In the learning submodule, the input information includes the processing time, delivery time, slack value, slack percentage of the lot and the statistical values ​​of these four input information (statistical values ​​include maximum value, minimum value, average value, variance, and ranking in the case), the weight of early completion, the weight of delayed completion, the serial number of the bonding wire type, the total number of this type of bonding wire type, the bonding frame type, the total number of this type of bonding frame type, the number of available machines, and the total number of lots in the case, totaling 32 types of data. Z-score standardization is used to preprocess the input data, as shown in formula (1).

[0081] Step 2.1.2: Use machine learning methods to learn the relationship between the lot's own attributes and global attributes and its scheduling priority. The machine learning methods used include but are not limited to multi-layer perceptron (MLP), enhanced multi-layer perceptron (EMLP), convolutional neural network (CNN), long short-term memory (LSTM), Transformer, etc.

[0082] The machine learning network uses machine learning methods to extract the features of each lot and its global attributes, and outputs the priority coefficient of a single lot. Then the priority coefficients are arranged in descending order to form a priority list of the lot, which serves as the estimated optimal solution to assist the subsequent decomposition method. Specifically, the structure of the machine learning model is: the input layer has 32 nodes; one or more hidden layers; and the output layer has one node. This method shows the relationship between lot-related attributes and priority coefficients, as shown in formula (2). Formula (3) represents the ReLU activation function.

[0083]

[0084] In formula (1), x’ , x , μ , σRespectively represent the standardized data, original data, feature mean and feature standard deviation. In formula (2), , w 1. w 2. b 1. b 2 respectively represent the priority coefficient of lot, input layer weight matrix, hidden layer weight matrix, input layer bias vector and hidden layer bias vector, T represents the transpose of the matrix. In formula (3), z Represents the input data of the ReLU activation function.

[0085] Step 2.1.3: Output an estimated scheduling plan based on the relationship between the priority coefficients of the lots;

[0086] According to the data of a certain packaging and testing factory in Shanghai, the processing situation of the bonding process scheduling process is designed. The number of lots is set to 100, the number of bonding machine units is set to 8, and the processing capacity of the bonding machine units is consistent. The relevant information of the lots in the case is shown in Table 1. Through Random (random generation), MLP, EMLP (adding dropout), CNN, LSTM, Transformer estimation, and using the BAHA algorithm to solve 5 different cases 10 times, the average of the initial and final scheduling solutions of different cases using different machine learning methods is recorded in Table 2.

[0087] Table 1 Lot related information

[0088]

[0089] Table 2 Initial and final mean values ​​of case studies of bonding process scheduling solutions estimated by different networks

[0090]

[0091] Table 2 Case initial and final mean values ​​of bonding process scheduling solutions estimated by different networks - continued

[0092]

[0093] Step 2.2: In the decomposition submodule of the bonding process scheduling module, based on the estimated scheduling plan generated by the learning submodule, an adaptive overlapping decomposition method is used to decompose the lot scheduling task in the complete bonding process scheduling problem into N w There are overlapping scheduling sub-problems. The adaptive overlapping scheduling windows based on the scheduling scale in the decomposition sub-module are as follows: Figure 4 shown.

[0094] The specific operations of step 2.2 are:

[0095] Step 2.2.1: Calculate the parameters required for adaptive overlapping decomposition based on the scale of the scheduling problem, including the number of scheduling sub-problems after decomposition N w , calculate the sequence number of the left and right lots of each scheduling subproblem NL i and NR i ;

[0096] According to formula (4), the lot scheduling problem is decomposed into N w The number of overlapping lots is the same as the number of bonding machine groups. According to formula (5) and formula (6), the sequence numbers of the left and right lots of each scheduling subproblem are calculated. NL i and NR i ;

[0097]

[0098] In formula (4), formula (5), and formula (6), N w represents the number of scheduling subproblems decomposed, B is a fixed multiple, n , K are the number of lots and the number of bonding machine units respectively; N w -1 The number of slots in the scheduling subproblem is n / N w + K ; NL i and NR i and i The overlapping parts of adjacent sub-scheduling problems are K lot;

[0099] Step 2.2.2: Based on the estimated scheduling solution and the parameters required for adaptive overlap decomposition, the bonding process scheduling problem is decomposed into N w have K The scheduling subproblem of overlapping lots;

[0100] Step 2.3: In the scheduling submodule of the bonding process scheduling module, the bidirectional artificial hummingbird algorithm is applied to solve each scheduling subproblem decomposed in the decomposition submodule one by one. The flow chart of the bidirectional artificial hummingbird algorithm in the scheduling submodule is as follows: Figure 5Then, these scheduling sub-problems are merged in order from front to back, and the scheduling results of the subsequent sub-problems are used to update the overlapping parts with the previous sub-problems, and finally a complete bonding process scheduling solution is obtained.

