TW200817891A - Adaptive multivariate fault detection - Google Patents

Abstract

A method and apparatus for detecting faults. A set of data samples is received, the set of data samples including multiple process variables. One or more multivariate statistical models are adapted, wherein adapting includes applying a change to at least one univariate statistic of the one or more multivariate statistical models if the change is greater than a threshold value. The one or more multivariate statistical models are used to analyze subsequent process data to detect faults.

Description

200817891 IX. Description of the Invention: TECHNICAL FIELD OF THE INVENTION The specific embodiments of the present invention relate to error detection, and in particular to error detection of a variety of misleading features. [Prior Art] Many companies use a device that includes multiple sensors and controllers. These sensors and controllers are carefully product quality during processing. A means of monitoring such multiple sensors and controlling statistical processing monitoring (a statistical analysis performed on sensor measurements and processing I), which enables self-detection error detection. A "fault" can be used to avoid an upcoming adjustment for the manufacture of an offset (eg, an indication of the prevention of a machine's operational parameters with a desired value. Therefore, one of the goals of statistical processing monitoring is Generates a predicate and/or detection error. I) During processing monitoring, when one of the most recently processed data or the second statistical model deviates by a quantity that is large enough to cause a threshold of another trust, the detection A mistake. A model quantity 'value' represents the deviation between the statistical features collected during the actual processing monitoring and the statistical features predicted by the model. The rigid quantity is the only mathematical method to eliminate this deviation. Commonly used squared prediction errors (Squared Prediction Error, SPE, Qres, or Q), and Hotelling's T2 〇 系 关于 # 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密 精密The fault or error), or the number of faults or misrepresented defects before the multi-statistical measurement of the super-measurement is a purely processed amount of data. Each model model measurement package its general reference 5 200817891 Each model measurement has an individual trust threshold, which is also referred to as a trust limit or control limit, where the value represents an acceptable upper limit for the model measurement. If a model measurement exceeds its individual trust threshold during processing monitoring, it should be inferred that the processing data has deviated from the threshold due to an error. The fact of an obstacle to accurate error detection is that the manufacturing process generally drifts over time, even without any problems. For example, this operation in half of the conductor processing chamber typically drifts between successive cleanings of the chamber and successive replacements of the consumed chamber components. Common statistical processing of error detection The monitoring method is subject to the drawback of distinguishing between normal drift and an error. In particular, some error detection methods use a static model that assumes that the processing situation remains constant over the survival of the tool. Such a model cannot be resolved between the expected changes in time and the unpredicted deviations caused by an error. To avoid the processing drift triggering many false alarms, this control limit must be set to a width sufficient to accommodate the drift. Therefore, the model cannot detect subtle errors.

Gallagher, Neal B. et al. "Development and evaluation of multivariate statistical process control tools for a semiconductor etch process: improving robustness through model updating" , ADCHEM 1 997, Banff, Canada; and Li, Weihua et al., "Recursive PCA for adaptive process monitoring", J. Process Control, Vol. 10, pp. 471-486 (2 0 0 0), each of which describes a method of drifting back to a condition that should be processed by cyclically adapting a model to 6 200817891 in the processing data. The Gallagher publication describes the adaptation of mean and covariance statistics. If more than one trust limit is measured for one of the modules Q or T2, Gallagher attempts to resolve the error and normal drift by identifying the occurrence of an error. The Li publication describes the number of principal components in the adaptation, skew variance, Principal Component Matrix, and principal component analysis (PCA) models. The adaptation method suggested by Gallagher and Li detects gradual errors.

