Condition inference system, condition inference method, and program

Abstract

A condition inference system comprising: an identified results acquisition unit that acquires identified results in which time-series trends in measurement values measured by each among a plurality of sensors provided to a target are identified with respect to each of the plurality of sensors; a domain knowledge information acquisition unit that acquires, from a storage unit, domain knowledge information in which sets with respect to the plurality of sensors of the identified results and types of conditions of the target are associated with each other; and a condition inference unit that infers the type of condition of the target via Bayesian inference on the basis of the sets with respect to the plurality of sensors of the identified results acquired by the identified results acquisition unit, and the domain knowledge information acquired by the domain knowledge information acquisition unit.

Description

State estimation system, state estimation method, and program

[0001] The present invention relates to a state estimation system, a state estimation method, and a program. This application claims priority under Japanese Patent Application No. 2025-018195, filed in Japan on February 6, 2025, the contents of which are incorporated herein by reference.

[0002] Exposure equipment is used to manufacture devices such as integrated circuits. Techniques for detecting abnormalities in exposure equipment are known (Patent Document 1).

[0003] U.S. Patent Application Publication No. 2008 / 0128644

[0004] One aspect of the present invention is a state estimation system comprising: an identification result acquisition unit that acquires identification results in which the time-series trend of measured values ​​measured by each of a plurality of sensors provided on an object is identified for each of the plurality of sensors; a domain knowledge information acquisition unit that acquires domain knowledge information from a storage unit in which the set of the identification results for the plurality of sensors and the type of state of the object are associated; and a state estimation unit that estimates the type of state of the object by performing Bayesian estimation based on the set of the identification results for the plurality of sensors acquired by the identification result acquisition unit and the domain knowledge information acquired by the domain knowledge information acquisition unit.

[0005] One aspect of the present invention is a state estimation method comprising: an identification result acquisition step of acquiring an identification result in which the time-series trend of measured values ​​measured by each of a plurality of sensors provided on the object is identified for each of the plurality of sensors; a domain knowledge information acquisition step of acquiring domain knowledge information from a storage step in which the set of the identification results for the plurality of sensors and the type of state of the object are associated; and a state estimation step of estimating the type of state of the object by performing Bayesian estimation based on the set of the identification results for the plurality of sensors acquired in the identification result acquisition step and the domain knowledge information acquired in the domain knowledge information acquisition step.

[0006] One aspect of the present invention is a program for causing a computer to execute: an identification result acquisition step in which the time-series trend of measured values ​​measured by each of a plurality of sensors provided on an object is identified for each of the plurality of sensors; a domain knowledge information acquisition step in which domain knowledge information is obtained from a storage step in which the pairs of the identification results for the plurality of sensors and the type of state of the object are associated; and a state estimation step in which the type of state of the object is estimated by performing Bayesian estimation based on the pairs of the identification results for the plurality of sensors obtained by the identification result acquisition step and the domain knowledge information obtained by the domain knowledge information acquisition step.

[0007] This figure shows an example of an overview of the state estimation system 1 according to the embodiment. This figure shows an example of an overview of the processing flow by the state estimation system 1 according to the embodiment. This figure shows an example of an overview of the learning of the classifier according to the embodiment. This figure shows an example of an overview of the learning of the classifier when there are a sufficient number of labels according to the embodiment. This figure shows an example of an overview of the learning of the classifier when there are only a few labels according to the embodiment. This figure shows an example of an overview of the method for creating a knowledge matrix according to the embodiment. This figure shows an example of a graph visualizing the value of the conditional probability calculated by calculating the average for J experts according to the embodiment. This is a conceptual diagram of the estimation of the probability of failure modes by Bayesian estimation according to the embodiment. This figure shows an example of an overview of the processing of estimating the probability of failure modes by Bayesian estimation in the state estimation system 1 according to the embodiment. This figure shows an example of the functional configuration of the state estimation system 1 according to the embodiment. This figure shows an example of the processing flow of failure mode estimation according to the embodiment. This figure shows an example of a case where the probability representing a state is specified as a component of the knowledge matrix according to a modified embodiment. This figure shows an example of a knowledge matrix according to the first embodiment. This figure shows an example of the state identification result of sensor log data according to the second embodiment. This figure shows another example of the state identification result of sensor log data according to the second embodiment. This figure shows another example of the state identification result of sensor log data according to the second embodiment. This figure shows an example of the state of sensor log data for each sensor for each failure according to the third embodiment. This figure shows the failure mode estimation results according to the third embodiment. This figure shows the sensor log data for each sensor according to the fourth embodiment. This figure shows the scores for each failure mode according to the fourth embodiment. This figure shows an example of the overview of the state estimation system 1a according to the modified example. This figure shows an example of the functional configuration of the state estimation system 1a according to the modified example. This figure shows an example of the failure mode hybrid estimation process flow according to the modified example.

[0008] (Embodiments) Hereinafter, embodiments will be described in detail with reference to the drawings. [Outline of State Estimation System 1] Figure 1 is a diagram showing an example of an overview of the state estimation system 1 according to this embodiment. The state estimation system 1 estimates the state of failure as the state of the exposure apparatus 2. Multiple sensors 3 are attached to predetermined parts of the exposure apparatus 2.

[0009] Conventional machine learning (AI) can identify the time-series state (e.g., upward trend, downward trend) of measurements from multiple sensors installed in a target device. However, conventional machine learning has difficulty identifying the type of failure (failure mode) occurring in the target device based on the time-series state of the sensor measurements. In other words, conventional machine learning could not identify the failure mode.

[0010] The state estimation system 1 estimates the failure mode of the exposure apparatus 2 based on a mathematical method that combines state identification of sensor log data with expert knowledge (knowledge matrix). Sensor log data is a time series of sensor measurements. The state (also referred to as class) of the sensor log data is identified by identifying the trend in the time series of sensor measurements. Expert knowledge is reflected in a pre-created knowledge matrix. The knowledge matrix is ​​a matrix that describes the relationship between the state of multiple sensor log data and the failure mode. The state estimation system 1 estimates the probability of the failure mode by fusing the state identification results for each of the multiple sensor log data with the knowledge matrix using Bayesian estimation.

[0011] The object in which the failure occurs is the exposure device 2. It is assumed that each of the multiple sensors 3 is functioning normally. There is a correlation between the type of failure (failure mode) occurring in the exposure device 2 and the trends in the sensor log data of each of the multiple sensors 3. Therefore, the failure mode is reflected in the trends in the sensor log data of each of the multiple sensors 3.

[0012] Furthermore, the multiple sensors 3 may be attached to predetermined locations on the exposure apparatus 2 such that there is a correlation between the measured values ​​obtained by each of the multiple sensors 3, corresponding to the malfunction state of the exposure apparatus 2. As another example, the multiple sensors 3 may be attached to predetermined locations on the exposure apparatus 2 such that the measured values ​​obtained by each of the multiple sensors 3 are independent of each other and correspond to the malfunction state of the exposure apparatus 2.

