KR20110091111A - System and method for dignosis of semiconduct manufacturing apparatus - Google Patents

Abstract

PURPOSE: A failure diagnosis system of a semiconductor apparatus and a method using the same are provided to enhance the quality of a process, processing yield, and equipment productivity by finding out the reason of malfunction of each part comprising a semiconductor manufacturing apparatus in real time using a computation intelligence system. CONSTITUTION: A sensor information gathering unit(200) collects sensor information about each part of a semiconductor manufacturing apparatus. A part signal-monitoring unit(300) collects prediction sensor information by forming monitoring model about the each part from the collected sensor information and outputs a failure trust value of the collected prediction sensor information. A failure judgment unit(400) judges manufacturing based on the failure trust value.

Description

Fault diagnosis system and method of semiconductor device {SYSTEM AND METHOD FOR DIGNOSIS OF SEMICONDUCT MANUFACTURING APPARATUS}

The present invention relates to a failure diagnosis system for a semiconductor device, and more particularly, to a failure diagnosis system and method for a semiconductor device for diagnosing a failure by monitoring sensor values of each component constituting a semiconductor manufacturing device.

In general, the semiconductor process includes an etching process, a thin film deposition process, a cleaning process, a photo process, and the like, and a series of unit processes are performed sequentially or in combination to manufacture a semiconductor device.

In general, a semiconductor process is performed in a vacuum chamber, and after the semiconductor substrate is placed in the chamber, plasma is generated on the seated substrate to form a thin film or etch the substrate.

The state of the plasma when an abnormality occurs in a component constituting the semiconductor device, for example, a mass flow controller (MFC), a radio frequency source (RF source), a bias power, etc. while the substrate is being processed. Is changed, resulting in a problem that the deposition or etching characteristics are different.

To solve this, a separate in-situ diagnostic tool is used to monitor the plasma, but it does not provide a cause of failure. Another method is to diagnose the cause of failure by modeling the relationship between the electrical signal of the component and the in-situ interaction with neural networks. In this case, the installation of expensive in-situ system is required, and there is a difficulty in generating a specific failure pattern in advance, and thus, the failure cause is not provided in case of unknown failure. On the other hand, when collecting and processing the parts sensor information collected from a number of plasma equipment in a single processing place, it takes a considerable time to collect and transmit a large amount of sensor information, due to the large amount of sensor information to troubleshoot Algorithm for this is complicated, which leads to a problem of lowering the accuracy of the diagnosis of the failure of the plasma.

In order to solve the above problems, an object of the present invention is to provide a failure diagnosis system and method for a semiconductor device capable of immediately responding to the failure by increasing the accuracy of failure diagnosis of the semiconductor device.

Another object of the present invention is to provide a failure diagnosis system and method for a semiconductor device for improving production efficiency by detecting and immediately diagnosing the cause of failure of a semiconductor device in real time.

In order to achieve the above object, the failure diagnosis system of a semiconductor device of the present invention is a sensor information collection unit for collecting the sensor information for each component of the semiconductor device in real time, and a monitoring model for each component from the collected sensor information And a component signal monitoring unit configured to collect the predicted sensor information, and output a failure belief value of the collected prediction sensor information, and a failure determination unit determining whether the failure is based on the failure belief value obtained from the component signal monitoring unit.

The monitoring model may be an autocorrelation model that outputs predicted sensor information of a magnetic component using sensor information of a magnetic component as input data, and the autocorrelation model may be formed to be equal to the number of components.

The monitoring model may be a cross-correlation model that outputs predictive sensor information of a magnetic component using signals other than the sensor information of the magnetic component as input data.

The monitoring model may be a self-cross-correlation model that outputs predicted sensor information of the magnetic component using sensor information of the magnetic component and sensor information other than the magnetic component as input data.

The surveillance model may be a statistical model or a statistical time series model for each component.

The component signal monitoring unit may accumulate component sensor information by a CUSUM control chart and apply the accumulated component sensor information to a failure belief function.