Spitzlsperger, Gerhard et al., Tokyo (2004), ISSM, "Detection of a via etch process using adaptive multivariate methods", which reveals the use of human expertise. To adapt only the univariate means and the scaling coefficient that are expected to drift. However, by adapting only univariate means and scaled coefficients, this approach does not provide for the adaptation of such covariances between variables within a model. The various common adaptation methods described above are susceptible to cumulative computational rounding errors, which are caused by the periodic adaptation. This results in an inaccurate statistical value for the model that can cause both false alarms and faults to detect errors. SUMMARY OF THE INVENTION One aspect of the present invention relates to a method for detecting errors, comprising receiving processing data, the processing data including a plurality of processing variables, and adapting one or more multivariate statistical models according to the processing data of the 7 200817891, Wherein the adapting process includes applying a change to at least one singular statistic of the one or more multivariate statistical models, if the change is greater than a threshold, and adjusting the one or more A statistical model of variables is used to analyze the data of subsequent connections to detect errors. Another aspect of the present invention is directed to a machine-storable medium containing data that, when accessed by a machine, causes the machine to perform a method, the method comprising receiving processing data, the processing data comprising a plurality of processing numbers, And adapting one or more multivariate statistical models based on the processed data, wherein the adapting process includes applying a change to the at least one univariate statistic of the one or more multivariate statistics, if the change is greater than a threshold And analyzing the subsequent processing data using the one or more adapted multivariate statistical models to detect errors. Still another aspect of the present invention relates to a statistical processing monitoring system including a database for storing one or more multivariate statistical models and an error detector coupled to at least one manufacturing machine and the data The error detector is configured to receive processing material from the at least one manufacturing machine, wherein the processing data includes a plurality of processing variables to adapt at least one of the multivariate statistical models, wherein the adapting process includes applying a change to the At least one univariate statistic of one or more multivariate statistical models is obtained if the change is greater than a threshold and used to analyze the subsequent processing data using the one or more adapted multivariate statistical models Detect errors. The package change makes it possible to take the variable modulus value or change it to 8 200817891 [Embodiment] > Tian Shu is a method and device for detecting errors & "in the case of receiving processing data containing complex processing variables 2 Including the temperature and pressure decane flow (silane n〇w), etc., the first δ-ten model is adapted according to the processing data. The modified text will not exceed the ~ threshold value, and the change is applied to , and at least a single variable is counted. In a specific implementation of f multiple processing variables - measured drift and pre-determined. The adapted multivariate statistical model can then be processed by the subsequent detection of each error. In the following description, a number of details are presented. However, it is to be understood that the present invention may be embodied in the following specific details, and the details of the structures and devices are in the form of a block diagram detail in order to avoid obscuring the invention. It is represented by an algorithm and a symbolic symbol in a computer memory bay. The familiar data processing artist uses the narrative and presentation of these algorithms to give the other skilled artisans the most efficient nature. The method, the steps of the steps or instructions that lead to a self-consistent result, are those that require physical manipulation of physical quantities. Although necessary, these quantities are usually stored in a computer system. Electrical, magnetic forms that have been combined, compared, or otherwise manipulated. These signals have been shown to be represented as bits, values, components, etc. In a processing 〇-^ or a suitable package, the interval It is used to divide the implementation of the technique. The technique of displaying the intermediate arithmetic is often used to convey the order. It is not the symbol of the signal, the specific real variable. Multiple and many if the amount of statistics based on an implementation of the analysis for the artist will be specific (not f ί) 9 200817891 characters, terms, numbers count Ronggu! ^ Μ, the temple is very convenient, mainly based on common usage The factors. However, it should be noted that the terms and similar terms are related to the appropriate physical quantity, and only apply to these convenient labels. Unless otherwise stated, they can be in _ ^ a. It is clear from the metaphor that the terminology used in the text is "processing", "calculation", "蚪管★"冲#" or "decision" or "display", etc., which represents a computer system 哎 _ Μ , . The operation and processing of the electronic computing device in the meta-cheek-like electronic device, which is manipulated and converted into the physical information in the register and memory of the Raytheon anger. & is the physical quantity in the memory of the computer system, or the temporary thief or its information storage, transmission or display device. This month is also a device for performing the operations described herein. The device can be selectively activated or re-sent in response to a demand, or can be a computer that is stored in the computer. This computer program

C, 1 is stored on a computer readable medium such as, but not limited to, any type of magnetic disk including a floppy disk, a compact disk, a read-only memory (CD_R0MS) and a magnetic disk drive, a read-only memory (RQMS) , random access memory ports (RAMs), erasable stylized read-only memory (EPROMs), electronic sigma-translated private read-only memory (EEpR〇Ms), magnetic or optical cards or the like Media, which is suitable for storing electronic instructions and each computer is equipped with a computer system bus. The algorithms and modules described in this section are not related to any other device in any particular computer. Various general-purpose systems may be adapted to construct more specialized equipment to perform the steps of the method in accordance with the teachings of the present invention, or one or 5:3'. The architecture required for these various systems will be detailed below.

200817891 stated. In addition, the present invention is not limited to any particular programming language. The invention can be implemented by a machine readable medium embodied in a machine readable medium for storing or transmitting information. For example, a machine includes a machine readable storage medium (eg, a read only memory (memory access memory (RAM), a disk storage medium, an optical flash memory device, etc.), a machine readable transmission medium (f audio) Or other forms of propagated signals (e.g., carrier waves, red digit signals, etc.), etc. The following description provides details of a unified system for monitoring operational and/or error detection (unstable manufacturing processes) on a manufacturing device. In a specific embodiment, the statistical processing is used in the manufacture of electronic devices, such as semiconductors. Manufacturing requires many manufacturing steps, sputtering, and chemical vapor deposition involving different types of manufacturing processes to be performed in three different types. On different types of machines. In addition, this statistical process is used to monitor the manufacture of other products (such as cars). This other also requires a lot of different processing handled by various manufacturing machines. Figure 1 depicts the statistical processing monitoring system 1 00 One of the statistical processing monitoring systems 100 includes a statistical process to monitor its via the data communication link 1 60 with one or more manufacturing The plurality of processing controllers 150 are coupled. The statistical processing supervisor can be included in a factory (e.g., a manufacturing plant). Alternatively, the statistical processing monitoring system 100 can include a description of the work. The device readable medium in the form: ROM), the storage medium, the gas, the optical, the external signal, the processing device for processing the monitoring and monitoring system. For example, the management system is the manufacturing step of the product. Body embodiment. The apparatus 105, 110 and the control system 100 have manufacturing machines 11 in the manufacturing plant, for example, all of the manufacturing machines 110 that can be operated on one or more specific processes.

