State detection device, state detection method, learning model generation method, and computer program
Through non-integer order differential equation model and machine learning technology, combined with the control and observation data of semiconductor manufacturing devices, a learning model is generated, which solves the problem of low state detection accuracy in the prior art, and realizes accurate detection of the state of semiconductor manufacturing devices and Exception recognition.
Patent Information
- Application Number
- CN202380034707.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-04-18
- Filing Date
- 2023-04-17
- Publication Date
- 2024-12-06
AI Technical Summary
The existing state detection technology is difficult to accurately detect the state of the semiconductor manufacturing device, especially in multiple operation modes and complex mechanical operations, resulting in low abnormal detection accuracy.
The non-integer order differential equation model is used to combine machine learning, and the control data and observation data of the semiconductor manufacturing device are obtained, the learning model is generated and state detection is performed, and the parameters and fitness of the inferred model are used for state classification.
Accurate detection of the status of the semiconductor manufacturing device is realized, high accuracy can be maintained in multiple operating modes, abnormal states and equipment failures of the robotic arm are identified, and the accuracy and reliability of state detection are improved.
Smart Images

Figure CN119096208A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to a state detection device, a state detection method, a method for generating a learning model, and a computer program. Background Art
[0002] Patent document 1 proposes a state determination device that acquires data related to industrial machinery, creates multiple partial time series data based on the acquired data related to the industrial machinery by sliding the time series data of physical quantities in the data related to the industrial machinery in the direction of the time axis, extracts multiple learning data including the multiple partial time series data, performs machine learning using the extracted learning data, and thereby generates a learning model.
[0003] Patent Document 1: Japanese Patent Application Publication No. 2020-128013 Summary of the invention
[0004] The present disclosure provides a state detection device, a state detection method, a learning model generation method, and a computer program, which are expected to detect the state of a target device such as a semiconductor manufacturing device with high accuracy.
[0005] According to one embodiment of the present disclosure, a state detection device comprises: a first acquisition unit, which acquires input data to an object device and observation data of the action of the object device; a determination unit, which determines parameters of an inference model based on the input data and observation data acquired by the first acquisition unit, and the inference model infers the observation data based on the input data; and a detection unit, which detects the state of the object device based on the parameters determined by the determination unit.
[0006] According to the present disclosure, it is expected that the state of a target device such as a semiconductor manufacturing device can be detected with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0007] Figure 1 This is a schematic diagram for explaining an example of the information processing system involved in this embodiment.
[0008] Figure 2 This is a block diagram showing a configuration example of a state detection device according to the present embodiment.
[0009] Figure 3 It is a schematic diagram showing an example of a state detection model involved in this embodiment.
[0010] Figure 4 This is a flowchart showing an example of the procedure of the learning data generation process performed by the state detection device according to the present embodiment.
[0011] Figure 5This is a schematic diagram for explaining an outline of a parameter determination process of an estimation model according to the present embodiment.
[0012] Figure 6 This is a flowchart showing an example of the procedure of the state detection model generation process performed by the state detection device according to the present embodiment.
[0013] Figure 7 This is a flowchart showing an example of the procedure of the state detection process performed by the state detection device according to the present embodiment.
[0014] Figure 8 This is a schematic diagram showing an example of a state detection model included in the state detection device according to Modification 1. DETAILED DESCRIPTION
[0015] Hereinafter, specific examples of the information processing system according to the embodiments of the present disclosure will be described with reference to the accompanying drawings. In addition, the present disclosure is not limited to these examples, and is intended to include all changes within the meaning and scope of the claims and equivalent to the claims.
[0016] <System Overview>
[0017] Figure 1 1 is a schematic diagram for explaining an example of an information processing system involved in the present embodiment. The information processing system involved in the present embodiment is configured to include a state detection device 1 and a semiconductor manufacturing device (target device) 3. The semiconductor manufacturing device 3 shown in the figure is a device for conveying wafers in a semiconductor manufacturing process. The semiconductor manufacturing device 3, for example, moves a fork holding a wafer by a mechanism such as a motor and a pulley (not shown), thereby conveying the wafer held by the fork. The state detection device 1 is a device that detects the state of the operation of the semiconductor manufacturing device 3 and controls the operation of the semiconductor manufacturing device 3 according to the detected operation.
[0018] The state detection device 1 and the semiconductor manufacturing device 3 are connected, for example, via a cable such as a communication line or a signal line. The state detection device 1 inputs control data for driving a motor that moves the fork to the semiconductor manufacturing device 3. The control data includes information such as the drive amount of the motor. The semiconductor manufacturing device 3 drives the motor according to the input control data, thereby moving the fork to transport the wafer. In addition, the semiconductor manufacturing device 3 is equipped with various sensors for observing the transport action of the wafer. The semiconductor manufacturing device 3 detects information related to the action of the device corresponding to the control data, such as the movement amount, the movement angle, the speed or the acceleration, etc., through the sensor, and outputs observation data containing the detected information to the state detection device 1. In addition, the sensor for obtaining the observation data can include, for example, a camera for taking images (still images or moving images), or a microphone for acquiring sounds. The observation data can include data such as images taken by the camera or sounds acquired by the microphone.
[0019] The state detection device 1 obtains control data input to the semiconductor manufacturing device 3, observation data output from the semiconductor manufacturing device 3 that operates according to the control data, and stores the control data and observation data as a record related to semiconductor manufacturing. Each time an action of the semiconductor manufacturing device 3 is completed, the state detection device 1 reads out the time series information of the control data and observation data related to the action, and detects the state of the semiconductor manufacturing device 3 based on the time series information. For example, an action of the semiconductor manufacturing device 3 can be various action units such as a fork starting to move from a certain location and stopping at another location, a fork holding a wafer at a certain location and then putting it down to another location, or a fork starting to move and then returning to the original location. In addition, the state of the semiconductor manufacturing device 3 detected by the state detection device 1 is set to two states of abnormality and no abnormality in this embodiment, but is not limited to this. The state of the semiconductor manufacturing device 3 can be, for example, a state in which the degree of abnormality from no abnormality to abnormality is classified into five levels, and in addition, for example, in addition to abnormality and no abnormality, it can also include a state immediately after maintenance or immediately after component replacement.
[0020] In addition, the state of the semiconductor manufacturing device 3 detected by the state detection device 1 can include, for example, a numerical value indicating the degree of abnormality. The numerical value indicating the degree of abnormality can be, for example, a decimal in the range of 0.0 to 1.0 or a percentage in the range of 0% to 100%. In addition, the numerical value indicating the degree of abnormality can be, for example, various numerical values such as the period until the predicted failure, the period required for repair, or the cost required for repair. In addition, the detected state can also include a state expressed by combining multiple numerical values, such as "failure occurs with a probability of more than 70% within 90 days".
