KR20170035806A - Method and system for root cause analysis of deviation associated with equipment - Google Patents

Abstract

The present invention discloses a system for analyzing a basic cause of deviation related to equipment. The system includes a sensor resolution action (SeRA) entity for detecting deviation related to the equipment. A deviation resolution graph (DRG) entity determines a deviation resolution sub-graph (DRsG) for solving deviation based on the DRG. The DRsG includes an observation related to the solution task and the equipment. The recommendation entity recommends the DRsG to solve the deviation related to the equipment.

Description

TECHNICAL FIELD The present invention relates to a method and system for root cause analysis of equipment-related deviations,

The present invention relates to root cause analysis and more specifically relates to a mechanism for root cause analysis of deviations related to equipment.

Within a fish facility / manufacturing facility, many tasks are controlled by various inputs (e.g., data) generated from sensors installed in process equipment and building facilities. This data includes both structured and unstructured types. These tasks can perform a number of functions including process optimization and identification of unusual events and disturbances in the operation of a production facility / manufacturing facility.

Within production facilities / manufacturing facilities, events and disturbances of various sizes continue to affect process operations. The events and disturbances are handled by root cause analysis of deviations related to the process. Currently, these events and disturbances are being handled by various systems and methods. However, while conventional systems and methods are somewhat effective in root-cause analysis of deviations associated with equipment, they are not as effective as memory power, loss of contextual information due to disconnected data storage, mining of structured data, Including mining of non-data, prioritization of resolution action, consideration of equipment state change, cost, accuracy and speed.

Therefore, there remains a need for a robust system and method for root cause analysis of deviations related to equipment.

The technical problem to be solved by the present invention is to provide a method for root cause analysis of equipment-related deviations.

Another technical problem to be solved by the present invention is to provide a mechanism for detecting equipment-related deviations.

Another aspect of the present invention is to provide a mechanism for determining a deviation resolution sub-graph (DRsG) to solve a deviation based on a deviation resolution graph (DRG).

Another technical problem to be solved by the present invention is to provide a mechanism for recommending DRsG for resolving equipment-related deviations.

The technical objects of the present invention are not limited to the technical matters mentioned above, and other technical subjects not mentioned can be clearly understood by those skilled in the art from the following description.

According to an aspect of the present invention, there is provided a system for root cause analysis of deviations associated with equipment, the apparatus comprising: a Sensor Resolution Action (SeRA) entity for detecting a deviation associated with at least one equipment; Wherein the at least one DRsG is a DRG entity that resolves deviations based on at least one Deviation Resolution Graph (DRG) by determining a Deviation Resolution sub-graph (DRsG) A DRG entity comprising a Resolution action of the at least one device and at least one observation associated with the at least one device, and a Recommendation to recommend the at least one DRsG to resolve the deviation associated with the at least one device Entity.

In some embodiments of the invention, the observation is an original observation.

In some embodiments of the invention, the observation is a secondary observation.

In some embodiments of the invention, the resolving operation includes a repair operation.

In some embodiments of the invention, the resolution includes a secondary operation.

In some embodiments of the invention, the resolving operation includes a resolving operation.

In some embodiments of the invention, the resolving operation includes a reset operation.

In some embodiments of the invention, the DRG is formed based on data received from a plurality of sources. In an embodiment of the invention, the data is structured data. In an embodiment of the invention, the data is unstructured data.

In some embodiments of the invention, the deviation is identified based on the SeRA graph. The SeRA graph shows the relationship between equipment groups and resolution work.

In some embodiments of the invention, the SeRA graph is formed based on an analysis of the set of DRsG of the equipment. The context analysis identifies the equipment that contains the signature of the repair operation and forms a causal relationship that is stored for automated recommendation of the DRsG.

In some embodiments of the invention, the context analysis engine calculates a deviation score based on whether the time series signal of the at least one device includes a change point. The confidence engine determines the confidence score based on the deviation score. The fuzzy score globalizer determines the deviation vector based on the score. The sensor selector establishes a causal relationship between the deviation and the instrument based on the SeRA graph.

In some embodiments of the present invention, when the change point is not available in the time series signal, a window-let dynamic time warping (WDTW) is performed to calculate the deviation score .

In some embodiments of the present invention, a Window-let Gamma (WGamma) is performed to calculate the deviation score of the time series signal of the instrument, when the change point is available in the time series signal.

In some embodiments of the invention, a Fisher discriminant index (FDI) and a dynamic time warping (DWT) are performed when the change point is available in the time series signal, Is calculated.

In some embodiments of the invention, it is calculated based on the long term trend and the short term aging trend.

In some embodiments of the present invention, the DRsG is triggered by the sensor selector when the deviation vector satisfies a threshold.

In some embodiments of the present invention, the time series signal is obtained from video data of the at least one device.

In some embodiments of the present invention, the SeRA entity provides at least one recommendation for a deviation of the at least one device based on the video data.

According to an aspect of the present invention, there is provided a method for analyzing root cause of deviations associated with equipment, the method comprising: determining, by a Sensor Resolution Action (SeRA) entity, And resolving the deviation based on at least one DRG by determining at least one Deviation Resolution sub-graph (DRsG) by a Deviation Resolution Graph (DRG) entity, Wherein the at least one DRsG includes at least one resolution task and at least one observation associated with the at least one device, and recommending the at least one DRsG by the recommendation entity to resolve the deviation associated with the at least one device .

According to an aspect of the present invention, there is provided a computer program product including computer executable program code recorded on a computer-readable non-volatile storage medium for root cause analysis of deviations associated with equipment, A sensor resolution action (SeRA) entity detects deviations associated with at least one device and is determined by a Deviation Resolution Graph (DRG) entity to include at least one Deviation Resolution sub - DRSG) to resolve the deviation based on at least one DRG, wherein the at least one DRsG comprises at least one resolution operation and at least one observation associated with the at least one device, wherein the recommendation entity By recommending said at least one DRsG, A it comprises solving the above-mentioned deviation.

The embodiments of the present invention will be appreciated and understood when considered in connection with the following detailed description and the accompanying drawings. However, the following detailed description, which refers to the preferred embodiments and numerous specific details thereof, is intended to illustrate the invention and not to limit the invention. Many modifications and variations of the embodiments of the present invention may be made without departing from the spirit of the invention, and the embodiments described herein include all such modifications.

1 is a schematic diagram of a system for root cause analysis of equipment-related deviations in accordance with an embodiment of the present invention.
2 is a diagram showing different objects according to an embodiment of the present invention in accordance with properties.
FIG. 3 shows a deviation resolution graph according to an embodiment of the present invention, wherein "P" indicates a pass (PASS) and "F" indicates a failure (FAIL).
4 shows a deviation resolution subgraph according to an embodiment of the present invention.
Figure 5 illustrates the traversing of the deviation resolution subgraph according to an embodiment of the present invention.
FIG. 6 is a flow chart for explaining a root cause analysis method of equipment-related deviations according to an embodiment of the present invention.
Figure 7 shows a time series with change points according to an embodiment of the present invention.
Figure 8 shows a time series separated into sub-series based on a change point, in accordance with an embodiment of the present invention.
9 shows a time series with narrow-range and wide-area subsequences according to an embodiment of the present invention.
10 is a flow diagram of a method for obtaining a final deviation point vector according to an embodiment of the present invention.
FIG. 11 shows a final DRG (DRSG) reinforced with information of a sample SeRA graph, a sample DRSG, and a SeRA graph according to an embodiment of the present invention.
Figure 12 illustrates a proposed pipeline that is assisted by monitoring structured MES (Manufacturing Execution System) data in accordance with an embodiment of the present invention to trigger DRSG.
13 is a view showing attributes of sensors and recipe objects according to an embodiment of the present invention.
FIG. 14A illustrates a DRSG evolution / alternative DRSG due to a state change to the same original observation-secondary observation (OO-SO) according to an embodiment of the present invention, and FIG. 14B illustrates a DRSG evolution / alternative DRSG according to an embodiment of the present invention A flowchart for describing recommendations based on equipment status.
FIG. 14C is a graph for describing recommendation based on equipment status according to an embodiment of the present invention. FIG.
15 shows a time series having a Change Point Present (CPP) and a Change Point Absent (CPA) characteristic corresponding to the first and second differences according to the embodiment of the present invention.
16 is a heat map diagram of a proposed WDTW matrix for sensor time series data classified under CPA captured before and after repair work, in accordance with an embodiment of the present invention.
17 is an exemplary category-by-category time series data according to an embodiment of the present invention.
Figure 18 shows a proposed complete fuzzy point totalization (FSG) based score totalization unit in accordance with an embodiment of the present invention.
19A shows a FSG surface according to an embodiment of the present invention.
19B is a schematic diagram of a system for providing recommendations for intra-equipment deviations based on video data.
Figure 20 illustrates a computing environment that implements a mechanism for root cause analysis of deviations associated with equipment in accordance with an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS Embodiments of the present invention and their various features and advantages are described in further detail in connection with non-limiting embodiments illustrated by the accompanying drawings and described by the following description. The description of well known components and processing techniques is omitted in order not to obscure the embodiments of the present invention. Moreover, the various embodiments of the present invention are not necessarily mutually exclusive, and some embodiments of the present invention may be combined with one or more other embodiments to form new embodiments. As used herein, " or " means not to exclude, unless otherwise indicated. The examples of the present invention are merely for understanding the manner in which the embodiments of the present invention are practiced, and those skilled in the art will be able to practice the embodiments of the present invention. Accordingly, it is to be understood that the examples do not limit the scope of the invention.

Entities, modules and units are used interchangeably throughout this specification.

Embodiments of the present invention disclose a system for fundamental analysis of equipment-related deviations. The system includes a SerRA entity that detects equipment-related deviations. The DRG entity determines DRsG to resolve DRG-based deviations. DRsG includes troubleshooting and observations related to the equipment. Recommendation Entity solves equipment related deviations by recommending DRsG.

Unlike conventional methods and systems, the proposed method can perform root cause analysis of equipment related deviations in an easy, reliable, accurate, fast, efficient, and flexible way. The proposed method can be used to rank the DRSG using different attributes such as time, execution cost, quality of results and frequency of execution, thus improving the accuracy of the root cause analysis of equipment related deviations. The proposed method can obtain a confidence level for the deviation score of the sensor data using the fuzzy inference system. The proposed method provides better analysis based on gradual changes in equipment conditions during Preventive Maintenance (PM) cycles. The proposed method provides better analysis during abrupt changes in equipment condition. The proposed method can effectively record DRSG and sensor analysis data in a Comprehensive Elucidation Table (ComET) scheme.

The proposed method can use a different modeling method than the bezier network method. The proposed system can be used as a recommendation system by both the user and the machine. The proposed method can obtain the reliability of fault detection based on a statistical model without requiring extensive manual intervention. The proposed method uses a heuristic-less analysis method for out-of-spec (OOS) detection that does not require recipe information. The proposed method provides for extracting the necessary support data and solving two equally feasible recommendations, instead of querying the user / analysis system with a second (sequential) query / query. The proposed method can provide an accurate understanding of the deviation and resolution action (RA). The proposed method can provide fast, accurate, and reliable recommendations for resolution work to resolve deviations. The proposed method can provide reliable real-time identification of robust deviations / defects and timely triggering / recommendation of appropriate RAs. The proposed method can improve the expert system power.