C

In one embodiment, each manufacturing machine 110 is a machine that manufactures electronic devices, such as an etcher, a chemical vapor deposition furnace, a photolithography device, an implanter, and the like. Alternatively, the manufacturing machine 110 may be of a type that manufactures other products, such as automobiles. In a specific embodiment, each of the manufacturing machines 110 can be of a single type. Alternatively, the manufacturing machine 110 can include a number of different types of equipment, each of which can perform different processes. Each manufacturing machine 110 can include multiple sensors for monitoring operation on the manufacturing machine 110. One type of sensor included in the manufacturing machine π 可 may be a temperature sensor. Examples of other sensors include a pressure sensor, a flow rate sensor, or any other sensor that monitors the physical properties of a working component made by the manufacturing machine 110 or the physical condition of a manufacturing process. Each manufacturing process performed on a manufacturing machine 110 is characterized by various physical conditions and attributes detected by the sensor and various operational parameters that are collected to correlate as processing data. Each distinct physical condition or attribute detected by the sensor, as well as each operational parameter, can be a distinct processing variable of the processed data. Examples of process variables representative of detector data include process chamber pressure, susceptor temperature, RF forward power, and RF reflected power. Examples of process variables representing operational parameters include flow rate settings (e.g., chemical reagents) and (for a chamber exhaust vacuum pump) section 12 200817891 flow valve settings. The sensor, manufacturing machine, and process controller can be monitored during the process to collect the process variables in a timely manner at successive points. In one embodiment, each process variable is applied to a particular location, and one or more process variables can be applied to a particular process. In a specific embodiment, the sensing and operating parameters of the different steps in a process represent distinct processing variables (modeled as additional dimensions of the model space). This can be useful, for example, if multiple steps are performed at a manufacturing location in a machine with different operating parameter settings. Example In a three-step manufacturing process, one of the susceptor temperatures during the three-step process is considered to be three distinct processing variables. It may be beneficial to have individual scales of the processing steps, such as when multiple layers are deposited on a single processing work piece, or when different steps of a process are exposed to different processing conditions (eg, pressure, temperature) and many more). The process controller 150 controls the operational parameters of the manufacturing machine 110. The ί processing controller controls the process chamber temperature of the manufacturing machine, the vacuum, the gas injection system, and so on. The process controller 150 can store a one or process recipe 160. Each recipe 160 defines the operational parameters of the machine 110 on each step of the process. In a specific embodiment, recipe 1 60 is loaded into manufacturing machine 110 by process controller 150. The data communication link 1 60 can include a common communication link and it is wireless or wired. The data may be passed in the pure (r a w ) or processed format at the machine 110, the processing controller 150, and the statistical processing monitoring set. In one embodiment, a Semiconductor Equipment Standard (SECS) interface can be used. In other embodiments, a process can be used. In the middle of a measuring device, the degree will be divided into a single 如J, such as an air pump, a plurality of manufacturing can be manufactured, or a communication model, a high-speed SECS message service (HSMS) interface, etc. Wait. The statistical processing monitoring device 105 can be a single server for analyzing incoming processing data from the manufacturing machine 110, the sensor 155, and the processing controller 150. Alternatively, the statistical processing monitoring device 1 〇 5 may comprise multiple servers and/or computers. In one embodiment, the statistical processing monitoring device 105 includes an error detector 1 25, an error detector 130, and an error reporter 150. The statistical processing monitoring device 1〇5 also includes a storage device 175. In a specific embodiment, the statistical processing monitoring device ι〇5 is included in one or more processing controllers 150. Alternatively, the statistical processing monitoring device 105 can also be a distinguishable and/or independent device. The storage device 175 includes a processing measurement database ι2, one or more multivariate statistical models 135, error characteristics 14A, and error classifications 145. In a specific embodiment, the storage device 175 is in the embodiment, and the particular device of the storage device includes: ...a single storage device for a computer or a feeding device..., the storage: Name 175 can be external to the statistical processing monitoring device 1〇5. In a specific 175 comprising multiple storage devices

135. Once the multivariate statistical model model 1 3 5 , the storage can be stored in the processing measurement data processing data can be used to display the processing for the manufacturing machines running on the specialized manufacturing machine, etc., the Store the processed data with a multivariate statistical model of 丨3 5 . One can be stored in the storage device 175 14 Γ

200817891 In a specific embodiment, a training session is used to collect data that produces a plurality of multivariate statistical models. The training session includes one of a collection of processing operations on a particular machine that is known and/or controlled to accomplish one of the specific manufacturing operations. The processing data received from the processing run during the training period can be used to generate statistics (e.g., meail, variable arrays, etc.). The statistics are collectively used to generate one or more variable statistical models that are generally directed to a particular process on a particular machine. Each multivariate statistical model 135 includes one or more model measures. The modulo is measured as a scalar value' which characterizes a set of processing data and a shift between the models. In a specific embodiment, the model metering includes a Square Prediction Error (which is generally referred to as SPE, Qres, Q), and Hotelling's T2. Model metrology also includes combined measurements (such as combined multivariate indexing (CMI)). Each of these measurements corresponds to a different method of calculating the probability that the processed data being monitored has the same statistics as the training resources used to establish the model. The above statistics and measurements can be calculated based on general statistical algorithms. Converting an M-dimension handles the variable space to the main N-dimension two between each other' where Μ is the number of processing variables, and n is M. The PCA calculates a set of Μ eigenvectors (M eigenvect〇rs) and eigenvalues, where each eigenvector transforms the number of data to an individual dimension of the main component space, and each feature number is often compared to a corresponding The variable represented by the feature value. In order to simplify the space of the component (reducing the dimension of the main component space), the main component analysis can be utilized for the most or more multivariate models corresponding to the orbital deviation or the estimated roots ( 15 200817891 The N feature vector of the large feature value is held in the model: the other feature vector is discarded or ignored. The number N of main components held in the model is a parameter of the model selected by the user. The number of major components (n) may be selected based on a model that interprets less data when a smaller value N is used and when a larger value N is used than is exceeded by a specified model. Once one or more multivariate statistical models have been generated, they can be used by the error detector 125 to monitor the processing run on the manufacturing machine 1 1 . The error detector 1 25 performs various statistical processing monitoring methods. And analyzing the processing data, each of the methods is based on at least one variable statistical model. In a specific embodiment, the error detector 1 2 5 receives directly from The manufacturing machine 110, the sensor 155, and/or the processing controller 150 processes the measurement data (processing data). In another embodiment, the error detector 1 25 can receive processing measurements. In another embodiment, the error detector 125 receives processing data from the source of the two. (J In order to detect an error, the error detector 1 25 calculates Statistics of the processed data of each process being monitored, and comparing the calculated statistics with the corresponding statistics of the appropriate multivariate statistical model. The statistics are compared for a model measurement or for multiple model measurements (eg, T2, SPE, CMI) If one or more of the models are measured beyond a predefined threshold (referred to as a trust limit or control limit), an error can be detected. In a particular embodiment, each model meter has a threshold for the user to select. The value of the threshold can be used to indicate the risk of an error warning (if the threshold is too low) and the risk of not detecting an error (if the threshold is too high) Eclectic 0, the mid-sum re-measurement is different, if any 1 j of the measurement exceeds the threshold value, an error occurs. In addition, only if the specific measurement T-heart T has exceeded the threshold value or only if the multiple measurement exceeds the threshold The value may indicate that a particular error has been identified by the error detector 125, and the error is analyzed by the error detector. The error detector 130 compares the error with one of the error signatures. Each error signature represents one (each ) 曰, , , and j specific errors

