Learning device, industrial system, inference device, learning method, and inference method

Abstract

A learning device (100) comprises: an adaptive model training unit (106) for generating an adaptive model trained with the relationship between a first target label and target data measured in a target device (103); an inference error acquisition unit (107) for calculating a base model inference error, which is between the first target label and a second target label inferred by using a base model trained with the relationship between a reserve label and reserve data measured in a reserve device (101), and an adaptive model inference error, which is between the first target label and a third target label inferred by using the adaptive model; and an error model generation unit (108) for generating an error model to be used for integrating a base model inference result and an adaptive model inference result obtained by inputting new target data respectively into the base model and the adaptive model, through training based on the target data, the magnitude of the base model inference error, and the magnitude of the adaptive model inference error. The error model is a model which, upon input of the new target data, outputs a weight used at the time of the integration.

Description

Learning device, industrial system, inference device, learning method, and inference method

[0001] The present disclosure relates to a learning device, an industrial system, an inference device, a learning method, and an inference method.

[0002] In recent years, machine learning technology has been used in industrial systems for condition monitoring or anomaly detection. The data that a learning model learns (machine learning) is signals (sensor data) acquired by sensors attached to industrial equipment in the industrial system, and inferences are made based on the sensor data acquired by the sensors during operation of the industrial system.

[0003] When generating a learning model (machine learning model), a learning device must perform machine learning using a large amount of training data, which requires a lot of time and effort. For this reason, a technology has been considered in which a learning device generates a learning model by performing machine learning using training data collected in advance by operating other industrial systems, and the generated learning model is used in multiple different industrial systems of the same type. This technology eliminates the need to prepare training data for each industrial system and perform machine learning for each industrial system to generate a learning model, thereby reducing the time and effort required to prepare a learning model.

[0004] Patent Literature 1 discloses a method for constructing an inference model capable of appropriately executing a predetermined inference in a new environment. The inference device in Patent Literature 1 provides target data to each of multiple inference models derived from local learning data obtained in different environments to execute the predetermined inference, and causes each inference model to execute the predetermined inference, thereby obtaining an inference result for each inference model. The inference device then determines values ​​of each integration parameter according to the environmental data, weights the inference results of each inference model using the determined values ​​of each integration parameter, and integrates the weighted inference results to generate an inference result for the target environment.

[0005] Japanese Patent Application Laid-Open No. 2021-22079

[0006] However, the technology of Patent Document 1 does not acquire data in the new target environment, and therefore has the problem of being unable to generate appropriate inference results with high inference accuracy for a new environment that differs from the environment in which data was acquired in advance.

[0007] The present disclosure has been made in consideration of the above, and aims to obtain a learning device that can generate a learning model that can generate appropriate inference results with high inference accuracy even in a new environment that is different from the environment in which data has been acquired in advance.

[0008] To solve the above-mentioned problems and achieve the object, the learning device of the present disclosure includes an adaptive model learning unit that generates an adapted model that learns the relationship between target data, which is data measured by a target device that is the target of learning and inference when the target device is operating, and a first target label, which is the label of the target data.The learning device of the present disclosure also includes an inference error acquisition unit that infers a second target label corresponding to the target data using a base model that learns the relationship between preliminary data, which is data measured by a spare device that is the target of learning and is a different device from the target device when the spare device is operating, and a spare label, which is the label of the preliminary data, and calculates an error between the first target label and the second target label as a base model inference error, and infers a third target label corresponding to the target data using the adaptive model and calculates an error between the first target label and the third target label as an adaptive model inference error. The learning device of the present disclosure also includes an error model generation unit that generates an error model used to integrate a base model inference result, which is the output when new target data, which is data newly measured by the target device when the target device is newly operated, is input to the base model, and an adaptive model inference result, which is the output when the new target data is input to the adaptive model, by learning based on the target data, the base model inference error, and the adaptive model inference error. The error model learns the relationship between the magnitude of the base model inference error, the magnitude of the adaptive model inference error, and the target data, and when new target data is input, outputs an error model inference result, which is data corresponding to the weighting of the base model inference result and the adaptive model inference result at the time of integration.

[0009] The learning device according to the present disclosure has the advantage of being able to generate a learning model that can generate appropriate inference results with high inference accuracy even in a new environment that differs from the environment in which data was previously acquired.

[0010] FIG. 1 is a diagram showing a configuration example of a learning device according to the first embodiment. FIG. 2 is a diagram showing a configuration example of a spare sensor unit according to the first embodiment. FIG. 3 is a diagram showing a configuration example of a target sensor unit according to the first embodiment. FIG. 4 is a diagram for explaining the operation of a base model learning unit according to the first embodiment. FIG. 5 is a diagram for explaining the operation of an adaptive model learning unit according to the first embodiment. FIG. 6 is a diagram showing a configuration example of an inference error acquisition unit according to the first embodiment. FIG. 7 is a diagram for explaining the operation of an error model generation unit according to the first embodiment. FIG. 8 is a diagram showing a configuration example of an inference device according to the first embodiment. FIG. 9 is a diagram for explaining the operation of an error model generation unit and a dimension reduction unit according to the first embodiment. FIG. 1 shows a configuration example of an error model generation unit according to a second embodiment. FIG. 2 shows a configuration example of an error model generation unit according to a third embodiment. FIG. 3 shows a configuration example of an inference device according to a third embodiment. FIG. 4 shows a configuration example of an error model generation unit according to a fourth embodiment. FIG. 5 shows a configuration example of an inference device according to a fourth embodiment. FIG. 6 shows a configuration example of an adaptive region model inference unit and a region threshold determination unit provided in the error model generation unit according to the fourth embodiment. FIG. 7 shows a configuration of a servo control device which is a first example of an industrial device according to a fifth embodiment. FIG. 8 shows a configuration of a laser cutting device which is a second example of an industrial device according to the fifth embodiment. FIG. 9 shows a configuration of a robot device which is a third example of an industrial device according to the fifth embodiment. FIG. 10 shows a configuration example of a processing circuit provided in the inference device according to the first embodiment when implemented by a processor and a memory. FIG. 11 shows a configuration example of a processing circuit provided in the inference device according to the first embodiment when configured by dedicated hardware.

[0011] A learning device, an industrial system, an inference device, a learning method, and an inference method according to embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0012] First Embodiment Fig. 1 is a diagram illustrating an example of the configuration of a learning device 100 according to a first embodiment. The learning device 100 includes a base model learning unit 105, an adaptive model learning unit 106, an inference error acquisition unit 107, and an error model generation unit 108. Fig. 1 illustrates a backup device 101 that is the target of learning, a target device 103 that is the target of learning and inference, a backup sensor unit 102, and a target sensor unit 104. The backup sensor unit 102 may be disposed inside the backup device 101. The target sensor unit 104 may be disposed inside the target device 103.