[0013] Figure 2 shows an example of the processing flow outlined by the state estimation system 1 according to this embodiment. The state estimation system 1 acquires sensor log data from each of the multiple sensors 3 provided in the exposure apparatus 2. There are two types of sensor log data: training sensor log data and identification sensor log data. Training sensor log data is used to create a classifier (i.e., training). Identification sensor log data is used to identify the state of the sensor log data.

[0014] A classifier is created for each of the multiple sensors 3. In other words, a classifier is created for each sensor. Therefore, the classifier is also referred to as a classifier for each sensor or a local classifier. It is created based on machine learning using training data, which consists of training sensor log data and class labels for each sensor. The classifier calculates the class probability, which is the probability (also referred to as likelihood) that the sensor log data belongs to a certain class. The class probability is a probability that indicates the trend of the sensor log data.

[0015] On the other hand, the knowledge matrix is ​​created based on the expertise of experts regarding the exposure apparatus 2. These experts are well-versed in the relationship between the failure modes of the exposure apparatus 2 and the trends in the sensor log data of each of the multiple sensors 3. For example, P knowledge matrices are created based on the expertise of J experts (J: a natural number). These P knowledge matrices are then subjected to statistical processing to create a probabilistically described knowledge matrix.

[0016] The state estimation system 1 estimates the probability (also referred to as likelihood) of a failure mode by fusing the state identification results for each of the multiple sensors with a knowledge matrix using Bayesian estimation.

[0017] [Classifier Training] This section describes the training of a classifier that calculates class probabilities from sensor log data. Figure 3 shows an example of the overview of classifier training according to this embodiment. Training sensor log data is acquired from each of the multiple sensors 3. Therefore, training sensor log data is acquired in quantities equal to the number of multiple sensors 3. Let the number of multiple sensors 3 be D (D: natural number). A total of N sets (N: natural number) of training sensor log data, each consisting of D training sensor log data, are acquired. The N sets of training sensor log data are acquired from the multiple sensors 3 at different times, for example.

[0018] For each of the D training sensor log data points contained within each of the N training sensor log data sets, a class label is assigned to indicate the state (class) of the sensor log data. The state of the sensor log data can be classified into seven states, as an example: up-tend, down-tend, step, steady, oscillation, asymptote, and random walk.

[0019] The state of the sensor log data can be classified according to the type of exposure device 2, the types of multiple sensors 3, and the number of failure modes, and is not limited to the seven states described above. Therefore, the number of states in the sensor log data may also be other than seven.

[0020] The classifier is created for each sensor based on machine learning, using training data consisting of training sensor log data and class labels for each sensor. Any machine learning model may be used.

[0021] Below, we will describe an example of classifier training for two cases: one where a sufficient amount of training data is available (a sufficient number of labels), and another where only a small amount of training data is available (a small number of labels).

[0022] [When there are a sufficient number of labels] Figure 4 shows an example of the overview of classifier training when there are a sufficient number of labels according to this embodiment. When a sufficient number of training data can be prepared, classifier parameters that satisfy the following equation (1) are trained.

[0023]

[0024] In equation (1), x represents the training sensor log data, and y Ｌ represents the class label. L is a function that represents the classifier, and the subscript φ of L represents the classifier's parameters.

[0025] Alternatively, features may be extracted from the training sensor log data x and input into the classifier. In this case, the value obtained by converting the training sensor log data x to F(x), etc., is input into the classifier instead of x in equation (1). F(x) is a function that represents the feature extractor that extracts features from the training sensor log data x. In Figure 4, the solid arrows labeled "(1)" schematically show the data flow when the training sensor log data x is directly input into the classifier. The dashed arrows labeled "(2)" schematically show the data flow when the value of F(x), which represents the feature extractor that extracts features from the training sensor log data x, is directly input into the classifier.

[0026] Any feature extractor may be used, as long as it extracts the time-series features of the sensor log data. The feature extractor may extract features from the training sensor log data x as numerical features or as image features. When the feature extractor extracts numerical features, these numerical features may be one or more of the following: frequency, autocorrelation, or slope. When the feature extractor extracts image features, these numerical features may be, for example, a recurrence plot.

[0027] When numerical feature quantities are input, the identifier is, for example, a Support-Vector Machine (SVM) or a random forest. When image feature quantities are input, the identifier is a deep neural network such as a Convolutional Neural Network (CNN). As described above, any machine learning model may be used for the identifier.

[0028] [When there are only a small number of labels] FIG. 5 is a diagram showing an example of an overview of learning of an identifier when there are only a small number of labels according to the present embodiment. When there are only a small number of labels, time-series feature extraction is performed using simulation data, and learning of the identifier is performed. An example when the identifier is a deep neural network will be described.

[0029] First, pre-training is performed. In pre-training, an autoencoder is learned using simulation data. As shown in the drawing, the autoencoder includes an encoder and a decoder. As the simulation data, for example, data simulating learning sensor log data generated by performing numerical simulation using a predetermined function is used. For example, learning sensor log data in a vibration state is generated by adding noise to a combination of trigonometric functions. The upward trend and the downward trend are each generated by adding noise to a linear function. A library of pre-generated simulation data may be used as the simulation data.

[0030] Next, the compressed feature quantity (schematically indicated by the letter z in FIG. 5) in the intermediate layer of the autoencoder for which pre-training has been completed is used as an input to the identifier. A class label is assigned to the compressed feature quantity. Using the set of the compressed feature quantity and the class label assigned to the feature quantity as teacher data, fine-tuning is performed on the fully-connected layer (FC layer) included in the deep neural network of the identifier.

[0031] In addition, when there are a sufficient number of labels for the compressed feature amounts in the intermediate layer of the autoencoder, feature amounts (schematically indicated by the character θ in FIG. 5) extracted by the feature extractor as described above may be added and fine-tuning may be performed.

[0032] [Creation of knowledge matrix] Next, a method for creating a knowledge matrix will be described. FIG. 6 is a diagram showing an example of an outline of a method for creating a knowledge matrix according to the present embodiment. A case where J experts create a knowledge matrix showing the relationship between D sensors and M types of failure modes will be described.

[0033] In the knowledge matrix, the rows are associated with the sensor numbers, and the columns are associated with the failure modes, and it is a D-row, M-column matrix. Therefore, the column of y Ｇ = m represents the states of the sensor log data of the D sensors when the failure mode is the m-th failure mode. In FIG. 6, y Ｇ ∈ {1,..., M} represents the failure mode, and χ ∈ {1,..., D} represents the sensor number. Each component of the knowledge matrix indicates the state of the sensor log data, such as {up-trend, down-trend, step, steady, oscillation, asymptote, random walk}, for example. Also, M （ｉ） is the knowledge matrix created by the i-th expert.

[0034] The knowledge matrices created by each of the J experts are aggregated. In this aggregation, as an example of statistical processing, the average of the J knowledge matrices is calculated. Here, the average of the J knowledge matrices is a matrix having, as components, the average values of the J knowledge matrices for each component of the J knowledge matrices. As a result of this aggregation, a conditional probability as shown in Expression (2) is obtained.

