CN118378024A - Abnormality detection method and device based on wafer manufacturing and electronic device - Google Patents

Abstract

The present application relates to an abnormality detection method, device and electronic device based on wafer manufacturing, and the abnormality detection method includes: preprocessing the training data, and then training the target deep learning model based on the preprocessed training data and bidirectional restricted adversarial learning. Acquire and preprocess the time series data to be detected, and determine the reconstruction error corresponding to the time series data to be detected according to the pre-trained target deep learning model; determine the abnormal sequence according to the reconstruction error, and evaluate the abnormal score in the abnormal sequence according to the sliding window of the preset sequence length, and determine the abnormal data of the target wafer equipment, which is conducive to capturing and identifying various types of abnormal situations, thereby improving the recognition efficiency of abnormal data in the time series data to be detected.

Description

Anomaly detection method, device and electronic device based on wafer manufacturing

Technical Field

The present application relates to the field of semiconductor chips, and in particular to an abnormality detection method, device and electronic device based on wafer manufacturing.

Background technique

In the existing semiconductor chip manufacturing process, wafer manufacturing is a complex, expensive and lengthy process involving hundreds of process steps, which requires simultaneous monitoring of corresponding process parameter changes. With the increase in the precision and complexity of semiconductor equipment, higher requirements are placed on the quality of fault detection and classification in the wafer manufacturing process, that is, the need to promptly discover and identify potential anomalies and faults.

At present, the relevant technologies for fault detection and classification commonly used in the wafer manufacturing process are to extract the characteristic value information in the raw data of the sensor by creating a model and perform statistics on the characteristic value information. Due to the complexity and precision of semiconductor chips, traditional rule-based or statistical detection methods often cannot meet the requirements of accuracy and efficiency. In the wafer manufacturing process, the raw data of different sensors are diverse, and it is difficult for chip engineers to mark and judge whether a single time point of different sensors is abnormal, resulting in low accuracy of abnormality detection in the wafer manufacturing process.

With regard to the problem in related technologies that the accuracy of abnormality detection in the wafer manufacturing process needs to be improved, no effective solution has been proposed so far.

Summary of the invention

In this embodiment, a method, device and electronic device for abnormality detection based on wafer manufacturing are provided to solve the problem in the related art that the accuracy of abnormality detection in the wafer manufacturing process needs to be improved.

In a first aspect, a method for detecting anomalies based on wafer manufacturing is provided in this embodiment, and the method includes:

Acquire time series data to be detected; the time series data to be detected includes sensor data of wafer equipment;

Determine the reconstruction error corresponding to the time series data to be detected according to a pre-trained target deep learning model; the target deep learning model is trained according to a preset adversarial learning method;

An abnormal sequence is determined according to the reconstruction error, and an abnormal score in the abnormal sequence is evaluated according to a sliding window of a preset sequence length to determine abnormal data of a target wafer device.

In some of the embodiments, before acquiring the time series data to be detected, the process includes:

Acquire multiple training time series data in a wafer manufacturing process, and process the multiple training time series data to obtain multiple unified training time series data;

Determine dense sequence data in the plurality of unified training time series data; the dense sequence data includes sequence data conforming to a normal distribution;

According to the dense sequence data and the preset adversarial learning method, the preset deep learning model is trained to determine a target deep learning model; the preset deep learning model is based on the waveform of the wafer sensor.

In some of the embodiments, before determining the reconstruction error corresponding to the time series data to be detected according to the pre-trained target deep learning model, the method includes:

Acquire initial timing data from the sensor on the wafer according to the preset sampling frequency.

The initial time series data is processed according to the dense sequence data to obtain the time series data to be detected.

In some embodiments, determining the reconstruction error corresponding to the time series data to be detected according to a pre-trained target deep learning model includes:

Mapping the time series data to be detected to a preset latent space to obtain a coding vector;

When the encoding vector is within a preset threshold range, processing the encoding vector and generating a reconstructed sequence;

The difference between the data of the reconstructed sequence and the time series data to be detected is determined as a reconstruction error.

In some embodiments, determining the abnormal sequence according to the reconstruction error includes:

Normalizing the reconstruction errors corresponding to multiple time points in the time series data to be detected to obtain error scores for the multiple time points;

Determine an array of abnormal scores according to preset abnormal values; the preset abnormal scores are obtained by processing the error scores according to a preset statistical method;

Determine that the sequence corresponding to the abnormal score array is an abnormal sequence.

In some embodiments, evaluating the abnormal scores in the abnormal sequence according to the sliding window of the preset sequence length to determine the abnormal data of the target wafer equipment includes:

Dividing the abnormal sequence according to a sliding window of a preset sequence length to obtain multiple subsequences;

Determine the abnormal scores exceeding the preset statistical value in the subsequence as abnormal points of wafer equipment, and determine the abnormal scores other than the abnormal points of wafer equipment in the subsequence as normal points of wafer equipment;

Setting a first label for the abnormal point of the wafer equipment in the subsequence, and setting a second label for the normal point of the wafer equipment;

The data of the abnormal point of the wafer device corresponding to the first tag is determined as the abnormal data of the wafer device.

In some embodiments, evaluating the abnormal scores in the abnormal sequence according to the sliding window of the preset sequence length to determine the abnormal data of the target wafer device further includes:

According to the sliding window, the abnormal sequence is divided to obtain a first wafer sequence and a second wafer sequence; the first wafer sequence and the second wafer sequence are adjacent subsequences respectively;

Performing a difference operation on a first abnormal score in the first wafer sequence and a second abnormal score in the second wafer sequence to obtain a target sequence abnormal score; the first abnormal score and the second abnormal score include an abnormal score ranked first in the abnormal sequence;

When the abnormal score of the target sequence does not exceed a preset score threshold, setting a second label for the abnormal score in the first wafer sequence;

Determine the abnormal scores other than those set as the second label as target wafer device abnormal data.

