TWI786788B - Diagnosis system of machine - Google Patents

Abstract

A diagnosis system of machine is used to diagnose a machine. The diagnosis system of machine includes a database of big data, a sensing device, and a processor. The database of big data stores at least one recognition model of wafer. The sensing device is configured to sense time-dependent characteristics when the machine processing a first wafer, and correspondingly generate a first time-dependent characteristic data. The processor is configured to update the at least one recognition model of wafer according to the first time-dependent characteristic data.

Description

Machine Diagnosis System

The present disclosure relates to a machine diagnosis system, especially a system for analyzing and diagnosing problems of semiconductor manufacturing machines.

As the semiconductor manufacturing process continues to shrink, the requirements for product quality and machine stability are getting higher and higher. At present, semiconductor factories mainly introduce Fault Detection and Classification System (FDC) into the wafer manufacturing process. Collect the data reported by each sensor on the machine (for example: gas flow, pressure, voltage, current, etc.).

However, semiconductor machines are highly precise and have many sensors. In order to effectively monitor the status of the machine, most sensors collect data at a fast frequency, so the data is huge. The monitoring method of the current FDC system is to collect summary values (average value, maximum value, minimum value, area, etc.) for specific sensor report data collection section data during the manufacturing process of each wafer for specification card control, follow-up data Analytics can also only be processed analytically through aggregated values.

In addition, the current FDC system collects the reported values of many sensors on the machine, but generally only sets summary values for some important parameters for data collection and subsequent data analysis and processing, and cannot comprehensively analyze the parameters of all sensors. Sexual analysis, so will lose a lot of information of the original data.

The summary data needs to be pre-defined by the engineer to collect the sensor summary value, which is usually set based on the engineer’s experience, but the causes of machine abnormalities are ever-changing, and the existing sensor summary value may not be able to monitor the abnormality. The cause of the problem can only be found after abnormal analysis.

In view of this, one purpose of the present disclosure is to propose a machine diagnosis system capable of solving the above-mentioned problems.

In order to achieve the above purpose, according to an embodiment of the present disclosure, a machine diagnosis system includes a big data database, a sensing device, and a processor. The big data database stores at least one wafer identification model. The sensing device is configured to sense time-dependent characteristics when the machine processes the first wafer, and correspondingly generate first time-dependent characteristic data. The processor is configured to update at least one wafer identification model based on the first time-dependent feature data.

In one or more embodiments of the present disclosure, the sensing device is further configured to sense the time-dependent characteristics of another machine when processing the second wafer, and correspondingly generate second time-dependent characteristic data, and the processor further It is configured to: compare the first time-dependent feature data with the second time-dependent feature data according to at least one wafer identification model, and generate a comparison result correspondingly; and generate visual chart information according to the comparison result.

In one or more embodiments of the present disclosure, the big data database stores a plurality of time-dependent characteristic data about processing a plurality of wafers, and the processor is further configured to divide the plurality of time-dependent characteristic data into several groups ; respectively calculating average values of several groups; and generating at least one wafer identification model according to the several average values.

In one or more implementations of the present disclosure, the processor is further configured to: find an average value closest to the first time-dependent feature data, and classify the first time-dependent feature data into a group corresponding to the average value; recalculating the group average; and updating at least one wafer identification model based on the recalculated average.

In one or more embodiments of the present disclosure, the at least one wafer identification model includes a normal wafer identification model, the big data database stores a plurality of time-dependent characteristic data about processing a plurality of wafers, and the processor is further configured To: mark several time-dependent normal characteristic data from several time-dependent characteristic data according to the marking data; generate a normal wafer identification model according to the several time-dependent normal characteristic data; and generate a normal wafer identification model according to the normal wafer identification model and the first time-dependent The feature data judges whether the machine is abnormal when processing the first wafer.