The proposed method can form an expert system with a deviation resolution / resolution task recommendation system for a production process / manufacturing process in an industry (such as a semiconductor industry, an industrial plant, a petroleum refinery, etc.).

BRIEF DESCRIPTION OF THE DRAWINGS Referring now to the drawings, and more particularly to Figs. 1 to 20, wherein like reference numerals refer to like corresponding elements, preferred embodiments are shown.

1 is a schematic diagram of a system 100 for root cause analysis of equipment 200 related deviations in accordance with an embodiment of the present invention. The system 100 includes a SeRa entity, a DRG entity 104, a recommendation entity 106, and a comprehensive description table 108. The SeRA entity 102 senses the deviations associated with the equipment 200. After detection of the deviation, the DRG entity 104 determines the Deviation Resolution sub-Graph (DRsG) to resolve the deviation.

In an embodiment of the present invention, DRsG includes a Resolution Action (RA) and an observation associated with the device 200. In some embodiments of the invention, the observation is an original observation. The original observation identifies the beginning of the deviation. The original observation triggers resolution work for deviations. In general, the original observation is performed using a metrology equipment (e.g., OOS (non-specification) critical dimension equipment, etc.).

In some embodiments of the invention, the observation is a secondary observation. Secondary observations are performed within the process of deviation response. Secondary observations are also referred to as in-process observations. In general, secondary observations are performed on process equipment (e.g., high-reflection power devices, etc.).

In some embodiments of the invention, the resolution task (RA) refers to any modification applied to the device 200 to resolve the deviation. These attributes include the equipment ID to which the modification was applied, the modified part or module, the name of the modified parameter and its value, and so on.

Based on the DRsG, the recommendation entity 106 recommends DRsG to resolve the deviations associated with the equipment 200. [ The ComET 108 also stores the SeRA graph and DRG.

In some embodiments of the invention, the DRG is formed based on data received from a plurality of sources. In some embodiments of the invention, the data is constructed data. In some embodiments of the invention, the data is unconstructed data.

The unstructured data is obtained from various sources such as deviation reports, vendor manuals, standard operating procedures, emails, user logs, and the like. The unstructured data includes a description of deviations that are regularly generated. According to various analyzes, unstructured data will account for approximately 79% of the total data in the near future. The unstructured data typically includes information about the original observation recorded at the beginning of the deviation (e.g., the critical dimension of the device is outside the specification, the action taken to resolve the issue (e.g., the radio frequency used in the recipe (E.g., reduced RF power), intermediate / secondary observations (e.g., high reflected power), and the results of the final inspection.

The structured data includes information related to the device 200. The structured data may include, for example, sensor read information, equipment availability, recipe per wafer, measurement values, sensor time series data for a particular wafer run, do. Unlike unstructured data, structured data is recorded in real time in accordance with the wafer process schedule.

In some embodiments of the present invention, structured data and unstructured data are used to recommend correct resolution of deviations. The proposed method also performs real-time fault detection and identification of resolution tasks using a combination of the structured data and unstructured data.

In some embodiments of the invention, the deviation is identified based on the SeRA graph. The SeRA graph shows the relationship between equipment groups and resolution work.

In some embodiments of the present invention, the SeRA graph is formed based on the anteroposterior analysis over a set of DRsGs of the equipment. The post-war analysis identifies equipment with a signature of the repair operation and forms a causal relationship stored for automated recommendation of DRsG.

The post-analysis engine (not shown) calculates a deviation score based on whether the time-series signal of the equipment includes the change point. A confidence engine (not shown) determines a confidence score based on the deviation score. The fuzzy score globalizer (not shown) determines a deviation vector based on the confidence score. A sensor selector (not shown) establishes a causal relationship between the deviation and the equipment based on the SeRA graph. The sensor selector triggers DRsG based on the deviation vector when a deviation is detected.

In some embodiments of the present invention, when the change point is not available in the time series signal, a window-let dynamic time warping (WDTW) is performed to calculate the deviation score .

In some embodiments of the present invention, a Window-let Gamma (WGamma) is performed to calculate the deviation score of the time series signal of the instrument, when the change point is available in the time series signal.

In some embodiments of the invention, a Fisher discriminant index (FDI) and a dynamic time warping (DWT) are performed when the change point is available in the time series signal, Is calculated.

In some embodiments of the invention, it is calculated based on the long term trend and the short term aging trend.

In some embodiments of the present invention, the DRsG is triggered by the sensor selector when the deviation vector satisfies a threshold.

In some embodiments of the present invention, the time series signal is obtained from video data of the at least one device.

In some embodiments of the invention, the DRSG provides an action flow for any negative observation. The DRSG provides the flexibility to analyze work even if the situation is unique. In addition, the DRSG recommends modifying the workflow even if deviations have not occurred in the past. DRSG is aided by data analysis of structured data. Data analysis assists in process, flag deviation and real-time monitoring of the associated DRSG.

In some embodiments of the invention, the proposed method is used to convert unstructured data, such as deviation reports, to objects. This can be achieved by using Natural Language Processing (NLP) or statistical models (e.g., Term-Frequency-Inverse-Document-Frequency (TF-IDF), entropy Based method, and the like). In order to increase the visibility of all related tasks to be performed in order to understand and solve the deviations, the proposed method classifies deviations into the following main object types.

Different objects with their respective attributes are shown in Fig. 2 for performing data analysis. Objects have various attributes such as date, time, equipment ID to which the object is linked, observed wafer / lot ID, previous and subsequent object IDs, and so on. FIG. 3 shows a DRG according to an embodiment of the present invention, where P denotes pass (PASS) and F denotes fail (FAIL).

Also, if the deviation response is converted into the above-mentioned object, the recommendation process links them. This connection is mainly established according to the date and time characteristics of the objects. For well-written deviation reports, the time series of objects can be tracked by the location of different work groups in the document. In the case of structured data, objects generally have system-generated timestamps associated with them. Once the connection is established, form a chain of objects that leads from the beginning to the end of the deviation. Figure 3 illustrates this object chain. This chain is referred to as a deviation resolution graph (DRG). As shown in FIG. 4, the deviation resolution includes one original observation (OO ₁ ), four secondary observations (SO ₁ - SO ₄ ), and four resolution tasks (RA _A - RA _D ). Between secondary observations, a number of resolution tasks are arranged. To further analyze the deviation response, the method divides the DRG into smaller subsets based on the analysis rules of the deviation resolution graph.

Analysis rules for deviation resolution graphs: In a typical DRG, multiple repairs and secondary observations are interspersed between each other. In order to interpret the resolution work that has a significant impact on resolving issues, the steps are listed below.

From the deviation resolution graph, the proposed method can be used to generate sections starting with negative observation and ending with the next positive observation of the same type. The method also ends the graph section by adding the last node (E) after the last positive observation node. This graph section is referred to as a deviation resolution sub-graph (DRSG). Figure 4 shows the four DRSGs obtained from Figure 3.

In one embodiment of the present invention, where there is no positive observation of the corresponding negative observation, the proposed method includes the entire section from the positive observation node to the end of the DRG. Thus, in the observation of SO4, the secondary deviation correction graph is shown in Fig. 4C. Based on the above procedure, the DRG becomes a DRSG after the addition of the E node at the end.

In some embodiments of the present invention, the proposed method treats all other in-process observations of different types as branching points. These bifurcation points will be used as entry / exit points from other deviations.

In some embodiments of the present invention, each DRSG is treated as a tree branch having an observation node acting as an entry / exit point of another DRSG according to the following rules.

In any observation node (SO _2), if the result is, if the value in the opposite forward (e.g. SO ₂ (P) rather than when in the SO ₂ (F)), the flow exits the DRSG that meet the following criteria: Arrives at the graph: the target DRSG must have at least one instance of the same node, i. E. At least one node should be SO ₁ (P).

In selecting a plurality of target DRSGs satisfying the criteria for the observation node and its value, the same ranking method is listed as follows.

Before leaving, the user is given an option as to whether to continue with DRSG or not and the procedure is overridden by the user's decision.

If no DRSG satisfies the criteria, the flow moves to the next node after obtaining the user's approval. If the user approves, ComET 108 is updated with another DRSG having the opposite node. For example, in the above example, ComET 108 obtains another entry with an SO ₂ (F) node replacing SO ₂ (P). This is because the error value of SO2 and the example of the solution flow that proceeds even if it has that information can be used for later deviation resolution.

This procedure is shown in Fig. This includes DRSG for SO _1. At the second node, SO ₂ (P), instead of passing the expected value for each original DRSG, an error value is obtained. Then, based on the user identification, the flow branches to the other of DRSG SO1 having a SO ₂ (F) or node, or continue with the same DRSG. If the decision is to continue with the same DRSG, the ComET 108 is updated with the new DRSG. If the decision is to branch, the next step is directed by the new DRSG. Using the above rules, DRSGs can be generated and traversed for the desired execution of the resolution flow. Such a multi-tree structure can be stored in the form of a table or database having an appropriate index. A collection of such databases is referred to as a Comprehensive Elucidation Table (ComET) scheme.

For example, common equipment in a manufacturing environment includes hundreds of sensors that measure various attributes such as temperature, pressure, gas flow, incident and reflected RF power, optical emission spectra, current and voltage, and the like. There are also actuators such as recipe-controlled pressure valves, gas flow valves, solenoids, heaters, RF power controllers, etc., performed in the equipment. The MES of a semiconductor fab includes sensor data as well as actuator data. If the instrument includes a complete set of sensors, it can track the time series deviation of at least one sensor / actuator. However, knowing which sensors to monitor for given deviations and identifying deviations at the sensor level may require expertise. In the next section, the proposed method automatically uses the deviation and defect event model learning and sensing at the sensor level and uses this information to improve the applicability of DRG and DRSG.

Analysis of Before / After Resolution (B / A RA): The analysis of B / A RA includes learning and monitoring phases. In the learning phase, for each DRSG, the proposed method searches for all changes in sensor activity immediately after the first node of the DRSG (i.e., before the resolution operation) and immediately after the end of the DRSG (i.e., after the resolution operation). If any significant change in the behavior is observed and the monitored sensors exhibit a similar deviation behavior, it means that the corresponding node of the DRSG can be raised with certain confidence.

6 is a flowchart 600 illustrating a method for analyzing causes of deviations associated with equipment in accordance with an embodiment of the present invention. In step 602, the method includes detecting deviations associated with the equipment. In one embodiment of the invention, the method allows the SeRA entity 102 to detect deviations associated with the equipment. In step 604, the method includes determining DRsG based on the DRG to resolve the deviation. In an embodiment of the invention, the method may allow the DRG entity 104 to determine DRsG based on the DRG and resolve the deviation. In step 606, the method recommends DRsG to resolve deviations associated with the equipment. In one embodiment of the invention, the method allows recommending entity 106 to recommend DRsG to resolve deviations associated with the equipment.