In a second aspect, in this embodiment, a wafer manufacturing-based abnormality detection device is provided, the device comprising: a data acquisition module, an error determination module, and a data evaluation module;

The data acquisition module is used to acquire the time series data to be detected; the time series data to be detected includes sensor data of the wafer equipment;

The error determination module is used to determine the reconstruction error corresponding to the time series data to be detected according to a pre-trained target deep learning model; the target deep learning model is trained according to a preset adversarial learning method;

The data evaluation module is used to determine an abnormal sequence according to the reconstruction error, and evaluate the abnormal score in the abnormal sequence according to a sliding window of a preset sequence length to determine the abnormal data of the target wafer equipment.

In a third aspect, an electronic device is provided in this embodiment, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the wafer manufacturing-based anomaly detection method described in the first aspect is implemented.

In a fourth aspect, in this embodiment, a storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the abnormality detection method based on wafer manufacturing described in the first aspect above is implemented.

Compared with the related art, the wafer manufacturing-based abnormality detection method, device and electronic device provided in this embodiment preprocesses the training data, and then trains the target deep learning model based on the preprocessed training data and bidirectional restricted adversarial learning to process the time series data to be detected, which is conducive to capturing and identifying various types of abnormal situations, thereby improving the accuracy and efficiency of abnormality detection of the time series data to be detected. Afterwards, the reconstruction error corresponding to the time series data to be detected is obtained, and the time series data to be detected is monitored in real time through a sliding window, thereby improving the recognition efficiency of abnormal data in the time series data to be detected.

The details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

FIG1 is a hardware structure block diagram of a terminal of an abnormality detection method based on wafer manufacturing provided in this embodiment;

FIG2 is a flow chart of an abnormality detection method based on wafer manufacturing provided in an embodiment of the present application;

FIG3 is a flow chart of an abnormality detection method based on wafer manufacturing provided in this specific embodiment;

FIG4 is a structural diagram of a deep learning model provided in this specific embodiment;

FIG5 is a structural block diagram of the abnormality detection device based on wafer manufacturing according to this embodiment.

Detailed ways

In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with general skills in the technical field to which this application belongs. The words "one", "a", "the", "these" and the like in this application do not indicate a quantitative limitation, and they may be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "multiple" involved in this application refers to two or more. "And/or" describes the association relationship of associated objects, indicating that there may be three relationships, for example, "A and/or B" may mean: A exists alone, A and B exist at the same time, and B exists alone. Generally, the character "/" indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

Wafer manufacturing is an important part of the semiconductor manufacturing process. Its main purpose is to process silicon wafers into wafers that can be used for integrated circuits. The wafer manufacturing process involves multiple complex processes and technologies, including silicon ingot growth, chip cutting, heat treatment, and subsequent wafer processing. During the wafer processing process, circuits and electronic components such as transistors, capacitors, and logic switches are made on the wafer. This process usually includes steps such as cleaning the wafer, surface oxidation, chemical vapor deposition, coating, exposure, development, etching, ion implantation, and metal sputtering to complete the processing and production of several layers of circuits and components.

The complexity and dynamics of wafer manufacturing lead to the following problems in existing eigenvalue information extraction methods: 1. Some abnormal patterns are not statistically significant, and fault detection for such patterns is not accurate enough. 2. Different process recipes require experienced engineers to create different models. 3. Changes in process recipes require a lot of time to update the model. 4. Some process steps take too long, and eigenvalue extraction is not timely enough.

Due to the complexity and precision of semiconductor chips, traditional rule-based or statistical detection methods often cannot meet the requirements of accuracy and efficiency. In the wafer manufacturing process, the raw data of different sensors are diverse, and it is difficult for chip engineers to mark and judge whether a single time point of different sensors is abnormal. With the development of artificial intelligence algorithms such as machine learning and deep learning, unsupervised learning algorithms that do not require labels have become the mainstream of fault detection algorithms. The methods for detecting and classifying faults in the prior art include: methods based on mean and standard deviation, methods based on probability models, methods based on distance, and methods based on clustering, etc. Among them, the methods based on mean and standard deviation are simple and easy to implement and have high computational efficiency; however, the processing effect for non-Gaussian distributed data is poor, and the robustness to outliers is poor. Methods based on probability models, such as Gaussian sum models, can model complex data distributions and take into account the correlation between features; however, they have high requirements for assumptions about data distribution; and have high computational complexity. Distance-based methods, such as the k-value nearest neighbor algorithm, do not rely on prior knowledge of data distribution; can adapt to different types of data; but have high computational complexity; and are not applicable to high-dimensional data sets. Clustering-based methods: can find dense areas in the data set and treat sparse areas as anomalies; however, the computational overhead is high for high-dimensional data and complex data distribution.

Based on the problems existing in the prior art, this application adopts a deep learning algorithm for anomaly detection in the wafer manufacturing process. Among them, the deep learning algorithm has powerful automatic learning and characterization capabilities, and can learn and identify complex abnormal patterns from a large amount of data. Once the constructed and trained model is deployed, by automatically collecting the sensor parameters of the equipment at a specific frequency, the data flow changes in the wafer manufacturing process can be monitored in real time, and then the patterns that do not match the normal working mode can be identified, and the operator or the automation system can be promptly alerted so that corrective measures can be taken. This ability to respond quickly and handle anomalies can effectively prevent the impact of failures on production and greatly improve the yield of chip production.