In one or more embodiments of the present disclosure, at least one wafer identification model includes an abnormal wafer identification model, the big data database stores several time-dependent characteristic data about processing several wafers, and the processor is further configured To: mark several time-dependent abnormal characteristic data from several time-dependent characteristic data according to the marking data; generate an abnormal wafer identification model according to the several time-dependent abnormal characteristic data; and according to the abnormal wafer identification model and the first time-dependent The feature data judges whether the machine is normal when processing the first wafer.

To sum up, in the machine diagnosis system disclosed in this disclosure, through the establishment of a big data database, the time-dependent original data from semiconductor machines can be comprehensively processed. In the machine diagnosis system disclosed in this disclosure, the difference analysis and comparison between machines can be performed through the processor. In addition, the processor can group several time-dependent feature data through unsupervised machine learning, so as to achieve the purpose of identifying the wafers processed by the machine as normal or abnormal. Moreover, the processor can mark normal wafers and abnormal wafers in the system in advance through engineers, and mine abnormal wafers based on the normal wafer identification model corresponding to the normal time-dependent characteristic data of normal wafers, or based on the corresponding Abnormal wafer identification model based on abnormal time-dependent feature data to mine normal wafers.

The above description is only used to explain the problems to be solved by the present disclosure, the technical means to solve the problems, and the effects thereof, etc. The specific details of the present disclosure will be introduced in detail in the following implementation methods and related drawings.

The following will disclose multiple implementations of the present disclosure with diagrams, and for the sake of clarity, many practical details will be described together in the following description. However, it should be understood that these practical details should not be used to limit the present disclosure. That is to say, in some implementations of the present disclosure, these practical details are unnecessary. In addition, for the sake of simplifying the drawings, some well-known structures and components will be shown in a simple and schematic manner in the drawings. The same reference numbers will be used throughout the drawings to refer to the same or similar elements.

Please refer to FIG. 1 , which is a schematic diagram of a machine diagnosis system 100 according to an embodiment of the present disclosure. As shown in FIG. 1 , the machine diagnosis system 100 of this embodiment includes a big data database 110 , a sensing device 120 and a processor 130 . The big data database 110 is configured to store at least one wafer identification model 112 and several time-dependent feature data CDs. The wafer identification model 112 includes a normal wafer identification model 113 and an abnormal wafer identification model 114 . The time-dependent characteristic data CD includes first time-dependent characteristic data CD1 and second time-dependent characteristic data CD2.

Please refer to FIG. 2 , the sensing device 120 is configured to sense the time-dependent characteristics when the first machine MC1 processes the first wafer W1 , and correspondingly generate first time-dependent characteristic data CD1 . The sensing device 120 is further configured to sense the time-dependent characteristics when the second machine MC2 processes the second wafer W2, and correspondingly generate second time-dependent characteristic data CD2. Alternatively, the sensing device 120 is further configured to sense the time-dependent characteristics of several wafers processed by several tools, and correspondingly generate several time-dependent characteristic data CDs. The processor 130 is configured to update the wafer identification model 112 according to the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2. The processor 130 is further configured to compare the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2 according to the wafer recognition model 112 , and correspondingly generate a comparison result CR. The processor 130 is further configured to generate visual chart information V according to the comparison result CR.

Various implementations of the machine diagnosis system 100 of the present disclosure will be described in detail below.

Implementation method (1): Implement machine difference analysis and comparison.

Please refer to FIG. 2 , in this embodiment, the plant has a first machine MC1 and a second machine MC2 with the same model and specification for processing the first wafer W1 and the second wafer W2 respectively. The first machine MC1 has a sensing device 120, and the second machine MC2 has another sensing device 120. The above-mentioned sensing device 120 is configured to respectively sense the time-dependent characteristics of the first machine MC1 processing the first wafer W1 and The time-dependent feature when the second machine MC2 processes the second wafer W2 is sensed, and the first time-dependent feature data CD1 and the second time-dependent feature data CD2 are correspondingly generated.

In some embodiments, the first wafer W1 is substantially several wafers, and the second wafer W2 is substantially several wafers.

In some embodiments, so-called time-dependent feature data is a function of changes in features over time (ie, a time series). For example, the time-dependent characteristic data may be a time series of voltage values. The above are examples for simple illustration only, and are not intended to be limiting.