[0101] The specific operations of step 2.3 are:

[0102] Step 2.3.1: The overall number of iterations and population size are I total and P total , the number of iterations and the number of populations of the scheduling subproblem are calculated by formula (7) and formula (8) respectively: I S and P S， Use BAHA to refine your search.

[0103] In order to speed up the solution and prevent the solution quality from seriously declining. I total and P total If they are 2000 and 100 respectively, the number of iterations for solving the scheduling subproblem is I S and population P S As shown in formula (7) and formula (8). Where, I min and P min The minimum number of iterations and population size for the scheduling subproblem is 200 and 60 respectively. The overlap between the solved scheduling subproblem and the next problem is the final K The lots will be used as lots in the new scheduling problem. In addition, the processing state of the unit of the solved part of the previous scheduling subproblem is used as the initial state of the unit of the next scheduling problem. Since the decoding process is the most time-consuming part in the entire solution process, it can be inferred that the running time of different solution algorithms is mainly positively correlated with the number of lots in the scheduling problem, the number of decomposed parts, the number of populations, and the number of iterations. Therefore, we can preliminarily calculate the ratio of the running time with decomposition submodules and without decomposition submodules, as shown in formula (9).

[0104]

[0105] In formula (7) and formula (8), I S and P S are the number of iterations and the number of iteration populations for solving the scheduling subproblem. In formula (9), t and tno Theoretical time estimates required to solve the schedule with and without decomposed submodules, respectively.

[0106] Table 3 Comparison of computational complexity before and after decomposition

[0107]

[0108] Step 2.3.2: Use BAHA to perform optimization search on the decomposed scheduling subproblems in the decomposed submodule.

[0109] The specific operations of step 2.3.2 are:

[0110] Step 2.3.2.1: Initialize the initial population through bidirectional initialization.

[0111] Bidirectional initialization refers to generating initial solutions through two methods: directional generation based on machine learning methods and heuristic rules, and random generation. Machine learning methods include MLP, EMLP, CNN, and LSTM. Heuristic rules include the shortest delivery time priority rule and the shortest delivery time slack time priority rule (delivery time minus processing time). The same number of initial solutions (hummingbird individuals) are generated using each of the above seven methods.

[0112] Step 2.3.2.2: Determine the number of iterations. If the current number of iterations is less than I S , then two-way guided foraging or two-way territorial foraging will be performed randomly with equal probability, otherwise proceed to step 2.3.6;

[0113] In the process of two-way guided foraging, each hummingbird will select a new guide individual with the same probability in two ways: one is random selection, and the other is selection based on fitness value weight (that is, the reciprocal of fitness value as the proportion of the total is used as the probability of being selected). There are three ways of two-way guided foraging: selecting one, multiple, K The lots are crossovered by POX to generate two new hummingbird individuals, and the hummingbird individual with the minimum delivery penalty is retained.

[0114] In the two-way field foraging, all hummingbird individuals are selected and selected by fitness weights. P S hummingbirds, mixed and randomly selected P S hummingbird individuals are selected as individuals participating in the field foraging process. K The lots are randomly exchanged to generate a new hummingbird individual, and the hummingbird individual with the minimum delivery penalty is retained.

[0115] Step 2.3.2.3: Determine the number of iterations. If the current number of iterations is 0.1× I S If the number of times is greater than 1, perform enhanced two-way elite foraging, otherwise proceed to step 2.3.5.

[0116] In the bidirectional elite foraging, 2× N e hummingbird individuals conduct elite taboo searches, N e The value is shown in formula (10). Each selected hummingbird individual undergoes 2× P S / N e Sub-territory foraging.

[0117]

[0118] Step 2.3.2.4: Perform bidirectional migration for food and proceed to step 2.3.3;

[0119] In bidirectional migration foraging, 0.1× P S The number of temporary hummingbird individuals is shuffled and replaced with the 0.1× with the worst target value in the population P S Number of hummingbird individuals.