The various operations, operations, blocks, steps, etc. of the proposed method may be performed in the described order or in a different order or simultaneously. Further, in some embodiments of the present invention, some operations, operations, blocks, steps, etc. may be omitted, added, changed, or skipped without departing from the spirit of the present invention.

FIG. 7 shows a time series having an alteration according to an embodiment of the present invention.

Numerical series (numerical time - series data : One of the first steps is when the sensor detects a jump (eg, a gas flow, an RF power sensor), a tipping point (eg, a light emission spectrum), a peak (E.g., a light emission spectrum, for example). Two categories are defined: Change Point Absent (CPA) and Change Point Present (CPP). This information helps to perform fine tuning analysis. Figure 7 shows an example of a curve with a peak and a jump change point.

In one embodiment of the present invention, when the change point is not available in the time-series signal, a window-let dynamic time warping (WDTW) is performed to calculate the deviation score of the time-series signal of the equipment.

In one embodiment of the invention, when the change point is available in a time-series signal, a window-let gamma (WGamma) is performed to calculate the deviation score of the time-series signal of the equipment.

In one embodiment of the present invention, when the change point is available in a time-series signal, a Fisher discriminant index (FDI) and a dynamic time warping (DWT) are performed to determine the deviation score of the time- .

CPA : In the absence of a change point (CP), the WDTW process involving continuously dividing B / A time series data and calculating dynamic time warping (DTW) scores between two sensor time series at each level do. The WDTW matrix may be monitored for any out-of-spec (OOS) value over time in any interested time window of interest. The advantage of using a window-ruling approach with a single DTW score (calculated for the entire time series) is that the time window of the deviation can be narrowed as desired without knowledge of the recipe's lower level information in the time series. This enables the proposed invention to have no recipe information. Also, due to windowing, avoid the averaging of the results when considering the entire time series or fixed window size. Based on the regions in the WDTW matrix in which the cause-effect manifestations are observed, they are identified as region of interest (ROI). This ROI is monitored for each successive wafer process and the deviation for detecting defects is checked, thus triggering the corresponding resolution operation. The value of the integrated WDTW score from the ROI of the WDTW matrix is analyzed by a confidence engine and given a confidence score based on the dynamics of the process and process equipment.

개의 매트릭스를 생성하고, 이들의 특이값 분해(Singular-Value-Decomposition; SVD)를 계산하는 것으로 이어진다. 여기서, n은 각각의 레벨에서 양분된 시간 윈도 내의 최대 샘플 수를 나타내며, i는 센서 번호를 나타낸다. SVD가 얻어지면, 최대 특이값과 최소 특이값의 비율을 계산되고, 이는

로 지칭되며 여기서 A와 B는 편차 이전 및 이후를 의미하며 (p, q)는 계산이 수행되는 시간 윈도를 의미한다. 단일 점수(전체 시계열을 단일 윈도로 고려할 때)는 상술한 것(단일 DTW 점수)과 유사한 문제를 유발할 수 있으며, 편차의 정확한 위치를 찾아내지 못할 수 있다. 그러므로, 그러한 엔트리로부터, 매트릭스가 형성되고, 이는 윈도-렛 감마(Window-let Gamma) 매트릭스로 지칭되며, 모니터링될 수 있으며, CPP의 경우 WDTW와 유사한 매트릭스가 형성된다. WGamma 매트릭스의 ROI로부터의 통합된 값들은 추가 분석을 위해 확신 엔진으로 전달된다. CPP - Temporary deviation monitoring : If there is a change point (CP), there are two possibilities: temporal deviation and amplitude bias / slope deviation. The temporary deviation is described first. Following the detection of the transient deviation between before and after the time series, successive B / A time series data is divided and divided at each level

Matrix, and then calculates their Singular-Value-Decomposition (SVD). Where n represents the maximum number of samples in the time window divided at each level and i represents the sensor number. When the SVD is obtained, the ratio of the maximum singular value to the minimum singular value is calculated,

Where A and B are before and after the deviation, and (p, q) is the time window in which the calculation is performed. A single score (when considering the entire time series as a single window) may cause problems similar to those described above (a single DTW score) and may not be able to locate the exact position of the deviation. Therefore, from such an entry, a matrix is formed, referred to as a Window-let Gamma matrix, which can be monitored and a matrix similar to WDTW in the case of CPP is formed. The integrated values from the ROI of the WGamma matrix are passed to the assurance engine for further analysis.

CPP : Amplitude Deviation Monitoring : To monitor amplitude deviations (both in deflection and in slope changes), the time series is divided into sub-series using CP, i.e., the start and end points of the window, as shown in Fig. And they are classified as " narrow range " and " wide range " based on the kurtosis of the subsequence value as shown in FIG. Narrower subseries can only provide amplitude bias deviations, whereas broader subseries can provide amplitude and tilt bias deviations. Depending on this, a FDI based method in addition to the DTW or FDI based method or the DTW method is used to obtain the deviation score. Score: S _FDI and S _{FDI - The DTW} is delivered to the assurance engine for further analysis.

Category (categorical) / nominal (nominal) time-series data: In order to calculate the deviation score for the numerical time series, the proposed method along the lines similar to the previous window-let-Min-Hash matrix (Window-let Min-Hash matrix ; Lt; RTI ID = 0.0 > WMH) < / RTI > The resulting deviation score is passed to a sensor selector module for further analysis.

Even after obtaining the deviation score, there is a unique problem in the confidence level of the ability to accurately sense the deviation. To address this, the method utilizes a confidence engine that provides a confidence score for each deviation score calculated above. The deviation score calculated in the numerical time series is passed to the assurance engine. Due to the nature of the deviations occurring in the time series data by category, the corresponding deviation scores through the WMH matrix are passed directly to the sensor selector module. Thus the input to the confidence engine is a set of four deviation scores derived from the numerical time series: S _WDTW , S _WGamma , S _FDI and S _FDI-DTW arranged in vectors _.

Confidence Engine (CE): A simple occurrence of a high deviation score for a time series may not be a good basis for triggering the corresponding DRSG. One of the main reasons for this observation is the change in the state of the equipment, which can occur slowly or rapidly. The proposed method deals with such equipment state changes. The result from the convincing engine is one of the key inputs of this method. The assurance engine includes a historical data store (HDS) module and a fuzzy socre-globalization (FSG) sub-module.

The HDS module analyzes historical data to obtain a set of local confidence scores. The local confidence score is used to obtain a global confidence score. The overall confidence score is multiplied by the corresponding deviation score to obtain the final deviation score. The following local confidence scores are calculated by the sensor bunching unit, the long term dynamics / aging unit and the short term dynamics / aging unit.

The long term dynamics / aging unit operates by considering the long term dynamics of the deviation score through the regression model and by providing an insight as to how the current value is fitted to the learned long term characteristic. Since there are four deviation scores, this sub-module calculates four corresponding local confidence scores, β _i , (i = 1, 2, 3, 4) .

The short-term dynamics / aging unit operates by considering the short-term dynamics of the deviation score as a certain forgetting factor and as a variance / standard deviation of the deviation score over the range of the previous process. In addition, the short-term dynamics / aging unit provides insight into how the current value is fitted to the short-term characteristic. Since there are four deviation scores, this sub-module calculates the corresponding four local confidence scores γ _i , (i = 1, 2, 3, 4) .

The outputs of these sub-modules are integrated into the following forms:

이는 전반적 확신 점수를 얻기 위해 FSG 모듈로 전달된다.

This is communicated to the FSG module to obtain overall confidence scores.

를 계산한다. 이러한 값들은 네 개의 편차 점수

와 곱해져서

로 얻어지는 센서 당(per sensor) 최종 편차 점수가 얻어진다. 이 최종 편차 벡터는 센서 선택기 유닛에 의해 이용되어 센서와 편차 사이의 인과 관계(causality relationship)를 성립하고, 나아가 최종 편차 벡터(

)에 기반하여 DRSG RA를 트리거링한다.The Fuzzy Score-Globalization (FSG) sub-module operates based on a fuzzy inference system (FIS) that receives the integrated vector. The FSG sub-module has four corresponding overall confidence values,

. These values are the four deviation scores

Multiplied by

The final sensor deviation per sensor is obtained. This final deviation vector is used by the sensor selector unit to establish a causality relationship between the sensor and the deviation,

) To trigger the DRSG RA.

Sensor Selector (SS) : The sensor 'group' and sensor resolution-action (SeRA) graphs showing the relationship between the sensor (group) and the resolution work are inferred through the B / A RA analysis pipeline and sensor aggregation unit. do. In the SeRA graph, the sensor and resolving operations form nodes. If a causal relationship between a given resolution task and sensor (group) data is observed, the association between the given resolution task and the sensor (group) (edge) is added. Also, each edge is annotated with an association weight. Considering the edge, if the causality is high (for example, if the relationship is more frequent or dominating when inferred based on past observations and / or by a specialist in the field), a higher relative association Weights are assigned to such edges. Association weights are necessary and sufficient conditions for the association graph. This is because the mapping between the set of resolution tasks (i.e., S _RA ) and the set of sensors (i.e., S _S ) need not necessarily be bijective. Thus, the weights quantify the indications of confidence to model the association.

가 문턱값보다 크다면;

에 기초하여, 대응하는 편차 DRSG가 트리거링된다. 여기서 arg()는 편차가 관찰된 방향의 '편각'를 의미하고 이는 편차 점수 또는 이의 조합({F _WDTW , F _WGamma , F _FDI 및 F _{FDI - DTW} } 및/또는 S _WMH )으로 직접 맵핑될 수 있고, 이는 사양 외(OOS)이거나 비정상이다.An example SeRA graph for the following scenario is shown in FIG. 11 with an event-graph for the overall view. The RF power sensor measurement is out of spec (OOS) in the negative direction (current value is generally less than required). This may be due to damage of the power line to the RF coil or damage of the RF coil. The final (sample) DRSG derived from a combination of historical data query on the power of the DRSG and real-time RA triggering of the SeRA's power is also shown in FIG. The final deviation score (vector) is used with the final DRG (DRSG) to track and trigger the resolution task if anomalies are observed. The same information is stored in Comet 108 in the form of a table. While a given sensor group is being monitored (focused monitoring), if observed

Is greater than the threshold value;

The corresponding deviation DRSG is triggered. Where arg () means the 'angle of declination' of the direction in which the deviation is observed and can be mapped directly to a deviation score or a combination thereof ({ F _WDTW , F _WGamma , F _FDI and F _{FDI - DTW} } and / or S _WMH ) , Which is out of specification (OOS) or abnormal.

Figure 12 illustrates a proposed pipeline that is assisted by monitoring structured MES data in accordance with an embodiment of the invention to trigger the DRSG. A complete data flow diagram summarizing the proposed pipeline that is assisted by monitoring the structured MES data to trigger the DRSG is shown in FIG. Previous sequential wafer process results are stored for OOS testing and learning of regression and dispersion models of long-term and short-term trends. The assurance engine's HDS is supplied either in-house or externally based on a number of infrastur- tural factors. The output of the sensor selector is combined with the DRSG and thus tabulated with ComET 108 to form an ES with real-time (wafer process level) automatic fault detection and DRSG (RA) triggering capability.