In some embodiments, for example, the time-dependent feature includes at least one of features such as gas flow of the tool when processing the wafer, pressure of the processing chamber, temperature of the processing chamber, magnetic field, voltage, and current. Therefore, the above-mentioned time-dependent feature data includes at least one of the gas flow value, the pressure value of the processing chamber, the temperature value of the processing chamber, the magnetic field value, the voltage value, and the current value. It should be noted that the present disclosure is only an example for simple description, but not intended to be limiting.

The method 300 for implementing machine difference analysis and comparison will be described in detail below. The method 300 includes steps S301, S302, S303, S304, S305, and S306.

Please refer to Figure 3. Firstly, execute S301: receive first time-dependent characteristic data CD1 and second time-dependent characteristic data CD2. Specifically, after receiving the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2, the processor 130 stores the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2 in the big data database 110 .

Next, execute S302: generate the wafer identification model 112 according to the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2. Specifically, the processor 130 generates the wafer identification model 112 correspondingly according to the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2. In some embodiments, the wafer identification model 112 is a classification standard used to determine whether a wafer is a normal wafer or an abnormal wafer. More specifically, the processor 130 calculates the first time-dependent feature data CD1 and the second time-dependent feature data CD2 according to the wafer identification model 112 (for example, linear regression (linear regression), k-means algorithm etc.), so as to determine whether the processing of the first wafer W1 by the first machine MC1 and the processing of the second wafer W2 by the second machine MC2 are normal or abnormal.

In some implementations, the processor 130 can also correspondingly generate the normal wafer identification model 113 or the abnormal wafer identification model 114 according to the first time-dependent feature data CD1 and the second time-dependent feature data CD2 . The normal wafer identification model 113 or the abnormal wafer identification model 114 is generated based on an algorithm. The detailed method for generating the normal wafer identification model 113 or the abnormal wafer identification model 114 will be described in the following embodiments (2), Paragraphs of Embodiment (3) and Embodiment (4) are described in detail.

Next, step S303 is executed: receiving another first time-dependent characteristic data CD1 and another second time-dependent characteristic data CD2. Specifically, the first machine MC1 and the second machine MC2 process the next batch of first wafers W1 and the next batch of second wafers W2 respectively. The sensing device 120 senses the time-dependent characteristics of the next batch of first wafers W1, and correspondingly generates another first time-dependent characteristic data CD1. Another sensing device 120 senses the time-dependent characteristics of the next batch of second wafers W2, and correspondingly generates another second time-dependent characteristic data CD2.

Next, step S304 is executed: updating the wafer identification model 112 according to another first time-dependent characteristic data CD1 and another second time-dependent characteristic data CD2. Specifically, after the processor 130 receives another first time-dependent characteristic data CD1 and another second time-dependent characteristic data CD2, it uses an operation similar to step S302 to process all first time-dependent characteristic data CD1 and all second time-dependent characteristic data. Calculations are performed according to the time characteristic data CD2 to update the wafer identification model 112 .

More specifically, when the processor 130 receives another first time-dependent characteristic data CD1, it adds another first time-dependent characteristic data CD1 to the group of the first time-dependent characteristic data CD1. When the processor 130 receives another second time-dependent characteristic data CD2, it adds another second time-dependent characteristic data CD2 to the group of the second time-dependent characteristic data CD2. Then, the processor 130 recalculates the updated first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2 to obtain a new wafer identification model 112 , that is, the update of the wafer identification model 112 is completed.

In some implementations, after receiving another first time-dependent feature data CD1 and another second time-dependent feature data CD2, the processor 130 uses operations similar to step S302 to process all first time-dependent feature data CD1 and all The second time-dependent feature data CD2 is calculated to update the normal wafer identification model 113 and the abnormal wafer identification model 114 .