[0013] The backup device 101 and the target device 103 are different devices (industrial devices) of the same type, and the backup sensor unit 102 and the target sensor unit 104 are configured to include sensors of the same type. In the first embodiment, "same type" means having similar functions. Therefore, in the first embodiment, "devices of the same type" refers to devices having similar functions, and "sensors of the same type" refers to sensors that can measure a certain physical quantity with the same accuracy.

[0014] The standby device 101 is an industrial device that acquires standby data (later-described standby data 203) used for learning by a base model learning unit (backup model learning unit) 105. The target device 103 is an industrial device that acquires target data (later-described target data 303) used for learning by an adaptive model learning unit (target model learning unit) 106. The standby device 101 and the target device 103 are, for example, processing devices that process a processing target (workpiece).

[0015] The preliminary data 203 is data measured by the preliminary sensor unit 102 in the preliminary device 101 when the preliminary device 101 is operating. The target data 303 is data measured by the target sensor unit 104 in the target device 103 when the target device 103 is operating.

[0016] In the first embodiment, the number of pieces of preliminary data 203 is greater than the number of pieces of target data 303. Therefore, in the first embodiment, the standby device 101 is an industrial device that is trained using a large number of pieces of preliminary data 203, and the target device 103 is an industrial device that is trained using a smaller number of pieces of target data 303 than the standby data 203.

[0017] The backup device 101 is used for learning using the backup data 203, and the target device 103 performs processing on a processing target that will become a product. When the target device 103 is operated, a learning model trained using the backup data 203 and a learning model trained using the target data 303 are used.

[0018] The sensor included in the standby sensor unit 102 measures the state of the standby device 101 or the state of the processing target of the standby device 101. The sensor included in the target sensor unit 104 measures the state of the target device 103 or the state of the processing target of the target device 103.

[0019] The spare sensor unit 102 transmits the measured data as spare data 203 to the base model learning unit 105. A spare label (a spare label 401 described later) corresponding to the spare data 203 is also attached to the spare data 203, and this spare label 401 is input to the base model learning unit 105. The spare data 203 and the spare label 401 are used for machine learning by the base model learning unit 105.

[0020] The target sensor unit 104 transmits the measured data as target data 303 to the adapted model training unit 106. In addition, a target label (target label 501 described later) corresponding to the target data 303 is attached to the target data 303, and this target label 501 is input to the adapted model training unit 106. The target data 303 and the target label 501 are used for machine learning by the adapted model training unit 106. The target label 501 is a first target label.

[0021] The base model training unit 105 generates a base model (a base model 403 described later) by learning the relationship between the preliminary data 203 and the preliminary labels 401. The adapted model training unit 106 generates an adapted model (an adapted model 503 described later) by learning the relationship between the target data 303 and the target labels 501.

[0022] The inference error acquisition unit 107 infers a target label corresponding to the target data 303 and calculates the error between the inferred target label and a target label 501 that actually corresponds to the target data 303 .

[0023] Specifically, the inference error acquisition unit 107 calculates the error (base model inference error 603 described later) between the target label inferred by inputting the target data 303 into the base model (preliminary model) 403 and the target label 501 corresponding to the target data 303.

[0024] In addition, the inference error acquisition unit 107 calculates the error (adapted model inference error 606 described later) between the target label inferred by inputting the target data 303 into the adapted model (target model) 503 and the target label 501 corresponding to the target data 303.

[0025] The error model generation unit 108 generates an error model (error model 701 described later) to be used to integrate the result of inference using the base model 403 (base model inference result Br described later) and the result of inference using the adapted model 503 (adapted model inference result Ar described later).

[0026] The error model 701 is a model that calculates an error model inference result for target data (new target data 801, described later). The error model inference result is information indicating the relationship between the magnitude of the error when inference is performed using the base model 403 and the magnitude of the error when inference is performed using the adapted model 503. In other words, the error model inference result is information indicating the relationship between the magnitude of the base model inference result Br, which is the inference result when new target data 801 is input to the base model 403, and the magnitude of the adapted model inference result Ar, which is the inference result when new target data 801 is input to the adapted model 503. The error model inference result corresponds to the predicted value of the error in the base model inference result Br and the predicted value of the error in the adapted model inference result Ar.

[0027] When the base model inference result Br and the adapted model inference result Ar are integrated, the error model inference result is used to determine the weighting to be set for the base model inference result Br and the weighting to be set for the adapted model inference result Ar. That is, the error model generation unit 108 generates an error model 701 for inferring the weighting to be used when the inference results of each inference model are integrated.

[0028] The error model 701 calculates an error model inference result such that the smaller the error when calculated using the adapted model 503, the higher the priority of the adapted model 503. Furthermore, in areas where there is no pre-trained data, the error model 701 calculates an error model inference result such that the priority of the base model 403 is higher because the base model 403 has been trained using a larger amount of data.

[0029] The learning device 100 is connected to the standby device 101 when acquiring the preliminary data 203 from the standby device 101 and generating the base model 403. The learning device 100 is also connected to the target device 103 when acquiring the target data 303 from the target device 103 and generating the adapted model 503.

[0030] For example, learning device 100 generates base model 403 using spare device 101 at a manufacturer that produces spare device 101 and target device 103. After that, learning device 100, an inference device 800 (described later), and target device 103 are purchased, and learning device 100 generates adapted model 503 and error model 701 using target device 103 at the purchaser. Then, inference device 800 infers the state of target device 103, etc., using base model 403, adapted model 503, and error model 701 at the purchaser.

[0031] Alternatively, the spare device 101, the target device 103, the learning device 100, and the inference device 800 may be purchased, and the learning device 100 may learn the base model 403 using the spare device 101 at the place of purchase.

[0032] The learning device 100 may learn the base model 403 and the adapted model 503 at any location. The learning device 100 may also acquire the preliminary data 203 measured by the preliminary sensor unit 102 via another external device. The learning device 100 may also acquire the target data 303 measured by the target sensor unit 104 via another external device.

[0033] Furthermore, the learning device 100 may store a base model 403 that has been trained in advance by another learning device. In this case, the learning device 100 does not need to include the base model training unit 105.

[0034] 2 is a diagram illustrating a configuration example of the spare sensor unit 102 according to the first embodiment. The spare sensor unit 102 includes a sensor 201 that measures the state of the spare device 101 or the state of the object to be processed by the spare device 101, and outputs the measured signal as spare data 203. The spare data 203 is, for example, data detected (measured) by a current sensor, data detected by an optical sensor, data detected by a force sensor, etc.

[0035] The spare sensor unit 102 may include a feature extraction unit 202. Fig. 2 shows a case where the spare sensor unit 102 includes the feature extraction unit 202. The feature extraction unit 202 extracts features from the signal measured by the sensor 201, and outputs a set of the extracted features as spare data 203.

[0036] 3 is a diagram illustrating a configuration example of the target sensor unit 104 according to the first embodiment. The target sensor unit 104 includes a sensor 301 that measures the state of the target device 103 or the state of the processing target of the target device 103, and outputs the measured signal as target data 303.

