Time series fault detection, fault classification, and transition analysis

Abstract

This paper discloses a method and system for instantaneous time series analysis of data. One method includes the following steps: receiving time series data; generating a training dataset including randomized data points; generating combinations of randomized data points using these randomized data point sets within a time window; calculating distance values ​​based on these randomized data point combinations; generating a classifier based on multiple calculated distance values; and using the classifier to determine the probability that new time series data generated during a new execution period of the process matches the new time series data. A system for performing this method is also disclosed.

Description

[0001] This application is a divisional application of the invention patent application filed on September 19, 2017, with application number 201780057453.X and invention title "Time Series Fault Detection, Fault Classification and Transition Analysis Using K-Nearest Neighbor and Logistic Regression Methods". Technical Field

[0002] This disclosure relates to artificial neural networks, and more specifically to time-series fault detection, fault classification, and transition analysis for analyzing systems. Background Technology

[0003] For example, a semiconductor processing technology comprises multiple steps within a defined time interval. The process may include transitions from one step to another. Time-series data is the data collected within this time interval, including these transitions (e.g., time-series transitions). Generally, statistical methods (e.g., Statistical Process Control, SPC) are used to analyze sensor data used in semiconductor manufacturing processes. However, SPC and other statistical methods for monitoring processing cannot monitor time-series transitions. Statistical methods cannot detect short-duration signal disturbances in data received from sensors over time. Statistical methods also provide false alarms (e.g., the entire signal does not match the target signal because the smallest portion of the signal is outside the guard band) and do not allow for adjustment of the sensitivity of anomaly detection. Attached Figure Description

[0004] This disclosure is illustrative by way of example in the accompanying drawings and is not intended to be limiting.

[0005] Figure 1 This illustrates one implementation of a network architecture.

[0006] Figure 2 This illustration depicts one implementation of a method for time series transformation analysis.

[0007] Figure 3 This illustration depicts another implementation of a method for time series transformation analysis.

[0008] Figure 4 Plot the time series data used for time series transformation analysis.

[0009] Figure 5A Plot the combination of randomized data points and time windows used for time series transformation analysis.

[0010] Figure 5B Plot the distances between randomized data points used for time series transformation analysis.

[0011] Figure 6 Plot the distance to the training set used for time series transformation analysis.

[0012] Figures 7A-7B Plot the distance to the training set used for time series transformation analysis.

[0013] Figure 8 Plotting logistic regression used for time series transformation analysis.

[0014] Figure 9 Plot the effect of θ on logistic regression used for time series transformation analysis.

[0015] Figures 10A-10B Plot the probabilities of matching time series data used for time series transformation analysis.

[0016] Figure 11A-11D Plot the probabilities of matching time series data used for time series transformation analysis.

[0017] Figure 12A Plot the time series data used for time series transformation analysis.

[0018] Figure 12B Plot the distance to the training set used for time series transformation analysis.

[0019] Figure 12C Plotting logistic regression used for time series transformation analysis.

[0020] Figures 13A-13D Plot the probabilities of matching time series data used for time series transformation analysis.

[0021] Figure 14 Plot the time series data from multiple inputs used for time series transformation analysis.

[0022] Figures 15A-15D Plot the probabilities of matched time series data for multiple inputs used in time series transformation analysis.

[0023] Figure 16 An exemplary computer system is illustrated. Detailed Implementation

[0024] Embodiments of this disclosure pertain to methods and systems for time series transformation analysis of data. For example, the data sample may be sensor data from a semiconductor processing device. In one embodiment, the method and system can detect the probability that new time series data matches previous time series data. Time series transformation analysis can be performed using k-Nearest Neighbor (kNN) analysis and logistic regression (LR) as described in the embodiments. Embodiments of this disclosure are extendable in terms of adjusting the sensitivity of time series transformation analysis.

[0025] As manufacturing processes (e.g., manufacturing processes) include increasingly shorter step times, smaller components, and tighter tolerances, transitions (e.g., how to move from step A to step B in a manufacturing process) become increasingly critical. Problems can arise if process transitions are overshooted or undershooted (e.g., transitioning too quickly from 10 degrees to 20 degrees, or too slowly from 10 degrees to 20 degrees, etc.). Repeatable performance includes consistent transitions. Traditional monitoring methods (e.g., SPC) cannot monitor instantaneous time series and cannot detect short-duration signal disturbances in data received from sensors over time (referred to herein as sensor time series data). These short anomalies can cause defects (e.g., defects on the wafer) or reduce yield.

[0026] Time series transformation analysis provides the ability to monitor time series transformations. It can detect rare, unusual, and / or unexpected changes in sequences that are not detectable by traditional methods (e.g., the shape, value, position, etc. of the curves plotted against the sample of time series data); [referencing...] Figure 4 In one implementation, monitoring of time series transitions is performed by estimating the expected transition trajectory from historical time series data and comparing the trajectory of new time series data with that of historical time series data. Time series transition analysis can also detect short anomalies and provide tuning parameters to sensitize or desensitize the accuracy of detection. Time series transition analysis can also overcome the false alarm rate of conventional methods. For example, guard band analysis may provide false alarms when the entire signal does not match the target signal because the smallest part of the signal is outside the guard band, while time series transition analysis provides the probability of the signal matching the target signal and does not provide false alarms. In another implementation, time series transition analysis can be used to detect short-duration signal disturbances (e.g., capturing disturbance identifiers (e.g., similarity search)) to search for all instances of the FDC.