Next, step S305 is executed: comparing the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2, and correspondingly generating a comparison result CR. Specifically, the processor 130 compares the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2 judged according to the wafer identification model 112 , and correspondingly generates a comparison result CR. More specifically, the comparison result CR includes: the processor 130 judges that each first wafer W1 and each second wafer W2 are normal wafers or abnormal wafers; Statistical results of the time-dependent feature data CD1 and the second time-dependent feature data CD2; and results of judging whether a problem occurs with the first machine MC1 or the second machine MC2. For example, the statistical results include the correlation coefficient (for example, R value or R2 value) between the first time-dependent feature data CD1 and the ^second time-dependent feature data CD2 in each set of data, the first time-dependent feature data CD1 and the second time-dependent feature data The significant difference between the two time-dependent characteristic data CD2, the time point when the maximum difference appears between the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2, the first time-dependent characteristic data CD1 and the second time-dependent characteristic data The maximum or minimum value of the characteristic data CD2, etc. Specifically, each set of data includes at least one first time-dependent characteristic data CD1 and one second time-dependent characteristic data CD2.

The following is a simple example: In a usage scenario, when the processor 130 reads that the second time-dependent characteristic data CD2 of each set of data is significantly different from the first time-dependent characteristic data CD1 at a certain time point, and according to the When the wafer recognition model 112 is judged as normal/abnormal, abnormal/normal or abnormal/abnormal respectively (for example, when the processor 130 has judged that the first wafer W1 is a normal wafer (or abnormal wafer) and judged that the second wafer W1 is Under the premise that W2 is an abnormal wafer, the temperature of the processing chamber of the second machine MC2 is always 20 degrees Celsius higher than the temperature of the processing chamber of the first machine MC1 between the 7th second and the 8th second above), it is judged that there is a problem in the second machine MC2, especially in the process between the 7th second and the 8th second.

Next, step S306 is executed: generating visual chart information V according to the comparison result CR. Specifically, the processor 130 may draw the visual chart information V according to the comparison result CR. In some implementations, the visual chart information V includes several line graphs and several tables. For example, the processor 130 can draw the statistical result list of the first time-dependent feature data CD1 and the second time-dependent feature data CD2 into a table, and display the first wafer W1 and the second wafer W1 in each set of data. W2 is judged as the result of a normal wafer or an abnormal wafer. The processor 130 may also output each group of data that has a significant difference between the read first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2, and display the difference between the first machine MC1 and the second machine. MC2 judges it as the result of a problem. The processor 130 can also draw the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2 into a line graph, so as to facilitate the engineers to read and take further measures on the machines with problems.

In some implementations, the first machine MC1 and the second machine MC2 can be any machinery and equipment used to manufacture wafers or process wafers. The model specification of machine MC2 is limited.

By executing steps S301, S302, S303, S304, S305, and S306, the machine diagnosis system 100 can determine the first machine from the comparison between the first time-dependent characteristic data CD1 and the second time-dependent characteristic data CD2 Whether there is a problem with MC1 or the second machine MC2.

Embodiment (2): Using unsupervised machine learning (un-supervised learning) to judge the wafer as a normal wafer or an abnormal wafer.

In this embodiment, the factory has tools for processing several wafers. The tool has a sensing device 120 for sensing a plurality of time-dependent characteristics of a plurality of wafers, and correspondingly generating a plurality of time-dependent characteristic data CD. Several time-dependent feature data CDs are stored in the big data database 110 .

Next, the processor 130 divides the time-dependent feature data CD into several groups. In this operation, the number of groups can be set by the engineer in the machine diagnosis system 100. After the engineer inputs the number of groups to be classified, the processor 130 groups the time-dependent feature data CD according to the required number of groups. (clustering), but the disclosure is not limited thereto. In some implementations, the processor 130 can calculate the time-dependent characteristic data CD by itself, and then divide them into several groups according to distinguishing several main characteristics of the time-dependent characteristic data CD. In this embodiment, the number of groups is two, but this disclosure does not intend to limit the number of whole groups.

Hereinafter, the method 400 of how the processor 130 groups the time-dependent feature data CD will be described in detail by taking the implementation in which the number of groups is two as an example. The method 400 includes steps S401, S402, S403, S404, S405.

