Abnormality detection device, abnormality detection system and abnormality detection method

Abstract

The embodiments of the present disclosure relate to an abnormality detection device, an abnormality detection system, and an abnormality detection method. An abnormality detection device includes: a first extraction unit, which extracts input waveform data from waveform data input to the first extraction unit; a first determination unit, which is used to determine whether the input waveform data includes detection object waveform data; a second extraction unit, which is used to extract and output the detection object waveform data from the input waveform data when the first determination unit determines that the input waveform data includes the detection object waveform data; and a second determination unit, which is used to determine whether the detection target device has an abnormality based on whether the detection object waveform output from the second extraction unit indicates an abnormality.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] The disclosure of Japanese Patent Application No. 2019-025169 filed on February 15, 2019 including the specification, drawings and abstract is incorporated herein by reference in its entirety. Technical Field

[0003] The present invention relates to an abnormality detection device. For example, the present invention is applicable to an abnormality detection device for determining whether a detection target device is abnormal using a detection object waveform included in time series waveform data output from the detection target device. Background Art

[0004] In a manufacturing site such as a factory, abnormality of equipment such as a manufacturing process or an operating state is detected by monitoring the waveform of sensor data of a sensor installed in the equipment to be monitored.

[0005] However, in a manufacturing site, it is required not to continuously monitor the waveform of sensor data of a device to be monitored, but to monitor the waveform of sensor data only when the device is performing the operation to be monitored.

[0006] Therefore, in the manufacturing site, the waveform when the operation to be monitored is performed is extracted from the waveform of the sensor data of the equipment to be monitored as the detection target waveform. In order to detect abnormalities more accurately, it is important to use the extracted detection target waveform to detect abnormalities of the equipment.

[0007] Japanese Unexamined Patent Application Publication No. 2010-038884 (hereinafter referred to as "Patent Document 1") discloses a related technology for extracting a detection target waveform from input waveform data. According to Patent Document 1, it is detected whether there is a portion (trigger point) in the input waveform data that satisfies the trigger condition, and if there is a portion that satisfies the trigger condition, the waveform of the portion is extracted as the detection target waveform. In addition, according to Patent Document 1, the trigger condition is set based on, for example, rising edge / falling edge, setup time / hold time violation, runt, transition, and pulse width. Summary of the invention

[0008] However, in the technology disclosed in Patent Document 1, when a waveform satisfying the trigger condition is mixed in input waveform data (even if the detection target waveform is not included in the input waveform), the waveform is mistakenly recognized as the detection target waveform.

[0009] Other objects and novel features will become apparent from the description of this specification and the accompanying drawings.

[0010] According to one embodiment, anomaly detection determines whether the input time series data includes data to be detected. When it is determined that the data to be detected is included, the detection object time series data is identified from the input time series data, and it is determined whether the detection object time series data is data indicating anomaly.

[0011] According to the above-mentioned embodiments, it can be helpful to solve the above-mentioned problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a diagram showing an exemplary configuration of an abnormality detection device according to the first embodiment.

[0013] Figure 2 : is a diagram showing an example of a detection target waveform related to the first embodiment.

[0014] Figure 3 : is a diagram showing an example of a detection target waveform related to the first embodiment.

[0015] Figure 4 is a flowchart showing an exemplary flow of an abnormality detection process according to the first embodiment.

[0016] Figure 5 is a diagram illustrating an exemplary method for generating learning data for the object waveform detection algorithm AL[1] according to the first embodiment.

[0017] Figure 6 is a flowchart showing an exemplary operation flow of the input buffer according to the first embodiment.

[0018] Figure 7 is an image diagram showing an example of the relationship among N1, M, and L related to the first embodiment.

[0019] Figure 8 is a flowchart showing an exemplary operation flow of the waveform extraction unit according to the first embodiment.

[0020] Fig. 9 is a diagram showing an exemplary waveform data thinning process performed by the waveform extraction unit according to the first embodiment.

[0021] Fig.10 is a flowchart showing an exemplary operation flow of the waveform determination unit according to the first embodiment.

[0022] Fig.11 is a flowchart showing an exemplary operation flow of the waveform extraction unit according to the first embodiment.

[0023] Fig.121 is an image diagram showing an exemplary process of extracting a detection object waveform by the waveform extraction unit according to the first embodiment.

[0024] Fig.13 is a flowchart showing an exemplary operation flow of the waveform determination unit according to the first embodiment.

[0025] Fig.14 is a diagram showing an exemplary autoencoder for a waveform determination unit according to the first embodiment.

[0026] Fig.15 is a diagram showing an exemplary operation of the waveform determination unit according to the second embodiment.

[0027] Fig.16 : is a flowchart showing an exemplary flow of an abnormality detection process performed by the entire abnormality detection device related to the third embodiment.

[0028] Fig.17 is a diagram showing an exemplary configuration of an abnormality detection system according to a fourth embodiment.

[0029] Fig.18 is a diagram showing an exemplary configuration of an abnormality detection device related to the fifth embodiment.

[0030] Fig.19 1 is an image diagram showing an exemplary process of extracting a detection object waveform by a waveform extraction unit according to the fifth embodiment.

[0031] Fig. 20 is a diagram showing an exemplary configuration of an abnormality detection device according to a sixth embodiment.

[0032] Fig.21 104 is an image diagram showing an example of a determination process of the waveform determination units 104X and 104Y according to the sixth embodiment.

[0033] Fig. 22 : is a diagram conceptually showing an exemplary configuration of abnormality detection devices of the first to third embodiments, the fifth embodiment, and the sixth embodiment.

[0034] Fig.23 is a diagram conceptually showing an exemplary configuration of the abnormality detection system of the fourth embodiment.

[0035] Fig.24 is a diagram conceptually showing another exemplary configuration of the abnormality detection apparatus of the first to third embodiments, the fifth embodiment, and the sixth embodiment. DETAILED DESCRIPTION

[0036] For the sake of clarity of description, the following description and drawings are appropriately omitted and simplified. In the drawings, the same elements are represented by the same reference numerals, and repeated descriptions are omitted as needed.

[0037] In addition, the elements described as function blocks for performing various processes in the drawings may be configured by hardware such as a CPU (central processing unit), a memory, and other circuits, and may be implemented by software such as a program loaded into the memory. Therefore, it should be understood by those skilled in the art that these function blocks may be implemented in various forms by separate hardware, separate software, or a combination thereof, and the present invention is not limited to any one of them. In the drawings, the same elements are represented by the same reference numerals, and their repeated descriptions are omitted as required.

[0038] In addition, various types of non-transitory computer-readable media can be used to store the program described above and provide the program to the computer. Non-transitory computer-readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (e.g., floppy disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks, CD-ROMs (Compact Disc Read Only Memory)), CD-Rs (CD-Recordable), CD-R / Ws (CD-Rewritable), solid-state memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs, Flash ROMs, RAMs (Random Access Memory)). The program can also be supplied to the computer via various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can provide programs to the computer via wired or wireless communication paths (such as wires and optical fibers).

[0039] First embodiment

[0040] Configuration of the First Embodiment

[0041] Figure 1 is a diagram showing an exemplary configuration of an abnormality detection device according to the first embodiment. Figure 1 Main components of the abnormality detection device are shown, and other components are omitted.

[0042] It is assumed that the detection target device for detecting an abnormality by the abnormality detection device according to the first embodiment is a manufacturing device in a manufacturing system for manufacturing semiconductor devices. However, the present embodiment is not limited to this, and the abnormality detection device according to the first embodiment can be applied as a device for detecting an abnormality in various manufacturing devices in various manufacturing systems.

[0043] The abnormality detection device according to the first embodiment detects abnormality of a manufacturing device based on a detection target waveform included in time series waveform data output from the manufacturing device. Assume that there is only one detection target waveform. Figure 2 and Figure 3 : is a diagram showing an example of a detection target waveform according to the present first embodiment. Figure 2 The waveform data including the non-detection target waveform and the detection target waveform (normal waveform) are shown, and Figure 3 Waveform data including a non-detection target waveform and a detection target waveform (abnormal waveform) are shown.

