Semiconductor packaging and bonding process fast response scheduling method and system based on super heuristic algorithm

Abstract

The application relates to a kind of semiconductor packaging workshop bonding process fast response scheduling method and system based on super heuristic algorithm, method includes establishing semiconductor packaging bonding process scheduling optimization model;Through the order splitting of large batch order by batch strategy;Through super heuristic algorithm adaptive generation scheduling scheme, wherein reinforcement learning algorithm is used as senior selection strategy, heuristic operator is used as the method of bottom problem domain;Output optimal scheduling scheme, arrange production.The scheduling method and system of the application, according to the characteristics of semiconductor packaging bonding process scheduling problem, design simple heuristic method, to build LLH method set, so as to realize the optimization solution of semiconductor packaging bonding process scheduling problem, both retains the good global optimization performance of meta heuristic algorithm, avoids the uncertainty caused by the adjustment of algorithm parameters in meta heuristic algorithm by artificial experience, can effectively improve the efficiency of algorithm design and bonding process scheduling efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of workshop operation scheduling, in particular to a semiconductor packaging and testing bonding process fast response scheduling method based on a hyper-heuristic algorithm. BACKGROUND

[0002] The semiconductor packaging and testing process flow mainly includes scribing, wafer mounting, bonding, plastic packaging, deburring, plating, printing, cutting, forming, appearance inspection, finished product testing, packaging and delivery, etc. Among them, the wire bonding process has a wide application in many electronic products and is a key technology of wafer manufacturing system. Research on efficient bonding area scheduling method is of great significance to improve the performance of wafer manufacturing system, improve enterprise efficiency and enhance the speed of enterprise response to market. In the wafer manufacturing system, the wire bonding processing area includes several equipment units, each equipment unit includes multiple wire bonding machines, and a magazine composed of several hundred wafers is taken as the minimum production unit to enter the equipment unit for wire bonding. In actual production, the wire bonding process is organized in a face-to-order mode (MTO), and due to the large scale and variety of orders, an efficient scheduling method must be adopted.

[0003] In the field of production scheduling, the production scheduling of the semiconductor packaging and testing bonding process is usually summarized as a large-scale (PMSP) parallel machine scheduling problem, which is a classic NP-hard problem. For parallel machine scheduling problems, existing research usually adopts simple rules or heuristic algorithms.

[0004] The inventor points out the problems of the above algorithms: first, although simple rules and heuristic algorithms have certain advantages in solving problems, their performance still has room for improvement. Rule-based scheduling algorithms have the advantages of stable performance and easy implementation, but for large-scale bonding process production scheduling problems, they perform poorly. Heuristic algorithms usually perform well in solving complex problems, they can find near-optimal solutions within a reasonable time without exhausting all possible combinations. However, heuristic algorithms usually cannot guarantee global optimal solutions, but only provide local optimal solutions or near-optimal solutions. This means that in some cases they may not find the best solution, and in practical applications they have poor adaptability due to changes in scenarios. In addition, the performance of heuristic algorithms is highly dependent on the selected heuristic information or rules, and different heuristic information may lead to different results, so careful selection or adjustment of these information is required. SUMMARY

[0005] To solve the technical problems existing in the prior art, the present application provides a semiconductor packaging and testing bonding process fast response scheduling method based on a hyper-heuristic algorithm, which can improve the scheduling efficiency of the bonding process.

[0006] To achieve the above object, the technical solution of the present application is as follows:

[0007] A semiconductor packaging and testing bonding process fast response scheduling method based on a hyper-heuristic algorithm, comprising the following steps:

[0008] Step one, establish a semiconductor packaging and testing bonding process scheduling optimization model;

[0009] Step two, screen out large batch orders, and split the orders by batch strategy;

[0010] Step three, generate a scheduling scheme by a hyper-heuristic algorithm, use the dqn deep reinforcement learning algorithm as a high-level selection strategy, and use the heuristic operator as a method for the underlying problem domain;

[0011] Step four, output the optimal bonding process scheduling scheme, including the processing priority ranking scheme of the orders to be processed and the selection priority ranking scheme of the processing equipment units, and arrange production.

[0012] The semiconductor packaging and testing bonding process scheduling optimization model established in step one includes symbol definition, objective function, constraint condition and decision variable.

[0013] The batch strategy in step two is to use the minimum batch strategy to split the orders whose magazine workpiece quantity is greater than the maximum batch constraint until the batch interval is met.

