A method for determining anomalies based on graph attention mechanisms

Abstract

The application relates to a method for determining an anomaly based on a graph attention mechanism, comprising: acquiring multivariate time series data, wherein the time series data is data generated by a plurality of sensors of a device in a semiconductor manufacturing process; acquiring final embedding vectors corresponding to the plurality of sensors according to a graph structure learning model, wherein the final embedding vectors are used to represent the association relationship between the plurality of sensors; inputting a target subsequence in the time series data in a time sequence into a time series feature extraction model to obtain a time series feature vector; splicing the time series feature vector and the final embedding vector and inputting the spliced result into a prediction model to obtain predicted data corresponding to the plurality of sensors; and determining whether an anomaly occurs at a target time point corresponding to measured data according to the predicted data and the measured data following the target subsequence in the time series data.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing, and more particularly to a method for identifying anomalies based on a graph attention mechanism. Background Technology

[0002] In semiconductor manufacturing, multiple sensors typically monitor the production status simultaneously on the same machine. These sensors on the same machine are interconnected; when an anomaly occurs, some of these interconnected sensors may simultaneously generate abnormal data, thus triggering an alarm.

[0003] Existing technologies model these sensor data separately using univariate time-series prediction, predicting data individually for each sensor and then comparing it with the actual data to determine whether to trigger an alarm. However, this method may produce false alarms, meaning that a sensor may erroneously generate an alarm message due to its own malfunction or environmental noise when no actual anomaly has occurred. This creates redundancy in the entire monitoring system and increases the maintenance costs of production equipment. Summary of the Invention

[0004] This specification describes one or more embodiments of a method for anomaly identification based on a graph attention mechanism. This method simultaneously models data from multiple sensors using a multivariate temporal anomaly detection model, and compares all predicted data with the actual data to obtain better prediction results. Simultaneously, a temporal feature extraction model is used to explicitly extract temporal information from the multivariate temporal data, effectively utilizing this information to achieve better prediction accuracy.

[0005] This specification provides a method for anomaly detection based on a graph attention mechanism, including:

[0006] Acquire multivariate time-series data, which is data generated by multiple sensors of equipment during semiconductor manufacturing;

[0007] The final embedding vectors corresponding to the multiple sensors are obtained according to the graph structure learning model, and the final embedding vectors are used to represent the correlation between the multiple sensors.

[0008] The target subsequence with the earliest time sequence in the time series data is input into the time series feature extraction model to obtain the time series feature vector;

[0009] The temporal feature vector and the final embedded vector are concatenated and input into the prediction model to obtain the prediction data corresponding to the multiple sensors;

[0010] Based on the predicted data and the measured data of the target subsequence in the time series data, determine whether there is an anomaly at the target time point corresponding to the measured data.

[0011] In one possible implementation, it also includes:

[0012] When an anomaly occurs at the target time point, the abnormal sensor and the corresponding anomaly type are determined based on the predicted data and the measured data.

[0013] In one possible implementation, determining whether an anomaly occurs at the target time point corresponding to the measured data is based on the predicted data and the measured data of the target subsequence following the time series data includes:

[0014] The error is determined based on the predicted data and the measured data. Based on the comparison between the error and a preset first threshold, it is determined whether an anomaly occurs at the target time point corresponding to the measured data.

[0015] In one possible implementation, determining the abnormal sensor and its corresponding abnormality type based on the predicted data and the measured data includes:

[0016] Based on the data corresponding to any target sensor among the plurality of sensors in the predicted data and the measured data, an error is determined. Based on the comparison result of the error and the preset threshold corresponding to the target sensor, it is determined whether the target sensor has an anomaly and the corresponding anomaly type.

[0017] In one possible implementation, the graph structure learning model includes a graph embedding layer, a graph structure learning layer, and a graph relation aggregation layer.