[0035]

[0036] In Expression (2), let the local class indicating the state of the sensor log data be y ｄ Ｌ ∈ {1,..., K} (d = 1,..., D), and assume that there are K states in the state of the sensor log data. M ｄ，ｍ （ｊ）represents the component of the d-th row and m-th column of the knowledge matrix of the j-th expert. The conditional probability shown in Equation (2) indicates the probability that the state of the sensor log data is the k-th state when the sensor number is the d-th and the failure mode is the m-th failure mode. The right side of Equation (2) shows that the average of each component of the knowledge matrix for J experts is calculated. Note that "C" is a normalization factor.

[0037] Among multiple experts, there may be differences in opinions regarding the state of sensor log data for each failure mode. Even if there are differences in opinions among multiple experts, the differences in opinions regarding the state of sensor log data for each failure mode are expressed as probabilities by calculating the average of each component of the knowledge matrix for multiple experts.

[0038] FIG. 7 shows an example of a graph visualizing the values of the conditional probabilities calculated by calculating the average for J experts as described above. In FIG. 7, with the failure mode as the global class, the probability of the local class for each global class is shown by the graph.

[0039] [Estimation of Failure Mode Probability by Bayesian Estimation] Next, the estimation of the failure mode probability by Bayesian estimation in the state estimation system 1 will be described. FIG. 8 is a conceptual diagram of the estimation of the failure mode probability by Bayesian estimation. The problem of estimating the failure mode probability by Bayesian estimation is to Ｌ (note that it is marked with a tilde in FIG. 8) and the local class y identified by the discriminator from the sensor log data ｄ Ｌ to be matched.

[0040] In the state estimation system 1, the posterior probability of the global class (failure mode) is calculated based on Equation (3).

[0041]

[0042] In Equation (3), X ＊ represents the identification data, and y ｍ Ｇ＊ , Y Ｌ＊ , Y (tilde)Ｌ＊ These represent the global class for the identification data, the local class derived from the sensor log data, and the local class derived from the knowledge matrix, respectively. Conditional probability P ＬＣ This represents the conditional probability attributable to the local classifier, and the conditional probability P ＫＭ This represents the conditional probability due to the knowledge matrix. Also, P(y ｍ Ｇ The ) represents the probability of a global class (failure mode) occurring. The delta function is calculated based on the knowledge matrix for the local class y Ｌ (Note that a tilde is added in Figure 8) and the local class y identified by the classifier from the sensor log data. ｄ Ｌ The condition is imposed that these two conditions are consistent. As shown in equation (3), the posterior probability of a global class (failure mode) is expressed using the product of the conditional probability attributable to the local classifier and the conditional probability attributable to the knowledge matrix.

[0043] Here, the posterior probability of the global class (failure mode) shown by equation (3) is rewritten according to the local class model. Examples of local class models include the AND model and the OR model, which are described below. The local class model may be either the AND model or the OR model. The AND model may be adopted if it is considered that the multiple sensors 3 are mounted on the exposure apparatus 2 such that the measured values ​​measured by each of the multiple sensors 3 are independent of each other.

[0044] The AND model is a model that treats local classes as independent of each other, as shown in equations (4) and (5).

[0045]

[0046]

[0047] Equation (4) is the conditional probability P attributable to the local classifier. ＬＣ In this, we show that the local classes are independent of each other. Equation (5) is the conditional probability P due to the knowledge matrix. ＫＭIn this, we show that the local classes are independent of each other. If the local classes are independent of each other, then from equations (4) and (5), we can obtain equation (6) shown below from equation (3). P(y ｍ Ｇ ) represents the prior probability of the global class and is a quantity given in advance. In the example, we assume that all global classes are uniform.

[0048]

[0049] Taking the sum over local classes in equation (6) yields equation (7).

[0050]

[0051] In the AND model, the state estimation system 1 uses the posterior probability of the global class (failure mode) shown by equation (7) as the failure mode estimation result.

[0052] On the other hand, the OR model is a model in which D-dimensional local classes are connected by OR. The conditional probability P attributable to the local classifiers. ＬＣ , and the conditional probability P due to the knowledge matrix ＫＭ Assuming that the local classes in each are connected by OR, equation (3) yields equation (8) shown below. Note that in equation (8), the overall normalization is taken into consideration as appropriate, and the left side is written as being proportional to the right side.

[0053]

[0054] Taking the sum over local classes in equation (8) yields equation (9).

[0055]

[0056] In the case of the OR model, the state estimation system 1 uses the posterior probability of the global class (failure mode) shown by equation (9) as the failure mode estimation result.

[0057] Figure 9 shows an overview of the process for estimating the probability of failure modes using Bayesian estimation in the state estimation system 1 described above. In the state estimation system 1, the state of each sensor log data is identified from the sensor log data of multiple sensors 3 by a local classifier. In other words, each local class is identified. In addition, the state estimation system 1 creates a knowledge matrix based on statistical processing from the knowledge of multiple experts. Based on this knowledge matrix, the probability of each local class for each global class is calculated.

[0058] The posterior probability of a global class (failure mode) is calculated by matching the local class, which is calculated based on the knowledge matrix, with the local class, which is identified by the classifier from the sensor log data, and then using the product of the conditional probability attributable to the local classifier and the conditional probability attributable to the knowledge matrix.

[0059] [Functional Configuration of State Estimation System 1] Figure 10 shows an example of the functional configuration of the state estimation system 1 according to this embodiment. The state estimation system 1 comprises a processing unit 10 and a storage unit 11.

[0060] The processing unit 10 includes a sensor information acquisition unit 100, an identification unit 101, an identification result acquisition unit 102, a domain knowledge information acquisition unit 103, a state estimation unit 104, and an output unit 105. Each of these functional units is realized, for example, by the CPU reading a program from ROM (Read Only Memory), expanding it into RAM (Random Access Memory), and executing processing according to that program. The ROM and RAM are included in the storage unit 11.

[0061] The sensor information acquisition unit 100 acquires detection target sensor information D1 from each of the multiple sensors 3. The detection target sensor information D1 shows the time series of measurements taken by sensor 3-i (i = 1, 2, ..., D: D is the number of multiple sensors 3). Here, one of the multiple sensors 3 is referred to as sensor 3-i.

[0062] The identification unit 101 identifies the time-series trend of the measured values ​​measured by each of the multiple sensors 3, indicated by the detection target sensor information D1 acquired by the sensor information acquisition unit 100. Based on multiple trained models A1, the identification unit 101 identifies the time-series trend of the measured values ​​indicated by the detection target sensor information D1 acquired by the sensor information acquisition unit 100, for each of the multiple sensors 3.