[0021] In addition, the state of the semiconductor manufacturing device 3 detected by the state detection device 1 includes not only the state related to the future prediction, such as the possibility of abnormality of the semiconductor manufacturing device 3, but also the state related to the past inference, such as the occurrence of a certain phenomenon in the past or the probability of the occurrence of a certain phenomenon. The state related to the past inference can be various values such as whether the semiconductor manufacturing device 3 has been maintained, the estimated number of days of maintenance performed a few days ago, or the percentage of the possibility of maintenance performed within 10 days. In addition, the detected state can also include a state expressed by combining multiple numerical values, such as "maintenance is performed within 10 days with a probability of less than 10%".
[0022] The state detection device 1 involved in this embodiment performs a process of determining the parameters of the inference model for inferring the action of the semiconductor manufacturing device 3 based on the time series information of the control data and the observation data. The inference model is a model that accepts the control data as input and outputs the prediction data obtained by predicting the observation data. In this embodiment, a model of a non-integer differential equation is used. In addition, a non-integer differential equation can be called a fractional differential equation. Non-integer differential equations are known mathematical concepts (for example, see Nobumasa Sugimoto, "On Calculus of Non-Integer Order", Mathematical Society of Japan, February 2017, Mathematical Communications Vol. 21, No. 4, p. 5-22), so a detailed description is omitted. However, the inference model is not limited to a model of a non-integer differential equation. For example, it can be a model of an integer differential equation or a model represented by an arbitrary equation.
[0023] The state detection device 1 inputs control data to the inference model of the determined parameters, obtains the inference data output by the inference model and compares it with the original observation data, thereby calculating the fitness of the model. In the present embodiment, the fitness uses RMSE (Root Mean Squared Error), but it is not limited to this, and the fitness can also use an indicator different from RMSE. The state detection device 1, for example, compares the calculated fitness with a predetermined threshold value, and detects the state of the semiconductor manufacturing device 3 based on the comparison result. For example, when the calculated RMSE exceeds the threshold value, the state detection device 1 can detect that the semiconductor manufacturing device 3 is abnormal.
[0024] In addition, the state detection device 1 according to the present embodiment detects the state of the semiconductor manufacturing device 3 based on the parameters of the determined inference model. In the information processing system according to the present embodiment, the parameters of the inference model are accepted as input, and a learning model for classifying the state of the semiconductor manufacturing device 3 is generated in advance by machine learning. The state detection device 1 includes a pre-generated learning model, inputs the parameters of the determined inference model into the learning model, obtains the classification result of the state output by the learning model, and thereby detects the state of the semiconductor manufacturing device 3.
[0025] In addition, the learning model used by the state detection device 1 can also perform machine learning as follows: the above-mentioned fitness and parameters are accepted as inputs, and the state of the semiconductor manufacturing device 3 is classified. The state detection device 1 determines the parameters of the inference model, inputs the control data into the inference model to calculate the fitness, and inputs the calculated fitness and the determined parameters of the inference model into the learning model that has been machine-learned in advance. The state detection device 1 can detect the state of the semiconductor manufacturing device 3 by acquiring the output information of the learning model. In addition, the learning model can also perform machine learning as follows: information different from the fitness and parameters is accepted as input, and the state of the semiconductor manufacturing device 3 is classified.
[0026] When the state detection device 1 detects that the state of the semiconductor manufacturing apparatus 3 is abnormal, for example, the state detection device 1 stops the operation of the semiconductor manufacturing apparatus 3 and displays a warning message on a display unit such as a liquid crystal display.
[0027] In this embodiment, the target device whose state is detected by the state detection device 1 is the semiconductor manufacturing device 3 that conveys wafers, but the present invention is not limited thereto. The target device for state detection may be any semiconductor manufacturing device or may be a device other than a semiconductor manufacturing device.
[0028] <Device Structure>
[0029] Figure 2 1 is a block diagram showing a configuration example of the state detection device 1 according to the present embodiment. The state detection device 1 according to the present embodiment is configured to include a processing unit 11, a storage unit (memory) 12, a communication unit (transceiver) 13, a display unit 14, and an operation unit 15. In addition, in the present embodiment, the description is made assuming that the processing is performed by one state detection device 1, but the processing of the state detection device 1 may be performed in a distributed manner by a plurality of devices.
[0030] The processing unit 11 is composed of a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), a GPU (Graphics Processing Unit), or a quantum processor, a ROM (Read Only Memory), and a RAM (Random Access Memory). The processing unit 11 reads and executes a program 12a stored in the storage unit 12 to perform various processes such as a process for controlling the operation of the semiconductor manufacturing device 3 and a process for detecting the state of the semiconductor manufacturing device 3.
[0031] The storage unit 12 is configured using a large-capacity storage device such as a hard disk. The storage unit 12 stores various programs executed by the processing unit 11 and various data required for the processing of the processing unit 11. In the present embodiment, the storage unit 12 stores a program 12a executed by the processing unit 11. In addition, the storage unit 12 is provided with a history storage unit 12b storing a history of control data and observation data, a model information storage unit 12c storing information related to a learning model for classifying the state of the semiconductor manufacturing device 3, and a learning data storage unit 12d storing learning data used for machine learning for generating the learning model.
[0032] In the present embodiment, the program (computer program, program product) 12a is provided in a manner recorded on a recording medium 99 such as a memory card or an optical disk, and the state detection device 1 reads the program 12a from the recording medium 99 and stores it in the storage unit 12. However, the program 12a may be written to the storage unit 12, for example, during the manufacturing stage of the state detection device 1. In addition, for example, the program 12a may be obtained by the state detection device 1 through communication from a remote server device or the like. For example, the program 12a may be read by a writing device from the content recorded on the recording medium 99 and written to the storage unit 12 of the state detection device 1. The program 12a may be provided in a manner of distribution via a network, or may be provided in a manner of being recorded on the recording medium 99.
[0033] The history storage unit 12b of the storage unit 12 stores and accumulates information that establishes correspondence between control data input from the state detection device 1 to the semiconductor manufacturing device 3 and observation data output from the semiconductor manufacturing device 3 based on the control data in a time series. The history storage unit 12b maintains the time series information of the acquired control data and observation data at least during one action performed by the semiconductor manufacturing device 3, that is, until the state detection device 1 detects that the data required for the state of the semiconductor manufacturing device 3 are consistent. After the state detection device 1 detects the state of the semiconductor manufacturing device 3 based on the time series information of one action, the history storage unit 12b may discard the time series information of the control data and observation data, or may continue to maintain it.