[0044] like Figure 1 As shown, the abnormality detection device according to the first embodiment includes a signal input unit 101, an input buffer 102, a waveform extraction unit 103, an output selection unit 108, and a waveform determination unit 104, a waveform extraction unit 105, an output selection unit 109, a waveform determination unit 106, and a control unit 107. The waveform extraction unit 103 is an example of a first waveform extraction unit, the waveform determination unit 104 is an example of a first waveform determination unit, the waveform extraction unit 105 is an example of a second waveform extraction unit, and the waveform determination unit 106 is an example of a second waveform determination unit. In the drawings described below, the waveform extraction units 103 and 105 are also respectively represented as waveform extraction units #1 and #2, and the waveform determination units 104 and 106 are also respectively represented as waveform determination units #1 and #2. The output selection units 108 and 109 are also respectively represented as output selection units #1 and #2.

[0045] The signal input unit 101 receives a monitoring signal from a detection target device (a manufacturing device for semiconductor devices in this first embodiment), and performs a predetermined signal process on the input monitoring signal through various filtering circuits, analog-to-digital converters, etc., thereby generating time series waveform data. The signal input unit 101 outputs the generated waveform data to the input buffer 102. The monitoring signal is a signal representing the state of the process of the manufacturing device. The monitoring signal is, for example, a sensor signal from a variety of sensors set in the manufacturing device, or a sensor signal from a variety of sensors added to the manufacturing device. The monitoring signal can be an analog signal, or it can be a digital signal input to various communication devices. The various sensors can be, for example, a flow rate sensor for monitoring the flow rate of a gas, a pressure sensor for monitoring the chamber pressure, a power sensor for monitoring the RF power of a plasma, an EPD (end point detector) for monitoring the progress of etching, etc., but they can be other sensors. The waveform data is described as one system, but it can be multiple systems.

[0046] The input buffer 102 includes a circular buffer (not shown) and an output part (not shown), and after accumulating waveform data within a predetermined period of time using the circular buffer, the time series waveform data output from the signal input unit 101 is copied to the output part. When N1 points of the waveform data are accumulated in the output part, the input buffer 102 outputs the waveform data consisting of N1 points to the waveform extraction units 103 and 105. The waveform data accumulated in the output part is shifted by the number of points M. Thereafter, the input buffer 102 repeats this operation every time N1 points of the waveform data are accumulated in the output part. Here, when the number of points of the detection object waveform is L, the relationship between N1, M and L is N1≥L+M. For example, N1=2*L, M=L.

[0047] The waveform extraction unit 103 receives the waveform data composed of N1 points output from the input buffer 102, extracts the waveform data when the waveform determination unit 104 becomes a data input waiting state, and outputs the waveform data to the output selection unit 108. Depending on the shape of the detection object waveform or the non-detection object waveform, the waveform extraction unit 103 can sparse or compress the waveform data composed of N1 points into waveform data composed of N2 (N1>N2, for example, N2=N1 / 10) points, and the waveform data composed of N2 points has a coarse granularity by, for example, setting the average value of every 10 points as one data. Thus, the amount of calculation of the waveform determination unit 104 in the subsequent stage can be reduced, and the calculation time can be shortened.

[0048] The output selection unit 108 outputs the waveform data output from the waveform extraction unit 103 to the waveform determination unit 104. However, when the waveform data is used to generate the object waveform detection algorithm AL[1] used in the waveform determination unit 104, the output selection unit 108 outputs the waveform data output from the waveform extraction unit 103 to the outside of the device. Outside the device, learning data described later is generated by using the waveform data including the detection object waveform (which may be a normal waveform or an abnormal waveform) in the waveform data output from the waveform extraction unit 103. The object waveform detection algorithm AL[1] is generated using the generated learning data. The output selection unit 108 switches the output destination based on the instruction of the control unit 107.

[0049] The waveform determination unit 104 determines whether the detection target waveform is included in the waveform data output from the waveform extraction unit 103 using the target waveform detection algorithm AL[1] based on machine learning learned in advance. The detection result of the detection target waveform is output to the waveform extraction unit 105.

[0050] The detection object waveform is usually specified based on the knowledge of the engineer who manages the manufacturing equipment. However, when the engineer specifies the detection object waveform, it is necessary to manually extract and collect a large amount of learning data for generating the object waveform detection algorithm AL[1]. However, collecting such learning data is very labor-intensive and time-consuming.

[0051] Therefore, in the first embodiment, a small amount of waveform data (at least one waveform data set) obtained by manually cutting out only the detection object waveform from the waveform data output from the waveform extraction unit 103 is prepared outside the device. A program executed on a PC (personal computer) automatically generates waveform data obtained by randomly shifting the cut detection object waveform in the time axis direction within a range of approximately twice the number of points of the detection object waveform, and sets the waveform data as learning data for the target waveform detection algorithm AL[1]. When there is a change in the amplitude direction and the time axis direction of the detection object waveform, waveform data obtained by randomly expanding and contracting the detection object waveform in the amplitude direction and the time axis direction respectively within an assumed normal range is generated, and this is also used as learning data for the target waveform detection algorithm AL[1]. Then, outside the device, the generated waveform data set is used as learning data for generating the object waveform detection algorithm AL[1].

[0052] The object waveform detection algorithm AL[1] can be, for example, an algorithm using an autoencoder which is a type of NN technology in neural network (NN) technology. The autoencoder has an input layer, one or more intermediate layers, and an output layer having the same number of nodes as the input layer. The autoencoder is a method of learning feature points compressed using an error back propagation method so that input data and output data are equal to each other. Therefore, when waveform data included in the learning data is input to the input layer, waveform data similar to the waveform data input to the input layer is reproduced on the output layer. On the contrary, when waveform data having components not included in the learning data is input to the input layer, waveform data similar to the input waveform data cannot be reproduced, and therefore the waveform data differs between the input and the output.

[0053] The object waveform detection algorithm AL[1] also determines whether the degree of deviation between the input waveform data and the output waveform data of the autoencoder is equal to or less than a predetermined first degree of deviation threshold, and determines whether the waveform data includes the detection object waveform based on the determination result. Here, if the degree of deviation is equal to or less than the first degree of deviation threshold, it is determined that the detection object waveform is included. Here, as the degree of deviation, the Euclidean distance or the like can be used. The Euclidean distance d between the input waveform data x(i) and the output waveform data y(i) of the autoencoder is given by the following formula.

[0054] [Equation 1]

[0055]

[0056] As the object waveform detection algorithm AL[1], a method using a classification type neural network such as a convolutional neural network (CNN) can also be conceived.

[0057] The object waveform detection algorithm AL[1] continues the determination operation continuously. Therefore, by reducing the number of input nodes of the waveform determination unit 104 (i.e., the number of N2 points of waveform data), the amount of calculation during the normal operation of determining the detection object waveform by the object waveform detection algorithm AL[1] can be suppressed, and the power consumed by the abnormality detection device can be suppressed. In addition, since the calculation time can be shortened, the waveform determination unit 104 can be implemented even with an arithmetic unit with low processing power.

[0058] The waveform extraction unit 105 inputs waveform data consisting of N1 points from the input buffer 102. When the waveform extraction unit 105 receives the detection result of the detection object waveform indicating that the input waveform data includes the detection object waveform from the waveform determination unit 104, the waveform extraction unit 105 accurately extracts the detection object waveform having an appropriate number of points from the input waveform data by using a trigger condition that the amplitude level of the waveform data is greater than a predetermined signal level (trigger level). The waveform extraction unit 105 outputs the extracted waveform data of the detection object waveform to the output selection unit 109.

[0059] The output selection unit 109 outputs the waveform data of the detection object waveform output from the waveform extraction unit 105 to the waveform determination unit 106. However, when the waveform data is used to generate an abnormality detection algorithm AL[2] used by the waveform determination unit 106 as described later, the output selection unit 109 outputs the waveform data of the detection object waveform output from the waveform extraction unit 105 to the outside of the device. The abnormality detection algorithm AL[2] is generated outside the device by using the waveform data of the detection object waveform as a normal waveform among the waveform data of the detection object waveform output from the waveform extraction unit 105 as learning data. The output selection unit 109 switches the output destination based on the instruction of the control unit 107.

[0060] The waveform determination unit 106 determines whether the detection object waveform output from the waveform extraction unit 105 is an abnormal waveform by using an abnormality detection algorithm AL[2] based on machine learning based on advance learning, and determines whether the manufacturing equipment has an abnormality. Here, if the detection object waveform is an abnormal waveform, it is determined that the manufacturing equipment has an abnormality. The abnormality detection result indicating whether the manufacturing equipment has an abnormality is output to the outside of the device. These abnormalities are related to the products and manufacturing processes manufactured in the equipment, and are related to the operating conditions of the equipment itself.

