A method for generating knowledge points based on univariate time series data
By generating knowledge points directly from FDC univariate time-series data, and utilizing the variational autoencoder (VAE) and gated recurrent unit (GRU) model, the inefficiency of knowledge graph generation in semiconductor manufacturing is solved, achieving more efficient knowledge utilization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN ZHIXIAN FUTURE IND SOFTWARE CO LTD
- Filing Date
- 2023-02-03
- Publication Date
- 2026-05-15
AI Technical Summary
In the semiconductor manufacturing process, existing technologies struggle to efficiently utilize large amounts of non-textual data, such as univariate time-series data processed by FDC, resulting in a cumbersome and inefficient knowledge graph generation process.
Instead of generating event description text, knowledge points are generated directly from FDC univariate time series data. Using variational autoencoder (VAE) and gated recurrent unit (GRU) models, time information is extracted, prediction data is generated, and anomaly types are determined to form knowledge points for generating knowledge graphs.
It simplifies the knowledge graph generation process in semiconductor manufacturing, improving data processing efficiency and knowledge utilization.
Smart Images
Figure CN116451786B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor manufacturing, and in particular to a method for generating knowledge points based on univariate time-series data. Background Technology
[0002] The semiconductor manufacturing process generates various types of data, which can be broadly categorized into two types based on their source: one type comes from various equipment, such as equipment operating status, equipment parameters, equipment manuals, and operating logs; the other type comes from the testing data of the produced chips, such as the chip's electrical characteristics and defects. In current production processes, often only a small portion of this data is analyzed to guide production, and the majority of the data remains underutilized.
[0003] In knowledge graph generation, the conventional approach involves extracting nodes and causal / sequential relationships between them from open-domain text documents using Natural Language Processing (NLP) to construct the knowledge graph. However, semiconductor manufacturing generates a large amount of non-textual data, such as univariate time-series data processed by a fault detection and classification system (FDC). Using conventional knowledge graph generation methods requires first generating event description text using a language generation model, and then extracting knowledge from this text to construct the knowledge graph. This data processing involves language generation and deconstruction, which is cumbersome and inefficient. Summary of the Invention
[0004] This specification describes one or more embodiments of a method for generating knowledge points based on univariate time-series data. Instead of generating event description text from FDC univariate time-series data, it directly generates knowledge from the FDC univariate time-series data. This knowledge can then be used to generate a knowledge graph, thereby simplifying the knowledge graph generation process throughout semiconductor manufacturing. After generating the knowledge graph, it can guide the entire semiconductor production process.
[0005] Firstly, a method for generating knowledge points based on univariate time series data is provided, including:
[0006] Acquire univariate time-series data, wherein the time-series data is data generated by sensors of equipment during semiconductor manufacturing;
[0007] The target subsequence with the earliest time sequence in the time series data is input into the prediction model to obtain the prediction data;
[0008] Based on the predicted data and the measured data of the target subsequence in the time series data, the anomaly type related to the device sensor is determined;
[0009] Based on the anomaly type, the category of the time-series data, the device number of the device, and the number of the wafer being processed by the device, corresponding knowledge points are determined. These knowledge points are used to generate or update a knowledge graph in the semiconductor field.
[0010] In one possible implementation, the timing data is data generated by sensors in the semiconductor manufacturing process and then processed by a fault detection and classification system (FDC).
[0011] In one possible implementation, inputting the target subsequence that appears first in the time series data into the prediction model to obtain prediction data includes:
[0012] The target subsequence is divided into several segments of the same size;
[0013] The aforementioned segments are sequentially encoded using the encoder of a variational autoencoder (VAE) in chronological order to obtain a sequence of latent space variables.
[0014] The temporal information of the latent space variable sequence is extracted using a gated recurrent unit (GRU) model.
[0015] The latent space variable sequence and the time information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain the prediction data.
[0016] In one possible implementation, the types of data generated by the sensor include at least: temperature, humidity, voltage, current, and pressure.
[0017] In one possible implementation, the abnormality type includes at least: too low, too high, continuously decreasing, continuously increasing, and fluctuating.