[0027] Fault detection classification (FDC) for time-series data monitors data from a single sensor, which may lead to inaccurate classification. More information can be extracted by monitoring multiple signals that vary over time (e.g., a valve position change preceding a pressure spike may indicate a problem in the pressure control logic, a pressure spike preceding a valve position change may indicate a problem with the pressure sensor, etc.). The techniques disclosed herein process coupled signals that vary over time.

[0028] Time series transition analysis can combine k-nearest neighbor (kNN) methods (e.g., the kNN algorithm) with logistic regression (LR) binary classifiers to monitor time series data. The combination of kNN and LR can be used to detect specific excursions in time series data. Time series transition analysis can use kNN to simplify short-term time series transitions for each time window (e.g., a 1-second sliding window over a 100-second time interval) to a single dimension to determine the distance from the expected behavior. Time series transition analysis can use LR to build a binary classifier that generates the probability that new time series data has or does not have the target pattern (e.g., whether the new time series data is outside the distance determined by the kNN method).

[0029] Time series transition analysis can be used to detect transitions between setpoint changes in a process based on time series data, and to detect deviations in new time series data from the expected transition trajectory. The expected transition trajectory can be defined by the time series data.

[0030] Figure 1 A network architecture 100 according to one implementation is illustrated. Initially, the time series transformation analysis system 102 identifies data sources 106A-N (e.g., sensors), which define the system and / or are used to monitor the system (e.g., physical processing system 104). The physical processing system 104 may be a semiconductor processing device, such as a chamber for an etching reactor, a deposition chamber, etc. A user can select time series data (e.g., samples) from various data sources in the data sources 106A-N (e.g., via a graphical user interface (GUI)) via a client machine 110. The time series transformation analysis system 102 generates a training dataset and calculates distance values ​​based on the training dataset and the time series data.

[0031] In one embodiment, the user may also select deviation 108 (i.e., a defined parameter of abnormal system behavior) via client machine 110, and deviation 108 may be stored in permanent storage unit 112 by time series transformation analysis system 102.

[0032] For example, the physical processing system 104 may include manufacturing tools or be connected to manufacturing tools directly or via a network (e.g., a local area network (LAN)). Examples of manufacturing tools include semiconductor manufacturing tools used to manufacture electronic devices, such as etchers, chemical vapor deposition furnaces, and so on. The steps involved in manufacturing such equipment may include dozens of manufacturing steps involving different types of manufacturing processes, which may be referred to as recipes.

[0033] The physical processing system 104 may include any type of computing device (including desktop computers, laptop computers, programmable logic controllers (PLCs), handheld computers, or similar computing devices) to control the system. The data source 106 (e.g., a sensor) may be part of the physical processing system 104 and / or the manufacturing tool, or may be connected to the physical processing system 104 and / or the manufacturing tool (e.g., via a network).

[0034] The client machine 110 can be any type of computing device, including desktop computers, laptops, mobile communication devices, mobile phones, smartphones, handheld computers, or similar computing devices.

[0035] In one embodiment, the entity processing system 104, data source 106, permanent storage unit 112, and client machine 110 are connected to the time series transformation analysis system 102, either directly or indirectly via a hardware interface (not shown) or a network (not shown). The network can be a local area network (LAN) (e.g., an intranet within a company), a wireless network, a mobile communication network, or a wide area network (WAN) (e.g., the Internet or a similar communication system). The network can include any number of networking and computing devices, such as wired and wireless devices.

[0036] The functional division presented above is merely illustrative. In other embodiments, the functions may be combined into monolithic components or further subdivided into any combination of components. For example, the client machine 110 and the time series transformation analysis system 102 may be hosted on a single computer system, a separate computer system, or a combination thereof.

[0037] Figure 2 This illustration depicts one embodiment of a method 200 for time series transformation analysis. Method 200 can be executed by processing logic, which may include hardware (e.g., circuitry, dedicated logic, programmable logic, microcode, etc.), software (e.g., instructions running on a processing device), or a combination thereof. In one embodiment, method 200 is performed by… Figure 1 The time series transformation analysis system 102 is used to perform this.

[0038] exist Figure 2At block 202, the processing logic of the time series transformation analysis system 102 receives time series data 402 (e.g., target signal), such as... Figure 4 As shown in the diagram. One or more sensors can generate time-series data 402 during a process (e.g., a manufacturing process). The time-series data 402 may include a first plurality of data points. The first plurality of data points may include data points at samples of the time-series data 402. For example, as... Figure 4 As shown, samples can be taken at n=25 and n+1=50. The values ​​of time series data 402 may include t(n) and t(n+1) at approximately [0, 4].