[0061] As described above, the learning data for generating the anomaly detection algorithm AL[2] is collected from the waveform data output from the waveform extraction unit 105. Therefore, the learning data is effectively prepared. The anomaly detection algorithm AL[2] can be, for example, an algorithm using an autoencoder. When using an autoencoder, in the anomaly detection algorithm AL[2], the waveform data of the detection object waveform output from the waveform extraction unit 105 is input into the autoencoder, and it is determined whether the deviation between the input waveform data and the output waveform data is equal to or less than a predetermined second deviation threshold, and based on the determination result, it is determined whether the detection object waveform is an abnormal waveform. Here, if the deviation is equal to or less than the second deviation threshold, it is determined that the detection object waveform is not an abnormal waveform but a normal waveform.

[0062] As described above, the waveform determination unit 104 determines whether the input waveform data extracted by the input buffer 102 includes the detection target waveform, and the waveform determination unit 106 performs abnormality determination on the input waveform data determined to include the detection target waveform.

[0063] Operation of the First Embodiment

[0064] Hereinafter, the operation of the abnormality detection device according to the present first embodiment is described.

[0065] Anomaly Detection Process

[0066] Figure 4 1 is a flowchart showing an exemplary flow of an abnormality detection process performed by the entire abnormality detection device according to the first embodiment. Figure 4 As shown in FIG. 1 , when N1 points of the waveform data are accumulated in the output portion of the input buffer 102, the input buffer 102 outputs the waveform data composed of N1 points to the waveform extraction units 103 and 105. In step S11, the waveform extraction unit 103 extracts the waveform data composed of N1 points output from the input buffer 102, and outputs the waveform data to the waveform determination unit 104 via the output selection unit 108. In this case, the waveform extraction unit 103 can thin out the waveform data composed of N1 points to become waveform data composed of N2 points (N1>N2).

[0067] Subsequently, the waveform determination unit 104 inputs the waveform data output from the waveform extraction unit 103 to an automatic encoder using an object waveform detection algorithm AL[1], and calculates the degree of deviation between the waveform data input to the automatic encoder and the waveform data output to the automatic encoder. In steps S12 and S13, the waveform determination unit 104 determines whether the waveform data includes a detection object waveform based on the calculated degree of deviation. In this case, when the degree of deviation is equal to or less than a first degree of deviation threshold, the waveform determination unit 104 determines that the detection object waveform is included.

[0068] When the waveform determination unit 104 determines that the detection target waveform is not included (No in step S13), the input buffer 102 shifts the waveform data in the output unit of the input buffer 102 by M points (step S14), and the process returns to step S11.

[0069] On the other hand, when the waveform determination unit 104 determines that the detection object waveform is included (yes in step S13), the waveform extraction unit 105 accurately extracts the detection object waveform from N1 points of the waveform data output by the input buffer 102 by using trigger conditions, etc., and outputs the detection object waveform to the waveform determination unit 106 via the output selection unit 109 (step S15).

[0070] Subsequently, the waveform determination unit 106 inputs the waveform data of the detection object waveform output from the waveform extraction unit 105 into the autoencoder by using the anomaly detection algorithm AL[2], and calculates the deviation degree of the waveform data between the input and output of the autoencoder (step S16). Next, the waveform determination unit 106 uses the anomaly detection algorithm AL[2] to determine whether the deviation degree is equal to or less than the second deviation degree threshold (step S17 is yes), and when the deviation degree is equal to or less than the second deviation degree threshold (step S17 is yes), it is determined that the manufacturing equipment is normal (step S18), and when the deviation degree exceeds the second deviation degree threshold (step S17 is no), it is determined that the manufacturing equipment has an abnormality (step S19). Thereafter, the process returns to step S11 through step S14.

[0071] Method for generating learning data for object waveform detection algorithm AL[1]

[0072] Figure 5 1 is a diagram showing an example of a method for generating learning data for an object waveform detection algorithm AL[1] according to the first embodiment. Here, it is assumed that an engineer performs generation of learning data on a PC outside the device, but is not limited to this, and an algorithm generation unit provided outside the device can generate learning data.

[0073] like Figure 5 As shown, first, a waveform portion of the detection object waveform is extracted from waveform data including the detection object waveform, and the waveform data including the detection object waveform is from the waveform data output by the waveform extraction unit 103. Here, the detection object waveform is a normal waveform, but may be an abnormal waveform. Subsequently, waveform data obtained by randomly shifting the extracted detection object waveform in the time axis direction within a range of approximately twice the number of points of the detection object waveform is automatically generated by a program executed on a PC.

[0074] Here, as described above, depending on the shapes of the detection object waveform and the non-detection object waveform, the waveform extraction unit 103 can thin out the waveform data composed of N1 points input from the input buffer 102 to become waveform data composed of N2 (N1>N2) points with coarse granularity. In this case, waveform data randomly shifted along the time axis direction is generated based on the waveform data thinned by the waveform extraction unit 103. In other words, the granularity of the automatically generated waveform data is the same as the granularity of the waveform data output from the waveform extraction unit 103.

[0075] Subsequently, when there are changes in the amplitude direction and the time axis direction of the detection object waveform, waveform data obtained by randomly expanding and contracting the detection object waveform in the amplitude direction and the time axis direction respectively is also generated within the assumed normal range. The waveform data set generated above is used as learning data for the object waveform detection algorithm AL[1].

[0076] Operation of Input Buffer 102

[0077] Figure 6 1 is a flowchart showing an example of the operation flow of the input buffer 102 according to the first embodiment. Figure 6 As shown, the input buffer 102 receives the waveform data output from the signal input unit 101 (step S21), accumulates the input waveform data within a predetermined period using a ring buffer, and copies the waveform data to the output section (step S22).

[0078] When N1 points of the waveform data are accumulated in the output portion, that is, when the input of the waveform data consisting of N1 points is completed (Yes in step S23), the input buffer 102 outputs the waveform data consisting of N1 points to the waveform extraction units 103 and 105 (step S24). In step S25, the input buffer 102 shifts the waveform data in the output portion by the number of points M. Thereafter, the process returns to step S22.

[0079] Figure 7 is an image diagram showing an example of the relationship between N1, M and L when the number of points of the detection object waveform is L. The relationship between N1, M and L should be N1 ≥ L + M. Figure 7 In the example of , N1 = 2×L, and M = L. However, in the case where the waveform fluctuation of the non-detection target waveform is large before and after the detection target waveform, N1 needs to be set to a value slightly larger than L.

[0080] Further, in Figure 7In the example of , the input buffer 102 outputs the waveform data consisting of N1 points to the waveform extraction units 103 and 105 five times. Among them, the fourth waveform data includes the detection object waveform. Therefore, the waveform determination unit 104 of the subsequent stage determines that the fourth waveform data includes the detection object waveform.

[0081] Operation of the Waveform Extraction Unit 103

[0082] Figure 8 1 is a flowchart showing an example of the operation flow of the waveform extraction unit 103 according to the first embodiment. Figure 8 As shown, in step S31, the waveform extraction unit 103 inputs the waveform data consisting of N1 points output from the input buffer 102.

[0083] Here, for example, in the case where the shapes of the detection object waveform and the non-detection object waveform are very different from each other, the contour can also be used to classify them. Here, it is assumed that the shapes of the detection object waveform and the non-detection object waveform are very different. Therefore, the waveform extraction unit 103 generates waveform data consisting of N2 points (N1>N2) with coarse granularity from the waveform data consisting of N1 points (step S32). This makes it possible to reduce the amount of calculation of the waveform determination unit 104 and shorten the calculation time. As a calculation for generating waveform data consisting of N2 points, for example, various processes such as a sparse process can be used, in which the average value of every 10 points is used as one data, the value of one point in every 10 points (in either case, N2=N1 / 10) is used as one data, or linear interpolation between the data of two points is used.

[0084] Subsequently, the waveform extraction unit 103 waits until the waveform determination unit 104 becomes a waveform data input waiting state (step S33). When it becomes a waiting state (yes in step S33), the waveform data composed of N2 points is output to the waveform determination unit 104 through the output selection unit 108 (step S34). Next, in step S35, the waveform extraction unit 103 waits until N1 points of the waveform data are accumulated in the output part of the input buffer 102, that is, until the update of the input buffer 102 is completed. If the update of the input buffer 102 has been completed (yes in step S35), the process returns to step S31.