[0018] In one possible implementation, determining the anomaly type related to the device sensor based on the predicted data and the measured data following the target subsequence in the time-series data includes:
[0019] Based on the predicted data and the measured data, the probability value corresponding to the anomaly type is determined. When the probability value of any anomaly type is greater than a preset first threshold, the anomaly type is determined to be an anomaly type related to the device sensor.
[0020] In one possible implementation, the prediction model is trained based on a training set and a test set, both of which contain the univariate time series data, and the time points of the data in the test set are after the data in the training set.
[0021] In one possible implementation, the prediction model is trained through the following process:
[0022] The data sequence in the training set is divided into several segments of the same size;
[0023] The aforementioned segments are sequentially encoded using the encoder of a variational autoencoder (VAE) in chronological order to obtain a sequence of latent space variables.
[0024] The temporal information of the latent space variable sequence is extracted using a gated recurrent unit (GRU) model.
[0025] The latent space variable sequence and the time information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain the prediction data;
[0026] The values of the parameters in the variational autoencoder (VAE) and the gated recurrent unit (GRU) model are adjusted by minimizing the error between the predicted data and the data in the test set.
[0027] In one possible implementation, the types of error include at least: root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE).
[0028] In one possible implementation, after determining the corresponding knowledge point based on the anomaly type, the category of the time-series data, the device number, and the number of the wafer being processed by the device, the method further includes:
[0029] The knowledge points are added to a temporary knowledge base, and the knowledge points in the temporary knowledge base are deduplicated.
[0030] The knowledge graph is generated or updated using the knowledge points in the temporary knowledge base after deduplication.
[0031] Secondly, a device for generating knowledge points based on univariate time-series data is provided, including:
[0032] The data acquisition unit is configured to acquire univariate time-series data, wherein the time-series data is data generated by sensors of equipment during semiconductor manufacturing; the data prediction unit is configured to input the target subsequence with the earliest time sequence in the time-series data into a prediction model to obtain predicted data; the anomaly detection unit is configured to determine the anomaly type related to the equipment sensors based on the predicted data and the measured data following the target subsequence in the time-series data; and the knowledge point generation unit is configured to determine the corresponding knowledge point based on the anomaly type, the category of the time-series data, the equipment number of the equipment, and the number of the wafer being processed by the equipment, wherein the knowledge point is used to generate or update a knowledge graph in the semiconductor field.
[0033] In one possible implementation, the device further includes:
[0034] The deduplication unit is configured to add the knowledge points to a temporary knowledge base and perform deduplication processing on the knowledge points in the temporary knowledge base; the knowledge graph generation unit is configured to generate or update the knowledge graph using the deduplicated knowledge points in the temporary knowledge base.
[0035] This invention proposes a method for generating knowledge points based on univariate time-series data. Using an end-to-end approach, it does not generate event description text from FDC univariate time-series data, but directly generates knowledge from FDC univariate time-series data. Then, a knowledge graph can be generated based on the knowledge, thereby simplifying the process of generating knowledge and knowledge graphs throughout the semiconductor manufacturing process. Attached Figure Description
[0036] To more clearly illustrate the technical solutions of the various embodiments disclosed in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only a few embodiments disclosed in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0037] Figure 1 This is a framework diagram of a method for generating knowledge points based on univariate time series data disclosed in an embodiment of the present invention;
[0038] Figure 2 This is a flowchart of a method for generating knowledge points based on univariate time series data, as disclosed in an embodiment of the present invention.
[0039] Figure 3 This is a flowchart of the method for predicting data disclosed in an embodiment of the present invention;
[0040] Figure 4 This is a schematic block diagram of an apparatus for generating knowledge points based on univariate time series data, as disclosed in an embodiment of the present invention. Detailed Implementation
[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0042] Fault Detection and Classification (FDC) systems primarily collect, integrate, and monitor semiconductor production-related data to achieve real-time analysis and effectively capture production anomalies.
[0043] A semiconductor manufacturing machine is equipped with numerous sensors, each monitoring a specific parameter such as temperature, humidity, voltage, current, or pressure. Each sensor differs in its data collection accuracy and acquisition method. After the sensors upload the raw production data to relevant data processing equipment, the FDC system installed on this equipment processes the various types of raw data, generating several types of univariate time-series data. This univariate time-series data can be used for status monitoring of semiconductor production.