[0120] Step 2.3.2.5: Output the individual with the best fitness value, that is, the optimal solution for the scheduling subproblem;

[0121] Step 2.3.3: Merge the optimal scheduling solutions of these scheduling sub-problems in order from front to back, and use the scheduling results of the subsequent sub-problems to update the overlapping parts with the previous sub-problems, and finally obtain a complete bonding process scheduling solution;

[0122] Step 3: In the semiconductor packaging and testing area global scheduling module, the semiconductor packaging and testing area global scheduling plan is completed using the bonding process scheduling plan obtained in the bonding process scheduling module and the semiconductor packaging and testing area lot data and machine (unit) data in the semiconductor packaging and testing area data module, and output to the GUI module;

[0123] The specific operations of step 3 are:

[0124] Step 3.1: In the heuristic rule automatic scheduling submodule, the bonding process scheduling scheme obtained in the bonding process scheduling module and the semiconductor packaging and testing area lot data and machine (unit) data in the semiconductor packaging and testing area data module are extracted, and the shortest delivery period priority rule and delivery period slack time rule are used to complete the generation of the preliminary global scheduling scheme, and the scheme is transmitted to the expert experience fine-tuning scheduling submodule;

[0125] Step 3.2: In the expert experience fine-tuning scheduling submodule, there are two heuristic rule bases: lot selection rule base and equipment selection rule base. The scheduling specialist selects appropriate simple rules or designs compound rules from these two rule bases according to the current specific needs of the closed-test area to adjust the preliminary global scheduling plan and finally generate an optimized global scheduling plan.

[0126] Among them, the lot selection rule base includes the rule that the shorter the processing time, the higher the priority, the longer the processing time, the first-come-first-served rule, the earlier the delivery date, the higher the equipment precision required by the lot, the lower the equipment precision required by the lot, and the higher the importance of the lot user; the equipment selection rule base includes the rule that the shortest processing time is prioritized, the shortest machine change time is prioritized, the highest equipment precision is prioritized, the lowest equipment precision is prioritized, the most order types that can be processed are prioritized, and the least order types that can be processed are prioritized. The design of composite rules is achieved by selecting multiple rules in the same library at the same time and assigning different priorities to each rule.

[0127] Step 4: In the GUI module, you can view the Gantt chart of the schedule and the operation data related to the lot and machine (unit) during the schedule process.

[0128] A semiconductor packaging and testing area fast response scheduling system based on experience learning, including a semiconductor packaging and testing area data module, a bonding process scheduling module, a semiconductor packaging and testing global scheduling module and a GUI module; used to complete a semiconductor packaging and testing area fast response scheduling method;

[0129] The semiconductor packaging and testing area data module is used to store and transmit lot and machine (unit) data;

[0130] The bonding process scheduling module is used to generate the scheduling plan for the bonding process. It consists of three parts: the learning submodule, the decomposition submodule, and the scheduling submodule. In the learning submodule, machine learning is used to estimate the scheduling plan; in the decomposition submodule, the large-scale bonding process problem is decomposed into multiple scheduling subproblems through the adaptive overlapping decomposition method; in the scheduling submodule, BAHA is used to solve a single scheduling problem.

[0131] The semiconductor packaging and testing area global scheduling module includes a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule, which are used to automatically generate a preliminary global scheduling plan and fine-tune the final global scheduling plan respectively;

[0132] The GUI module is used to display Gantt charts and operating data.

[0133] The system also includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements a semiconductor packaging and testing area data module, a bonding process scheduling module including a learning submodule, a decomposition submodule and a scheduling submodule, a semiconductor packaging and testing area global scheduling module including a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule, and a GUI module.

