Abnormality determination device

The abnormality detection device in semiconductor manufacturing equipment addresses erroneous judgments by calculating and comparing monitoring statistics with reference secular variations, ensuring accurate abnormality detection and efficient identification of malfunction causes.

WO2026239708A1PCT designated stage Publication Date: 2026-11-19HITACHI HIGH TECH CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2026/006407
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-14
Filing Date
2026-02-20
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing methods for determining abnormalities in semiconductor manufacturing equipment are prone to erroneous judgments due to age-related changes and temporary disturbances in time-series sensor data, making it difficult to efficiently identify the causative components of malfunctions.

Method used

An abnormality detection device that calculates actual and predicted monitoring statistics, compares them with reference secular variations, and determines the presence of abnormalities while assessing the risk of misjudgment, using a processor to analyze time-series sensor data from semiconductor manufacturing equipment.

Benefits of technology

Provides accurate abnormality detection results and identifies components with high misjudgment risk, improving the efficiency of identifying the cause of malfunctions by prioritizing investigations on relevant sensors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2026006407_19112026_PF_FP_ABST
    Figure JP2026006407_19112026_PF_FP_ABST
Patent Text Reader

Abstract

Provided is a technology for presenting an accurate abnormality presence / absence determination result in abnormality determination of time-series sensor data indicating a change over time. An abnormality determination device for determining an abnormality of a component of a semiconductor manufacturing device: receives input of time-series sensor data indicating a state transition of the component, and a normal period and a diagnosis period of the data; calculates a statistic of an abnormality determination object from the time-series sensor data; calculates a measured monitoring statistic; calculates a change amount of the statistic caused by deterioration over time of the component; calculates a measured change-over-time amount; calculates a predicted monitoring statistic corresponding to the diagnosis period by using the measured change-over-time amount corresponding to the normal period; compares the measured monitoring statistic corresponding to the diagnosis period with the predicted monitoring statistic corresponding to the diagnosis period to perform abnormality determination; compares the measured change-over-time amount corresponding to the normal period with a reference change-over-time amount defining a range that the measured change-over-time amount can take to calculate an erroneous determination risk of the abnormality determination; and outputs an abnormality determination result and the erroneous determination risk.
Need to check novelty before this filing date? Find Prior Art

Description

Abnormality determination device

[0001] This disclosure relates to an abnormality detection device, and more particularly to an abnormality detection device for determining abnormalities in semiconductor manufacturing equipment used to process semiconductor wafers.

[0002] Semiconductor manufacturing equipment processes multiple semiconductor wafers sequentially. When a malfunction occurs in semiconductor manufacturing equipment, it is necessary to quickly identify the causative component. One method for improving efficiency in this process is to determine whether there are any sudden abnormal changes in time-series sensor data obtained from sensors that observe the state of each component constituting the semiconductor manufacturing equipment, between the normal period during normal operation of the semiconductor manufacturing equipment and the diagnostic period after a malfunction, and to prioritize the investigation of components related to sensors that show abnormal changes. However, in semiconductor manufacturing equipment, due to the accumulation of by-products associated with processing, there is time-series sensor data that shows age-related changes specific to each component, regardless of malfunctions. Therefore, if a certain threshold is set for the amount of change between the normal period and the diagnostic period to determine whether there is an abnormal change, misjudgment due to the influence of age-related changes may occur. Patent Document 1 provides a method applicable to determining whether there is an abnormal change in time-series sensor data that shows age-related changes specific to each component. In Patent Document 1, based on sensor data from the equipment's past normal period, future values ​​of the sensor data considering the influence of age-related changes are predicted, and the presence or absence of an abnormal change is determined based on a comparison of the predicted value and the observed value.

[0003] Japanese Patent Publication No. 2020-149368

[0004] However, in semiconductor manufacturing apparatuses, the determination of the presence or absence of an abnormal change according to the prior art frequently causes erroneous determinations, making it difficult to improve the efficiency of identifying the component that is the cause of a malfunction in the semiconductor manufacturing apparatus. A semiconductor manufacturing apparatus is composed of a large number of components, and mutual interference or the like occurs between sensors that observe the states of the components. Therefore, even time-series sensor data during a normal period may exhibit temporary disturbance. In this case, the secular change is overestimated or underestimated from the normal period, and a predicted value based thereon fluctuates, which may cause an erroneous determination of the presence or absence of an abnormal change. Further, in a semiconductor manufacturing apparatus, time-series sensor data for a certain period after maintenance work such as cleaning may be regarded as invalid for observing the state of a component, and may be excluded from the normal period. In this case, since the interval between the normal period and the diagnosis period of the time-series sensor data is widened, the fluctuation of the predicted value due to the overestimated or underestimated secular change also increases.

[0005] The present disclosure has been made in view of the above circumstances, and provides a technique for presenting an accurate abnormality presence / absence determination result in abnormality determination of time-series sensor data indicating a secular change.

[0006] According to one embodiment, an abnormality determination device includes a data input circuit, a data output circuit, a storage device, and a processor, and is an abnormality determination device that determines an abnormality of a component of a semiconductor manufacturing apparatus, wherein time-series sensor data indicating a state transition of the component, and at least one normal period and a diagnosis period among periods of the time-series sensor data are input to the data input circuit, the processor calculates a statistic to be subjected to abnormality determination from the time-series sensor data to calculate an actually measured monitoring statistic, calculates a variation of the statistic caused by secular degradation of the component from the time-series sensor data to calculate an actually measured secular variation, the storage device stores a reference secular variation that specifies a possible range of the actually measured secular variation, the processor calculates a predicted monitoring statistic corresponding to a diagnosis period using the actually measured secular variation corresponding to a normal period, compares the actually measured monitoring statistic corresponding to the diagnosis period with the predicted monitoring statistic corresponding to the diagnosis period to determine presence or absence of an abnormality, compares the actually measured secular variation corresponding to the normal period with the reference secular variation corresponding to the normal period to calculate an erroneous determination risk of the abnormality presence / absence determination, and the data output circuit outputs the abnormality presence / absence determination result and the erroneous determination risk.