[0034] The model information storage unit 12c stores information related to the model used by the state detection device 1 for state detection of the semiconductor manufacturing device 3. The information related to the model may include, for example, structural information indicating how the model is structured, and information such as the values of parameters inside the model. The state detection device 1 involved in this embodiment uses two models for state detection. The first model is a model of a non-integer differential equation that accepts input of control data and outputs prediction data obtained by predicting observed data. The parameters of the first model are determined each time the state detection device 1 performs state detection of the semiconductor manufacturing device 3. The first learning model is hereinafter referred to as an "inference model".
[0035] The second model is a learning model that accepts the parameters of the first model (inference model) as input and classifies the state of the semiconductor manufacturing device 3. The second model is a learned model that has been pre-learned to determine the parameters by machine learning, and can adopt various learning models of various structures such as SVM (Support Vector Machine), decision tree, random forest or DNN (Deep Neural Network). In the present embodiment, the state detection device 1 uses, for example, learning data prepared in advance by the developer or operator of the information processing system to perform machine learning processing, thereby pre-generating a learning model. The state detection device 1 stores information such as the parameters of the pre-generated learning model in the model information storage unit 12c. However, the state detection device 1 may also obtain a learned learning model generated by other devices without performing machine learning processing, and store the acquired learning model in the model information storage unit 12c for use. The second learning model is hereinafter referred to as the "state detection model".
[0036] The learning data storage unit 12d stores learning data used for machine learning that generates a state detection model. In the present embodiment, the learning data is data that establishes a correspondence between the parameters of the inference model and the values representing the state of the semiconductor manufacturing device 3. The learning data is data that can be called teacher data used in so-called supervised machine learning. The learning data is generated in advance, for example, by collecting control data and observation data by causing the semiconductor manufacturing device 3 to operate. The developer or operator of the information processing system involved in the present embodiment performs a task of assigning a value representing the state of the semiconductor manufacturing device 3, a so-called annotation task, to the collected control data and observation data. The state detection device 1 obtains data that establishes a correspondence between the control data and the observation data and the values representing the state of the semiconductor manufacturing device 3, and determines the parameters of the inference model corresponding to the control data and the observation data. The state detection device 1 can generate learning data by establishing a correspondence between the value representing the state of the semiconductor manufacturing device 3 that is established with the control data and the observation data and the determined parameters.
[0037] The communication unit 13 of the state detection device 1 is connected to the semiconductor manufacturing device 3 via a cable such as a communication line or a signal line, and transmits and receives data with the semiconductor manufacturing device 3 via the cable. In the present embodiment, the communication unit 13 transmits the control data provided by the processing unit 11 to the semiconductor manufacturing device 3. In addition, the communication unit 13 receives the observation data transmitted from the semiconductor manufacturing device 3, and provides the received observation data to the processing unit 11.
[0038] The display unit 14 is configured using a liquid crystal display or the like, and displays various images and characters based on the processing of the processing unit 11. In the present embodiment, the display unit 14 displays, for example, various information related to the operating state of the semiconductor manufacturing device 3, and displays a warning message or the like when an abnormality is detected in the operating state of the semiconductor manufacturing device 3.
[0039] The operation unit 15 receives the user's operation and notifies the processing unit 11 of the received operation. For example, the operation unit 15 receives the user's operation through an input device such as a mechanical button or a touch panel provided on the surface of the display unit 14. In addition, for example, the operation unit 15 may be an input device such as a mouse and a keyboard, and these input devices may also be detachable from the state detection device 1.
[0040] In addition, the storage unit 12 may be an external storage device connected to the state detection device 1. In addition, the state detection device 1 may be a multi-computer including a plurality of computers, or a virtual machine virtually constructed by software. In addition, the state detection device 1 is not limited to the above-mentioned structure, and for example, the display unit 14 and the operation unit 15 may not be provided.
[0041] In addition, in the state detection device 1 involved in this embodiment, the processing unit 11 reads out and executes the program 12a stored in the storage unit 12, thereby controlling the processing unit 11a, the parameter determination unit 11b, the state detection unit 11c, and the machine learning processing unit 11d as functional units of the software and being implemented by the processing unit 11.
[0042] The control processing unit 11a generates control data and transmits it to the semiconductor manufacturing device 3 according to a predetermined semiconductor manufacturing sequence, thereby performing processing to control the operation of the semiconductor manufacturing device 3. For example, the control processing unit 11a transmits control data including a command to drive the motor of the semiconductor manufacturing device 3 and information on the driving amount to the semiconductor manufacturing device 3, thereby driving the motor of the semiconductor manufacturing device 3 to move the fork holding the wafer and transport the wafer. In addition, the control processing unit 11a obtains the observation data transmitted from the semiconductor manufacturing device 3 as a response to the control data, establishes a correspondence between the control data and the observation data, and stores them in the history storage unit 12b of the storage unit 12. The control processing unit 11a repeatedly transmits the control data and obtains the observation data, and accumulates the time series information of the combination of the control data and the observation data in the history storage unit 12b.
[0043] When an action of the semiconductor manufacturing device 3 under the control of the control processing unit 11a is completed, the parameter determination unit 11b reads the control data and observation data related to the action from the history storage unit 12b. The parameter determination unit 11b performs processing to determine the parameters of the prediction model for predicting the observation data based on the control data. In this embodiment, the prediction model is a model represented by a non-integer order differential equation represented by the following formula (1).
[0044] [Mathematical formula 1]
[0045]
[0046] In formula (1), x(t) is the time series input data (control data) input to the inference model, and y(t) is the time series observation data corresponding to the input control data. k and x0 is the coefficient, ν k is the order of differentiation. k , ν k And x0 is a parameter determined by parameter determination unit 11b. In formula (1), K is a constant that determines the number of terms in the non-integer order differential equation, and is a value predetermined by the developer of the information processing system according to the present embodiment based on the characteristics of semiconductor manufacturing apparatus 3, etc.
[0047] The parameter determination unit 11b sequentially inputs the control data of the time series to the inference model of the non-integer order differential equation that appropriately determines the initial value of the parameter, and sequentially obtains the inference data of the time series output by the inference model. The parameter determination unit 11b compares the observation data corresponding to the control data with the inference data inferred by the inference model to calculate the error, and determines the parameters of the inference model by the difference method based on the calculated error. The parameter determination unit 11b can determine the parameters of the inference model by, for example, calculating the gradient related to each parameter based on the error, and using the steepest descent method, quasi-Newton method or SGD (Stochastic Gradient Descent, stochastic gradient descent method) and other methods based on the calculated gradient. In addition, the method for determining the parameters of the inference model is a prior art, so the detailed description is omitted.