[0063] Multiple pre-trained models A1 consist of pre-trained models A1-i (i = 1, 2, ..., D: D is the number of sensors 3). Pre-trained model A1-i is a machine learning model that has learned the relationship between the time series of measured values ​​shown by the training sensor information L1 and the trend of that time series. As described above, if there are only a few labels, pre-trained model A1-i may be trained using features extracted based on the results of numerical simulations of the time series of measured values ​​measured by sensors 3-i (i = 1, 2, ..., D: D is the number of sensors 3). If there are a sufficient number of labels, the results calculated by numerical simulations are not necessary.

[0064] The identification result acquisition unit 102 acquires the identification result C1 from the identification unit 101. The identification result C1 is the result of identifying the time-series trend of the measured values ​​measured by each of the multiple sensors 3 provided in the exposure apparatus 2 for each of the multiple sensors 3.

[0065] The domain knowledge information acquisition unit 103 acquires multiple domain knowledge information B1 from the storage unit 11. The multiple domain knowledge information B1 consists of domain knowledge information B1-i (i = 1, 2, ..., J: J is the number of experts). Domain knowledge information B1-i is information that associates sets of multiple sensors 3 of the identification result C1 with the type of state (failure mode) of the exposure device 2.

[0066] Domain knowledge information B1-i is, for example, a knowledge matrix created based on the knowledge of the i-th expert described above. Therefore, domain knowledge information B1-i is created based on the expert's knowledge of the exposure apparatus 2. Note that domain knowledge information B1-i is not limited to a matrix format, as long as the sets of multiple sensors 3 in the identification result C1 are associated with the type of state (failure mode) of the exposure apparatus 2.

[0067] The state estimation unit 104 estimates the type of state (failure mode) of the exposure apparatus 2 by performing Bayesian estimation based on the set of identification results C1 for multiple sensors 3 acquired by the identification result acquisition unit 102 and the multiple domain knowledge information B1 acquired by the domain knowledge information acquisition unit 103.

[0068] The output unit 105 outputs the failure mode estimation result from the state estimation unit 104.

[0069] The memory unit 11 stores various types of information. For example, the memory unit 11 stores multiple trained models A1 and multiple domain knowledge information B1. The memory unit 11 is configured using a storage device such as a magnetic hard disk drive or a semiconductor storage device.

[0070] State estimation system 1 is, for example, a personal computer (PC). State estimation system 1 may be implemented as a single computer or as a virtual server. Each functional unit of state estimation system 1 may be distributed across multiple servers. In other words, state estimation system 1 may be configured across multiple devices. State estimation system 1 may also be implemented as a cloud server.

[0071] For example, the sensor information acquisition unit 100, the identification unit 101, and the output unit 105, which are functional units of the processing unit 10, may be provided in a separate device from the state estimation system 1. In other words, these sensor information acquisition unit 100, the identification unit 101, and the output unit 105 may be omitted from the configuration of the state estimation system 1. Also, the storage unit 11 may be provided as a separate storage device (such as a database server) from the state estimation system 1.

[0072] [Failure Mode Estimation Process] The flow of the failure mode estimation process, in which the state estimation system 1 estimates the failure mode of the exposure apparatus 2 based on Bayesian estimation, will be described. Figure 11 is a diagram showing an example of the flow of the failure mode estimation process according to this embodiment. As an example, the state estimation system 1 repeatedly executes the failure mode estimation process while the exposure apparatus 2 is operating. As another example, the state estimation system 1 may execute the failure mode estimation process when the exposure apparatus 2 is being maintained.

[0073] Step S10: The sensor information acquisition unit 100 acquires the detection target sensor information D1 from each of the multiple sensors 3.

[0074] Step S20: The identification unit 101 identifies the time-series trend of the measured values ​​measured by each of the multiple sensors 3, indicated by the detection target sensor information D1 acquired by the sensor information acquisition unit 100. Based on the trained model A1, the identification unit 101 identifies the time-series trend of the measured values ​​indicated by the detection target sensor information D1 acquired by the sensor information acquisition unit 100, for each of the multiple sensors 3.

[0075] The identification result acquisition unit 102 acquires the identification result C1 from the identification unit 101. The identification result acquisition unit 102 supplies the identification result C1 to the state estimation unit 104.

[0076] Step S30: The domain knowledge information acquisition unit 103 acquires multiple domain knowledge information B1 from the storage unit 11.

[0077] Step S40: The state estimation unit 104 estimates the failure mode by performing Bayesian estimation based on the set of identification results C1 for multiple sensors 3 acquired by the identification result acquisition unit 102 and the multiple domain knowledge information B1 acquired by the domain knowledge information acquisition unit 103.

[0078] The state estimation unit 104 estimates the failure mode based on equation (3) described above. Here, the state estimation unit 104 calculates the conditional probability P for the local class. ＬＣThis is calculated based on the identification result C1. The state estimation unit 104 calculates the conditional probability P caused by the knowledge matrix. ＫＭ This is calculated based on multiple domain knowledge information B1.

[0079] As described above, the state estimation unit 104 calculates statistical values ​​of multiple domain knowledge information B1 and uses them in Bayesian estimation. Therefore, the state estimation unit 104 uses in Bayesian estimation a conditional probability for each of the multiple sensors 3, which is a conditional probability based on the statistical values ​​of multiple domain knowledge information B1 acquired by the domain knowledge information acquisition unit 103.

[0080] Step S50: The output unit 105 outputs the failure mode estimation result from the state estimation unit 104. For example, the output unit 105 displays the estimation result on a display unit (not shown). The display unit is, for example, a display on a terminal device (personal computer, tablet, or smartphone, etc.) used by the maintenance personnel of the exposure apparatus 2. Alternatively, the output unit 105 may output the estimation result to an external device (server, etc.). In that case, the external device may notify the terminal device used by the maintenance personnel of the exposure apparatus 2 of the estimation result. With this, the state estimation system 1 completes the failure mode estimation process.

[0081] In this embodiment, an example has been described in which multiple domain knowledge information B1 is pre-stored in the storage unit 11, but the embodiment is not limited to this. Statistical values ​​of the multiple domain knowledge information B1 may be calculated in advance and stored in the storage unit 11. The statistical values ​​of the multiple domain knowledge information B1 are, for example, the average of J knowledge matrices.

[0082] In this embodiment, we have described an example in which the components of a knowledge matrix created based on the knowledge of multiple experts represent the time-series state of the measured values ​​of each sensor for a given failure mode. In other words, we have described an example in which the determined state of the time-series state of the measured values ​​of each sensor for a given failure mode is specified by an expert, but we are not limited to this. The components of the knowledge matrix may also specify probabilities that represent the time-series state of the measured values ​​of each sensor for a given failure mode.

[0083] Figure 12 shows an example where, instead of a fixed state, a probability representing a state is specified as an element of the knowledge matrix. For comparison, Figure 12 shows side by side the case where a fixed state is specified, as in the embodiment, and the case where a probability representing a state is specified. As shown in the figure, for example, for a failure mode of "regulating valve system failure", the state of the sensor log data of sensor 1 may be specified as a probability, such as "0.9" for a downward trend, "0.1" for asymptotic, and "0" for other classes, instead of a fixed state such as "downward trend or stuck".