[0039] Back Figure 2 At block 204, the processing logic of the time series transformation analysis system 102 generates a training dataset including randomized data points 502 (e.g., random samples), such as... Figure 5A As shown in the diagram. Randomized data points 502 may include a distribution of expected ranges for one or more deviations relative to time series data 402. The distribution may be a normal distribution or another distribution. In one embodiment, 100 random samples are generated, each representing a deviation relative to the time series data. Figure 5A As shown in the example, randomized data points 502 include deviations at each of the data points (e.g., n and n+1). For example, data points are clustered around [0, 4] at n = 25 and n+1 = 50. Randomized data points 502 can be used as a training set for a pattern of time series data 402. Each randomized data point in randomized data points 502 can correspond to one of a first plurality of data points from time series data 402.

[0040] Back Figure 2 At block 206, the processing logic of the time series transformation analysis system 102 uses the set of randomized data points 502 within the time window 506 to generate combinations of randomized data points, such as... Figure 5A As shown in the diagram. For example, a combination of randomized data points may include one of randomized data points 502a within an instance of time window 506 from 0 to n (e.g., 25) and one of randomized data points 502b within an instance of time window 506 from n to n+1 (e.g., 25 to 50). In one implementation, the processing logic may generate randomized data points at the end of time window 506 (e.g., refer to...). Figure 5A In another embodiment, the processing logic may generate randomized data points at the midpoint of time window 506. In another embodiment, the processing logic may generate randomized data points at the beginning of time window 506.

[0041] Time window 506 can be a sliding time window, and the process can occur within a certain time interval greater than the sliding time window. The sliding time window can be a period of time stretched backward from the current time. For example, a two-second sliding window can include any sample or data point that has occurred in the past two seconds. In one embodiment of the sliding time window, the first instance can be 0-25, the second instance can be 25-50, and so on. Thus, the window slides in 25-second increments. In another embodiment of the sliding time window, the first instance can also be 0-25, the second instance can be 1-26, then 2-27, and so on. Thus, the time window slides in 1-second increments (or other time units).

[0042] Randomized data point combinations can be generated for each of the multiple instances of the sliding time window 506. Each of the multiple instances can span different time periods within the time interval (e.g., the randomized data point combination includes samples from a first data point at position n and a second data point at position n+1).

[0043] Back Figure 2 At block 208, the processing logic of the time series transformation analysis system 102 calculates distance values ​​based on combinations of randomized data points. A first distance value can be calculated for a first combination of randomized data points. The first distance value can represent the distance between a first subset of a set of multiple randomized data points and a combination of the first subset of the first set of multiple data points. The calculation of distance values ​​can be performed for each of the multiple instances of a sliding time window.

[0044] like Figure 5B As shown, randomized data points can be combined to provide randomized data point combinations 507, each of which includes a first randomized data point from t(n) and a second randomized data point from t(n+1). These randomized data point combinations 507 can be used to calculate distance values ​​using the k-nearest neighbor algorithm.

[0045] The processing logic can use the k-nearest neighbor (kNN) algorithm to calculate distance thresholds for each instance within time window 506. For example, a first distance threshold can be generated for time window 506 at time t = 25 (e.g., using data points from time 0 to 25), a second distance threshold can be generated for time window 506 at time t = 26 (e.g., using data points from time 1 to 26), and so on. The calculation of the distance thresholds may include calculating the Euclidean distance (see reference 507) between each of the multiple randomized data point combinations 507 and each of the remaining randomized data point combinations 507 from the training dataset. Figure 5B The calculation of the distance threshold may include identifying the minimum Euclidean distance from the calculated Euclidean distances. The minimum Euclidean distance may be the distance threshold.

[0046] When using kNN-type algorithms, the training dataset can be used to estimate the distance between the deviation sample and the training data. For each training sample, which includes a combination of randomized data points, the Euclidean distance between this sample and all other samples in the training set can be calculated, and the k-th minimum value can be stored. For sample j, this is achieved through equation d. j =small k (x j The distance is calculated using the formula -X), where X is an n×m matrix. The value n represents the number of training samples (e.g., 100 random samples). The value of m can represent the number of time samples or data points (e.g., ...). Figure 4-6 (The example illustrates two time samples at n=25 and n+1=50). Variable x j It can be an m-element vector (e.g., [0, 4]) and can represent the j-th column in X. This process is repeated for all samples in the training set to generate a neighborhood or limit vector L with n elements. The neighborhood or limit vector L can be used to generate a well-separated training set to train a simple classifier. Random samples from the training set can be selected to compute k for each sample. nn =small k (x j -X). Random samples not from the training set can be selected, and k is calculated for visualization purposes. nn value.

[0047] like Figure 6 As shown, random samples (sample category 602a) were selected from the deviation pattern, and k was estimated. nn Measurement. Random samples exhibiting non-deviation behavior (sample class 602b) were selected, and k was estimated. nn Metrics. Sample class 602a shows a smaller distance to the training set than sample class 602b. Figure 6 In this context, the two-dimensional signal has been reduced to a one-dimensional quantity that appears to be linearly separable.

[0048] The above processing has been described for samples consisting of two data points. However, this same processing can be generalized to multi-dimensional data to reduce multi-dimensional inputs to a single metric (e.g., the k-th distance between the sample and the training data). This is followed by testing against all possible values. Figure 6 At that time, a minimum value is located at an offset position of approximately [0, 4]. Figures 7A-7B Plot the kNN metric for multiple input patterns and show the minimum value that appears at approximately [0, 4].