U

The situation. In a specific embodiment, the error signature 丨4 〇 A 1 and the hummer have an aligned list of processing variables for a particular error-to-large statistical contribution. Error detection 1 30 can compare the permuted list of individual stored error signatures with the respective processing variables having the greatest contribution to the current error. Tianzi Changyu Temple

When a high level similarity between one of the error characteristics 140 and the current error is reported, a match is derived. Each of the X error characteristics 140 is associated with one or more error classifications 145. The error classification 145 may indicate the cause of the error - the actual problem or the possible cause of the current error. For example, if the error signature indicates that the maximum feed processing variable is the decane flow rate, the misclassification may indicate that a value that feeds decane into a processing chamber has been erratic. The error report 165 generates an error report indicating which of the error categories 145 corresponds to a current error. The error report can be transmitted to one or more clients that are networked to the statistical processing monitoring device 105 (not shown, and can be, for example, a local computer, a remote computer, a personal digital assistant (pDAs), a pager, a mobile phone, etc. Wait). Error reporting 1 65 can also cause the manufacturing machine to shut down, warn a machine, or perform other appropriate actions. 17 200817891 Even in the absence of errors, manufacturing processes typically drift over time, and this operation in a semiconductor processing chamber typically drifts between successive cleaning of the chamber and continuous replacement of depleted processing chamber components. By processing the drift, the changes in the processing variables caused by the drift are not erroneously translated into errors. Figure 2 depicts a flow chart of a particular implementation of the method of error by adapting one or more multivariate statistical models. The method can be performed by a process, which can include hardware (e.g., circuitry, dedicated logic logic, microcode, etc.), software (e.g., instructions on a processing device), or a combination of the above. In a specific embodiment, the method can be performed by the statistical processing monitoring device 105 of Figure 1. Referring to Figure 2, method 200 begins with receiving processing data logic (block 2 1 0). The processing data can be from a process running on a manufacturing machine and can include multiple processing variables. At block 215, the processing data is determined to determine if an error is for one or more multivariate systems. In the particular embodiment described, the processing data is analyzed to detect an error prior to execution. Alternatively, the data can be analyzed after adjustment. When the processing data indicates that one of the multivariate models or one of the multiple measures (e.g., T2, SPE, CMI, etc.) has exceeded, an error is indicated for one of the multivariate statistical models. In one embodiment, two or more multivariate statistical models are used for error detection. If at least one of the models is wrong, an error can be identified accordingly. Even if the error is not recognized, the processing can be directly detected by the logic, and the analyzer on the device running 200 can perform the statistical threshold and the system detection can be transmitted to a use. For example, if one model detects an error and the other model does not. In addition, unless at least two models indicate a possible error, an error will not be reported. In a specific embodiment, two or more multivariate statistical models distinguish one from the other in at least one mode. For example, a different type of processing data may be used, or a different amount of processing data may be used, for example, a first model containing a model containing a newer generation may include only the last 100 wafers. The method of drifting is used to distinguish (the statistics are adapted and which ones are used for further explanation. The model is not adapted to drift, and to drift. The multivariate statistical model can also be transformed. The multivariate statistical model Each of the different training materials or additional types used simultaneously during a training session is included in the design test or from the inclusion of additional operational criteria, as described above with reference to Figure 2, in the block.

