KR102362159B1 - Method for Fault Detection and Fault Diagnosis in Semiconductor Manufacturing Process - Google Patents

Abstract

In the present embodiments, residual data is generated by removing data restored through a restoration model from the collected time series data, denoise residual data is generated by removing sensor noise from the residual data, and a failure is detected by analyzing the denoise residual data. By classifying, there is provided a failure detection and classification device capable of visualizing the timing of occurrence of an abnormal pattern that causes defects in the product manufacturing process and the shape of the abnormal pattern.

Description

Method for Fault Detection and Fault Diagnosis in Semiconductor Manufacturing Process

The technical field to which the present invention pertains relates to a failure detection and classification method.

The content described in this section merely provides background information for the present embodiment and does not constitute the prior art.

In the semiconductor process, Fault Detection and Classification (FDC) uses machine learning or data mining techniques to establish a model for judging process results based on a sensor data stream that reflects the process status of each wafer.

Establishing an accurate failure detection and classification model can prevent failures occurring during the wafer manufacturing process from leaking to subsequent processes by early detection, and can contribute to yield improvement through continuous monitoring of the process. .

The existing failure detection and classification model extracts a small number of features from the trace data and uses them for classification, so pattern information is lost, or the characteristics of the semiconductor process are not considered in the feature extraction process, so the characteristics of normal trace data and abnormal trace data are similar. There are errors that are formed.

In an ideal semiconductor processing facility, the trace patterns are almost identical for each wafer. As equipment deteriorates and chamber contamination progresses, traces may vary between wafers, so that a normal wafer may be recognized as an abnormal wafer.

Korean Patent Publication No. 10-0679721 (Jan. 31, 2007) Korean Patent Publication No. 10-1744194 (2017.05.31.)

Embodiments of the present invention process trace data in consideration of fluctuations between wafers due to process drift and shift and sensor noise generated in the sensor measurement process, thereby generating abnormal patterns that cause defects in semiconductor processes Its main purpose is to visualize the morphology of time points and abnormal patterns.

Other objects not specified in the present invention may be additionally considered within the scope that can be easily inferred from the following detailed description and effects thereof.

According to one aspect of this embodiment, in the failure detection and classification method by the failure detection and classification device, collecting time series data sensing products for each process parameter in the manufacturing process, the collected time series data through a restoration model Failure detection comprising the steps of generating residual data by removing the restored data, generating denoise residual data by removing sensor noise from the residual data, and classifying a failure by analyzing the denoise residual data and classification methods.

According to another aspect of this embodiment, in the failure detection and classification device including one or more processors and a memory for storing one or more programs executed by the one or more processors, the processor senses a product for each process parameter in a manufacturing process Collects one time series data, the processor removes data restored through a restoration model from the collected time series data to generate residual data, and the processor removes sensor noise from the residual data to generate denoise residual data and the processor analyzes the denoise residual data to provide a failure detection and classification device, characterized in that the failure is classified.

As described above, according to the embodiments of the present invention, residual data is generated by removing data restored through a restoration model from the collected time series data, and denoise residual data is generated by removing sensor noise from the residual data, , by analyzing the denoise residual data to classify failures, there is an effect of visualizing the time of occurrence of an abnormal pattern that causes defects in the product manufacturing process and the shape of the abnormal pattern.

Even if it is an effect not explicitly mentioned herein, the effects described in the following specification expected by the technical features of the present invention and their potential effects are treated as if they were described in the specification of the present invention.

1 is a block diagram illustrating an apparatus for detecting and classifying a failure according to an embodiment of the present invention.
2 is a diagram illustrating a restoration model of a failure detection and classification apparatus according to an embodiment of the present invention.
3 and 4 are diagrams illustrating residual data generated by an apparatus for detecting and classifying a failure according to an embodiment of the present invention.
5 is a diagram illustrating an operation of setting a threshold value in residual data by an apparatus for detecting and classifying a failure according to an embodiment of the present invention.
6 is a flowchart illustrating a failure detection and classification method according to another embodiment of the present invention.
7 and 8 show simulation results performed according to embodiments of the present invention.

