CN118378024A

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

Info

Publication number: CN118378024A
Application number: CN202410805475.1A
Authority: CN
Inventors: 邱海祥; 姜辉; 王啸飞; 沈海超; 樊星
Original assignee: Hangzhou Guangli Microelectronics Co ltd
Current assignee: Hangzhou Guangli Microelectronics Co ltd
Priority date: 2024-06-21
Filing date: 2024-06-21
Publication date: 2024-07-23

Abstract

The application relates to an abnormality detection method and device based on wafer manufacturing and an electronic device, wherein the abnormality detection method comprises the following steps: the training data is preprocessed, and then the target deep learning model is obtained through training according to the preprocessed training data and the bidirectional limiting countermeasure learning. Acquiring and preprocessing time sequence data to be detected, and determining a reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model; and determining an abnormal sequence according to the reconstruction error, evaluating the abnormal score in the abnormal sequence according to a sliding window with a preset sequence length, and determining the abnormal data of the target wafer equipment, so that the method is beneficial to capturing and identifying various abnormal conditions, and further improves the identification efficiency of the abnormal data in the time sequence data to be detected.

Description

Abnormality detection method and device based on wafer manufacturing and electronic device

Technical Field

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

Background

In the existing semiconductor chip manufacturing process, wafer manufacturing is a complex, expensive and lengthy process involving hundreds of process steps, requiring simultaneous monitoring of corresponding process parameter variations. With the increasing precision and complexity of semiconductor devices involved, higher demands are being placed on the quality of detection and classification of faults during wafer fabrication, i.e., the need to discover and identify potential anomalies and faults in time.

At present, related technology is used for detecting and classifying faults commonly used in the wafer manufacturing process, wherein a model is created to extract characteristic value information in sensor original data, and statistics is carried out on the characteristic value information. Due to the complexity and precision of semiconductor chips, conventional rule-based or statistical-based detection methods often fail to meet the accuracy and efficiency requirements. In the wafer manufacturing process, the original data of different sensors are diversified, and a chip engineer is difficult to mark and judge whether a single time point of each sensor is abnormal, so that the abnormality detection precision in the wafer manufacturing process is lower.

Aiming at the problem that the abnormality detection precision in the wafer manufacturing process is to be improved in the related art, no effective solution is proposed at present.

Disclosure of Invention

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

In a first aspect, in this embodiment, there is provided an abnormality detection method based on wafer manufacturing, the method including:

acquiring time sequence data to be detected; the time sequence data to be detected comprises sensor data of wafer equipment;

determining a reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model; the target deep learning model is obtained through training according to a preset countermeasure learning method;

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

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

Acquiring a plurality of training time sequence data in the wafer manufacturing process, and processing the training time sequence data to obtain unified training time sequence data;

determining dense sequence data in a plurality of unified training sequence data; the dense sequence data comprises sequence data conforming to normal distribution;

Training the preset deep learning model according to the dense sequence data and the preset countermeasure learning method to determine a target deep learning model; the preset deep learning model is formed based on waveforms of the wafer sensor.

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

Acquiring initial time sequence data on a sensor of the wafer according to a preset sampling frequency,

And processing the initial time sequence data according to the dense sequence data to obtain time sequence data to be detected.

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

mapping the time sequence data to be detected to a preset potential space to obtain a coding vector;

when the coding vector is within a preset threshold range, processing the coding vector and generating a reconstruction sequence;

and determining a difference value between the data of the reconstruction sequence and the time sequence data to be detected as a reconstruction error.

In some of these embodiments, the determining an abnormal sequence from the reconstruction error includes:

Normalizing the reconstruction errors corresponding to a plurality of time points in the time sequence data to be detected to obtain error scores of the time points;

determining an abnormal score array according to a preset abnormal value; the preset abnormal score is obtained by processing the error score according to a preset statistical method;

and determining the sequence corresponding to the abnormal score array as an abnormal sequence.

In some embodiments, the evaluating the anomaly score in the anomaly sequence according to the sliding window with the preset sequence length to determine the anomaly data of the target wafer device includes:

dividing the abnormal sequence according to a sliding window with a preset sequence length to obtain a plurality of subsequences;

determining abnormal scores exceeding a preset statistical value in the subsequence as abnormal points of the wafer equipment, and determining abnormal scores except the abnormal points of the wafer equipment in the subsequence as normal points of the wafer equipment;

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

and determining the data of the abnormal point of the wafer equipment corresponding to the first label as abnormal data of the wafer equipment.

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

Dividing 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 respectively adjacent subsequences;

Performing differential 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 anomaly score and the second anomaly score include anomaly scores in the anomaly sequence with the scores ranked in front;

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;

and determining the abnormality scores except the second label as the abnormality data of the target wafer equipment.

In a second aspect, in this embodiment, there is provided an abnormality detection apparatus based on wafer manufacturing, the apparatus including: the system comprises a data acquisition module, an error determination module and a data evaluation module;

The data acquisition module is used for acquiring time sequence data to be detected; the time sequence data to be detected comprises sensor data of wafer equipment;

The error determining module is used for determining a reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model; the target deep learning model is obtained through training according to a preset countermeasure learning method;

the data evaluation module is used for determining an abnormal sequence according to the reconstruction error, evaluating the abnormal score in the abnormal sequence according to a sliding window with a preset sequence length, and determining the abnormal data of the target wafer equipment.