[0085] Fig. 9 is an image diagram showing an exemplary waveform data thinning process performed by the waveform extraction unit 103 according to the first embodiment. Fig. 9 In the example of , the waveform extraction unit 103 generates waveform data consisting of N2 points (N2=N1 / 10) by using the value of one point out of every 10 points of the waveform data consisting of N1 points.

[0086] Operation of the Waveform Determination Unit 104

[0087] Fig.10 1 is a flowchart showing an example of the operation flow of the waveform determination unit 104 according to the first embodiment. Fig.10 As shown, in step S41, the waveform determination unit 104 inputs the waveform data output from the waveform extraction unit 103. Thereafter, the waveform determination unit 104 uses the object waveform detection algorithm AL[1] based on machine learning learned in advance to determine whether the detection object waveform is included in the waveform data output from the waveform extraction unit 103, and outputs the detection result of the detection object waveform indicating whether the detection object waveform is included in the waveform data to the waveform extraction unit 105.

[0088] As reference Figure 5 As described, the waveform data output from the waveform extraction unit 103 is used to generate learning data for generating the object waveform detection algorithm AL[1]. As the object waveform detection algorithm AL[1], for example, there is a method of using an autoencoder which is one of the neural network technologies. Hereinafter, assuming that the autoencoder is used as the object waveform detection algorithm AL[1], the process after step S42 is described.

[0089] In step S42, the waveform determination unit 104 inputs the waveform data output from the waveform extraction unit 103 to the auto encoder, and calculates the degree of deviation of the waveform data input and output to the auto encoder by using the object waveform detection algorithm AL[1].

[0090] Subsequently, the waveform determination unit 104 determines whether the degree of deviation is equal to or less than a first degree of deviation threshold (step S43); when the degree of deviation is equal to or less than the first degree of deviation threshold (yes in step S43), it determines that the waveform data includes the detection object waveform (step S44); and when the degree of deviation exceeds the first degree of deviation threshold (no in step S43), it determines that the waveform data does not include the detection object waveform (step S45).

[0091] Next, in step S46, the waveform determination unit 104 outputs the detection result of the detection target waveform indicating whether the waveform data includes the detection target waveform to the waveform extraction unit 105. Thereafter, the process returns to step S41.

[0092] Operation of the Waveform Extraction Unit 105

[0093] Fig.11 1 is a flowchart showing an example of the operation flow of the waveform extraction unit 105 according to the first embodiment. Fig.11As shown, in step S51, the waveform extraction unit 105 inputs the waveform data output from the input buffer 102. Next, in step S52, the waveform extraction unit 105 confirms based on the detection result of the detection target waveform output from the waveform determination unit 104: whether the waveform determination unit 104 determines that the waveform data input from the input buffer 102 includes the detection target waveform.

[0094] When the waveform determination unit 104 determines that the detection object waveform is included (Yes in step S52), the waveform extraction unit 105 extracts only the detection object waveform from the waveform data input from the input buffer 102 using the trigger condition for extracting the detection object waveform and the like (step S53), and outputs the waveform data of the extracted detection object waveform to the waveform determination unit 106 via the output selection unit 109 (step S54). Thereafter, in step S55, the waveform extraction unit 105 enters a state of waiting for waveform data to be input from the input buffer 102.

[0095] On the other hand, when the waveform determination unit 104 determines that the detection object waveform is not included (No in step S52), the waveform extraction unit 105 discards the waveform data input from the input buffer 102 and transitions to an input waiting state for waveform data from the input buffer 102 (step S55).

[0096] Fig.12 FIG. 1 is an image diagram showing an exemplary detection object waveform extraction process performed by the waveform extraction unit 105 according to the first embodiment. Fig.12 In the example of , the waveform extraction unit 105 accurately extracts the detection target waveform having an appropriate number of points by using a trigger condition that the amplitude level of the detection target waveform exceeds a predetermined signal level.

[0097] Operation of the Waveform Determination Unit 106

[0098] Fig.13 1 is a flowchart showing an example of the operation flow of the waveform determination unit 106 according to the first embodiment. Fig.13 As shown, in step S61, the waveform determination unit 106 inputs the waveform data of the detection object waveform output from the waveform extraction unit 105. Then, the waveform determination unit 106 determines whether the detection object waveform is an abnormal waveform by using the abnormality detection algorithm AL[2] based on machine learning of advance learning, and determines whether the manufacturing equipment has an abnormality. Then, the waveform determination unit 106 outputs the abnormality detection result indicating whether the manufacturing equipment has an abnormality to the outside of the device.

[0099] The learning data for generating the abnormality detection algorithm AL[2] is prepared by collecting the waveform data output from the waveform extraction unit 105. As the abnormality detection algorithm AL[2], for example, there is a method using an autoencoder. Hereinafter, assuming that the autoencoder is used as the abnormality detection algorithm AL[2], the process after step S62 is described.

[0100] The waveform determination unit 106 inputs the waveform data of the detection target waveform output from the waveform extraction unit 105 to the auto encoder using the abnormality detection algorithm AL[2], and calculates the degree of deviation of the waveform data of the input and output of the auto encoder (step S62).

[0101] Subsequently, the waveform determination unit 106 determines whether the degree of deviation is equal to or less than a second degree of deviation threshold (step S63), and when the degree of deviation is equal to or less than the second degree of deviation threshold (yes in step S63), it determines that the waveform of the detection object is a normal waveform and the manufacturing equipment is normal (step S64), and when the degree of deviation exceeds the second degree of deviation threshold (no in step S63), it determines that the waveform of the detection object is an abnormal waveform and the manufacturing equipment is abnormal (step S65).

[0102] Next, the waveform determination unit 106 outputs the abnormality detection result indicating whether the manufacturing equipment is abnormal to the outside of the equipment (step S66). Thereafter, the process returns to step S61.

[0103] Fig.14 is a diagram showing an exemplary autoencoder for the waveform determination unit 106 according to the first embodiment. Fig.14 As shown, the autoencoder consists of an input layer, one or more intermediate layers (in Fig.14 The automatic encoder is a method of learning feature points compressed by an error back propagation method to make input data and output data equal. Therefore, when waveform data equivalent to the learning data is input to the input layer, the same waveform data is reproduced on the output layer. On the contrary, when waveform data that has not been learned is input to the input layer, the same waveform data cannot be reproduced, and thus the waveform data between the input and the output differs.

[0104] exist Fig.14 In the example, the autoencoder uses waveform data of a detection object waveform as learning data, and learns characteristic points of the waveform data of the detection object waveform so that when waveform data equivalent to the learning data is input to the input layer, the same waveform data is reproduced in the output layer.

[0105] Effects of the First Embodiment

[0106] In the first embodiment, first, the waveform determination unit 104 determines whether the detection target waveform is included in the time series waveform data from the detection target device, the number of points of which is greater than the number of points in the detection target waveform. When the waveform determination unit 104 determines that the waveform data includes the detection target waveform, the waveform extraction unit 105 extracts the detection target waveform from the waveform data. The waveform determination unit 104 determines whether the detection target device has an abnormality based on the waveform data of the detection target waveform output from the waveform extraction unit 105.

[0107] Therefore, even if the waveform data output from the waveform extraction unit 103 includes a mixture of the detection object waveform and the waveform other than the non-detection object waveform, only the detection object waveform can be extracted without erroneously determining the non-detection object waveform as the detection object waveform, and abnormality detection of the detection target device can be performed. When the waveform data output from the waveform extraction unit 103 does not include the detection object waveform, the waveform extraction unit 105 does not perform extraction processing of the detection object waveform, so that the amount of calculation can be suppressed.

[0108] The waveform determination units 104 and 106 perform the above-mentioned processing by using an algorithm obtained by using the waveform data obtained by learning the waveform data including the detection object waveform. In this case, as the learning data for the waveform determination unit 104, a small amount of waveform data (at least one waveform data set) obtained by only manually cutting out the detection object waveform is prepared, and the waveform data obtained by randomly shifting the detection object waveform cut out in the time axis direction within the range of about twice the length of the detection object waveform is used. When the amplitude direction and the time axis direction of the waveform data of the detection object waveform change, the waveform data randomly expanded and contracted in the amplitude direction and the time axis direction are set as learning data. Therefore, the engineer does not have to manually extract and collect a large amount of learning data for the waveform determination unit 104, and the engineer can reduce the workload and time.