[0049] Back Figure 2At block 210, the processing logic of the time series transformation analysis system 102 generates a classifier based on calculated distance values. The processing logic can generate a classifier by determining distance thresholds based on multiple calculated distances. Classifier generation can be performed for each of multiple instances of the sliding time window 506. Logistic regression can be used to generate the classifier.

[0050] The processing logic can be determined from the training data using logistic regression 802 (logit fit); such as... Figure 8 (As shown in the diagram) (e.g., generating a logarithmic fit 802 for the training data, which will generate the probability of signal mismatch). The training data may include raw time series data as well as randomized combinations of data points and their calculated distance values. The equation p(y|X) = 1 / (1+e -β*X This can be used to determine logistic regression 802. Training data is used to estimate β. For example... Figure 8 As shown, logistic regression 802 may include a location of a transition pattern from a first data point (sample class 602a) to a second data point (sample class 602b). The transition pattern may reflect near a reflection point 804 centrally located on the transition pattern. Time series data 402 can be detected as a step function with key transitions (e.g., a short step gradient deposition). Time series transition analysis can be used to overcome false positives caused by boundary methods.

[0051] Instantaneous time series analysis can utilize tuning parameters. Time series transformation analysis can control how much out-of-specification samples contribute to the distance. Increasing the contribution of out-of-specification samples makes the system more sensitive. Increasing the reflection point 804 makes the system less sensitive. Adjusting the slope of the logistic regression 802 changes the probability of samples approaching reflection 804. The logistic regression 802 may have a reflection limit (e.g., a vertical line), and any sample exceeding the reflection limit can be considered as not matching the expected behavior. In one implementation, compared to Figure 8 The transition pattern shown may require a shallower or less shallow transition pattern. θ can be used as a tuning parameter to adjust the transition to a shallower or less shallower pattern.

[0052] Figure 9 The logistic regression 802 with θ902 adjusted to generate a shallower transition is illustrated. With a shallower transition, it is possible to apply this to all inputs t. n and t n+1 Estimate the probability. When using beta estimation, the probability can be... Figures 7A-7B Maximize the minimum value determined in the algorithm.

[0053] The processing logic can receive a first parameter (e.g., θ902) to adjust the sensitivity of the probability determination. For example, θ902a can have a value of one, θ902b can have a value of two, and θ902c can have a value of five. The processing logic can adjust the shallowness of the transformation pattern around the reflection point 804 based on the first parameter 902. The tuning knob can be used to set the tuning to low sensitivity, high sensitivity, etc., by changing θ902.

[0054] Back Figure 2 At block 212, the processing logic of the time series transformation analysis system 102 uses a classifier to determine the probability that new time series data matches the original time series data. The processing logic can receive new time series data and calculate a second distance value between the original time series data within time window 506 and the new time series data within time window 506. The processing logic can use the classifier to determine whether the new time series data within time window 506 has a second distance value exceeding a distance threshold, and generate a fault or notification in response to the determination that the new time series data within time window 506 exceeds the distance threshold.

[0055] Figures 10A-10B For all values ​​[t] n , t n+1 The value indicates the probability of the input matching the time series data 402 (e.g., having a maximum value at [0, 4]).

[0056] Figure 11A-11D This plot illustrates the probability of various new time series data 1002 matching time series data 402 when using time series transformation analysis. Figure 11A In this context, the new time series data 1102a has a pattern that substantially matches the pattern of the time series data 402, resulting in a probability of approximately 1. Figure 11B In the data, the new time series data 1102b at n=25 is larger than expected, therefore the probability of matching time series data 402 is approximately 0.93. Figure 11C In the data, the new time series data 1102c is higher than time series data 402 at n=25 and lower than time series data 402 at n=50, therefore the probability of matching the target signal is approximately 0.5. Figure 11D In the new time series data 1102d, the value is significantly higher at n=25 and significantly lower at n=50, therefore the probability of matching the target is 0.

[0057] Figure 3 This illustration depicts one embodiment of a method 300 for time series transformation analysis. Method 300 can be executed via processing logic, which may include hardware (e.g., circuitry, dedicated logic, programmable logic, microcode, etc.), software (e.g., instructions running on a processing device), or a combination thereof. In one embodiment, method 300 is performed by… Figure 1 The time series transformation analysis system 102 is used to perform this.

[0058] At block 302, the processing logic receives time-series data 1102a-d including the first plurality of data points (refer to...). Figure 11A-11D Each of the multiple data points can be associated with a different time period. Time-series data can be generated by sensors during the process.

[0059] At block 304, the processing logic compares a first subset of a first plurality of data points from time series data 1102a-d within time window 1106 with a second subset of a second plurality of data points from previous time series data 402. Time window 1106 can be a sliding time window that extends backward by a specific amount in time from the current time point. The sliding time window can extend backward by a specific amount in time from the current time point.

[0060] At block 306, the processing logic calculates the distance value representing the combined distance between the first subset of the first plurality of data points and the second subset of the second plurality of data points.