[0044] Because semiconductor manufacturing involves a large amount of non-textual data, such as univariate time-series data processed by the Fault Detection and Classification (FDC) system, conventional knowledge graph generation methods would require first generating event description text from the data using a language generation model, and then extracting knowledge from the event description text to construct the knowledge graph. This data processing involves language generation and deconstruction, which is cumbersome and inefficient.
[0045] Based on this, one or more embodiments of this specification describe a method for generating knowledge points based on univariate time-series data. Instead of generating event description text from the univariate time-series data, this method directly generates knowledge points from the univariate time-series data, and then a knowledge graph can be generated based on these knowledge points, thereby simplifying the process of generating a knowledge graph throughout the semiconductor manufacturing process. According to one embodiment, Figure 1 This paper presents a framework for a method to generate knowledge points based on univariate time-series data. First, the input univariate time-series data is encoded using an encoder to obtain a latent space vector sequence. This sequence is then input into a gated recurrent unit (GRU) model to extract the time information of the latent space variable sequence. Next, the latent space variable sequence and the time information are input into a decoder for decoding to obtain predicted data. Finally, based on the difference between the predicted data and the corresponding measured data in the univariate time-series data, it is determined whether there are any anomalies in the production process and their corresponding types. Based on the anomaly type, the equipment number of the production equipment, and the number of the wafer being processed by the production equipment, knowledge points are determined.
[0046] The following will provide further explanation and description with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.
[0047] Figure 2 This is a flowchart illustrating a method for generating knowledge points based on univariate time-series data, as disclosed in an embodiment of the present invention. Figure 2 As shown, the method includes at least the following steps: Step 201, acquiring univariate time-series data, wherein the time-series data is data generated by sensors of equipment during semiconductor manufacturing; Step 202, inputting the target subsequence with the earliest time sequence in the time-series data into a prediction model to obtain prediction data; Step 203, determining the anomaly type related to the equipment sensors based on the prediction data and the measured data following the target subsequence in the time-series data; Step 204, determining the corresponding knowledge point based on the anomaly type, the category of the time-series data, the equipment number of the equipment, and the number of the wafer being processed by the equipment, wherein the knowledge point is used to generate or update a knowledge graph in the semiconductor field.
[0048] In step 201, univariate time-series data is obtained, wherein the time-series data is data generated by sensors of equipment during semiconductor manufacturing.
[0049] Specifically, the time-series data refers to the raw data generated by sensors in the semiconductor manufacturing process, which is then processed by the FDC system. As mentioned earlier, a semiconductor production machine is equipped with many sensors, each monitoring a parameter on the machine, such as temperature, humidity, voltage, current, pressure, etc. The value of any parameter output by the sensor is processed by the FDC system to obtain a set of univariate time-series data.
[0050] In step 202, the target subsequence with the earliest time sequence in the time series data is input into the prediction model to obtain the prediction data.
[0051] The prediction data is a prediction of the data that follows the target subsequence in the time series data.
[0052] In a specific embodiment, for a time series data (x1, x2, ..., x...) 100 Take the target subsequence (x1, x2, ..., x) that is sorted first. 40 The data x is input into the prediction model to obtain the results. 41 Predicted data x ′ 41 .
[0053] In another specific embodiment, for a time series data (x1, x2, ..., x...) 100 Take the target subsequence (x1, x2, ..., x) that is sorted first. 40 The data sequence (x) is input into the prediction model to obtain the prediction result. 41 ,x 42 ,…,x 50 Predicted data (x) ′ 41 ,x ′42 ,…x ′ 50 ).
[0054] In one embodiment, the method for determining the forecast data is as follows: Figure 3 As shown.
[0055] In step 301, the target subsequence is divided into several segments of the same size. For example, the target subsequence (x1, x2, ..., x...) is divided into several segments of the same size. 40 Divide into (x1, x2, ..., x) 10 ), (x 11 ,x 12 ,…,x 20 ), (x 21 ,x 22 ,…,x 30 ) and (x 31 ,x 32 ,…,x 40 ).