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

In a fourth aspect, in the present embodiment, there is provided a storage medium having stored thereon a computer program which, when executed by a processor, implements the wafer manufacturing-based anomaly detection method described in the first aspect.

Compared with the related art, the abnormality detection method, the abnormality detection device and the abnormality detection electronic device based on wafer manufacturing provided in the embodiment perform preprocessing on training data, further obtain a target deep learning model through training according to the preprocessed training data and based on bidirectional limiting countermeasure learning, process time sequence data to be detected, and are beneficial to capturing and identifying various types of abnormal conditions, and further improve accuracy and efficiency of abnormality detection of the time sequence data to be detected. And then, obtaining a reconstruction error corresponding to the time sequence data to be detected, and monitoring the time sequence data to be detected in real time through a sliding window, so that the recognition efficiency of abnormal data in the time sequence data to be detected is improved.

The details of one or more embodiments of the application are set forth in the accompanying drawings and the description below to provide a more thorough understanding of the other features, objects, and advantages of the application.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application and do not constitute a limitation on the application. In the drawings:

Fig. 1 is a hardware block diagram of a terminal of an abnormality detection method based on wafer manufacturing according to the present embodiment;

FIG. 2 is a flowchart of an anomaly detection method based on wafer fabrication according to an embodiment of the present application;

FIG. 3 is a flow chart of an anomaly detection method based on wafer fabrication provided in this embodiment;

FIG. 4 is a block diagram of a deep learning model provided by the present embodiment;

fig. 5 is a block diagram of the abnormality detection apparatus according to the present embodiment.

Detailed Description

The present application will be described and illustrated with reference to the accompanying drawings and examples for a clearer understanding of the objects, technical solutions and advantages of the present application.

Unless defined otherwise, technical or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terms "a," "an," "the," "these" and similar terms in this application are not intended to be limiting in number, but may be singular or plural. The terms "comprising," "including," "having," and any variations thereof, as used herein, are intended to encompass non-exclusive inclusion; for example, a process, method, and system, article, or apparatus that comprises a list of steps or modules (units) is not limited to the list of steps or modules (units), but may include other steps or modules (units) not listed or inherent to such process, method, article, or apparatus. The terms "connected," "coupled," and the like in this disclosure are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The term "plurality" as used herein means two or more. "and/or" describes an association relationship of an association object, meaning that there may be three relationships, e.g., "a and/or B" may mean: a exists alone, A and B exist together, and B exists alone. Typically, the character "/" indicates that the associated object is an "or" relationship. The terms "first," "second," "third," and the like, as referred to in this disclosure, merely distinguish similar objects and do not represent a particular ordering for objects.

Wafer fabrication is an important step in the semiconductor manufacturing process, which mainly aims at processing silicon wafers into wafers usable for integrated circuits, and involves a plurality of complex procedures and techniques including ingot growth, dicing, heat treatment, and subsequent wafer processing procedures. In wafer processing, circuits and electronic devices, such as transistors, capacitors, logic switches, etc., are fabricated on the wafer. This process typically includes the steps of cleaning the wafer, surface oxidation, chemical vapor deposition, film coating, exposure, development, etching, ion implantation, metal sputtering, etc., to complete the processing and fabrication of layers of circuits and devices.

The complexity and dynamics of wafer fabrication result in the following problems with existing eigenvalue information extraction methods: 1. some abnormal patterns are not statistically significant and failure detection for such patterns is not accurate enough. 2. Different process recipes require experienced engineers to create different models. 3. The change of process recipe takes a lot of time to update the model. 4. And part of the process steps are too long, and the characteristic value extraction is not enough in time.

Due to the complexity and precision of semiconductor chips, conventional rule-based or statistical-based detection methods often fail to meet the accuracy and efficiency requirements. In the wafer manufacturing process, the raw data of different sensors are diversified, and it is difficult for a chip engineer to mark and judge whether a single time point of the different sensors is abnormal. With the development of artificial intelligent algorithms such as machine learning, deep learning and the like, an unsupervised learning algorithm without labels becomes the main stream of the current fault detection algorithm. The method for detecting and classifying faults in the prior art comprises the following steps: mean and standard deviation based methods, probabilistic model based methods, distance based methods, clustering based methods, and the like; the method based on the mean value and the standard deviation is simple and easy to implement and has higher calculation efficiency; but the processing effect on non-gaussian distributed data is poor and the robustness against outliers is poor. Probability model-based methods, such as gaussian and model, which model complex data distributions and take into account correlations between features; but the assumption of data distribution requires higher demands; the computational complexity is high. Distance-based methods, such as k-value nearest neighbor algorithms, do not rely on a priori knowledge of the data distribution; different types of data can be adapted; but the computational complexity is high; and not applicable to high-dimensional datasets. Clustering-based methods: a density region in the dataset can be found and the sparse region is considered abnormal; but the computational overhead for high-dimensional data and complex data distributions is large.

Based on the problems existing in the prior art, the application adopts a deep learning algorithm for detecting the abnormality in the wafer manufacturing process. The deep learning algorithm has strong automatic learning and characterization capability, and can learn and identify complex abnormal modes from a large amount of data. Once the constructed and trained model is deployed, the sensor parameters of the acquisition equipment are automatically acquired at a specific frequency, so that the data flow change in the wafer manufacturing process can be monitored in real time, a mode which is not matched with the normal working mode is further identified, and an alarm is timely given to an operator or an automation system so as to take corrective measures. The capability of quick response and abnormality handling can effectively prevent the influence of faults on production, and greatly improve the yield of chip production.