Hereinafter, in the description of the present invention, if it is determined that the subject matter of the present invention may be unnecessarily obscure as it is obvious to those skilled in the art with respect to related known functions, the detailed description thereof will be omitted, and some embodiments of the present invention will be described. It will be described in detail with reference to exemplary drawings.

For wafer processing, a sensor that monitors process parameters input into the chamber in real time is installed, and when wafer processing is completed, trace data is generated for each process parameter. The fault detection and classification device analyzes trace data to detect a fault.

While the existing method does not consider a situation in which inter-wafer fluctuations and sensor noise are mixed in trace data, the failure detection and classification apparatus according to the present embodiment can accurately diagnose a failure in consideration of inter-wafer fluctuations.

1 is a block diagram illustrating an apparatus for detecting and classifying a failure according to an embodiment of the present invention.

The fault detection and classification device 110 includes at least one processor 120 , a computer readable storage medium 130 , and a communication bus 170 .

The processor 120 may control to operate as the failure detection and classification device 110 . For example, the processor 120 may execute one or more programs stored in the computer-readable storage medium 130 . The one or more programs may include one or more computer-executable instructions, which, when executed by the processor 120 , cause the fault detection and classification apparatus 110 to perform operations in accordance with the exemplary embodiment. can be configured.

Computer-readable storage medium 130 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. The program 140 stored in the computer-readable storage medium 130 includes a set of instructions executable by the processor 120 . In one embodiment, computer-readable storage medium 130 includes memory (volatile memory, such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, flash It may be memory devices, other types of storage media accessed by the failure detection and classification apparatus 110 and capable of storing desired information, or a suitable combination thereof.

Communication bus 170 interconnects various other components of fault detection and classification device 110 , including processor 120 and computer readable storage medium 140 .

The fault detection and classification device 110 may also include one or more input/output interfaces 150 and one or more communication interfaces 160 that provide interfaces for one or more input/output devices (not shown). The input/output interface 150 and the communication interface 160 are connected to the communication bus 170 . The input/output device (not shown) may be connected to other components of the failure detection and classification device 110 through the input/output interface 150 .

The failure detection and classification device 110 processes trace data in consideration of process drift and wafer-to-wafer variation due to shift and sensor noise generated in the sensor measurement process. The failure detection and classification device 110 generates residual data by removing data restored through a restoration model from time series data, removes sensor noise from the residual data to generate denoise residual data, and analyzes denoise residual data By classifying failures by doing this, it is possible to visualize the time of occurrence of an abnormal pattern that causes defects in the product manufacturing process and the shape of the abnormal pattern.

The processor 120 collects time series data of sensing products for each process parameter in the manufacturing process. The processor 120 generates residual data by removing data restored through a restoration model from the collected time series data. The processor 120 removes sensor noise from the residual data to generate denoise residual data. The processor 120 classifies the failure by analyzing the denoise residual data.

2 is a diagram illustrating a restoration model of a failure detection and classification apparatus according to an embodiment of the present invention.

The reconstruction model is an artificial neural network with multiple layers connected, adding noise to the input layer, setting the number of neurons in the hidden layer to be smaller than the input layer to reduce the dimension, and the information in the hidden layer is When mapped to the output layer, it is restored to the same dimension as the input layer.

The reconstruction model is a denoising auto-encoder (DAE) composed of an encoder and a decoder. The input layer with added noise is mapped to the hidden layer, and the network is trained to reconstruct the original input. After being mapped to the hidden layer, the input is restored while being mapped to the output layer. Autoencoder is a Feed Forward Neural Network consisting of one or more hidden layers for nonlinear feature extraction between the input layer and the output layer. learn

에 의해 부분적으로 노이즈가 첨가된

로 변형된다. 도 2를 참조하면 x₃과 x₅에 노이즈가 첨가되었다.The input data X is a stochastic mapping

Partially noise added by

is transformed into Referring to FIG. 2 , noise is added to x ₃ and x ₅ .

로 맵핑된다. W는 입력 레이어와 은닉 레이어 간의 가중치 매트릭스이고, b는 대응하는 바이어스 벡터이고, h는 설정된 활성화 함수이다.A hidden representation of an input with added noise

is mapped to W is the weight matrix between the input layer and the hidden layer, b is the corresponding bias vector, and h is the set activation function.