ComET Configuration : The SeRA graph from the assurance engine and sensor selector, including output from the DRSG, B / A RA analysis, is stored in the ComET 108. In addition, ComET 108 includes the deviation response and additional relevant data extracted from the knowledge of the field. ComET 108 forms the basis for discovering and learning relationships between the various processes within the manufacturing process. From the point of view of relational databases, it is also possible to use the following tables (i.e., main tables, sensor selector tables, rescue tables, cost tables, time tables, data collection frequency tables, can be described as a schema.

The main table contains DRSG information forming a ComET scheme. The main table will store nodes with all attributes. The sensor selector table contains the information of the sensor object as shown in Fig. The deviation ID (s) attribute links the sensor from the main table to the DRSG. The information from the SeRA graph obtained from the confidence engine and the sensor selector as well as the B / A RA analysis is stored as an attribute in this table.

The recipe table stores information about the recipe used in the equipment. For example, the recipe table stores attributes such as number of sub steps, duration, pressure, temperature, power, gas flow, alarm or interlock level of each parameter, and so on. The recipe table may be pre-formed from equipment or process specification documents. An example of a recipe object is shown in Fig.

The OO-SO table represents a secondary observation that is a necessary and sufficient condition for the original observation to pass. This table can actually be generated by merging the original observation (OO) and secondary observation tables of the deviation ID. Generally within the manufacturing environment, the original observation is performed on the product wafer. This can be done, for example, on a statistical sampling basis. If this OO is OOS, then the SO associated with the equipment is used to identify the pass / fail status of the original observation.

Cost Table: OO requires a product wafer to be run and can therefore be exposed at a high cost. In addition, OO may require that the wafer undergo some or all of the steps preceding the point at which the OO is performed. On the other hand, the SO can be low cost because it is performed on a short loop requiring a step through a limited step.

Time table : The time required to prepare the wafer for the SO is generally much lower compared to the OO since the OO wafer requires the entire process to that point, while the SO passes through the short loop.

Data collection frequency table : Generally, SO is higher in frequency of data collection than OO. Thus, changes in equipment conditions can be obtained with much better granularity over time.

Equipment-independent process changes: Some of the worse OOs can be upstream processes, wafer-related issues, and bad motion techniques such as incorrect recipes. This is irrelevant to the state of the process equipment in question. Since SO is associated with only a small number of process steps, it is largely unaffected by these issues and can be performed on a plurality of wafers at low cost and in a short time.

The use of this OO-SO table is to know the candidates of the subgraph for solving. For example, if SO ₂ and SO ₄ are necessary and sufficient conditions for OO ₁ , then all solution subgraphs starting with SO ₂ and SO ₄ will be the candidate subgraphs of the resolution work for any deviations associated with OO ₁ . Of these candidates, SO ₂ and SO ₄ are recommended to the user, sorted based on frequency, time, cost, etc., as described in more detail below.

Similar equipment table : Within a manufacturing environment, multiple equipment can be used interchangeably to perform the same process operation. In other words, there is a group of equipment that is considered equivalent for a particular process. This information exists in the form of a table and is available in the MES database. This information will be used to perform data analysis in the proposed method. For example, if viewed as the equipment A and B is similar to, a graph of the resolution corresponding to equipment B, OO ₁ for any deviation of the OO ₁ in the equipment A can be used for analysis. This is particularly useful when new equipment enters the manufacturing environment and there are insufficient past deviations associated with it.

Advantages of the ComET scheme : General troubleshooting procedures in semiconductor equipment or manufacturing processes are carried out by interspersing multiple observations between resolution tasks. In many cases, it is difficult to know the exact effect of work on observations. Therefore, all or some of the work prior to the observation can affect the results. The proposed form does not need to know the exact nature of this relationship because it stores all the operations that cause the change in observation state (from fail to pass).

Resolution subgraphs for deviations of the same type can be sorted according to different attributes of the graph, such as number of bifurcations, time and cost of resolution, availability of resources, frequency of occurrence, and the like. Depending on the preferred alignment method, different machine learning algorithms, such as, for example, a beacon network, neural network, and decision tree, may be adjusted to obtain the desired recommended resolution flow.

Since it can be sorted according to the frequency, it becomes possible to explain the state change of the equipment which occurs suddenly between the PMs or after the part replacement suddenly occurs. This mechanism is described in detail in the following sections.

According to the above rules, each negative observation node also becomes the starting node of the DRSG. As a result, it is possible to identify DRSGs that have not occurred independently in the past. That is, SO ₄ (F) in FIG. 3 occurs as part of the main resolution flow. However, even that produced DRSG in accordance with the rules prescribed, as shown in Figure 4 Next, look for a list of actions to be performed when the SO ₄ is the only observed deviations. That _{is, SO 4 (F) ->} RA-> SO 1 (F) -> RA C -> SO 1 (P) -> RA D -> SO 3 (F) -> RA A -> SO 3 (P) - > E (from Fig. 4c).

Each branch point depends on the observation result corresponding to the node. Based on this, the recommended solution flow may vary. Therefore, a dynamic solution is established that can adapt to the outcome of the proceeding procedure. The proposed method provides dynamic reconfiguration of the recommended DRSG.

Due to the SeRA graph obtained from the assurance engine and sensor selectors as well as the real-time (wafer process level) monitoring via the connection of the sensor (group) to the DRSG and B / A RA analysis, ComET 108 identifies the deviation and real- to provide. This provides real-time identification and recommendation of defects.

Pruning of DRSG: When initially created, there may be variations in the length and content of the DRSG in the same observation. For example, let's compare the first two DRSGs in FIG. In order to eliminate this change, the following method for concatenation of the subgraph (ie, the opinion of the field expert and the pass / fail possibility of in-process observation) is proposed.

Opinion of the field expert : It is important in the DRSG to draw expert knowledge from the field in determining specific tasks / observations and their sequence. This knowledge is far more reliable than questioning the experts on probability and likelihood. This is because knowledge of the necessary and sufficient conditions of operation / observation can be confirmed by the equipment supplier's instructions. Therefore, this is less likely to be erroneous. Examples of such queries are:

With respect to the first subgraph of Figure 4, one may ask, "Does in-process observation such as SO ₂ , SO _3, etc. need to be taken?" According to the answer, Can be simplified.

In relation to the second sub-graph in Fig. 4, "RC is applied to a necessary and sufficient condition for solving the SO ₁ (F)? Or does it require both RA _A and RA _B ? "According to the answer, the proposed method can promote / demote the second subgraph to the first subgraph.

Based on the answers of the experts, recommendations for deviation responses can be strengthened.

Pass / fail probability of in-process observations: If in-process observations of the solving subgraph (e.g. SO ₂ , SO ₃ , SO ₄ in the first subgraph of FIG. 4) If it has more than pass / fail probability, it can be removed from the subgraph. Thus, in the same example, if SO ₂ passes the probability> θ and SO ₃ fails with probability> θ and SO 4 fails with 0 <probability θ-Δ (θ>Δ> 0), SO ₂ and SO ₃ Can be removed and the SO ₄ can remain in the DRSG.

Recommendation Entity 106 : Once the ComET data is generated, it can be used to create a recommendation entity 106. The recommendation entity 106 may determine whether the user has identified an appropriate aligned DRGS in case of identifying a failing observation or if the B / A RA analysis, assurance engine and sensor selector observe the deviation of the sensor measurements to trigger the flag . Since all data related to the deviation and the DRSG are stored in the ComET 108, the in-list recommendation entity 108 of the DRSG is a question regarding the query for the correct OO / SO / sensor ID.

Also, unlike conventional / classic RSs that ask the user for details if one or more recommendations of the RA are searched for the query, the recommending entity 106 may determine the minimum possible number of the most accurate RA recommendations (ideally unique A) can be automatically collected and augmented with the necessary information. For example, consider a scenario where query Q1 is queried. If two feasible recommendations were found: R1 and R2 are searched using Comet 108 (and DRSG) traces with OO, SO or sensor ID, where R1 is true only in sub-condition SC1 Whereas R2 is true only in the mutually exclusive subcondition SC2. At this point, the classic RS makes a second query to the user. "If the R1 is true or R2 is true, then the corresponding recommendation is provided if the user enters the answer of the second query. / YMS data, it will automatically refer to the MES data (or inventory management system (IMS) or yield management system (YMS)) to determine the time / place , And collect data when R1 or R2 is true. Entering a wafer ID or timestamp with respect to the time / place where the problem / event / observation occurred can be done if R1 is true Or it can be entered much easier by the user, as opposed to more detailed questions such as determining whether R2 is true. In addition, in a realizable situation, Is faster and sometimes more accurate than relying on the user every time, so the proposed method can help implement a fast and highly automated solution work recommendation system, thanks to real-time support from background systems such as MES have.

The remaining major tasks are to align the DRSG and to describe the most unique characteristics of semiconductor equipment, namely state changes.

Grade setting (fine tuning) of DRSG: In a typical semiconductor manufacturing environment, the same defect observation can have a plurality of DRSGs, such as SO ₁ (F) shown in FIG. When such a situation occurs, the following method (frequency of occurrence, duration of DRSG execution, cost of DRSG execution, quality of result obtained by DRSG termination) is suggested at the time of recommendation to the user.

Frequency of occurrence : The solution flow corresponding to the same observation can be sorted according to the number of times they have been executed. For example from Figure 4 example, if the second sub-graph to resolve the SO ₁ (F) runs more than the first sub-graph, the second to the subgraph more high grade set and to correct the SO1 (F) Will be offered as a first recommendation. A potential disadvantage of this approach is that it can ignore changes in equipment status.

Duration of DRSG execution : Another alternative to using frequency is to sort the subgraphs according to the time it takes to complete the execution. It should be noted that ComET 108 contains information of the time required to complete the execution of each object. From this, it is possible to obtain the total time required and sort the recommended DRSG accordingly.

Execution cost of DRSG : Another alternative to frequency and execution time is to set the DRSG rating by the cost associated with their execution. These costs consist of the operating costs, such as demand resources, replacement parts, and the number of people required to perform the work.

Quality of results obtained at the end of the DRSG: The DRSG can be sorted according to the quality of the end result. This can be defined by the function Q as shown below.

Where V _i , TV _i , USL _i , LSL _i and W _i are the observation value, the target value, the upper spec limit, the lower spec limit and i Of the first process. Q may be replaced by an equivalent value, which means the quality of the final result. Thus, if the first subgraph of FIG. 4 has a better Q value than the second subgraph, then the first subgraph should be recommended if Q has priority over frequency, time and cost. Due to the scalable nature of ComET 108, the Q value is associated with the deviation ID for reference and stored in the main table.