[0014] As a preferred technical solution, step three includes: generating a population of N groups of individuals, randomly selecting a feasible solution group P and calculating the population fitness f; initialization; operating the last action, generating individuals and corresponding fitness values by calculation, determining the immediate reward according to the fitness value, at this time, the current state is regarded as "next state", then determine the state to which the current state belongs, and calculate the value of the new state; solution acceptance reservation; determine the reward and new state; judge whether the algorithm is ended, if the termination condition is met, the algorithm is ended, and the optimal solution is output, otherwise, update the target network, send the new state into the network and obtain the action, return to the initialization, operate the last action and the next step.

[0015] As a preferred technical solution, in step three, initialization is to randomly select a group of feasible solutions from the feasible solution group, and set this group of feasible solutions as p i , the fitness value is f i , P B is the optimal solution individual, F B is the fitness value of the optimal solution, P B is equal to p i , F B is equal to f i , the state is set to 0, and a random action in the action space is selected.

[0016] As a preferred technical solution, in step three, the acceptance reservation of the solution is to compare the fitness value of the new solution with the fitness value of the old solution, if the fitness value of the new solution is greater than the fitness value of the old solution, the new solution and the fitness value thereof are saved, and the state is updated to the new solution; if the fitness value of the new solution is less than the fitness value of the old solution, the simulated annealing algorithm is considered to be used for reevaluation; if the calculated probability is greater than a random value, the new solution is also saved and the state is updated, otherwise, the solution is abandoned.

[0017] As a preferred technical solution, the system adopting the semiconductor packaging bonding process fast response scheduling method based on the super heuristic algorithm comprises a data interaction module, a data preprocessing module and a scheduling scheme generation and optimization module, the data interaction module comprises a scheduling information acquisition and import submodule and a scheduling result output submodule, the scheduling information acquisition and import submodule receives the to-be-processed order data information and equipment unit data information input by a user and outputs the same to the data preprocessing module; and finally receives the scheduling result and outputs the scheduling result; the data preprocessing module comprises an order splitting submodule and a split batch order output submodule, the order splitting submodule is used for splitting the imported to-be-processed order, adopts a batch splitting strategy, outputs the split batch order to the split batch order output submodule after batch splitting, and outputs the order data after order splitting to the scheduling scheme generation and optimization module by the split batch order output submodule; the scheduling scheme generation and optimization module is used for receiving the input order data and scheduling model, generating a scheduling scheme and optimizing the scheduling scheme by the super heuristic algorithm, and finally outputting the scheduling result to the data interaction module.

[0018] As a preferred technical solution, the system adopting the semiconductor packaging bonding process fast response scheduling method based on the super heuristic algorithm further comprises a memory and a processor.

[0019] Compared with the prior art, the application has the following beneficial effects:

[0020] The semiconductor packaging bonding process fast response scheduling method based on the super heuristic algorithm has the following beneficial effects: the advanced heuristic strategy adopts the DQN deep reinforcement learning algorithm, the low-level heuristic candidate pool is composed of six simple and efficient heuristic methods, the super heuristic algorithm can always provide a satisfactory solution within a calculable time for different sizes of semiconductor packaging bonding process orders, the problem that the existing method is difficult to solve different order sizes is solved, and the scheduling efficiency of the bonding process is improved. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the semiconductor packaging bonding process fast response scheduling method based on the super heuristic algorithm of the application;

[0022] Figure 2is a framework diagram of the super-heuristic algorithm in the scheduling method of the present application;

[0023] Figure 3 is a scheduling result comparison diagram of an embodiment of the present application;

[0024] Figure 4 is a scheduling result line chart of an embodiment of the present application;

[0025] Figure 5 is a scheduling result comparison diagram of an embodiment of the present application;

[0026] Figure 6 is a scheduling time comparison diagram of an embodiment of the present application;

[0027] Figure 7 is a structural schematic diagram of the scheduling system of the present application. DETAILED DESCRIPTION

[0028] The technical solutions of the present application are further described below in combination with specific embodiments:

[0029] As shown in Figure 1 , a semiconductor packaging and testing bonding process fast response scheduling method based on a super-heuristic algorithm includes the following steps:

[0030] Step 1: Establish a semiconductor packaging and testing bonding process scheduling optimization model;

[0031] Specifically, the corresponding scheduling optimization model is set as follows:

[0032] The prerequisite for the bonding process scheduling is:

[0033] 1) All equipment units and order tasks are in a processable state at 0 time;

[0034] 2) There is only one wafer product in each batch order;

[0035] 3) The job cannot be interrupted during processing, and only one sub-batch can be processed at the same time on a machine;

[0036] 4) All orders only need one process to complete;

[0037] 5) The processing time of sub-batch orders composed of the same wafer product is different on different equipment units;

