CN103310285A

CN103310285A - Performance prediction method applicable to dynamic scheduling for semiconductor production line

Info

Publication number: CN103310285A
Application number: CN2013102395010A
Authority: CN
Inventors: 乔非; 马玉敏; 徐灵璐
Original assignee: Tongji University
Current assignee: Tongji University
Priority date: 2013-06-17
Filing date: 2013-06-17
Publication date: 2013-09-18

Abstract

The invention discloses a performance prediction method for dynamic scheduling of a semiconductor production line, which uses an extreme learning machine (Extreme Learning Machine, ELM) for prediction modeling. The invention considers feeding control and scheduling rules in a unified way, and predicts key short-term scheduling indicators such as equipment utilization rate and moving steps based on the real-time status of the system, and provides a basis for dynamic real-time scheduling. Introduce the new feed-forward neural network of ELM into the semiconductor manufacturing system, and establish a predictive model through the data available on the production line. The test results show that the ELM method can quickly obtain ideal prediction results. Compared with the traditional neural network modeling method, it has obvious advantages and application prospects in parameter selection and learning speed, and provides new ideas for online optimization control.

Description

A Performance Prediction Method Applicable to Dynamic Scheduling of Semiconductor Production Lines

技术领域technical field

本发明属于半导体制造领域，尤其是一种基于极限学习机的半导体生产线调度系统。The invention belongs to the field of semiconductor manufacturing, in particular to a semiconductor production line scheduling system based on an extreme learning machine.

背景技术Background technique

半导体晶圆生产线是公认的最为复杂的生产线之一，其具有设备多、产品种类多、可重入、半成品报废与重加工、机器失效等特点，故给生产线的调度与控制带来了极大的复杂性。The semiconductor wafer production line is recognized as one of the most complicated production lines. It has the characteristics of many equipments, many types of products, re-entry, semi-finished product scrapping and reprocessing, and machine failure, so it brings great challenges to the scheduling and control of the production line. complexity.

一个完整的半导体生产线调度包括投料控制、工件调度。投料控制用于确定物料进入生产系统的速率，工件调度则是指对竞争使用某设备的工件，当设备出现空闲时，按何决策从中选择下一个要加工的工件。由于预测模型具有展示系统未来动态的行为和能力，根据系统实时状态，任意地给出未来的控制策略，观察对象在不同策略下性能指标的输出变化。一方面性能预测对于控制评价提供了依据和参考，能够从生产管理的角度简单明了地表示调度结果的优劣，进而判断所得到的调度方案是否可行。另一方面根据一些不确定因素，比如设备情况变化、客户需求变化，加工瓶颈的变化等不确定因素，会大大影响生产线的状态，恶化性能指标，此时需根据不确定因素调整生产线的控制策略，来保证调度方案的进行，达到实时控制的要求。因此对于一个给定资源结构和调度策略的生产系统,计算其性能是有必要的。通过建立预测模型指导生产具有重要意义和实用价值，为高效的任务调度提供支持，保证服务的高性能运行，从而提高服务质量，以便更科学而经济地开展生产活动。A complete semiconductor production line scheduling includes feeding control and workpiece scheduling. Feeding control is used to determine the rate at which materials enter the production system. Workpiece scheduling refers to the decision to select the next workpiece to be processed when the equipment is idle for the workpieces that compete to use a certain equipment. Since the prediction model has the behavior and ability to display the future dynamics of the system, according to the real-time state of the system, the future control strategy is arbitrarily given, and the output changes of the performance indicators of the object under different strategies are observed. On the one hand, performance prediction provides a basis and reference for control evaluation, and can simply and clearly express the advantages and disadvantages of scheduling results from the perspective of production management, and then judge whether the obtained scheduling scheme is feasible. On the other hand, some uncertain factors, such as changes in equipment conditions, changes in customer demand, and changes in processing bottlenecks, will greatly affect the state of the production line and deteriorate performance indicators. At this time, it is necessary to adjust the control strategy of the production line according to uncertain factors. , to ensure the progress of the scheduling scheme and meet the requirements of real-time control. Therefore, it is necessary to calculate the performance of a production system with a given resource structure and scheduling strategy. Guiding production by establishing a predictive model is of great significance and practical value, providing support for efficient task scheduling, ensuring high-performance operation of services, thereby improving service quality, and carrying out production activities more scientifically and economically.