In addition, in order to achieve the above object, the fault diagnosis method of the semiconductor device of the present invention comprises the steps of collecting the sensor information for each component of the semiconductor device in real time, forming a monitoring model to collect the predicted sensor information and And outputting a failure belief value from the predicted sensor information, and determining whether the component is defective by comparing the output failure confidence value with a threshold value.

The monitoring model outputs predicted sensor information of the magnetic component using the sensor information of the magnetic component as input data, or outputs predicted sensor information of the magnetic component or signals of the magnetic component using signals other than the sensor information of the magnetic component as input data. And predictive sensor information of the magnetic component can be output using signals other than the magnetic component as input data.

The fault belief value may be obtained by applying the predicted sensor information to a CUSUM control chart and accumulating it, and then applying it to a fault belief function.

When the collection of sensor information for each component of the semiconductor device is completed, the RMSE may be obtained using the monitoring model of the sensor information of all components at an arbitrary time point, and failure may be determined from the obtained RMSE.

The present invention has the effect of improving equipment productivity, process yield, and process quality by diagnosing the failure of each component constituting the semiconductor device in real time and finding the cause of the failure by using a computer intelligent system.

In addition, the present invention can additionally perform a fault diagnosis after finishing the process, thereby increasing the reliability of fault diagnosis of the semiconductor device.

1 is a block diagram showing a failure diagnosis system of a semiconductor device according to the present invention.
2 is a structural diagram showing a structure of a neural network according to the present invention.
3 and 4 show a surveillance model according to the invention.
5 is a view showing the structure of the failure believer according to the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It will be apparent to those skilled in the art that the present invention may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, It is provided to let you know. Like reference numerals refer to like elements throughout.

1 is a configuration diagram showing a failure diagnosis system of a semiconductor device according to the present invention, Figure 2 is a structural diagram showing a structure of a neural network according to the present invention, Figures 3 and 4 is a view showing a monitoring model according to the present invention 5 is a view showing the structure of the failure believer according to the present invention.

Referring to FIG. 1, a failure diagnosis system of a semiconductor device according to the present invention includes a sensor information collecting unit 200 which collects sensor information of each component of a semiconductor device in real time, and the respective sensor components from the collected sensor information. By forming a monitoring model to collect the predicted sensor information, and compares the failure belief value obtained from the component signal monitoring unit 300 and the component signal monitoring unit 300 to output the failure confidence value of the collected predicted sensor information by comparing with a critical point It includes a failure determination unit 400 for determining whether or not a failure.

The sensor information collecting unit 200 collects sensor information of a component driving the semiconductor device, for example, the plasma processing apparatus 100. The plasma processing apparatus 100 may generate a plasma by applying upper and lower RF power and pressure to a reaction gas injected in a chamber, and the sensor information collecting unit 200 generates a plasma component while the substrate is processed. Information can be collected in real time.

When the sensor information of each component constituting the plasma processing apparatus 100 is collected in the sensor information collection unit 200 as described above, the component signal monitoring unit 300 forms a monitoring model to monitor the sensor information collected in real time. can do. As such a monitoring model, a neural network or fuzzy logic model, which is a computational intelligence model, may be used.

Of course, in addition to the computational intelligence model, as a monitoring model, a statistical model or an auto-regressive model or an auto-moving average model, which is a statistical time series method, may be applied. It can consist of standard deviations.

As shown in FIG. 2, the neural network 500 used as a computational intelligence model divides part sensor information into training data and test data, and learns part state patterns using the training data. It is a system that predicts the failure state of parts by using the learned model.

The neural network 500 is composed of an input layer 520, a hidden layer 540, and an output layer 560. The neural network 500 uses the weights between neurons to adjust the relationship between the input and output patterns. Use this to predict output or find the most similar group of patterns for a later input pattern. The number of neurons in the input layer 520 matches the number of all elements constituting the feature vector, and the number of neurons in the output layer 560 is one. Of course, the number of neurons of the output layer 560 is not limited to this, and may be composed of two or more.

When using the neural network as described above, a generalized delta rule may be used as a learning algorithm used for learning, and an optimized learning algorithm may be used by modifying the generalized delta law.