로 맵핑된다. W'는 입력과 은닉 간의 가중치 매트릭스이고, b'는 대응하는 바이어스 벡터이다.The hidden expression Y is the restored data

is mapped to W' is the weight matrix between the input and the hidden, and b' is the corresponding bias vector.

에 대한 손실함수

를 계산하고, 전체 훈련 데이터 세트의 평균 손실이 최소가 되도록 확률적 경사 하강(Stochastic gradient descent, SGD)을 통해 학습 파라미터 세트

를 학습하여 최적의 학습 파라미터 세트를 산출한다. 학습 파라미터 세트를 산출하는 공식은 수학식 1과 같이 표현된다.input data

loss function for

, and set the training parameters through stochastic gradient descent (SGD) so that the average loss of the entire training data set is minimal.

to calculate an optimal set of learning parameters. The formula for calculating the learning parameter set is expressed as Equation (1).

Hereinafter, an operation in which the failure detection and classification apparatus generates residual data will be described with reference to FIGS. 3 and 4 .

3 and 4 are diagrams illustrating residual data generated by an apparatus for detecting and classifying a failure according to an embodiment of the present invention.

In the step of generating residual data by the failure detection and classification device, a time series data set for a normal product including variation between products and sensor noise is learned through a restoration model. The restoration model restores time series data from which sensor noise has been removed.

By subtracting the time-series data restored by the failure detection and classification device from the time-series data for normal products, time-series data with sensor noise and abnormal patterns from which variations between products are removed are extracted.

라고 하고, 공정 파라미터

의 트레이스는

이고, T는 공정 주기이다.Multivariate pre-processing trace data of M process parameters measured by the sensor during processing of the lth wafer in the product manufacturing process

, and the process parameters

's trace is

, and T is the process cycle.

The time series data set for the process parameter m and the l-th product is expressed as in Equation (2). Assume the product is a wafer. The product-to-product variation may be a wafer-to-wafer (W2W) variation.

, 제품 간의 변동

, 센서 노이즈

, 이상 패턴

으로 구성된다. 웨이퍼 l이 정상이면 이상 패턴이 발생하지 않기 때문에 P_lm은 제로 벡터이다.ideal trace

, variation between products

, sensor noise

, the abnormal pattern

is composed of If wafer l is normal, P _lm is a zero vector because no abnormal pattern occurs.

The failure detection and classification apparatus according to the present embodiment generates residual data from which variation information between ideal time-series data and products is removed.

Using the DAE training dataset consisting of time-series data of normal products, the reconstruction model is trained individually for each process parameter m. The DAE training data set contains normal data with product-to-product variation and sensor noise.

으로 정의될 수 있다. 잔차 데이터는

로 정의될 수 있다.The restored time series data is

can be defined as The residual data is

can be defined as

Figure 3 (a) is the time-series data before processing, Figure 3 (b) is the restored time-series data, Figure 3 (c) is the residual data.

Referring to the restored time series data of FIG. 3B , sensor noise and abnormal patterns are removed through the restoration model, and variations between the ideal time series data and the product can be restored. The residual data in (c) of FIG. 3 shows sensor noise and anomaly patterns.

Figure 4 (a) is the data before processing for each parameter of the actual CVD (Chemical Vapor Deposition) process, Figure 4 (b) is the restored data. It can be seen that the abrupt change that occurred at the 150th sensing and the micro-perturbation that occurred at the 250th sensing are included in the residual data. That is, the restoration model restores only the variation between products and the ideal time series data, leaving only sensor noise and abnormal patterns in the residual data.

Hereinafter, an operation in which the failure detection and classification apparatus generates denoise residual data will be described with reference to FIG. 5 .

The generating of the denoise residual data by the failure detection and classification device includes setting a first noise threshold based on the size of the residual data and removing sensor noise using the first noise threshold.

The first noise threshold is expressed as a statistical representative value calculated for each time point from the residual data and an adjustment parameter that changes the statistical representative value through cross-validation. It is set according to the result of comparison with the level.

In the step of generating the denoise residual data by the failure detection and classification apparatus, a second noise threshold is set based on a duration of the residual data, and an abnormal pattern is detected using the second noise threshold.