目前预测建模方法主要有纯计算方法、仿真建模方法两类方法，然而对于复杂的半导体制造过程有着诸多约束和缺陷。首先，使用传统建模方法建立指导生产过程调度的精确数学模型变得越来越困难，而因假设降低了实际复杂度，又难以保证所建模型的精度；其次，仿真建模方法需要大量的时间和资金并且模型的运行时间过长，并且无法满足对生产现场中动态不确定因素做出快速有效反应的需要。制造生产过程中产生了大量的离在线数据，其中隐含了反映实际调度环境特点及调度知识的大量有效信息。人工神经网络具有自组织、自适应、并行处理和非线性的等特性，作为数据挖掘技术的一种，提取大量数据中的有用信息来描述认知、决策及控制的智能行为，被广泛应用于制造业控制领域。然而传统的神经网络的学习速度不能够满足高实时性的要求，无法满足对生产现场中动态不确定因素做出快速有效反应的需要致使成为其应用的重要瓶颈，尤其对于实时在线预测。At present, predictive modeling methods mainly include pure calculation methods and simulation modeling methods. However, there are many constraints and defects for complex semiconductor manufacturing processes. First of all, it is becoming more and more difficult to establish an accurate mathematical model to guide the production process scheduling using traditional modeling methods, and it is difficult to guarantee the accuracy of the built model because of the assumption that the actual complexity is reduced; secondly, the simulation modeling method requires a lot of Time and money and the running time of the model is too long, and it cannot meet the need to react quickly and effectively to the dynamic uncertainties in the production site. A large amount of offline data is generated in the manufacturing process, which contains a large amount of effective information reflecting the characteristics of the actual scheduling environment and scheduling knowledge. Artificial neural network has the characteristics of self-organization, self-adaptation, parallel processing and nonlinearity. As a kind of data mining technology, it extracts useful information from a large amount of data to describe the intelligent behavior of cognition, decision-making and control, and is widely used in field of manufacturing control. However, the learning speed of the traditional neural network cannot meet the high real-time requirements and the need to respond quickly and effectively to dynamic uncertain factors in the production site, which has become an important bottleneck for its application, especially for real-time online prediction.

极限学习机（ELM），是一种简单易用、有效的单隐层前馈神经网络的新型学习算法。该算法不同于传统的学习算法的提出，这种学习方法在保证网络具有良好泛化性能的同时,极大程度地提高了前向神经网络的学习速度,并且避免了基于梯度下降学习方法的许多问题,如局部极小、迭代次数过多、参数选择敏感等。因此ELM算法逐步受到关注，但现在处于发展阶段，在光谱分析、岩性识别、电力线路建设、指纹人脸识别等方面得到了应用。Extreme Learning Machine (ELM) is an easy-to-use and effective new learning algorithm for single hidden layer feed-forward neural network. This algorithm is different from the traditional learning algorithm. This learning method greatly improves the learning speed of the feed-forward neural network while ensuring that the network has good generalization performance, and avoids many problems of the gradient-based learning method. Problems, such as local minimum, too many iterations, sensitive parameter selection, etc. Therefore, the ELM algorithm has gradually attracted attention, but it is now in the development stage, and has been applied in spectral analysis, lithology identification, power line construction, fingerprint and face recognition, etc.

发明内容Contents of the invention

针对上述的技术缺陷和应用时的不利影响，本发明的目的是提供一种应用极限学习机，解决半导体生产线动态调度的性能预测问题，保证短期调度方案的可行性与有效性。In view of the above-mentioned technical defects and adverse effects during application, the purpose of the present invention is to provide an application of extreme learning machine, solve the performance prediction problem of dynamic scheduling of semiconductor production lines, and ensure the feasibility and effectiveness of short-term scheduling schemes.

本发明的技术方案如下：Technical scheme of the present invention is as follows:

一种基于极限学习机的用于半导体生产线动态调度的性能预测方法，包括：A performance prediction method for dynamic scheduling of semiconductor production lines based on extreme learning machines, including:

（1）采集半导体生产线历史数据，建立训练样本集与测试样本集。(1) Collect historical data of the semiconductor production line, and establish a training sample set and a test sample set.

（2）将对输入的投料方式及调度规则文字符号进行编码，使网络可以接受。为使输入数据有相同的量纲，对输入量进行归一化处理。为使输入数据有相同的量纲，对输入量进行归一化处理，将数据限制在[0,1]区间内，归一化公式为[x-min(x_i)]/[max(x_i)-min(x_i)]，其中，x指输入变量，i为样本编号。(2) Encode the input material feeding method and dispatching rule text symbols so that the network can accept it. In order to make the input data have the same dimension, the input quantity is normalized. In order to make the input data have the same dimension, the input quantity is normalized, and the data is limited to the [0,1] interval. The normalization formula is [x-min( _xi )]/[max(x _i )-min( _xi )], where x refers to the input variable and i is the sample number.