[0038] Define related symbols, which are specifically defined as follows:

[0039] O is a set of wafer orders to be processed, O = {O1, O2, …, O n}; n is the total number of orders; M is a set of bonding machine devices, M = {M1, M2, …, M m}; Oij is the jthsub-order of the ithbatch order; C max is the maximum completion time, B max is the maximum number of sub-orders of each batch order, n is the total number of orders, m is the total number of devices, P i,Mj is the processing time of order i on device M j , X 1b , i , j is the assignment variable of sub-order b of order i on device m j , if sub-order b of order i is assigned to device M j , it is 1, otherwise it is 0; B bi is the number of clips of sub-order O bi , B num is the total number of all sub-orders after batching, B max is the maximum number of sub-orders of each batch order, B min is the minimum number of sub-orders of each batch order, M o is the total number of molds; the optimization objective is to minimize the maximum completion time, and the objective function is as follows:

[0040] min(C max ) (1)

[0041] The constraints of the optimization objective are as follows:

[0042] The completion time of each device unit is not greater than the maximum completion time:

[0043]

[0044] The total quantity of each order after batching remains unchanged:

[0045]

[0046] Each sub-order of each order can only be processed on one device unit:

[0047]

[0048] All sub-orders after batching need to be processed:

[0049]

[0050] The total number of sub-orders of all batch orders after batching:

[0051] B num = KB max (6)

[0052] The batch size of each sub-order of each order should meet the maximum batch size constraint and the minimum batch size constraint:

[0053] B min ≤B bi ≤B max (7)

[0054] The maximum number of batches of workpieces is not greater than the corresponding maximum number of molds:

[0055] B max ≤M o (8)

[0056] A 0,1 variable indicating whether the bth sub-batch order in the i th order is processed on the equipment unit M j

[0057] x 1bij , x 2bij ∈{0,1} (9)

[0058] Step two, screening large batch orders, and splitting large batch orders through batch strategy;

[0059] The batch strategy is to use the minimum batch strategy to split the order whose number of clips is greater than the maximum batch constraint until the batch interval is met.

[0060] Step three, self-adaptive generation of scheduling scheme through hyper-heuristic algorithm, using dqn deep reinforcement learning algorithm as high-level selection strategy and heuristic operator as method of bottom problem domain;

[0061] Specifically, the state information: the purpose of state design is to reflect the influence on the environment after the action is executed, and to predict the selection of the next action. The state is represented as the change degree of the optimal fitness value of the previous and next generations and the proportion of the number of high-quality fitness values in the newly generated solutions, wherein the optimal fitness value of the current generation is fit, the optimal fitness value of the next generation is fit', and the change degree of the optimal fitness value of the previous and next generations is represented as state=(fit'-fit) / fit.

[0062] Action space: in this method, the action space is an integer (1, 2, 3, …, N A ), wherein each integer represents a reference to each low-level heuristic operator, and selecting which integer is equivalent to selecting which low-level heuristic operator;

[0063] Reward function: the reward function is used to represent the immediate return value. The objective function of this problem is to minimize the maximum completion time. If the maximum completion time of the current solution is less than that of the previous generation, the current solution is better, reward=1; if the solution is not improved, reward=0, and if the quality decreases, reward=-1.

[0064] The specific method of the low-level heuristic operation is as follows:​

[0065] LLH1: Randomly select a bit in the order code string, change it to another random device unit corresponding to the order selectable device unit;

[0066] LLH2: Randomly select two bits in the device unit code string, exchange them with each other;

[0067] LLH3: Randomly select two bits in the order code string, exchange them with each other;

[0068] LLH4: Randomly select two bits in the order code string, reverse the order of the string substring between the two bits to obtain a new order code sequence;

[0069] LLH5: Randomly select two bits in the order code string, move the string substring between the two bits to the front of the order code;

[0070] LLH6: Randomly select two bits in the order code, randomly shuffle the substring between the two bits;

[0071] The LLH method set for the semiconductor packaging bonding process scheduling problem is constructed by using the above method;

[0072] Step 3.1: Generate a population of N individuals, randomly select a feasible solution group P and calculate the population fitness f.

[0073] Step 3.2: Initialization.