[0037] The target sensor unit 104 may include a feature extraction unit 302. Fig. 3 shows a case where the target sensor unit 104 includes the feature extraction unit 302. The feature extraction unit 302 extracts features from the signal measured by the sensor 301, and outputs a set of the extracted features as target data 303.

[0038] Here, the feature quantity is a quantity that indicates the characteristics of the sensor signal. The feature quantity is, for example, a statistical quantity such as the mean value, variance, standard deviation, kurtosis, skewness, median, percentile value, maximum value, minimum value, crest factor, or number of peaks of the sensor signal. The feature quantity may also be the result of frequency analysis of the sensor signal, the result of filter bank analysis of the sensor signal (the result of summarizing the results of frequency analysis in a specific frequency range), the result of cepstrum analysis of the sensor signal, or the like. The feature quantity extraction units 202 and 302 can use a general feature quantity extraction method.

[0039] The backup device 101 may be a single specific device, or multiple devices may be used as the backup device 101 to acquire the backup data 203 .

[0040] 4 is a diagram for explaining the operation of the base model learning unit 105 according to the first embodiment. The base model learning unit 105 receives the preliminary data 203 and the preliminary label 401 attached to the preliminary data 203.

[0041] For example, if data collected by a current sensor when a ball screw is operated is preliminary data 203, the label indicating the state of the ball screw, etc. is preliminary label 401. Also, if data collected by an optical sensor when a laser cutting device processes an object to be processed is preliminary data 203, the label indicating the processing results, etc. of the object to be processed is preliminary label 401. Also, if data collected by a force sensor when a robot device tightens a screw is preliminary data 203, the label indicating the tightening results, etc. of the screw is preliminary label 401.

[0042] The base model learning unit 105 learns the relationship between the preliminary data 203 and the preliminary labels 401 attached to the preliminary data 203, thereby generating a base model 403 that indicates the relationship between the preliminary data 203 and the preliminary labels 401. The base model 403 is a model for inferring the preliminary labels 401 from the preliminary data 203. The base model learning unit 105 outputs the generated base model 403.

[0043] 5 is a diagram for explaining the operation of the adaptive model training unit 106 according to the first embodiment. The adaptive model training unit 106 receives the target data 303 and a target label 501 attached to the target data 303. The target label 501 is data similar to the preliminary label 401.

[0044] The adapted model learning unit 106 generates an adapted model 503 by learning the relationship between the target data 303 and the target label 501 attached to the target data 303. The adapted model 503 is a model for inferring the target label 501 from the target data 303. The adapted model learning unit 106 outputs the generated adapted model 503.

[0045] Here, the spare label 401 and the target label 501 may be attached by a person who observes the state of the apparatus (the spare apparatus 101 and the target apparatus 103) or the state of the processing object. Alternatively, the spare label 401 and the target label 501 may be attached automatically based on the observation results of the spare sensor unit 102 or the target sensor unit 104. Alternatively, the spare label 401 and the target label 501 may be attached automatically based on the observation results of a sensor or system other than the spare sensor unit 102 or the target sensor unit 104 that observes the state of the apparatus or the processing object.

[0046] A general regression model or classification model can be used for the base model 403 and the adapted model 503. The type of model is selected for the base model 403 and the adapted model 503 depending on the application or purpose. An example of the base model 403 and the adapted model 503 is a neural network. Other examples of the base model 403 and the adapted model 503 include a recurrent neural network, a deep neural network, a linear regression model, a Gaussian process regression model, a support vector regression model, a linear discriminant model, a support vector machine, a logistic regression, a mixture distribution model, Bayesian estimation, a decision tree (classification tree, regression tree), etc.

[0047] 6 is a diagram showing an example of the configuration of the inference error acquiring unit 107 according to the first embodiment. The inference error acquiring unit 107 includes a base model inference unit 601, a base model error calculation unit 602, an adapted model inference unit 604, and an adapted model error calculation unit 605.

[0048] The base model inference unit 601 inputs the target data 303, instead of the preliminary data 203, into the base model 403, and infers a base model inference result Br (second target label) corresponding to the target data 303. The base model inference result Br is the state of the device or the state of the processing target, and is an inferred target label.

[0049] The base model error calculation unit 602 calculates the error between the base model inference result Br and the target label 501 as the base model inference error 603 .

[0050] The adapted model inference unit 604 inputs the target data 303 into the adapted model 503, and infers an adapted model inference result Ar (third target label) corresponding to the adapted model 503. The adapted model inference result Ar represents the state of the device or the state of the processing target, and is an inferred adapted label.

[0051] The adapted model error calculation unit 605 calculates the error between the adapted model inference result Ar and the target label 501 as an adapted model inference error 606 .

[0052] 7 is a diagram for explaining the operation of the error model generation unit 108 according to the first embodiment. The error model generation unit 108 receives the target data 303, the base model inference error 603, and the adapted model inference error 606. The error model generation unit 108 performs learning based on the target data 303, the base model inference error 603, and the adapted model inference error 606 to generate an error model 701.

[0053] The error model 701 is a model that calculates an error model inference result from new target data 801, which is new target data. The error model 701 includes, for example, a model that calculates information corresponding to the error of the inference result when the base model 403 is applied to the new target data 801 (a base error model inference result Eb1, which will be described later), and a model that calculates information corresponding to the error of the inference result when the adaptive model 503 is applied to the new target data 801 (an adaptive error model inference result Ea1, which will be described later). Note that the error model 701 is not limited to a model that calculates two error model inference results, and may be a model that calculates one error model inference result.

[0054] 8 is a diagram showing a configuration example of an inference device 800 according to the first embodiment. The inference device 800 includes a base model inference unit 802, an adapted model inference unit 803, an error model inference unit 804, and an integration unit 805, and further stores a base model 403, an adapted model 503, and an error model 701. Note that at least one of the base model 403, the adapted model 503, and the error model 701 may be stored in an external device.

[0055] The inference device 800 receives new target data 801, which is a signal newly measured by the target sensor unit 104 in the target device 103, from the target sensor unit 104. The new target data 801 is data newly measured by the target sensor unit 104 in the target device 103 when the target device 103 starts operating again. The target device 103 starts operating again when the learning of the adaptive model 503 and the error model 701 for the target device 103 is completed and the target device 103 processes or transports an actual product.

[0056] When inference device 800 receives new target data 801, it estimates a target label (new target label) corresponding to new target data 801 and outputs it as inference result 806. That is, inference device 800 outputs inference result 806 corresponding to new target data 801.

[0057] In the inference device 800, new target data 801 is input to a base model inference unit 802, an adapted model inference unit 803, and an error model inference unit 804. The base model inference unit 802 has the same functions as the base model inference unit 601. By inputting the new target data 801 to the base model 403, the base model inference unit 802 calculates a base model inference result Br, which is the result of inference by the base model 403. That is, the base model inference unit 802 estimates the base model inference result Br as a target label corresponding to the new target data 801 using the base model 403, thereby acquiring the base model inference result Br.