[0084] In some cases, it is not possible to definitively state which state the time-series measurement values ​​of each sensor represent for a given failure mode; it can only be specified as a probability. Even in such cases, the knowledge matrix can reflect the expertise of specialists as probabilities.

[0085] Furthermore, while we have described an example where the mean is used as the statistical value of a knowledge matrix created based on the knowledge of multiple experts, this is not the only example. A weighted mean may also be used as the statistical value of the knowledge matrix. In this case, for example, the weight of the knowledge matrix created based on the knowledge of experts with particularly long experience may be set higher than the weight of the knowledge matrix created based on the knowledge of other experts with less experience.

[0086] Furthermore, the knowledge matrix may be created based on the knowledge of a single expert. In that case, based on the knowledge of that single expert, the probabilities representing the time-series state of the measured values ​​of each sensor for each failure mode may be specified as components of the knowledge matrix.

[0087] (First Embodiment) In this embodiment, a knowledge matrix was created first. Figure 13 shows an example of a knowledge matrix according to this embodiment. In the illustrated knowledge matrix, there are six failure modes, "regulating valve system failure", "pump system failure", "flow system failure 1", "flow system failure 2", "failure part 1", and "failure part 2", corresponding to each column. In the illustrated knowledge matrix, there are eight sensors numbered from "1" to "8", corresponding to each row. From the illustrated knowledge matrix, for example, if sensor "1" is "down or stuck", "3" is "up", and "8" is "down", then the failure mode is "regulating valve system failure".

[0088] (Second Embodiment) In this embodiment, the state of the sensor log data was first identified using a local classifier. Figure 14 shows an example of the identification result of the state of the sensor log data according to this embodiment. The identification result is that the state of the sensor log data was identified as either sensor skip or normal. In Figure 14(A), the sensor measurement value is shown against time. In Figure 14(B), the predicted probability of sensor skip is shown against time. The predicted probability of sensor skip is the probability identified using the local classifier for the sensor.

[0089] In Figures 14(A) and 14(B), the numerical values ​​on the horizontal axis represent the number of times a measurement was taken, and not the time itself. Furthermore, the identification result is the result of identifying the state based on the time series of sensor measurements (sensor log data) over a predetermined period of time. The same applies to Figures 15 and 16.

[0090] According to the measurements shown in Figure 14(A), sensor skips occurred during periods T1 and T2. The sensor log data was normal during periods other than T1 and T2. The predicted probability of sensor skips, shown in Figure 14(B), was high during the periods T1 and T2, respectively, when the sensor skips occurred. Therefore, it can be seen that the local classifier was able to identify sensor skips with high accuracy as a state of the sensor log data.

[0091] Figure 15 shows another example of the identification result of the state of sensor log data according to this embodiment. This identification result is the result of identifying the state of the sensor log data as normal or abnormal. An abnormal state is any state of the sensor log data other than normal. In Figure 15(A), the sensor measurement value is shown against time. In Figure 15(B), the label (true value) for the sensor log data is shown against time. This label is either "normal" or "abnormal". In Figure 15(C), the predicted probability of normal and abnormal states is shown against time.

[0092] According to the labels shown in Figure 15(B), the state of the sensor log data shown in Figure 15(A) is abnormal during period T11, excluding the beginning and end of the measurement. According to the predicted probabilities shown in Figure 15(C), in the first half of period T11, there is a mix of normal and abnormal predictions, with a higher probability of normal prediction. On the other hand, in the latter half of period T11, the probability of abnormal prediction is generally high, indicating high accuracy in estimating the state.

[0093] Figure 16 shows another example of the identification result of the state of sensor log data according to this embodiment. This identification result identifies the state of the sensor log data as normal, upward trend, and downward trend. In other words, in the identification result shown in Figure 16, the abnormal state shown in the identification result in Figure 15 is further divided into upward trend and downward trend. In Figure 16(A), the sensor measurement value is shown against time. In Figure 16(B), the label (true value) for the sensor log data is shown against time. This label is one of the values ​​of "normal", "upward trend", or "downward trend". In Figure 16(C), the predicted probability of normal, "upward trend", and "downward trend" is shown against time.

[0094] According to the labels shown in Figure 16(B), the state of the sensor log data shown in Figure 16(A) is either "upward trend" or "downward trend" during period T11, excluding the beginning and end of the measurement. According to the prediction probabilities shown in Figure 16(C), in the first half of period T11, predictions for normal, "upward trend," and "downward trend" are mixed, with the prediction probability for normal being high. On the other hand, in the latter half of period T11, the prediction probabilities for "upward trend" and "downward trend" are generally high, indicating high accuracy in estimating the state.

[0095] (Third Embodiment) In this embodiment, the failure mode of the exposure apparatus 2 was estimated by the state estimation system 1. Figure 17 shows the state of the sensor log data of each sensor for each failure according to this embodiment. Three failures that actually occurred in the exposure apparatus 2 were used as the target of estimation. As shown in the figure, the failure modes of these three failures were "regulating valve system failure", "pump system failure", and "flow rate system failure 2". In the "wait and see" case, the failure mode was not specified. At the time of the occurrence of these three failures, the state of the sensor log data of each sensor was confirmed as shown in the figure. The number of sensors is 11.

[0096] Figure 18 shows the failure mode estimation results according to this embodiment. Figures 18(A) to 18(E) each show the estimated failure mode scores for the failures shown in Figure 17. These scores indicate the probability of the estimated failure mode. In Figures 18(A), 18(B), 18(C), and 18(D), the score for the correct failure mode is the highest, indicating that the failure mode was correctly estimated. On the other hand, in Figure 18(E), the score for the incorrect failure mode is the highest, indicating that the failure mode was not correctly estimated. Based on the estimation results shown in Figure 18, it can be seen that the failure mode was estimated generally correctly.

[0097] (Fourth Embodiment) In this embodiment, the state estimation system 1 detected the failure when the failure mode was "regulating valve system failure". Figure 19 is a diagram showing the sensor log data of each sensor according to this embodiment. In Figure 19, the sensor log data for five of the sensors attached to the exposure apparatus 2 is shown. According to Figure 19(E), the probability of the failure mode being "regulating valve system failure" is high during the period when the sensor log data of sensor 9 shows a downward trend. The axis on the right side of Figure 19(E) shows the probability of the failure mode being "regulating valve system failure".

[0098] Figure 20 shows the scores for each failure mode in this embodiment. As shown in the figure, the score for the "regulating valve system failure" failure mode is estimated to be the highest. This indicates that the state estimation system 1 correctly estimated the "regulating valve system failure" failure mode. Based on the estimation results, the state estimation system 1 detected the "regulating valve system failure" failure mode and notified the user.

[0099] In this embodiment, an example of a case in which the failure state (failure mode) of the exposure apparatus 2 is estimated has been described, but it is not limited to this. The object whose state is estimated is not limited to the exposure apparatus 2. Also, the state of the object is not limited to a failure state. Domain knowledge information may be created based on the knowledge of an expert on the object, depending on the object and the type of state to be estimated.