[0074] Randomly select a group of feasible solutions from the feasible solution group, and let this group of feasible solutions be p i , and its fitness value is f i . Let P B be the optimal solution individual, and F B be the fitness value of the optimal solution fit. Let P B equal p i , F B equal f i , and the state is set to 0, and a random action in the action space is selected;

[0075] Step 3.3: Operate the last action action, generate individuals and corresponding fitness values by calculation. According to the fitness value, we can determine the immediate reward reward. At this time, the current state is regarded as the "next state". Next, we need to determine the state to which the state belongs, and calculate the value of the new state state';

[0076] Step 3.4: Acceptance and retention of solutions. Compare the fitness value of the new solution with that of the old solution. If the fitness value of the new solution (fit') is greater than that of the old solution (fit), save the new solution and its fitness value, and update the state to the new solution (state' = state, fit' = fit). If the fitness value of the new solution is less than that of the old solution (fit' < fit), consider using the simulated annealing algorithm to re-evaluate. If the calculated probability (p) is greater than a random value, then also save the new solution and update the state (state' = state, fit' = fit), otherwise, discard the solution.

[0077] Step 3.5: Determine the reward and the new state state’. Let EP represent the experience pool, and store the above values, then EP nE = [state, action, reward, state’], where nE represents the number of data groups in the experience pool;

[0078] Step 3.6: Determine whether the algorithm ends. If the termination condition is met, the algorithm ends and outputs the optimal solution. Otherwise, update the target network, send the new state into the network and obtain the action, and return to Step 3.3.

[0079] Step Four, output the optimal scheduling plan for the bonding process, including the processing priority ranking plan for the orders to be processed and the selection priority ranking plan for the processing equipment units, and arrange production.

[0080] In this embodiment, the data related to the bonding process orders is shown in Table 1 below, which records the production quantities of 8 orders, and the processing times of the magazines of each order on different equipment units.

[0081] Table 1 Bonding process workshop scheduling order processing time data

[0082]

[0083]

[0084] The reinforcement learning-based hyperheuristic algorithm (RL-HH) adopted in this invention is compared with three simple rules and the genetic algorithm in terms of performance. Through 6 independent experimental verifications, it can be seen that RL-HH is superior to the simple rules and the traditional genetic algorithm in terms of performance, as Figure 4 shown;

[0085] Figure 5 is the bar chart of the standard deviation of the scheduling results of each algorithm. It can be seen that the RL-HH algorithm is superior to the simple rules and the traditional genetic algorithm in terms of stability.

[0086] Figure 6The scheduling time comparison chart of the embodiment is shown in the figure, and it can be seen that the response time of the super heuristic algorithm is shorter.

[0087] Based on the same inventive concept, the embodiment of the application also provides a semiconductor packaging and bonding process fast response scheduling system based on a super heuristic algorithm. Figure 7 As shown in the figure, the system comprises a data interaction module, a data preprocessing module, a scheduling scheme generation and optimization module, the data interaction module comprises a scheduling information acquisition and import submodule and a scheduling result output submodule, the scheduling information acquisition and import submodule receives the user inputted processing order data information and equipment unit data information, and outputs them to the data preprocessing module; and finally receives the scheduling result, and outputs the scheduling result;

[0088] The data preprocessing module comprises an order splitting submodule and a split batch order output submodule, the order splitting submodule is used for splitting the imported processing order, adopts a batch strategy, and outputs to the split batch order output submodule after batch completion; and the order data after order splitting is outputted by the split batch order output submodule to the scheduling scheme generation and optimization module.

[0089] The scheduling scheme generation and optimization module is used for receiving the inputted order data and scheduling model, generating a scheduling scheme, optimizing the scheduling scheme through a super heuristic algorithm, and finally outputting the scheduling result to the data interaction module.

[0090] The scheduling system further comprises a memory and a processor, and the processor executes a computer program to realize the steps in the semiconductor packaging and bonding process fast response scheduling method based on a super heuristic algorithm.

[0091] The embodiment is only a further explanation of the application, and is not a limitation of the application, and those skilled in the art can make non-creative modifications to the embodiment according to needs after reading the specification, but as long as the modifications are within the scope of the claims of the application, they are protected by the Patent Law.