The method embodiments provided in the present embodiment may be executed in a terminal, a computer, or similar computing device. For example, the method is run on a terminal, and fig. 1 is a block diagram of a hardware structure of the terminal based on the abnormality detection method for wafer manufacturing according to the present embodiment. As shown in fig. 1, the terminal may include one or more (only one is shown in fig. 1) processors 102 and a memory 104 for storing data, wherein the processors 102 may include, but are not limited to, a microprocessor MCU, a programmable logic device FPGA, or the like. The terminal may also include a transmission device 106 for communication functions and an input-output device 108. It will be appreciated by those skilled in the art that the structure shown in fig. 1 is merely illustrative and is not intended to limit the structure of the terminal. For example, the terminal may also include more or fewer components than shown in fig. 1, or have a different configuration than shown in fig. 1.

The memory 104 may be used to store a computer program, for example, a software program of application software and a module, such as a computer program corresponding to the abnormality detection method based on wafer fabrication in the present embodiment, and the processor 102 executes the computer program stored in the memory 104 to perform various functional applications and data processing, that is, to implement the above-described method. Memory 104 may include high-speed random access memory, and may also include 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 memory remotely located relative to the processor 102, which may be connected to the terminal via a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The transmission device 106 is used to receive or transmit data via a network. The network includes a wireless network provided by a communication provider of the terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, simply referred to as a NIC) that can connect to other network devices through a base station to communicate with the internet. In one example, the transmission device 106 may be a Radio Frequency (RF) module for communicating with the internet wirelessly.

In the embodiment, the sensor parameters of the wafer manufacturing equipment are automatically acquired at a specific frequency, the change of data is monitored in real time by using the deep learning algorithm provided by the application, and the accurate and timely abnormal detection result is timely alarmed and fed back to a production line, so that a chip engineer or an automatic system can be reminded to intervene in the production process, the influence of faults on the production is effectively prevented, and the yield of chip production is improved. Fig. 2 is a flowchart of an abnormality detection method based on wafer manufacturing according to an embodiment of the present application, as shown in fig. 2, the flowchart includes the following steps:

step S210, obtaining time sequence data to be detected; the timing data to be detected includes sensor data of the wafer apparatus.

The method comprises the steps of setting sensors on process nodes and equipment for manufacturing wafers, and acquiring sensor data by a processor according to preset frequency, namely time sequence data to be detected. The acquired sensor data, that is, the time series data to be detected, covers various parameters such as temperature, pressure, flow, concentration, speed and the like, and reflects the real-time state in the manufacturing process. By collecting the sensor data of the wafer, the state of the wafer manufacturing process can be monitored in real time, and potential problems can be found in time. Specifically, the method for collecting time sequence data to be detected can include: wired connection acquisition, wireless connection acquisition, edge calculation acquisition and the like.

Step S220, determining a reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model; the target deep learning model is obtained through training according to a preset countermeasure learning method.

The processor trains a target deep learning model through a preset countermeasure learning method. Specifically, training a deep learning model based on a wafer manufacturing sensor waveform according to a preset countermeasure learning method to obtain optimal model parameters, namely a target deep learning model. Specifically, training time sequence data of a sensor in the wafer manufacturing process are obtained, and the training time sequence data are reconstructed and abnormal identification is carried out through a preset countermeasure learning method. The preset countermeasure learning method comprises a generator and a discriminator, and training data are alternately optimized through the generator and the discriminator to obtain a target deep learning model. Inputting the time sequence data to be detected into a target deep learning model which is trained in advance, and processing the time sequence data to be detected through a generating module based on an countermeasure learning method in the target deep learning model to obtain a reconstruction sequence; and determining a reconstruction sequence and an input sequence by a discrimination module based on an countermeasure learning method in the target deep learning model, calculating a reconstruction error between the reconstruction sequence and the input sequence at each time point, and further carrying out anomaly identification by the reconstruction error. Illustratively, the error of each point between the reconstructed sequence and the input sequence is calculated by employing an L2 regularization method; the L2 regularization method comprises the following steps: a regularization term is added to the loss function, i.e., the loss function is equal to the original loss plus λ times the sum of squares of the model parameters. Where λ is a regularization parameter for controlling the strength of regularization.

And step S230, determining an abnormal sequence according to the reconstruction error, evaluating the abnormal score in the abnormal sequence according to a sliding window with a preset sequence length, and determining the abnormal data of the target wafer equipment.

After obtaining the reconstruction error according to the target deep learning model, the processor determines an abnormal sequence in the time sequence data to be detected according to the reconstruction error. Specifically, normalization processing is performed on the reconstruction errors to obtain normalization scores, and abnormal sequences are determined according to the normalization scores. Illustratively, the sequence with normalized score between 0 and 1 at each time point is an outlier sequence. After the abnormal sequence is determined, a sliding window with a preset sequence length is introduced to divide an abnormal score array in the abnormal sequence into a plurality of signal segments to obtain a plurality of pieces of window data, and the abnormal scores of the plurality of pieces of window data in the abnormal sequence are evaluated according to a four-bit statistical method to obtain the abnormal data of the target wafer equipment.