[0007] Furthermore, according to another embodiment, the abnormality determination device comprises a data input circuit, a data output circuit, a memory device, and a processor, and is an abnormality determination device for determining abnormalities in components of semiconductor manufacturing equipment, wherein the data input circuit receives time-series sensor data showing the state changes of the component, and at least one normal period and a diagnostic period from the period of the time-series sensor data as input, the processor calculates a statistical quantity to be determined abnormally from the time-series sensor data and calculates an actual monitoring statistical quantity, calculates a change in a statistical quantity caused by the aging deterioration of the component from the time-series sensor data and calculates an actual aging change quantity, the memory device stores a reference aging change quantity that specifies the range that the actual aging change quantity can take, the processor calculates a predicted monitoring statistical quantity corresponding to the diagnostic period using the actual aging change quantity corresponding to the normal period and the reference aging change quantity, compares the predicted monitoring statistical quantity with the actual monitoring statistical quantity corresponding to the diagnostic period and determines whether there is an abnormality, and the data output circuit outputs the abnormality determination result.

[0008] According to this disclosure, a technology is available that provides accurate abnormality detection results for time-series sensor data showing changes over time. Furthermore, if a risk of misjudgment is indicated, operators can prioritize investigating components related to sensors that have been determined to have abnormal changes or those that have not changed but are determined to have a high risk of misjudgment, thereby improving the efficiency of identifying the causative component. Issues, configurations, and effects other than those described above will be clarified by the following description of the embodiments.

[0009] Figure 1 is a schematic diagram of the semiconductor manufacturing apparatus according to one embodiment. Figure 2A is a partial left view of the overall configuration diagram of the semiconductor manufacturing apparatus and the anomaly detection device for the semiconductor manufacturing apparatus according to one embodiment. Figure 2B is a partial right view of the overall configuration diagram of the semiconductor manufacturing apparatus and the anomaly detection device for the semiconductor manufacturing apparatus according to one embodiment. Figure 3 is a flowchart showing an example of the processing flow for creating time-series sensor data according to one embodiment. Figure 4 is a diagram showing an example of the data structure of the time-series sensor data storage unit according to one embodiment. Figure 5 is a diagram showing an example of time-series sensor data showing aging changes according to one embodiment. Figure 6 is a diagram showing an example of time-series sensor data showing abnormal changes according to one embodiment. Figure 7 is a flowchart showing an example of the processing flow for determining the presence or absence of an anomaly and calculating the risk of misjudgment according to one embodiment. Figure 8 is a diagram showing an example of the calculation results for determining the presence or absence of an anomaly and calculating the risk of misjudgment on time-series sensor data according to one embodiment. Figure 9 is a diagram showing an example of the display screen on the judgment result display monitor 511 according to one embodiment. Figure 10 is a flowchart showing an example of the processing flow for calculating the reference aging change amount according to one embodiment. Figure 11 is a diagram showing an example of the data structure of the reference aging change amount storage unit according to one embodiment. Figure 12 is a flowchart showing an example of the processing flow for the calculation of whether or not an abnormality is present according to Example 2.

[0010] The embodiments of this disclosure will be described below with reference to the drawings. In all drawings used to illustrate the embodiments, the same parts will be denoted by the same reference numerals, and repeated descriptions of the same parts may be omitted. In addition, the drawings may be more schematic than the actual embodiments in order to make the explanation clearer, but they are merely examples and do not limit the interpretation of this disclosure.

[0011] (1) Semiconductor Manufacturing Apparatus As shown in the schematic configuration diagram of the semiconductor manufacturing apparatus in Figure 1, the semiconductor manufacturing apparatus 10 in this embodiment has multiple recipes set in advance, and processes the wafer (sample, substrate) 101 according to one of the specified recipes. The semiconductor manufacturing apparatus 10 also has a group of sensors 102 that observes the state of multiple components that make up the semiconductor manufacturing apparatus 10. The group of sensors 102 includes multiple sensors, and the state of each component (e.g., temperature and pressure) during recipe execution or idle can be obtained from each sensor as sensor waveform data. There are cases where one sensor is provided for one component, or where multiple sensors are provided for one component. Alternatively, there may be cases where one sensor is provided for multiple components.

[0012] (2) Anomaly detection device Figure 2A is a left-side partial view of the overall configuration diagram of a semiconductor manufacturing apparatus and an anomaly detection device for a semiconductor manufacturing apparatus according to one embodiment. Figure 2B is a right-side partial view of the overall configuration diagram of a semiconductor manufacturing apparatus and an anomaly detection device for a semiconductor manufacturing apparatus according to one embodiment. When the left side of Figure 2B is combined with the right side of Figure 2A, the overall configuration diagram of the anomaly detection device for a semiconductor manufacturing apparatus is obtained.

[0013] As shown in the overall configuration diagrams of the semiconductor manufacturing apparatus and the semiconductor manufacturing apparatus anomaly detection apparatus in Figures 2A and 2B, the anomaly detection apparatus 2 comprises a group of computers (computer 30, computer 40, ...) consisting of a processor 31 and a storage device 32, and a server 50. The processor 31 performs the acquisition of sensor waveform data and the calculation of feature quantities corresponding to each semiconductor manufacturing apparatus (10, 11) in the semiconductor manufacturing apparatus group 1. The storage device 32 stores information necessary for the processing of the processor 31 and sensor waveform data. The server 50 consists of an input / output interface 51 for specifying the target for anomaly detection and displaying the detection result, a processor 52 for processing anomaly detection, and a storage device 53 for storing information necessary for processing anomaly detection.