[0048] The parameter determination unit 11b inputs control data again to the estimation model based on the determined parameters to obtain estimation data, calculates the error between the estimation data and the observation data again, and compares the calculated error with a preset threshold value. When the calculated error is larger than the threshold value, the parameter determination unit 11b updates the parameter based on the error. The parameter determination unit 11b repeatedly calculates the error and updates the parameter, and when the calculated error is smaller than the threshold value, the parameter at this time is finally determined as the parameter of the estimation model.
[0049] The state detection unit 11c detects the state of the semiconductor manufacturing device 3 using two methods in this embodiment. In the first method, the state detection unit 11c uses the inference model in which the parameters determined by the parameter determination unit 11b are set and the control data and observation data used for the determination of the parameters to calculate the fitness of the inference model. In this embodiment, the state detection unit 11c inputs the control data into the inference model to obtain the prediction data, and calculates the RMSE of the observation data and the prediction data as the fitness. In addition, in this embodiment, since RMSE is used as the fitness, the smaller the fitness value is, the more suitable the inference model is. The fitness can also use the inverse of RMSE, in which case the larger the fitness value is, the more suitable the inference model is. The state detection unit 11c detects the state of the semiconductor manufacturing device 3 based on the comparison between the calculated RMSE and a predetermined threshold value. In addition, when the parameter determination unit 11b calculates the RMSE as an error when determining the parameters, the state detection unit 11c does not need to recalculate the RMSE, and can obtain the RMSE calculated by the parameter determination unit 11b and compare it with the threshold value. For example, the state detection unit 11 c can detect that the state of the semiconductor manufacturing apparatus 3 is abnormal when the RMSE exceeds a threshold value, and can detect that there is no abnormality when the RMSE does not exceed the threshold value.
[0050] In the second method, the state detection unit 11 c detects the state of the semiconductor manufacturing apparatus 3 using the parameters determined by the parameter determination unit 11 b and the state detection model stored in the model information storage unit 12 c . Figure 3 : is a schematic diagram showing an example of a state detection model involved in this embodiment. The state detection model involved in this embodiment is a classification model that accepts the parameters of the inference model as input and outputs a classification result obtained by classifying the state of the semiconductor manufacturing device 3. In this example, the inference model takes μ0, ..., μ K ,ν1,…,ν K , x0 as input, and classifies whether the state of the semiconductor manufacturing device 3 is "abnormal" or "no abnormality". The state detection device 1 generates a state detection model by performing so-called supervised machine learning in advance, and stores information related to the generated state detection model in the model information storage unit 12c of the storage unit 12.
[0051] The state detection unit 11c inputs the parameters of the inference model determined by the parameter determination unit 11b into the state detection model, and obtains information related to the state of the semiconductor manufacturing device 3 output by the state detection model. For example, the state detection model can output two numerical values (confidence, likelihood, etc.) corresponding to the two states of "abnormality" and "no abnormality", and the state detection unit 11c can set the state corresponding to the larger of the two numerical values as the state of the semiconductor manufacturing device 3.
[0052] In addition, the output of the state detection model may also be categories added to multiple states of the semiconductor manufacturing device 3. For example, when there are three states (categories) of fault, maintenance, and normal as the states of the semiconductor manufacturing device 3, the cases where "T" is assigned to fault, "M" is assigned to maintenance, and "R" is assigned to normal can be set as categories. In addition, in this example, the categories are represented by text such as "T", "M", and "R", but it is not limited to this, and the categories may also be represented by numerical values such as "0", "1", and "2". The state detection model performs machine learning in advance to predict these categories.
[0053] In addition, the output of the state detection model may be a numerical value within a specified range related to the state of the semiconductor manufacturing device 3. For example, various numerical values such as the probability of failure of the semiconductor manufacturing device 3, the probability of maintenance, the probability of failure within a specified period, the number of days until failure, and the number of days after maintenance can be output from the state detection model.
[0054] The state detection unit 11c detects the state of the semiconductor manufacturing apparatus 3 by the first method and the second method described above, and finally detects that the state of the semiconductor manufacturing apparatus 3 is abnormal when the detection result that the state of the semiconductor manufacturing apparatus 3 is abnormal is obtained by at least one of the methods. When the state detection unit 11c obtains the detection result that there is no abnormality by both the first method and the second method, it finally detects that there is no abnormality in the state of the semiconductor manufacturing apparatus 3. When the state detection unit 11c detects that the state of the semiconductor manufacturing apparatus 3 is abnormal, it performs processing such as displaying a warning message on the display unit 14 and stopping the operation of the semiconductor manufacturing apparatus 3.
[0055] In addition, the state detection unit 11c may perform state detection (detection of abnormality) using both the fitness (RMSE) and the parameters instead of performing state detection (detection of abnormality) separately as described above. For example, the state detection unit 11c may use a learning model that has been pre-machined as follows, accept the fitness and the parameters as input, and classify the state of the semiconductor manufacturing device 3.
[0056] Before performing state detection of the semiconductor manufacturing device 3, the machine learning processing unit 11d performs machine learning processing using the learning data stored in the learning data storage unit 12d of the storage unit 12, thereby generating a state detection model. The learning data stored in the learning data storage unit 12d is data that establishes a correspondence between the parameters of the inference model and the state of the semiconductor manufacturing device 3. The machine learning processing unit 11d generates a state detection model by performing so-called supervised machine learning processing using the learning data. The supervised machine learning processing is a conventional technology, so in this embodiment, a detailed description is omitted. The machine learning processing unit 11d stores information related to the state detection model generated by machine learning in the model information storage unit 12c.
[0057] <State detection model generation process>
[0058] In the information processing system according to the present embodiment, before the state detection device 1 performs the state detection of the semiconductor manufacturing device 3, the following processing is performed to collect and generate learning data for generating a state detection model by machine learning. In addition, in the information processing system, the following processing is performed in advance to generate a state detection model by machine learning using the learning data. In the present embodiment, the description is made assuming that the state detection device 1 performs the processing related to the generation of these state detection models, but the present invention is not limited to this, and these processing may also be performed by a device other than the state detection device 1.
[0059] Figure 41 is a flowchart showing an example of the sequence of the learning data generation process performed by the state detection device 1 according to the present embodiment. The processing unit 11 of the state detection device 1 according to the present embodiment stores and accumulates the control data and the observation data in the history storage unit 12b as the semiconductor manufacturing device 3 operates. The machine learning processing unit 11d of the processing unit 11 appropriately obtains the control data and the observation data of one operation of the semiconductor manufacturing device 3 from the data stored in the history storage unit 12b (step S1).