The method embodiment provided in this embodiment can be executed in a terminal, a computer or a similar computing device. For example, when running on a terminal, FIG1 is a hardware structure block diagram of a terminal of an abnormality detection method based on wafer manufacturing provided in this embodiment. As shown in FIG1 , the terminal may include one or more (only one is shown in FIG1 ) processors 102 and a memory 104 for storing data, wherein the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA. The above-mentioned terminal may also include a transmission device 106 and an input and output device 108 for communication functions. It can be understood by those skilled in the art that the structure shown in FIG1 is only for illustration and does not limit the structure of the above-mentioned terminal. For example, the terminal may also include more or fewer components than those shown in FIG1 , or have a different configuration than that shown in FIG1 .

The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the abnormality detection method based on wafer manufacturing in this embodiment. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

The transmission device 106 is used to receive or send data via a network. The above network includes a wireless network provided by the communication provider of the terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly.

In this embodiment, a wafer manufacturing-based anomaly detection method is provided, which automatically collects sensor parameters of wafer manufacturing equipment at a specific frequency, monitors data changes in real time by using the deep learning algorithm proposed in the present invention, and promptly alarms and feeds back accurate and timely anomaly detection results to the production line, which can remind chip engineers or automation systems to intervene in the production process, effectively prevent the impact of faults on production, and improve the yield of chip production. Figure 2 is a flow chart of the wafer manufacturing-based anomaly detection method provided in an embodiment of the present application. As shown in Figure 2, the process includes the following steps:

Step S210, obtaining timing data to be detected; the timing data to be detected includes sensor data of the wafer equipment.

Among them, sensors are set on the process nodes and equipment of wafer manufacturing, and the processor collects sensor data according to a preset frequency, which is the time series data to be detected. Exemplarily, the collected sensor data, that is, the time series data to be detected, covers multiple parameters such as temperature, pressure, flow, concentration, speed, etc., reflecting the real-time status of the manufacturing process. By collecting sensor data of the wafer, the status of the wafer manufacturing process can be monitored in real time and potential problems can be discovered in time. Specifically, the method of collecting time series data to be detected may include: wired connection collection, wireless connection collection, and edge computing collection.

Step S220, determining the reconstruction error corresponding to the time series data to be detected based on a pre-trained target deep learning model; the target deep learning model is trained according to a preset adversarial learning method.

Among them, the processor obtains the target deep learning model through the preset adversarial learning method training. Specifically, the deep learning model based on the wafer manufacturing sensor waveform is trained according to the preset adversarial learning method to obtain the optimal model parameters, which is the target deep learning model. Specifically, the training time series data of the sensor in the wafer manufacturing process is obtained, and the training time series data is reconstructed and anomalies are identified through the preset adversarial learning method. The preset adversarial learning method includes a generator and a discriminator, and the training data is alternately optimized by the generator and the discriminator to obtain the target deep learning model. The time series data to be detected is input into the pre-trained target deep learning model, and the time series data to be detected is processed by the generation module based on the adversarial learning method in the target deep learning model to obtain a reconstructed sequence; thereafter, the reconstructed sequence and the input sequence are determined by the discriminant module based on the adversarial learning method in the target deep learning model, and the reconstruction error between the reconstructed sequence and the input sequence is calculated at each time point, and then the anomaly is identified through the reconstruction error. Exemplarily, the error of each point between the reconstructed sequence and the input sequence is calculated by adopting the L2 regularization method; wherein the L2 regularization method includes: adding a regularization term to the loss function, that is, the loss function is equal to the original loss plus λ multiplied by the square sum of the model parameters. Wherein λ is a regularization parameter used to control the strength of regularization.

Step S230, determining an abnormal sequence according to the reconstruction error, and evaluating the abnormal scores in the abnormal sequence according to a sliding window of a preset sequence length to determine abnormal data of the target wafer device.

Among them, after the processor obtains the reconstruction error according to the target deep learning model, the processor determines the abnormal sequence in the time series data to be detected according to the reconstruction error. Specifically, by normalizing the reconstruction error, a normalized score is obtained, and the abnormal sequence is determined according to the normalized score. Exemplarily, a sequence with a normalized score between 0 and 1 at each time point is an abnormal sequence. After determining the abnormal sequence, a sliding window of a preset sequence length is introduced to divide the abnormal score array in the abnormal sequence into multiple signal segments, and multiple window data are obtained. The abnormal scores of multiple window data in the abnormal sequence are evaluated according to the quartile statistical method, and then the target wafer equipment abnormal data is obtained.

Through the above steps, the target deep learning model based on bidirectional restricted adversarial learning is used to process the time series data to be detected, which is conducive to capturing and identifying various types of abnormal situations, thereby improving the accuracy and efficiency of abnormal detection of the time series data to be detected. Afterwards, the reconstruction error corresponding to the time series data to be detected is obtained, and the time series data to be detected is monitored in real time through a sliding window, thereby improving the recognition efficiency of abnormal data in the time series data to be detected.

In some of the embodiments, before step S210 of acquiring the time series data to be detected, the steps include: step S201 to step S203.

Step S201, acquiring a plurality of training time series data in a wafer manufacturing process, and processing the plurality of training time series data to obtain a plurality of unified training time series data.

Among them, when training the target deep learning model, the processor obtains multiple training time series data in the wafer manufacturing process. Exemplarily, multiple training time series data are the latest time series data obtained in real time, which ensures the real-time nature of the training data and further improves the accuracy of the subsequent target deep learning model. Thereafter, multiple training time series data are processed. Specifically, the method for processing the training time series data includes: detecting whether the timestamps and data values of multiple training time series data are missing, and if so, the missing timestamps and data values are supplemented, and further, the missing timestamps and data values are supplemented according to the previous training time series data, or the mode of the data values in the multiple training time series data is calculated, and then the missing data values are supplemented according to the mode. After the processor performs the supplementation processing on the multiple training time series data, the multiple supplemented training time series data are intercepted with a unified time length according to the preset timestamps, thereby obtaining multiple unified training time series data.