Claims

1. A method for fast response scheduling of semiconductor package bonding process based on super heuristic algorithm, characterized in that, The method comprises the following steps: Step one, establishing a semiconductor packaging and bonding process scheduling optimization model; The semiconductor packaging and bonding process scheduling optimization model established in step one comprises symbol definition, objective function, constraint condition and decision variable; Step two, screening out large batch orders, and performing order splitting on the large batch orders through a batching strategy; The batching strategy in step two is a minimum batch size batching strategy, which splits orders whose magazine workpiece quantity is greater than the maximum batch size constraint until the batch size interval is met; Step three, generating a scheduling scheme through a hyper-heuristic algorithm, using a dqn deep reinforcement learning algorithm as a high-level selection strategy and a heuristic operator as a method for the underlying problem domain; Specifically, state information: the purpose of state design is to reflect the influence on the environment after the action is executed, and to predict the selection of the next action; the state is represented as the change degree of the optimal fitness value of the previous and next generations and the proportion of the number of high-quality fitness value solutions in the newly generated solutions, wherein the optimal fitness value of the current generation is fit, the optimal fitness value of the next generation is fit', and the change degree of the optimal fitness value of the previous and next generations is represented as state=(fit'-fit) / fit; Action space: the action space is an integer (1, 2, 3, …, NA), wherein each integer represents a reference to each low-level heuristic operator, and selecting which integer is equivalent to selecting which low-level heuristic operator; Reward function: the reward function is used to represent the immediate return value, and the objective function is to minimize the maximum completion time; if the maximum completion time of the current solution is less than that of the previous generation, the current solution is better, reward=1; if the solution is not improved, reward=0, and if the quality decreases, reward=-1; The low-level heuristic operation is as follows: LLH1: randomly select a bit in the order code string, and change it to another random device unit corresponding to the order selectable device unit; LLH2: randomly select two bits in the device unit code string, and exchange them with each other; LLH3: randomly select two bits in the order code string, and exchange them with each other; LLH4: randomly select two bits in the order code string, and reverse the order of the string substring between the two bits to obtain a new order code sequence; LLH5: randomly select two bits in the order code string, and move the string substring between the two bits to the front of the order code; LLH6: randomly select two bits in the order code, and randomly disorder the order of the substring between the two bits; The LLH method set of the semiconductor packaging and bonding process scheduling problem is constructed by using the above method. The third step comprises: generating a population of N groups of individuals, randomly selecting a feasible solution group P and calculating the population fitness f; initialization; operating the last action, generating individuals and corresponding fitness values by calculation, determining the immediate reward according to the fitness value, at this time, the current state is regarded as the "next state", then determining the state to which the current state belongs and calculating the value of the new state; solution acceptance reservation; determining the reward and the new state; judging whether the algorithm is ended, if the termination condition is met, the algorithm is ended, and the optimal solution is output, otherwise, updating the target network, sending the new state into the network and obtaining the action, returning to the initialization, operating the last action and the following steps; In the third step, the initialization is to randomly select a set of feasible solutions from the feasible solution set, and let this set of feasible solutions be p i , and the fitness value is f i , let P B be the optimal solution individual, F B be the fitness value of the optimal solution, let P B equal p i , F B equal f i , and the state be set to 0, and a random action in the action space be selected; In the third step, the initialization is to randomly select a set of feasible solutions from the feasible solution set, and let this set of feasible solutions be p i , and the fitness value is f i , let P B be the optimal solution individual, F B be the fitness value of the optimal solution, let P B equal p i , F B equal f i , and the state be set to 0, and a random action in the action space be selected; The solution acceptance reservation is to compare the fitness value of the new solution with the fitness value of the old solution, if the fitness value of the new solution is greater than the fitness value of the old solution, the new solution and the fitness value thereof are saved, and the state is updated to the new solution; if the fitness value of the new solution is less than the fitness value of the old solution, the simulated annealing algorithm is considered to be used for reevaluation; if the calculated probability is greater than a random value, the new solution is also saved and the state is updated, otherwise, the solution is abandoned; The fourth step is to output the optimal scheduling scheme of the bonding process, including the processing priority sequencing scheme of the to-be-processed order and the selection priority sequencing scheme of the processing equipment unit, and arranging production.

2. The system for semiconductor package bonding process quick response scheduling method based on super heuristic algorithm according to claim 1, wherein, The system comprises a data interaction module, a data preprocessing module and a scheduling scheme generation and optimization module, the data interaction module comprises a scheduling information acquisition and import submodule and a scheduling result output submodule, the scheduling information acquisition and import submodule receives the to-be-processed order data information and the equipment unit data information input by a user and outputs them to the data preprocessing module; and finally receives the scheduling result and outputs the scheduling result; the data preprocessing module comprises an order splitting submodule and a split batch order output submodule, the order splitting submodule is used for splitting the imported to-be-processed order, adopts a batch strategy, outputs to the split batch order output submodule after batch completion, and outputs the order data after order splitting from the split batch order output submodule to the scheduling scheme generation and optimization module; The scheduling scheme generation and optimization module is used for receiving the input order data and the scheduling model, generating a scheduling scheme and optimizing the scheduling scheme through a hyper-heuristic algorithm, and finally outputting the scheduling result to the data interaction module.

3. The system for semiconductor package bonding process quick response scheduling method using super heuristic algorithm-based in claim 2, wherein, The system further comprises a memory and a processor.

Images