[0100] The state estimation system 1 may estimate the state of various devices or machines other than the exposure apparatus 2, various vehicles such as automobiles, aircraft, or ships, or factories (plants) made up of multiple combinations of equipment and machinery, such as malfunctions. These objects are typically fitted with various sensors to monitor their state. For example, in automobiles, sensors are installed in a manner that links them together as modules.

[0101] As another example, the state estimation system 1 may estimate the state of a living organism to which a biosensor is attached (such as a state of health or condition). In other words, the state estimation system 1 may be used for biological monitoring.

[0102] As described above, the state estimation system 1 according to this embodiment comprises an identification result acquisition unit 102, a domain knowledge information acquisition unit 103, and a state estimation unit 104. The identification result acquisition unit 102 acquires an identification result C1 in which the time-series trend of measured values ​​measured by each of the multiple sensors 3 provided on the target (in this embodiment, the exposure apparatus 2) is identified for each of the multiple sensors 3. The domain knowledge information acquisition unit 103 acquires domain knowledge information (in this embodiment, multiple domain knowledge information B1) from the storage unit 11, which associates the sets of identification results C1 for the multiple sensors 3 with the type of state of the target (in this embodiment, the exposure apparatus 2) (in this embodiment, the failure mode). The state estimation unit 104 estimates the type of state of the target (in this embodiment, the exposure apparatus 2) (in this embodiment, the failure mode) by performing Bayesian estimation based on the sets of identification results C1 for the multiple sensors 3 acquired by the identification result acquisition unit 102 and the domain knowledge information (in this embodiment, multiple domain knowledge information B1) acquired by the domain knowledge information acquisition unit 103.

[0103] With this configuration, the state estimation system 1 according to this embodiment can perform Bayesian estimation based on the set of multiple sensors 3 of the identification result C1 and domain knowledge information, thereby improving the accuracy of estimating the type of state of the target.

[0104] (Modified Versions) Modified versions of this embodiment will be described in detail below with reference to the drawings. In this modified version, we will describe a case where the state estimation system estimates the failure mode of the exposure apparatus 2 not only based on the results of Bayesian estimation, but also based on a Large Language Model (LLM). The state estimation system according to this modified version will be referred to as state estimation system 1a. Note that the same reference numerals will be used for components identical to those in the above-described embodiment, and descriptions of identical components and operations may be omitted.

[0105] Figure 21 shows an example of an overview of the state estimation system 1a according to this modified example. In the state estimation system 1a, the failure mode of the exposure apparatus 2 is estimated based on Bayesian estimation using the state identification results of the sensor log data and domain knowledge information, similar to the state estimation system 1. However, in the state estimation system 1a, the sets of state identification results for multiple sensors 3 and the domain knowledge information (multiple domain knowledge information B1) are each represented by graphs. Furthermore, in the state estimation system 1a, the failure mode of the exposure apparatus 2 is estimated not only based on the results of Bayesian estimation but also based on LLM.

[0106] A graph that represents the identification result of the state of sensor log data is referred to as the identification result graph. The identification result graph is generated based on the identification result of the state of sensor log data. A graph that represents domain knowledge information is referred to as the domain knowledge graph. As an example, the domain knowledge graph is generated based on a pre-created knowledge matrix. The domain knowledge graph is generated as a Bayesian network.

[0107] The identification result graph is a directed acyclic graph having multiple nodes (first nodes) associated with each of the multiple sensors 3, multiple nodes (second nodes) associated with each of the local classes (sensor log data states), and directed edges. An identification result graph is generated for each global class (failure mode). Edges connect the first and second nodes in the direction from the first node to the second node, based on the identification result. If one node included in the first node is connected to two or more nodes included in the second node, the class probability of the local class based on the identification result is set as the weight of the edge. In Figure 21, for convenience, the thickness of the edge indicates the class probability.

[0108] The domain knowledge graph is a directed acyclic graph having multiple nodes (third nodes) associated with each of the multiple sensors 3, multiple nodes (fourth nodes) associated with each of the local classes (states of sensor log data), and directed edges. The domain knowledge graph is generated for each global class (failure mode). Edges connect the third node and the fourth node in the direction from the third node to the fourth node, based on expert knowledge (knowledge matrix). If the knowledge matrix is ​​obtained by statistically processing the knowledge of multiple experts and includes the class probability of the local class, then one node included in the third node and two or more nodes included in the fourth node may be connected, and the class probability of the local class based on the knowledge matrix is ​​set as the weight of the edge.

[0109] By representing and structuring the data using a directed acyclic graph, the state identification results of sensor log data and domain knowledge information can be expressed in a format suitable for reading by LLM. Furthermore, since directed acyclic graphs can be described as Bayesian networks, they are suitable representations for use in Bayesian estimation.

[0110] Furthermore, the state estimation system 1a performs Bayesian estimation based on the identification result graph and the domain knowledge graph. In addition, the state estimation system 1a inputs the results of the Bayesian estimation and the identification result graph into the LLM, which then estimates the failure mode of the exposure apparatus 2. The LLM outputs the estimated failure mode along with the basis for the estimation as text. The state estimation system 1a can add qualitative explanations from the LLM to the estimation results, along with quantitative scoring using Bayesian estimation. Furthermore, the natural language interface of the LLM can improve convenience for the user. While the LLM alone may cause hallucination and stochastic output fluctuations, in the state estimation system 1a, the results of the Bayesian estimation are also input into the LLM, creating a complementary relationship between the LLM and Bayesian estimation.

[0111] A directed acyclic graph can be represented, for example, in JSON (JavaScript Object Notation) format. However, directed acyclic graphs may be represented in formats other than JSON, as long as they are readable by the LLM (Language-Learning Module).

[0112] [Functional Configuration of State Estimation System 1a] Figure 22 shows an example of the functional configuration of the state estimation system 1a according to this modified example. The state estimation system 1a comprises a processing unit 10a and a storage unit 11.

[0113] The processing unit 10a includes a sensor information acquisition unit 100, an identification unit 101, an identification result acquisition unit 102a, a domain knowledge information acquisition unit 103a, a state estimation unit 104a, an LLM 12a, an output unit 105a, and an input unit 106a. Comparing the state estimation system 1a according to this modified example (Figure 22) with the state estimation system 1 according to the embodiment (Figure 10), the identification result acquisition unit 102a, the domain knowledge information acquisition unit 103a, the state estimation unit 104a, the LLM 12a, the output unit 105a, and the input unit 106a are different. Here, the functions of the other components (sensor information acquisition unit 100 and identification unit 101) are the same as in the first embodiment. The description of functions that are the same as in the embodiment will be omitted, and the modified example will focus on the parts that differ from the embodiment.

[0114] When the identification result acquisition unit 102a acquires the identification result C1 from the identification unit 101, it generates an identification result graph C1a from the identification result C1.

[0115] The domain knowledge information acquisition unit 103a acquires the domain knowledge graph B1a from the storage unit 11.