Through the steps, the time sequence data to be detected is processed based on the target deep learning model of bidirectional limiting countermeasure learning, so that various abnormal conditions can be captured and identified, and the accuracy and efficiency of abnormal detection of the time sequence data to be detected are improved. And then, obtaining a reconstruction error corresponding to the time sequence data to be detected, and monitoring the time sequence data to be detected in real time through a sliding window, so that the recognition efficiency of abnormal data in the time sequence data to be detected is improved.

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

Step S201, a plurality of training time sequence data in the wafer manufacturing process are obtained, and the plurality of training time sequence data are processed to obtain a plurality of unified training time sequence data.

The processor acquires a plurality of training time sequence data in the wafer manufacturing process when training the target deep learning model. The plurality of training time sequence data are illustratively the latest time sequence data obtained in real time, so that the instantaneity of the training data is ensured, and the accuracy of the target deep learning model obtained later is further improved. Thereafter, the plurality of training time series data is processed, specifically, the method for processing the training time series data includes: detecting whether the time stamps and the data values of the plurality of training time sequence data are missing, if yes, complementing the missing time stamps and the data values, and further complementing the missing time stamps and the missing data values according to the previous training time sequence data, or calculating the mode of the data values in the plurality of training time sequence data, and further complementing the missing data values according to the mode. After the processor performs completion processing on the plurality of training time sequence data, intercepting operation of uniform time length is performed on the plurality of training time sequence data after completion according to a preset time stamp, and then a plurality of uniform training time sequence data are obtained.

Step S202, dense sequence data in a plurality of unified training sequence data are determined; the dense sequence data includes sequence data conforming to a normal distribution.

After acquiring a plurality of unified training time sequence data, the processor performs normalization processing on the plurality of pieces of time sequence data to acquire the sequence data conforming to normal distribution. By means of a 3Sigma statistical method, the range of unified training time sequence data is limited within three times of standard deviation of normal distribution, abnormal sequence data are eliminated, dense sequence data are obtained, a preset deep learning model is trained through the dense sequence data, and accuracy of a target deep learning model is improved.

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

After the processor acquires a plurality of dense sequence data, the dense sequence data is input into a preset deep learning model for training based on a preset countermeasure learning method. Further, the preset countermeasure learning method in the preset deep learning model comprises a generating module and a judging module, wherein each module is obtained by connecting two to three full-connection layers. And inputting the dense sequence data to a generating module and a judging module so as to train a preset deep learning model. When model training is completed, generating four models, namely a target deep learning model comprises: encoder, decoder, true and false discriminant, and measurement discriminant. Specifically, the encoder and decoder are two generators; the encoder maps the temporal sequence to a potential space, and the decoder converts the potential space to a reconstructed temporal sequence; the true-false discriminator is used for discriminating whether the input sequence is a true time sequence to be detected or a false time sequence generated by the decoder, and the measurement discriminator is used for measuring the effectiveness of the encoder in mapping the input sequence to the potential space. Further, the processor collects a plurality of training time sequence data at intervals and inputs the training time sequence data into the deep learning model for training; and performing iterative optimization through the generation module and the discrimination module to obtain the target deep learning model.

Through the steps, when the target deep learning model is trained, training time sequence data are collected, the training time sequence data are processed, dense sequence data which accord with normal distribution are obtained according to a statistical method, and then the deep learning model is trained through the dense sequence data, so that the target deep learning model is obtained. The training time sequence data is subjected to operations such as complement, truncation, normalization and the like, so that training data with higher accuracy is obtained, and the accuracy of a target deep learning model obtained through training is improved.

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

The processor automatically collects sensor data on equipment in the wafer manufacturing process according to a preset sampling frequency, namely initial time sequence data; judging whether the data length of the initial time sequence data is smaller than the length of the dense sequence data, if so, supplementing the length of the initial time sequence data according to the length of the dense sequence data. And detecting and complementing the missing time stamp and the data value of the initial time sequence data, and normalizing the initial time sequence data after complementing the data to obtain the time sequence data to be detected.

Through the steps, the data preprocessing operation is carried out on the initial time sequence data to be detected, the time sequence data to be detected is obtained, the consistency of the time sequence data to be detected and the training time sequence data is further achieved, and the efficiency and the accuracy of the subsequent processing of the time sequence data to be detected through the target deep learning model are improved.

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

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

Specifically, after loading the time sequence data to be detected into the target deep learning model, the processor maps the time sequence data to be detected into a potential space through a generating module in the target deep learning model, specifically a coder in the generating module, and obtains a coding vector corresponding to the time sequence data to be detected. Further, the time sequence data to be detected is mapped to the effectiveness of the potential space by a judging module in the target deep learning model, particularly a measuring judging device in the judging module through a measuring encoder.

In step S222, when the encoded vector is within the preset threshold range, the encoded vector is processed, and a reconstruction sequence is generated.

Specifically, the processor judges whether the coded vector is within a preset threshold range, if so, the coded vector is converted to a reconstructed time sequence, namely a reconstructed sequence, through a generating module in the target depth learning model, specifically a decoder in the generating module.

In step S223, a difference between the data of the reconstructed sequence and the time sequence data to be detected is determined as a reconstruction error.