[0060] The parameter determination unit 11b of the processing unit 11 inputs the control data obtained in step S1 to the estimation model of the non-integer order differential equation (step S2), and obtains the estimation data output by the estimation model (step S3). The parameter determination unit 11b calculates the error between the observation data obtained in step S1 and the estimation data obtained in step S3 (step S4). The parameter determination unit 11b determines whether the error calculated in step S4 is less than a predetermined threshold value (step S5). When the error is greater than the threshold value (S5: No), the parameter determination unit 11b updates the parameters of the estimation model so that the error becomes smaller (step S6), and returns the processing to step S2. The parameter determination unit 11b repeatedly updates the parameters by repeatedly performing the processing of steps S2 to S6, and when it is determined that the error is less than the threshold value (S5: Yes), it transfers the processing to step S7.
[0061] The machine learning processing unit 11d obtains information related to the state of the operation of the semiconductor manufacturing device 3 for the group of control data and observation data obtained in step S1 (step S7). Here, the information related to the state of the operation of the semiconductor manufacturing device 3 is the presence or absence of an abnormality in the semiconductor manufacturing device 3 in this embodiment. The developer or manager of the information processing system involved in this embodiment can determine the presence or absence of an abnormality in the semiconductor manufacturing device 3 based on the group of control data and observation data. For example, the situation in which the developer or manager sets data "1" for an abnormality and sets data "0" for no abnormality can be set as information related to the state of the operation of the semiconductor manufacturing device 3. This operation based on the developer or manager is called the so-called annotation operation. The annotation operation can be performed before the learning data generation process shown in this flowchart is implemented, and the annotation operation can be performed simultaneously by accepting input from the developer or manager during the learning data generation process.
[0062] The machine learning processing unit 11d stores the parameters of the inference model determined by the processing of steps S2 to S6 and the state of the semiconductor manufacturing device 3 acquired in step S7 as learning data in the learning data storage unit 12d (step S8), and ends the processing. Figure 4The process in the flowchart shown is a procedure for generating one piece of learning data, and the state detection device 1 can generate a plurality of pieces of learning data by repeatedly performing this process.
[0063] Figure 5 This is a schematic diagram for explaining an outline of a parameter determination process of an estimation model according to the present embodiment. Figure 5 A graph is shown in which the horizontal axis is time t and the vertical axis is the value of the control data x or the observation data y. Figure 5 In the figure, the control data x input to the semiconductor manufacturing device 3 is represented by a thin solid line, the observation data output by the semiconductor manufacturing device 3 is represented by a thick solid line, and the estimated data obtained by estimating the observation data y based on the control data x by the estimation model is represented by a dotted line. In the illustrated chart, the control data x, the observation data y, and the estimated data are shown for one action of the semiconductor manufacturing device 3. In addition, the error of the observation data y and the estimated data at a certain moment in one action is represented by a thin double arrow.
[0064] When there are N sets of control data x and observation data y for one action of the semiconductor manufacturing device 3, the state detection device 1 uses the inference model to obtain inference data for each of the N control data x, thereby obtaining N inference data. The state detection device 1 can obtain N error data by calculating the difference between the two corresponding data for the N observation data and the N inference data. The state detection device 1 can calculate RMSE by calculating, for example, the square root of the average of the squares for the N error data, and set the value of the RMSE to the error for the entire action (in Figure 4 The error calculated in step S4 of the flowchart).
[0065] In addition, in this example, the case of one-dimensional control data x and one-dimensional observation data y is used as an example for explanation, but the control data and the observation data can be vectors of two or more dimensions. In addition, the dimension of the control data does not need to be consistent with the dimension of the observation data. For example, the control data can be one-dimensional and the observation data can be three-dimensional. In this case, the inference model becomes a model that accepts a one-dimensional value as input and outputs a three-dimensional vector.
[0066] The state detection device 1 repeatedly updates the parameters of the non-integer order differential equation of the estimation model in the direction in which the calculated error becomes smaller, thereby obtaining the parameters of the estimation model that accurately estimates the observed data of the operation of the semiconductor manufacturing device 3 relative to the control data. The state detection device 1 generates data as learning data in which the parameters of the estimation model obtained in this way are associated with information indicating the state of the semiconductor manufacturing device 3 based on the control data and the observed data at that time. The state detection device 1 generates and accumulates as much learning data as possible, and performs machine learning processing using the learning data, thereby generating a state detection model.
[0067] Figure 6 1 is a flowchart showing an example of the sequence of the state detection model generation process performed by the state detection device 1 involved in the present embodiment. The machine learning processing unit 11d of the processing unit 11 of the state detection device 1 of the present embodiment appropriately obtains one piece of learning data from a plurality of learning data stored in the learning data storage unit 12d of the storage unit 12 (step S21). The machine learning processing unit 11d inputs the data of the parameters contained in the learning data obtained in step S1 to the learning model (not learned or in the process of learning) (step S22). The learning model is a structure that outputs a numerical value or category corresponding to the state of the semiconductor manufacturing device 3 with respect to the input of the parameter, and the machine learning processing unit 11d obtains the output of the learning model with respect to the data input in step S22 (step S23). The machine learning processing unit 11d calculates the error (cross entropy error) between the numerical value or category corresponding to the state of the semiconductor manufacturing device 3 contained in the learning data obtained in step S21 and the numerical value or category output by the learning model obtained in step S23 (step S24). The machine learning processing unit 11 d updates the parameters of the learning model based on the error calculated in step S24 so as to reduce the error (step S25 ).
[0068] The machine learning processing unit 11d determines whether the processing of steps S21 to S25 has been completed for all the learning data stored in the learning data storage unit 12d (step S26). If the processing has not been completed for all the learning data (S26: No), the machine learning processing unit 11d returns the processing to step S21, obtains other learning data and performs the same processing. If the processing has been completed for all the learning data (S26: Yes), the machine learning processing unit 11d determines whether the last calculated error is less than a predetermined threshold (step S27). If the error is greater than the threshold (S27: No), the machine learning processing unit 11d returns the processing to step S21, and performs the processing of steps S21 to S26 using the same learning data. If the error is less than the threshold (S27: Yes), the machine learning processing unit 11d stores the information of the learning model including the finally determined parameters as the information of the state detection model in the model information storage unit 12c (step S28), and ends the processing.