[0109] The learning data for the waveform determination unit 106 is acquired from the waveform data output from the waveform extraction unit 105. Therefore, the learning data for the waveform determination unit 106 can be prepared efficiently.

[0110] Second embodiment

[0111] Configuration and Operation of Second Embodiment

[0112] In the second embodiment, compared with the first embodiment described above, the configuration itself is similar, but the operation of the waveform determination unit 104 is different.

[0113] In the first embodiment described above, the waveform determination unit 104 calculates the degree of deviation of the waveform data input and output of the autoencoder, and when the degree of deviation is equal to or less than a predetermined first degree of deviation, determines that the waveform data input to the autoencoder includes the detection object waveform.

[0114] On the other hand, in the second embodiment, the waveform determination unit 104 uses an arithmetic value such as a normalized error (|(x(i)-y(i)| / (|x(i)|+α)) between the waveform data x(i) input to the autoencoder and the waveform data y(i) output from the autoencoder, where α is an optional constant for excluding division by 0. The waveform determination unit 104 determines that the detection object waveform is included in the waveform data when the normalized error is continuously equal to a predetermined error threshold or less than the predetermined error threshold plus a predetermined number.

[0115] Further, when it is determined that the waveform data includes the detection object waveform, the waveform determination unit 104 can instruct the waveform extraction unit 105 to exclude the range other than the range in which the normalized error is continuously equal to the error threshold or less than the error threshold plus a predetermined number from the range of extracting the detection object waveform. As a result, the possibility that the waveform extraction unit 105 erroneously extracts a non-detection object waveform can be reduced.

[0116] Fig.15 is a diagram showing an exemplary operation of the waveform determination unit 104 according to the second embodiment. Fig.15 The left side of the figure shows the waveform data at the input of the autoencoder. Fig.15 The center of shows the waveform data at the output of the autoencoder, and Fig.15 The right side of shows the normalized error between the waveform data at the input and the waveform data at the output of the autoencoder.

[0117] exist Fig.15 In the example on the right side of , there is a range R2 in which a predetermined number of normalized errors are continuously equal to or less than the error threshold. Therefore, the waveform determination unit 104 determines that the detection object waveform is included in the range R2 of the waveform data. Further, since it is determined that the detection object waveform is included in the range R2, the range R2 is determined to be the range for extracting the detection object waveform, and the ranges R1 and R3 other than the range R2 are excluded from the range for extracting the detection object waveform, and the waveform determination unit 104 instructs the waveform extraction unit 105.

[0118] Effects of the Second Embodiment

[0119] In the first embodiment described above, the waveform determination unit 104 determines that the waveform data includes the detection object waveform when the deviation of the waveform data between the input and output of the automatic encoder is less than a predetermined first deviation threshold value. Therefore, within the range where the waveform extraction unit 105 extracts the detection object waveform, there is a problem that when the waveform including the detection object waveform changes, it is erroneously determined that the detection object waveform is included, or when the detection object waveform includes a large abnormal portion, it is erroneously determined that the detection object waveform is not included.

[0120] On the other hand, in the second embodiment, when the normalized error between the waveform data input to the autoencoder and the waveform data output from the autoencoder continuously becomes equal to a predetermined error threshold or less than a predetermined error threshold-a predetermined number, the waveform determination unit 104 determines that the waveform data includes the detection object waveform. Therefore, when the waveform including the non-detection object waveform changes, the influence of the waveform fluctuation appears as a normalized error to a large extent, so that it can be correctly determined that the detection object waveform is not included. In addition, since the determination is performed based on the continuity of the normalized error in a predetermined section, even when a large abnormal part is included in the detection object waveform, it can be correctly determined that the detection object waveform is included. The remaining effects are the same as those of the first embodiment mentioned above.

[0121] Third embodiment

[0122] The first embodiment described above assumes that the detection target waveform is one, and the third embodiment assumes that there are a plurality of detection target waveforms different from each other.

[0123] Configuration of the Third Embodiment

[0124] Compared with the first embodiment described above, this third embodiment has the same configuration itself, but the object waveform detection algorithm AL[1] used for the waveform determination unit 104 and the abnormality detection algorithm AL[2] used for the waveform determination unit 106 are different.

[0125] The object waveform detection algorithm AL[1] used for the waveform determination unit 104 is configured by a classification type neural network such as a convolutional neural network. For each of the multiple detection object waveforms, the object waveform detection algorithm AL[1] calculates the probability that the detection object waveform is included in the waveform data by a soft max function or the like. When there is a detection object waveform whose calculated probability exceeds a predetermined probability threshold (e.g., 70%), the object waveform detection algorithm AL[1] determines that the detection object waveform is included in the waveform data. On the other hand, when there is no detection object waveform whose calculated probability exceeds the probability threshold, the object waveform detection algorithm AL[1] determines that any of the multiple detection object waveforms is not included in the waveform data.

[0126] For each of the plurality of detection target waveforms, the learning data for the target waveform detection algorithm AL[1] is obtained by using Figure 5 The detection result of the detection object waveform includes a result indicating whether the detection object waveform is included, and when the detection object waveform is included, includes information identifying the detection object waveform. The detection result of the detection object waveform is output to the waveform extraction unit 105 and the waveform determination unit 106.

[0127] The waveform determination unit 106 maintains multiple anomaly detection algorithms AL[2] corresponding to each of the multiple detection object waveforms, switches the anomaly detection algorithm AL[2] according to the detection result of the detection object waveform, and uses the switched anomaly detection algorithm AL[2] to determine whether the manufacturing equipment is abnormal.

[0128] Operation of the Third Embodiment

[0129] Fig.16 1 is a flowchart showing an example of the flow of an abnormality detection process performed by the entire abnormality detection device according to the third embodiment. Figure 4 The flowchart is different.

[0130] like Fig.16 As shown, when waveform data is extracted and output by the waveform extraction unit 103 in step S11, the waveform determination unit 104 calculates the probability that the waveform data output from the waveform extraction unit 103 includes the detection target waveform using the target waveform detection algorithm AL[1] for each detection target waveform among the plurality of detection target waveforms (step S71). In step S72, the waveform determination unit 104 determines whether there is a detection target waveform whose calculated probability exceeds the probability threshold.

[0131] If there is no detection target waveform whose probability exceeds the probability threshold value (No in step S72 ), the waveform determination unit 104 determines that none of the plurality of detection target waveforms is included in the waveform data, and returns to the process of step S11 via step S14 .

[0132] On the other hand, when there is a detection target waveform whose probability exceeds the probability threshold (Yes in step S72), the waveform determination unit 104 determines that the detection target waveform is included in the waveform data, and the waveform determination unit 106 performs switching to the abnormality detection algorithm AL[2] according to the detection target waveform (step S73). Figure 4 The flowchart is the same as

[0133] Effects of the Third Embodiment

[0134] In the third embodiment, for each of the multiple detectable target waves, the waveform determination unit 104 uses the object waveform detection algorithm AL[1] composed of a classification type neural network to calculate the probability that the detection target waveform is included in the wave data, and determines whether any of the multiple detection target waveforms is included based on whether the calculated probability exceeds a predetermined probability threshold. When the waveform determination unit 104 determines that any of the multiple detection target waveforms is included, the waveform determination unit 106 switches the abnormality detection algorithm AL[2] according to the determined detection target waveform, and determines whether the detection target waveform is an abnormal waveform.

[0135] Therefore, even in the case where there are a plurality of detection target waveforms, the plurality of detection target waveforms can be extracted respectively, and abnormality detection of the detection target device can be performed. The remaining effects are the same as those of the first embodiment mentioned above.

[0136] Fourth embodiment

[0137] The fourth embodiment is an embodiment of an abnormality detection system using one of the abnormality detection devices according to the first to third embodiments described above.