[0116] The state estimation unit 104a inputs the result of Bayesian estimation of the state type (failure mode) of the exposure apparatus 2 and the identification result graph C1a generated by the identification result acquisition unit 102a to the LLM 12a. Although this example describes the state estimation unit 104a inputting the identification result graph C1a to the LLM 12a, the identification result graph C1a may also be directly input to the LLM 12a from the identification result acquisition unit 102a.

[0117] LLM12a receives instructions from the input unit 106a and generates text indicating the estimated state type (failure mode) of the exposure apparatus 2 and the basis for the estimation, based on the domain knowledge graph B1a. This text is also referred to as the estimation basis text. LLM12a is a model such as GPT (Generative Pre-trained Transformer), Gemini®, BERT (Bidirectional Encoder Representations from Transformers)®, Claude®, or Llama®. LLM12a has a Retrievable-Augmented Generation (RAG) function. The LLM 12a uses RAG to retrieve information necessary for estimating the failure mode from the domain knowledge graph B1a, and generates the failure mode estimation result and estimation basis text based on the retrieved information. The LLM 12a may be installed on a separate server from the state estimation system 1a. In that case, the LLM 12a installed on the separate server will use RAG to retrieve the domain knowledge graph B1a stored in the storage unit 11. Figure 22 shows an example in which the LLM 12a retrieves the domain knowledge graph B1a stored in the storage unit 11, but the LLM 12a may also retrieve the domain knowledge graph B1a acquired by the domain knowledge information acquisition unit 103a.

[0118] The input unit 106a receives a request from the user and generates a predetermined prompt for input to the LLM 12a. This prompt is text that instructs the system to estimate the failure mode from the Bayesian estimation result of the failure mode and the identification result graph C1a, and to generate text that shows the basis for the estimation.

[0119] The prompt is preferably generated based on prompt engineering. For example, the prompt may include text such as, "You are a world-class AI expert specializing in fault diagnosis of semiconductor manufacturing equipment," "Your job is to provide a comprehensive diagnosis by integrating quantitative scores and qualitative reasoning," and "Your main conclusions should be consistent with the results of Bayesian estimation." Therefore, the prompt may include instructions to align the estimation results from LLM12a with the results of Bayesian estimation.

[0120] The output unit 105a outputs the failure mode estimation result and estimation basis text that shows the basis for the estimation.

[0121] The memory unit 11 stores, for example, multiple trained models A1 and a domain knowledge graph B1a.

[0122] Domain knowledge graph B1a is, for example, created in advance based on multiple domain knowledge information B1 (knowledge matrices). Alternatively, domain knowledge graph B1a may be created directly from expert knowledge without going through knowledge matrices.

[0123] [Failure Mode Hybrid Estimation Process] The flow of the failure mode hybrid estimation process, in which the state estimation system 1a estimates the failure mode of the exposure apparatus 2 based on Bayesian estimation and LLM, will be explained. Figure 23 is a diagram showing an example of the flow of the failure mode hybrid estimation process according to this modified example. Note that the processes in steps S110 and S120 are the same as the processes in steps S10 and S20 in Figure 11, so the explanation will be omitted.

[0124] However, in the process of step S120, when the identification result acquisition unit 102a acquires the identification result C1 by the identification unit 101, it generates an identification result graph C1a from the identification result C1. The identification result acquisition unit 102a supplies the identification result graph C1a to the state estimation unit 104a.

[0125] Step S130: The domain knowledge information acquisition unit 103a acquires the domain knowledge graph B1a from the storage unit 11.

[0126] Step S140: The state estimation unit 104a estimates the failure mode by performing Bayesian estimation based on the identification result graph C1a generated by the identification result acquisition unit 102a and the domain knowledge graph B1a acquired by the domain knowledge information acquisition unit 103a.

[0127] Step S150: The state estimation unit 104a inputs the result of Bayesian estimation of the state type (failure mode) of the exposure apparatus 2 and the identification result graph C1a generated by the identification result acquisition unit 102a to the LLM 12a. The LLM 12a generates the estimation result of the state type (failure mode) of the exposure apparatus 2 and the text of the basis for the estimation based on the domain knowledge graph B1a.

[0128] Step S160: The output unit 105a outputs the failure mode estimation result and estimation basis text showing the basis for the estimation. For example, the output unit 105 displays the failure mode estimation result and estimation basis text on a display unit (not shown). With this, the state estimation system 1a completes the failure mode hybrid estimation process.

[0129] The input unit 106a may also cause the LLM 12a to generate a score indicating the reliability of the estimation, along with the estimation result and the text supporting the estimation.

[0130] The input unit 106a may output the failure mode estimation result by Bayesian estimation along with the failure mode estimation result by LLM 12a to the output unit 105a. As described above, since the result of Bayesian estimation of the failure mode is input to LLM 12a, the failure mode estimation result by LLM 12a tends to match the failure mode estimation result by Bayesian estimation. Even if the failure mode estimation result differs between LLM 12a and Bayesian estimation, both the LLM 12a estimation result and the Bayesian estimation result may be presented to the user. The user can refer to both estimation results.

[0131] The input unit 106a may, instead of having the LLM 12a generate the failure mode estimation result, generate text indicating the basis for the Bayesian estimation of the failure mode. In that case, the state estimation unit 104a outputs the Bayesian estimation of the failure mode and the text indicating the basis for the estimation by the LLM 12a to the output unit 105a.

[0132] In this modified example, an example has been described in which the identification result of the state of the sensor log data and the domain knowledge information are represented by directed acyclic graphs (identification result graph and domain knowledge graph), respectively, but the invention is not limited to this example. At least one of the identification result of the state of the sensor log data and the domain knowledge information does not have to be represented by a directed acyclic graph. In that case, Bayesian estimation is performed using the identification result of the state of the sensor log data and the domain knowledge matrix, as in the embodiment, and the identification result graph and the domain knowledge graph may be used only for processing by LLM12a. Furthermore, the identification result graph does not have to be used as input to LLM12a, and the identification result of the state of the sensor log data may be represented in a format other than a graph and used as input. Also, the domain knowledge that LLM12a searches for using RAG does not have to use a domain knowledge graph, and the domain knowledge may be represented in a format other than a graph and used.

[0133] Therefore, in the state estimation system 1a, the Bayesian estimation result of the state type (failure mode) of the exposure apparatus 2 by the state estimation unit 104a and the set of identification results C1 for multiple sensors 3 acquired by the identification result acquisition unit 102a are input to the LLM 12a, and the domain knowledge information acquisition unit 103a acquires domain knowledge information (multiple domain knowledge information B1) to cause the LLM 12a to generate text that shows the basis for the estimation of the state type (failure mode) of the exposure apparatus 2. A graph does not need to be used in this process.

[0134] The state estimation system 1a may store new domain knowledge information in the storage unit 11 through interaction between the LLM 12a and the user. The new domain knowledge information may be stored in association with the domain knowledge graph B1a, or it may be stored individually without being associated with the domain knowledge graph B1a. Furthermore, the new domain knowledge information does not have to be stored as a graph, but may be stored as a knowledge matrix.