Step S202, determining dense sequence data in a plurality of unified training time series data; the dense sequence data includes sequence data that conforms to a normal distribution.

Among them, after obtaining multiple unified training time series data, the processor performs normalization processing on the multiple time series data to obtain sequence data that conforms to the normal distribution. Exemplarily, through the 3Sigma statistical method, it is beneficial to limit the range of the unified training time series data to within three times the standard deviation of the normal distribution, thereby excluding abnormal sequence data and obtaining dense sequence data. The preset deep learning model is trained through the dense sequence data, which is beneficial to improve the accuracy of the target deep learning model.

Step S203, training a preset deep learning model according to the dense sequence data and a preset adversarial learning method to determine a target deep learning model; the preset deep learning model is based on the waveform of the wafer sensor.

Among them, after the processor obtains multiple dense sequence data, based on the preset adversarial learning method, the dense sequence data is input into the preset deep learning model for training. Further, the preset adversarial learning method in the preset deep learning model includes a generation module and a discrimination module, each module is connected by two to three layers of fully connected layers. Input dense sequence data to the generation module and the discrimination module to train the preset deep learning model. When the model training is completed, four models are generated, that is, the target deep learning model includes: encoder, decoder, true and false discriminator, and measurement discriminator. Specifically, the encoder and the decoder are two generators; the encoder maps the time series to the latent space, and the decoder converts the latent space to the reconstructed time series; the true and false discriminator is used to distinguish whether the input sequence is a real time series to be detected or a false time series generated by the decoder, and the measurement discriminator is used to measure the effectiveness of the encoder mapping the input sequence to the latent space. Further, the processor collects multiple training time series data at intervals and inputs them into the deep learning model for training; thereafter, it is iteratively optimized through the generation module and the discrimination module to obtain the target deep learning model.

Through the above steps, when training the target deep learning model, the training time series data is collected and processed, and dense sequence data that conforms to the normal distribution is obtained according to the statistical method, and then the deep learning model is trained by the dense sequence data to obtain the target deep learning model. By completing, truncating and normalizing the training time series data, more accurate training data is obtained, which is conducive to improving the accuracy of the target deep learning model obtained by training.

In some of the embodiments, before determining the reconstruction error corresponding to the time series data to be detected based on the pre-trained target deep learning model in step S220, it includes: acquiring initial time series data on the sensor of the wafer according to a preset sampling frequency, processing the initial time series data according to the dense sequence data, and obtaining the time series data to be detected.

Among them, the processor automatically collects sensor data on the equipment in the wafer manufacturing process according to the preset sampling frequency, which is the initial time series data; it determines whether the data length of the initial time series data is less than the length of the dense sequence data. If so, the length of the initial time series data is padded according to the length of the dense sequence data. Afterwards, the missing timestamps and data values of the initial time series data are detected and padded, and the initial time series data after data padded is normalized to obtain the time series data to be detected.

Through the above steps, data preprocessing operations are performed on the initial time series data to be detected to obtain the time series data to be detected, thereby achieving consistency between the time series data to be detected and the training time series data, which is beneficial to improving the efficiency and accuracy of subsequent processing of the time series data to be detected through the target deep learning model.

In some embodiments, step S220 includes: step S221 to step S223.

Step S221, mapping the time series data to be detected to a preset latent space to obtain a coding vector.

Specifically, after the processor loads the time series data to be detected into the target deep learning model, the time series data to be detected is mapped to the latent space through the generation module in the target deep learning model, specifically the encoder in the generation module, to obtain the encoding vector corresponding to the time series data to be detected. Further, the effectiveness of the encoder in mapping the time series data to be detected to the latent space is measured through the discriminator module in the target deep learning model, specifically the measurement discriminator in the discriminator module.

Step S222: When the coding vector is within a preset threshold range, the coding vector is processed and a reconstructed sequence is generated.

Specifically, the processor determines whether the coding vector is within a preset threshold range. If so, the coding vector is converted into a reconstructed time series through the generation module in the target deep learning model, specifically the decoder in the generation module, that is, a reconstructed sequence is generated.

Step S223: determine the difference between the data of the reconstructed sequence and the time series data to be detected as the reconstruction error.

Among them, the processor distinguishes the data of the reconstructed sequence from the time series data to be detected through the discrimination module in the target deep learning model, specifically the true and false discriminator in the discrimination module; then calculates the difference between the data of the reconstructed sequence and the time series data to be detected at each time point, which is the reconstruction error. Specifically, calculating the difference between the data of the reconstructed sequence and the time series data to be detected at each time point by the L2 regularization method helps to avoid the problem that the target deep learning model is too sensitive to noise or abnormal scores in the training data, thereby improving the generalization ability of the target deep learning model.

Through the above steps, the generation module and the discrimination module in the target deep learning model are used to map, reconstruct and judge the time series data to be detected, thereby improving the accuracy and efficiency of determining the reconstruction error.

In some embodiments, determining the abnormal sequence according to the reconstruction error in step S230 includes:

Step S231 , normalizing the reconstruction errors corresponding to multiple time points in the time series data to be detected to obtain error scores for the multiple time points.

After the reconstruction errors corresponding to the multiple time points are determined, the multiple reconstruction errors are normalized to obtain error scores corresponding to the multiple reconstruction errors.

Step S232, determining an abnormal score array according to a preset abnormal value; obtaining the preset abnormal value by processing the error score according to a preset statistical method; determining that the sequence corresponding to the abnormal score array is an abnormal sequence.

The processor presets an abnormal value, and then determines abnormal score data in a plurality of error scores according to the abnormal value. Exemplarily, the abnormal value in the abnormal score array is in the range of 0 to 1.