[0138] Configuration and Operation of the Fourth Embodiment

[0139] Fig.17 4 is a diagram showing an exemplary configuration of an abnormality detection system according to a fourth embodiment. Fig.17 As shown, the abnormality detection system according to the fourth embodiment includes abnormality detection devices 10A and 10B corresponding to any of the abnormality detection devices according to the first to third embodiments, manufacturing equipment 20A and 20B, abnormality indicators 30A and 30B, MES (manufacturing execution system) 40, SCADA (supervisory control and data acquisition) 50, algorithm generation unit 60, and algorithm storage unit 70. SCADA 50 is an exemplary data acquisition unit. Further, in the drawings described below, the abnormality detection devices 10A and 10B are also represented by abnormality detection devices #A and abnormality detection devices #B, respectively, the manufacturing equipment 20A and 20B are also represented by manufacturing equipment #A and manufacturing equipment #B, respectively, and the abnormality indicators 30A and 30B are also represented by abnormality indicators #A and abnormality indicators #B, respectively.

[0140] The manufacturing equipment 20A is, for example, a manufacturing equipment for manufacturing semiconductors. The abnormality detection device 10A determines whether an abnormality has occurred on the manufacturing equipment 20A based on the monitoring signal input from the manufacturing equipment 20A, and outputs an abnormality detection result indicating whether an abnormality has occurred on the manufacturing equipment 20A to the abnormality indicator 30A. The abnormality indicator 30A displays the abnormality detection result of the manufacturing equipment 20A. Note that the relationship between the manufacturing equipment 20B, the abnormality detection device 10B, and the abnormality indicator 30B is the same as the relationship between the manufacturing equipment 20A, the abnormality detection device 10A, and the abnormality indicator 30A.

[0141] In this fourth embodiment, the algorithm generation unit 60 generates the target waveform detection algorithm AL[1] used by the waveform determination unit 104 , and generates the abnormality detection algorithm AL[2] used by the waveform determination unit 106 .

[0142] When the target waveform detection algorithm AL[1] is generated, the output selection unit 108 outputs the waveform data output from the waveform extraction unit 103 to the SCADA 50, and the SCADA 50 collects the waveform data output from the output selection unit 108 and outputs it to the algorithm generation unit 60. The algorithm generation unit 60 generates the waveform data by referring to Figure 5 The described method extracts the detection object waveform from the waveform data including the detection object waveform in the waveform data output by SCADA50, generates learning data while randomly shifting the extracted detection object waveform, and uses the generated learning data to generate the object waveform detection algorithm AL[1].

[0143] When the abnormality detection algorithm AL[2] is generated, the output selection unit 109 outputs the waveform data of the detection object waveform output from the waveform extraction unit 105 to the SCADA 50, and the SCADA 50 collects the waveform data of the detection object waveform output from the output selection unit 109 and outputs it to the algorithm generation unit 60. The algorithm generation unit 60 generates the abnormality detection algorithm AL[2] using the waveform data of the normal waveform among the waveform data of the detection object waveform output from the SCADA 50 as learning data.

[0144] The algorithm storage unit 70 stores various algorithms including the object waveform detection algorithm AL[1] and the abnormality detection algorithm AL[2] generated by the algorithm generation unit 60.

[0145] The abnormality detection device 10A includes the same Figure 1, and further includes an algorithm switching unit 110. In order to switch the object waveform detection algorithm AL[1] used by the waveform determination unit 104, the algorithm switching unit 110 reads the new object waveform detection algorithm AL[1] from the algorithm storage unit 70 and causes the waveform determination unit 104 to retain it. In order to switch the abnormality detection algorithm AL[2] used for the waveform determination unit 106, the algorithm switching unit 110 reads the new abnormality detection algorithm AL[2] from the algorithm storage unit 70 and causes the waveform determination unit 106 to retain it. The configuration of the abnormality detection device 10B is the same as that of the abnormality detection device 10A.

[0146] Effects of the Fourth Embodiment

[0147] In the fourth embodiment, the algorithm generation unit 60 generates learning data for the target waveform detection algorithm AL[1], and generates the target waveform detection algorithm AL[1] and the abnormality detection algorithm AL[2] using the learning data.

[0148] Therefore, compared with the first embodiment in which an engineer who manages the manufacturing equipment 20A or 20B generates learning data or an algorithm, the workload and time of the engineer can be reduced. The remaining effects are the same as those of the first embodiment mentioned above.

[0149] Fifth embodiment

[0150] Configuration and Operation of the Fifth Embodiment

[0151] Fig.18 : is a diagram showing an exemplary configuration of an abnormality detection device according to the fifth embodiment. Note that Fig.18 Only the main components constituting the abnormality detection device are shown, and other components are omitted. Fig.18 As shown, the abnormality detection device according to the fifth embodiment is different from the first embodiment described above in that the waveform extraction unit 103, the waveform determination unit 104, and the waveform extraction unit 105 are replaced by the waveform extraction unit 111, the waveform determination unit 112, and the waveform extraction unit 113. The waveform extraction unit 113 is also different from the first embodiment described above in that the waveform data is input from the waveform extraction unit 111 via the output selection unit 108 instead of being input from the input buffer 102. The waveform extraction unit 111 is an example of a first waveform extraction unit, the waveform determination unit 112 is an example of a first waveform determination unit, and the waveform extraction unit 113 is an example of a second waveform extraction unit. In the drawings described below, the waveform extraction units 111 and 113 are also represented by waveform extraction units #1 and #2, respectively, and the waveform determination unit 112 is also represented by the waveform determination unit #1.

[0152] Fig.191 is an image diagram showing an exemplary detection object waveform extraction process of the waveform extraction unit 113 according to the fifth embodiment. The waveform determination unit 112 uses an algorithm using a dynamic time warping (DTW) method as the object waveform detection algorithm AL[1]. When the waveform determination unit 112 uses the DTW method as the object waveform detection algorithm AL[1], it is necessary to obtain reference waveform data.

[0153] First, the process of obtaining reference waveform data for the DTW method is described. From the point where the amplitude level of the waveform data output by the input buffer 102 exceeds the start trigger level, the waveform extraction unit 111 extracts waveform data corresponding to r0 points. Here, the trigger condition for determining the start point is not limited to the case where the trigger level is exceeded, and for example, it can be considered Fig.19 . In order to fully include the detection object waveform in the waveform data having r0 points, when the number of points of the detection object waveform is L, r0 is set to a value sufficiently larger than L. The engineer who manages the abnormality detection target device specifies the range [p0, q0] of the detection object waveform from the waveform data of r0 points extracted by the waveform extraction unit 111 based on the engineer's knowledge, etc. Here, p0 represents the starting point and q0 represents the end point. Then, the object waveform detection algorithm AL[1] is given in advance as reference waveform data, information indicating that the waveform data output from the waveform extraction unit 111 is waveform data having r0 points, and information indicating the range [p0, q0] of the detection object waveform included in the waveform data of r0 points.

[0154] Next, the process of extracting the waveform of the detection object is described. From the point where the amplitude level of the waveform data output by the input buffer 102 exceeds the start trigger level as the starting point, the waveform extraction unit 111 extracts the waveform data within a predetermined range, and outputs the extracted waveform data to the waveform determination unit 112 and the waveform extraction unit 113 through the output selection unit 108. Here, it is assumed that the waveform data corresponding to r1 points are extracted from the starting point. Here, r1=r0, or r1>r0 or r1<r0. According to the waveform data, the waveform extraction unit 111 can extract the waveform data from the starting point to the end point where the amplitude level of the waveform data is lower than the end trigger level.

[0155] As the object waveform detection algorithm AL[1], the waveform determination unit 112 uses the DTW method to determine the correspondence between the reference waveform data and the waveform data (hereinafter referred to as the target waveform data) output from the waveform extraction unit 111, so that the distance (cost) between two points at each point of the object waveform data and the reference waveform data is minimized. Subsequently, the range [p1, q1] of the target waveform data corresponding to the range [p0, q0] of the reference waveform data is determined as the range of the detection object waveform. Then, the waveform determination unit 112 outputs the detection object range information indicating the range [p1, q1] of the detection object waveform to the waveform extraction unit 113.

[0156] In this case, the waveform determination unit 112 can confirm the validity of the range [p1, q1] determined as the detection object waveform. As a method of confirming the validity, for example, the Euclidean distance between the reference waveform data and the target waveform data of a predetermined number of points (e.g., 20 points) after the starting point of the range of the detection object waveform is calculated. Further, the Euclidean distance between the reference waveform data and the target waveform data is calculated by using a predetermined number of points (e.g., 20 points) before the end point of the range of the detection object waveform. If both Euclidean distances are less than a predetermined distance threshold, for example, 0.5 or less, it can be determined that the range of the detection object waveform is appropriate, and if any of the Euclidean distances is equal to or greater than the distance threshold, it can be determined that the range of the detection object waveform is inappropriate. In addition, only when the range of the detection object waveform is appropriate, the detection object range information can be output to the waveform extraction unit 113, and when the range of the detection object waveform is inappropriate, the object waveform data can be discarded, and the detection object range information will not be output to the waveform extraction unit 113.