The second noise threshold is set to a value corresponding to the second significance level after searching for a section in the residual data that successively exceeds the first noise threshold by one or more, calculating the length of the section crossing, and converting it into a histogram.

In the step of generating the denoise residual data by the failure detection and classification apparatus, the residual data having a value smaller than the second noise threshold is set to zero.

5 is a diagram illustrating an operation of setting a threshold value in residual data by the failure detection and classification apparatus according to an embodiment of the present invention.

The failure detection and classification apparatus first sets a first noise threshold for each time point in the residual data, and if the size of the residual data is smaller than the first noise threshold, it is regarded as sensor noise. Even if the size of residual data exceeds the first noise threshold by setting a second noise threshold, which is a reference for the minimum maintenance period of the abnormal pattern, if the size of the residual data does not continue for a period equal to or greater than the second noise threshold, it is regarded as noise. Otherwise, it is determined that an abnormal pattern has occurred.

과 표준편차

를 구하고 제1 노이즈 임계치

를 설정한다. k는 제1 노이즈 임계치의 크기를 조절하는 조정 파라미터이고, 교차검증을 이용하여 결정된다.Sample mean for each time point t from the residual data

and standard deviation

to find the first noise threshold

to set k is an adjustment parameter for adjusting the magnitude of the first noise threshold, and is determined using cross-validation.

Cross Validation divides the data set into h disjoint subsets, then uses the (h-1) subset for model training and the remaining 1 subset for model validation. After repeating the experiment h times, the first noise threshold is compared with the first significance level ε after calculating the ratio of the cumulative number of times. Cross-validation is performed by gradually increasing k until the ratio equals or exceeds ε. The minimum k in which the ratio of the number of times of residual data exceeding the first noise threshold exceeds ε is selected.

The first significance level ε corresponds to a type I error in the hypothesis test as the ratio of falsely determining that noise is not noise even though it is true. If the first significance level is 0.01, it means that noise as large as 1% can be mistaken for an anomaly pattern such as a peak.

The failure detection and classification apparatus according to the present embodiment removes noise statistically using the first significance level, but in order to solve the problem that there is a risk factor that erroneously determines that the noise is correct but not noise, the residual data is not noise. The noise is determined once again by calculating the run length of the determined residual data.

Due to the nature of the semiconductor process, where the occurrence rate of defective wafers is extremely small compared to normal wafers, the threshold training data set is mostly generated from normal wafer data. In the residual data of the threshold training set, except for the residual data that is likely to be some anomaly pattern, the duration of the residual data larger than the first noise threshold will be very small in probability, so it is necessary to remove all of these residual data.

The failure detection and classification apparatus finds all residual data sections that successively exceed one or more first noise thresholds for each residual data, and calculates a duration corresponding to the length of the sections. After constructing a histogram of the duration found in all residual data, a second noise threshold corresponding to the second significance level δ is found. Thereafter, all residual data of a section having a duration smaller than the second noise threshold in the residual data is regarded as noise and changed to zero.

A process of generating denoise residual data (DRT) of a new wafer l using a first noise control limit (NCL) and a second noise threshold (run length limit, RLL) will be described as follows.

를 생성한다.First, the time series data X _lm collected for each process parameter m (m=1, ... ,M) is input to the trained restoration model (DAE) to output the restored time series data, and residual data

to create

All sections in which the residual data exceed the first noise threshold NCL _m are found, and a duration set LS _lm and a viewpoint set IS _lm of these sections are generated.

RL _lmτ is the duration of the τ-th interval, TI _lmτ is the set of time points in the τ-th interval, and NA _lm is the total number of intervals in which RT _lm exceeds NCL _m . Time TI ₁₉₁ ={106,107, ... ,137} and RL ₁₉₁ =32 of the first section of RT ₁₉ of FIG. 5 .

For all points where RT _lm is less than or equal to NCL _m and for TI _lmτ of sections where duration RL _lmτ is less than the second noise threshold RLL _m among sections in which RT _lm exceeds NCL _m , all residual data are replaced with zero and the denoise residual is Generate data DRT _lm .

The failure detection and classification device classifies the failure by analyzing the denoise residual data.