[0056] In step 302, the segments are encoded sequentially in chronological order using the encoder of a variational autoencoder (VAE) to obtain a sequence of latent space variables.
[0057] As will be understood by those skilled in the art, VAE is a probabilistic model based on variational inference, belonging to generative models. It consists of two network parts: one part is called the encoder, which maps a high-dimensional input to a low-dimensional latent space variable, and the other part is called the decoder, which maps the low-dimensional latent variable back to the high-dimensional output.
[0058] In step 303, the temporal information of the latent space variable sequence is extracted using the Gated Recurrent Unit (GRU) model.
[0059] As will be understood by those skilled in the art, GRU is a model based on the Long Short-Term Memory (LSTM) network, which combines the input gate and forget gate of LSTM into an update gate and adds a reset gate.
[0060] In step 304, the latent space variable sequence and the time information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain the prediction data.
[0061] Then return Figure 2In step 203, based on the predicted data and the measured data of the target subsequence in the time series data, the anomaly type related to the device sensor is determined.
[0062] The abnormal types include at least: too low, too high, continuously decreasing, continuously increasing, and fluctuating.
[0063] Specifically, based on the predicted data and the measured data, the probability value corresponding to the anomaly type is determined. When the probability value of any anomaly type is greater than a preset first threshold, the anomaly type is determined to be an anomaly type related to the device sensor.
[0064] Determining the probability value corresponding to an anomaly type can be achieved in various ways, and no specific method is specified here. For example, a softmax function can be used after the decoder in a variational autoencoder (VAE) to output the probability values corresponding to various anomaly types.
[0065] In one embodiment, based on a single measured data point and a single predicted data point, the probability values corresponding to the abnormality types of "too low" and "too high" are determined respectively. When the probability value of any abnormality type is greater than a preset first threshold, the abnormality type is determined to be an abnormality type related to the device sensor.
[0066] In another embodiment, based on multiple consecutive measured data and corresponding multiple predicted data, the probability values corresponding to the anomaly types of too low, too high, continuous decrease, continuous increase, and fluctuation are determined respectively. When the probability value of any anomaly type is greater than a preset first threshold, the anomaly type is determined to be an anomaly type related to the device sensor.
[0067] The aforementioned first threshold can be set by engineers based on actual circumstances or experience, or it can be calculated by some data models based on historical data. Different thresholds can also be set for different anomaly types based on actual circumstances.
[0068] In step 204, based on the anomaly type, the category of the time-series data, the device number of the device, and the number of the wafer being processed by the device, corresponding knowledge points are determined. These knowledge points are used to generate or update a knowledge graph in the semiconductor field.
[0069] In some embodiments, the knowledge point can be in the form of a triple, specifically (anomaly type, device number, wafer number). For example, a specific knowledge point could be (overheating, machine 2, wafer 3), where wafer represents a wafer.
[0070] In some embodiments, the prediction model is trained based on a training set and a test set, both of which contain the univariate time series data, and the time points of the data in the test set are after the data in the training set.
[0071] In some more specific embodiments, the prediction model is trained through the following process:
[0072] The data sequence in the training set is divided into several segments of the same size; these segments are then encoded sequentially using the encoder of a Variational Autoencoder (VAE) in chronological order to obtain a latent space variable sequence; the temporal information of the latent space variable sequence is extracted using a Gated Recurrent Unit (GRU) model; the latent space variable sequence and the temporal information are input into the decoder of the VAE for decoding to obtain predicted data; and the values of the parameters in the VAE and the GRU model are adjusted by minimizing the error between the predicted data and the data in the test set.
[0073] The types of errors include at least: root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE).
[0074] In some embodiments, after step 204, the method further includes: step 205, adding the knowledge points to a temporary knowledge base and performing deduplication processing on the knowledge points in the temporary knowledge base; step 206, using the deduplicated knowledge points in the temporary knowledge base to generate or update the knowledge graph.