The sensor information of the j-th component, that is, information at a future time t + k of Sj may be predicted using the current time t and the past information t-m of the same component sensor information. The current and past information constitutes the input of the learning pattern, and the future time information constitutes the output of the learning pattern. When the number of component sensor information to be modeled is n, n neural networks are required.

In this way, the neural network structure for outputting the predictive sensor information of the magnetic component signal using the magnetic component sensor information as the input data is called an auto-correlation model.

As a result, the component signal monitoring unit 300 may collect prediction information based on sensor information collected in real time from the sensor information collecting unit 200.

In the above, the state of the part is monitored using an autocorrelation model, but the state of the part can be monitored using a neural network having a different structure. A cross-correlated model, which is a neural network structure for outputting predictive sensor information of magnetic components by calculating signals and inputting signals other than magnetic component sensor information in the input layer of the neural network, may be used.

In addition to the autocorrelation model and the cross-correlation model, the magnetic layer sensor, which is a neural network structure that outputs the predictive sensor information of the magnetic component by calculating the magnetic component sensor information and the sensor information other than the magnetic component as input data in the input layer of the neural network. An auto-cross correlated model may also be used.

In the above, although any one of each model is formed as a monitoring model, the present invention is not limited thereto and may have the following structure.

As shown in FIG. 3, the autocorrelation model and both the cross-correlation model and the self-cross-correlation model can be used as the monitoring model 600 to monitor the state of the parts, as shown in FIG. 4. Obviously, a monitoring model of a combination of a model, a cross-correlation model, and a self-cross-correlation model may be used.

In addition, in order to improve the measurement efficiency of neural networks, we can optimize the learning tolerance, the number of neurons in the hidden layer, the size of the initial weight, the slope of the neuron activation function, and so on. Can be.

As described above, when the prediction signal is obtained, the component signal monitoring unit 300 may output a failure belief value from the obtained prediction signal. Here, the failure confidence value may be obtained by accumulating the parts sensor information by applying the predictive sensor information to the CUSUM control chart, and outputting the accumulated parts sensor information as a belief function.

와 x는 정상 플라즈마에서 수집된 부품 센서정보의 통계적 평균치와 매 샘플링 순간에 수집되는 센서정보를 의미한다.The CUSUM control chart may be defined as in Equation 1. Where h and l mean when the variation of the variable increases or decreases,

And x denote the statistical mean value of the component sensor information collected in the normal plasma and the sensor information collected at each sampling instant.

As described above, when the sensor information is processed by the CUSUM control chart, the processed sensor information may be applied to the failure belief function shown in Equation 2.

와

는 고장믿음의 값을 조정하기 위한 제어변수들이며, 그 값은 부품 센서정보의 변화, 예컨대 통계적 표준편차를 참조하여 임의로 결정될 수 있다.here,

Wow

Are control variables for adjusting the value of the failure belief, and the value can be arbitrarily determined by referring to the change in the part sensor information, for example, the statistical standard deviation.

In the above, the predicted sensor information is applied to the CUSUM control chart and then applied to the failure belief function (first failure generator), but the output value can be directly obtained from the failure belief function without going through the CUSUM control chart (second failure). Faith generator). When the fault belief value is generated in the second fault belief generator, S _{h / l} in Equation 2 may be replaced with an absolute value of the difference between the predicted value and the mean value from the monitoring model.

The first failure faith generator is effective for detecting a progressive failure, and the second failure faith generator is effective for detecting a sudden failure. Therefore, as illustrated in FIG. 5, the failure faith generator 700 may be configured by combining the first failure faith generator and the second failure faith generator 720 and 740 to further increase failure diagnosis efficiency. .

As described above, the failure determination unit 400 may perform the failure determination according to whether the output value calculated by the failure belief function is greater than a predetermined reference value. For example, when the output value calculated by the failure belief function has a value between 0 and 1, any reference value may be set to 0.5, and when the output value is 0.5 or more, it may be considered that a failure has occurred. Of course, any reference value may be changed to a value other than 0.5.

As described above, the failure diagnosis system of the semiconductor device according to the present invention may diagnose failure of each component in real time, and may increase the process yield by checking the cause of the failure.