[0018] In one possible implementation, the graph structure learning model, the temporal feature extraction model, and the prediction model are trained using sample time-series data, the training comprising:

[0019] The feature data of the multiple sensors are input into the graph embedding layer to obtain multiple initial embedding vectors corresponding to the multiple sensors;

[0020] The multiple initial embedding vectors are input into the graph structure learning layer to obtain the relationship weights between the multiple initial embedding vectors.

[0021] The multiple initial embedding vectors and relation weights are input into the graph relation aggregation layer to obtain the final embedding vectors corresponding to the multiple sensors;

[0022] The sample subsequence with the highest time order in the sample time series data is input into the time series feature extraction model to obtain the sample time series feature vector;

[0023] The sample time-series feature vector is concatenated with the final embedding vector and input into the prediction model to obtain the sample prediction data corresponding to the multiple sensors;

[0024] By minimizing the error between the predicted sample data and the measured sample data that follows the target subsequence in the sample time series data, the values ​​of the parameters in the graph embedding layer, graph structure learning layer, graph relationship aggregation layer, time series feature extraction model, and prediction model are adjusted.

[0025] In one possible implementation, the relationship weights between the plurality of initial embedding vectors are obtained by cosine similarity calculation.

[0026] In one possible implementation, the plurality of initial embedding vectors and relation weights are input into the sensor relation aggregation layer to obtain the final embedding vectors corresponding to the plurality of sensors, including:

[0027] For any target sensor among the plurality of sensors, the embedding vectors corresponding to all its neighboring sensors are weighted and summed with their corresponding relation weights, and then added to the embedding vector corresponding to the target sensor to obtain the final embedding vector corresponding to the target sensor.

[0028] Wherein, the neighboring sensors of the target sensor are the sensors corresponding to the Kth largest embedding vectors in terms of relational weight with the embedding vector corresponding to the target sensor, where K is a preset value.

[0029] In one possible implementation, the temporal feature extraction model is a Long Short-Term Memory Neural Network (LSTM) or a Recurrent Neural Network (RNN).

[0030] 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).

[0031] This invention proposes a method for anomaly detection based on a graph attention mechanism. It simultaneously models data from multiple sensors using a multivariate temporal anomaly detection model. Simultaneously, a temporal feature extraction model explicitly extracts time information from the multivariate temporal data. All predicted data are then compared with actual data to obtain better prediction results and reduce false alarms. Furthermore, compared to multiple univariate models, using a single multivariate model reduces maintenance costs. Attached Figure Description

[0032] 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.

[0033] Figure 1 This is a framework diagram of a method for determining anomalies based on a graph attention mechanism disclosed in an embodiment of the present invention;

[0034] Figure 2 This is a flowchart of a method for determining anomalies based on a graph attention mechanism disclosed in an embodiment of the present invention;

[0035] Figure 3 This is a flowchart of the method for training a graph structure learning model, a temporal feature extraction model, and a prediction model disclosed in an embodiment of the present invention. Detailed Implementation

[0036] 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.

[0037] According to one embodiment, Figure 1 This illustrates the framework of a method for anomaly identification based on a graph attention mechanism. For example... Figure 1 As shown, the framework used in this method mainly consists of six parts: a graph embedding layer, a graph structure learning layer, a graph relation aggregation layer, a temporal feature extraction model, a prediction model, and an anomaly evaluation layer. The graph embedding layer, graph structure learning layer, and graph relation aggregation layer together are referred to as the graph structure learning model (not shown in the figure).

[0038] The graph embedding layer learns an embedding vector representation for each sensor, initially a random embedding vector. As the model trains, a suitable embedding vector representation for each sensor is obtained, and the dimension of the embedding vector representation can be set according to actual needs. The graph structure learning layer learns the relationships between sensors and uses a directed graph to represent these relationships. The embedding vector of each sensor is a node in the graph, and the relationships between sensors are edges in the graph. For example, if sensor A is considered to have an effect on sensor B, then there will be an edge from A to B. The graph relationship aggregation layer aggregates the embedding vector representations of each sensor based on an attention mechanism according to the learned graph structure and relationships, obtaining the final embedding vector representation of each sensor. The temporal feature extraction model learns the temporal relationships in multivariate time-series data. The prediction model predicts the data for the next time point based on the final embedding vectors of each sensor learned by the graph structure learning model and the multivariate time-series data detected by the sensors, obtaining the predicted data. The anomaly evaluation layer determines whether an anomaly occurs at the corresponding time point based on the predicted data and the actual data at that time point.