The step of classifying the failure by the failure detection and classification device analyzing the denoise residual data is a distance-based classification model that assumes that the distance between the data of the abnormal product and the data of the normal product is greater than the distance between the data of the normal product. categorize failures. The classification model classifies the failures using the distance between the denoise residual data.

The failure detection and classification device inputs denoise residual data to a classification model, the classification model outputs a score, and classifies the output score according to a third significance level. The third significance level is established using a time-series data set for a normal product including product-to-product variation and sensor noise.

Using a single-class k-Nearest Neighbor (kNN) algorithm, the DRT information is summarized as a failure detection and classification index (FDC-Index) to determine whether the manufactured wafer is defective. The single-class kNN algorithm is a distance-based classification method for anomaly detection, based on the assumption that the distance between the failed wafer data and the normal wafer data will be much larger than the distance between the normal wafer data. As the input of the kNN algorithm, data obtained by concatenating denoise residual data of M process parameters is used. If we define the combined denoise residual data DRT of wafer l as DRT _l =[DRT _l11 , ... ,DRT _l1T , ... ,DRT _lM1 , ... ,DRT _lMT ], then the failure detection and classification index of wafer l is expressed as in Equation 5.

k is the number of neighbors, and n(l,j) means the j-th nearest neighbor in wafer l. FDC-Index _l is the sum of the squares of Euclidean distances between wafer l and its k nearest neighbors among the normal wafer data in the DAE training data set. As the FDC-Index _l score of wafer l is larger, it is different from the denoise residual data (DRT) of normal data, so that the possibility that the wafer is defective increases.

The upper limit of FDC-Index for determining whether wafer l is defective is set with the DAE training data set. An FDC-Index value is calculated for each wafer in the DAE training data set, and an index threshold FDC-Index _α that satisfies the third significance level α is set based on the calculated result. When the time series data before processing of wafer l is collected, denoise residual data (DRT) is generated for each process parameter and DRT _l is input into the kNN algorithm to obtain the FDC-Index score. If FDC-Index _l ≤ FDC-Index _α , wafer l is classified as normal, otherwise it is classified as malfunctioning.

If wafer l is determined to be defective as a result of classification through FDC-Index, the cause of the failure is diagnosed using DRT information. The denoise residual data is composed of an anomaly pattern. The denoise residual data may include some non-removed noise. A section in which the denoise residual data is greater than 0 is highly likely to be a section in which an abnormal pattern occurs.

After checking the denoise residual data for each process parameter to find all sections in which the denoise residual data is greater than 0, the size and pattern of the denoise residual data are checked. As a result of the confirmation, it is determined that an abnormal pattern has occurred in sections where the size of the denoise residual data is large and long-lasting, and for the sections in which the denoise residual data is small and maintained only for a short period, time series data before processing and DAE training data of wafer l Time-series data before processing of the set are compared temporally to see if there is a difference in the pattern. If a difference is found in the pre-processing time series data pattern, it is determined that an abnormal pattern has occurred in the corresponding section.

The denoise residual data of the section in which the abnormal pattern is determined may inform the time of occurrence of the abnormality of the process parameter and the shape of the abnormal pattern.

6 is a flowchart illustrating a failure detection and classification method according to another embodiment of the present invention. The failure detection and classification method may be performed by a failure detection and classification device.

In step S210, the processor collects time series data of sensing products for each process parameter in the manufacturing process.

In step S220, the processor removes the data restored through the restoration model from the collected time series data to generate residual data. The reconstruction model is a network in which a plurality of layers are connected, and the dimensionality is reduced by adding noise to the input layer, setting the number of neurons in the hidden layer to be smaller than that of the input layer, and the reduced dimensionality of the hidden layer is combined with the input layer. The original input is restored by mapping to the output layer of the same dimension. The reconstruction model may be a denoising autoencoder that trains the network to reconstruct the original input.

In the step of generating residual data (S220), a time series data set for a normal product including variation between products and sensor noise is learned through the restoration model, and the restoration model restores time series data from which sensor noise is removed, and normal By subtracting the restored time-series data from the time-series data for products, time-series data with sensor noise and abnormal patterns from which variations between products are removed are extracted.