[0014] The semiconductor manufacturing equipment group 1 is connected to the computer group (computers 30, 40, ...) directly or via a wired or wireless network. The computer group (computers 30, 40, ...) and the server 50 are also connected via a wired or wireless network. As a result, each computer (30, 40) stores sensor waveform data acquired from each semiconductor manufacturing equipment (10, 11) in a storage device 32, and the processor 31 is configured to perform calculations using the sensor waveform data stored in the storage device 32 at high speed. The server 50 is configured to enable the display of analysis and diagnostic results across the semiconductor manufacturing equipment group 1.

[0015] As will be described later, the processor 31 includes a preprocessing unit 310 and a feature calculation unit 311.

[0016] As will be described later, the storage device 32 includes a sensor waveform storage unit 320 in which sensor waveform data is stored.

[0017] The input / output interface 51 includes an input interface (data input circuit) 510 and a judgment result display monitor (data output circuit) 511, as will be described later.

[0018] As described later, the processor 52 includes a measured monitoring statistics calculation unit 520, a measured year-on-year change calculation unit 521, a predicted monitoring statistics calculation unit 522, an anomaly presence / absence determination unit 523, a false determination risk calculation unit 524, a reference year-on-year change setting unit 525, a reference year-on-year change update unit 526, and a reference year-on-year change interpolation unit 527.

[0019] As will be described later, the storage device 53 includes a time-series sensor data storage unit 530, a reference aging change amount storage unit 531, and a diagnostic result storage unit 532.

[0020] (3) Creation of Time-Series Sensor Data Time-series sensor data is data that shows the changes in the feature quantities of the sensor waveform data for each recipe execution, and is created for each set of sensor, individual component, and recipe. In other words, time-series sensor data can be considered as time-series data that shows the changes in the state of components such as sensors and individual components. Referring to Figure 3, an example of the process of creating time-series sensor data from the semiconductor manufacturing apparatus 10 will be explained. Figure 3 shows the flow of processing for each recipe execution. The sensor waveform data acquired from the semiconductor manufacturing apparatus 10 is stored in the sensor waveform storage unit 320. Steps S1 to S5 will be explained in order below.

[0021] (First step S1: Acquisition of sensor waveform data) First, the processor 31 acquires sensor waveform data from one of the sensors in the previous recipe execution run from the sensor waveform storage unit 320. The sensor waveform data acquired from the semiconductor manufacturing equipment 10 is configured to store the sensor values ​​within the recipe in a time series such as processing time and time.

[0022] (Second step S2: Preprocessing sensor waveform data) Next, the preprocessing unit 310 performs preprocessing on the sensor waveform data of the selected sensor. The preprocessing method is arbitrary, but for example, it may involve extracting processing time intervals, removing missing values, or standardizing the data to make it easier to handle.

[0023] (Third step S3: Calculation of features) Next, the feature calculation unit 311 calculates features that represent the characteristics of the sensor waveform data using the pre-processed sensor waveform data. For example, statistical quantities such as the mean and standard deviation of the sensor waveform data can be used as features.

[0024] (Fourth step S4: Storing feature quantities) As shown in Figure 4 described later, the calculated feature quantities 49 are transmitted to the time-series sensor data storage unit 530 of the server 50, along with the following information: the device ID 41 for identifying the semiconductor manufacturing equipment, the start time of the recipe execution 42, the cumulative number of wafers processed by the semiconductor manufacturing equipment 43, the sensor ID 44 which is information for identifying the sensor, the individual component ID 45 which is information for identifying the individual component being observed, the recipe ID 46 which is information for identifying the recipe, the number of recipe executions 47 which is the number of times the recipe has been executed by the individual component, and the device status 48 (normal / abnormal) of the semiconductor manufacturing equipment during the recipe execution, and stored in the time-series sensor data storage unit 530.

[0025] The status of the semiconductor manufacturing equipment is determined based on the wafer inspection results after recipe execution. If there is a time gap between recipe execution and inspection, the status of the semiconductor manufacturing equipment may be initially set to "inspecting" and then updated to "normal" or "abnormal" after the inspection.

[0026] (Step 5 S5: Determining completion of calculations for all sensors) Next, it is determined whether the feature calculations have been completed for all sensors. If they have been completed (Yes), the processing for that recipe execution run ends.

[0027] If the calculation in step 5, S5, is not completed (No), the process moves to selecting another sensor and continues with the flow of processing for each recipe execution (S1-S5).

[0028] Figure 4 shows an example of the data structure of the time-series sensor data storage unit 530. In the example data structure shown in Figure 4, the following items are provided: device ID 41, execution start time 42, cumulative number of wafers processed 43, sensor ID 44, individual component ID 45, recipe ID 46, recipe execution count 47, device status 48, and feature quantity 49. In the data structure shown in Figure 4, for example, if data is extracted by specifying sensor ID 44, individual component ID 45, and recipe ID 46, time-series sensor data representing the change in feature quantity 49 for each recipe execution count 47 can be obtained.

[0029] (4) Determination of the presence or absence of abnormal changes and calculation of the risk of misjudgment Next, an example of time-series sensor data (TSSD) will be explained using Figures 5 and 6. Figure 5 is a diagram showing an example of time-series sensor data showing changes over time according to one embodiment. Figure 6 is a diagram showing an example of time-series sensor data showing abnormal changes according to one embodiment.

[0030] Time-series sensor data (TSSD) obtained from semiconductor manufacturing equipment (10, 11) shows age-related changes specific to each component type, regardless of equipment malfunctions. Figure 5 is a graph of time-series sensor data (TSSDA: feature quantity 49A, TSSDB: feature quantity 49B) observed by current sensors of different individual components (A, B) of the same component type. In the graph of Figure 5, the vertical axis represents feature quantities 49A and 49B, the horizontal axis represents the number of recipe executions (47), and the black circles ● represent the time-series sensor data TSSDA and TSSDB for feature quantities 49A and 49B.