[0134] The above-mentioned embodiment only expresses one implementation mode of the present invention, and its description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A fast response scheduling method for semiconductor packaging and testing area based on experience learning, characterized in that: Construct a fast response scheduling system for semiconductor packaging and testing area, which includes semiconductor packaging and testing area data module, bonding process scheduling module, semiconductor packaging and testing global scheduling module and GUI module. The semiconductor packaging and testing area data module is used to store and transmit lot and machine or unit data; the bonding process scheduling module includes learning submodule, decomposition submodule and scheduling submodule; the semiconductor packaging and testing area global scheduling module includes heuristic rule automatic scheduling submodule and expert experience fine-tuning scheduling submodule; The GUI module is used to display the Gantt chart and operation data; specifically includes the following steps: Step 1: Extract the bonding process lot data and machine or unit data from the semiconductor packaging and testing area data module, and pass them to the bonding process scheduling module; Step 2: In the bonding process scheduling module, solve the bonding process scheduling plan through the learning submodule, decomposition submodule and scheduling submodule, and output it to the semiconductor packaging and testing area global scheduling module; Step 2.1: In the learning submodule of the bonding process scheduling module, use machine learning methods to extract the relationship between the own attributes and global attributes of the learning lot and the scheduling priority from the experience library of the bonding process, and generate an estimated scheduling plan based on the lot priority order according to the lot attribute data provided by the data module; Step 2.2: In the decomposition submodule of the bonding process scheduling module, based on the estimated scheduling plan generated by the learning submodule, an adaptive overlapping decomposition method is used to decompose the lot scheduling task in the complete bonding process scheduling problem according to the scale of the bonding process, that is, the number of lots and the number of bonding units. N w overlapping scheduling sub-problems; Step 2.3: In the scheduling sub-module of the bonding process scheduling module, each scheduling sub-problem decomposed in the decomposition sub-module is solved one by one by applying the bidirectional artificial hummingbird algorithm BAHA; then, these scheduling sub-problems are merged in order from front to back, and the scheduling results of the subsequent sub-problems are used to update the overlapping parts with the previous sub-problems, and finally a complete bonding process scheduling plan is obtained; Step 3: In the global scheduling module of the semiconductor packaging and testing area, the bonding process scheduling plan obtained in the bonding process scheduling module and the semiconductor packaging and testing area lot data and machine or unit data in the semiconductor packaging and testing area data module are used to complete the semiconductor packaging and testing area global scheduling plan, and output it to the GUI module; Step 4: In the GUI module, the Gantt chart of the scheduling and the operation data related to the lot and machine or unit during the scheduling process can be viewed.