[0069] <Status Detection Process>
[0070] The state detection device 1 according to the present embodiment performs the following processing: controls the action of the semiconductor manufacturing device 3, and uses the learned state detection model to detect the state of the semiconductor manufacturing device 3 (in the present embodiment, the presence or absence of abnormality) for a completed action. For example, the state detection processing of the state detection device 1 is performed each time an action of the semiconductor manufacturing device 3 is completed. In addition, the state detection device 1 according to the present embodiment also performs state detection based on the fitness of the inference model, which determines parameters based on the control data and observation data of the semiconductor manufacturing device 3.
[0071] Figure 7 4 is a flowchart showing an example of the sequence of the state detection processing performed by the state detection device 1 involved in the present embodiment. The control processing unit 11a of the processing unit 11 of the state detection device 1 involved in the present embodiment controls the operation by sending control data to the semiconductor manufacturing device 3. The control processing unit 11a obtains the observation data sent from the semiconductor manufacturing device 3 as the operation result, and stores the control data and the observation data in the history storage unit 12b. The state detection unit 11c of the processing unit 11 obtains the control data and the observation data of one operation amount of the semiconductor manufacturing device 3 that was operated immediately before from these data stored in the history storage unit 12b (step S41).
[0072] The parameter determination unit 11b of the processing unit 11 inputs the control data obtained in step S41 to the estimation model of the non-integer order differential equation (step S42), and obtains the estimation data output by the estimation model (step S43). The parameter determination unit 11b calculates the error between the observation data obtained in step S41 and the estimation data obtained in step S43 (step S44). The parameter determination unit 11b determines whether the error calculated in step S44 is less than a prescribed threshold value (step S45). When the error is greater than the threshold value (S45: No), the parameter determination unit 11b updates the parameters of the estimation model to reduce the error (step S46), and returns the process to step S42. The parameter determination unit 11b repeatedly updates the parameters by repeatedly performing the processes of steps S42 to S46, and when it is determined that the error is less than the threshold value (S45: Yes), the process is transferred to step S47. In addition, the processes of steps S42 to S46 performed by the parameter determination unit 11b in this flowchart are the same as those in Figure 4 The processing of steps S2 to S6 performed by the parameter determination unit 11 b in the flowchart shown is the same.
[0073] The state detection unit 11c calculates the fitness for the inference model whose parameters are determined by the processing of steps S42 to S46 (step S47). At this time, the state detection unit 11c can input control data to the inference model that determines the parameters to obtain inference data, calculate the RMSE of the observed data and the inference data, and set the calculated RMSE as the fitness.
[0074] In addition, the state detection unit 11c inputs the parameters of the inference model determined by the processing of steps S42 to S46 into the state detection model pre-stored in the model information storage unit 12c (step S48). The state detection unit 11c obtains information related to the state of the semiconductor manufacturing device 3 output by the state detection model (step S49). In this embodiment, the information output by the state detection model is information indicating whether there is an abnormality in the operation of the semiconductor manufacturing device 3.
[0075] The state detection unit 11c determines whether there is an abnormality in the operation of the semiconductor manufacturing device 3 based on the fitness calculated in step S47 and the information obtained in step S48 (step S50). For example, when the RMSE calculated in step S47 exceeds a preset threshold value, or when the information obtained from the state detection model in step S49 is "abnormality", the state detection unit 11c can determine that there is an abnormality in the operation of the semiconductor manufacturing device 3. In addition, when the RMSE calculated as the fitness in step S47 does not exceed the preset threshold value, and the information obtained from the state detection model in step S49 is "no abnormality", the state detection unit 11c can determine that there is no abnormality in the operation of the semiconductor manufacturing device 3. When there is no abnormality in the operation of the semiconductor manufacturing device 3 (S50: No), the state detection unit 11c ends the processing. When there is an abnormality (S50: Yes), the state detection unit 11c displays a warning message on the display unit 14 (step S51), for example, and ends the processing.
[0076] <Modification>
[0077] The state detection device 1 according to the present embodiment is not limited to the above-described structure, and various modifications can be adopted. Several modifications of the state detection device 1 are described below. However, the structure of the state detection device 1 is not limited to the above-described structure and the structures of the following modifications.
[0078] (Variant 1)
[0079] Figure 8 1 is a schematic diagram showing an example of a state detection model provided in the state detection device 1 according to the modification example 1. The state detection model according to the modification example 1 receives the parameters of the inference model as input, and receives the environmental information as input, and outputs information related to the state of the semiconductor manufacturing device 3. The environmental information includes, for example, the operation mode (high-speed mode or low-speed mode, etc.) of the semiconductor manufacturing device 3, information related to the wafer processed by the semiconductor manufacturing device 3, or information related to the environment around the semiconductor manufacturing device 3 (temperature or humidity, etc.). The semiconductor manufacturing device 3 sends the environmental information to the state detection device 1 together with the observation data. The state detection device 1 stores the environmental information in the history storage unit 12b in correspondence with the control data and the observation data.
[0080] The learning data used for the machine learning for generating the state detection model according to the modification 1 is data in which the parameters and environmental information of the inference model are associated with the state of the semiconductor manufacturing apparatus 3. The state detection apparatus 1 generates the learning data by associating the information on the state of the semiconductor manufacturing apparatus 3 set by the developer or the manager, etc., with the parameters determined based on the control data and observation data stored in the history storage unit 12b and the environmental information stored in the history storage unit 12b.
[0081] In addition, when the state detection device 1 involved in variant example 1 performs state detection on the semiconductor manufacturing device 3, it determines the parameters of the inference model based on the control data input to the semiconductor manufacturing device 3 and the observation data obtained from the semiconductor manufacturing device 3, obtains environmental information from the semiconductor manufacturing device 3, and inputs the parameters and environmental information into the state detection model, thereby obtaining information related to the state of the semiconductor manufacturing device 3 output by the state detection model.
[0082] (Variant 2)
[0083] The state detection device 1 involved in the modification example 2 does not update the parameters of the inference model for all parameters, but only updates the parameters of the inference model for a part of the parameters. That is, the state detection device 1 involved in the modification example 2, for example, when there are ten parameters of the inference model, determines seven parameters based on the control data and the observation data, and uses the preset values for the remaining three parameters. The fixed parameters that are not dynamically determined are preset by, for example, the developer of the information processing system involved in this embodiment.
[0084] In addition, the state detection device 1 according to Modification 2 may prepare a plurality of fixed parameters corresponding to the environmental information when using the environmental information described in Modification 1. The state detection device 1 reads the fixed parameters corresponding to the environmental information acquired from the semiconductor manufacturing device 3 and sets them as part of the parameters of the inference model, and determines the remaining parameters based on the control data and the observation data.