[0031] As shown in the graphs for individual components (A and B) in Figure 5, current sensors exhibit a common aging pattern: electrical resistance increases with component degradation, and current increases due to feedback control for output maintenance. In this example, the time-series sensor data TSSDA graph for component A shows no temporary disturbances. On the other hand, the time-series sensor data TSSDB graph for component B shows a temporary disturbance BB.

[0032] Other factors include aging due to the accumulation of by-products during processing. However, initial values ​​of time-series sensor data (TSSD) may differ due to individual component variations. Furthermore, if mutual interference between sensors occurs, time-series sensor data (TSSD) may be temporarily disrupted even when the semiconductor manufacturing equipment is operating normally.

[0033] On the other hand, as shown in Figure 6, when a semiconductor manufacturing device malfunctions, the time-series sensor data (TSSD: feature quantity 49) related to the faulty component shows a sudden abnormal change 66 that differs from the aging change (see A in Figure 5) of the time-series sensor data (TSSDA: feature quantity 49A) observed by the current sensor of individual component A.

[0034] Server 50 processes time-series sensor data (TSSD) to determine whether there are any abnormal changes and to calculate the risk of false judgment. An example of this will be explained with reference to Figure 7. Figure 7 is a flowchart showing an example of the processing flow for determining the presence or absence of abnormalities and calculating the risk of false judgment. Steps T1 to T10 will be explained in order below.

[0035] (First step T1: Input of normal period and diagnostic period) First, a pair of normal period (Pn) and diagnostic period (Pa) to be processed are input from the input interface 510. The normal period (Pn) is a portion of the data contained in the time-series sensor data storage unit 530 during which the semiconductor manufacturing equipment 10 is in a normal state. The diagnostic period (Pa) is a portion of the data during which the semiconductor manufacturing equipment 10 is in an abnormal state. This normal period (Pn) and diagnostic period (Pa) may be specified by the recipe execution start time 42 or by the cumulative number of wafers processed 43.

[0036] (Second step T2: Acquisition of time-series sensor data) The processor 52 acquires time-series sensor data that has data in the specified normal period (Pn) and diagnostic period (Pa).

[0037] (Third step T3: Calculation of actual monitoring statistics) Next, the actual monitoring statistics calculation unit 520 calculates the actual monitoring statistics for each recipe execution run of the acquired time-series sensor data. An example of actual monitoring statistics is the level of the time-series sensor data. In this case, the value of the feature quantity of the time-series sensor data itself may be used as the actual monitoring statistics, or the average value of the neighboring period for each recipe execution run may be used as the actual monitoring statistics to reduce noise. In addition, variance or outlier ratio may be used as the actual monitoring statistics, in which case it can be calculated from the neighboring period for each recipe execution run.

[0038] (Fourth step T4: Calculation of measured aging change) Next, the measured aging change calculation unit 521 calculates the change in the measured monitoring statistic for each recipe execution during the normal period (Pn) (change in the measured monitoring statistic). This change is highly likely to be an aging change unrelated to any malfunctions observed during the normal operation of the semiconductor manufacturing equipment, and will be referred to here as the measured aging change. For example, if a level is used for the measured monitoring statistic, the change in that level for each recipe execution (trend of time-series sensor data) becomes the measured aging change. Here, the trend of the time-series sensor data (level) corresponding to the measured monitoring statistic means, for example, the change in the level, the slope of the level, the tendency of the change in the level, etc.

[0039] (Step 5 T5: Calculation of Predictive Monitoring Statistics) Next, the predictive monitoring statistics calculation unit 522 calculates predictive monitoring statistics for the diagnostic period (Pa) using the measured monitoring statistics and measured year-on-year change during the normal period (Pn). Predictive monitoring statistics are predictions of monitoring statistics during the diagnostic period (Pa) assuming that the time-series sensor data maintains the measured year-on-year change from the normal period (Pn) to the diagnostic period (Pa). As an example, first, the average value of the measured year-on-year change during the normal period (Pn) is calculated as the average year-on-year change per recipe execution. Subsequently, one recipe execution count within the normal period (Pn) and one recipe execution count within the diagnostic period (Pa) are selected as the calculation targets. The change obtained by multiplying the difference in the number of execution counts of these two recipes by the average year-on-year change is considered to be the change that would occur if the average year-on-year change was maintained. By adding this change to the measured monitoring statistics for the target recipe executions during the normal period (Pn), the predicted monitoring statistics for the target recipe executions during the diagnostic period (Pa) can be obtained. By repeating this calculation until all recipe executions in the diagnostic period (Pa) are included, the predicted monitoring statistics for the entire diagnostic period (Pa) can be obtained.

[0040] (Step 6 T6: Abnormality Determination) Next, the abnormality determination unit 523 compares the measured monitoring statistic and the predicted monitoring statistic corresponding to the diagnostic period (Pa) to determine whether there is an abnormal change. Based on this, the abnormality determination unit 523 generates a result of the abnormality determination (also called the abnormality determination result). For example, a two-sample t-test may be performed on the measured monitoring statistic and the predicted monitoring statistic, or a threshold may be set for the effect size representing the difference. Here, "two-sample t-test" means testing whether there is a difference in the means of samples drawn from two independent populations.

[0041] If a change is detected, it indicates that an abnormal change has occurred between the normal period (Pn) and the diagnostic period (Pa) that cannot be explained by the measured amount of change over time.

[0042] (Step 7 T7: Obtaining Reference Age Change Amount) Next, the processor 52 obtains the reference age change amount by specifying the sensor ID 44 and recipe ID 46 from the reference age change amount storage unit 531. Here, the reference age change amount can be set, for example, based on the measured age change amount observed during the normal operation of the device in the past.