The processor distinguishes the data of the reconstruction sequence and the time sequence data to be detected through a distinguishing module in the target deep learning model, in particular a true and false distinguishing device in the distinguishing module; and then calculating the difference between the data of the reconstruction sequence at each time point and the time sequence data to be detected, namely, the reconstruction error. Specifically, the difference value between the data of the reconstructed sequence at each time point and the time sequence data to be detected is calculated through an L2 regularization method, so that the problem that the target deep learning model is excessively sensitive to noise or abnormal scores in training data is solved, and the generalization capability of the target deep learning model is improved.

Through the steps, the time sequence data to be detected is mapped, reconstructed and judged through the generation module and the judgment module in the target deep learning model, and therefore accuracy and efficiency of determining reconstruction errors are improved.

In some of these embodiments, determining an anomaly sequence from the reconstruction error in step S230 includes:

step S231, carrying out normalization processing on the reconstruction errors corresponding to a plurality of time points in the time sequence data to be detected, and obtaining error scores of the time points.

And after the reconstruction errors corresponding to the time points are determined, carrying out normalization processing on the reconstruction errors to obtain error scores corresponding to the reconstruction errors.

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

The processor presets an abnormal value, and further determines abnormal value data in a plurality of error values according to the abnormal value. Illustratively, the outliers in the anomaly score array are in the range of 0to 1.

Through the steps, the reconstruction errors are normalized, the error scores corresponding to the reconstruction errors are determined, the abnormal score arrays are further determined according to the error scores, the sequences corresponding to the abnormal score arrays are determined to be abnormal sequences, and therefore accuracy of determining the abnormal sequences is improved.

In some embodiments, in step S230, the evaluation of the anomaly score in the anomaly sequence according to the sliding window with the preset sequence length, to determine the anomaly data of the target wafer device, includes:

Step S233, dividing the abnormal sequence according to a sliding window with a preset sequence length to obtain a plurality of subsequences.

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

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

The processor takes the abnormal score points exceeding the preset statistical points in the subsequence as abnormal points of the wafer equipment. Further, the preset statistical point is determined according to a quartile statistical method, a quartile statistical value is calculated in the subsequence, and the quartile statistical value is the preset statistical point. And judging abnormal scores exceeding the quartile statistics in the subsequences, wherein the abnormal scores correspond to abnormal points of the wafer equipment. After determining the abnormal point of the wafer equipment, setting a first label, namely an abnormal label, for the abnormal point of the wafer equipment; and meanwhile, setting a second label, namely a normal label, at a normal point of the wafer except for an abnormal point of the wafer equipment in the subsequence. Illustratively, the first label of the wafer apparatus outlier is 1 and the second label of the wafer outlier is 0.

In step S235, the data of the abnormal point of the wafer apparatus corresponding to the first tag is determined as the abnormal data of the target wafer apparatus.

Through the steps, the abnormal sequence is subjected to sliding division through the sliding window to obtain a plurality of subsequences, and abnormal data of the target wafer equipment in the subsequences are determined through a quartile statistical method. The abnormal data of the target wafer equipment are evaluated in a sliding window mode, so that the recall rate of the abnormal data is increased, and the accuracy of abnormal detection is further improved.

In some embodiments, in step S230, the method for determining the anomaly data of the target wafer device further includes:

Step S236, dividing 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 respectively adjacent subsequences.

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 anomaly score and the second anomaly score include anomaly scores in the anomaly sequence that have a top ranking of scores.

The processor determines maximum anomaly scores in the first wafer sequence and the second wafer sequence respectively, wherein the maximum anomaly scores are the first anomaly scores and the second anomaly scores respectively. And performing differential operation on the first abnormal score and the second abnormal score to obtain a differential value between the first abnormal score and the second abnormal score, and dividing the differential value by the value of the first abnormal score to obtain the target sequence abnormal score.

In step S238, when the anomaly score of the target sequence does not exceed the preset score threshold, a second label is set for the anomaly score in the first wafer sequence. And determining the abnormality scores except the second label as the abnormality data of the target wafer equipment.

The processor judges whether the abnormal score of the target sequence exceeds a preset score threshold, and when the abnormal score of the target sequence does not exceed the preset score threshold, the processor can understand that the abnormal score in the first wafer sequence is a normal score, namely, a second label is set for the abnormal score in the first wafer sequence. The processor updates the abnormal scores in the abnormal sequence again, and determines the abnormal scores except the second label as the abnormal data of the target wafer equipment.

Through the steps, the recall rate of the abnormal data can be increased by the sliding window, and the false detection rate of the abnormal data can be increased. Therefore, the recall judging method is adopted, namely, labels of abnormal scores in a plurality of subsequences are updated through the difference operation, so that the false detection rate of abnormal detection is reduced, and the accuracy of the abnormal detection is improved.

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

Fig. 3 is a flowchart of an abnormality detection method based on wafer manufacturing according to the present embodiment, and as shown in fig. 3, the abnormality detection method based on wafer manufacturing includes the following steps:

step S310, preprocessing training data and training a deep learning model.