（3）采用极限学习机方法构建预测模型。对于ELM仅需要确定的神经网络的隐层节点个数，不需要调整网络的输入权值和隐元的偏置及其他参数，采用试凑法选取适合的隐节点个数。所以假设对于给定的N个不同样本的训练集

其中x为n维输入变量，即x_i∈Rⁿ，t为m维输出变量，即t_i∈R^m，i为样本编号；激活函数为g(x)，隐含层节点数为

，则(3) Using the extreme learning machine method to build a prediction model. For ELM, only the number of hidden layer nodes of the neural network needs to be determined, and there is no need to adjust the input weights of the network, the bias of hidden elements and other parameters, and the trial and error method is used to select the appropriate number of hidden nodes. So suppose for a given training set of N different samples

Where x is an n-dimensional input variable, that is, x _i ∈ R ⁿ , t is an m-dimensional output variable, that is, t _i ∈ R ^m , i is a sample number; the activation function is g(x), and the number of hidden layer nodes is

,but

A）随机获取初始输入权值w_i和偏置b_i，i=1,...,

。A) Randomly obtain the initial input weight w _i and bias b _i , i=1,...,

.

B）计算隐层输出矩阵H。B) Calculate the hidden layer output matrix H.

$H h (({w w}_{11},, . . . . . .,, {w w}_{\overset{~ ~}{N N}},, {b b}_{11},, . . . . . .,, {b b}_{\overset{~ ~}{N N}},, {x x}_{11},, . . . . . .,, {x x}_{N N})) = = {[\begin{matrix} g g (({w w}_{11} \cdot &Center Dot; {x x}_{11} + + {b b}_{11})) & \cdot &Center Dot; \cdot &Center Dot; \cdot &Center Dot; & g g (({w w}_{\overset{~ ~}{N N}} \cdot &Center Dot; {x x}_{11} + + {b b}_{\overset{~ ~}{N N}})) \\ \cdot &Center Dot; & \cdot &Center Dot; \\ \cdot &Center Dot; & \cdot &Center Dot; \cdot &Center Dot; \cdot &Center Dot; & \cdot &Center Dot; \\ \cdot &Center Dot; & \cdot &Center Dot; \\ g g (({w w}_{11} \cdot \cdot {x x}_{N N} + + {b b}_{11})) & \cdot &Center Dot; \cdot &Center Dot; \cdot \cdot & g g (({w w}_{\overset{~ ~}{N N}} \cdot \cdot {x x}_{N N} + + {b b}_{\overset{~ ~}{N N}})) \end{matrix}]}_{N N \times \times \overset{~ ~}{N N}}$

C）计算输出权值β：

T＝[t₁,…,t_N]^T，其中

是矩阵H的Moore-Penrose广义逆矩阵。C) Calculate the output weight β:

T=[t ₁ ,…,t _N ] ^T , where

is the Moore-Penrose generalized inverse matrix of matrix H.

（4）运用测试样本测试预测模型的网络性能，将预测结果反归一化处理后得到输出值o_i其中，i＝1,…,m，m指输出值o的维数，即有m个输出；与测试样本输出值对比，判断是否能够满足精度要求。精度要求可以采用平均相对误差为基准，对于m个输出变量，

其中n为测试样本个数，j为样本编号，i为输出变量编号，t为预期的输出变量。将测试的平均相对误差结果与预期的相对误差相比较,保证m个变量中最大平均相对误差小于等于预测相对平均误差，即预测精度

其中δ′为设定的精度要求。(4) Use test samples to test the network performance of the prediction model, and denormalize the prediction results to obtain the output value o _i where, i=1,...,m, m refers to the dimension of the output value o, that is, there are m Output; compared with the output value of the test sample, it is judged whether the accuracy requirement can be met. The accuracy requirement can be based on the average relative error. For m output variables,

Among them, n is the number of test samples, j is the sample number, i is the output variable number, and t is the expected output variable. Compare the average relative error result of the test with the expected relative error to ensure that the maximum average relative error among the m variables is less than or equal to the predicted relative average error, that is, the prediction accuracy

Where δ' is the set accuracy requirement.

（5）如果测试结果的预测精度能够满足要求，建立预测模型过程结束，得到所需的预测模型；如不满足，则转到步骤（4），重新选择的神经网络的隐层节点个数再次训练模型。(5) If the prediction accuracy of the test results can meet the requirements, the process of establishing the prediction model is over, and the required prediction model is obtained; if not, go to step (4), and the number of hidden layer nodes of the re-selected neural network is again Train the model.

所述历史数据包括输入量和输出量，其中输入量为：投料方式、调度规则以及系统实时状态缓冲区队长、在制品数量WIP、生产线上个调度周期内产生移动步数等；输出量为系统短期调度性能指标，包括日生产率、设备利用率、瓶颈设备排队队长和调度周期内产生的移动量。The historical data includes input volume and output volume, wherein the input volume is: feeding mode, scheduling rules, system real-time status buffer captain, WIP quantity WIP, number of moving steps generated in the last scheduling cycle of the production line, etc.; output volume is the system Short-term scheduling performance indicators, including daily production rate, equipment utilization rate, queue length of bottleneck equipment, and movement generated during the scheduling period.