[0157] The waveform extraction unit 113 extracts a waveform within the range [p1, q1] of the detection target waveform from the waveform data output by the waveform extraction unit 111 as the detection target waveform, based on the detection target segment information, and outputs the waveform data of the extracted detection target waveform to the waveform determination unit 106 via the output selection unit 109. Similar to the first embodiment described above, the waveform determination unit 106 can detect anomalies in the manufacturing equipment by using the anomaly detection algorithm AL[2].

[0158] Effects of the Fifth Embodiment

[0159] In the fifth embodiment, as an object waveform detection algorithm AL[1], the waveform determination unit 112 uses the DTW method to determine the range [p1, q1] of the detection object waveform included in r1 points of the waveform data, and the waveform extraction unit 113 extracts the waveform within the range [p1, q1] of the detection object waveform from the r1 points of the waveform data as the detection object waveform.

[0160] Therefore, the waveform extraction unit 113 can accurately extract the detection object waveform without using a trigger condition, etc. The remaining effects are the same as those of the first embodiment mentioned above.

[0161] In the fifth embodiment, it is assumed that the detection object waveform is included in the waveform data within a predetermined range from the point where the amplitude level of the waveform data exceeds the start trigger level, but it is considered that the detection object waveform is not included according to the waveform data. Therefore, a waveform extraction unit 111 and a waveform determination unit 112 may be provided at a subsequent stage of the waveform determination unit 104 of the first embodiment mentioned above. The waveform extraction unit 111 and the waveform determination unit 112 may perform the above-mentioned process with respect to the waveform data including the detection object waveform determined by the waveform determination unit 104 to obtain the range [p1, q1] of the detection object waveform.

[0162] Sixth embodiment

[0163] As with the third embodiment described above, the sixth embodiment assumes that there are a plurality of detection target waveforms that are different from each other. In the following description, it is assumed that there are two detection target waveforms.

[0164] Configuration and Operation of Sixth Embodiment

[0165] Fig. 20 : is a diagram showing an example of the configuration of an abnormality detection device according to the sixth embodiment. Note that Fig. 20 Only the constituent elements constituting the main part of the abnormality detection device are shown, and other constituent elements are omitted. Fig. 20As shown, compared with the first embodiment described above, the abnormality detection device according to the sixth embodiment has the following configuration: two sets of waveform extraction units 103, output selection units 108, and waveform determination units 104 corresponding to two detection object waveforms are provided. That is, instead of the waveform extraction unit 103, the output selection unit 108, and the waveform determination unit 104, a first set of waveform extraction units 103X, an output selection unit 108X, and a waveform determination unit 104X, and a second set of waveform extraction units 103Y, an output selection unit 108Y, and a waveform determination unit 104Y are provided. The waveform extraction units 103X and 103Y are examples of the first waveform extraction unit, and the waveform determination units 104X and 104Y are examples of the first waveform determination unit. In the figures described below, waveform extraction units 103X and 103Y are also represented as waveform extraction units #1X, #1Y, respectively, waveform determination units 104X and 104Y are also represented as waveform determination units #1X, #1Y, respectively, and output selection units 108X and 108Y are also represented as output selection units #1X, #1Y, respectively.

[0166] Fig.21 1 is an image diagram showing an exemplary process of determining whether a detection target waveform is included by the waveform determination units 104X and 104Y according to the sixth embodiment. A first group of the waveform extraction unit 103X, the output selection unit 108X, and the waveform determination unit 104X are provided to determine whether one detection target waveform (hereinafter referred to as detection target waveform X) of two detection target waveforms is included in the waveform data output from the input buffer 102.

[0167] A second set of a waveform extraction unit 103Y, an output selection unit 108Y, and a waveform determination unit 104Y is provided to determine whether the waveform data output from the input buffer 102 includes the other detection target waveform (hereinafter referred to as detection target waveform Y) of the two detection target waveforms.

[0168] exist Fig.21 In the example of , the waveform data output from the input buffer 102 includes two detection target waveforms X and Y. Therefore, the waveform determination unit 104X determines that the waveform data includes the detection target waveform X, and the waveform determination unit 104Y determines that the waveform data includes the detection target waveform Y.

[0169] In this case, the waveform determination units 104X and 104Y can respectively determine the two detection object waveforms X and Y substantially at the same time. Therefore, the waveform extraction unit 105 can extract the two detection object waveforms X and Y from the waveform data substantially at the same time. However, the present invention is not limited to this, and the waveform extraction unit 105 can extract the two detection object waveforms X and Y separately.

[0170] In the sixth embodiment, there are two detection object waveforms. Therefore, similar to the third embodiment described above, the waveform determination unit 106 maintains two abnormality detection algorithms AL[2] corresponding to each of the two detection object waveforms, and switches the two abnormality detection algorithms AL[2] according to the detection results of the detection object waveforms, and uses them.

[0171] In the sixth embodiment, two examples of detection target waveforms are described, but three or more detection target waveforms may be used. In this case as well, a set of waveform extraction unit 103, output selection unit 108, and waveform determination unit 104 may be provided for several detection target waveforms.

[0172] Effects of the Sixth Embodiment

[0173] In the sixth embodiment, a set of waveform extraction units 103, output selection units 108 and waveform determination units 104 corresponding to each detection object waveform in multiple detection object waveforms are provided, and each set of waveform determination units 104 determines whether the corresponding detection object waveform is included in the waveform data.

[0174] Therefore, even in the case where there are multiple detection object waveforms, multiple detection object waveforms can be extracted respectively, and abnormality detection of the detection target device can be performed. In addition, it can be determined that the waveform data includes multiple detection object waveforms at substantially the same time, and they can be extracted at substantially the same time. The remaining effects are the same as those of the first embodiment mentioned above.

[0175] Concepts of the first to third embodiments, the fifth embodiment, and the sixth embodiment

[0176] Fig. 22 : is a diagram showing an example of a configuration of an abnormality detection apparatus, which conceptually shows the first to third embodiments and the fifth and sixth embodiments described above. Fig. 22 The abnormality detection device shown includes a first waveform extraction unit 911 , a first waveform determination unit 912 , a second waveform extraction unit 913 , and a second waveform determination unit 914 .

[0177] The first waveform extraction unit 911 inputs the time series waveform data output from the detection target device, and outputs the input waveform data having a larger number of points than the detection target waveform. At this time, the first waveform extraction unit 911 may output the input waveform data as it is, or may output the input waveform data after the number of points is thinned out. The first waveform extraction unit 911 corresponds to the waveform extraction units 103, 103X, 103Y, and 111.

[0178] The first waveform determination unit 912 uses the first algorithm AL[1] to determine whether the detection target waveform is included in the input waveform data output from the first waveform extraction unit 911. The first waveform determination unit 912 corresponds to the waveform determination units 104, 104X, 104Y, and 112. The first algorithm AL[1] corresponds to the target waveform detection algorithm AL[1].

[0179] When the first waveform determination unit 912 determines that the input waveform data includes the detection target waveform, the second waveform extraction unit 913 receives the input waveform data, and extracts and outputs the detection target waveform from the input waveform data. The second waveform extraction unit 913 corresponds to the waveform extraction units 105 and 113.

[0180] The second waveform determination unit 914 determines whether the detection target waveform output from the second waveform extraction unit 913 is an abnormal waveform by using the second algorithm AL[2], and determines whether the detection target device is abnormal. The second waveform determination unit 914 corresponds to the waveform determination unit 106. The second algorithm AL[2] corresponds to the abnormality detection algorithm AL[2].

[0181] Concept of the Fourth Embodiment

[0182] Fig.23 is a diagram showing an example of a configuration of an abnormality detection system, which conceptually shows the fourth embodiment described above. Fig.23 The anomaly detection system shown includes an anomaly detection device 91 , a data acquisition unit 92 and an algorithm generation unit 93 .