[0043] (Step 8 T8: Calculation of misjudgment risk) The misjudgment risk calculation unit 524 then calculates the misjudgment risk of the abnormality determination result based on a comparison between the reference aging change amount and the measured aging change amount. The reference aging change amount defines the range that the aging change amount can take. The calculation method will be described later, but it is also possible to define upper and lower limits that the aging change amount can take, or to assume a probability distribution for the aging change amount and define parameters that define that distribution. In addition, the reference aging change amount may be constant for each pair of sensor ID 44 and recipe ID 46, or it may change according to the number of recipe executions 47.

[0044] Hereafter, this embodiment will deal with the case where the amount of change over time follows a Gaussian distribution, and its parameters (mean μ and standard deviation σ) change every 47 recipe executions. In this case, at each recipe execution, the measured amount of change over time is compared with the parameters of the reference amount of change over time, and if it deviates from the 3σ interval, there is a risk of misjudgment; if it is within the 3σ interval, there is no risk of misjudgment.

[0045] FIG. 8 is a diagram illustrating an example of the calculation results of the abnormality presence / absence determination and the misjudgment risk for time-series sensor data according to an embodiment. FIG. 8 shows an example of a case where (abnormal change: absent, misjudgment risk: present). In FIG. 8, the vertical axis represents the monitored statistic 81, the horizontal axis represents the number of recipe executions 47, black dots ● represent the actually measured monitored statistic, and white dots ○ represent the predicted monitored statistic. BB indicates temporary disturbance, and 66 indicates a sudden abnormal change.

[0046] For example, in the time-series sensor data (TSSD) of FIG. 8, when the monitored statistic 81 is set as the level, the actually measured secular variation is underestimated due to the temporary disturbance BB in the normal period (Pn), so that the predicted monitored statistic (○) is similar to the actually measured monitored statistic (●) in the diagnosis period (Pa), resulting in a misjudgment of "no abnormal change" in the sixth step T6 (abnormality presence / absence determination). In the eighth step T8 (calculation of misjudgment risk), the actually measured secular degradation amount used to calculate this predicted monitored statistic (○) deviates downward from the 3σ interval, so that this abnormality presence / absence determination result can be calculated as having "misjudgment risk". As such, the misjudgment risk calculation unit 524 is configured to be capable of calculating the misjudgment risk of the abnormality presence / absence determination result as a discrete value or a continuous value.

[0047] (Ninth Step T9: Store Abnormality Determination Result and Misjudgment Risk) Thereafter, the abnormality presence / absence determination result and the misjudgment risk are stored in the diagnosis result storage unit 532, together with the apparatus ID (41), sensor ID (44), component individual ID (45), recipe ID (46) for identifying the time-series sensor data (TSSD), as well as the normal period (Pn) and the diagnosis period (Pa).

[0048] (Tenth Step T10: Determination of Completion of Calculation for All Sensors) The processor 52 determines whether the calculation of abnormality presence / absence judgment and misjudgment risk has been completed for all time-series sensor data (TSSD) that have data in the specified normal period (Pn) and diagnosis period (Pa). If the calculation is completed (Yes), the process is terminated. Regarding the tenth step T10, if the processing of the calculation of abnormality presence / absence judgment and misjudgment risk has not been completed for all time-series sensor data (No), the process proceeds to select another uncompleted time-series sensor data, and continues the processing of determining the presence or absence of abnormal change and calculating the misjudgment risk (T1-T9) for the other uncompleted time-series sensor data.

[0049] After all processing is completed (T10: Yes), the time-series sensor data, the abnormality presence / absence judgment result, and the misjudgment risk are displayed on the judgment result display monitor 511.

[0050] That is, the first embodiment can be summarized as follows.

[0051] The data input circuit 510 receives time-series sensor data indicating the state transition of the component, at least one normal period (Pn) and at least one diagnosis period (Pa) among the periods of the time-series sensor data.

[0052] The processor 52 calculates the statistic for abnormality judgment target from the time-series sensor data, and calculates the actually measured monitoring statistic. In addition, the processor 52 calculates the statistic variation caused by the aged deterioration of the component from the time-series sensor data, and calculates the actually measured aged variation.

[0053] The storage device 53 stores a reference aged variation that specifies the possible range of the actually measured aged variation by the processor 52.

[0054] The processor 52 uses the actually measured aged variation corresponding to the normal period (Pn) to calculate the predicted monitoring statistic corresponding to the diagnosis period (Pa).

[0055] The processor 52 compares the actually measured monitoring statistic corresponding to the diagnosis period (Pa) with the predicted monitoring statistic corresponding to the diagnosis period (Pa), and determines the presence or absence of abnormality.

[0056] The processor 52 compares the measured aging change amount corresponding to the normal period (Pn) with the reference aging change amount corresponding to the normal period (Pn) and calculates the risk of misjudgment in determining whether an anomaly exists. The monitor 511, which is a data output circuit, outputs the anomaly determination result and the risk of misjudgment.

[0057] Figure 9 shows an example of the display screen in the judgment result display monitor 511 according to one embodiment. An example of the display screen DSP in the judgment result display monitor 511 is shown in Figure 9. In the example in Figure 9, the following is shown.

[0058] 1) Judgment conditions JC: Device ID 41: R1, Recipe ID 46: X, Normal period Pn: 5000-12000, Diagnostic period Pa: 22000-29000. Here, the values ​​for the normal period Pn and the diagnostic period Pa are the values ​​for the cumulative number of wafers processed 43, and the period is expressed by the value of the cumulative number of wafers processed 43.

[0059] 2) Sensor ID 44: Current sensor A: ISA, Current sensor B: ISB, Pressure sensor C: PCC.

[0060] 3) Graphs of time-series sensor data TSSD for each sensor (current sensor A, current sensor B, pressure sensor C): Current sensor A: TSSDA, Current sensor B: TSSDB, Pressure sensor C: TSSDC.