[0061] At block 308, the processing logic determines whether the distance value exceeds the distance threshold (see reference). Figure 11A-11D ).

[0062] At block 310, the processing logic outputs a notification in response to a determination that the distance value exceeds a distance threshold. In one embodiment, the notification includes an indication of the probability that new time series data matches time series data 402 (e.g., 0.996 for new time series data 1102a, 0.926 for new time series data 1102b, 0.502 for new time series data 1102c, and 0 for new time series data 1102d). In one embodiment, the notification includes an indication of which segments of the new time series data (e.g., time window intervals corresponding to the new time series data) have a probability threshold (e.g., 0.5, 0.85) below the probability threshold for matching time series data 402. In one embodiment, the notification may be displayed via a graphical user interface (e.g., via graphics, charts, text, etc.). In one embodiment, the notification is one or more of an audible, visual, etc. alert. In one embodiment, the notification is sent via one or more of a telephone, email, text, etc. In one implementation, the output of the notification causes one or more of the tools, devices, components, facilities, etc., to perform one or more of the following actions: stop activity, pause activity, slow down activity, shut down, etc.

[0063] Time series transformation analysis can be used for anomaly detection. In one implementation, Fault Detection and Classification (FDC) automatically searches recipe sensor data for known defects and / or anomaly markers. User setup costs can be low because expected behavior can be inferred from historical behavior. The database of known defects can be independent of the recipe setpoint. The same database can be applied to multiple recipes. Defect databases can be developed internally in a controlled environment and deployed to the field. Known defects can have corrective actions that allow for rapid resolution of known defects. Fault repair knowledge can be captured for anomaly markers (e.g., highlighting tracks or sensor data of interest to users, allowing users to label or classify anomaly markers and add corrective actions). Typical use cases include post-processing recipe data for known defects and capturing knowledge about fault repair and new defects.

[0064] Time series shift analysis can be used for time series deviation detection to search for time series exhibiting anomalous behavior that cannot be detected by traditional methods such as SPC and standard fault monitoring methods. User setup costs can be low because expected behavior is inferred from historical behavior. The algorithm can be designed to tolerate false alarms inherent in other methods, such as simple protection band monitoring. Time series deviations can be stored and used to search historical or future data. Fault repair knowledge can be captured. Common use cases include post-processing recipe data for known defects, knowledge capture regarding fault repair and new defects, analysis of instantaneous time series, and repeatability analysis.

[0065] Time series transformation analysis can be used to identify problems when a process experiences errors. For example, a chamber may be experiencing intermittent pressure spikes, but finding the root cause and solution can be difficult due to one or more of the following reasons: lack of data derived from the tool, or inability to reproduce the error internally or in the field. When using time series transformation analysis, a subset of the tool's historical cycle data can be searched for deviations. Deviations can be found (e.g., a deviation search identifies multiple strokes that do not match expected behavior), spikes in the tool data can be matched, and the search can be repeated for deviations. The occurrence of several deviations allows for efficient troubleshooting and problem resolution. Problems can be identified as functional issues of specific components (e.g., fluctuations in pressure readings caused by specific valve opening and closing of pumps).

[0066] Time series transition analysis can also be used to detect instability. For example, the tool might use a lower power identifier on the formulation. Candidate formulations can be continuously cycled on the tool. Manually analyzing all strokes may be impractical, thus potentially missing intermittent low-probability and / or frequent issues. With anomaly detection and time series deviation detection, power and reflected power can be analyzed for all steps across all strokes of a candidate formulation. The analysis can quickly identify suspicious behavior on a certain percentage of strokes. Some observed defects may have potential yield impacts. Feedback sent to the process development team can suggest formulation changes and process repetition. Deviations can be reduced by approximately 5%.

[0067] Figure 12A The time series data used for time series transformation analysis is plotted. As shown in Figure 12, n samples or data points of time series data 402 are truncated, instead of... Figure 4 The two samples in the time window shown are illustrated. In one implementation, samples are truncated at [5, 10, 15...95], resulting in 19 samples, thus making Method 200 or 300 a 19-dimensional problem rather than a 2-dimensional problem. When using time series transformation analysis (e.g., Method 200, Method 300), a training set is generated for the target signal at each sample point.

[0068] like Figure 12B As shown, random samples (category 602a) were selected from the deviation pattern, and k was estimated. nn Measurement. Random samples exhibiting non-deviation behavior (class 602b) were selected, and k was estimated. nn Measurement. For example... Figure 12C The diagram shows how to generate a logistic regression 802 using θ = 5. Using logistic regression 802, the probabilities of various input signals can be estimated for different matching time series data 402, such as... Figures 13A-13D As shown in the diagram. Time series data 402 is a pattern used to generate a trained classifier. New time series data 1302a-d are new signals additionally generated by performing a specific process associated with the original time series data 402. Figure 13A In this study, the new time series data 1302a is offset relative to time series data 402, and the matching probability is approximately 0.6. Figure 13B In the data, the new time series data 1302b has a higher match rate than the time series data 402 at n=0 to n=50, and the match probability is about 0.7.