ϋ F the same number, use different trust restrictions, maintain the model and distinguish each model. All processing data processed by the example, a second preventive maintenance (preventive all processing data, and a third model can also be adapted by using different adaptations to handle the variables, which are suitable, etc.), reference block 220 In an embodiment, at least one multivariate less than one multivariate statistical molding is adapted to be established in one of the modes not mentioned above using a single set of training material. In addition, different models can be used for training materials. For example, if a modulo variable 'or a statistical model requires additional information for a longer training period to be desirable. In 220, the processing logic determines that the drift is 19 200817891 No is measured for one or more processing variables. If a particular statistic (eg, median, standard offset, etc.) of a processing variable has been gradually adjusted, the drift is measured. If the measurement does not drift, the method ends. If the drift is detected for one or more processing variables, the method proceeds to block 225. In another embodiment, the method proceeds to block 2 2 5 regardless of whether drift has been measured. f

In block 225, processing logic determines if a predetermined condition has occurred and is then indicated as the adaptation trigger. In a specific embodiment, the adaptation trigger includes a specific time interval (e.g., one hour, one day, etc.). The adaptation trigger includes a specific number of processing runs, a predetermined number of data samples, and the like. The adaptation trigger can occur once the particular number of processing runs are completed, the predetermined number of data samples are generated, and the like. The one or more adaptation triggers can be combined to cause an adaptation to be performed, e.g., if a predetermined number of data settings are generated, or if one of the prior adjustments has expired. If the adaptation trigger does not occur, the method ends. If the adaptation trigger has occurred, the method proceeds to block 230. In block 230, one or more of the multivariate statistical models are adapted. Different algorithms can be used to adjust to drift, and one of the algorithms includes an exponentially weighted moving average (EWMA). Other suitable adaptation algorithms include the forgetting factor and windowing; 5, the use of & recursive moving average. Also fly the algorithm. ^ Other Tuning In a specific embodiment, all of the 20 200817891 statistics for each of these processing variables are adapted. The specific statistics for the 'specified processing variables not being adapted and/or over one year old processing variables will not be adapted. In a specific embodiment, processing logic

One or more of the first group of processing variables, each of which is expected to drift in the normal operation of the L manufacturing machine. The first set of processing variables are adapted in the model while maintaining all other processing variable statistics. The processing logic can also identify one or more of the second set of variables, each of which is pre-stated to remain unchanged without errors. All of the processing variables outside of the second group can be adapted in the model while maintaining the processing variables outside the second group. In a specific embodiment, the processing variables of the first group and the processing variables outside the second group can be used together. This will allow the processing logic to detect both gradual faults and sudden faults, and avoid erroneous adaptations to a gradual error. One or more statistics can be adapted for each adapted process variable. Examples of statistics that can be adapted to handle variables include mean, variation, covariance, correlation (e.g., associative array), major component feature vectors and feature values, and the number of major components. In a specific embodiment, different statistics can be adapted for different processing variables. The statistics can be adapted to the processing variables (if any) based on user input, or automatically without user input (e.g., based on a selection algorithm). For example, the number and variation of a particular process variable may be expected to drift, while the covariance of such process variables with other process variables may not be expected to drift. Thus, the number and variation of the appropriate processing variables can be drifted while the covariance statistics between the processing variables and other processing variables remain fixed. In other examples, for all other processing variables, all statistics 21 200817891 can be expected to drift. Therefore, all statistics for those other processing variables can be adapted. ''