[0039] 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.

[0040] Figure 2 This is a flowchart illustrating a method for determining anomalies based on a graph attention mechanism, as disclosed in an embodiment of the present invention. Figure 2 As shown, the method includes at least the following steps: Step 201, acquiring multivariate time-series data, wherein the time-series data is data generated by multiple sensors of equipment during semiconductor manufacturing; Step 202, acquiring the final embedding vectors corresponding to the multiple sensors according to a graph structure learning model, wherein the final embedding vectors are used to characterize the correlation between the multiple sensors; Step 203, inputting the target subsequence with the earliest time sequence in the time-series data into a time-series feature extraction model to obtain a time-series feature vector; Step 204, concatenating the time-series feature vector with the final embedding vector and inputting it into a prediction model to obtain the prediction data corresponding to the multiple sensors; Step 205, determining whether an anomaly occurs at the target time point corresponding to the measured data based on the prediction data and the measured data of the target subsequence in the time-series data.

[0041] In step 201, multivariate time-series data is acquired, which is data generated by multiple sensors of the equipment during the semiconductor manufacturing process.

[0042] Time-series data refers to the raw data generated by sensors in semiconductor manufacturing equipment, processed by a Functional Data Conversion (FDC) system. A semiconductor manufacturing machine is equipped with many sensors, each monitoring a specific parameter such as temperature, humidity, voltage, current, or pressure. The value of any parameter output by a sensor over a given period is processed by the FDC system to obtain a set of univariate time-series data. By combining the time-series data from all sensors, multivariate time-series data is obtained.

[0043] Multivariate time series data can be represented as D = {X} 1 X 2 , ..., X n}, X i ∈R k X represents the value of k sensors at the i-th time point. i It is a k-dimensional vector, where k is the number of sensors, and i = 1, 2, ..., n.

[0044] In step 202, the final embedding vectors corresponding to the multiple sensors are obtained according to the graph structure learning model. The final embedding vectors are used to characterize the correlation between the multiple sensors.

[0045] Specifically, the graph structure learning model includes a graph embedding layer, a graph structure learning layer, and a graph relation aggregation layer. The final embedding vector corresponding to the i-th sensor can be represented as... i = 1, 2, ..., k. The dimension of the embedding vector can be preset according to the specific situation, but it should be understood that the dimension of the embedding vector corresponding to each sensor should be the same.

[0046] In step 203, the target subsequence with the earliest time sequence in the time series data is input into the time series feature extraction model to obtain the time series feature vector.

[0047] Specifically, if a time window of length ω is preset, then the target subsequence used to predict the target time point as the t-th time point can be represented as S. t ={X t-ω X t-ω+1 , ..., X t-1}. S t The input is fed into the temporal feature extraction model to obtain the temporal feature vector W. t It should be understood that t > ω here.

[0048] In one embodiment, the temporal feature extraction model is a Long Short-Term Memory Neural Network (LSTM) or a Recurrent Neural Network (RNN).

[0049] In step 204, the temporal feature vector and the final embedded vector are concatenated and input into the prediction model to obtain the prediction data corresponding to the multiple sensors.

[0050] Specifically, the time-series feature vector W t With the final embedding vector to After vector concatenation, the data is input into the prediction model to obtain the predicted data of multiple sensors at time point t.

[0051] In one embodiment, the prediction model is a feedforward neural network (FNN). S t and to After concatenating the vectors, a full connection is established with the FNN to obtain the predicted data corresponding to the t-th time point.