Through the above steps, the reconstruction error is normalized, the error score corresponding to the reconstruction error is determined, and then the abnormal score array is determined according to the error score, and the sequence corresponding to the abnormal score array is determined to be an abnormal sequence, thereby improving the accuracy of determining the abnormal sequence.

In some embodiments, in step S230, the abnormal scores in the abnormal sequence are evaluated according to a sliding window of a preset sequence length to determine the abnormal data of the target wafer device, including:

Step S233: divide the abnormal sequence according to a sliding window of a preset sequence length to obtain multiple subsequences.

The processor presets a sliding window of sequence length, and performs sliding division on the abnormal sequence according to the sliding window to obtain multiple subsequences. Further, the multiple subsequences are window data of fixed time.

Step S234, determining the abnormal scores exceeding the preset statistical value in the subsequence as wafer equipment abnormal points, determining the abnormal scores other than the wafer equipment abnormal points in the subsequence as wafer normal points; setting the first label for the wafer equipment abnormal points in the subsequence, and setting the second label for the wafer normal points.

Among them, the processor regards the abnormal score points in the subsequence that exceed the preset statistical points as wafer equipment abnormal points. Furthermore, the preset statistical points are determined according to the quartile statistical method, and the quartile statistical values are calculated in the subsequence, and the quartile statistical values are the preset statistical points. The abnormal score that exceeds the quartile statistical value in the subsequence is judged, and the abnormal score corresponds to the wafer equipment abnormal point. After the wafer equipment abnormal point is determined, a first label, that is, an abnormal label, is set for the wafer equipment abnormal point; at the same time, a second label, that is, a normal label, is set for the normal wafer points in the subsequence except the wafer equipment abnormal point. Exemplarily, the first label of the wafer equipment abnormal point is 1, and the second label of the wafer normal point is 0.

Step S235 , determining that the data of the abnormal point of the wafer device corresponding to the first tag is the abnormal data of the target wafer device.

Through the above steps, the abnormal sequence is divided into multiple subsequences by sliding the sliding window, and the target wafer device abnormal data in the multiple subsequences is determined by the quartile statistical method. Evaluating the target wafer device abnormal data by sliding the window is conducive to increasing the recall rate of abnormal data, and further conducive to improving the accuracy of abnormality detection.

In some embodiments, the method of evaluating the abnormal scores in the abnormal sequence according to the sliding window of the preset sequence length in step S230 to determine the abnormal data of the target wafer device further includes:

Step S236: divide the abnormal sequence according to the sliding window to obtain a first wafer sequence and a second wafer sequence; the first wafer sequence and the second wafer sequence are adjacent subsequences respectively.

The processor divides the abnormal sequence into a plurality of adjacent first wafer sequences and second wafer sequences according to the sliding window.

Step S237, performing a difference operation on the first abnormal score in the first wafer sequence and the second abnormal score in the second wafer sequence to obtain a target sequence abnormal score; the first abnormal score and the second abnormal score include the abnormal score ranked first in the abnormal sequence.

The processor determines the maximum abnormal score in the first wafer sequence and the second wafer sequence, which are the first abnormal score and the second abnormal score, respectively. The first abnormal score and the second abnormal score are subjected to a difference operation, that is, the difference between the first abnormal score and the second abnormal score is obtained, and the difference is divided by the value of the first abnormal score, which is the target sequence abnormal score.

Step S238: When the target sequence abnormality score does not exceed the preset score threshold, a second label is set for the abnormality score in the first wafer sequence, and the abnormality scores other than those set as the second label are determined as target wafer device abnormality data.

The processor determines whether the target sequence abnormal score exceeds a preset score threshold. When the target sequence abnormal score does not exceed the preset score threshold, it can be understood that the abnormal score in the first wafer sequence is a normal score, that is, a second label is set for the abnormal score in the first wafer sequence. The processor re-updates the abnormal score in the abnormal sequence and determines the abnormal score other than the one set as the second label as abnormal data of the target wafer device.

Through the above steps, the sliding window will increase the recall rate of abnormal data, but also increase the false detection rate of abnormal data. Therefore, the above critical recall method, that is, updating the labels of abnormal scores in multiple subsequences through difference operations, is conducive to reducing the false detection rate of abnormal detection, and thus helps to improve the accuracy of abnormal detection.

The present embodiment is described and illustrated by means of specific examples below.

FIG3 is a flow chart of an abnormality detection method based on wafer manufacturing provided in this specific embodiment. As shown in FIG3 , the abnormality detection method based on wafer manufacturing includes the following steps:

Step S310, preprocessing training data and training the deep learning model.

Specifically, the processor collects and loads the latest multiple time series data of sensor A during the wafer manufacturing process. The multiple time series data here are the multiple training time series data in the aforementioned embodiment. Thereafter, the missing timestamps and data values of the multiple time series data are detected and completed, and the multiple time series data are truncated to a unified length, and the length of the historical data sequence is saved. The length of the historical data sequence here is the dense sequence data in the aforementioned embodiment. After normalizing these multiple time series data, the mean sequence of historical dense data is calculated and saved. The above-mentioned deep learning model carefully constructed for the sensor waveform in the semiconductor field is loaded, including two generators and two discriminators. The two generators can be regarded as encoders and decoders, respectively. The collected dense sequences are batch-input into the model, and the generator and discriminator are used to alternately optimize, and the trained optimal model parameters are saved.

Preferably, the operation of detecting and completing missing time and data of multiple time series data may include: calculating the interval of timestamps from the front to the back of a time series data, obtaining the mode of the timestamp interval, and using the mode as the data collection period. Dividing the data collection period by the interval of the previous and next timestamps in the time series data, if the divisible result is equal to n, then completing n-1 forward values between the time series data corresponding to the previous timestamp and the next timestamp, where n is a positive integer greater than or equal to 1.