ίJ Referring now to Figure 2, in block 2 3 5, the processing logic determines whether the adaptation will change one or more univariate statistics by at least one threshold, and then refers to the terminology. When adapting a multivariate statistical model, having cumulative computational rounding errors will result in the risk of making erroneous changes to the calculated value of the univariate statistics that have not actually changed. Thus, the accumulated error in the univariate statistic can & be a disproportionate (dlspr0p〇rti〇nate) error in the calculation of the multivariate statistic (e.g., covariance). To improve such an occurrence, if the adaptation will not change the one or more univariate at least the adaptation threshold, the method proceeds to block 245. If one or more univariate statistics are to be changed the adaptation threshold ’ value, the method proceeds to block 240. In one embodiment, the adjustment threshold for each individual univariate statistic is a fixed value. ...the adjustment threshold for a single variable statistic is a corresponding value, such as a predetermined fraction of a current value. For example, in a specific embodiment φ # 曰,, 版 々 甲 ,, the individual univariate statistics of the adjustment threshold may be one billionth of the current value of the individual 1 (Γ9). One or more single variables can be shared by the same adjustment threshold. In addition, the single variable, the specific 岑4 owner has their own adjustment threshold. In block 240, The univariate statistics that change the amount greater than or equal to the value of the adjustment threshold are changed. The adjustment of the univariate statistics that does not change the iron threshold is adjusted until the change is true. 22 200817891 The adapted (various) multivariate statistical model is used to analyze the next processed data to detect errors at block 24 5. The method then ends. In a specific embodiment, the recent The processing data is used to adapt a multivariate statistical model after the processing data does not deviate enough to indicate the wrong model. Alternatively, the most recent processing data can be used to adapt before performing error detection. In other embodiments, error detection is performed twice on the processed data, once before the adaptation and once after the adaptation, according to the method described in Figure 3 below. Figure 3 depicts a detection error. A flowchart of a specific embodiment of method 300. The method can be performed by processing logic, which can include hardware (eg, circuitry, dedicated logic, programmable logic, microcode, etc.), software (eg, in a processing device) The above-running instruction), or a combination of the above. In a specific embodiment, the method 300 can be performed by the statistical processing monitoring device 105 of Fig. 1. Referring now to Figure 3, the method 00 begins with the processing of the received processing data. Logic (block 3 1 0). In block 315, a first error detection algorithm is applied to the processing data. In a specific embodiment, the first error detection algorithm is relatively loose (not The sensitive error detection threshold. In block 320, the processing logic determines if an error is indicated. In a particular embodiment, if an error is indicated, the method proceeds to block 3 3 5 And if no error is indicated, the method proceeds to block 3 2 5. In an alternative embodiment, the method proceeds to block 325 which indicates if an error is indicated. In block 335, an error is reported. An error may include notifying a user by transmitting a message to a client via 23 200817891, issuing a warning sound on the manufacturing machine, stopping a process, etc. The method then ends. In block 3 2 5, one or more A multivariate statistical model can be adapted. Next, a second error detection algorithm can be applied (block 3 30). In a specific embodiment, the second detection algorithm uses the adapted model to determine a Whether the error has occurred. In a specific embodiment, the second error detection algorithm uses a correspondingly rigorous (sensitive) error detection threshold. The use of two error detection algorithms reduces the chance of triggering false alarms and increases the chances of detecting actual errors. This operation in a semiconductor processing chamber typically undergoes a sudden shift after machine repair, correction maintenance (eg, replacement or classification of a component), or prevention of maintenance (eg, cleaning of a processing chamber), the owner of which is collectively referred to For machine maintenance. To avoid this being identified as a sudden shift in error, it is intended to "reset" all or part of the model after machine maintenance. Figure 4 depicts a flow diagram of one embodiment of a method of detecting an error by resetting one or more statistical models after machine maintenance. The method can be performed by processing logic, which can include hardware (e.g., circuitry, special purpose logic, programmable logic, microcode, etc.), software (e.g., instructions that are executed on a processing device), or a combination of the above. In a specific embodiment, method 400 can be performed by statistical processing monitoring device 105 of FIG. Referring to Figure 4, method 400 begins by detecting machine maintenance (block 405). This machine maintenance can be detected for a particular manufacturing machine or for multiple manufacturing machines. In block 41 0, one or more multivariate statistical models are automatically reset. A model reset can be automatically initiated when a value of a process variable is changed in a manner that implies that 2008 2008891 machine maintenance has been performed. Examples of such process variable changes include a counter that is reset to zero (eg, indicating a component replacement), an out of service that is longer than a predetermined time period, or a particular processing set point value. change. Alternatively, one of the one or more multivariate statistical models can be manually initialized by a user, such as when a manufacturing machine is replied to operate after machine maintenance. In a specific embodiment, the model is reset via a reply to a state that coincides with one of the original states produced when a training session is completed and before any adjustments are performed (block 4 15). In another embodiment, the model is reset by adapting all or part of the model to a new operational condition (block 420), as described above with reference to Figure 2. For example, statistics for selected processing variables can be adapted to reflect this new operational situation. In a particular embodiment, the processing variables selected for adaptation include those variables that are expected to change because a particular type of maintenance actually performs. Alternatively, the processing variables selected for adaptation may be those having a maximum error contribution and/or an error contribution greater than a threshold value. The calculation may be based on the error statistics generated by the processing data generated after maintenance of the machine to the model as it existed prior to maintenance of the machine. In one embodiment, a number of processing variables selected for adaptation is repeatedly incremented until one or more model error statistics fall below a predetermined threshold. In a specific embodiment, resetting the multivariate statistical model includes initializing a reset training session (block 42 5 ). From the reset training session 25 200817891 The data can be used to recalculate all or part of the multivariate statistical model. In a specific embodiment, the reset training session uses actual processing from a product on a manufacturing machine. Processing data. In a specific embodiment, the variable error guess is disabled during the reset training period. This will avoid. Many false alarms occur. In addition, multivariate error detection is effective for a particular misclassification and/or error signature during the heavy training session. Therefore, errors that may be false alarms can be suppressed, but actual errors can still be monitored. The reset training session is ended once sufficient processing data has been collected to recreate at least one multivariate statistical model. In one embodiment, when the processing data indicates that a new multivariate model has converged to a stable set of statistics, the reset training period ends' and error detection resumes. This will occur when the specific statistics for the new variable statistical model are changed by less than the threshold value by introducing new processing data. Alternatively, the reset training session can be completed when a predetermined number of training material samples are generated by the manufacturing machine and when the new model is incorporated. In another embodiment, the reset training time U may end when one or more model error gauges are lowered by a predetermined threshold when comparing the new model with the processing data. In yet another specific example, the reset training period may end when the ratio of one or more model error statistics exceeds a predetermined threshold interpolation by a predetermined threshold when comparing the new model with the training capital . In yet another embodiment, the reset training period may end when a frequency in which the selected processing variable is different and whose median value is greater than a frequency deviation of their target deviation by a predetermined threshold. The multi-displacement system is used to apply a first manufacturing machine. In a specific embodiment, the technique described below with reference to resetting a multivariable type is used to apply a first manufacturing machine. One of the current statistical models to a second manufacturing machine. To convert a multivariate statistical model from the first device to the second manufacturing machine, a copy of the multivariate type is generated and associated with the second manufacturing machine. In an embodiment, an initial of the model for the second manufacturing machine is a current state of the model on the first machine. An adaptation and/or training session can then be initialized to adapt the model to the second machine. Figure 5 depicts a graphical representation of an exemplary form of computing system 500 in which a set of instructions is executable and used To cause any one or more of the methods discussed herein. In alternative embodiments, the machine can be connected (e.g., networked) to a regional corporate network, or to other machines in the Internet. The machine can operate as a client-server network environment or a client interception operation is one point in a peer-to-peer (or decentralized) network environment (machine. The machine can be a personal computer, desktop computer A set-top box (STB), a personal digital assistant PDA, a mobile circuit application, a server, a network router, a switch, or a bridge can perform a set of instructions that specify actions taken by the machine (non-sequence Further, when describing only a single machine, a "machine" should also be employed to include an individual or connected (multiple sets of) instructions to perform any one or more of the methods described herein. The example computer system 5 00 includes a processing device (at the statistical model storage variable manufacturing machine statistical mode, a specific state resetting training machine, the machine specific real network, making a force, or peer) On-board box, network, or any sequence or any set of items in the project line) 27 200817891 502, main memory 5〇4 (such as read-only memory, flash memory, dynamic random access Recall (such as synchronous dynamic random access memory, or Rambus dynamic random access memory), static memory 5〇6 (such as flash memory, static random access memory, etc.), and data The storage device 518 communicates with the other through the bus bar 530.