[0058] The adapted model inference unit 803 has the same function as the adapted model inference unit 604. By inputting new target data 801 to the adapted model 503, the adapted model inference unit 803 calculates an adapted model inference result Ar, which is the result of inference by the adapted model 503. That is, the adapted model inference unit 803 estimates the adapted model inference result Ar as an adaptation label corresponding to the new target data 801 using the adapted model 503, and thereby obtains the adapted model inference result Ar.

[0059] Furthermore, the error model inference unit 804 inputs the new target data 801 into the error model 701, and calculates the error model inference result E1, which is the result of inference by the error model 701. That is, the error model inference unit 804 estimates the error model inference result E1 using the error model 701 as information on the integration weighting corresponding to the new target data 801, and thereby acquires the error model inference result E1.

[0060] When the base model 403 is more reliable than the adapted model 503, the error model inference unit 804 calculates an error model inference result E1 in which the base model inference result Br is weighted more heavily than the adapted model inference result Ar.

[0061] On the other hand, if the base model 403 has a lower reliability than the adapted model 503, the error model inference unit 804 calculates an error model inference result E1 in which the base model inference result Br is weighted less than the adapted model inference result Ar.

[0062] Examples of the error model inference result E1 include the base error model inference result Eb1 and the adaptive error model inference result Ea1, which will be described later. The error model inference result E1 corresponds to, for example, the base model inference error 603 and the adaptive model inference error 606 for the new target data 801.

[0063] The integrating unit 805 calculates an inference result 806 by weighting and integrating the base model inference result Br and the adapted model inference result Ar based on the error model inference result E1. The integrating unit 805 weights the inference result with a higher reliability, between the base model inference result Br and the adapted model inference result Ar, more heavily than the inference result with a lower reliability. The integrating unit 805 outputs the inference result 806 to a display device (not shown) or the like.

[0064] At least one of learning device 100 and inference device 800 may be located on target device 103. Alternatively, at least one of spare device 101 and target device 103 may be connected to a network, and at least one of learning device 100 and inference device 800 may be partly or entirely located on a processing circuit on the network or on a processing circuit on the Internet cloud.

[0065] When learning the error model 901, the learning device 100 may learn the error model (the error model 901 described later) after performing dimensional compression on the target data 303. In this case, the learning device 100 includes a dimensional compression unit (the dimensional compression unit 902 described later) that performs dimensional compression on the target data 303.

[0066] 9 is a diagram for explaining the operations of the error model generating unit 108 and the dimensional compression unit 902 according to the first embodiment. The dimensional compression unit 902 performs dimensional compression on the target data 303. The dimensional compression unit 902 outputs the dimensionally compressed target data 303 to the error model generating unit 108. Details of the dimensional compression will be described later.

[0067] The error model generation unit 108 generates an error model 901 by performing learning based on the target data 303 whose dimensions have been compressed by the dimension compression unit 902, the base model inference error 603, and the adapted model inference error 606. The error model 901 is a model that calculates an inference result 806 from new target data 801, similar to the error model 701.

[0068] Fig. 10 is a diagram showing a second example configuration of inference device 1000 according to the first embodiment. Of the components in Fig. 10, components that achieve the same functions as inference device 800 shown in Fig. 8 are given the same reference numerals, and duplicated explanations will be omitted. Fig. 10 shows the configuration of inference device 1000, which is an example of inference device 800.

[0069] In addition to the components included in the inference device 800, the inference device 1000 includes a dimensionality reduction unit 902. The dimensionality reduction unit 902 included in the inference device 1000 has the same functions as the dimensionality reduction unit 902 included in the learning device 100.

[0070] The inference device 1000 also includes an error model inference unit 1001, which is an example of the error model inference unit 804, and stores an error model 901, which is an example of the error model 701. The error model 901 may be stored in an external device.

[0071] In the inference device 1000, new target data 801 is input to a base model inference unit 802, an adapted model inference unit 803, and a dimensionality reduction unit 902. Upon receiving the new target data 801, the dimensionality reduction unit 902 performs dimensionality reduction on the new target data 801. The dimensionality reduction unit 902 outputs the dimensionally reduced new target data 801 to an error model inference unit 1001.

[0072] In this way, the dimensionality reduction unit 902 performs dimensionality reduction on the new target data 801 by applying the same process as the dimensionality reduction performed on the target data 303 when the error model 901 was trained. The error model inference unit 1001 inputs the dimensionality-reduced new target data 801 to the error model 901, and calculates an error model inference result E2, which is the result of inference by the error model 901. That is, the error model inference unit 1001 estimates the error model inference result E2 using the error model 901 as integration weighting information corresponding to the dimensionally reduced new target data 801, thereby obtaining the error model inference result E2. Examples of the error model inference result E2 include the base error model inference result Eb1 and the adaptive error model inference result Ea1.

[0073] Generally, when a learning model is trained, the greater the dimension of the data to be trained, the greater the amount of calculation required for training and inference. Also, the greater the dimension of the data to be trained, the greater the amount of data required for training.

[0074] In embodiment 1, the learning device 100 learns the error model 901 using the dimensionally compressed target data 303, thereby reducing the amount of calculation required for learning and inference, enabling high-speed learning and inference, and further reducing the amount of data required for learning.

[0075] The dimension reduction unit 902 can apply, for example, a dimension reduction technique using principal component analysis as the dimension reduction process. In another example, the dimension reduction unit 902 can apply, as the dimension reduction process, general dimension reduction techniques such as dimension reduction using linear discriminant analysis (LDA), kernel principal component analysis, t-distributed stochastic neighbor embedding (t-SNE), non-negative matrix factorization (NMF), an autoencoder, or uniform manifold approximation and projection (UMAP).

[0076] Furthermore, if the target data 303 is time series data, the dimension reduction unit 902 can apply dimension reduction by extracting features, or can further apply the general dimension reduction method described above to the target data 303 from which features have been extracted.

[0077] In industrial systems, multiple industrial devices with the same configuration are often manufactured. In the past, when creating a learning model taking into account the individual differences between industrial devices, it was necessary to acquire and label a large amount of data for each individual industrial device before training the learning model.

[0078] The learning device 100 of the first embodiment acquires and labels a large amount of data (acquisition of preliminary data 203) for the standby device 101, which is an industrial device, and then trains the base model 403. Furthermore, the learning device 100 acquires and labels a smaller amount of data (acquisition of target data 303) for the target device 103 than the amount of data acquired for the standby device 101, and then trains the adapted model 503 and the error model 701. That is, in the first embodiment, a large amount of data is acquired and labeled for training the base model 403, but when training the adapted model 503 and the error model 701, less data is acquired and labeled than when training the base model 403. As a result, the learning device 100 can provide highly accurate training models (the base model 403, the adapted model 503, and the error model 701) that prevent accuracy degradation due to individual differences, even if there are individual differences in the devices (the standby device 101 and the target device 103), the sensors 201 and 301, the processing target, etc.