Preferably, in order to keep the length of each time series data consistent, it is necessary to obtain the time series data with the shortest data length among multiple time series data; and according to the shortest data length, truncate these multiple time series data to obtain time series data of uniform length, and save it as the historical data sequence length.

Preferably, the 3sigma statistical method is used to process multiple time series data after completion and truncation, and the time points corresponding to the time series data outside 3sigma are judged as outliers. The same time point of multiple time series data is input into 3sigma. As long as there is a time point outside 3sigma, the time point is determined to be an outlier, and the entire time series including the outlier is determined as an outlier sequence. Thereafter, the outlier sequences are filtered out of the multiple time series data, and dense sequences are collected. These multiple dense sequences are horizontally normalized and the mean of each timestamp is calculated to obtain the mean sequence of historical dense data.

Step S320: pre-processing the time series data to be detected.

Specifically, the processor automatically collects sensor parameters of sensor A on the device at a preset specific frequency to obtain online real-time window data, where the window data is the time series data to be detected in the aforementioned embodiment. When the length of the window data is less than the length of the historical data sequence, it is padded according to the mean sequence of historical dense data, and then the window data is preprocessed through operations such as truncation of the time series data to be trained in step S320 to obtain the time series data to be detected.

Step S330: Perform anomaly detection and identification on the time series data to be detected based on the trained deep learning model.

Specifically, after the online real-time window data is preprocessed to obtain the time series data to be detected, the time series data to be detected is loaded into the trained deep learning model, that is, the target deep learning model in the aforementioned embodiment. When evaluating the anomalies in the time series data to be detected by the target deep learning model, the reconstructed sequence of the time series data to be detected is first obtained by the two generators in the deep learning model, namely the encoder and the decoder, and then the anomaly is identified by calculating the reconstruction error between the reconstructed sequence and the input sequence. The generator here is the generation module in the aforementioned embodiment. Exemplarily, the error at each time point between the reconstructed sequence and the input sequence is calculated by the L2 regularization method. Thereafter, the normalized score of the reconstruction error is calculated, and the anomaly score at the time point corresponding to the normalized score between 0 and 1 is used as the final anomaly score array.

Afterwards, a sliding window is introduced to perform data segmentation operations on the anomaly score array of online data, and the array is divided into multiple signal segments, thereby obtaining multiple window data of fixed seconds, and the multiple window data are the subsequences in the aforementioned embodiment. The quartile statistics are calculated in the multiple window data, and the points exceeding the quartile statistics are set as anomalies. At the same time, the output anomaly score array is converted into an anomaly label array. The label of the anomaly point is 1, which is the first label in the aforementioned embodiment; the label of the normal point is 0, which is the second label in the aforementioned embodiment. The sliding window method is used to evaluate the anomaly score, which increases the recall rate of the anomaly.

Furthermore, for the conversion of anomaly scores to anomaly labels, a sliding window is used to divide the complete time series into multiple subsequences of equal length. Each subsequence is used as a sample to detect whether each subsequence contains anomalies, so as to increase the recall rate of anomalies. Specifically, points in the subsequence whose anomaly scores exceed the quartile are set as anomalies. Exemplarily, the window sequence length is set to one-third of the complete sequence, and the sliding step size is set to one-thirtieth of the complete sequence. First, the lower quartile value Q1 and the upper quartile value Q3 in each sliding window of the input sequence are calculated, and the points in the window data that are greater than the upper quartile plus the lower quartile value Q1 and the upper quartile value Q3 are added. All time points (Q3-Q1) are judged as abnormal, and abnormal array 1 is generated, as shown below:

in, represents anomaly array 1, Q1 represents the lower quartile value, Q3 represents the upper quartile value; t represents the timestamp; Determined based on the yield of timing data.

Subtract the online window data that is less than the lower quartile All time points (Q3-Q1) are judged as abnormal, and abnormal array 2 is generated, as shown below:

in, represents anomaly array 2, Q1 represents the lower quartile value, Q3 represents the upper quartile value; t represents the timestamp; Determined based on the yield of the timing data. Mark the abnormal points and non-abnormal points in the abnormal array, where the abnormal points are marked as 1 and the non-abnormal points are marked as 0. Furthermore, if there is an abnormal point in the two abnormal arrays, it is considered an abnormal point.

At the same time, the sliding window will increase the recall rate for anomalies, and will also increase the false detection rate. Therefore, further, a post-processing method of critical recall is used to reduce the false detection rate for anomaly detection. Specifically, the maximum anomaly score in the front and rear sliding windows is subjected to a difference operation, where the front and rear sliding windows are the first wafer sequence and the second wafer sequence in the aforementioned embodiment, respectively. If the difference operation result obtained is less than the set threshold, all the labels of the abnormal points in the front window, that is, the first wafer sequence, are reassigned to the normal second label. The final post-processed two-dimensional anomaly array is output and the abnormal points in the timing window of the visualization page are marked in red in real time for chip engineers to observe and monitor the changes in data in the window in real time.

Furthermore, for the critical recall method, the outliers are processed to reduce the false positive rate, thereby improving the accuracy of anomaly recognition. Specifically, the maximum anomaly score of each sliding window sequence is first saved. For each anomaly sequence, we use the maximum anomaly score to represent it, that is, the maximum anomaly score of the previous and next intervals is calculated by the difference operation to obtain M, as shown below:

Among them, k represents the kth abnormal sequence, Indicates the maximum anomaly score of the previous sliding window sequence; represents the maximum anomaly score of the next sliding window sequence, Represents the difference result. If M does not exceed the preset abnormal threshold θ, all point labels in the preceding interval are assigned 0. By re-identifying the abnormal points, it is helpful to reduce false positives.