[0183] Abnormality detection device 91 and Fig. 22 The data acquisition unit 92 acquires the input waveform data output from the first waveform extraction unit 911 and the waveform data of the detection object waveform output from the second waveform extraction unit 913, and outputs them to the algorithm generation unit 93. The acquisition unit 92 corresponds to the SCADA 50.

[0184] The algorithm generation unit 93 generates first learning data using input waveform data including the detection object waveform among the input waveform data output from the data acquisition unit 92, and generates a first algorithm AL[1] using the first learning data. In addition, the algorithm generation unit 93 generates a second algorithm AL[2] by using waveform data of a normal waveform of the detection object waveform among the waveform data of the detection object waveform output from the data acquisition unit 92 as second learning data. The algorithm generation unit 93 corresponds to the algorithm generation unit 60. The first learning data corresponds to the learning data for the object waveform detection algorithm AL[1]. The second learning data corresponds to the learning data for the abnormality detection algorithm AL[2].

[0185] Other concepts of the first to third embodiments, the fifth embodiment, and the sixth embodiment

[0186] Fig.24 is a diagram showing an example of other structures of the abnormality detection apparatus, which conceptually shows the first to third embodiments, the fifth embodiment, and the sixth embodiment described above. Fig.24 The abnormality detection device shown includes an input unit 915 , a first determination unit 916 , a data identification unit 917 , and a second determination unit 918 .

[0187] The input unit 915 receives the time series data, extracts the time series data of a predetermined period from the input time series data, and outputs the extracted time series data as the input time series data. The input unit 915 corresponds to the input buffer 102.

[0188] The first determination unit 916 determines whether the input time series data includes the time series data to be detected. This determination may be made based on a first algorithm using a neural network technique.

[0189] The first determination unit 916 corresponds to the waveform determination units 104 , 104X, 104Y, and 112 .

[0190] The first algorithm corresponds to the object waveform detection algorithm AL[1].

[0191] When input time series data is input and the determination result of the first determination unit 916 indicates that the input time series data includes data to be detected, the data identification unit 917 identifies the time series data corresponding to the data to be detected from the input time series data, and outputs the identified time series data as the detection object time series data. The detection object time series data may be data having a period shorter than a predetermined period. The data identification unit 917 corresponds to the waveform extraction units 105 and 113.

[0192] The second determination unit 918 determines whether the detection object time series data is data indicating an abnormality. This determination can be performed based on a second algorithm using a neural network technique. The second determination unit 918 corresponds to the waveform determination unit 106. The second algorithm corresponds to the abnormality detection algorithm AL[2].

[0193] Although the invention made by the inventors has been specifically described based on the embodiments, the present invention is not limited to the embodiments that have been described, and it is needless to say that various modifications can be made without departing from the gist thereof.

[0194] For example, in the embodiment described above, an output selection unit is provided in the subsequent stage of the waveform extraction unit, and a control unit for controlling the output selection unit is provided, but the present embodiment is not limited thereto. For example, the waveform extraction unit may have a function of autonomously switching the output destination, in which case the output selection unit and the control unit are not required.

[0195] In the above-described embodiment, it is assumed that the detection target device for detecting abnormality is a manufacturing device such as a manufacturing device for a semiconductor device, but the present embodiment is not limited thereto. The detection target device may be a device other than a manufacturing device.

Claims

1. An anomaly detection device, comprising: a first extraction unit configured to receive the waveform data and extract input waveform data from the waveform data; A first determining unit is configured to determine whether the detection object waveform data is included in the input waveform data; a second extracting unit configured to extract the detection object waveform data from the input waveform data based on the determination result of the first determining unit; as well as a second determining unit configured to determine whether the detection object waveform data extracted by the second extracting unit indicates abnormality, wherein the first determining unit determines by using a first algorithm based on first learning data; wherein the second determining unit determines by using a second algorithm based on second learning data; wherein the first learning data is generated based on input waveform data including the detection object waveform data among the input waveform data extracted by the first extraction unit, and The second learning data is generated based on the detection object waveform data determined to be normal among the detection object waveform data extracted by the second extraction unit.

2. The abnormality detection device according to claim 1, wherein the first extraction unit performs sparse processing on the waveform data to generate the input waveform data; The first determination unit determines whether the detection target waveform data is included in the input waveform data as thinned-out waveform data.

3. The abnormality detection device according to claim 1, The first determination unit determines by using a first algorithm utilizing a first autoencoder, calculates the degree of deviation between the input waveform data to be input to the first autoencoder and the output data of the first autoencoder, and determines whether the detection object waveform data is included in the input waveform data based on the degree of deviation.

4. The abnormality detection device according to claim 1, The first determination unit determines by using a first algorithm utilizing a first autoencoder, calculates a normalized error between the input waveform data to be input to the first autoencoder and the output data of the first autoencoder, and determines based on the normalized error that the detection object waveform data is included in the input waveform data.

5. The abnormality detection device according to claim 4, wherein the first determining unit generates detection object range information, the detection object range information indicating a range corresponding to the detection object waveform data in the input waveform data, The second extraction unit extracts the detection object waveform data based on the detection object range information.

6. The abnormality detection device according to claim 1, wherein the first determining unit determines the type of the waveform data of the detection object by using an algorithm of a classification type neural network, The second determination unit includes a plurality of abnormality detection algorithms for the type of the detection object waveform data, switches the abnormality detection algorithm according to the determined type of the detection object waveform data, and determines whether the extracted detection object waveform data indicates abnormality by using the abnormality detection algorithm.

7. The abnormality detection device according to claim 1, wherein the first determining unit includes reference waveform data, and determines a correspondence between the input waveform data and the reference waveform data by using a dynamic time warping method to generate detection object range information, the detection object range information indicating a range corresponding to the detection object waveform data from the input waveform data, The second extraction unit extracts the detection object waveform data based on the detection object range information.

8. The abnormality detection device according to claim 7, When the first determination unit determines that the range is valid, the first determination unit notifies the range corresponding to the detection object waveform data.

9. The abnormality detection device according to claim 1, wherein the detection object waveform includes a plurality of detection object waveforms that are different from each other, A plurality of sets of the first extracting units and the first determining units are provided to correspond to each of the plurality of detection target waveforms, and each of the first determining units determines whether a corresponding detection target waveform is included in the input waveform data.

10. The abnormality detection device according to claim 1, The second determination unit uses a second algorithm utilizing a second autoencoder to calculate the degree of deviation between the extracted detection object waveform data to be input into the second autoencoder and the output data of the second autoencoder, and determines whether the detection object waveform data extracted by the second extraction unit indicates an abnormality.

11. An anomaly detection system comprising: The abnormality detection device according to claim 1; a data acquisition unit configured to acquire the input waveform data from the first extraction unit and to acquire the detection object waveform data from the second extraction unit, an algorithm generating unit configured to generate first learning data based on the input waveform data including the detection object waveform data, generate second learning data based on the detection object waveform data indicating normality, and generate a first algorithm and a second algorithm by using the first learning data and the second learning data, respectively, The first determining unit and the second determining unit perform determination by using the first algorithm and the second algorithm, respectively.

12. The anomaly detection system according to claim 11, The algorithm generation unit extracts the detection object waveform data from the input waveform data, and uses the waveform data obtained by shifting the extracted detection object waveform data along the time axis within a range where the number of points is greater than the extracted detection object waveform data as the first learning data.

13. The anomaly detection system according to claim 12, wherein the algorithm generation unit further generates, as the first learning data, waveform data obtained by expanding or contracting the detection object waveform data including the shifted detection object waveform data in the amplitude direction or the time axis direction.

14. An anomaly detection method using an anomaly detection device, comprising: Inputting input waveform data, wherein the input waveform data is time series waveform data output from the device to be detected; determining whether the detection object waveform data is included in the input waveform data; When it is determined that the detection object waveform data is included in the input waveform data, extracting the detection object waveform data from the input waveform data; determining whether the detection object waveform data indicates an abnormality, so as to determine whether the device has an abnormality, wherein it is determined whether the detection object waveform data is included in the input waveform data by using a first algorithm based on first learning data; wherein determining whether the detection object waveform data indicates abnormality by using a second algorithm based on second learning data; wherein the first learning data is generated based on input waveform data including the detection object waveform data among the input waveform data, and The second learning data is generated based on the detection object waveform data determined to be normal among the detection object waveform data.

Images