Please refer to Figure 4. First, execute step S401: determine the number of groups. Specifically, in some implementations, the engineer can set the machine diagnosis system 100 to set the number of groups to be classified as two.

Next, step S402 is executed: using an algorithm to calculate the time-dependent characteristic data CD to establish the wafer identification model 112 . Specifically, the processor 130 calculates the time-dependent feature data CD by using a clustering algorithm (for example, a k-means clustering algorithm). More specifically, the processor 130 randomly selects two time-dependent feature data CDs, and assigns them as the mean values of each group, thereby generating the wafer identification model 112 . In some implementations, the number of time-dependent feature data CDs randomly selected by the processor 130 is substantially equal to the number of groups.

Next, step S403 is executed: another time-dependent feature data CD is randomly selected for calculation. Specifically, the processor 130 randomly selects the next time-dependent characteristic data CD, and calculates the difference between the time-dependent characteristic data CD and the average values of the two groups, so as to find the time-dependent characteristic data CD closest to the time-dependent characteristic data CD. an average value of .

Next, step S404 is executed: grouping another time-dependent characteristic data CD. Specifically, the processor 130 classifies the time-dependent characteristic data CD into the group corresponding to the closest average value.

Next, step S405 is executed: updating the wafer identification model 112 . Specifically, the processor 130 recalculates the average value of the group to obtain a new average value of the group, and updates the wafer identification model 112 according to the new average value obtained from the recalculation.

Through the above steps, the processor 130 can complete the basic grouping operation on a time-dependent feature data CD.

In some implementations, the processor 130 will repeatedly execute steps S403, S404, and S405, and make each time-dependent feature data CD be classified into the above two groups through the repeated execution of the calculation of the grouping and classification algorithm. one of the groups. It should be noted that, in step S403, the processor 130 may select the time-dependent feature data CD that has been calculated by the clustering and classification algorithm, and execute steps S404 and S405 again to re-group the time-dependent feature data CD. If the processor 130 repeatedly executes steps S403 , S404 , and S405 so that each time-dependent feature data CD does not change the group to which it belongs, the processor 130 completes the grouping of the time-dependent feature data CDs.

Finally, the engineer can judge which group is the group to which the normal wafer belongs or the group to which the abnormal wafer belongs according to his experience. For example, it can be judged according to the graphical features presented by the time-dependent characteristic data CD on the two-dimensional line graph (for example, the temperature time series in a group all present a peak value in a certain period of time, and this peak value should not appear theoretically. In the temperature time series corresponding to the processing of normal wafers by the machine now).

In some implementations, step S403 may include multi-layer calculations. Specifically, in step S403, not only the clustering calculation such as the k-means clustering and classification algorithm may be implemented, but the second or more layers of the algorithm may be implemented successively. For example, after the k-means grouping and classification algorithm is calculated, the calculation can be performed by finding the peak value or trough value feature generated in a specific time period, so as to complete the grouping of the time-dependent feature data CD. In a word, the present disclosure does not intend to limit the type, frequency and sequence of algorithms implemented for grouping the time-dependent feature data CDs.

Embodiment (3): Mining abnormal wafers based on time-dependent normal feature data.

In this embodiment, the same machine base as in Embodiment (2) is used, and the machine base has the same sensing device 120 as in Embodiment (2). The time-dependent characteristics of several wafers are sensed by the sensing device 120 , and a plurality of time-dependent characteristic data CDs are correspondingly generated. Several time-dependent feature data CDs are stored in the big data database 110 .

In this embodiment, engineers manually mark several time-dependent normal feature data corresponding to normal wafers in all time-dependent feature data CDs, and then learn through the principle of machine learning, so that the machine diagnosis system 100 Anomalous wafers can be unearthed.

The method 500 for mining abnormal wafers based on time-dependent normal feature data will be described in detail below. The method 500 includes steps S501, S502, and S503.