Explanation of Equipment Condition Change : The frequency of occurrence of DRSG can be used to determine the recommended path of work. Other machine learning algorithms may also be used to determine recommendation, such as Beginnings, Neural Networks, Decision Trees, Dips, and Reinforcement Learning Algorithms. However, this approach assumes the static behavior of the system. The manufacturing equipment is essentially non-stationary. The state of the equipment changes by usage and specific external events. The state change rate may be gradual in the process chamber, such as by build-up of residual processing material between the preventive maintenance (PM) cycles, or by wear of the part, And the like. As a result, the frequency of repair can change over time, and it is important to consider this phenomenon to predict the correct deviation resolution subgraph. In the following, an approach that considers not only equipment state changes between PM cycles but also abrupt state changes (i.e., equipment state changes between PMs and equipment state changes after parts replacement) are suggested.

Generally within a manufacturing environment, the PM schedule is based on the number of runs, the processed wafers, or the time taken. As a result, the repair frequency can vary within the PM cycle. 14A shows two subgraphs with the same deviation (high average critical dimension). The work-up for the first subgraph RA _A is the tuning of the matching circuit of the etching equipment, whereas the second subgraph RA _B is the chamber cleaning. Now, if many wafers pass through the chamber and if OO ₁ is observed at the end of the PM cycle, then RA _B is likely to be a reasonable workaround for RA _A. If OO ₁ was observed at the beginning of the PM cycle, the reverse would be possible. If the analysis of the recommendation solver subgraph is based solely on the frequency or machine learning algorithm, then the predicted answer will not be given. The proposed method takes into account changes in the repair frequency within the PM cycle when multiple solution subgraphs exist for the same observation, using ComET 108 as follows.

First, a method of obtaining a set of reference PM resolution subgraph vectors is described.

1. Assume that data is available for n PM cycles and this cycle is determined by the number of wafer runs. From this data, we find the median of the number of wafer runs between two instances of the same subgraph. In subgraph SG1, eta is a median value of the number of wafer runs.

2. Divide the total PM cycle execution by η and obtain the sum of the occurrence cases of SG1 at each interval. In this way, one vector is obtained per PM cycle. Note that some of the PMs may have different orders if they have a run count that is greater than or less than the specified run count.

Here, SG1 _PM1 is a vector for the first PM cycle PM1 and the dimension of the vector is nSG1 PM1.

1. When vectors of different subgraphs are obtained, they are stored in memory for use as a reference.

The reference pattern samples different equipment state spaces. If a new deviation occurs in the PM cycle at the same observation (for example a high average critical dimension), the following steps are followed.

1. In the subgraph, vectors are formed in the current PM cycle in a manner similar to that described above. For example, in the execution of m times, RA _A will have the following vectors.

2. Next, the distance between the current PM and the reference PM is obtained by truncating the reference PM dimension to nSG1PMcurrent . For example, the distance value between PM1 and the current PM is given by:

는 0으로 나눠지는 것을 방지하기 위한 매우 작은 양의 실수이다. 또한, p= 1는 맨해튼 거리(city block/Manhattan distance)이고 p=2는 유클리드 거리(Euclidean distance)이다. p의 선택은 요구되는 정확성에 기초할 수 있고 작업 유형에 따라 다를 수 있다.here,

Is a very small positive number to prevent it from being divided by zero. Also, p = 1 is the city block / Manhattan distance and p = 2 is the Euclidean distance. The choice of p may be based on the required accuracy and may vary depending on the type of operation.

1. To obtain a predicted value for the next index of the current PM of SG1, a weighted sum of the reference PMs is obtained.

Finally, the values of the indexes are compared with respect to different SGs and are provided to the user in descending order. It is predicted that the SG having the highest value is most recommended to be executed.

14B is a flowchart 1400b for explaining a recommendation method based on equipment status according to an embodiment of the present invention. At step 1042a, the method includes obtaining an intermediate value of execution among the two instances of DRsG. In step 1404a, the method includes converting the PM cycle to a vector that uses the intermediate value as a bin size. In step 1406a, the method includes finding a similarity metric between the current PM vector and the previous PM cycle's vector. At step 1408a, the method includes obtaining an estimate for the next bin for each DRsG of the current PM cycle. In step 1410a, the method compares the predictions over similar DRsGs and selects the maximum value.

The proposed method utilizes the previous information of the DRsG implementation to predict the occurrence of the next DRsG.

The various operations, operations, blocks, steps, etc. of the proposed method may be performed in the described order or in a different order or simultaneously. Further, in some embodiments of the present invention, some operations, operations, blocks, steps, etc. may be omitted, added, changed, or skipped without departing from the spirit of the present invention.

FIG. 14C is a graph for describing recommendation based on equipment status according to an embodiment of the present invention. FIG.

Changes in equipment status after part replacement : The method described above is for a gradual change in the PM cycle. In addition, sudden changes such as part replacement or improvement of the equipment configuration can occur and the deviation profile can be changed. In Fig. 14A, for example, if the matching circuit of the etching equipment has been upgraded, the first subgraph is less likely to be a problem. In this case, a sudden change in the frequency of the associated subgraph is expected. As a result, the reference PM vector obtained above may not be valid. In this case, the following points are proposed.

1. After repair, ask the expert whether this is a periodic repair where the repeat of the deviation is predicted, or an unusual repair that can correct the deviation in the near future.

1. If it is a regular repair, there is no change in the frequency trend when compared to historical data.

2. If it is a unusual one-time repair, it leads to the following:

a. Ask the expert whether an approximate wafer run can be obtained prior to the repetition of defects. This information can be obtained from previous experience or vendor manuals.

b. Next, we ask the expert what DRSG is connected to the unusual repair. And flag the linked DRSG as least recommended for the number of runs obtained above. In the present example, when OO ₁ (high average critical dimension) occurs, the first subgraph SG1 will be recommended as the last option.

c. If the number of runs between two successive runs of SG1 is less than or equal to the median value of the run of SG1 (in this case, eta), remove the set of flags described above and implement the mechanism described in the section "Equipment State Changes Between PMs". Note that the user will have to execute SG1 if there is no resolution of OO1 (F) during the previous SG.

In addition, unstructured data can be mined by event chains, observation / metering (measurement) and resolution operations to form DRGs or multiple DRSGs. In addition, DRG-based ComET can store and retrieve information for further analysis. The analytical platform can be aided by a chain of statistical analysis or a more precise machine learning based analysis chain. In addition, the analytical accuracy and robustness (i. E., Good indicators) can be extended by constraining the solution space or by fine-tuning the model / structure of the analysis module. This can generally be assisted by including knowledge of the area. Once the necessary information base is formed to form the ES, the information may be asked in the future to quickly retrieve the necessary information.

The proposed method also provides a pipeline for automatically triggering the appropriate DRSG. For example, MES data is interpreted by including health / status information of the equipment and by monitoring time series data collected in real time by built-in sensor measurements.

In one embodiment of the present invention, the B / A RA analysis pipeline allows the sensor time series data to be learned before and after the resolution work and to focus on monitoring structured MES data. The discovery and characterization of causal relationships between resolution work and MES data is exploited in robust multi-resolution and multi-step inference pipelines capable of real-time fault and RA triggering at the wafer process level. In the process of the wafer W _B and W _A is set to Recipe R _e, equipment E _e are key events since the 'old' and 'of solving tasks (groups / chain) to correct the detected issue and due to the occurrence of M _e, respectively B &quot; and &quot; A &quot;, respectively. S ^B = {S _i ^B }, S ^A = {S _i ^A }, i = 1, ... , M, which may be a set of M sensors before and after the resolution work. The data recorded by the sensors may have numerical or categorical / nominal properties.

Sensor measurement - Numerical time series: Due to the heterogeneous nature of the sensor and process recipe, a single deviation numerical indicator does not fit the purpose. For example, because DTW-based techniques are not affected by bias and delay (time warping) in the time series, DTW-based techniques are based on the assumption that the time series is the change point (jump, , Continuous vibration, etc.). The time series is analyzed based on the CP so that the time series can be narrowed down to the applicability of the characteristic similarity index / algorithm based on the presence / absence of the change point (CP). Based on the CP, the proposed method defines two categories: Change Point Present (CPP) and Change Point Absent (CPA).

The time series is analyzed based on the following procedure.

)가 추정될 수 있다. 예를 들어, 노이즈 분포가 가우시안이라면, 노이즈 표준 편차 추정은 MAD/0.6745로 주어진다. 센서 시계열(ts)의 제1 차이(ts')를 계산한다.I. Calculate the median absolute deviation (MAD) of the wavelet coefficients at the finest level. Based on the shape of the noise distribution (e.g., Gaussian, Laplace, Poisson, uniform), the noise standard deviation

) Can be estimated. For example, if the noise distribution is Gaussian, the noise standard deviation estimate is given as MAD / 0.6745 . The first difference ts' of the sensor time series ts is calculated.

, 여기서 I는 진폭 스케일(시계열-값 축) 상의 계급 간격(class-interval) (히스토그램 빈, histogram bins)의 세트이고

이다. 계급 간격/빈 사이즈는 다음에 의해 계산된다:

II. The histogram of ts ' is given as:

, Where I is a set of class-intervals (histogram bin) on the amplitude scale (time series-value axis)

to be. Rank spacing / bin size is calculated by:

)에 집중된다. 만약 히스토그램의 0이 아닌 값이 이 영역의 외부에서 발견된다면, 점프(변경점)이 감지된다. 시계열은 CPP로 표시된다.III. In a smooth time series with no jump discontinuities, the histogram is a zero region (i.e.,

). If a nonzero value of the histogram is found outside this area, a jump (change point) is detected. Time series are denoted by CPP.

IV. Calculate ts '' (the second difference of ts ) and repeat the third and fourth steps. If a nonzero value is found outside the zero domain, the change is detected. Time series are denoted by CPP. In addition, the method can be used to detect high-order cusps (i.e., acceleration cusps), and higher order differences can be considered (i.e., the third difference ts ''' ).

V. If the condition is false, the time series is denoted by CPA.

A time series having CPP and CPA characteristics together with the first and second differences is shown in Fig. If the time series is classified as CPP or CPA, additional processing is performed. The method can also be used to detect and quantify any significant deviations before and after the repair of the time series. Since a significant number of mixed / heterogeneous tasks occur in a given recipe-run / wafer process, the single deviation score across the information contained in the time series of the entire recipe-run / But does not include information that can be used to distinguish the causes at.

이 i번째 센서에 대하여 N 개의 샘플/값들의 시계열 데이터 벡터인 것으로 한다. 해결 작업 전후의 i 번째 센서(S_i ^B 및 S_i ^A)의 시계열 데이터를 나타내기 위하여, 적절한 윗첨자ts_i ^B 및 ts_i ^A가 각각 표기에 추가된다. p번째 및 q번째 샘플 사이에서 분할

를 정의한다.For all practical purposes it is useful to narrow down the information contained in deviation scores from a time series of the same nature to a specific resolution. This partitioning can be assisted by including recipe sub-step information. Deviation scores can be calculated on a time series basis for each recipe sub-step. Although, in most cases, such a resolution level is sufficient, certain deviations may be revealed only in certain areas of the recipe sub-stage. This deviation and information about the area containing the information may disappear. To solve this problem, the proposed method utilizes a window-let based deviation detection method.