[0052] In step 205, based on the predicted data and the measured data of the target subsequence that follows the time series data, it is determined whether there is an anomaly at the target time point corresponding to the measured data.

[0053] Based on the predicted data corresponding to the t-th time point (target time point) The measured data X that follows the target subsequence in the time series data t To determine whether an anomaly occurs at the t-th time point.

[0054] In one embodiment, the anomaly evaluation layer bases the prediction data... and the measured data X t The error is determined, and based on the comparison between the error and a preset first threshold, it is determined whether an anomaly occurs at the target time point corresponding to the measured data. For example, when the error is greater than the first threshold, it is determined that an anomaly occurs at the t-th time point.

[0055] There are several ways to calculate the error, such as using the root mean square error (RMSE), mean square error (MSE), or mean absolute error (MAE), and no specific method is used here.

[0056] Through steps 201 to 205, data generated by multiple sensors are simultaneously modeled using a multivariate temporal anomaly detection model, and temporal relationships are extracted using a temporal feature extraction model. All predicted data are then compared with the actual data to obtain better prediction results and reduce false alarms.

[0057] In some possible implementations, the method further includes: step 206, when an anomaly occurs at the target time point, determining the anomaly sensor and the corresponding anomaly type based on the predicted data and the measured data.

[0058] In one embodiment, based on the prediction data from any target sensor among the plurality of sensors... and the measured data X t The data corresponding to that, i.e. and X t The error is determined by analyzing the components in the dimension corresponding to the target sensor. Based on the comparison between the error and a preset threshold corresponding to the target sensor, it is determined whether the target sensor is malfunctioning and the corresponding malfunction type. For example, if the error is greater than the threshold, it is determined that the target sensor malfunctioned at time point t. Simultaneously, the corresponding malfunction type is determined based on the type of machine parameter detected by the target sensor.

[0059] The preset thresholds for the aforementioned target sensors can be set by engineers based on actual conditions or experience, or they can be calculated by data models based on historical data. Different thresholds can also be set for different sensors based on actual conditions.

[0060] Since the error described in step 206 is the error between scalars, when calculating the error, the difference between the two can be directly calculated, or the absolute value of the difference can be calculated; there is no limitation here.

[0061] In some possible implementations, the method further includes: step 207, determining corresponding knowledge points based on the anomaly sensor, the anomaly type, the device number of the device, and the number of the wafer being processed by the device, wherein the knowledge points are used to generate or update a knowledge graph in the semiconductor field.

[0062] In some embodiments, since multiple anomaly types may exist simultaneously at a certain point in time, corresponding to multiple target sensors malfunctioning, the knowledge point can be in the form of a tuple. Specifically, the tuple is in the form of (anomaly type 1, ... anomaly type m, device number, wafer number). For example, a specific knowledge point can be (excessive temperature, excessive pressure, machine 2, wafer 3), where wafer represents a wafer.

[0063] Figure 2 The steps included are using a trained model to predict data and detect anomalies; the method for training the model is as follows: Figure 3 As shown.

[0064] Figure 3This is a flowchart illustrating the methods for training a graph structure learning model, a temporal feature extraction model, and a prediction model disclosed in embodiments of the present invention. Figure 3 As shown, the graph structure learning model, the temporal feature extraction model, and the prediction model are trained using sample temporal data. The training includes at least the following steps: Step 301, inputting the feature data of the multiple sensors into the graph embedding layer to obtain multiple initial embedding vectors corresponding to the multiple sensors; Step 302, inputting the multiple initial embedding vectors into the graph structure learning layer to obtain the relationship weights between the multiple initial embedding vectors; Step 303, inputting the multiple initial embedding vectors and relationship weights into the graph relationship aggregation layer to obtain the final embedding vectors corresponding to the multiple sensors. Step 304: Input the time-ordered subsequence of the sample time series data into the time series feature extraction model to obtain the sample time series feature vector; Step 305: Concatenate the sample time series feature vector with the final embedding vector and input it into the prediction model to obtain the sample prediction data corresponding to the multiple sensors; Step 306: Adjust the values ​​of the parameters in the graph embedding layer, graph structure learning layer, graph relationship aggregation layer, time series feature extraction model and prediction model by minimizing the error between the sample prediction data and the measured sample data that follows the target subsequence in the sample time series data.