In step S230, the processor removes sensor noise from the residual data to generate denoise residual data.

In the generating of the denoise residual data ( S230 ), a first noise threshold is set based on the size of the residual data, and sensor noise is removed using the first noise threshold. The first noise threshold is expressed as a statistical representative value calculated for each time point from the residual data and an adjustment parameter that changes the statistical representative value through cross-validation. It is set according to the result of comparison with the level.

In the generating of the denoise residual data ( S230 ), a second noise threshold is set based on the duration of the residual data, and an abnormal pattern is detected using the second noise threshold. The second noise threshold is set to a value corresponding to the second significance level after searching for a section in the residual data that successively exceeds the first noise threshold by one or more, calculating the length of the section crossing, and converting it into a histogram.

In the step of generating the denoise residual data ( S230 ), residual data having a value smaller than the second noise threshold is set to zero.

In step S240, the processor classifies the failure by analyzing the denoise residual data.

In the step of classifying the failure by analyzing the denoise residual data (S240), the failure is classified through a distance-based classification model on the premise that the distance between the data of the abnormal product and the data of the normal product is greater than the distance between the data of the normal product. do. The classification model classifies the failures using the distance between the denoise residual data. The denoise residual data is input to a classification model, and the classification model outputs a score, and classifies the output score according to a third significance level. A third significance level is established using a time-series data set for normal products, including product-to-product variation and sensor noise.

In step S250, after analyzing the denoise residual data to classify the failure, the processor analyzes the denoise residual data to diagnose the cause of the failure.

The step of diagnosing the cause of the failure (S250) is to search for a section in which the denoise residual data is greater than 0 for each process parameter, and compare it with time-series data sensed based on the size and pattern of the denoise residual data in time to find an abnormal pattern is determined, and the occurrence time of the abnormal pattern is output.

In the step of diagnosing the cause of the failure ( S250 ), the denoise residual data for the failure is analyzed to output the process parameter causing the failure, the occurrence time of the abnormal pattern, and the type of the abnormal pattern.

7 and 8 show simulation results performed according to embodiments of the present invention. 7 is a diagnosis result for data of an etching process.

7(a) is denoise residual data for BCI ₃ flow among process parameters, and FIG. 7(b) is pre-processing time series data for BCI ₃ flow among process parameters.

Denoise residual data during the total wafer processing time of the failed wafer 12 is large. Through the denoise residual data, it can be diagnosed that a clear and persistent abnormal pattern occurred in the pre-processing time series data.

FIG. 7(c) is denoise residual data for pressure among process parameters, and FIG. 7(d) is pre-processing time series data for pressure among process parameters.

The denoise residual data of failed wafers 4, 7, 13, and 18 significantly increased and then decreased in the periods of 0 sec. to 6 sec. and 53 sec. to 55 sec. It can be diagnosed that high pressure fluctuations occurred in both periods.

7(e) is denoise residual data for RF power among process parameters, and FIG. 7(f) is pre-processing time series data for RF power among process parameters.

Diagnosis results showed that abnormal fluctuations occurred during the processing of faulty wafers 4, 13, and 17. It can be seen that the faulty wafers 4 and 13 had an anomalous fluctuation in the form of a peak that suddenly dropped and rose in the period of 49 seconds to 54 seconds.

Failed wafer 17 is estimated to have peaked in 22 s to 24 s and 62 s to 66 s periods. The two circles in (f) of FIG. 7 show pre-processing time-series data for a period in which abnormal fluctuations occurred during processing of the failed wafer 17. In the first period, the time-series data of the failed wafer 17 and the normal time-series data show a large difference. Since it is not visible, it is judged to be a false warning, and it can be understood that a downward peak occurred in the second period.

8 is a diagnosis result for data of a CVD process.

8A is denoise residual data of failed wafers 3 and 7 in which an abnormal pattern occurred. Denoise residual data is shown to exceed 0 for defective wafer 3 in the interval of 226 seconds to 264 seconds, and defective wafer 7 in the interval of 246 seconds to 264 seconds. As a result of comparing the pre-processing time series data of FIG. 8 (b) with the pre-processing time series data of the DAE training wafer, a small change occurred during the processing of the faulty wafer 3, and during the processing of the faulty wafer 7, in the section where the parameter value should be kept constant It can be seen that a sudden change has occurred. Considering that the parameter observation period of the sensor is 0.2 seconds, it can be understood that the abnormal pattern occurred 45.2 seconds and 49.2 seconds after the start of processing of faulty wafers 3 and 7, respectively.