Please refer to Figure 5. Firstly, execute step S501: mark several time-dependent normal characteristic data from several time-dependent characteristic data CD according to the marking data. Specifically, the processor 130 receives several time-dependent feature data CDs and stores them in the big data database 110, and then the engineer sets up the machine diagnosis system 100, and the setting is the marking data input by the engineer. For example, the labeling data may be a value range (for example, a voltage range input) of the time-dependent characteristic data CD that should be judged as a normal wafer, and the value range is one of the above-mentioned labeling data. Then, the processor 130 marks several time-dependent normal characteristic data by using the marking data. It should be noted that this disclosure does not intend to limit the type of marking data, more specifically, as long as it is any number, function (formula) that can standardize the time-dependent characteristic data CD that should be judged as a normal wafer Or computer programming syntax is within the spirit and scope of this disclosure.

In some embodiments, the engineer can also manually mark the expected several time-dependent normal characteristic data in the machine diagnosis system 100 (for example, the engineer selects several time-dependent characteristic data CDs from several time-dependent characteristic data CDs according to his experience. normal characteristic data), instead of marking several time-dependent normal characteristic data after inputting the marking data to the machine diagnosis system 100 .

Next, step S502 is executed: generating a normal wafer identification model 113 according to several time-dependent normal feature data. Specifically, the processor 130 uses an algorithm (for example, regression) to calculate the marked time-dependent normal feature data, so as to generate the normal wafer identification model 113 . Specifically, the good wafer identification model 113 is calculated to predict the time series of time-dependent characteristics that theoretically good wafers should behave.

In some embodiments, the normal wafer identification model 113 may be like an AR, MA, ARMA or ARIMA model, which are examples only and are not intended to be limiting. In other words, any model that can be used to predict the time series of time-dependent characteristics that a theoretically normal wafer should behave is within the spirit and scope of the present disclosure.

Next, step S503 is executed: judging whether the tool is abnormal when processing the first wafer W1 according to the normal wafer identification model 113 and the first time-dependent feature data CD1 . Specifically, after the processor 130 generates the normal wafer identification model 113 by calculating several time-dependent normal characteristic data, it receives the corresponding first time-dependent characteristic data CD1 from the first wafer W1 sensed by the sensing device 120 , and then compare the first time-dependent feature data CD1 with the normal wafer identification model 113 . If the error between the first time-dependent feature data CD1 and the normal wafer identification model 113 is less than or equal to (or less than) a preset threshold, the processor 130 determines that the first wafer W1 is a normal wafer. If there is an error greater than (or greater than or equal to) a preset threshold between the first time-dependent feature data CD1 and the normal wafer identification model 113, the processor 130 judges that the first wafer W1 is an abnormal wafer, that is, it reaches abnormal mining. Wafer efficacy.

By executing steps S501 , S502 , and S503 , mining abnormal wafers based on time-dependent normal feature data can be completed.

Embodiment (four): Mining normal wafers based on time-dependent abnormal feature data.

In this embodiment, the same machine base as in Embodiment (3) is used, and the machine base has the same sensing device 120 as in Embodiment (3). The time-dependent characteristics of several wafers are sensed by the sensing device 120 , and a plurality of time-dependent characteristic data CDs are correspondingly generated. Several time-dependent feature data CDs are stored in the big data database 110 .

In this embodiment, engineers manually mark several time-dependent abnormal feature data corresponding to abnormal wafers in all time-dependent feature data CDs, and then learn through the principle of machine learning, so that the machine diagnosis system 100 Normal wafers can be unearthed.

The method 600 of how to mine normal wafers based on time-dependent anomaly feature data will be described in detail below. The method 600 includes steps S601, S602, and S603.