[0079] The learning device 100 of embodiment 1 is preferably applied when a large amount of preliminary data 203 is acquired for the backup device 101 and a smaller amount of target data 303 than the preliminary data 203 is acquired for the target device 103.

[0080] According to the first embodiment, the number of times that learning data (target data 303) for the target device 103 is acquired and labeled can be reduced, thereby reducing the number of steps required to build a learning model.

[0081] When the learning device 100 of the first embodiment is applied to the target device 103, the base model 403 may be trained in advance. By storing the pre-trained base model 403, the learning device 100 does not need to store the preliminary data 203 and the preliminary label 401. This allows the learning device 100 to be configured with a smaller storage capacity than when learning is performed using the preliminary data 203, the preliminary label 401, the target data 303, and the target label 501.

[0082] The amount of calculation required for model training depends on the number of data items to be trained. The data used by the training device 100 when training the adapted model 503 and the error model 701 are the target data 303 and the target label 501. Therefore, in the first embodiment, the amount of calculation can be reduced compared to when training is performed using the preliminary data 203, the preliminary label 401, the target data 303, and the target label 501, and therefore the inference device 800 can be constructed quickly.

[0083] In an industrial system, it is easy to acquire a large amount of data from the standby device 101, but it may be possible to acquire only a small amount of data from the target device 103. In the first embodiment, a large amount of data is acquired from the standby device 101, but the learning device 100 and the inference device 800 can be realized by acquiring a minimum amount of data from the target device 103, and therefore this can be applied even when it is difficult to acquire a large amount of data from the target device 103.

[0084] 11 is a flowchart showing an example of the operation of the learning device 100 according to the first embodiment. The learning device 100 learns the base model 403, the adapted model 503, and the error model 701 by executing the processes of steps S1 to S11 shown in FIG.

[0085] The learning device 100 acquires preliminary data 203 from the preliminary device 101 (step S1). The learning device 100 assigns preliminary labels 401 to the preliminary data 203 (step S2). Note that the process of assigning the preliminary labels 401 to the preliminary data 203 may be performed by another device or by a person. The base model learning unit 105 of the learning device 100 learns the relationship between the preliminary data 203 and the preliminary labels 401 and generates a base model 403 (step S3).

[0086] The learning device 100 acquires target data 303 from the target device 103 (step S4). The learning device 100 assigns a target label 501 to the target data 303 (step S5). Note that the process of assigning the target label 501 to the target data 303 may be performed by another device or by a person. The adapted model training unit 106 of the learning device 100 learns the relationship between the target data 303 and the target label 501 and generates an adapted model 503 (step S6). Note that the processes of steps S1 to S3 and the processes of steps S4 to S6 may be performed in any order.

[0087] The inference error acquisition unit 107 of the learning device 100 infers information corresponding to the target data 303 using the base model 403. That is, the inference error acquisition unit 107 inputs the target data 303 into the base model 403 and calculates the base model inference result Br (step S7). The inference error acquisition unit 107 calculates the base model inference error 603, which is the error between the base model inference result Br and the target label 501 (step S8).

[0088] The inference error acquiring unit 107 infers information corresponding to the target data 303 using the adapted model 503. That is, the inference error acquiring unit 107 inputs the target data 303 to the adapted model 503 and calculates the adapted model inference result Ar (step S9). The inference error acquiring unit 107 calculates the adapted model inference error 606, which is the error between the adapted model inference result Ar and the target label 501 (step S10). Note that the processing of steps S7 and S8 and the processing of steps S9 and S10 may be executed first, or the first may be executed first.

[0089] The error model generation unit 108 of the learning device 100 generates an error model 701 based on the target data 303, the base model inference error 603, and the adapted model inference error 606 (step S11).

[0090] 12 is a flowchart showing an example of the operation of inference device 800 according to embodiment 1. Inference device 800 calculates inference result 806 by executing the operations of steps S21 to S25 shown in FIG.

[0091] The inference device 800 reads the base model 403, adapted model 503, and error model 701 learned by the learning device 100. The inference device 800 acquires new target data 801 from the target device 103 (step S21).

[0092] The base model inference unit 802 of the inference device 800 infers information corresponding to the new target data 801 using the base model 403. That is, the base model inference unit 802 inputs the new target data 801 into the base model 403 and calculates the base model inference result Br (step S22).

[0093] Furthermore, the adapted model inference unit 803 of the inference device 800 infers information corresponding to the new target data 801 using the adapted model 503. That is, the adapted model inference unit 803 inputs the new target data 801 to the adapted model 503 and calculates the adapted model inference result Ar (step S23).

[0094] The error model inference unit 804 of the inference device 800 infers information corresponding to the new target data 801 using the error model 701. That is, the error model inference unit 804 inputs the new target data 801 into the error model 701 and calculates the error model inference result E1 (step S24). Note that the processing of steps S22, S23, and S24 may be performed in any order.

[0095] The integration unit 805 of the inference device 800 integrates the base model inference result Br and the adapted model inference result Ar based on the error model inference result E1 to calculate an inference result 806 (step S25).

[0096] Next, detailed description will be given of exemplary configurations of the error model generation unit 108 and the inference device 800. As a premise, the base model 403, which has learned a larger number of pieces of preliminary data 203 than the target data 303, can obtain more accurate inference results for a wider range of new target data 801 than the adapted model 503, which has learned only a smaller number of pieces of target data 303 than the preliminary data 203. On the other hand, due to individual differences between the backup device 101 and the target device 103, the tendency of the new target data 801 acquired from the target device 103 changes from the preliminary data 203, and depending on the new target data 801, the adapted model 503 can produce more accurate estimation results than the base model 403.

[0097] The labels used in training the learning model can be normalized and treated as falling within the range of 0 to 1 without losing generality. In this case, if the range of the label is 0 to 1, the range of the magnitude of the error will also be 0 to 1. For this reason, in the following, unless otherwise specified, it will be assumed that the range of the label is normalized to 0 to 1.

[0098] 13 is a diagram illustrating an example of the configuration of the error model generating unit 1300 according to the first embodiment. In FIG. 13, the configuration of the error model generating unit 1300, which is an example of the error model generating unit 108, is illustrated.

[0099] The error model generation unit 1300 includes a base error information setting unit 1301 , a base error model learning unit 1302 , an adaptive error information setting unit 1304 , and an adaptive error model learning unit 1305 .

[0100] The error model generation unit 1300 receives the target data 303, the base model inference error 603, and the adaptive model inference error 606 as input. The error model generation unit 1300 performs learning based on the target data 303, the base model inference error 603, and the adaptive model inference error 606, thereby generating the base error model 1303 and the adaptive error model 1306 as the error model 701.