[0085] (Variant 3)
[0086] As the inference model involved in this embodiment, a model based on a non-integer order differential equation of formula (1) is adopted. However, the inference model is not based on the model of formula (1), and a model based on various mathematical formulas can be adopted. The following describes several mathematical formulas that can be adopted as inference models. In addition, in the following mathematical formulas, γ1…γ K and μ1…μ k are coefficients, θ1…θ k and ν1…ν Kis the order of differentiation. K is an integer that determines the complexity of the model and is a value predetermined by the developer of the information processing system of this embodiment. f and g are functions for preprocessing data by elementary functions, linear calculations, statistical processing, principal component analysis, or machine learning models, for example. The parameters contained in these functions can also become parameters of the inference model.
[0087] [Mathematical formula 2]
[0088]
[0089] This formula (2) is a generalized mathematical formula of formula (1). In formula (2), the case where L=0, γ0=1, f(y, t)=y, and g(x, t)=x-x0 is the same as formula (1).
[0090] [Mathematical formula 3]
[0091]
[0092] In the formula (3), x0 and y0 can also be parameters of the estimation model.
[0093] [Formula 4]
[0094]
[0095] In the formula (4), t0 can also be a parameter of the estimation model.
[0096] [Formula 5]
[0097]
[0098] This formula (5) is used when x and y are 1 components. n 、b n , x0 and y0 can be either parameters of the inference model or predetermined constants.
[0099] [Mathematical formula 6]
[0100]
[0101] In formula (6), A is a d1-dimensional matrix, B is a d2-dimensional matrix, y(t) and y0 are d1-dimensional vectors, and x(t) and x0 are d2-dimensional vectors.
[0102] [Formula 7]
[0103]
[0104] In the formula (7), F and G are neural networks, and A and B are weight parameters of the neural network. The weight parameters can be determined in advance by a method such as machine learning.
[0105] (Variant 4)
[0106] In the information processing system according to the present embodiment described above, each time the state detection of the semiconductor manufacturing device 3 is performed, the state detection device 1 updates the parameters of the inference model based on the control data and the observation data, and detects the state based on the updated parameters, but the present invention is not limited thereto. The state detection device 1 of the information processing system according to the modification 4 does not update the parameters of the inference model each time the state detection of the semiconductor manufacturing device 3 is performed, but detects the state of the semiconductor manufacturing device 3 using the inference model with parameters predetermined.
[0107] The state detection device 1 according to the modification 4 includes an estimation model with predetermined parameters. The state detection device 1 acquires control data and observation data for the semiconductor manufacturing device 3 that is the object of state detection, and inputs the acquired control data into the estimation model to acquire prediction data of the observation data. The state detection device 1 compares the observation data and the prediction data, and can determine that the state of the semiconductor manufacturing device 3 is abnormal, for example, when the difference between the observation data and the prediction data exceeds a threshold.
[0108] The inference model provided in the state detection device 1 according to the modification 4 determines parameters based on, for example, control data and observation data acquired during past operations of the semiconductor manufacturing device 3 to be detected. Alternatively, for example, the inference model may determine parameters based on control data and observation data acquired in another device having the same structure as the semiconductor manufacturing device 3 to be detected.
[0109] Summary
[0110] In the information processing system according to the present embodiment having the above structure, the control processing unit (first acquisition unit) 11a of the state detection device 1 acquires the control data (input data) input to the semiconductor manufacturing device (target device) 3 and the observation data of the operation of the semiconductor manufacturing device 3. The parameter determination unit (determination unit) 11b determines the parameters of the estimation model for estimating the observation data from the control data based on the acquired control data and observation data, and the state detection unit (detection unit) 11c detects the state of the semiconductor manufacturing device 3 based on the determined parameters. As a result, the information processing system can be expected to realize the state detection of the semiconductor manufacturing device 3 with higher accuracy compared with the case where the state is detected only based on the input and output data with respect to the semiconductor manufacturing device 3.
[0111] For example, when there are multiple operation modes of the semiconductor manufacturing device 3, the input and output data are distributed in multiple clusters (groups) or continuous ranges, and it is difficult to detect abnormal or normal states based on statistics, etc. In the information processing system involved in this embodiment, since the relationship between the input and output data itself can be learned, it is not easily affected even when there are multiple operation modes.
[0112] For example, the semiconductor manufacturing device 3 has a robot arm that moves three-dimensionally. When the robot arm grabs irregularly arranged objects and arranges them in an orderly manner, the robot arm performs completely different actions each time. Therefore, when input or output data that deviates from previous statistics is obtained, it is generally difficult to determine whether this is an abnormality of the robot arm or the robot arm is normal but the position or weight of the object is a value that has not been previously determined. The information processing system involved in this embodiment can be expected to perform state detection such as excessive friction of the robot arm joints or the end of the life of the motor based on the relationship between the input and output data, regardless of the configuration of the object and how the robot arm moves.
[0113] In addition, in the information processing system involved in the present embodiment, a state detection model (learning model) is pre-generated by machine learning using learning data that establishes a correspondence between the parameters of the inference model and the state of the semiconductor manufacturing device 3. The state detection model is a learning model that accepts and outputs the state of the semiconductor manufacturing device 3 using the parameters of the inference model as input. The state detection device 1 has a pre-generated state detection model, inputs the parameters of the inference model determined based on the control data and the observation data into the state detection model, obtains the state of the semiconductor manufacturing device 3 output by the state detection model, and thereby detects the state. As a result, the information processing system can expect to infer the state of the device based on the change in the parameters even when the fitness does not change significantly but only the parameters change.
[0114] In addition, in the information processing system according to the present embodiment, for the inference model that determines the parameters, the state detection unit (second acquisition unit, calculation unit) 11c of the state detection device 1 acquires the inference data for the control data, and calculates the fitness (e.g., RMSE) based on the observation data and the inference data. The state detection unit 11c detects the state of the semiconductor manufacturing device 3 based on the calculated fitness. In addition to performing state detection based on the parameters of the inference model, the state detection based on the fitness of the inference model is also performed, so that the information processing system can be expected to perform state detection of the semiconductor manufacturing device 3 with higher accuracy.
[0115] In addition, in the information processing system involved in the present embodiment, a model of a non-integer order differential equation is used as an inference model. Compared with the model of the integer order differential equation, the model of the non-integer order differential equation can accurately model the input and output characteristics of the target device such as the semiconductor manufacturing device 3, so it can be expected that the state detection device 1 can accurately detect the state of the semiconductor manufacturing device 3. In the non-integer order differential equation, as a special case of limiting the order of differentials to non-negative integers, the usual integer order differential equations (ordinary differential equations, etc.) are included. Therefore, if compared with the case of the same number of parameters, the non-integer order differential equation has a larger solution space than the integer order differential equation, and the descriptive ability as a model is higher. In addition, the non-integer order differential equation has a known result for the analysis of viscoelastic bodies, and it can be expected to be superior to the usual differential equation model in the modeling of friction of contact parts including grease, etc.