2. The semiconductor packaging and testing area fast response scheduling method based on experience learning according to claim 1 is characterized in that: The specific operations of step 2 are: Step 2.1: In the learning submodule of the bonding process scheduling module, a machine learning method is used to extract the relationship between the lot's own attributes and global attributes and the scheduling priority from the bonding process experience library, and an estimated scheduling plan based on the lot priority order is generated according to the lot attribute data provided by the data module; The specific operations of step 2.1 are: Step 2.1.1: Extract the lot's own attributes and global attributes, and perform data preprocessing on the attribute values; input information into the learning submodule, the information includes multiple types of key data, and use Z-score standardization to preprocess the input data, as shown in formula (1); Step 2.1.2: Use machine learning methods to learn the relationship between the lot's own attributes and global attributes and its scheduling priority; The machine learning network uses machine learning methods to extract the features of each lot and its global attributes, and outputs the priority coefficient of a single lot; then it arranges them in descending order according to the priority coefficient to form a priority list of the lot, which serves as the estimated optimal solution to assist the subsequent decomposition method; the structure of the machine learning model is: the input layer has multiple nodes; one or more hidden layers; the output layer has one node; the relationship between the lot-related attributes and the priority coefficient is shown, as shown in formula (2); formula (3) represents the ReLU activation function; In formula (1), x’ , x , μ , σ Respectively represent the standardized data, original data, feature mean and feature standard deviation; in formula (2), , w 1. w 2. b 1. b 2 respectively represent the priority coefficient of lot, input layer weight matrix, hidden layer weight matrix, input layer bias vector and hidden layer bias vector, T Represents the transpose of a matrix; In formula (3), z Represents the input data of the ReLU activation function; Step 2.1.3: Output an estimated scheduling plan based on the relationship between the priority coefficients of the lots; Step 2.2: In the decomposition submodule of the bonding process scheduling module, based on the estimated scheduling scheme generated by the learning submodule, an adaptive overlapping decomposition method is used to decompose the lot scheduling task in the complete bonding process scheduling problem into N w overlapping scheduling subproblems; The specific operations of step 2.2 are: Step 2.2.1: Calculate the parameters required for adaptive overlapping decomposition based on the scale of the scheduling problem, including the number of scheduling sub-problems after decomposition N w , calculate the sequence number of the left and right lots of each scheduling subproblem NL i and NR i ; According to formula (4), the lot scheduling problem is decomposed into N w There are scheduling sub-problems, and the number of lots in the overlapping part is consistent with the number of bonding machine groups; According to formula (5) and formula (6), the sequence numbers of the left and right lots of each scheduling subproblem are calculated. NL i and NR i ; In formula (4), formula (5), and formula (6), N w represents the number of scheduling subproblems decomposed, B is a fixed multiple, n , K are the number of lots and the number of bonding machine units respectively; N w -1 The number of slots in the scheduling subproblem is n / N w + K ; NL i and NR i and i The overlapping parts of adjacent sub-scheduling problems are K lot; Step 2.2.2: Based on the estimated scheduling solution and the parameters required for adaptive overlap decomposition, the bonding process scheduling problem is decomposed into N w have K The scheduling subproblem of overlapping lots; Step 2.3: In the scheduling submodule of the bonding process scheduling module, the bidirectional artificial hummingbird algorithm is applied to solve each scheduling subproblem decomposed in the decomposition submodule one by one; then, the optimal scheduling solutions of these scheduling subproblems are merged in a forward-to-backward order, and the scheduling solutions of the subsequent subproblems are used to update the overlapping parts with the previous subproblems, and finally a complete bonding process scheduling solution is obtained; The specific operations of step 2.3 are: Step 2.3.1: The overall number of iterations and population size are I total and P total , the number of iterations and the number of populations of the scheduling subproblem are calculated by formula (7) and formula (8) respectively: I S and P S ; The number of iterations to solve the scheduling subproblem I S and population P S As shown in formula (7) and formula (8), where: I min and P min is the minimum value of the number of iterations and population of the scheduling subproblem; the overlap between the solved scheduling subproblem and the next problem, i.e. the final K The lot will be used as the lot in the new scheduling problem; in addition, the processing state of the unit of the solved part of the previous scheduling sub-problem is used as the initial state of the unit of the next scheduling problem; in the whole solution process, the main factors of the running time of different solution algorithms are deduced through the decoding process; the ratio of the running time with decomposition submodule and without decomposition submodule is preliminarily calculated, as shown in formula (9); In formula (7) and formula (8), I S and P S are the number of iterations and the number of iteration populations for solving the scheduling subproblem respectively; in formula (9), t and t no are the theoretical time estimates required to solve the schedule with and without decomposed submodules, respectively; Step 2.3.2: Use BAHA to perform optimization search on the decomposed scheduling subproblems in the decomposed submodule; The specific operations of step 2.3.2 are: Step 2.3.2.1: Initialize the initial population through bidirectional initialization; Bidirectional initialization refers to generating initial solutions through two methods: directional generation based on machine learning methods and heuristic rules, and random generation; heuristic rules include the shortest delivery time priority rule and the shortest delivery time slack time priority rule; using the above two methods to generate the same number of initial solutions; Step 2.3.2.2: Determine the number of iterations. If the current number of iterations is less than I S , then two-way guided foraging or two-way territorial foraging will be performed randomly with equal probability, otherwise proceed to step 2.3.6; In the process of two-way guided foraging, each hummingbird will select a new guide individual with the same probability in two ways: one is random selection, and the other is selection based on fitness value weight. There are three ways of two-way guided foraging: selecting one, multiple, K Perform POX crossover on lots to generate two new hummingbird individuals, and keep the hummingbird individual with the minimum delivery penalty. In the two-way field foraging, all hummingbird individuals are selected and selected by fitness weights. P S hummingbirds, mixed and randomly selected P S hummingbird individuals are selected as individuals participating in the territory foraging process; during the territory foraging process, two, multiple, K The lots are randomly exchanged to generate a new hummingbird individual, and the hummingbird individual with the minimum delivery penalty is retained; Step 2.3.2.3: Determine the number of iterations. If the current number of iterations is 0.1× I S multiples of, conduct enhanced two-way elite foraging, otherwise proceed to step 2.3.5; In the bidirectional elite foraging, 2× N e hummingbird individuals conduct elite taboo searches, N e The values ​​are shown in formula (10); each selected hummingbird individual undergoes 2× P S / N e sub-territorial foraging; Step 2.3.2.4: Perform bidirectional migration for food and proceed to step 2.3.3; In bidirectional migration foraging, 0.1× P S The number of temporary hummingbird individuals is shuffled and replaced with the 0.1× with the worst target value in the population P S Number of hummingbird individuals; Step 2.3.2.5: Output the individual with the best fitness value, that is, the optimal solution for the scheduling subproblem; Step 2.3.3: Merge the optimal scheduling solutions of these scheduling sub-problems in order from front to back, and use the scheduling results of subsequent sub-problems to update the overlapping parts with the previous sub-problems, and finally obtain a complete bonding process scheduling solution.