本发明的有益效果是：The beneficial effects of the present invention are:

本发明基于极限学习机算法对半导体生产线的短期性能预测建立了神经网络预测模型，能够对系统中的实时变化及时作出反应，减少了重调度的需要。本发明把系统看成黑箱，提供的建模方法把利用生产线可获得的历史数据及离在线数据，挖掘其中有用知识，实现实时在线优化控制。本发明根据半导体生产线特征考虑短期因素包括缓冲区等待工件数量、生产线在制品数量、生产线已产生移动步数等，及短期性能指标包括生产率、设备利用率、排队队长、移动步数等，及将半导体的调度的投料控制与工件调度统一考虑，形成系统的预测建模研究。The invention establishes a neural network prediction model for short-term performance prediction of a semiconductor production line based on an extreme learning machine algorithm, can respond to real-time changes in the system in time, and reduces the need for rescheduling. The invention regards the system as a black box, and provides a modeling method that utilizes historical data and offline data that can be obtained from the production line, digs out useful knowledge in them, and realizes real-time online optimization control. According to the characteristics of the semiconductor production line, the present invention considers short-term factors including the number of workpieces waiting in the buffer zone, the number of work-in-progress in the production line, the number of moving steps in the production line, etc., and the short-term performance indicators include productivity, equipment utilization, queue length, and moving steps. The feeding control and workpiece scheduling of semiconductor scheduling are considered together to form a systematic predictive modeling research.

本发明提供的性能预测建模方法效率高、实时性好、实现方便，非常实用动态实时调度，为在线优化控制提供了有效的途径。The performance prediction modeling method provided by the invention has high efficiency, good real-time performance, convenient realization, very practical dynamic real-time scheduling, and provides an effective way for online optimization control.

附图说明Description of drawings

图1为半导体生产线简化模型示意图；Figure 1 is a schematic diagram of a simplified model of a semiconductor production line;

图2为本发明实施例生产调度系统的实现ELM预测建模的流程图；Fig. 2 is the flow chart of the realization of ELM predictive modeling of the production scheduling system of the embodiment of the present invention;

图3为本发明实施例建立神经网络的结构示例图。FIG. 3 is a structural example diagram of establishing a neural network according to an embodiment of the present invention.

具体实施方式Detailed ways

为更好地阐述本发明的调度系统的控制方法，请参阅图1。To better illustrate the control method of the dispatching system of the present invention, please refer to FIG. 1 .

图1为半导体生产线简化模型示意图，以生产工艺分类，有三个设备群，共五台设备，分别为扩散设备群、离子注入设备群和光刻设备群。扩散设备群包括第一扩散设备Ma、第二扩散设备Mb；离子注入设备群包括第一离子注入设备Mc、第二离子注入设备Md；光刻设备群包括光刻设备Me。在每个设备群前，分别设有第一缓冲区B_Mab、第二缓冲区B_Mcd和第三缓冲区B_Me用来储存需要等待加工的工件信息。通过上述的设备群，可以实现六道工序，包括：扩散、离子注入、光刻、离子注入、扩散、光刻。本领域技术人员可知，同一设备群中的设备可以被同时选中或单个选中工作。Figure 1 is a schematic diagram of a simplified model of a semiconductor production line, classified by production process, there are three equipment groups, a total of five equipment, namely the diffusion equipment group, ion implantation equipment group and lithography equipment group. The diffusion equipment group includes first diffusion equipment Ma and second diffusion equipment Mb; the ion implantation equipment group includes first ion implantation equipment Mc and second ion implantation equipment Md; the lithography equipment group includes lithography equipment Me. In front of each equipment group, a first buffer B_Mab, a second buffer B_Mcd and a third buffer B_Me are respectively set up to store workpiece information waiting to be processed. Through the above-mentioned equipment group, six processes can be realized, including: diffusion, ion implantation, photolithography, ion implantation, diffusion, and photolithography. Those skilled in the art know that the devices in the same device group can be selected to work simultaneously or individually.

具体而言，Minifab是根据实际生产线简化而来的一个简单的半导体生产线模型，如图所示。Minifab生产三种类型的工件，其加工路径是一样的。工件的工艺流程包括六个工序：工序一和工序五为扩散，可以在Ma或Mb上加工；工序二和工序四为离子注入，可以在设备Mc或Md上加工；工序三和工序六为光刻，在设备Me上加工。三个设备群都具有可重入性。设备Ma和Mb为多卡并行加工设备，最大加工批量为3卡，只有工序相同的工件才可以组批一起并行加工。Ma和Mb是完全可互换的两台设备。设备Mc和Md为单卡批加工设备，每台设备一次只能加工一卡工件。Mc和Md也是完全可互换的两台设备。设备Me也为单卡批加工设备。Specifically, Minifab is a simple semiconductor production line model simplified based on the actual production line, as shown in the figure. Minifab produces three types of workpieces with the same machining path. The technological process of the workpiece includes six processes: process 1 and process 5 are diffusion, which can be processed on Ma or Mb; process 2 and process 4 are ion implantation, which can be processed on equipment Mc or Md; process 3 and process 6 are optical Engraved, processed on the device Me. All three device groups are reentrant. Equipment Ma and Mb are multi-card parallel processing equipment, and the maximum processing batch is 3 cards. Only workpieces with the same process can be grouped together for parallel processing. Ma and Mb are two pieces of equipment that are completely interchangeable. Equipment Mc and Md are single-card batch processing equipment, and each device can only process one card of workpieces at a time. The Mc and Md are also two pieces of equipment that are completely interchangeable. Equipment Me is also a single-card batch processing equipment.