Specifically, the processor collects and loads the latest pieces of time sequence data of the sensor a in the wafer manufacturing process, where the pieces of time sequence data are the pieces of training time sequence data in the foregoing embodiments. Then, detecting and complementing the missing time stamps and the data values of the plurality of pieces of time sequence data, performing uniform length cutting operation on the plurality of pieces of time sequence data, and storing the historical data sequence length, wherein the historical data sequence length is the dense sequence data in the embodiment. The plurality of time sequence data are normalized, and then a historical dense data average value sequence is calculated and stored. The loading is directed to a deep learning model of a sensor waveform careful construction in the semiconductor field, and the deep learning model comprises two generators and two discriminators. The two generators can be regarded as an encoder and a decoder, respectively. And inputting the collected dense sequences into a model in batches, alternately optimizing by using a generator and a discriminator, and storing the trained optimal model parameters.

Preferably, the operation of detecting and complementing the missing time and data of the plurality of pieces of time series data may include: and calculating the time stamp interval from front to back for one piece of time sequence data, acquiring the mode of the time stamp interval, and taking the mode as a data acquisition period. And dividing the data acquisition period by the interval between the previous time stamp and the next time stamp in the time sequence data, and if the obtained dividing result is equal to n, supplementing n-1 forward values in the time sequence data corresponding to the previous time stamp and the next time stamp, wherein n is a positive integer greater than or equal to 1.

Preferably, in order to keep the length of each piece of time series data uniform, therefore, it is necessary to acquire one piece of time series data having the shortest data length among the pieces of time series data; and cutting off the plurality of pieces of time sequence data according to the shortest data length to obtain time sequence data with uniform length, and storing the time sequence data as the historical data sequence length.

Preferably, a 3sigma statistical method is used to process the pieces of time series data after the operations of complement, truncation and the like are performed, and a time point corresponding to the time series data outside 3sigma is judged as an outlier. The same time point of the plurality of pieces of time series data is input into 3sigma, as long as one time point is out of 3sigma, the time point is determined to be an outlier, and the whole time series including the outlier is determined to be an outlier. And screening the outlier sequences from the plurality of pieces of time sequence data, and collecting to obtain dense sequences. And carrying out transverse normalization on the plurality of dense sequences and calculating the average value of each time stamp to obtain a historical dense data average value sequence.

Step S320, preprocessing the time series data to be detected.

Specifically, the processor automatically collects sensor parameters of the sensor a on the device at a preset specific frequency to obtain on-line real-time window data, where the window data is the time sequence data to be detected in the foregoing embodiment. When the window data length is smaller than the historical data sequence length, the window data is supplemented according to the historical dense data average value sequence, and then the window data is preprocessed through operations such as cutting off the time sequence data to be trained in the step S320, so that the time sequence data to be detected is obtained.

And step S330, performing anomaly detection and identification on the time sequence data to be detected according to the trained deep learning model.

Specifically, when preprocessing on-line real-time window data to obtain time sequence data to be detected, loading the time sequence data to be detected into a trained deep learning model, namely the target deep learning model in the embodiment. When the abnormality in the time series data to be detected is evaluated through the target deep learning model, firstly, a reconstruction sequence of the time series data to be detected is obtained through two generators in the deep learning model, namely an encoder and a decoder, and then, the abnormality identification is carried out by calculating a reconstruction error between the reconstruction sequence and an input sequence. The generator is the generating module in the foregoing embodiment. Illustratively, the error at each point in time between the reconstructed sequence and the input sequence is calculated by the method of L2 regularization. And calculating a normalized score of the reconstruction error, and taking the abnormal score at the time point corresponding to the normalized score between 0 and 1 as a final abnormal score array.

And then, introducing a sliding window to perform data segmentation operation on the abnormal score array of the online data, and dividing the array into a plurality of signal segments, so as to obtain a plurality of pieces of window data with fixed seconds, wherein the plurality of pieces of window data are subsequences in the embodiment. A quartile statistic is calculated within the plurality of pieces of window data, and a point exceeding the quartile statistic is set as an outlier. Meanwhile, the output abnormal score array is converted into an abnormal label array. Wherein the label of the abnormal point is 1, namely the first label in the previous embodiment; the normal point label is 0, which is the second label in the previous embodiment. The abnormal score is evaluated by adopting a sliding window mode, so that the recall rate of the abnormality is increased.

Further, for the conversion from anomaly scores to anomaly labels, the complete time sequence is divided into a plurality of sub-sequences of equal length using a sliding window, each sub-sequence is used as a sample, and whether each sub-sequence contains an anomaly point is detected to increase the recall rate for anomalies. Specifically, points within a subsequence where the anomaly score exceeds four digits are set as anomalies. Illustratively, the window sequence length is set to one third of the complete sequence and the sliding step size is set to one thirty-half of the complete sequence. Calculating the lower quartile value Q1 and the upper quartile value Q3 in each sliding window of the input sequence, adding more than the upper quartile in the window dataAll time points of the times (Q3-Q1) are judged to be abnormal, and an abnormal array 1 is generated as follows:

Wherein, Representing an anomaly array 1, Q1 representing a lower quartile value, and Q3 representing an upper quartile value; t represents a time stamp; And determining according to the yield of the time sequence data.