[0101] The base error information setting unit 1301 is set in advance by the user with information (base error information) that specifies that the magnitude of the inference error by the base model 403 for the new target data 801 used during inference will be smaller than the magnitude of the inference error by the adapted model 503. The base error information is information that indicates the relationship between information on lattice points in space (data space) and information on labels at these lattice points. Note that a default value set in advance in the base error information setting unit 1301 may be used as the base error information.

[0102] The adaptive error information setting unit 1304 is configured in advance by the user with information (adaptive error information) that specifies that the magnitude of the inference error by the adapted model 503 for the new target data 801 used during inference will be larger than the magnitude of the inference error by the base model 403. The adaptive error information is information that indicates the relationship between information on lattice points in space and information on labels at these lattice points. Note that a default value configured in advance in the adaptive error information setting unit 1304 may be used as the adaptive error information.

[0103] The base error information setting unit 1301 sets the base error information set by the user in the base error model learning unit 1302. That is, the base error information setting unit 1301 sets in the base error model learning unit 1302 base error information indicating that inference using the new target data 801, inference by the base model 403 is more accurate than inference by the adapted model 503.

[0104] The base error information setting unit 1301 arranges data points in a grid pattern (at the vertices of a cube) in a spatial region where the preliminary data 203 and target data 303 acquired from the preliminary sensor unit 102 and target sensor unit 104 are distributed.

[0105] The base error information setting unit 1301 then generates base error information by setting the value set in the base error information. Specifically, the base error information setting unit 1301 generates base error information by setting a value such as "0" as a label for the magnitude of error for each data point to indicate that the magnitude of the error is small. That is, the base error information setting unit 1301 generates base error information indicating that the error in inference using the base model 403 is small. This is because the base model 403 is trained using a large amount of data. In this case, the label for the magnitude of error for the data points arranged in a grid pattern may be a value smaller than half the range of the magnitude of the inference error by the base model 403 and the adapted model 503. For example, if the range of the label is 0 to 1, the base error information setting unit 1301 can set a value less than 0.5 (e.g., "0") as the error label.

[0106] The base error model learning unit 1302 learns the relationship between the magnitude of the base model inference error 603 and the target data 303. In other words, the base error model learning unit 1302 learns the magnitude of the base model inference error 603 when the target data 303 is input to the base model 403.

[0107] During this learning, the base error model learning unit 1302 adds the base error information data to the learning data. At this time, the base error model learning unit 1302 performs learning by assigning a smaller weight to the base error information data than to the target data 303 and the base model inference error 603 data. This is because the base error information data is hypothetical data and is less reliable than the actually acquired target data 303 and the base model inference error 603 data. For example, if one data point contains the target data 303 and the base model inference error 603 data, the target data 303 and the base model inference error 603 data are learned with a larger weight than the base error information data.

[0108] As a result, when the base error model learning unit 1302 learns the base error model 1303, in the region of space where the target data 303 is distributed, it learns the relationship between the target data 303 and the base model inference error 603 in priority to the base error information.

[0109] Furthermore, when training the base error model 1303, the base error model training unit 1302 trains base error information in a space where the target data 303 does not exist. That is, when training the base error model 1303, the base error model training unit 1302 learns information that the magnitude of the error (a value less than 0.5) set in the data of the base error information is small in a space where the target data 303 does not exist. In other words, the base error model training unit 1302 calculates the base error model 1303 by learning that in a space where the target data 303 does not exist, the value of the base model inference error 603 corresponding to the target data 303 is a small value corresponding to the value (a value less than 0.5) set in the base error information.

[0110] The adaptive error information setting unit 1304 sets the adaptive error information set by the user in the adaptive error model learning unit 1305. That is, the adaptive error information setting unit 1304 sets in the adaptive error model learning unit 1305 adaptive error information indicating that inference using the new target data 801, the inference by the adaptive model 503 is less accurate than the inference by the base model 403.

[0111] The adaptive error information setting unit 1304 arranges data points in a grid pattern in a spatial region in which the preliminary data 203 and the target data 303 acquired from the preliminary sensor unit 102 and the target sensor unit 104 are distributed.

[0112] The base error information setting unit 1301 then generates base error information by setting the value set in the base error information. Specifically, the adaptive error information setting unit 1304 generates adaptive error information by setting a value such as "1" as a label for the magnitude of error for each data point to indicate that the magnitude of the error is large. That is, the adaptive error information setting unit 1304 generates adaptive error information indicating that the error in inference using the adapted model 503 is large. This is because the adapted model 503 is trained using a small amount of data. In this case, the error label for the data points arranged in a grid pattern may be a value greater than half the range of the magnitude of the inference error by the base model 403 and the adapted model 503. For example, if the range of the label is 0 to 1, the adaptive error information setting unit 1304 can set a value greater than 0.5 (e.g., "1") as the error label.

[0113] The adaptive error model learning unit 1305 learns the relationship between the magnitude of the adaptive model inference error 606 and the target data 303. That is, the adaptive error model learning unit 1305 learns the magnitude of the adaptive model inference error 606 when the target data 303 is input to the adaptive model 503.

[0114] During this learning, the adaptive error model learning unit 1305 adds the adaptive error information data to the learning data. At this time, the adaptive error model learning unit 1305 performs learning by assigning a smaller weight to the adaptive error information data than to the target data 303 and the adaptive model inference error 606 data. This is because the adaptive error information data is hypothetical data and is less reliable than the actually acquired target data 303 and the adaptive model inference error 606 data. For example, if one data point contains the target data 303 and the adaptive model inference error 606 data, the target data 303 and the adaptive model inference error 606 data are learned with a larger weight than the adaptive error information data.

[0115] As a result, when the adaptive error model learning unit 1305 learns the adaptive error model 1306, in the region of space where the target data 303 is distributed, it learns the relationship between the target data 303 and the adaptive model inference error 606 with priority over the adaptive error information.

[0116] Furthermore, when training the adaptive error model 1306, the adaptive error model training unit 1305 trains adaptive error information in a space where the target data 303 does not exist. That is, when training the adaptive error model 1306, the adaptive error model training unit 1305 learns information that the magnitude of the error (a value of 0.5 or more) set in the adaptive error information is large in a space where the target data 303 does not exist. In other words, the adaptive error model training unit 1305 calculates the adaptive error model 1306 by learning that in a space where the target data 303 does not exist, the value of the adaptive model inference error 606 corresponding to the target data 303 is a large value corresponding to the value (a value of 0.5 or more) set in the adaptive error information.

[0117] Fig. 14 is a diagram showing a third example configuration of inference device 1400 according to the first embodiment. Of the components in Fig. 14, components that achieve the same functions as inference device 800 shown in Fig. 8 are given the same reference numerals, and duplicated explanations will be omitted. Fig. 14 shows the configuration of inference device 1400, which is an example of inference device 800.

[0118] Like inference device 800, inference device 1400 includes a base model inference unit 802 and an adapted model inference unit 803. Also, like inference device 800, inference device 1400 stores a base model 403 and an adapted model 503.