8C is denoise residual data of the normal wafer 29 and the failed wafer 2 in which the abnormal pattern occurred. As a result of analyzing the pre-processing time series data of FIG. 8(d), the pre-processing time series data of the normal wafer 29 and the pre-processing time series data and pattern of the DAE training wafer in the section where the denoise residual data of the normal wafer 29 exceeds 0 There is no difference, and it can be seen that the faulty wafer 2 experienced severe vibration from 69.2 seconds to 79.2 seconds, which is a period of 346 seconds to 396 seconds. Even if the output of the denoise residual data is misclassified, the classification result can be corrected by additionally comparing the time series data before processing.

8E is denoise residual data of failed wafers 3, 4, 5, and 6 in which an abnormal pattern occurred. As a result of analyzing the time series data before processing in (f) of FIG. 8 , it can be confirmed that an abnormal pattern in the form of a peak occurred in the interval of 45.2 seconds to 52.8 seconds.

The failure detection and classification apparatus may be implemented in a logic circuit by hardware, firmware, software, or a combination thereof, or may be implemented using a general-purpose or special-purpose computer. The device may be implemented using a hardwired device, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or the like. In addition, the device may be implemented as a system on chip (SoC) including one or more processors and controllers.

The failure detection and classification apparatus may be mounted in the form of software, hardware, or a combination thereof on a computing device or server provided with hardware elements. A computing device or server includes all or part of a communication device such as a communication modem for performing communication with various devices or wired/wireless communication networks, a memory for storing data for executing a program, and a microprocessor for executing operations and commands by executing the program. It can mean a variety of devices, including

Although it is described that each process is sequentially executed in FIG. 6, this is merely an exemplary description, and those skilled in the art change the order described in FIG. 6 within the range that does not depart from the essential characteristics of the embodiment of the present invention Alternatively, various modifications and variations may be applied by executing one or more processes in parallel or adding other processes.

The operations according to the present embodiments may be implemented in the form of program instructions that can be performed through various computer means and recorded in a computer-readable medium. Computer-readable medium represents any medium that participates in providing instructions to a processor for execution. Computer-readable media may include program instructions, data files, data structures, or a combination thereof. For example, there may be a magnetic medium, an optical recording medium, a memory, and the like. A computer program may be distributed over a networked computer system so that computer readable code is stored and executed in a distributed manner. Functional programs, codes, and code segments for implementing the present embodiment may be easily inferred by programmers in the art to which this embodiment belongs.

The present embodiments are for explaining the technical idea of the present embodiment, and the scope of the technical idea of the present embodiment is not limited by these embodiments. The protection scope of this embodiment should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be interpreted as being included in the scope of the present embodiment.

Claims (16)