[0061] 4) Abnormal change 91: Abnormal change status of each sensor (current sensor A, current sensor B, pressure sensor C): Current sensor A: present (YES), Current sensor B: absent (NO), Pressure sensor C: absent (NO).

[0062] 5) Misjudgment Risk 92: Status of misjudgment risk for each sensor (current sensor A, current sensor B, pressure sensor C): Current sensor A: None (NO), Current sensor B: Present (YES), Pressure sensor C: None (NO).

[0063] Semiconductor manufacturing equipment is equipped with numerous sensors, and the time-series sensor data (TSSD) obtained from them is enormous. When a malfunction occurs in the semiconductor manufacturing equipment, the time-series sensor data (TSSD) that is judged to have an abnormal change (YES) or no abnormal change (NO) but with a risk of misjudgment (YES) during the specified normal period Pn and diagnostic period Pa is displayed at the top of the judgment result display monitor 511 (corresponding in this case to current sensor A and current sensor B), allowing operators to prioritize the investigation of components related to these sensors and improve the efficiency of identifying the causative component.

[0064] (5) Determination of Reference Amount of Change Over Time As described above, the reference amount of change over time defines the range that the amount of change over time can take. Figure 10 is a flowchart showing an example of the processing flow for calculating the reference amount of change over time according to one embodiment. In this embodiment, a method for calculating the reference amount of change over time based on time-series sensor data obtained during the normal operation of a semiconductor manufacturing apparatus will be explained with reference to Figure 10. In other words, the reference amount of change over time can be set, for example, based on the measured amount of change over time observed during the normal operation of a semiconductor manufacturing apparatus in the past. However, the calculation method is not limited to this embodiment, and for example, even if time-series sensor data is not obtained in advance, it is possible to determine it based on domain knowledge. Steps U1 to U6 will be explained in order below.

[0065] (First step U1: Input of sensor ID and recipe ID) First, the input interface 510 specifies the sensor ID 44 and recipe ID 46 for which the reference aging amount will be calculated.

[0066] (Second step U2: Acquisition of time-series sensor data) The processor 52 acquires a group of time-series sensor data from the time-series sensor data storage unit 530 that matches the specified sensor ID 44 and recipe ID 46, and covers a period in which the semiconductor manufacturing equipment is in a normal state. The group of time-series sensor data consists of multiple time-series sensor data (TSSD) associated with different individual component IDs 45.

[0067] (Third step U3: Calculation of actual monitoring statistics) Next, the actual monitoring statistics calculation unit 520 calculates actual monitoring statistics for each time-series sensor data (TSSD).

[0068] (Fourth step U4: Calculation of measured year-on-year change) Next, the measured year-on-year change calculation unit 521 calculates the measured year-on-year change for each measured monitoring statistic. The measured year-on-year change is, for example, the trend (slope) of the measured monitoring statistic.

[0069] (Step 5 U5: Calculation of Reference Age Change Amount) Next, the Reference Age Change Amount Setting Unit 525 calculates the reference age change amount based on multiple measured age change amounts obtained from the time-series sensor data group. Specifically, for each recipe execution run included in the time-series sensor data group, the average (average μ115) and standard deviation (standard deviation σ116) of the measured age change amounts are calculated as the reference age change amount.

[0070] (Step 6 U6: Storing the reference aging change amount) The calculated reference aging change amounts (115, 116) are stored in the reference aging change amount storage unit 531 along with the sensor ID 44, recipe ID 46, recipe execution count 47, and sample number 114, which is the number of measured aging change amounts used in the calculation.

[0071] An example of the data structure of the reference aging change amount storage unit 531 obtained by this process is shown in Figure 11. As shown in Figure 11, the items in the data structure of the reference aging change amount storage unit 531 include sensor ID 44, recipe ID 46, recipe execution count 47, sample size 114, mean μ 115, and standard deviation σ 116.

[0072] By specifying the pair of sensor ID 44 and recipe ID 46, the trend of the reference aging change amount (115, 116) in the time-series sensor data TSSD can be seen for every 47 recipe executions. Since the reference aging change amount (115, 116) is calculated from multiple measured monitoring statistics during the normal operation of the device, the influence of disturbances that may be included in the normal period Pn is suppressed. If the sample size 114 is small and there is a risk that the influence of disturbances will remain, interpolation of the reference aging change amount (115, 116) as described later may be performed, or it may be replaced with an initial value based on domain knowledge.

[0073] The reference aging change amount storage unit 531 can be updated sequentially in accordance with the update of the time-series sensor data storage unit 530. In this case, first, the processor 52 acquires time-series sensor data (TSSD) indicating that the device state is normal, which has been newly stored in the time-series sensor data storage unit 530. Next, it performs processing (third step U3 to fourth step U4) on this time-series sensor data (TSSD) to calculate the measured aging change amount for each recipe execution.

[0074] Next, the processor 52 specifies the sensor ID 44, recipe ID 46, and recipe execution number associated with the calculated measured aging change amount, and retrieves the corresponding reference aging change amount (115, 116) and sample size 114 from the reference aging change amount storage unit 531. The reference aging change amount update unit 526 sequentially updates the reference aging change amount (mean and sample variance) using the reference aging change amount (115, 116), sample size 114, and the newly calculated measured aging change amount. It also increments the sample size 114 by the number of measured aging change amounts newly used in the update. The updated reference aging change amount (115, 116) and sample size 114 are stored in the reference aging change amount storage unit 531.

[0075] (6) Interpolation of Reference Age Changes In the aforementioned abnormality determination and misjudgment risk calculation process (7th step T7), the processor 52 obtains the reference age change amounts (115, 116) by specifying a pair of sensor ID 44 and recipe ID 46 from the reference age change amount storage unit 531. As in this embodiment, when the reference age change amounts (115, 116) change according to the number of recipe executions 47 and are calculated from past time-series sensor data (TSSD), the reference age change amounts (115, 116) for the recipe execution count that you want to compare with the measured age change amount may be undefined. Also, even if it is defined, the number of samples 114 used in the calculation may be small, and the value may not be reliable.