[0119] The inference device 1400 also includes an error model inference unit 1401, which is an example of the error model inference unit 804, and an integration unit 1402, which is an example of the integration unit 805. The inference device 1400 also stores a base error model 1303 and an adaptive error model 1306, which are examples of the error model 701. At least one of the base model 403, the adaptive model 503, the base error model 1303, and the adaptive error model 1306 may be stored in an external device.

[0120] Error model inference unit 1401 includes base error model inference unit 1403 and adaptive error model inference unit 1404. Base error model 1303 is used in base error model inference unit 1403, and adaptive error model 1306 is used in adaptive error model inference unit 1404.

[0121] In the inference device 1400 , new target data 801 is input to a base model inference unit 802 , an adaptive model inference unit 803 , a base error model inference unit 1403 , and an adaptive error model inference unit 1404 .

[0122] The base error model inference unit 1403 inputs the new target data 801 into the base error model 1303, and calculates the base error model inference result Eb1, which is the result of inference by the base error model 1303. That is, the base error model inference unit 1403 estimates the base error model inference result Eb1 using the base error model 1303 as information on the integration weighting corresponding to the new target data 801, and thereby acquires the base error model inference result Eb1.

[0123] The adaptive error model inference unit 1404 inputs the new target data 801 into the adaptive error model 1306, and calculates the adaptive error model inference result Ea1, which is the result of inference by the adaptive error model 1306. That is, the adaptive error model inference unit 1404 estimates the adaptive error model inference result Ea1 using the adaptive error model 1306 as information on the integration weighting corresponding to the new target data 801, and thereby acquires the adaptive error model inference result Ea1.

[0124] The integrating unit 1402 calculates the inference result 806 by weighting and integrating the base model inference result Br and the adaptive model inference result Ar based on the base error model inference result Eb1 and the adaptive error model inference result Ea1. That is, the integrating unit 1402 determines the inference result 806 from the base error model inference result Eb1, the adaptive error model inference result Ea1, the base model inference result Br, and the adaptive model inference result Ar. The integrating unit 1402 outputs the inference result 806 to a display device (not shown) or the like. The inference device 1400 may also include a dimensionality reduction unit 902.

[0125] The base error model inference result Eb1 corresponds to a predicted value of the magnitude of the base model inference error 603 (the inferred value of the error by the base model 403) for the new target data 801. The adaptive error model inference result Ea1 corresponds to a predicted value of the magnitude of the adaptive model inference error 606 (the inferred value of the error by the adaptive model 503) for the new target data 801.

[0126] One example of a method for determining the inference result p, which is the inference result 806, is to calculate a weighted average (weighted mean) of the base model inference result Br and the adaptive model inference result Ar using the base error model inference result Eb1 and the adaptive error model inference result Ea1.

[0127] Here, the base error model inference result Eb1 is expressed as the base error model inference result e b and the adaptive error model inference result Ea1 is the adaptive error model inference result e t In addition, the base model inference result Br is set as the base model inference result p b and the adaptive model inference result Ar is the adaptive model inference result p t In this case, if the inference result 806 is taken as an inference result p, the integration unit 1402 determines the inference result p by, for example, the following equation (1).

[0128]

[0129] Another example of a method for determining the inference result p is the base error model inference result e b is the adaptive error model inference result e t If the base model inference result p b is used as the inference result 806. In this method, the base error model inference result e b is the adaptive error model inference result e t If the inference result p t is the inference result 806. In addition, the base error model inference result e b and the adaptive error model inference result e t When the magnitudes of the base model inference results p band the adaptive model inference result p t The average value of these is taken as the inference result 806.

[0130] In this way, the integration unit 1402 integrates the base model inference result p b and the adaptive model inference result p t For the error (base model inference result p b and the adaptive model inference result p t ) The base error model inference result e b and the adaptive error model inference result e t The integration unit 1402 determines the inference result 806 by using the base model inference result p b and the adaptive model inference result p t Among these, an inference result 806 that prioritizes an inference result that is considered to have a small error can be determined.

[0131] When the new target data 801 is included in the distribution of the target data 303, the error model inference unit 1401 generates the base error model inference result e having the magnitude of the error learned by the base error model 1303. b and outputs the adaptive error model inference result e having the magnitude of the error learned by the adaptive error model 1306. t will be output.

[0132] Here, consider a case where the new target data 801 is data with properties different from those of the target data 303, such that it deviates from the distribution of the target data 303. As described above, when training the base error model 1303, the error model generation unit 1300 prioritizes training the relationship between the target data 303 and the base model inference error 603 in the region of space where the target data 303 is distributed, and has learned information that the magnitude of the error set in the data of the base error information is small in the space where the target data 303 is not present. Also, when training the adaptive error model 1306, the error model generation unit 1300 prioritizes training the relationship between the target data 303 and the adaptive model inference error 606 in the region of space where the target data 303 is distributed, and has learned information that the magnitude of the error set in the data of the adaptive error information is large in the space where the target data 303 is not present.

[0133] From these facts, if the new target data 801 is data with properties different from the target data 303, such that it deviates from the distribution of the target data 303, the base error model inference result e b The value of the adaptive error model inference result e t Therefore, when the new target data 801 is out of the distribution of the target data 303, the integration unit 1402 calculates the inference result 806 according to the formula (1) to reduce the base model inference result p b The adaptive model inference result p t It will take priority over.

[0134] Therefore, if the inference device 1400 can predict that the new target data 801 exists in the region of space where the target data 303 is distributed and that the magnitude of the error in the inference result of the adapted model 503 is smaller than the magnitude of the error in the inference result of the base model 403, it can obtain an inference result 806 that prioritizes the inference result of the adapted model 503 over the inference result of the base model 403.

[0135] Furthermore, if the inference device 1400 predicts that the new target data 801 does not exist in the region of space where the target data 303 is distributed, or that the magnitude of the error in the inference result of the adapted model 503 is greater than the magnitude of the error in the inference result of the base model 403, it obtains an inference result 806 that prioritizes the inference result of the base model 403.

[0136] In other words, in the inference device 1400, the adapted model 503 is primarily responsible for inferring data that is considered to be accurately inferable and that is included in the trained target data 303 from which the new target data 801 has been trained. Also, in the inference device 1400, the base model 403, which has trained more data than the adapted model 503, is primarily responsible for inferring data that the adapted model 503 has not trained and data for which the inference of the adapted model 503 has a larger error than the inference of the base model 403.

[0137] It is generally known that when using a learning model, small differences in the input to the learning model can have a significant impact on the output from the learning model. For example, when a learning model is used to predict the operational results of an industrial system having industrial equipment, even if the industrial systems are the same, differences in the obtained data may occur due to individual differences between the industrial systems, such as the degree of deterioration of the drive unit, differences in friction due to the condition of the lubricant or grease, differences in ambient temperature, and installation errors and sensitivity differences of the sensors that acquire the data to be input to the learning model. Therefore, when a learning model generated by machine learning using data collected in one industrial system is installed as an inference device in another system, the difference between the training data and the obtained data may result in a large error in the output from the learning model, potentially resulting in inaccurate inference results.