In one embodiment, referring to FIG4 , FIG4 is a structural diagram of a deep learning model provided in this specific embodiment. The entire model mainly includes a data preparation component 41 , a model training component 42 and an abnormal output component 43 .

The data preparation component 41 mainly includes a sample pre-processing module, which performs pre-processing operations on batch sensor time series by data completion, data truncation and normalization, and then sends the pre-processed time series to the model for training in batches.

For the model training component 42, a real-time anomaly detection model based on bidirectional restricted adversarial learning is constructed for the sensor waveform in the semiconductor field, that is, the deep learning model in the aforementioned embodiment. The model includes two generators and two discriminator modules, each of which is composed of two to three layers of full connection as basic components. The generator and the discriminator are respectively embedded in the framework of adversarial learning, and the training time series data is input to train the model. After the model training is completed, four models will be generated: an encoder, a decoder, a true and false discriminator, and a measurement discriminator. Among them, the encoder and the decoder correspond to two generators. The encoder maps the time series to the latent space, and the decoder converts the latent space to the reconstructed time series. The true and false discriminator is used to distinguish whether it is a real time series input or a false time series generated by the decoder, and the measurement discriminator is used to measure the effectiveness of the encoder mapping the input sequence to the latent space.

Preferably, a batch of the latest data is pulled at regular intervals and trained using the deep learning model proposed above. The collected dense time series is input into the model, and the generator and the discriminator are alternately optimized, and the optimal model parameters are saved. For each training cycle, the number of iterations for training the two discriminators is first determined. For example, the number of iterations of the discriminator can be q times, where q is a positive integer greater than 1; then the generator is iteratively trained once. Thereafter, the loss of the two discriminators and the generator is calculated and back-propagated respectively. The learning rate decay strategy and the early stopping strategy are used for training. For example, for the learning rate decay strategy, if the training loss of 10 consecutive time series data epochs does not decrease, the learning rate is reduced. For the early stopping strategy, if the training loss of 10 consecutive epochs does not decrease after the learning rate decays, the training is stopped and the optimal model parameters are saved. Among them, the trained deep learning model is optimized and adjusted, specifically based on abnormal data optimization: remove noise data with poor representation effect to ensure data quality. After cleaning the abnormal data, retrain the model to eliminate the impact of the abnormality.

The abnormal output component 43 mainly includes three stages: abnormal score detection, label conversion, and critical recall. In the abnormal score detection stage, only two generators, namely the encoder and the decoder, are used to reconstruct the reconstructed sequence, and then the anomaly is detected by calculating the deviation between the reconstructed sequence and the input sequence. Specifically, the L2 norm loss is used to evaluate the reconstruction error.

Preferably, the real-time anomaly detection method is based on bidirectional restricted adversarial learning, and the reconstructed sequence is constructed based on the training sequence. When a new sequence comes in, the difference between the previously constructed reconstructed sequence and the new sequence is detected, the anomaly score at each time point is calculated, and the anomaly is filtered out according to the preset anomaly score threshold. Using the above-mentioned deep learning algorithm to monitor the changes in sensor parameters in real time, the time series anomalies can be quickly identified, with an accuracy of 0.99+ and a precision/recall/f1-score of 0.8+.

This application collects a large amount of sensor data in the chip production process, performs targeted data cleaning and data preprocessing, including completion, truncation, and normalization steps, which is conducive to improving the quality of input data. At the same time, according to the specific needs of wafer manufacturing in the semiconductor field, a highly adaptable anomaly detection algorithm is customized and designed. The deep learning algorithm is based on bidirectional restricted adversarial learning, which can capture and identify different types of anomalies and improve the accuracy and efficiency of anomaly detection. And, the trained deep learning model is optimized and adjusted to improve its accuracy and robustness. Exemplarily, the optimization and adjustment of the model can be carried out by optimizers, learning rate decay strategies, training early stopping strategies and other methods. The trained model is used to monitor the data flow in the actual chip production in real time, and the test results are fed back to the operator or the automation system. This helps to detect abnormal situations in time and take corresponding measures to improve the yield.

It should be noted that the steps shown in the above process or the flowchart in the accompanying drawings can be executed in a computer system such as a group of computer executable instructions.

In this embodiment, an abnormality detection device based on wafer manufacturing is also provided, which is used to implement the above-mentioned embodiments and preferred implementation modes, and the descriptions that have been made will not be repeated. The terms "module", "unit", "subunit", etc. used below can implement a combination of software and/or hardware of predetermined functions. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

FIG5 is a structural block diagram of an abnormality detection device based on wafer manufacturing according to this embodiment. As shown in FIG5 , the device includes: a data acquisition module 10 , an error determination module 20 , and a data evaluation module 30 .

The data acquisition module 10 is used to acquire the time series data to be detected; the time series data to be detected includes sensor data of the wafer equipment.

The error determination module 20 is used to determine the reconstruction error corresponding to the time series data to be detected based on a pre-trained target deep learning model; the target deep learning model is trained according to a preset adversarial learning method.

The data evaluation module 30 is used to determine the abnormal sequence according to the reconstruction error, and evaluate the abnormal score in the abnormal sequence according to the sliding window of the preset sequence length to determine the abnormal data of the target wafer equipment.

It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

This embodiment also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

Optionally, the electronic device may further include a transmission device and an input/output device, wherein the transmission device is connected to the processor, and the input/output device is connected to the processor.

Optionally, in this embodiment, the processor may be configured to perform the following steps through a computer program:

S1, obtaining time series data to be detected; the time series data to be detected includes sensor data of wafer equipment.

S2, based on the pre-trained target deep learning model, determine the reconstruction error corresponding to the time series data to be detected; the target deep learning model is trained according to the preset adversarial learning method.