A method for detecting and classifying a failure by a failure detection and classification device, the method comprising:
collecting time series data of sensing products for each process parameter in the manufacturing process;
generating residual data by removing data restored through a restoration model from the collected time series data;
generating denoise residual data by setting a first noise threshold based on the size of the residual data and removing sensor noise using the first noise threshold; and
Analyze the denoise residual data to monitor and visualize the occurrence time of the abnormal pattern and the shape of the abnormal pattern in real time
A fault detection and classification method comprising a. According to claim 1,
The reconstruction model is a network in which a plurality of layers are connected, and the dimensionality is reduced by adding noise to the input layer, setting the number of neurons in the hidden layer to be smaller than the input layer, and the reduced dimensionality of the hidden layer is the A failure detection and classification method, characterized in that it is an autoencoder that is mapped to an output layer of the same dimension as the input layer, and trains the network to restore the original input. According to claim 1,
The step of generating the residual data is
A time series data set for a normal product including fluctuations between products and sensor noise due to drift is learned through the restoration model, and the restoration model restores time series data from which the sensor noise is removed, and the normal product A failure detection and classification method, characterized in that by subtracting the restored time series data from the time series data for . According to claim 1,
The step of generating the residual data is
A time series data set for a normal product including variations between products and sensor noise due to shift is learned through the restoration model, and the restoration model restores time series data from which the sensor noise is removed, and the normal product A failure detection and classification method, characterized in that by subtracting the restored time series data from the time series data for . According to claim 1,
The first noise threshold is expressed as a statistical representative value calculated for each time point from the residual data and an adjustment parameter for changing the statistical representative value through cross-validation, and the adjustment parameter is the number of accumulated times exceeding the first noise threshold. A failure detection and classification method, characterized in that the ratio is set according to a result of comparing it with the first significance level. According to claim 1,
The generating of the denoise residual data includes:
A second noise threshold is set based on the duration of the residual data, and an abnormal pattern is detected using the second noise threshold. 7. The method of claim 6,
The second noise threshold is a value corresponding to the second significance level after searching the residual data for a section continuously exceeding the first noise threshold by one or more, calculating the length of the crossing section, and converting it into a histogram. Failure detection and classification method, characterized in that set. 7. The method of claim 6,
The generating of the denoise residual data includes:
and setting the residual data having a value smaller than the second noise threshold to zero. According to claim 1,
The step of classifying the failure by analyzing the denoise residual data includes:
The failure is classified through a classification model based on the distance between the denoise residual data, the denoise residual data is input to the classification model, the classification model outputs a score, and the output score is calculated according to a third significance level classify,
The third significance level is a failure detection and classification method, characterized in that it is set using a time series data set for a normal product including fluctuations between products and sensor noise. According to claim 1,
The step of monitoring and visualizing the occurrence time of the abnormal pattern and the shape of the abnormal pattern in real time includes the steps of classifying a failure by analyzing the denoise residual data and diagnosing the cause of the failure by analyzing the denoise residual data comprising steps,
Searching for a section in which the denoise residual data is greater than 0 for each process parameter, and comparing it temporally with the sensed time series data based on the size and pattern of the denoise residual data to determine whether the abnormal pattern occurs, A failure detection and classification method, characterized in that outputting the occurrence time of the abnormal pattern. 11. The method of claim 10,
The step of monitoring and visualizing the occurrence time of the abnormal pattern and the shape of the abnormal pattern in real time,
A method for detecting and classifying a failure, characterized in that by analyzing the denoise residual data for the failure, a process parameter causing the failure, an occurrence time of the abnormal pattern, and a shape of the abnormal pattern are output. A failure detection and classification apparatus comprising one or more processors and a memory storing one or more programs executed by the one or more processors,
The processor collects time series data sensing products for each process parameter in the manufacturing process,
The processor generates residual data by removing data restored through a restoration model from the collected time series data,
The processor sets a first noise threshold based on the size of the residual data and removes sensor noise using the first noise threshold to generate denoise residual data,
The processor analyzes the denoise residual data to monitor and visualize the occurrence time of the abnormal pattern and the shape of the abnormal pattern in real time. 13. The method of claim 12,
The processor is
A time series data set for a normal product including fluctuations between products due to drift and shift and sensor noise is learned through the restoration model, and the restoration model restores time series data from which the sensor noise is removed. and subtracting the restored time series data from the time series data for the normal product, and extracting time series data having sensor noise and abnormal patterns from which variations between the products are removed. 13. The method of claim 12,
The processor is
The failure detection and classification apparatus of claim 1, wherein a second noise threshold is set based on the duration of the residual data, and an abnormal pattern is detected using the second noise threshold. 13. The method of claim 12,
The processor is
The failure is classified through a distance-based classification model between the denoise residual data, the denoise residual data is input to the classification model, the classification model outputs a score, and the output score is calculated according to a third significance level classify,
The third significance level is a failure detection and classification device, characterized in that it is set using a time series data set for a normal product including fluctuations between products and sensor noise. 13. The method of claim 12,
The processor is
Searching for a section in which the denoise residual data is greater than 0 for each process parameter, comparing with the sensed time series data based on the size and pattern of the denoise residual data over time, determines whether an abnormal pattern occurs, and A failure detection and classification device, characterized in that outputting the occurrence time of the abnormal pattern.

Images