Please refer to Figure 6. Firstly, step S601 is executed: mark several time-dependent abnormal characteristic data from several time-dependent characteristic data CDs according to the marking data. Specifically, the processor 130 receives several time-dependent feature data CDs and stores them in the big data database 110, and then the engineer sets up the machine diagnosis system 100, and the setting is the marking data input by the engineer. For example, the flag data may be a value range (for example, inputting a voltage range) of the time-dependent characteristic data CD that should be judged as an abnormal wafer, and the value range is one of the above flag data. Then, the processor 130 marks a plurality of time-dependent abnormal characteristic data by using the marking data. It should be noted that this disclosure does not intend to limit the type of marking data, more specifically, as long as it is any number or function (formula) that can standardize the time-dependent characteristic data CD that should be judged as an abnormal wafer Or computer programming syntax is within the spirit and scope of this disclosure.

In some embodiments, the engineer can also manually mark the expected several time-dependent abnormal characteristic data in the machine diagnosis system 100 (for example, the engineer selects several time-dependent abnormal characteristic data from several time-dependent characteristic data CDs according to his experience. Abnormal characteristic data), instead of marking several time-dependent abnormal characteristic data after inputting the marking data to the machine diagnosis system 100 .

Next, step S602 is executed: generating an abnormal wafer identification model 114 according to several time-dependent abnormal feature data. Specifically, the processor 130 uses an algorithm (for example, regression) to perform calculations on the marked data of several time-dependent abnormal features, so as to generate the abnormal wafer identification model 114 . Specifically, the abnormal wafer identification model 114 is calculated to predict the time series of time-dependent characteristics that theoretically abnormal wafers should behave.

In some embodiments, the abnormal wafer identification model 114 may be a model such as AR, MA, ARMA or ARIMA, which are examples only and are not intended to be limiting. In other words, any model that can be used to predict the time series of time-dependent characteristics that theoretically abnormal wafers should exhibit is within the spirit and scope of the present disclosure.

Next, step S603 is executed: judging whether the tool is normal when processing the first wafer W1 according to the abnormal wafer identification model 114 and the first time-dependent characteristic data CD1 . Specifically, after the processor 130 generates the abnormal wafer identification model 114 by calculating several time-dependent abnormal characteristic data, it receives the corresponding first time-dependent characteristic data CD1 from the first wafer W1 sensed by the sensing device 120 , and then compare the first time-dependent feature data CD1 with the abnormal wafer identification model 114 . If there is an error between the first time-dependent feature data CD1 and the abnormal wafer identification model 114 that is less than or equal to (or less than) a preset threshold, the processor 130 determines that the first wafer W1 is an abnormal wafer. If there is an error greater than (or greater than or equal to) a preset threshold between the first time-dependent feature data CD1 and the abnormal wafer identification model 114, the processor 130 judges that the first wafer W1 is a normal wafer, that is, the mining is normal. Wafer efficacy.

By executing steps S601 , S602 , and S603 , mining normal wafers based on time-dependent abnormal feature data can be completed.

From the above detailed description of the specific implementation of this disclosure, it can be clearly seen that in the machine diagnosis system of this disclosure, through the establishment of a large data database, it is possible to comprehensively process the time-based information from semiconductor machines. Raw data. In the machine diagnosis system disclosed in this disclosure, the difference analysis and comparison between machines can be performed through the processor. In addition, the processor can group several time-dependent feature data through unsupervised machine learning, so as to achieve the purpose of identifying the wafers processed by the machine as normal or abnormal. Moreover, the processor can mark normal wafers and abnormal wafers in the system in advance through engineers, and mine abnormal wafers based on the normal wafer identification model corresponding to the normal time-dependent characteristic data of normal wafers, or based on the corresponding Abnormal wafer identification model based on abnormal time-dependent feature data to mine normal wafers.

The above content summarizes the features of several implementations, so that those skilled in the art can better understand the aspects of this case. Those skilled in the art should understand that without departing from the spirit and scope of the present application, the above content can be easily used as a basis for designing or modifying other changes, so as to achieve the same purpose and/or realize the embodiments described herein. same advantage. The above contents should be understood as examples of the present disclosure, and the scope of protection thereof should be subject to the scope of the patent application.