[0069] like Figure 14As shown, time series data 402 may include first data 1402 (e.g., time series data 1402) from a first sensor and second data 1404 (e.g., time series data 1404) from a second sensor. Processing logic can determine the temporal relationship between the first and second data (e.g., capturing temporarily-spaced covariate signals for FDC). Each time series data may have different patterns on its respective signal. Figure 14 In this context, a dip in time series data 1404 may be associated with an increase in time series data 1402 (e.g., it may cause an increase in time series data 1402). Time series transformation analysis (e.g., methods 200, 300) can be used to detect correlation patterns in time series data 1402 and 1404. In one implementation, a single training vector is generated by concatenating time series data 1402 and 1404 (e.g., generating a 39-dimensional question). When using logistic regression and the kNN algorithm, the probability of various input signals matching time series data 1402 and 1404 and the relationship between time series data 1402 and 1404 can be estimated. Figure 15A -D plots the probability that input signals 1502 and 1504 match time series data 1402 and 1404.

[0070] In one implementation, two training sets are generated (one training set for each time series data 402) and two-dimensional logistic regression 802 is used.

[0071] In one example, time-series transition analysis can receive data measured by three sensors. This data may include forward power data, reflected power data, and pressure data. The three signals from the three sensors and their covariance can indicate indicators of plasma strike deviation. Time-series transition analysis can identify anomalies within a certain time interval in the data measured by the three sensors. The deviation relative to expectations may primarily occur in the forward power signal at approximately 0.4 seconds into the time interval. The deviation can also cause a higher-than-normal reflected power indicator. This indicates a plasma strike problem. The pressure may show the correct shape but be offset by approximately 0.5 seconds. Pressure spikes can mark when the reflected power strike occurred. Time-series transition analysis of the data from the three sensors can identify where the anomalies affecting one or more other signal data begin to occur, determining what caused the plasma strike deviation.

[0072] Figure 16 This is a block diagram illustrating an exemplary computing device (or system) 1600. In one embodiment, the computing device (or system) 1600 may be... Figure 1The time series transformation analysis system 102. The computing device 1600 includes a set of instructions for causing the computing device 1600 to perform any or more of the methodologies discussed herein. The machine can operate as a server machine in a client-server network environment. The machine can be a personal computer (PC), a set-top box (STB), a server, a network router, a switch or bridge, or any machine capable of executing a set of instructions (sequentially or otherwise) specifying actions to be taken by the machine. Further, although only a single computing device is illustrated, the term "computing device" is also used to include any series of machines that individually or jointly execute a set (or more) of instructions to perform any or more of the methodologies discussed herein.

[0073] An exemplary computing device 1600 includes a processing system (processing device) 1602 that communicates with each other via a bus 1608, a main memory 1604 (e.g., read-only memory (ROM), flash memory, dynamic random access memory (DRAM) (e.g., synchronous DRAM (SDRAM)) etc.), a static memory 1606 (e.g., flash memory, static random access memory (SRAM) etc.), and a data storage device 1616.

[0074] Processing device 1602 represents one or more general-purpose processing devices, such as microprocessors, central processing units, etc. More specifically, processing device 1602 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, or a processor implementing other instruction sets or combinations thereof. Processing device 1602 may also be one or more special-purpose processing devices, such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), network processors, etc. Processing device 1602 is configured to perform the operations and steps discussed herein.

[0075] The computing device 1600 may further include a network interface device 1622. The computing device 1600 may also include a video display unit 1610 (e.g., a liquid crystal display (LCD) or a cathode ray tube (CRT)), a digit input device 1612 (e.g., a keyboard), a cursor control device 1614 (e.g., a mouse), and a signal generation device 1620 (e.g., a speaker).

[0076] Data storage device 1616 may include a computer-readable storage medium 1624 storing one or more instruction sets 1626 on the data storage device, these instruction sets implementing any or more of the methods or functions described herein. In one embodiment, instruction 1626 includes a time series transition analysis system 102. Computer-readable storage medium 1624 may be a non-transitory computer-readable storage medium including instructions that, when executed by a computer system, cause the computer system to perform a set of operations including time series transition analysis (e.g., method 200, method 300, etc.). Instruction 1626 may also reside ( wholly or at least partially) in main memory 1604 and / or processing device 1602 during execution of these instructions by computing device 1600, which also constitute computer-readable media. Instruction 1626 may further be transmitted or received on network 1628 via network interface device 1622.

[0077] While computer-readable storage medium 1624 is shown as a single medium in the exemplary embodiment, the term "computer-readable storage medium" is used to include a single medium or multiple media (e.g., a centralized or distributed database and / or associated caches and servers) that store one or more instruction sets. The term "computer-readable storage medium" is also used to include any medium capable of storing, encoding, or implementing a set of instructions executable by a machine, and said instruction set causing the machine to perform any or more of the methodologies of this disclosure. Accordingly, the term "computer-readable storage medium" is used to include (but is not limited to) solid-state memory, optical media, and magnetic media.