图2为本发明实施例生产调度系统的实现ELM预测建模的流程图。Fig. 2 is a flow chart of implementing ELM predictive modeling in the production scheduling system according to the embodiment of the present invention.

首先，在数据样本中确定输入变量（S101）和输出变量（S102）。步骤S101中，输入变量主要包括投料方式P、调度规则和系统实时状态。调度规则包括扩散设备群调度规则R_ab和光刻设备群调度规则R_e；系统实时状态系统实时状态是当前阶段对制造系统生产状况客观的描述，包括生产线在制品数量WIP，第一缓冲区B_M_ab缓冲区队长Q_ab，第二缓冲区B_Mcd缓冲区队长Q_cd，第三缓冲区B_M_e缓冲区队长Q_e，缓冲区总队长Q’，以及生产线在上个调度周期内产生的移动步数M，其中，M’是指生产线上所有硅片在上个调度周期内移动的总步数。First, an input variable (S101) and an output variable (S102) are determined in a data sample. In step S101, the input variables mainly include feeding mode P, scheduling rules and real-time system status. The scheduling rules include the diffusion equipment group scheduling rule R _ab and the lithography equipment group scheduling rule R _e ; the real-time status of the system is an objective description of the production status of the manufacturing system at the current stage, including the number of WIP of the production line, the first buffer zone B_M _ab buffer leader Q _ab , second buffer B_Mcd buffer leader Q _cd , third buffer B_M _e buffer leader Q _e , buffer total leader Q', and the number of moving steps generated by the production line in the last scheduling cycle M, where M' refers to the total number of steps that all silicon wafers on the production line moved in the last scheduling cycle.

其中，投料方式P决定了产品的种类及其到达生产线的时间和速率，对生产线的状态和性能有直接的影响，可选的，本实施例的投料方式为：固定时间间隔投料（Conrin）、泊松投料（Possion）、固定在制品投料（Conwip）。调度规则决定等待加工工件分配不同的加工优先级。本实施例中，扩散设备群调度规则R_ab和光刻设备群调度规则R_e是先入先出（FIFO）、最早交货期优先（EDD）、临界值（CR）、最短剩余加工时间（SRPT）、制造周期方差波动平滑（FSVCT）中的一种或几种。缓冲区队长则表示等待加工的工件个数，Among them, the feeding method P determines the type of product and the time and rate at which it reaches the production line, which has a direct impact on the state and performance of the production line. Optionally, the feeding method in this embodiment is: fixed time interval feeding (Conrin), Poisson feeding (Possion), fixed WIP feeding (Conwip). Scheduling rules decide to assign different processing priorities to workpieces waiting to be processed. In this embodiment, the diffusion equipment group scheduling rule R _ab and the lithography equipment group scheduling rule R _e are first in first out (FIFO), earliest delivery date first (EDD), critical value (CR), shortest remaining processing time (SRPT) ), one or more of manufacturing cycle variance fluctuation smoothing (FSVCT). The buffer captain indicates the number of workpieces waiting to be processed.

步骤S102中，输出变量包括生产率

设备利用率、瓶颈设备的排队队长和总移动量。In step S102, output variables include productivity

Device utilization, queue length and total movement of bottleneck devices.

生产率

为单位时间内加工完成的lot（批次）数量，其值越高，表示创造的价值越高，故其能够直接反映日投料计划的好坏。生产率

其中P′为调度周期内的总产量，T_c为调度周期时间。设备利用率表示设备处于加工状态的时间占其开机时间的比率，记为U_i，其中i∈(a,b,c,d,e)，表示设备的标识符，例如U_a表示设备Ma的设备利用率。描述为

其中T_opi为Mi设备的开机时间，T_ui为Mi设备加工产品时的使用时间。本实施例中，设备利用率指的是Ma设备利用率U_a、Mb设备利用率U_b、Mc设备利用率U_c、Md设备利用率U_d、Me设备利用率U_e。瓶颈设备的排队队长表示瓶颈加工设备前等待加工的工件个数，一般为设备前缓冲区内的工件数量，记为Q。若其过长会导致平均加工周期的增加，产品合格率降低等。本实施例中，以光刻设备Me作为瓶颈设备，则瓶颈设备的排队队长指Me设备的排队队长Q。总移动量指所有硅片在单位时间内移动的总步数，记作M。每产生一步移动，则相应批次完成一步加工程序。总移动量越高，生产线完成的加工任务越高。productivity