S3, determining an abnormal sequence according to the reconstruction error, and evaluating the abnormal score in the abnormal sequence according to a sliding window of a preset sequence length to determine abnormal data of the target wafer device.

It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and will not be repeated in this embodiment.

In addition, in combination with the wafer manufacturing-based anomaly detection method provided in the above embodiments, a storage medium may also be provided in this embodiment to implement the method. The storage medium stores a computer program; when the computer program is executed by a processor, any of the wafer manufacturing-based anomaly detection methods in the above embodiments is implemented.

It should be understood that the specific embodiments described herein are only used to explain the application, rather than to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of this application.

Obviously, the drawings are only some examples or embodiments of the present application. For ordinary technicians in the field, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that although the work done in this development process may be complicated and lengthy, for ordinary technicians in the field, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application.

The term "embodiment" in this application refers to a specific feature, structure or characteristic described in conjunction with the embodiment that can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor does it mean that it is mutually exclusive with other embodiments and is independent or optional. It is clearly or implicitly understood by those of ordinary skill in the art that the embodiments described in this application can be combined with other embodiments without conflict.

The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of patent protection. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the attached claims.

Claims (10)

1. A wafer manufacturing-based anomaly detection method, characterized in that the method comprises: Acquire time series data to be detected; the time series data to be detected includes sensor data of wafer equipment; Determine the reconstruction error corresponding to the time series data to be detected according to a pre-trained target deep learning model; the target deep learning model is trained according to a preset adversarial learning method; An abnormal sequence is determined according to the reconstruction error, and an abnormal score in the abnormal sequence is evaluated according to a sliding window of a preset sequence length to determine abnormal data of a target wafer device. 2. The wafer manufacturing-based abnormality detection method according to claim 1, characterized in that before acquiring the time series data to be detected, it comprises: Acquire multiple training time series data in a wafer manufacturing process, and process the multiple training time series data to obtain multiple unified training time series data; Determine dense sequence data in the plurality of unified training time series data; the dense sequence data includes sequence data conforming to a normal distribution; According to the dense sequence data and the preset adversarial learning method, the preset deep learning model is trained to determine a target deep learning model; the preset deep learning model is based on the waveform of the wafer sensor. 3. The wafer manufacturing-based anomaly detection method according to claim 2, characterized in that before determining the reconstruction error corresponding to the time series data to be detected according to the pre-trained target deep learning model, it includes: Acquire initial timing data from the sensor on the wafer according to the preset sampling frequency. The initial time series data is processed according to the dense sequence data to obtain the time series data to be detected. 4. The wafer manufacturing-based abnormality detection method according to claim 1, characterized in that: The step of determining the reconstruction error corresponding to the time series data to be detected according to the pre-trained target deep learning model includes: Mapping the time series data to be detected to a preset latent space to obtain a coding vector; When the encoding vector is within a preset threshold range, processing the encoding vector and generating a reconstructed sequence; The difference between the data of the reconstructed sequence and the time series data to be detected is determined as a reconstruction error. 5. The wafer manufacturing-based anomaly detection method according to claim 1, wherein determining the anomaly sequence according to the reconstruction error comprises: Normalizing the reconstruction errors corresponding to multiple time points in the time series data to be detected to obtain error scores for the multiple time points; Determine an array of abnormal scores according to preset abnormal values; the preset abnormal scores are obtained by processing the error scores according to a preset statistical method; Determine that the sequence corresponding to the abnormal score array is an abnormal sequence. 6. The wafer manufacturing-based anomaly detection method according to claim 1, characterized in that the step of evaluating the anomaly scores in the anomaly sequence according to the sliding window of the preset sequence length to determine the target wafer equipment anomaly data comprises: Dividing the abnormal sequence according to a sliding window of a preset sequence length to obtain multiple subsequences; Determine the abnormal scores exceeding the preset statistical value in the subsequence as abnormal points of wafer equipment, and determine the abnormal scores other than the abnormal points of wafer equipment in the subsequence as normal points of wafer equipment; Setting a first label for the abnormal point of the wafer equipment in the subsequence, and setting a second label for the normal point of the wafer equipment; The data of the abnormal point of the wafer device corresponding to the first tag is determined as the abnormal data of the wafer device. 7. The wafer manufacturing-based anomaly detection method according to claim 6, characterized in that the sliding window of the preset sequence length is used to evaluate the anomaly score in the anomaly sequence to determine the target wafer device anomaly data, and further comprises: According to the sliding window, the abnormal sequence is divided to obtain a first wafer sequence and a second wafer sequence; the first wafer sequence and the second wafer sequence are adjacent subsequences respectively; Performing a difference operation on a first abnormal score in the first wafer sequence and a second abnormal score in the second wafer sequence to obtain a target sequence abnormal score; the first abnormal score and the second abnormal score include an abnormal score ranked first in the abnormal sequence; When the abnormal score of the target sequence does not exceed a preset score threshold, setting a second label for the abnormal score in the first wafer sequence; Determine the abnormal scores other than those set as the second label as target wafer device abnormal data. 8. An abnormality detection device based on wafer manufacturing, characterized in that the device comprises: a data acquisition module, an error determination module, and a data evaluation module; The data acquisition module is used to acquire the time series data to be detected; the time series data to be detected includes sensor data of the wafer equipment; The error determination module is used to determine the reconstruction error corresponding to the time series data to be detected according to a pre-trained target deep learning model; the target deep learning model is trained according to a preset adversarial learning method; The data evaluation module is used to determine an abnormal sequence according to the reconstruction error, and evaluate the abnormal score in the abnormal sequence according to a sliding window of a preset sequence length to determine the abnormal data of the target wafer equipment. 9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the wafer manufacturing-based anomaly detection method according to any one of claims 1 to 7. 10. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the wafer manufacturing-based anomaly detection method according to any one of claims 1 to 7 are implemented.