[0078] Some parts of the following detailed description are presented in the symbolic representation of algorithms and operations on data bits within computer memory. These algorithmic descriptions and representations are means by which those skilled in the art of data processing most effectively communicate the substance of their work to others skilled in the art. An algorithm here (and generally) is conceived as a self-consistent sequence of steps leading to a certain result. These steps are those involving the manipulation of physical quantities. Typically, though not always, these quantities take the form of electrical or magnetic signals that can be stored, transmitted, combined, compared, and otherwise manipulated. It is sometimes appropriate (primarily for general purposes) to refer to these signals as bits, values, elements, symbols, characters, items, numbers, etc.

[0079] However, it should be remembered that all such terms and similar terms are to be associated with appropriate entities and are merely suitable labels applied to those entities. Unless otherwise specifically stated, as understood from the following discussion, it is to be understood that throughout this specification, discussions using terms such as “determine,” “identify,” “compare,” “send,” etc., refer to the actions and procedures by which a computer system (or similar electronic computing device) manipulates and transforms data represented by entities (e.g., electronic) in the computer system’s caches and memories into other data similarly represented by entities in the computer system’s memory or caches or other such information storage, transmission, or display devices.

[0080] Embodiments of this disclosure also relate to systems for performing the operations described herein. Such systems may be specifically constructed for the purposes described herein, or they may include a general-purpose computer selectively started or reconfigured by a computer program stored in the computer. Such computer programs may be stored in computer-readable storage media, such as (but not limited to) any type of disk (including floppy disks, optical disks, CD-ROMs, and magneto-optical disks), read-only memory (ROM), random access memory (RAM), EPROM, EEPROM, magnetic or optical cards, flash memory, or any type of medium suitable for storing electronic instructions.

[0081] The algorithms and displays presented herein are not, in essence, relating to any particular computer or other device. Various general-purpose systems may be used in conjunction with the programs taught herein, or more specialized devices may be constructed to perform the method steps. The structures of various systems used for these purposes will be apparent from the descriptions herein. Furthermore, implementations of this disclosure are not described with reference to any particular programming language. It will be understood that the teachings of the disclosure as described herein can be implemented using various programming languages.

[0082] It should be understood that the above description is intended to be illustrative and not restrictive. After reading and understanding the above description, those skilled in the art will understand many other embodiments. Therefore, the scope of protection of this disclosure should be determined by referring to the appended claims and the full scope of the equivalents conferred by such claims.

Claims

1. A method for time series transformation analysis of sensor data from a semiconductor processing device, comprising the following steps: Receive historical time series data, which is sensor data for semiconductor manufacturing processes generated by one or more sensors of the semiconductor processing device during deviations in one or more semiconductor processing processes performed via the semiconductor processing device; Generate training data, which includes multiple randomized data points associated with the expected range of the deviation behavior of the historical time series data; as well as A trained logistic regression classifier is generated by training a logistic regression classifier based on the training data using a processing device. The trained logistic regression classifier is associated with logistic regression, which indicates the location of a transition pattern from one or more first data points to one or more second data points. The one or more first data points indicate the deviation behavior, and the one or more second data points indicate the non-deviation behavior of the one or more semiconductor processing steps performed via the semiconductor processing device. The transition pattern reflects near a reflection point located on the transition pattern. The trained logistic regression classifier is capable of indicating the probability that new time-series data generated during a new execution of the one or more semiconductor processing steps matches the deviation behavior corresponding to the historical time-series data.

2. The method according to claim 1, wherein The sensor data includes forward power data, reflected power data, or pressure data, and The transition pattern corresponds to a transition between setpoint changes in one or more semiconductor processing processes of the semiconductor processing apparatus, and wherein the probability is associated with the deviation of the new time series data relative to the transition pattern.

3. The method according to claim 1, further comprising the steps of: Multiple randomized data point combinations are generated based on the set of multiple randomized data points within the time window; as well as Multiple distance values ​​are calculated based on the combination of the multiple randomized data points, wherein the training of the logistic regression classifier is further based on one or more of the multiple distance values. The calculation of the plurality of distance values ​​includes, for each of the randomized data point combinations, calculating a corresponding distance value between the corresponding randomized data point combination and each of the remaining randomized data point combinations in the plurality of randomized data point combinations.

4. The method of claim 3, wherein training the logistic regression classifier based on one or more of the plurality of distance values ​​comprises training the logistic regression classifier based on the minimum distance value among the plurality of distance values.

5. The method according to claim 1, further comprising the following steps: Receive tuning parameters; and The slope of the transition pattern around the reflection point is adjusted based on the tuning parameters, thereby adjusting the sensitivity of the detection accuracy of whether the new time series data matches the historical time series data.

6. The method of claim 1, wherein the logistic regression has a reflection limit, wherein a low matching probability is associated with the new time series data exceeding the reflection limit.

7. The method of claim 1, wherein the historical time series data includes first data from a first sensor and second data from a second sensor, wherein the trained logistic regression classifier is further based on the temporal relationship between the first data and the second data.