[0076] To address this, the reference aging change amount interpolation unit 527 can be executed in conjunction with the aforementioned process (7th step T7) to interpolate the reference aging change amounts (115, 116).

[0077] Specifically, the reference aging change interpolation unit 527 performs interpolation calculations on the reference aging change (115, 116) that the processor 52 wants to obtain if the reference aging change (115, 116) is undefined or the number of samples 114 is less than the specified required number of samples, and returns the result to the processor 52. If the number of samples is equal to or greater than the required number, the reference aging change (115, 116) is returned to the processor 52 as is.

[0078] One example of interpolation is to obtain two reference aging change values ​​(115, 116) with a sample size greater than or equal to the minimum sample size in the vicinity of the recipe execution time to be interpolated, and then perform linear interpolation from these two points. The misjudgment risk calculation unit 524 can use the interpolated reference aging change values ​​(115, 116) to perform processing even in recipe execution times where the reference aging change values ​​(115, 116) are undefined or the sample size 114 is small. In other words, if the value corresponding to the normal period of the reference aging change value is undefined, the unit interpolates the value of the reference aging change value corresponding to the normal period based on the reference aging change values ​​of other periods that are set (in this case, for example, the period in the vicinity of the recipe execution time to be interpolated).

[0079] In Example 1, "(4) Determination of the presence or absence of abnormal changes and calculation of the risk of misjudgment," a method for presenting accurate abnormality judgment results was shown by calculating the risk of misjudgment of the presence or absence of abnormalities based on the reference aging change amount (115, 116). Example 2 describes a method for presenting accurate abnormality judgment results by correcting the result of the abnormality judgment itself. Example 2 will be described below.

[0080] Figure 12 is a flowchart showing an example of the processing flow for the calculation of whether or not an abnormality is present according to Embodiment 2. Steps T11-T14 and T20 in Figure 12 are the same steps as steps T1-T4 and T10 in Figure 7, and the explanation of T11-T14 and T20 can be found by referring to the operation of steps T1-T4 and T10.

[0081] The determination of whether or not there is an abnormality in Example 2 is carried out as follows.

[0082] The following steps are performed in order: Step 1 T11 (input of normal period and diagnostic period), Step 2 T12 (acquisition of time-series sensor data), Step 3 T13 (calculation of measured monitoring statistics), and Step 4 T14 (calculation of measured year-on-year change).

[0083] (5th step T15: Calculation of predictive monitoring statistics) After the 4th step T14, the predictive monitoring statistics calculation unit 522 calculates the predictive monitoring statistics for the diagnostic period Pa using the measured monitoring statistics and measured year-on-year changes during the normal period Pn.

[0084] (Step 6 T16: Abnormality Determination) Next, the abnormality determination unit 523 compares the measured monitoring statistics and the predicted monitoring statistics corresponding to the diagnostic period Pa to determine whether there is an abnormal change. Based on this, the abnormality determination unit 523 generates an abnormality determination result (also called the abnormality determination result).

[0085] (Step 7 T17: Determination of whether to correct the abnormality detection result) Next, it is determined whether the abnormality detection result obtained in Step 6 T16 needs to be corrected. Specifically, if the measured aging change amount calculated in Step 4 T14 (equivalent to Step 4 T4 in Figure 7) is within the range of the reference aging change amount (115, 116), it is determined that no correction to the abnormality detection result is necessary (No). If the measured aging change amount calculated in Step 4 T14 is outside the range of the reference aging change amount (115, 116), it is determined that correction to the abnormality detection result is necessary (Yes).

[0086] If it is determined that no correction is needed to the abnormality detection result (No), the abnormality detection result is stored in the diagnostic result storage unit 532, and the process proceeds to the tenth step T20 (equivalent to the tenth step T10 in Figure 7), which will be described later.

[0087] On the other hand, if it is determined that correction of the abnormality determination result is necessary (Yes), the process transitions to step 5 T15 in order to re-execute the calculation of the predictive monitoring statistics in step 5 T15.

[0088] (Re-execution of step 5 T15) Next, the calculation of the predictive monitoring statistics is redone (i.e., the calculation of the predictive monitoring statistics is performed again). At this time, it is preferable to modify the average year-on-year change used in the calculation of the predictive monitoring statistics so that it is calculated using the reference year-on-year change (115, 116) instead of the measured year-on-year change.

[0089] In other words, if the measured annual change amount corresponding to the normal period Pn is outside the range of the reference annual change amount (115, 116), the predictive monitoring statistic corresponding to the diagnostic period Pa is calculated using the reference annual change amount (115, 116). Therefore, if the measured annual change amount corresponding to the normal period Pn is within the range of the reference annual change amount (115, 116), the predictive monitoring statistic corresponding to the diagnostic period Pa is calculated using the measured annual change amount.

[0090] (Re-execution of step 6 T16) Next, based on the corrected predicted monitoring statistics, the abnormality determination in step T6 is performed again to generate a corrected abnormality determination result. Then, the corrected abnormality determination result is stored in the diagnostic result storage unit 532. After that, step 7 T17 is executed, but since further correction of the corrected abnormality determination result is not necessary, the process proceeds to step T20.

[0091] (Step 10 T20: Determination of Completion of Calculations for All Sensors) The processor 52 determines whether the abnormality detection and false-detection risk calculations have been completed for all time-series sensor data (TSSD) that have data in the specified normal period Pn and diagnostic period Pa. If the calculations are completed (Yes), the process ends. If the process is not completed in Step 10 T20 (No), the process moves on to selecting other time-series sensor data and continues the process of determining whether there are abnormal changes in the other time-series sensor data (T11-T17).