Processor 502 represents one or more processing devices (e.g., microprocessors, central processing units, etc.) that are generally intended. In particular, the processor 502 can be a Complex Instruction Set Computing (CISC) microprocessor, a Reduced Instruction Set Computing (RISC) microprocessor, or a Very Long Instruction Word (VLIW) microprocessor, or can be implemented The processor of other instruction sets may be implemented as a processing unit of a combined instruction set. The processing benefit 502 can also be one or more specific intended processing devices (eg, an application specific integrated circuit (ASIC)), a field programmable logic gate array, a digital signal processor, Network processor, or similar). The processor 502 is configured to execute processing logic 526 for performing the operations and steps described herein. The computer system 500 further includes a network interface device 5〇8. The computer system 500 also includes a video display unit 51 (such as a liquid crystal display (LCD) or cathode ray tube (CRT)), alphanumeric input devices 5 12 (such as a keyboard), and indicator control devices (such as a mouse). And a signal generating device 5 16 (for example, a speaker). The data storage device 518 can be used to store a storage medium 5 3 1 , and can store one or more groups of any one or more of the methods described above. Or functional instructions. The software 5 2 2 汝 丄 丄 也 也 can also be fully or at least partially resident in the main memory 5 〇 4 and/or the processor 502 via the computer system 5 〇〇 28 2008 17891 The main memory 5〇4 and the processor 5〇2 may also constitute a machine-accessible storage medium. The software 522 is further transmitted and received over a network 52 through the network interface device 508. The machine accessible media 531 can also be used to store a set of body structures that define the user's usage status and user preferences for defining the user directory. The data structure set and user preferences can also be stored in other sections of the computer system, such as static memory 506. When the machine is accessible to the farmer's scale, it is shown as a single-media 531, in an exemplary embodiment, _ "project machine accessible storage medium" should be employed to include Store in a + 々 heavy media ί^ Save one or more sets of instructions for a single media or Dolly Green® C such as a centralized Ο and feeding device). The item "machine" is used to store, compile, or retrieve media, and should also be employed by one or more of the present invention. And the media, the body should be used to package people / the project "machine accessible storage media and magnetic media, and carrier signals:. t is not limited to) solid state memory, optics should be able to understand the above limitations. Many other and exciting intentions are not cited as gamma VIII...~Dan body implementation ratio and understanding of the above description and then the actual (3) can be read in the skilled person to read the following patent application y乍. Accordingly, the scope of the invention should be construed as being limited by the scope of the claims, and the invention also includes all the equivalents. BRIEF DESCRIPTION OF THE DRAWINGS 29 200817891 FIG. 1 is a specific embodiment of a statistical processing monitoring system; FIG. 2 is a flow chart showing a specific embodiment of a method for detecting errors by adapting one or more variables. Figure 3 depicts a flow diagram of a specific embodiment of a method of detecting an error; Figure 4 depicts a flow of a specific embodiment of a method for detecting an error by resetting one or more statistical models after machine maintenance Figure

Figure 5 depicts a pictorial representation of a machine in an exemplary form of computing system having a set of instructions executable to cause the machine to perform any one or more of the methods discussed herein. [Main component symbol description] 11 0 manufacturing machine 1 5 5 sensor 1 7 0 recipe 150 processing controller 160 data communication link 125 error detector 1 3 0 error detector 165 error reporter 120 processing measurement database 1 3 5 multivariate statistical model 140 error feature 145 error classification 30 200817891 175 storage device 1 0 5 statistical processing monitoring device 2 1 0 receiving processing data 2 1 5 determines whether the processing data indicates whether the error 220 drift is measured for the processing variable? 225 Predetermine whether the interval is reached? 230 Adapting Multivariate Statistical Models 23 5 Does the adaptation change the univariate statistics by at least one threshold? 240 Change univariate statistics 245 Use adapted multivariate statistics to analyze continuation processing data 3 1 0 Receive processing data 3 1 5 Apply the first error detection algorithm 320 Is the error indicated? 325 Adaptation Multivariate Statistical Model 3 3 0 Apply Second Error Detection Algorithm 335 Report Error 400 Detect Machine Maintenance 4 1 0 Automatically reset multivariate statistical model 4 1 5 Reply multivariate statistical model to a state consistent with an original state 420 Adapting the multivariate statistical model to the new operating condition 425 Initializing the reset training period 502 Processor 5 2 6 Processing logic 5 04 Main memory 31 200817891 522 Software 5 06 Static memory 508 Network interface 5 20 Network 5 1 0 video display 512 letters and numbers 5 1 4 index control device 5 1 6 signal generating device 5 1 8 auxiliary memory 5 3 1 machine accessible 522 software input device to set the storage medium 32

Claims (1)