100:Machine diagnosis system

110:Big data database

112:Wafer identification model

113:Normal wafer identification model

114: Abnormal Wafer Identification Model

120: Sensing device

130: Processor

300,400,500,600: method

S301, S302, S303, S304, S305, S306, S401, S402, S403, S404, S405, S501, S502, S503, S601, S602, S603: steps

CD: time-dependent characteristic data

CD1: The first time-dependent feature data

CD2: The second time-dependent feature data

CR: comparison result

MC1: The first machine

MC2: the second machine

V: Visual chart information

W1: first wafer

W2: second wafer

In order to make the above and other purposes, features, advantages and embodiments of the present disclosure more comprehensible, the accompanying drawings are described as follows: FIG. 1 is a functional block diagram illustrating a machine diagnosis system according to an embodiment of the present disclosure. FIG. 2 is another functional block diagram illustrating a machine diagnosis system according to an embodiment of the present disclosure. FIG. 3 is a flow chart illustrating a method for implementing machine difference analysis and comparison according to an embodiment of the present disclosure. FIG. 4 is a flowchart illustrating a method for grouping time-dependent feature data according to an embodiment of the present disclosure. FIG. 5 is a flow chart illustrating a method for mining abnormal wafers based on time-dependent normal feature data according to an embodiment of the present disclosure. FIG. 6 is a flow chart illustrating a method for mining normal wafers based on time-dependent anomaly feature data according to an embodiment of the present disclosure.

Domestic deposit information (please note in order of depositor, date, and number) none Overseas storage information (please note in order of storage country, institution, date, and number) none

100:Machine diagnosis system

110:Big data database

112:Wafer identification model

113:Normal wafer identification model

114: Abnormal Wafer Identification Model

120: Sensing device

130: Processor

CD: time-dependent characteristic data

CD1: The first time-dependent feature data

CD2: The second time-dependent feature data

Claims (5)

A tool diagnostic system for diagnosing a tool. The tool diagnostic system includes: a large data database storing at least one wafer identification model and a plurality of time-dependent characteristic data related to processing a plurality of wafers; The sensing device is configured to sense the time-dependent characteristics when the machine processes a first wafer, and correspondingly generates a first time-dependent characteristic data; and a processor is configured to: according to the first time-dependent characteristics updating the at least one wafer identification model; dividing the time-dependent characteristic data into a plurality of groups; calculating average values of the groups; and generating the at least one wafer identification model according to the average values. The tool diagnosis system as described in Claim 1, wherein the sensing device is further configured to sense a time-dependent feature when another machine processes a second wafer, and correspondingly generate a second time-dependent feature data, And the processor is further configured to: compare the first time-dependent feature data and the second time-dependent feature data according to the at least one wafer identification model, and generate a comparison result correspondingly; and according to the comparison As a result, a visual chart information is generated. The machine diagnostic system as described in claim 1, wherein the The processor is further configured to: find the average value closest to the first time-dependent feature data, and classify the first time-dependent feature data into the group corresponding to the average value; recalculate the group's an average value; and updating the at least one wafer identification model according to the recalculated average value. The machine diagnosis system as described in claim 1, wherein the at least one wafer identification model includes a normal wafer identification model, and the big data database stores a plurality of time-dependent characteristic data about processing a plurality of wafers, and The processor is further configured to: mark a plurality of time-dependent normal feature data from the time-dependent feature data according to a labeling data; generate the normal wafer identification model according to the time-dependent normal feature data; and generate the normal wafer identification model according to the normal wafer The circle identification model and the first time-dependent feature data determine whether the tool is abnormal when processing the first wafer. The machine diagnosis system as described in claim 1, wherein the at least one wafer identification model includes an abnormal wafer identification model, and the big data database stores a plurality of time-dependent characteristic data about processing a plurality of wafers, and The processor is further configured to: mark a plurality of time-dependent abnormal characteristic data from the time-dependent characteristic data according to a marking data; generate the abnormal wafer identification model according to the time-dependent abnormal characteristic data; and judging whether the tool is normal when processing the first wafer according to the abnormal wafer identification model and the first time-dependent feature data.

Images