[0092] In other words, Example 2 can be summarized as follows.

[0093] The data input circuit 510 receives time-series sensor data showing the state changes of the components, and at least one set of normal period Pn and diagnostic period Pa from the period of the time-series sensor data.

[0094] The processor 52 calculates statistical quantities for determining anomalies from time-series sensor data and calculates actual monitoring statistics.

[0095] The processor 52 calculates the amount of change in statistical quantities caused by the aging of components from time-series sensor data, and calculates the measured amount of aging change.

[0096] The memory device 53 stores a reference aging rate that specifies the range of possible measured aging rates, as determined by the processor 52. This reference aging rate is set based on measured aging rates observed during past normal operation of the device.

[0097] The processor 52 uses the measured amount of change over time and the reference amount of change over time corresponding to the normal period Pn to calculate a predictive monitoring statistic corresponding to the diagnostic period Pa.

[0098] The processor 52 compares the predicted monitoring statistics with the measured monitoring statistics corresponding to the diagnostic period Pa to determine whether or not there is an abnormality. The monitor 511, which is a data output circuit, outputs the result of the abnormality determination.

[0099] When determining whether an abnormality exists, the processor 52 calculates a predictive monitoring statistic for the diagnostic period Pa using the measured annual change amount if the measured annual change amount corresponding to the normal period Pn is within the range of the reference annual change amount. Furthermore, if the measured annual change amount corresponding to the normal period Pn is outside the range of the reference annual change amount, the processor 52 calculates a predictive monitoring statistic for the diagnostic period Pa using the reference annual change amount.

[0100] According to Example 2, if the measured aging change amount calculated in the fourth step T14 is outside the range of the reference aging change amount (115, 116), the result of the abnormality determination itself is corrected, so that an accurate abnormality determination result can be provided. As a result, an accurate abnormality determination result can be obtained in the abnormality determination of time-series sensor data showing aging changes.

[0101] The disclosure made by the Discloser has been described in detail based on the examples above, but it goes without saying that the disclosure is not limited to the above examples and can be modified in various ways.

[0102] 1: Semiconductor manufacturing equipment group 10: Semiconductor manufacturing equipment A 11: Semiconductor manufacturing equipment B 2: Anomaly detection device 30: Computer A 31: Processor A 32: Storage device A 40: Computer B 50: Server 51: Input / output interface 510: Data input circuit (input interface) 511: Data output circuit (monitor) 52: Processor 53: Storage device.

Claims

1. An abnormality detection device comprising a data input circuit, a data output circuit, a memory device, and a processor for determining abnormalities in components of a semiconductor manufacturing apparatus, wherein the data input circuit receives time-series sensor data indicating the state transition of the component, and at least one set of normal period and diagnostic period from the period of the time-series sensor data, the processor calculates a statistical quantity of the component subject to abnormality detection from the time-series sensor data and calculates an actual monitoring statistical quantity, calculates a change in a statistical quantity caused by the aging deterioration of the component from the time-series sensor data and calculates an actual aging change quantity, the memory device stores a reference aging change quantity that specifies the range that the actual aging change quantity can take, the processor calculates a predicted monitoring statistical quantity corresponding to the diagnostic period using the actual aging change quantity corresponding to the normal period, compares the actual monitoring statistical quantity corresponding to the diagnostic period with the predicted monitoring statistical quantity corresponding to the diagnostic period and determines whether or not there is an abnormality, An abnormality detection device characterized by comparing the measured amount of aging corresponding to the normal period with the reference amount of aging corresponding to the normal period, calculating the risk of misjudgment in determining the presence or absence of an abnormality, and the data output circuit outputting the result of the determination of the presence or absence of an abnormality and the risk of misjudgment.

2. An anomaly detection device according to claim 1, characterized in that the level of the time-series sensor data is used as the measured monitoring statistic, and the trend of the time-series sensor data is used as the measured amount of change over time.

3. An anomaly determination device according to claim 1, characterized in that the reference aging amount is set based on the measured aging amount observed during the normal operation of a semiconductor manufacturing device in the past.

4. An anomaly determination device according to claim 3, characterized in that, when the value corresponding to the normal period of the reference aging amount is undefined, the device interpolates the value corresponding to the normal period based on the reference aging amount of another set period.

5. An abnormality detection device comprising a data input circuit, a data output circuit, a memory device, and a processor for determining abnormalities in components of a semiconductor manufacturing apparatus, wherein the data input circuit receives time-series sensor data indicating the state transition of the component, and at least one set of normal period and diagnostic period from the period of the time-series sensor data; the processor calculates a statistical quantity to be determined for abnormality from the time-series sensor data and calculates an actual monitoring statistical quantity; the processor calculates a change in a statistical quantity caused by the aging deterioration of the component from the time-series sensor data and calculates an actual aging change quantity; the memory device stores a reference aging change quantity that specifies the range that the actual aging change quantity can take; the processor calculates a predicted monitoring statistical quantity corresponding to the diagnostic period using the actual aging change quantity and the reference aging change quantity corresponding to the normal period; the processor compares the predicted monitoring statistical quantity with the actual monitoring statistical quantity corresponding to the diagnostic period and determines whether there is an abnormality; and the data output circuit outputs the result of the determination of whether there is an abnormality.

6. An anomaly determination device according to claim 5, characterized in that, in determining the presence or absence of an anomaly, if the measured amount of change over time corresponding to the normal period is within the range of the reference amount of change over time, the device calculates a predictive monitoring statistic corresponding to the diagnostic period using the measured amount of change over time; and if the measured amount of change over time corresponding to the normal period is outside the range of the reference amount of change over time, the device calculates a predictive monitoring statistic corresponding to the diagnostic period using the reference amount of change over time.