It is the number of lots (batches) processed per unit time. The higher the value, the higher the value created, so it can directly reflect the quality of the daily feeding plan. productivity

Among them, P' is the total output in the scheduling cycle, and T _c is the scheduling cycle time. The equipment utilization rate indicates the ratio of the time that the equipment is in the processing state to its startup time, which is recorded as U _i , where i∈(a,b,c,d,e) represents the identifier of the equipment, for example, U _a represents the equipment utilization. described as

Among them, _Topi is the power-on time of the Mi device, and T _ui is the usage time of the Mi device when processing products. In this embodiment, the equipment utilization refers to Ma equipment utilization U _a , Mb equipment utilization U _b , Mc equipment utilization U _c , Md equipment utilization U _d , and Me equipment utilization U _e . The queue leader of the bottleneck equipment indicates the number of workpieces waiting to be processed in front of the bottleneck processing equipment, which is generally the number of workpieces in the buffer zone in front of the equipment, denoted as Q. If it is too long, it will lead to an increase in the average processing cycle and a decrease in the product qualification rate. In this embodiment, the lithography device Me is used as the bottleneck device, and the queue length of the bottleneck device refers to the queue length Q of the Me device. The total amount of movement refers to the total number of steps that all silicon chips move per unit time, denoted as M. Each time a movement occurs, the corresponding batch completes a step of processing procedures. The higher the total movement, the higher the processing task that the line can complete.

确定好输入、输出变量后，即可收集数据样本，相应的产生样本输出变量（S103）和样本输出变量（S104）。After the input and output variables are determined, data samples can be collected, and correspondingly generate sample output variables (S103) and sample output variables (S104).

生产系统处于正常工作状态后，基于投料方式和调度规则的组合，观察一个调度周期（本实施例为10天）之后输出的生产线性能指标。可选的，此时的投料方式为Conrin方式，调度规则为FIFO策略。假设，第一、第二离子注入设备Mc、Md在FIFO策略下已满足加工需求，不予考虑规则选取，对投料方式和Ma、Mb氧化扩散设备群及Me光刻瓶颈设备的优先规则进行选取。本领域人员可知，样本数目可以根据不同的生产线进行改变，而非仅限于上述的数目，同理，调度周期也可以根据实际情况进行选取。After the production system is in a normal working state, based on the combination of feeding mode and scheduling rules, observe the output production line performance indicators after a scheduling period (10 days in this embodiment). Optionally, the feeding mode at this time is the Conrin mode, and the scheduling rule is the FIFO strategy. Assume that the first and second ion implantation equipment Mc and Md have met the processing requirements under the FIFO strategy, and the selection of rules is not considered, and the priority rules for the feeding method, Ma, Mb oxidation diffusion equipment group and Me lithography bottleneck equipment are selected. . Those skilled in the art know that the number of samples can be changed according to different production lines, and is not limited to the above number. Similarly, the scheduling cycle can also be selected according to actual conditions.

然后，则为了能够让样本数据可以统一处理，则需要对输入的投料方式及调度规则文字符号进行编码，即对数据进行预处理（S105），主要包括编码、归一化等。Then, in order to allow the sample data to be processed uniformly, it is necessary to encode the input feeding method and dispatching rule text symbols, that is, to preprocess the data (S105), which mainly includes encoding, normalization, etc.

在本实施例中，编码方式如下：In this embodiment, the encoding method is as follows:

投料方式：1.Conrin，2.Possion，3.Conwip；Ma、Mb调度规则：1.FIFO，2.EDD，3.SRPT；Me调度规则：1.EDD，2.SRPT，3.CR，4.FSVCT。Feeding method: 1.Conrin, 2.Possion, 3.Conwip; Ma, Mb scheduling rules: 1.FIFO, 2.EDD, 3.SRPT; Me scheduling rules: 1.EDD, 2.SRPT, 3.CR, 4 .FSVCT.

归一化公式为[x-min(x_i)]/[max(x_i)-min(x_i)]，这里，i为样本编号，x为样本值。如此，即可建立如图3所示的神经网络结构，该结构包含9个输入变量，8个输出变量。The normalization formula is [x-min( _xi )]/[max( _xi )-min(xi ₎ ], where i is the sample number and x is the sample value. In this way, the neural network structure shown in Figure 3 can be established, which includes 9 input variables and 8 output variables.