[0135] Furthermore, a program to realize the function of any component in any device described above may be recorded on a computer-readable recording medium, and that program may be loaded into a computer system and executed. Here, "computer system" includes hardware such as an operating system or peripheral devices. "Computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, CDs (Compact Disc)-ROMs (Read Only Memory), and storage devices such as hard disks built into a computer system. Moreover, "computer-readable recording medium" also includes volatile memory within a computer system that acts as a server or client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line, which retains the program for a certain period of time. Such volatile memory may be, for example, RAM (Random Access Memory). The recording medium may be, for example, a non-temporary recording medium.

[0136] Furthermore, the above program may be transmitted from a computer system that stores this program in a memory device or the like to another computer system via a transmission medium or by transmission waves within the transmission medium. Here, the "transmission medium" for transmitting the program refers to a medium that has the function of transmitting information, such as a network like the Internet or a communication line like a telephone line. Also, the above program may be for the purpose of realizing a part of the functions described above. Furthermore, the above program may be one that can realize the above functions in combination with a program already recorded in the computer system, a so-called differential file. A differential file may also be called a differential program.

[0137] Furthermore, the functions of any component in any device described above may be implemented by a processor. For example, each process in the embodiment may be implemented by a processor that operates based on information such as a program, and a computer-readable recording medium that stores information such as a program. Here, the processor may be implemented by implementing the functions of each part in separate hardware, or by implementing the functions of each part in integrated hardware. For example, the processor includes hardware, and the hardware may include at least one of a circuit that processes digital signals and a circuit that processes analog signals. For example, the processor may be configured using one or more circuit devices or one or both of one or more circuit elements mounted on a circuit board. An IC (Integrated Circuit) may be used as the circuit device, and a resistor or capacitor may be used as the circuit element.

[0138] Here, the processor may be, for example, a CPU. However, the processor is not limited to a CPU, and various types of processors may be used, such as a GPU (Graphics Processing Unit) or a DSP (Digital Signal Processor). The processor may also be, for example, a hardware circuit using an ASIC (Application Specific Integrated Circuit). Furthermore, the processor may be composed of, for example, multiple CPUs, or of hardware circuits using multiple ASICs. The processor may also be composed of, for example, a combination of multiple CPUs and hardware circuits using multiple ASICs. Furthermore, the processor may include, for example, one or more amplifier circuits or filter circuits that process analog signals.

[0139] While embodiments of this disclosure have been described in detail above with reference to the drawings, the specific configuration is not limited to these embodiments and may include designs and other elements that do not depart from the gist of this disclosure.

[0140] 1, 1a... State estimation system, 102, 102a... Identification result acquisition unit, 103, 103a... Domain knowledge information acquisition unit, 104, 104a... State estimation unit, 11... Storage unit, 3... Multiple sensors, B1... Multiple domain knowledge information, C1... Identification result

Claims

An identification result acquisition unit acquires an identification result that identifies the time-series trend of the measured values ​​measured by each of the multiple sensors installed on the target, A domain knowledge information acquisition unit acquires domain knowledge information from a storage unit, which associates the set of identification results for the plurality of sensors with the type of state of the target. A state estimation unit estimates the type of state of the target by performing Bayesian estimation based on the set of identification results for the plurality of sensors acquired by the identification result acquisition unit and the domain knowledge information acquired by the domain knowledge information acquisition unit. A state estimation system equipped with the following features. 　 A sensor information acquisition unit that acquires sensor information from each of the plurality of sensors, which shows the time series of the measured values ​​measured by the sensor, An identification unit that identifies the time-series trend indicated by the sensor information acquired by the sensor information acquisition unit for each of the plurality of sensors, Furthermore, The identification result acquisition unit acquires the identification result from the identification unit. The state estimation system according to claim 1. 　 The identification unit identifies the trend of the time series indicated by the sensor information acquired by the sensor information acquisition unit for each of the multiple sensors, based on a trained model in which the relationship between the time series and the trend of the time series has been learned. The state estimation system according to claim 2. 　 The trained model is trained using features extracted based on the results of numerical simulations of the time series of measurements taken by the sensor. The state estimation system according to claim 3. 　 The domain knowledge information acquisition unit acquires a plurality of domain knowledge information from the storage unit, The state estimation unit uses in the Bayesian estimation a conditional probability for the identification result for each of the multiple sensors, which is a conditional probability based on the statistical values ​​of the multiple domain knowledge information acquired by the domain knowledge information acquisition unit. The state estimation system according to claim 1. 　 The Bayesian estimation result of the state estimation unit for the type of state of the target, and the set of identification results for the multiple sensors acquired by the identification result acquisition unit are input to the large-scale language model, and the large-scale language model is made to generate text that shows the basis for the estimation of the type of state of the target based on the domain knowledge information. The state estimation system according to claim 1. 　 The set of identification results for the plurality of sensors is represented by an identification result graph having nodes associated with each of the plurality of sensors and nodes associated with each of the time series trends. The domain knowledge information is represented by a domain knowledge graph having nodes associated with each of the multiple sensors and nodes associated with each of the time-series trends. The results of the Bayesian estimation of the type of state of the target by the state estimation unit and the identification result graph are input to the large-scale language model, and the large-scale language model is made to generate text that shows the basis for the estimation of the type of state of the target based on the domain knowledge graph. The state estimation system according to claim 6. 　 The aforementioned domain knowledge information is created based on expert knowledge about the subject. The state estimation system according to claim 1. 　 The plurality of sensors are attached to a predetermined part of the target, There is a correlation between the measured values ​​obtained by each of the aforementioned multiple sensors, depending on the state of the object. The state estimation system according to claim 1. 　 The state of the object mentioned above is the state of failure of the object mentioned above. The state estimation system according to claim 1. 　 An identification result acquisition step is to acquire an identification result in which the time-series trend of the measured values ​​measured by each of the multiple sensors installed on the target is identified for each of the multiple sensors, A domain knowledge information acquisition step involves acquiring domain knowledge information from a storage step, which associates the set of identification results for the plurality of sensors with the type of state of the target. A state estimation step which estimates the type of state of the target by performing Bayesian estimation based on the set of identification results for the plurality of sensors obtained in the identification result acquisition step and the domain knowledge information obtained in the domain knowledge information acquisition step, A method for estimating a state. 　 On the computer, An identification result acquisition step is to acquire an identification result in which the time-series trend of the measured values ​​measured by each of the multiple sensors installed on the target is identified for each of the multiple sensors, A domain knowledge information acquisition step involves acquiring domain knowledge information from a storage step, which associates the set of identification results for the plurality of sensors with the type of state of the target. A state estimation step which estimates the type of state of the target by performing Bayesian estimation based on the set of identification results for the plurality of sensors obtained in the identification result acquisition step and the domain knowledge information obtained in the domain knowledge information acquisition step, A program to execute.

Images