[0075] Figure 4 This is a schematic block diagram of an apparatus for generating knowledge points based on univariate time-series data, as disclosed in an embodiment of the present invention. The apparatus 400 includes:
[0076] The data acquisition unit 401 is configured to acquire univariate time-series data, wherein the time-series data is data generated by sensors of equipment during semiconductor manufacturing; the data prediction unit 402 is configured to input the target subsequence with the earliest time sequence in the time-series data into a prediction model to obtain prediction data; the anomaly detection unit 403 is configured to determine the anomaly type related to the equipment sensors based on the prediction data and the measured data following the target subsequence in the time-series data; the knowledge point generation unit 404 is configured to determine the corresponding knowledge point based on the anomaly type, the category of the time-series data, the equipment number of the equipment, and the number of the wafer being processed by the equipment, wherein the knowledge point is used to generate or update a knowledge graph in the semiconductor field.
[0077] In one embodiment, the apparatus further includes:
[0078] The deduplication unit 405 is configured to add the knowledge points to a temporary knowledge base and perform deduplication processing on the knowledge points in the temporary knowledge base; the graph generation unit 406 is configured to generate or update the knowledge graph using the knowledge points in the temporary knowledge base after deduplication processing.
[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.
[0080] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.
[0081] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for generating knowledge points based on univariate time series data, comprising: Acquire univariate time-series data, wherein the time-series data is data generated by sensors of equipment during semiconductor manufacturing; The target subsequence with the earliest time sequence in the time series data is input into the prediction model to obtain the prediction data; Based on the predicted data and the measured data of the target subsequence in the time series data, the anomaly type related to the device sensor is determined; wherein, the anomaly type includes at least: too low, too high, continuously decreasing, continuously increasing, and fluctuating; the determination of the anomaly type related to the device sensor includes: based on multiple consecutive measured data and corresponding multiple predicted data, determining the probability value corresponding to the anomaly type as too low, too high, continuously decreasing, continuously increasing, and fluctuating, respectively; when the probability value of any anomaly type is greater than a preset first threshold, the anomaly type is determined to be the anomaly type related to the device sensor; Based on the anomaly type, the category of the time-series data, the device number, and the number of the wafer being processed by the device, corresponding knowledge points in the form of triplets are determined. These knowledge points are used to generate or update a knowledge graph in the semiconductor field.
2. The method according to claim 1, characterized in that, The time-series data is data generated by sensors in the semiconductor manufacturing process and then processed by the Fault Detection and Classification (FDC) system.
3. The method according to claim 1, characterized in that, The step of inputting the target subsequence with the earliest time sequence from the time series data into the prediction model to obtain prediction data includes: The target subsequence is divided into several segments of the same size; Several segments are encoded sequentially using the encoder of a variational autoencoder (VAE) in chronological order to obtain a sequence of latent space variables. The temporal information of the latent space variable sequence is extracted using a gated recurrent unit (GRU) model. The latent space variable sequence and the time information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain the prediction data.
4. The method according to claim 1, characterized in that, The types of data generated by the sensor include at least: temperature, humidity, voltage, current, and pressure.
5. The method according to claim 1, characterized in that, The prediction model is trained based on a training set and a test set, both of which contain the univariate time series data, and the time points of the data in the test set are after the data in the training set.
6. The method according to claim 5, characterized in that, The prediction model is trained through the following process: The data sequence in the training set is divided into several segments of the same size; Several segments are encoded sequentially using the encoder of a variational autoencoder (VAE) in chronological order to obtain a sequence of latent space variables. The temporal information of the latent space variable sequence is extracted using a gated recurrent unit (GRU) model. The latent space variable sequence and the time information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain the prediction data; The values of the parameters in the variational autoencoder (VAE) and the gated recurrent unit (GRU) model are adjusted by minimizing the error between the predicted data and the data in the test set.
7. The method according to claim 6, characterized in that, The types of errors include at least: root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE).
8. The method according to claim 1, characterized in that, After determining the corresponding knowledge point based on the anomaly type, the category of the time-series data, the device number, and the number of the wafer being processed by the device, the method further includes: The knowledge points are added to a temporary knowledge base, and the knowledge points in the temporary knowledge base are deduplicated. The knowledge graph is generated or updated using the knowledge points in the temporary knowledge base after deduplication.