Subtracting less than the lower quartile from the on-line window dataAll time points of the times (Q3-Q1) are judged to be abnormal, and an abnormality array 2 is generated as follows:

Wherein, Representing an anomaly array 2, Q1 representing a lower quartile value, and Q3 representing an upper quartile value; t represents a time stamp; And determining according to the yield of the time sequence data. And marking the abnormal point and the non-abnormal point in the abnormal array, wherein the abnormal point is marked as 1, and the non-abnormal point is marked as 0. Further, two exception arrays have one point that is an exception, even if it is an exception point.

Meanwhile, the sliding window increases the recall rate for anomalies, and also increases the false detection rate. Therefore, further, the post-processing method for criticizing recall is adopted to reduce the false detection rate for abnormal detection. Specifically, the maximum anomaly score in the front and rear sliding windows is divided, where the front and rear sliding windows are the first wafer sequence and the second wafer sequence in the foregoing embodiments, respectively. And if the obtained difference division operation result is smaller than the set threshold value, reassigning all abnormal point labels in the front window, namely the first wafer sequence, to normal second labels. Outputting the final post-processed two-dimensional abnormal array, marking abnormal points in a time sequence window of the red visual page in real time, observing by a chip engineer, and monitoring the change of data in the window in real time.

Further, for the criticizing recall method, the abnormal points are processed to reduce the false detection rate, so that the abnormal recognition accuracy is improved. Specifically, the maximum anomaly score of each sliding window sequence is first saved, and for each anomaly sequence, we use the maximum anomaly score to represent it, i.e. the maximum anomaly score of the preceding and following intervals is subjected to a difference division operation to obtain M, as follows:

wherein k represents the kth abnormal sequence, Representing the maximum anomaly score for the previous sliding window sequence; representing the maximum anomaly score for the latter sliding window sequence, Indicating the difference division result. And if M does not exceed the preset abnormal threshold value theta, all the point labels of the front interval are assigned to 0. And the abnormal points are identified again, so that false alarms are reduced.

In one embodiment, referring to fig. 4, fig. 4 is a block diagram of the deep learning model provided in this embodiment. The whole model mainly comprises a data preparation component 41, a model training component 42 and an anomaly output component 43.

For the data preparation component 41, a sample preprocessing module is mainly included. And carrying out pretreatment operation on the time series of the batch sensors through data complementation, data truncation and normalization, and then sending the pretreated time series into a model for training.

For the model training component 42, a real-time anomaly detection model based on two-way limiting countermeasure learning, the deep learning model in the foregoing embodiment, is constructed for the semiconductor field sensor waveform. The model comprises two generators and two discriminator modules, each module is formed by taking two to three layers of full connection as a basic component. The generator and the discriminator are respectively embedded into a framework of countermeasure learning, and training time sequence data is input to train the model. After model training, four models are generated: encoder, decoder, true and false discriminant and measurement discriminant. Wherein the encoder and decoder correspond to two generators. The encoder maps the temporal sequence to the potential space and the decoder converts the potential space to a reconstructed temporal sequence. The true-false discriminator is used to discriminate whether the input real time sequence or the decoder generated false time sequence, and the measurement discriminator is used to measure the effectiveness of the encoder in mapping the input sequence to the potential space.

Preferably, a batch of up-to-date data is pulled at intervals, trained using the deep learning model set forth above. Inputting the collected dense time sequence into a model, alternately optimizing by using a generator and a discriminator, and storing optimal model parameters. For each training period, firstly determining the iteration times of training two discriminators, wherein the iteration times of the discriminators can be q times, and q is a positive integer greater than 1; the generator is then trained iteratively. Thereafter, loss calculation and back propagation are performed for the two discriminators and the generator, respectively. Training is carried out by adopting a learning rate attenuation strategy and an early stopping strategy. Illustratively, for the learning rate decay strategy, the learning rate is reduced if there are 10 consecutive time series data epoch for which the training loss is not reduced. For the early-stop strategy, if there is still no decrease in training loss for 10 epochs in succession after the learning rate decays, training is stopped and the optimal model parameters are saved. The trained deep learning model is optimized and parameter-adjusted, and the optimization is specifically based on abnormal data: and removing noise data with poor characterization effect, and ensuring data quality. The model is retrained after the abnormal data is cleaned to eliminate the influence of the abnormality.

For the abnormal output component 43, three stages of abnormal score detection, label conversion and criticizing recall are mainly included. In the anomaly score detection stage, only two generators, namely an encoder and a decoder, are used for reconstruction to obtain a reconstructed sequence, and the anomaly is detected by calculating the deviation between the reconstructed sequence and an input sequence. Specifically, the reconstruction error is evaluated using the L2 norm loss.

Preferably, the real-time anomaly detection method is based on two-way limiting countermeasure learning, and the reconstructed sequence is constructed based on a training sequence. When a new sequence comes in, detecting the difference between the reconstructed sequence constructed before and the new sequence, calculating the abnormal score of each time point, and screening out the abnormality according to a preset abnormal score threshold. By using the deep learning algorithm to monitor the change of the sensor parameters in real time, abnormal points in time can be rapidly identified, the Accuracy Accuracy reaches 0.99+, and the Accuracy/Recall/F1 value (Precision/Recall/F1-Score) reaches 0.8+ of identification Precision.