Is a time series data vector of N samples / values for the ith sensor. Appropriate superscripts ts _i ^B and ts _i ^A are added to the notation to indicate the time series data of the i- th sensor (S _i ^B and S _i ^A ) before and after the resolution work, respectively. Split between pth and qth samples

.

과

사이에서

로 계산되고,

의 윈도-렛 DTW(WDTW) 점수 행렬을 형성한다.WDTW method: Dynamic time distortion score

and

Between

Lt; / RTI &gt;

DTW (WDTW) &lt; / RTI &gt; score matrix of &lt; / RTI &gt;

의 기본값을 갖는 설정 가능한 수치)이다. 또한, W는 lcm(s_a)로 계산되되, 여기서 lcm은 최소 공배수를 의미하고, sa는 {K, K-1, K-2, …, l}의 세트이다. K는 스케일 해상도(scale-resolution)이고 W는 시간 해상도(temporal resolution)이다. k번째 열은 depth-k에서의 열로 지칭된다. depth-1 열은 전체 시계열 ts_i ^A 및 ts_i ^B 의 DTW 점수와 동일한 성분(entry)를 갖는 상수열(constant row)이다. depth-2 열은 두 개의 특이 성분(unique entry)을 갖는다.If the number of indexes used in the partition: N / x or (N / x) +1 is found to be a non-integer entity, then the number of indexes in a given column Adjustments are performed by applying a ceiling / flooring operation to explicitly cover the entire time series. The proposed window-and-let DTW matrix has columns K and W and K is a depth value (

Lt; RTI ID = 0.0 &gt; a &lt; / RTI &gt; Also, W is calculated as lcm (s _a ), where lcm is the least common multiple, and sa is {K, K-1, K-2, ... , l} . K is the scale-resolution and W is the temporal resolution. The kth column is referred to as the column at depth-k. The depth-1 column is a constant row with the same entry as the DTW score of the entire time series ts _i ^A and ts _i ^B. The depth-2 column has two unique entries.

와

로 계산된 DTW 점수와 동일한 값으로 지정된다. 유사하게, 다음 W/ 2 행은 ts_i ^A 및 ts_i ^B 사이의 다음 절반의 시계열 분할, 즉

와

로 계산된 DTW 점수와 동일한 값으로 지정된다. 일반적으로, 행렬의 (k, w)번째 성분은 다음과 같이 주어진다:The first W / 2 row is the time-series splitting of the first half between ts _i ^A and ts _i ^B ,

Wow

Is set to the same value as the DTW score calculated by the above equation. Similarly, the next W / 2 row is the time-series partition of the next half between ts _i ^A and ts _i ^B ,

Wow

Is set to the same value as the DTW score calculated by the above equation. In general, the (k, w) th component of the matrix is given by:

The heat-map diagram of the proposed WDTW matrix for the sensor time series data classified as CPA obtained before and after the thermal analysis describes the inference process during the monitoring step for defect detection and RA triggering as shown in FIG. The highest value of the elements of the WDTW matrix is represented by a pre-defined gray highlight area of the heat map diagram and a relatively low value is displayed by a relatively dark gray area.

)로 표시되는 시간 해상도를 의미한다. 세로축은 깊이 해상도를 의미하고 각각의 기본 단위는 깊이 인덱스(Depth Index; DI)(

)를 의미한다.As shown in FIG. 16, the abscissa indicates the time interval of each of the temporal resolution index (TRI) (

) &Lt; / RTI &gt; The vertical axis represents depth resolution, and each basic unit is represented by a depth index (DI) (

).

From 16, up to 6, and DI TRI amounts to 1 to lcm (6!) = 60. At a given depth index, the relationship between the TRI and the value shown in the heat map can be calculated as a DTW score using the above equation. For example, at DI = 1 , all values are necessarily interpreted as a single DTW score on a single window (i.e., in the entire time series). At DI = 2 , a TRI of 1 to 30 means the DTW score calculated in the first half of the time series and a TRI of 31 to 60 means the DTW score calculated in the next half of the time series. Fig. If you look at the last DI, namely six shown in Fig. 16, the maximum deviation between the digitized time series by the DTW score is observed in the window over the resolution index 20 to 30. This is referred to as Window-Of-Interest (WOI) in the proposed framework: domain (s) / zone (s) / duration of interest. The WOI may be a single or multiple adjacent windows or a separate window / zone.

If the value of WOI is OOS (out of specification) under conditions of two corresponding time series before and after each resolution work, the WOI and i-th sensor are referred to as "effect", the resolution task "arising" As time series data corresponding thereto. Therefore, a causal relationship between the resolution task (cause) and the time series deviation (effect) can be established. Details of how to investigate these causal relationships are described below. If the inverse causal relationship is also established to be true (with certain convictions) (ie, if the time series deviation of the ith sensor (cause) of a given WOI is mapped to a 'necessary resolution task' (effect) Can trigger the 'event' that the time series data of the i-th sensor can be monitored in real time in a continuous wafer process and a related workaround is needed if a deviation occurs. Given these criteria, the reasoning process (and the triggering of the resolution task) for defect detection is as follows.

1. The B / A analysis monitors the sensor generated by the cause and the B / A 'deviation' effect (the event that can be corrected by the resolution task) in the solution task and the causal relationship.

2. Monitoring is performed for a continuous wafer process with a sliding window in one wafer process and a WDTW matrix is constructed for each successive wafer process time series (related parameters such as recipe, chamber, etc. are fixed / Known).

3. Start at the lowest resolution (DI = 1 ) in the current WDTW matrix.

4. If the observed value is OOS (as compared to the value at DI = 1 of the WDTW matrix corresponding to the previous continuous wafer process), then from DI = 1 to DI = DI _max (6 in the drawing) To obtain the WOI at the highest resolution level (DI).

5. If the value of WOI is OOS compared to the corresponding value in the WDTW matrix corresponding to the previous continuous wafer process, then 'mark' the WOI. OOS can also be observed when there is a significant 'peak' when compared to adjacent values in a given DI (i.e., when the TRI of FIG. 16 is 20-30 and the values are 10-20 and 30-40 Compare with.

6. The score of the WOI (deviation score) is summed and passed to the assurance engine for the calculation of the deviation score combiner and the 'deviation detection' confidence. If the change / OOS is not detected, 0 is passed. If the sensor time series is of CPP nature, the WDTW operation / pipeline can not be applied and the value equal to zero is passed. This score is referred to as S _WDTW .

Due to the nature of the DTW score, the OOS boundary and threshold values are a single value in contrast to the two values (Upper-Spec-Limit (USL) and Lower-Spec-Limit (LSL) The OOS threshold should be set to a value that is the sum of the false alarms and the realities, such as the 3 sigma rule, the hypothesis testing framework (with maximum a posteriori (MAP) or maximum likelihood (ML) Can be set using all known standard methods of reducing the likelihood / probability of misses of events or critical events. The robustness of false alarms in OOS due to noise / Can be improved by operating a consensus time series before and after a single time series. More specifically, it can be calculated between time series data (i.e., corresponding to one wafer process) Unlike the decision based on the WDTW matrix, the WDTW matrix can be computed between the consensus time series data before the solution operation and the consensus time series data after the solution operation. The consensus time series is a simple average, a weighted average, a synchronized weighted averaging ), DTW-based merging, cluster centroids, etc. Considering the consensus time series reduces the effects of noise / outliers and thus ensures causality If appropriate, reliability can be further improved by examining similarly time-series data that are interdependent.

In order to find the CPP, the proposed method may be applied to various methods (e. G., Eigenvalue decomposition methods to detect significant temporal changes (i. E., X-axis, That is, the deviation representing the deviation of the value or the change of the slope / rate of change in the values of the Y axis, the value axis, and the amplitude axis).

The eigenvalue decomposition method is used to detect significant temporal changes (i.e., deviations in the X-axis, time axis, or index axis) using an SVD-based WGamma matrix.

SVD-based WGamma matrix: Define the block matrix S _i as

Here, ts _i ^A and ts _i ^B represent the time series data before and after the resolution work of the i-th sensor as a vector form. Through the synchronization process, the number of elements in the time series vector should be the same so that the metrics S _i can be formed. The heat of the Si must be linearly independent (within the tolerance limit) in order to allow the resolution task to have a significant impact on the time series deviation characteristics in terms of time delay, which may result in false alarm rate / probability and miss rate / May be determined based on a hypothesis-based threshold value calculation method that spans or past the historical data to a minimum.

1, but calculating the singular value decomposition (SVD) of S _i in [U D _{si si si} V] = SVD _(Si), where _si U (and its relative orthonormal (orthonormal counterpart) V _si) are embedded line ( column are singular vectors corresponding to the elements of the singular value-diagonal matrix D _si .

로 계산한다.2. The gamma-score (the ratio of the largest singular value to the smallest singular value of S _i )

.

가 문턱값보다 크다면, 편차가 감지된 것으로 확인된다.3. If the gamma score

Is greater than the threshold value, it is confirmed that the deviation is detected.

The presence of a plurality of change points and disparate temporal changes may result in erroneous setting of the threshold. The proposed method can be used to define a similar method to the WDTW matrix-based method. Defines the WGamma matrix.

Where each element is a gamma score computed as:

는 SVD에 의하여 계산된 특이값 매트릭스이다:here,

Is the singular value matrix computed by SVD:

Inference: Based on the WGamma matrix, reasoning, fault detection, and RA triggering can be identified using a heat map diagram. Elements of high WGamma are displayed in light gray and low elements are displayed in dark gray. The proposed scheme tracks light gray cells from the lowest resolution depth ( DI = 1 ) to the highest one ( DI = DI _max ). The WGamma matrix technique can be used to establish a causal relationship between the resolution work and the deviation in time series. If the inverse causal relationship is also true (with certain confidence) (ie, if the time series deviation at a given WOI (cause) is mapped to a 'required solution action (effect)', Event "in response to a deviation from the time-series data of the wafer 10. The WOI can be compared to a previous heatmap corresponding to the last batch of subsequent wafer processes, If there is no change or / OOS is observed, the sensor time series has a CPA property and is almost equal to zero (0), since the value of the indicated WOI (s) (deviation score) is merged and passed to the deviation score compositor and assurance engine. This value is defined as SW _Gamma .

The following process can detect the change point on the time axis and divide the time series data into a plurality of lower steps.

1. Calculate the first difference ts' and the second difference ts'' of the time series ts .

2. Using the MAD estimate of the wavelet coefficients of the time series at the highest level of resolution, calculate the noise floor three times the MAD estimate.

3. Detect the location of impulse peaks at ts ' and ts '' values greater than noise. A group of adjacent values may appear due to the impulse spread. This can be solved by setting the median of the adjacent values to the position of the peak.

여기서 α, β, δ는 ts에서 단차(step)/점프 및 속도 첨점(velocity cusp)를 표시하는 ts' 및 ts''에서 임펄스 피크의 시간축에서의 위치들이다. 또한, ts(p, q)는 인덱스 p 내지 q(p, q 포함)의 시계열의 구분을 나타낸다.4. Separate the time series from the position of the detected peak as follows:

Where α, β, and δ are positions on the time axis of the impulse peak at ts 'and ts', which represent the step / jump and velocity cusp at ts. Also, ts (p, q) represents a time series of indices p to q (inclusive of p and q).