On the other hand, when the process is completed and the collection of the entire sensor information for each part is completed, it is applied to the monitoring model of the parts signal part as an input and outputs the predicted value, and the RMSE (Root Mean Square Error) The calculation can be used to determine the failure. Diagnosis using the RMSE may be applied to monitoring the plasma by applying the entire sensor pattern collected during the process.

In this case, we construct a monitoring model for all parts sensor information collected at any time, input the collected sensor information into the monitoring model, output the prediction value, calculate the RMSE between the output prediction value and the actual value, and then calculate this RMSE. Calculate the fault belief value by applying it to the CUSUM control chart and the fault faith function of Equation 2 and diagnose the fault based on the calculated fault faith value.

The range of steady state of RMSE can be determined by Equation 3, where RMSE _m and RMSE _s mean the mean and standard deviation of multiple model RMSEs.

In addition, the range of the steady state of the RMSE can be adjusted by adding a scale variable by Equation 4, and RMSE _min and RMSE _max mean the minimum and maximum RMSE determined among the plurality of RMSEs. Here, the scale variable can be adjusted through experiments. For example, 0.01 (1%) to 0.1 (10%) of the RMSE value may be applied.

As described above, the failure diagnosis system of the semiconductor device according to the present invention can further add a failure diagnosis after the process, and when using a combination of real-time failure information and failure information after the process has the effect of increasing the reliability of the failure diagnosis. have.

Although described above with reference to the drawings and embodiments, those skilled in the art that the present invention can be variously modified and changed within the scope without departing from the spirit of the invention described in the claims below I can understand.

100: semiconductor device 200: sensor information collection unit
300: component signal monitoring unit 400: failure determination unit
500: computational intelligence model 600: surveillance model

Claims (10)

A sensor information collecting unit which collects sensor information about each component of the semiconductor device in real time;
A component signal monitor configured to form a monitoring model for each component from the collected sensor information, collect predictive sensor information, and output a failure belief value of the collected predictive sensor information; And
A failure determination unit determining whether a failure is based on a failure belief value obtained from the component signal monitoring unit;
Failure diagnosis system of a semiconductor device comprising a. The method according to claim 1,
The monitoring model is an autocorrelation model for outputting predictive sensor information of a magnetic component using sensor information of a magnetic component as input data, and the autocorrelation model is formed to be equal to the number of components. The method according to claim 1,
And the monitoring model is a cross-correlation model for outputting predictive sensor information of a magnetic component using signals other than the sensor information of the magnetic component as input data. The method according to claim 1,
And the monitoring model is a self-cross-correlation model for outputting predictive sensor information of the magnetic component using sensor information of the magnetic component and sensor information other than the magnetic component as input data. The method according to claim 1,
Wherein the monitoring model is a statistical model or a statistical time series model for each component. The method according to claim 1,
The component signal monitoring unit accumulates component sensor information by a CUSUM control chart, and applies the accumulated component sensor information to a failure belief function. Collecting sensor information about each component of the semiconductor device in real time;
Forming a monitoring model to collect predicted sensor information;
Outputting a fault belief value from the predicted sensor information; And
Determining whether a component is broken by comparing the output failure faith value with a threshold value;
Failure diagnosis method of a semiconductor device comprising a. The method according to claim 7,
The monitoring model outputs predicted sensor information of the magnetic component using the sensor information of the magnetic component as input data, or outputs predicted sensor information of the magnetic component or signals of the magnetic component using signals other than the sensor information of the magnetic component as input data. And a method for diagnosing a failure of a semiconductor device that outputs predictive sensor information of a magnetic component using signals other than the magnetic component as input data. The method according to claim 7,
The fault diagnosis method is a failure diagnosis method of a semiconductor device that is obtained by applying the accumulated sensor information to the CUSUM control chart and accumulates the failure confidence function. The method according to claim 7,
A method for diagnosing a failure of a semiconductor device, when the collection of sensor information on each part of the semiconductor device is completed, the RMSE is obtained by using a monitoring model on the sensor information of all parts at an arbitrary time point, and the failure is determined from the obtained RMSE.

Images