[0116] In addition, in this embodiment, the state detection device 1 performs state detection of the semiconductor manufacturing device 3 using two methods, namely, state detection based on parameters of the inference model and state detection based on fitness of the inference model, but the present invention is not limited thereto. The state detection device 1 may perform state detection using any one of the methods, or may perform state detection using three or more methods in addition to the two methods.
[0117] In addition, in the present embodiment, the target device for which the state detection device 1 performs state detection is set to the semiconductor manufacturing device 3, but the present invention is not limited thereto, and the target device may be any device different from the semiconductor manufacturing device 3. The target device may be any device that can perform certain actions based on input data and output the results of the actions as observation data.
[0118] In addition, in the present embodiment, the inference model is set to a model of a non-integer order differential equation, but it is not limited thereto. The inference model may be, for example, a model of an integer order differential equation, or, for example, a model of machine learning, or a model represented by any other mathematical formula. For example, in the case where the action of the object device of the state detection is not so complicated, it is possible to obtain sufficient accuracy even if a model of an integer order differential equation is used instead of a model of a non-integer order differential equation.
[0119] In addition, in the present embodiment, the state detection device 1 is a structure that uses a state detection model that has been pre-machined to perform state detection based on parameters of the inference model, but is not limited to this. The state detection device 1 may, for example, determine the parameters of the inference model based on control data and observation data, and perform state detection of the semiconductor manufacturing device 3 by comparing the determined parameters with a predetermined threshold. The threshold to be compared with the parameter can be predetermined by, for example, a developer or manager of the information processing system. In addition, for example, the state detection device 1 may also store the history of the parameters of the determined inference model in advance, and detect the state of the semiconductor manufacturing device 3 based on the change in the value of the parameter in a certain period. In addition, for example, the state detection device 1 detects the state of the semiconductor manufacturing device 3 based on whether one or more parameters contain a parameter whose value deviates from a prescribed range (so-called outlier), or based on the number of times an outlier is generated in a certain period. The method of detecting the state of the semiconductor manufacturing device 3 based on the parameters of the inference model can adopt any method.
[0120] The embodiments disclosed this time should be considered as illustrative in all aspects and not restrictive. The scope of the present disclosure is indicated by the claims rather than the above meaning, and is intended to include all modifications within the meaning and scope equivalent to the claims.
[0121] The matters described in each embodiment can be combined with each other. In addition, the independent claims and dependent claims described in the claims can be arbitrarily combined in all kinds of combinations without considering the reference form. In addition, although the form of recording the reference to two or more other claims in the claims (multiple dependent claim form) is used, it is not limited to this. It is also possible to record in the form of recording multiple dependent claims (multiple references to multiple claims) that at least reference one multiple dependent claim.
[0122] Description of Reference Numerals
[0123] 1…state detection device (computer); 3…semiconductor manufacturing device (object device); 11…processing unit; 11a…control processing unit (first acquisition unit); 11b…parameter determination unit (determination unit); 11c…state detection unit (detection unit, second acquisition unit, calculation unit); 11d…machine learning processing unit; 12…storage unit; 12a…program (computer program); 12b…record storage unit; 12c…model information storage unit; 12d…learning data storage unit; 13…communication unit; 14…display unit; 15…operation unit; 99…recording medium.
Claims
1. A state detection device, comprising: a first acquisition unit that acquires input data to a target device and observation data of an action of the target device; a determination unit that determines parameters of an inference model that infers the observation data from the input data based on the input data and the observation data acquired by the first acquisition unit; and The detection unit detects the state of the target device based on the parameter determined by the determination unit.
2. The state detection device according to claim 1, wherein: A learning model is provided for performing machine learning by receiving as input a parameter of an inference model of observation data for inferring an action of the object device based on input data to the object device, and outputting a state of the object device, The detection unit detects the state of the target device by inputting the parameter determined by the determination unit into the learning model and acquiring the state of the target device output by the learning model.
3. The state detection device according to claim 2, wherein: The learning model is a learning model generated in advance by machine learning using learning data that associates parameters of an estimation model of observation data that estimates an action of the object device based on input data to the object device with a state of the object device.
4. The state detection device according to claim 1, wherein: have: a second acquisition unit that inputs the input data acquired by the first acquisition unit to the estimation model whose parameters are determined by the determination unit, and acquires estimation data of the observation data output by the estimation model; as well as a calculation unit that calculates the fitness of the inference model based on the observation data acquired by the first acquisition unit and the inference data acquired by the second acquisition unit, The detection unit detects the state of the target device based on the parameter determined by the determination unit and the suitability calculated by the calculation unit.
5. The state detection device according to claim 1, wherein: The inference model is a model of non-integer order differential equations.
6. The state detection device according to claim 1, wherein: The state of the target device detected by the detection unit includes the degree of abnormality related to the operation of the target device.
7. The state detection device according to claim 1, wherein: Information on the state of the target device detected by the detection unit is displayed on the display unit.
8. A state detection method, The information processing device acquires input data to the target device and observation data of the operation of the target device. Based on the acquired input data and observation data, the parameters of the inference model are determined. The inference model infers the observation data based on the input data. The state of the target device is detected based on the determined parameters.
9. A method for generating a learning model, The information processing device acquires learning data that associates parameters of an inference model of observation data for inferring an action of the object device based on input data to the object device with the state of the object device. A learning model is generated by machine learning using the acquired learning data. The learning model accepts as input parameters of an inference model of observation data for inferring the action of the object device based on input data to the object device, and outputs the state of the object device.
10. A computer program that causes a computer to execute the following processing: acquiring input data to a target device and observation data of an action of the target device; Determining parameters of an inference model based on the acquired input data and the observed data, the inference model inferring the observed data based on the input data; and The state of the target device is detected based on the determined parameters.
11. A computer program that causes a computer to execute the following processing: acquiring learning data that associates parameters of an estimation model of observation data for estimating an action of an object device based on input data to the object device with a state of the object device; and A learning model is generated by machine learning using the acquired learning data. The learning model accepts as input parameters of an inference model of observation data for inferring the action of the object device based on input data to the object device, and outputs the state of the object device.
Citation Information
Patent Citations
State determination device and method
JP2020128013A