[0065] The time series data of the samples used to train the model can be represented as This represents the value of k sensors at time point i. It is a k-dimensional vector, where k is the number of sensors, i = 1, 2, ..., n. The time-series data used to train the model contains no outliers.

[0066] Step 301: Input the feature data of the multiple sensors into the graph embedding layer to obtain multiple initial embedding vectors corresponding to the multiple sensors.

[0067] The embedding vector of the i-th sensor can be represented as V i For V i There are several ways to initialize V. In one embodiment, V can be initialized in a way that allows initialization to be performed ... i All dimensions are initialized to 0; in another embodiment, random numbers can be used to pair V. i Initialization is performed; in yet another embodiment, it can be based on the sample time series data D. train For V i Perform initialization. For V i After initialization, the initial embedding vector is obtained.

[0068] Step 302: Input the multiple initial embedding vectors into the graph structure learning layer to obtain the relationship weights between the multiple initial embedding vectors.

[0069] In one embodiment, the relationship weights between the plurality of initial embedding vectors are obtained by cosine similarity calculation.

[0070] For example, the embedding vector V i and V j The cosine similarity between them can be calculated using formula (1):

[0071]

[0072] Among them, V i ·V j Let ||V|| denote the dot product of two vectors. i || represents vector V i The length of the module.

[0073] Step 303: Input the multiple initial embedding vectors and relation weights into the graph relation aggregation layer to obtain the final embedding vectors corresponding to the multiple sensors.

[0074] In one embodiment, for any target sensor among the plurality of sensors, the embedding vectors corresponding to all its neighboring sensors are weighted and summed with their corresponding relation weights, and then added to the embedding vector corresponding to the target sensor to obtain the final embedding vector corresponding to the target sensor; wherein, the neighboring sensors of the target sensor are the sensors corresponding to the Kth largest embedding vectors in terms of relation weight with the embedding vector corresponding to the target sensor, and K is a preset value.

[0075] In one example, the value of K is preset to 2. For the first sensor, the k-1 relation weights between it and the second to kth sensors are sorted. The two sensors corresponding to the top two ranked relation weights are the neighbor sensors of the first sensor, let's assume they are the second and third sensors. Let's set the weights of these two relations as α2 and α3. Then, the final embedding vector of the first sensor is... This can be expressed as shown in formula (2):

[0076]

[0077] Step 304: Input the sample subsequence with the highest time order in the sample time series data into the time series feature extraction model to obtain the sample time series feature vector.

[0078] Specifically, with Figure 2 The method shown is similar. With a preset time window of length ω, the sample subsequence for the r-th time point can be represented as... Will The input is fed into the temporal feature extraction model to obtain the sample temporal feature vector. It should be understood that r > ω here.

[0079] Step 305: Concatenate the sample time-series feature vector with the final embedding vector and input it into the prediction model to obtain the sample prediction data corresponding to the multiple sensors.

[0080] Specifically, the sample time-series feature vector With the final embedding vector to After vector concatenation, the data is input into the prediction model to obtain the sample prediction data corresponding to the r-th time point from multiple sensors.

[0081] Step 306, by minimizing the sample prediction data The measured sample data that follows the target subsequence in the sample time series data To address the error, the values ​​of the parameters in the graph embedding layer, graph structure learning layer, graph relationship aggregation layer, temporal feature extraction model, and prediction model are adjusted.

[0082] There are several ways to calculate the error, such as using the root mean square error (RMSE), mean square error (MSE), or mean absolute error (MAE), and no specific method is used here.