8. A method for time series transformation analysis of sensor data from a semiconductor processing device, comprising the following steps: Receive current time-series data for semiconductor manufacturing processes generated by one or more sensors of the semiconductor processing device during one or more semiconductor processing processes performed via the semiconductor processing device; The current time-series data is provided by a processing device as input to a trained logistic regression classifier, which is trained using training data based on historical time-series data, which is sensor data for semiconductor manufacturing processes generated by the one or more sensors during deviation behaviors of the one or more semiconductor processing processes. The training data contains a plurality of randomized data points associated with an expected range of deviation behaviors of the historical time-series data. The trained logistic regression classifier is associated with logistic regression indicating the location of a transition pattern from one or more first data points to one or more second data points, the one or more first data points indicating the deviation behavior, and the one or more second data points indicating non-deviation behaviors of the one or more semiconductor processing processes. The transition pattern reflects near a reflection point located on the transition pattern. Obtain one or more outputs from the trained logistic regression classifier; Based on the one or more outputs, determine the probability that the current time series data matches the deviation behavior corresponding to the historical time series data; as well as Based on the probability of satisfying the threshold probability, an action corresponding to the one or more semiconductor processing processes is performed.

9. The method according to claim 8, wherein The sensor data includes forward power data, reflected power data, or pressure data, and Performing the action includes one or more of the following: Provide alerts; Interrupt the activity of the semiconductor processing device; or Update the manufacturing parameters of the semiconductor processing device.

10. The method of claim 8, further comprising the following steps: Based on the probability, an alert is provided that includes one or more of the following: A first indication of the probability that the current time series data matches the historical time series data; or The second indication is a probability of one or more segments of the current time series data having a corresponding probability of matching the historical time series data.

11. The method of claim 8, wherein the trained logistic regression classifier is a binary classifier that indicates whether the current time series data is outside a distance threshold determined using k-nearest neighbor (kNN) based on the historical time series data, wherein the distance threshold is the smallest Euclidean distance among the Euclidean distances between a given combination of randomized data points calculated by combining the plurality of randomized data points and the remaining randomized data point combinations.

12. The method of claim 8, wherein the transition pattern corresponds to a transition between setpoint changes in one or more semiconductor processing processes of the semiconductor processing apparatus, and wherein the probability is associated with a deviation of the current time series data relative to the transition pattern.

13. The method of claim 8, further comprising the following steps: Receive tuning parameters; and The slope of the transition pattern around the reflection point is adjusted based on the tuning parameters, thereby adjusting the sensitivity of the detection accuracy of whether the current time series data matches the historical time series data.

14. The method of claim 8, wherein the logistic regression has a reflection limit, wherein a low matching probability is associated with the current time series data exceeding the reflection limit.

15. A non-transitory computer-readable storage medium including instructions that, when executed by a processing device, cause the processing device to perform operations for time-series transition analysis of sensor data from a semiconductor processing device, the operations including: Receive historical time series data, which is sensor data for semiconductor manufacturing processes generated by one or more sensors of the semiconductor processing device during deviations in one or more semiconductor processing processes performed via the semiconductor processing device; Generate training data, which includes multiple randomized data points associated with the expected range of the deviation behavior of the historical time series data; as well as A logistic regression classifier is trained based on the training data to generate a trained logistic regression classifier, wherein the trained logistic regression classifier is associated with logistic regression, which indicates the location of a transition pattern from one or more first data points to one or more second data points, the one or more first data points indicating the deviation behavior, the one or more second data points indicating the non-deviation behavior of the one or more semiconductor processing steps performed via the semiconductor processing device, wherein the transition pattern reflects near a reflection point located on the transition pattern, and the trained logistic regression classifier is capable of indicating the probability that new time series data generated during a new execution of the one or more semiconductor processing steps matches the deviation behavior corresponding to the historical time series data.

16. The non-transitory computer-readable storage medium according to claim 15, wherein... The sensor data includes forward power data, reflected power data, or pressure data, and The transition pattern corresponds to a transition between setpoint changes in one or more semiconductor processing processes of the semiconductor processing apparatus, and wherein the probability is associated with the deviation of the new time series data relative to the transition pattern.

17. The non-transitory computer-readable storage medium of claim 15, wherein the operation further comprises: Multiple randomized data point combinations are generated based on the set of multiple randomized data points within the time window; as well as Multiple distance values ​​are calculated based on the combination of the multiple randomized data points, wherein the training of the logistic regression classifier is further based on one or more of the multiple distance values. The calculation of the plurality of distance values ​​includes, for each of the randomized data point combinations, calculating a corresponding distance value between the corresponding randomized data point combination and each of the remaining randomized data point combinations in the plurality of randomized data point combinations.

18. The non-transitory computer-readable storage medium of claim 17, wherein training the logistic regression classifier based on one or more of the plurality of distance values ​​comprises training the logistic regression classifier based on the minimum distance value among the plurality of distance values.

19. The non-transitory computer-readable storage medium of claim 15, wherein the operation further comprises: Receive tuning parameters; as well as The slope of the transition pattern around the reflection point is adjusted based on the tuning parameters, thereby adjusting the sensitivity of the detection accuracy of whether the new time series data matches the historical time series data.

20. The non-transitory computer-readable storage medium of claim 15, wherein the logistic regression has a reflection limit, wherein a low matching probability is associated with the new time-series data exceeding the reflection limit.

Images