200817891 X. Patent application scope: 1 · A method for detecting errors, comprising: receiving processing data, the processing data comprising a plurality of processing variables; adapting one or more multivariate statistical models according to the processing data, wherein The process includes applying a change to the one or more multivariate statistical models to: less than a single variable statistic, if the change is greater than a threshold; and () using the one or more adapted A multivariate statistical model is used to analyze subsequent processing data to detect errors. 2. The method of claim 1, further comprising: determining whether the processed data indicates the target before adapting the one or more multivariate statistical models or after adapting the multivariate statistical models One of the one or more multivariate statistical models is wrong. 3. The method of claim 1, wherein the one or more multivariate statistical models are adapted based on a measured drift of one or more of the processing variables of the complex number. 4. The method of claim 1, wherein a processing variable from a first subset of the processing variables of the complex number is adapted, and wherein a processing variable from a second subset of the processing variables of the complex number Not adjusted. 33. The method of claim 1, further comprising: determining, by applying a first error detection algorithm, whether the processing variable is indicated before adapting the one or more multivariate statistical models An error; and after applying the second error detection algorithm, after adapting the one or more multivariate statistical models, determining whether the processing data indicates an error, wherein the control restriction of the first error detection algorithm is Widely controlled by the second error detection algorithm. 6. The method of claim 1, wherein the one or more multivariate statistical models comprise at least a first model and a second model, a quantity of historical data considered, and processing used The variable, the number of major components used, the trust limit, the forgetting factor, one of the methods of adjusting to the measured drift, the adapted processing variables, the training data used to generate the model, and the adjustment threshold are at least one of The second model is different from the first model. 7. The method of claim 1, further comprising: automatically resetting at least one of the multivariate statistical models upon detecting machine maintenance on a tool associated with the processing data. 8. The method of claim 7, wherein the resetting process comprises recalculating statistics for processing variables having an error contribution greater than one threshold, adapting at least a portion of a multivariate statistical model to a new one. 200817891 Operational situation, and replying to the multivariate statistical model to at least one of a state consistent with one of the original states produced during the time period. 9 - The method of claim 7, wherein the reset is over initialized The training session is reset, wherein the reset training period is aggregated to update at least one of the multivariate statistical models. The method of claim 1, wherein the adapting comprises applying a change to an associative array, one of the main components, a loading vector, a mean, a variation, and a main At least one of a component feature vector and a feature value. 1 1. A machine-accessible medium containing data, which, when executed by a machine, causes the machine to perform a method, the method comprising: receiving processing data, the processing data comprising a plurality of processing variables adapted according to the processing data a plurality of multivariate statistical models, the process comprising applying a change to the one or more multivariate statistical modules and less than a single variable statistic if the change is greater than a threshold; and using the one or more adapted The multivariate statistical model is used to process data to detect errors. 1 2 . The machine-accessible media as described in the scope of claim 1 of the patent application includes a training process for receiving more sub-objectives, negatives, and access by the co-party. The party 35 200817891 method further comprises: determining whether the processing data indicates an error for the one multivariate statistical model before adapting the one or more multivariate statistical models or after adapting the statistical model of the variable. 1 3 - The machine-accessible medium of claim 11, wherein the one or more multivariate statistical models are adapted based on a measured drift of one or more of the processing of the plurality. 14. The machine-accessible medium of claim 11, wherein the method further comprises: determining, by applying a first error detection algorithm, whether the processing variable is indicated before adapting the one or more variable statistical model An error; by applying a second error ^[贞 演 algorithm, after adapting the one or more variable statistical model, determining whether the processed data indicates an error, the control restriction of the first error detection algorithm is wider The control limitation of the second error algorithm. The machine-accessible medium of claim 11, wherein the one or more multivariate statistical models comprise at least a first model, a second model, and a quantity of historical data considered, used The number of variables, the number of major components used, the trust limit, the remains, the method of adapting to one of the measured drifts, the adaptation, and so on, and more or more of which are more than one of the parties. The second model is different from the first model in that it is at least one of the training data generated by the model, the training data used for the model, and the adjustment threshold. 16. The method of claim 11, wherein the method further comprises: automatically resetting the plurality of machines when detecting machine maintenance on a tool associated with the processing material. At least one of the variable statistical models. η 17. The machine-accessible medium of claim 11, wherein the adapting process further comprises applying a change to an associative array, a number of primary components, a loading vector, a median (mean) At least one of a change, a covariance, a main component feature vector, and a feature value. 1 8. A statistical processing monitoring system, comprising: an Ij database for storing one or more multivariate statistical models, and an error detector coupled to at least one manufacturing machine and the database, The error detector is configured to receive processing data from the at least one manufacturing machine, wherein the processing data includes a plurality of processing variables to adapt at least one of the one or more variable statistical models, wherein the adapting process includes applying a change to the At least one univariate statistic of one or more multivariate statistical models, if the change is greater than a threshold, and for analyzing the continuation using the one or more adapted multivariate statistical models Processing data to detect errors with 37 200817891. In the statistical processing monitoring system described in claim 18, the error detector is further used to adapt the one or more multivariate-models or after adapting the multivariate statistical models. Determining whether the material indicates that the one or more multivariate statistical models in the statistical processing monitoring system described in claim 18 is based on the One or more of the plural number is adjusted as soon as the measurement drifts. 2 1 · The error detector is further used to automatically reset the multivariate statistics when detecting machine maintenance in the at least one manufacturing machine, as in the statistical processing monitoring system described in claim 18 One of the models. U, its statistical assets. Error. , its rational change, at least 38 on its