数据经预处理后，即可在步骤S103和步骤S104中产生的样本输入、输出变量选出训练样本（S107）和测试样本（S106）。本实施例中，控制策略根据投料方式和调度规则的不同形成了36种组合，对每种组合随机抓取系统状态，采集50条训练样本，形成1800组输入输出数据，另采集200组作为测试样本。After the data is preprocessed, the training samples (S107) and test samples (S106) can be selected from the sample input and output variables generated in step S103 and step S104. In this embodiment, the control strategy forms 36 combinations according to the different feeding methods and scheduling rules. For each combination, the system status is randomly captured, 50 training samples are collected, and 1800 sets of input and output data are formed, and another 200 sets are collected as tests. sample.

在确定好训练样本后，确定隐含层神经元个数（S109）。由于不同的隐含层节点个数会导致预测性能的差异。After the training samples are determined, the number of hidden layer neurons is determined (S109). Due to the different number of hidden layer nodes, the prediction performance will be different.

然后，基于隐含层神经元个数构建ELM神经网络（S111），并确定最终的ELM神经网络模型（S113），即是ELM神经网络是由确定隐层节点的神经网络经ELM算法训练而得。Then, construct the ELM neural network based on the number of neurons in the hidden layer (S111), and determine the final ELM neural network model (S113), that is, the ELM neural network is obtained by training the neural network of the hidden layer nodes through the ELM algorithm .

分别选取测试样本中的输入数据（S108）和输出数据（S110），这里的输入数据即为上文中的输入变量数据，输出数据即为输出变量数据；然后，将输入数据输入值ELM神经网络模型（S113），经计算后，即可产生相应的预测结果（S115），然后将该预测结果与测试样本中的输出数据对比（S112）；接着，预测结果与测试样本中的输出数据的误差是否满足指定的精度要求（S114），若不满足精度要求，则返回至步骤（S109）重新进行确定隐含层神经元个数，本实施例中，以测试样本200个的各指标的相对误差总均值为标准，当隐含层节点数为135时，测试误差最小。若满足精度要求，则输出此ELM神经网络模型（S116）；在实际生产过程中，即可将半导体产线实时状态出入至该模型（S116），经该模型计算后，即可输出预测的性能指标（S117）。Select the input data (S108) and output data (S110) in the test sample respectively, where the input data is the input variable data above, and the output data is the output variable data; then, input the input data into the ELM neural network model (S113), after calculation, the corresponding prediction result can be generated (S115), and then the prediction result is compared with the output data in the test sample (S112); then, whether the error between the prediction result and the output data in the test sample is Satisfy the specified accuracy requirement (S114), if not, return to step (S109) to re-determine the number of neurons in the hidden layer. The mean value is the standard, and when the number of hidden layer nodes is 135, the test error is the smallest. If the accuracy requirements are met, then output the ELM neural network model (S116); in the actual production process, the real-time status of the semiconductor production line can be entered into and out of the model (S116), and the predicted performance can be output after the model is calculated indicator (S117).

根据测试样本测试网络的拟合性能，反归一化后预测结果与样本值对比，分析各指标值的相对平均误差。表1为200组样本数据测试的平均相对误差比较，从预测结果可以看出，ELM的预测精度均低于10%，到达理想的预测精度。通过与BP神经网络和RBF神经网络建模效果对比发现，预测精度非常接近，最大相差仅0.0638。然而ELM的训练应该是速度是BP的330.57倍，是RBF的60.45倍，测试速度是BP的7.40倍，RBF的9.38倍，可见ELM的学习和测试速度明显优于其他方法，表现出在线预测的较大优势。According to the fitting performance of the test sample test network, the predicted results after denormalization are compared with the sample values, and the relative average error of each index value is analyzed. Table 1 compares the average relative error of 200 sets of sample data tests. It can be seen from the prediction results that the prediction accuracy of ELM is lower than 10%, reaching the ideal prediction accuracy. By comparing the modeling effects of BP neural network and RBF neural network, it is found that the prediction accuracy is very close, and the maximum difference is only 0.0638. However, the training speed of ELM should be 330.57 times that of BP, 60.45 times that of RBF, and the test speed is 7.40 times that of BP, 9.38 times that of RBF. It can be seen that the learning and testing speed of ELM is significantly better than other methods, showing the effectiveness of online prediction. Big advantage.

表1200组样本数据测试的平均相对误差比较Table 1200 group of sample data test average relative error comparison

同时为了更明确的看到预测效果，从测试结果中选取四个测试样本，将ELM络的预测值与仿真得到的系统性能指标的实际值相比较，结果表2所示。从中可以看出，预测结果与实际值非常接近，效果理想。另外从结果可以看出，不同的实时系统状态即使采用同一种控制策略，实现制造系统的性能相差较大，或者相同的系统状态下采取不同的控制策略得到了不同的效果，这验证了动态实时调度中根据实时系统状态调整控制策略的重要性。At the same time, in order to see the prediction effect more clearly, four test samples were selected from the test results, and the predicted value of the ELM network was compared with the actual value of the system performance index obtained by simulation. The results are shown in Table 2. It can be seen that the predicted result is very close to the actual value, and the effect is ideal. In addition, it can be seen from the results that even if the same control strategy is adopted in different real-time system states, the performance of the manufacturing system is quite different, or different control strategies are adopted under the same system state to obtain different effects, which verifies the dynamic real-time The importance of adjusting control strategies according to real-time system state in scheduling.