Based on the statistical nature of the values of the time series in each sub-step, two processes can be applied to detecting significant value changes before and after the resolution. If the range of values is 'narrow', the proposed method can detect significant value changes using pseudo-clustering and FDI. If the range is 'wide', the proposed method utilizes a DTW + FDI based method to detect significant value changes. The time series sub-steps are classified into a 'narrow' or 'wide' range based on the kurtosis of the values of a given sub-step (shown in FIG. 9). If the kurtosis is greater than -1, the time series sub-steps are classified as a 'wide' range, otherwise they are classified as a narrow range.

Similar Clustering and FDI Scoring Method: In the case of 'narrow' range properties, the algorithm is described as follows.

1. Since the values in the partition ts _i ^B ( p, q ) and ts _i ^A ( p, q ) and the time series before and after the solution are derived from their corresponding divisions, ) - Grouping can be treated as pseudo-cluster because it is inherent. If the 'placement' in a one-dimensional space of a similar cluster is significantly different, there is a deviation between time series distinctions before and after repair.

2. Deviation scores can be calculated by FDI as follows:

3. The score can be obtained by any inter-cluster-intra-cluster distance measurement.

라면, 시계열 구분은 현저히 다른 것(즉, 해결 작업에 의해 영향을 받은 것으로)으로 표시될 수 있다. 용어

및

는 예를 들어, 앞서 기술된 MAD 기반 추정을 이용하여 추정하는 노이즈 플로어(표준 편차)이다.4. The higher the FDI score, the better the separation between similar clusters. if

, Then the time series breakdown may be marked as being significantly different (ie, affected by the resolution work). Terms

And

Is a noise floor (standard deviation) estimated using, for example, the MAD-based estimation described above.

DTW + FDI based method: In the case of 'wide' range characteristics, the DTW score can be used to detect whether two time series have significant deviations in the slope. The FDI score-based method can still be applied to the detection of bias induced deviations. Based on these points, the algorithm is described as follows.

One. ts _i ^B ( p, q ) and ts _i ^A (p, q) are represented by the separated time series before and after the resolution work, and the DTWs before these are calculated.

2. If the DTW score is higher than the threshold, there is a significant difference between the time series, and a causal relationship can be established between the resolution work performed and the time-series behavior during the appropriate lower step. The threshold values described above can be learned and fixed from past data or can be adjusted interactively. This method can be applied to detecting tilt deflection.

라면, 이는 편향으로부터 유도된 편차를 의미한다.3. Calculate the FDI score similar to the steps described above. if

, Which means a deviation derived from deflection.

In addition, the DTW method can not be applied to a time series sub-step with a 'narrow' range, and the method is based on the principle of bias / shift of values - a major cause of deviation in the time series of the 'narrow' Time series). The 'broad' range of time series sub-steps is also sensitive to slope variations / deflections. The DTW method is most suitable for differentiating these changes. In addition, FDI is used to detect deviations derived from deflection.

Inference: If the inverse causal relationship between significant deviations and resolution work is also true (with certain confidence), the time series deviations in a given sub-step (similar to WOI) can be mapped to the 'required resolution work'. Thus, the proposed method can be used to monitor real-time time-series data of successive wafers / processes and trigger the relevant resolution work events needed to generate deviations. Detection of the occurrence of the deviation (OOS) may be related to the value obtained from the analysis for the previous continuous wafer process.

가 편차 점수 조합기 및 확신 엔진에 전달된다. Course-I - The following deviation score

Is delivered to the deviation score compositor and the assurance engine.

가 편차 점수 조합기 및 확신 엔진에 전달된다. Course II - The following integrated deviation score

Is delivered to the deviation score compositor and the assurance engine.

Sensor measurement- categorical time series : FIG. 17 shows exemplary time series data by category. The proposed method describes the problem of narrowing the deviation to a specific resolution using a window-based technique. Since all or part of the time series / sequence by category can be treated as an array or character string as shown in Fig. 17, the deviation value is determined by string similarity metric - min-hash Technique. &Lt; / RTI &gt; Consider that the proposed method can be used to define an extended time series formed by the sum / concatenation of 2 and 3 grams as follows:

는

및

인 2그램이고,

은

,

및

인 3그램이다. n그램(n=2, 3)은 편차 수치를 계산에서의 차수(order) 정보를 보존한다. 예를 들어, 도 17로부터, 만약

라면,

이다. i번째 센서의 확장된 시계열 데이터에 있어 p번째 및 q번째 샘플 사이의(p, q 포함) 구분(partition)을 각각

및

로 나타낸다. min-hash에 의하여 계산된 유사도 값을

로 나타내고 따라서 편차 수치를

로 정의한다. 또한, 윈도-렛 MH(Min-Hash) 매트릭스를 다음과 같이 정의한다:here,

The

And

Lt; / RTI &gt; grams,

silver

,

And

3 grams. The n-gram ( n = 2, 3 ) preserves the order information in calculating the deviation values. For example, from Figure 17,

Ramen,

to be. In the extended time-series data of the i-th sensor, the (p, q) partition between the p-th and q-th samples is

And

Respectively. The similarity value calculated by min-hash

And thus the deviation value

. In addition, we define the window-min MH-matrix as:

이다.here

to be.

Based on the WMH matrix, inference and fault detection / RA triggering can be obtained using a heat map diagram.

Deviation Score _Compositor : The numeric sensor time series data, S _WDTW , S _WGamma , S _FDI and S _{FDI - DTW,} are combined by a deviation score combiner to form a deviation score vector of length 4: SD = [S _WDTW , S _WGamma , S _FDI and S _{FDI - DTW} ] This vector is passed to the assurance engine module for confidence-of-computed-score determination and final deviation score calculation. Depending on the manner in which the time series measurements are taken by category, the noise characteristics of the series / sequence are different. If a category-specific value is directly obtained and measured (for example, a switch on-off state), the measured noise may be treated as absent for all practical purposes. Conversely, if a numerical measurement is made and quantized to obtain a time series by category, the quantization noise is substantially uniform. Since the set of values affected by this noise is uniform over the entire interval, the raw deviation score can be treated as a final score without having to be convinced. Thus, S _WMH is provided directly to the sensor selection module for inference.

Confidence engine: A confidence engine is also referred to as a fuzzy-logic-based confidence engine. The fuzzy-logic-based confidence engine calculates the overall, robust, reliable and usable confidence score for triggering the resolution task, taking into account the dependency and dynamics of these deviation scores. The assurance engine includes an HDS unit and an FSG unit.

The HDS unit stores all historical data for analysis of sensor dynamics such as aging and long-term trend, outlier detection and exclusion, and local variation emmd. More precisely, (local) confidence scores are calculated by historical data and updated regularly (by periodic or manual triggering) by the HDS. These are inputs to the FSG module for the calculation of the final confidence score vector. It consists mainly of three units for the calculation of regional points: a sensor aggregation unit (α), a long term dynamics / aging model (β) and a short term dynamics model (γ).

The aggregating unit sorts the 'similar sensor' data for robust analysis (if present) and functions as an outlier exclusion unit. The 'similarity' rule is determined based on the equipment architecture / layout and sensor time-series characteristics (eg, correlation). This can be illustrated by the following example (with the degree of similarity being ' similarity '): placed at different locations in the chamber (front left, front right, rear left and rear right) Consider four temperature sensors with devices (actuators). In a given recipe under normal operating conditions, the time series data of all four sensors are highly &quot; highly correlated &quot;. If a particular temperature sensor (e.g., rear left) is OOS (out of spec) based on its current deviation score and the measurement of another temperature sensor is not OOS, then it is likely to be due to a defect in the sensor (front left) . The weights may be determined in proportion to the correlation value between the sensors. The correlation value comparison (high / low) is based on a comparison to the current value under a normal environment, rather than an absolute value. Also, if available, the same result can be confirmed using a thermostat input power monitor. Based on the above model, the data flow of the aggregation module is as follows:

1. The initial level of the set is obtained (if feasible) by a pair-wise string similarity measure (eg minhash, LSH, Levenstein edit distance) based measurement. Each sensor name and recipe name are concatenated to form a recipe-sensor name (string). Based on the similarity between the synthetic strings, the sensors (in consideration of the recipe) are bunched. This ensures that certain temperature sensors are tied together in a first approximation but the pressure and temperature sensors are not tied together. For example, a Recipe sensor name: Recipe1_ChamberRearLeftTemperatureMonitorValue, Recipe1_ChamberRearRightTemperatureMonitorValue, Recipe1_ChamberFrontLeftTemperatureMonitorValue and Recipe1_ChamberFrontRightTemperatureMonitor is being tied not Recipe1_ChamberPressureMonitor, Recipe1 _ RFForwardPower comprises the formed group.

The final cluster / group of sensors in the character space represents a 'similar' sensor under typical operating conditions. This representation is quantified by calculating the centroid and ISID (the ratio of the distribution within the cluster and the minimum value of the distance in the cluster) for each cluster. This can be used as a basis to ensure that the current sensor time-series characteristic does not show the same measurement value as any outliers. For the current sensor time series characteristic, based on the position placed in the characteristic space, the weight is calculated as follows:

Where d (C, S) is the distance between the current sensor time series and the center of the closest cluster / group (e.g., Euclidean distance, absolute value difference).

Alternatively, a rule-based set model derived from domain information (if available) may also be applied. In these scenarios, the scores are defined by an expert.

Aging model / long-term dynamics model _: The aging model tracks the long-term trend of the deviation score of each sensor. In contrast to aggregation units where the operating area is a comparison between sensors, the aging model uses a comparison within the sensor. Unlike the aggregate unit score α derived from each sensor, the aging model score is derived according to the kind of deviation (combined by the deviation score combiner) β _i where i = { WDTW , WGamma , FDI, FDI- DTW } There are four types of deviation scores generated. The aging model of each deviation score is calculated through a regression analysis (e.g., nonlinear) using historical data (the previous value of the deviation score).

Consider the previous value of the deviation score as time series data having the sequence number of the wafer process as a time series index. In order to model the long term trend, the deviation scores of continuous / adjacent wafer processes, including short-term dynamics, are not taken into account and rather the 'decimated value' (the value of the P wafer process selected in the interval) Where the value of P may be determined based on the equipment (part) aging model.

In one embodiment of the present invention, the average of the deviation scores for the P wafer process can be regarded as a single value of time series for estimation of regression model parameters. Based on the regression model extrapolation, the current value of the deviation score is expected. The predicted value is compared to the calculated value of the deviation score for the current consecutive wafer process pair considered. Their differences (if any) are quantified as follows:

Here, i = {WDTW, WGamma, FDI, FDI- DTW} a, d (P _i, S _i) i is the distance between the predicted value and the calculated value of the second deviation score (e.g. the Euclidean distance, absolute distance Etc.).