[0083] In each round of training, the model is trained using different sample subsequences and sample test data by sliding the time window of length ω.

[0084] 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.

[0085] 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.

[0086] 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 identifying anomalies based on a graph attention mechanism, comprising: Acquire multivariate time-series data, which is data generated by multiple sensors of equipment during semiconductor manufacturing; The graph structure learning model is used to obtain the final embedding vectors corresponding to the multiple sensors, and the final embedding vectors are used to represent the correlation between the multiple sensors; the graph structure learning model includes a graph embedding layer, a graph structure learning layer and a graph relationship aggregation layer. The target subsequence with the earliest time sequence in the time series data is input into the time series feature extraction model to obtain the time series feature vector; The temporal feature vector and the final embedded vector are concatenated and input into the prediction model to obtain the prediction data corresponding to the multiple sensors; Based on the predicted data and the measured data of the target subsequence in the time series data, determine whether there is an anomaly at the target time point corresponding to the measured data; The graph structure learning model, the temporal feature extraction model, and the prediction model are trained using sample temporal data. The training includes: inputting feature data from the multiple sensors into the graph embedding layer to obtain multiple initial embedding vectors corresponding to the multiple sensors; inputting the multiple initial embedding vectors into the graph structure learning layer to obtain relationship weights between the multiple initial embedding vectors; inputting the multiple initial embedding vectors and relationship weights into the graph relationship aggregation layer to obtain the final embedding vectors corresponding to the multiple sensors; inputting the time-ordered sample subsequence from the sample temporal data into the temporal feature extraction model to obtain sample temporal feature vectors; concatenating the sample temporal feature vectors with the final embedding vectors and inputting them into the prediction model to obtain sample prediction data corresponding to the multiple sensors; and adjusting the values ​​of the parameters in the graph embedding layer, graph structure learning layer, graph relationship aggregation layer, temporal feature extraction model, and prediction model by minimizing the error between the sample prediction data and the measured sample data that follows the target subsequence in the sample temporal data.

2. The method according to claim 1, characterized in that, Also includes: When an anomaly occurs at the target time point, the abnormal sensor and the corresponding anomaly type are determined based on the predicted data and the measured data.

3. The method according to claim 1, characterized in that, Based on the predicted data and the measured data of the target subsequence in the time series data, determine whether an anomaly occurs at the target time point corresponding to the measured data, including: The error is determined based on the predicted data and the measured data. Based on the comparison between the error and a preset first threshold, it is determined whether an anomaly occurs at the target time point corresponding to the measured data.

4. The method according to claim 2, characterized in that, Based on the predicted data and the measured data, the abnormal sensors and corresponding abnormality types are determined, including: Based on the data corresponding to any target sensor among the plurality of sensors in the predicted data and the measured data, an error is determined. Based on the comparison result of the error and the preset threshold corresponding to the target sensor, it is determined whether the target sensor has an anomaly and the corresponding anomaly type.

5. The method according to claim 1, characterized in that, The relationship weights between the multiple initial embedding vectors are obtained by cosine similarity calculation.

6. The method according to claim 1, characterized in that, The multiple initial embedding vectors and relation weights are input into the sensor relation aggregation layer to obtain the final embedding vectors corresponding to the multiple sensors, including: For any target sensor among the plurality of sensors, the embedding vectors corresponding to all its neighboring sensors are weighted and summed with their corresponding relation weights, and then added to the embedding vector corresponding to the target sensor to obtain the final embedding vector corresponding to the target sensor. Wherein, the neighboring sensors of the target sensor are the sensors corresponding to the Kth largest embedding vectors in terms of relational weight with the embedding vector corresponding to the target sensor, where K is a preset value.

7. The method according to claim 1, characterized in that, The temporal feature extraction model is either a Long Short-Term Memory Neural Network (LSTM) or a Recurrent Neural Network (RNN).

8. The method according to claim 1 or 3, characterized in that, The types of errors include at least: root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE).

Images