表2ELM络的预测值与仿真得到的系统性能指标的实际值相比较的Table 2 Comparison between the predicted value of the ELM network and the actual value of the system performance index obtained by simulation

上述的对实施例的描述是为便于该技术领域的普通技术人员能理解和应用本发明。熟悉本领域技术的人员显然可以容易地对这些实施例做出各种修改，并把在此说明的一般原理应用到其他实施例中而不必经过创造性的劳动。因此，本发明不限于这里的实施例，本领域技术人员根据本发明的揭示，对于本发明做出的改进和修改都应该在本发明的保护范围之内。The above description of the embodiments is for those of ordinary skill in the art to understand and apply the present invention. It is obvious that those skilled in the art can easily make various modifications to these embodiments, and apply the general principles described here to other embodiments without creative efforts. Therefore, the present invention is not limited to the embodiments herein, and improvements and modifications made by those skilled in the art according to the disclosure of the present invention should fall within the protection scope of the present invention.

Claims

1. performance prediction method that can be used for Dynamic Schedule of Semiconductor Fabrication Line comprises:

(1) gathers the semiconductor production line historical data, set up training sample set and test sample book collection;

(2) will encode to feeding mode and the scheduling rule letter symbol of input, network can be accepted; For making the input data that identical dimension be arranged, input quantity is carried out normalized;

(3) adopt the extreme learning machine method to make up forecast model; Only need the hidden node number of the neural network determined for ELM, do not need to adjust the input weights of network and biasing and other parameters of hidden unit, adopt method of trial and error to choose suitable hidden node number;

(4) network performance of performance test test sample forecast model obtains output valve and the contrast of test sample book output valve after the renormalization that will predict the outcome is processed, judge whether to satisfy accuracy requirement;

(5) if the precision of prediction of test result can meet the demands, set up the forecast model process and finish, obtain required forecast model; As not satisfying, then forward step (3) to, the hidden node number of the neural network that reselects is training pattern again.

2. the method for claim 1, it is characterized in that: historical data comprises input quantity and output quantity described in the step (1), and wherein input quantity comprises: feeding mode, scheduling rule and the real-time status buffer zone team leader of system, goods in process inventory WIP, production line produce mobile step number in last dispatching cycle; Output quantity is system's short term scheduling performance index, comprises the amount of movement that produces in day throughput rate, plant factor, the team leader of bottleneck device queue and dispatching cycle.

3. the method for claim 1, it is characterized in that: step is carried out normalized to input quantity described in (2), and in [0,1] interval, the normalization formula is [x-min (x with data limit _i)]/[max (x _i)-min (x _i)], wherein, x refers to input variable, i is sample number.

4. the method for claim 1 is characterized in that: suppose the training set for given N different samples in the step (3)

Wherein x is n dimension input variable, i.e. x _i∈ R ⁿ, t is m dimension output variable, i.e. t _i∈ R ^m, i is sample number; Activation function is g (x), and the hidden layer node number is

, then

A) obtain at random the initial input weight w _iWith biasing b _i, i=1 ...,

B) calculate hidden layer output matrix H;

H (w_{1}, . . ., w_{\tilde{N}}, b_{1}, . . ., b_{\tilde{N}}, x_{1}, . . ., x_{N}) = {[\begin{matrix} g (w_{1} \cdot x_{1} + b_{1}) & \cdot \cdot \cdot & g (w_{\tilde{N}} {\cdot x}_{1} + b_{\tilde{N}}) \\ \cdot & \cdot \\ \cdot & \cdot \cdot \cdot & \cdot \\ \cdot & \cdot \\ g (w_{1} {\cdot x}_{N} + b_{1}) & \cdot \cdot \cdot & g (w_{\tilde{N}} \cdot x_{N} + b_{\tilde{N}}) \end{matrix}]}_{N \times \tilde{N}}

C) calculate output weights β:

T=[t ₁..., t _N] ^T, wherein It is the Moore-Penrose generalized inverse matrix of matrix H.

5. the method for claim 1, it is characterized in that: output valve is with o described in the step (4) _iExpression, i=1 wherein ..., m, m refer to the dimension of output valve o, and m output is namely arranged.

6. the method for claim 1, it is characterized in that: it is benchmark that accuracy requirement described in the step (4) can be adopted average relative error, for m output variable, average relative error

, wherein n is the test sample book number, and j is sample number, and i is the output variable numbering, and t is the output variable of expection; The average relative error result of test is compared with the relative error of expection, guarantee that maximum average relative error is less than or equal to prediction relative average error, i.e. precision of prediction in m the output variable

The wherein accuracy requirement of δ ' for setting.