According to the application, through collecting a large amount of sensor data in the chip production process, targeted data cleaning and data preprocessing are performed, and the steps of complement, truncation and normalization are included, so that the quality of input data is improved. Meanwhile, according to the specific requirements of wafer manufacturing in the semiconductor field, an abnormality detection algorithm with strong adaptability is custom designed. The deep learning algorithm is based on bidirectional limiting countermeasure learning, can capture and identify different types of abnormal conditions, and can improve the accuracy and efficiency of abnormality detection. And optimizing and adjusting parameters of the trained deep learning model to improve accuracy and robustness of the deep learning model. Illustratively, optimizing and tuning the model may be by way of an optimizer, a learning rate decay strategy, a training early-stop strategy, and the like. And monitoring the data flow in the actual chip production by the trained model in real time, and feeding back the detection result to an operator or an automation system. This helps to find anomalies in time and take corresponding action to improve yield.

It should be noted that the steps illustrated in the above-described flow or in the flow diagrams of the figures may be performed in a computer system, such as a set of computer-executable instructions.

In this embodiment, an abnormality detection device based on wafer manufacturing is further provided, and the abnormality detection device is used to implement the foregoing embodiments and preferred embodiments, and is not described again. The terms "module," "unit," "sub-unit," and the like as used below may refer to a combination of software and/or hardware that performs a predetermined function. While the means described in the following embodiments are preferably implemented in software, implementations in hardware, or a combination of software and hardware, are also possible and contemplated.

Fig. 5 is a block diagram of the abnormality detection apparatus based on wafer manufacturing of the present embodiment, and as shown in fig. 5, the apparatus includes: a data acquisition module 10, an error determination module 20, a data evaluation module 30.

A data acquisition module 10, configured to acquire time-series data to be detected; the timing data to be detected includes sensor data of the wafer apparatus.

The error determining module 20 is configured to determine a reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model; the target deep learning model is obtained through training according to a preset countermeasure learning method.

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

The above-described respective modules may be functional modules or program modules, and may be implemented by software or hardware. For modules implemented in hardware, the various modules described above may be located in the same processor; or the above modules may be located in different processors in any combination.

There is also provided in this embodiment an electronic device comprising a memory having stored therein a computer program and a processor arranged to run the computer program to perform the steps of any of the method embodiments described above.

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

Alternatively, in the present embodiment, the above-described processor may be configured to execute the following steps by a computer program:

s1, acquiring time sequence data to be detected; the timing data to be detected includes sensor data of the wafer apparatus.

S2, determining a reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model; the target deep learning model is obtained through training according to a preset countermeasure learning method.

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

It should be noted that, specific examples in this embodiment may refer to examples described in the foregoing embodiments and alternative implementations, and are not described in detail in this embodiment.

In addition, in combination with the abnormality detection method based on wafer manufacturing provided in the above embodiment, a storage medium may be provided in the present embodiment. The storage medium has a computer program stored thereon; the computer program, when executed by a processor, implements any of the wafer-manufacturing-based anomaly detection methods of the above embodiments.

It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to be limiting. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure in accordance with the embodiments provided herein.

It is to be understood that the drawings are merely illustrative of some embodiments of the present application and that it is possible for those skilled in the art to adapt the present application to other similar situations without the need for inventive work. In addition, it should be appreciated that while the development effort might be complex and lengthy, it would nevertheless be a routine undertaking of design, fabrication, or manufacture for those of ordinary skill having the benefit of this disclosure, and thus should not be construed as a departure from the disclosure.

The term "embodiment" in this disclosure means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearances of such phrases in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive. It will be clear or implicitly understood by those of ordinary skill in the art that the embodiments described in the present application can be combined with other embodiments without conflict.

The above examples merely represent a few embodiments of the present application, which are described in more detail and are not to be construed as limiting the scope of the patent claims. It should be noted that it will be apparent to those skilled in the art that several variations and modifications can be made without departing from the spirit of the application, which are all within the scope of the application. Accordingly, the scope of the application should be assessed as that of the appended claims.

Claims

1. An anomaly detection method based on wafer manufacturing, the method comprising:

2. The wafer-based anomaly detection method of claim 1, wherein prior to the acquiring the timing data to be detected, comprising:

3. The wafer manufacturing-based anomaly detection method according to claim 2, wherein before determining the reconstruction error corresponding to the time series data to be detected according to a pre-trained target deep learning model, the method comprises:

4. The wafer-based manufacturing anomaly detection method of claim 1, wherein,

The determining the reconstruction error corresponding to the time sequence data to be detected according to a pre-trained target deep learning model comprises the following steps:

5. The wafer-based manufacturing anomaly detection method of claim 1, wherein the determining an anomaly sequence from the reconstruction error comprises:

6. The wafer fabrication-based anomaly detection method of claim 1, wherein the evaluating anomaly scores in the anomaly sequence according to a sliding window of a preset sequence length to determine target wafer device anomaly data comprises:

7. The wafer fabrication-based anomaly detection method of claim 6, wherein the evaluating anomaly scores in the anomaly sequence according to a sliding window of a preset sequence length to determine target wafer device anomaly data further comprises:

8. An anomaly detection device based on wafer manufacturing, the device comprising: the system comprises a data acquisition module, an error determination module and a data evaluation module;

9. An electronic device comprising a memory and a processor, wherein the memory has stored therein a computer program, the processor being arranged to run the computer program to perform the wafer-based manufacturing anomaly detection method of any one of claims 1 to 7.

10. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the wafer-manufacturing-based anomaly detection method of any one of claims 1 to 7.