Short-term dynamics model: The short-term dynamics model quantifies short-term sensor variation score variance. The quantification of dynamics is performed using the change in i-th deviation score and is called γ _i ( i = {WDTW, WGamma, FDI, FDI-DTW} ). It is clear that γ _i has a range of [ 0 , ∞). Even if the likelihood of change is close to ∞, theoretically it can occur and the likelihood of occurrence is close to zero when considering well-controlled equipment operations. The gamma _i is conditioned by mapping to a high value of 1, 0 so that the range lies within [0, 1]. This can be achieved by techniques such as the histogram transformation technique.

Fuzzy overall score screen (Fuzzy score - globalization ; FSG unit: The FSG unit is a fuzzy inference system that receives α, β _i and γ _i and outputs a global confidence C _i ( i = {WDTW, WGamma, FDI, FDI- DTW } The crisp logic threshold is based on the open-ended range and the nature of the points that need to be continuous and not necessarily uniformly distributed, (L), medium (M), and high (L), the middle range (M), and the high range (H) scores correspond to high overall convictions, however, with β and γ within the H range, it is possible to determine which of the M ranges α, and if it can be known, occasionally periodic adjustments Considering that the FIS is a "universal approximator", the FIS system has a decision surface that allows a given edge / boundary / It is possible to interpolate the entire range of the set of central rules, which motivate to suggest a fuzzy logic-based score wholezing unit contrary to a clear logic-based unit. Purpose (for example, a set of rules) and may be changed based on knowledge in the field, if necessary.

alpha beta _i γ _i C _i L L L L L L H L L H L L L H H M H L L L H L H M H H L M H H L H

를 제공한다.The proposed secure FSG based score-based generalization unit is shown in FIG. 18 with the proposed rule-based technique. The relationship of the FSG surface ( C _i variation) for α and β _i at γ _i = 0, 0.25, 0.5, 0.75 and 1.0, respectively, in a given rule base is shown in FIG. C i obtained from the FSG is the overall confidence score. This is multiplied by the respective deviation score to obtain the final deviation score vector

Lt; / RTI &gt;

Figure 19b is a schematic diagram of a system that provides recommendations for deviations within the equipment based on video data. In one embodiment of the invention, the video is converted into a time series signal and text information. Time series signals are sent for B / A analysis and text information is sent for DRsG formation. The proposed system can be used to provide additional capabilities that can be indexed and to navigate the video based on the tagged text.

20 illustrates a computing environment 2002 that implements a mechanism for root cause analysis of equipment-related deviations in accordance with an embodiment of the present invention. The computing environment 2002 includes at least one processing unit 2008 comprising a control unit 2004, an arithmetic logic unit (ALU) 2006, a memory 2010, a storage unit 2012, a plurality of networking Device 2016 and a plurality of input / output devices 2014. [ The processing unit 2008 receives an instruction from the control unit 2004 and performs processing. In addition, any logic and arithmetic operations associated with the execution of the instructions are computed with the aid of the ALU 2006.

The entire computing environment 2002 may include a plurality of homogeneous or heterogeneous cores, a plurality of different types of CPUs, special media and other accelerators. The processing unit 2008 processes the instructions of the above description. Further, the plurality of processing units 2008 may be arranged over a single chip or a plurality of chips.

Techniques including instructions and code necessary for implementation are stored in memory unit 2010 or storage unit 2012 or both. At execution time, the instructions may be fetched from the corresponding memory 2010 or storage 2012 and executed by the processing unit 2008. [

In the case of a hardware implementation, various networking devices 2016 or external input / output devices 2014 may be coupled to the computing environment 2002 to support implementation through the networking and input / output device units.

Thus, the proposed mechanism wipe out the memory of the associated resolution work when unusual repair is performed, and then builds the data when the state of the equipment returns to its original state in a gradual manner.

It will be understood by those of ordinary skill in the art to which this invention pertains that the description of the claims of the invention is illustrative and that numerous variations, changes, variations, substitutions and equivalents may be resorted to, You will understand that you can. In addition, those of ordinary skill in the art will appreciate that the use of the term is domain-specific and that the invention will vary depending upon the implementation in a particular industry.

Embodiments of the present invention may be implemented through at least one software program that runs on at least one hardware device and performs network management functions that control the components. The components shown in Figs. 1 to 20 include at least one hardware device, or a block that can be configured as a combination of a hardware device and a software module.

The description of the foregoing embodiments will reveal the general nature of embodiments of the present invention and that others may be readily modified and / or adapted for various uses without departing from the general concept by applying current knowledge, And should be understood to be within the meaning and range of embodiments of the present invention. It is to be understood that the phraseology and terminology employed herein is for the purpose of description and not of limitation. Therefore, while the preferred embodiments of the present invention have been described, it will be appreciated by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present invention as defined by the following claims.

100: System 102: SeRA entity
104: DRG entity 106: Recommendation entity
108: ComET 200: Equipment

Claims (32)

A Sensor Resolution Action (SeRA) entity that detects deviations associated with at least one device;
A DRG entity for resolving deviations based on at least one Deviation Resolution Graph (DRG) by determining at least one Deviation Resolution sub-graph (DRsG), wherein the at least one DRsG comprises at least one A DRG entity comprising a Resolution action and at least one observation associated with the at least one device; And
And a recommendation entity for recommending the at least one DRsG to resolve the deviation associated with the at least one device. The method according to claim 1,
Wherein the observation is one of an original observation and a secondary observation. The method according to claim 1,
Wherein the resolution work includes one of a repair job, a secondary action, a resolution job, and a reset job, the system for root cause analysis of deviations associated with at least one device. The method according to claim 1,
Wherein the DRG is formed based on data received from a plurality of sources, the data comprising at least one of structured data and unstructured data, System for Cause Analysis. The method according to claim 1,
Wherein the deviation is identified based on a SeRA graph, the SeRA graph representing a relationship between a group of equipment and the at least one DRG. 6. The method of claim 5,
Wherein the SeRA graph is formed based on a before-after analysis of the DRsG set of the at least one device. The method according to claim 6,
Wherein said back-and-forth analysis identifies said at least one piece of equipment comprising a signature of a repair operation and forms a causal relationship stored for automated recommendation of said at least one DRsG, A system for root cause analysis of deviations associated with. 8. The method of claim 7,
The causal relationship recommending the at least one DRsG is:
Calculating a deviation score based on whether or not the time series signal of said at least one equipment includes a change point,
Determining a confidence score based on the deviation score, by a confidence engine,
Determining a deviation vector based on the confidence score, by a fuzzy score globalize,
A root cause analysis of the deviations associated with at least one equipment, formed by establishing a causal relationship between the deviation and the at least one equipment based on the SeRA graph, by a sensor selector, For the system. 9. The method of claim 8,
When the change point is unavailable in the time series signal, a window-let dynamic time warping (WDTW) is performed to calculate the deviation score of the time series signal of the at least one equipment A system for root cause analysis of deviations associated with at least one equipment. 9. The method of claim 8,
Wherein the window-let gamma (WGamma) is performed to calculate the deviation score of the time series signal of the at least one piece of equipment when the change point is available in the time series signal A system for root cause analysis of deviations. 9. The method of claim 8,
When the change point is available in the time series signal, a Fisher discriminant index (FDI) and a dynamic time warping (DWT) are performed to calculate the deviation score of the time series signal of the at least one equipment A system for root cause analysis of deviations associated with at least one instrument to calculate. 9. The method of claim 8,
Wherein the confidence score is calculated based on a long term trend and a short term aging trend, wherein the confidence score is calculated based on a long term trend and a short term aging trend. 9. The method of claim 8,
Wherein the at least one DRsG is triggered by the sensor selector when the deviation vector satisfies a threshold. 9. The method of claim 8,
Wherein the time series signal is obtained from video data of the at least one piece of equipment. 15. The method of claim 14,
Wherein the SeRA entity provides at least one recommendation for a deviation of the at least one device based on the video data. The method according to claim 1,
A system for root cause analysis of deviations associated with at least one instrument, further comprising a Comprehensive Elucidation Table (ComET) for storing SeRA graphs and DRGs. A sensor resolution action (SeRA) entity detects deviations associated with at least one device,
Determining at least one Deviation Resolution sub-graph (DRsG) by a Deviation Resolution Graph (DRG) entity to resolve the deviation based on at least one DRG, DRsG comprises at least one resolution operation and at least one observation associated with the at least one device,
And recommending the at least one DRsG by a recommendation entity to resolve the deviation associated with the at least one device. 18. The method of claim 17,
Wherein the observation is one of an original observation and a secondary observation. 18. The method of claim 17,
Wherein the resolution work includes one of a repair job, a secondary job, a resolution job, and a reset job. 18. The method of claim 17,
Wherein the DRG is formed based on data received from a plurality of sources, the data being at least one of structured data and unstructured data. 18. The method of claim 17,
Wherein the deviation is identified based on a SeRA graph, the SeRA graph representing a relationship between a group of equipment and the at least one DRG. 22. The method of claim 21,
Wherein the SeRA graph is formed based on the anteroposterior analysis of the DRsG set of the at least one piece of equipment. 22. The method of claim 21,
Wherein the context analysis identifies the at least one piece of equipment that includes a signature of a repair operation and forms a causal relationship that is stored for automated recommendation of the at least one DRsG. Lt; / RTI &gt; 24. The method of claim 23,
The causal relationship recommending the at least one DRsG is:
Calculating a deviation score based on whether or not the time series signal of said at least one equipment includes a change point,
Determining a confidence score based on the deviation score, by a confidence engine,
The fuzzy score globalizer determines a deviation vector based on the confidence score,
A method for root-cause analysis of deviations associated with at least one equipment, which is formed by a sensor selector, by establishing a causal relationship between the deviation and the at least one equipment based on the SeRA graph. 25. The method of claim 24,
Wherein the windowed dynamic time warping (WDTW) is performed to calculate the deviation score of the time series signal of the at least one piece of equipment when the change point is not available in the time series signal Methods for root cause analysis. 25. The method of claim 24,
If the change point is available in the time series signal, a window-lead gamma (WGamma) is performed to calculate the deviation score of the time series signal of the at least one piece of equipment. Lt; / RTI &gt; 25. The method of claim 24,
(FDI) and dynamic time warping (DWT) are performed to calculate the deviation score of the time series signal of the at least one piece of equipment when the change point is available in the time series signal A method for root - cause analysis of related deviations. 25. The method of claim 24,
Wherein the confidence score is calculated based on a long term trend and a short term trend trend, the method for root cause analysis of deviations associated with at least one equipment. 25. The method of claim 24,
Wherein the at least one DRsG is triggered by the sensor selector when the deviation vector satisfies a threshold. 25. The method of claim 24,
Wherein the time series signal is obtained from video data of the at least one piece of equipment. 31. The method of claim 30,
Wherein the SeRA entity provides at least one recommendation for a deviation of the at least one device based on the video data. 18. The method of claim 17,
A method for root cause analysis of deviations associated with at least one instrument, the method further comprising storing SeRA graphs and DRGs by a comprehensive descriptive table (ComET).

Images