3. The semiconductor packaging and testing area fast response scheduling method based on experience learning according to claim 1 is characterized in that: The specific operations of step 3 are: Step 3.1: In the heuristic rule automatic scheduling submodule, the bonding process scheduling scheme obtained in the bonding process scheduling module and the semiconductor packaging and testing area lot data and machine or unit data in the semiconductor packaging and testing area data module are extracted, and the shortest delivery period priority rule and the delivery period slack time rule are used to complete the generation of the preliminary global scheduling scheme, and the scheme is transmitted to the expert experience fine-tuning scheduling submodule; Step 3.2: In the expert experience fine-tuning scheduling submodule, there are two heuristic rule bases: lot selection rule base and equipment selection rule base. The scheduling specialist selects appropriate simple rules or designs compound rules from these two rule bases according to the current specific needs of the closed-loop testing area to adjust the preliminary global scheduling plan and finally generate an optimized global scheduling plan.

4. The semiconductor packaging and testing area fast response scheduling method based on experience learning according to claim 1 is characterized in that: Machine learning methods include multi-module perceptron network MLP, enhanced multi-module perceptron network EMLP, convolutional neural network CNN, long short-term memory network LSTM, and Transformer.

5. The semiconductor packaging and testing area fast response scheduling method based on experience learning according to claim 2 is characterized in that: In the learning submodule in step 2.1.1, the input information includes the processing time, delivery time, slack value, slack percentage of the lot and the statistical values ​​of these four input information, the weight of early completion, the weight of delayed completion, the serial number of the bonding wire type, the total number of this type of bonding wire type, the bonding frame type, the total number of this type of bonding frame type, the number of available machines and the total number of lots in the case, totaling 32 types of data; among which, the statistical values ​​include the maximum value, minimum value, average value, variance, and rank in the case.

6. The semiconductor packaging and testing area fast response scheduling method based on experience learning according to claim 3 is characterized in that: The lot selection rule library includes the rule that the shorter the processing time, the higher the priority; the longer the processing time, the higher the priority; the first-come-first-served rule; the earlier the delivery date, the higher the priority; the higher the equipment precision required by the lot, the higher the priority; the lower the equipment precision required by the lot, the higher the priority; and the higher the importance of the lot user, the higher the priority; the equipment selection rule library includes the rule that the shortest processing time is the priority, the shortest machine change time is the priority, the highest equipment precision is the priority, the lowest equipment precision is the priority, the priority rule that the most order types can be processed is the priority, and the priority rule that the least order types can be processed is the priority; the design of composite rules is achieved by selecting multiple rules in the same library at the same time and assigning different priorities to each rule.

7. A fast response scheduling system for semiconductor packaging and testing area based on experience learning, characterized in that: Used to implement the semiconductor packaging and testing area fast response scheduling method based on experience learning as described in any one of claims 1-6, the system includes four parts: semiconductor packaging and testing area data module, bonding process scheduling module, semiconductor packaging and testing global scheduling module and GUI module; The semiconductor packaging and testing area data module is used to store and transmit lot and machine or group data; The bonding process scheduling module is used to generate a scheduling plan for the bonding process, which consists of three parts: a learning submodule, a decomposition submodule, and a scheduling submodule; In the learning submodule, machine learning is used to estimate the scheduling plan; In the decomposition submodule, the large-scale bonding process problem is decomposed into multiple scheduling subproblems through the adaptive overlapping decomposition method; in the scheduling submodule, BAHA is used to solve a single scheduling problem; The semiconductor packaging and testing area global scheduling module includes a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule, which are used to automatically generate a preliminary global scheduling plan and fine-tune the final global scheduling plan respectively; The GUI module includes a Gantt chart display submodule and an operation data submodule.

8. The semiconductor packaging and testing area fast response scheduling system based on experience learning according to claim 7 is characterized in that: The system also includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it implements a semiconductor packaging and testing area data module, a bonding process scheduling module including a learning submodule, a decomposition submodule and a scheduling submodule, a semiconductor packaging and testing area global scheduling module including a heuristic rule automatic scheduling submodule and an expert experience fine-tuning scheduling submodule, and a GUI module.

Images