[0138] The inference device 1400 of the first embodiment calculates the base error model inference result e corresponding to the inference error using the base model 403. b and the adaptive error model inference result e corresponding to the inference error using the adaptive model 503 t Based on this, the base model inference result Br and the adapted model inference result Ar are weighted and integrated. This allows the inference device 1400 to prioritize the adapted model 503 when the adapted model 503 is more reliable than the base model 403, and to prioritize the base model 403 when the base model 403 is more reliable than the adapted model 503. Therefore, the inference device 1400 can generate appropriate inference results with high inference accuracy.

[0139] According to the inference device 800 of embodiment 1, it is possible to obtain highly accurate inference results for the target device 103 compared to the cases of inference using only the base model 403, inference using only the adapted model 503, and integrating the inference results of the base model 403 and the adapted model 503 using a pre-set weight (fixed value).

[0140] In this way, the learning device 100 and the inference device 800 of embodiment 1 can improve the inference accuracy of a learning model when the learning model learned from data collected in other similar industrial systems is used in the target industrial system.

[0141] As described above, in the first embodiment, the error model generation unit 108 of the learning device 100 generates the error model 701 used to integrate the base model inference result Br and the adapted model inference result Ar by learning the relationship between the magnitude of the base model inference error 603, the magnitude of the adapted model inference error 606, and the target data 303. Therefore, the learning device 100 can generate a learning model that can generate appropriate inference results with high inference accuracy even for a new environment that differs from the environment in which data was previously acquired.

[0142] Furthermore, when new target data 801 is input, error model 701 outputs error model inference result E1, which is data corresponding to the weighting of base model inference result Br and adapted model inference result Ar at the time of integration. This allows inference device 800 to integrate the base model inference result Br and adapted model inference result Ar by weighting them based on error model inference result E1. Therefore, inference device 800 can calculate appropriate inference results with high inference accuracy even for a new environment different from the environment for which data was previously acquired.

[0143] Second Embodiment Next, a second embodiment will be described with reference to Fig. 15. In the second embodiment, weighting is performed using an anomaly detection model.

[0144] 15 is a diagram illustrating an example of the configuration of an error model generating unit 1500 according to the second embodiment. In FIG. 15, the configuration of the error model generating unit 1500, which is an example of the error model generating unit 108, is illustrated.

[0145] The error model generation unit 1500 includes an anomaly detection model learning unit 1507 , a base error information setting unit 1501 , a base error model learning unit 1502 , an adaptive error information setting unit 1504 , and an adaptive error model learning unit 1505 .

[0146] Similar to the error model generation unit 1300, the error model generation unit 1500 receives as input the target data 303, the base model inference error 603, and the adaptive model inference error 606. The error model generation unit 1500 performs learning based on the target data 303, the base model inference error 603, and the adaptive model inference error 606, thereby generating a base error model 1503 and an adaptive error model 1506 as the error model 701.

[0147] The anomaly detection model learning unit 1507 regards the target data 303 as normal data and learns an anomaly detection model. The anomaly detection method and the method for generating an anomaly detection model will be described later. The anomaly detection model is a model that determines whether the data (grid points) of the base error information are abnormal or normal. In other words, the anomaly detection model is a model that determines whether the grid points are in the space where the target data 303 existed.

[0148] When the base error information data is normal, grid points representing the base error information data exist in the space where the target data 303 exists. On the other hand, when the base error information data is abnormal, grid points representing the base error information data do not exist in the space where the target data 303 exists.

[0149] In the second embodiment, if the data of the base error information is normal, the target data 303 exists in the space where the data is distributed, and therefore the normal base error information is deleted. On the other hand, if the data of the base error information is abnormal, the target data 303 does not exist in the space where the data is distributed, and therefore the base error information is used as is. As a result, if the data of the base error information is normal, the target data 303 is adopted, and if the data of the base error information is abnormal, the base error information is adopted. The anomaly detection model learning unit 1507 outputs the anomaly detection model to the base error model learning unit 1502 and the adaptive error model learning unit 1505.

[0150] Similar to the base error information setting unit 1301, the base error information setting unit 1501 sets, in the base error model learning unit 1502, base error information that specifies that the magnitude of the inference error by the base model 403 for the new target data 801 used during inference will be smaller than the magnitude of the inference error by the adapted model 503.

[0151] Similar to the base error information setting unit 1301, the base error information setting unit 1501 arranges data points in a grid pattern in a spatial region where the data acquired from the spare sensor unit 102 and the target sensor unit 104 are distributed.

[0152] Then, similar to the base error information setting unit 1301, the base error information setting unit 1501 generates base error information in which the label for the magnitude of the error for each data point is set to a value less than 0.5 (for example, "0") to indicate that the magnitude of the error is small.

[0153] The base error information setting unit 1501 also uses the anomaly detection model generated by the anomaly detection model learning unit 1507 to evaluate the data of the base error information, and deletes base error information that is determined to be normal from the original base error information. That is, the base error information setting unit 1501 determines whether a grid point corresponding to the base error information exists in the space where the target data 303 is distributed, and if so, deletes this base error information. As a result, the base error information is no longer used in the space where the target data 303 is distributed.

[0154] The base error model learning unit 1502 learns the relationship between the magnitude of the base model inference error 603 and the target data 303. During this learning, the base error model learning unit 1502 adds data of base error information from which some data has been deleted to the learning data.

[0155] As a result, when the base error model learning unit 1502 learns the base error model 1503, it learns the relationship between the target data 303 and the base model inference error 603 in the spatial region where the target data 303 is distributed.

[0156] Furthermore, when training the base error model 1503, the base error model training unit 1502 trains base error information in a space where the target data 303 does not exist. That is, when training the base error model 1503, the base error model training unit 1502 learns information that the magnitude of the error (a value less than 0.5) set in the data of the base error information is small in a space where the target data 303 does not exist. In other words, the base error model training unit 1502 calculates the base error model 1503 by learning that in a space where the target data 303 does not exist, the value of the base model inference error 603 corresponding to the target data 303 is a small value corresponding to the value (a value less than 0.5) set in the base error information.

[0157] Similar to the adaptive error information setting unit 1304, the adaptive error information setting unit 1504 sets adaptive error information in the adaptive error model learning unit 1505, which specifies that the magnitude of the inference error by the adaptive model 503 for the new target data 801 used during inference will be larger than the magnitude of the inference error by the base model 403.

[0158] Similar to the adaptive error information setting unit 1304, the adaptive error information setting unit 1504 arranges data points in a grid pattern in the spatial region where the data acquired from the spare sensor unit 102 and the target sensor unit 104 are distributed.

[0159] Then, similar to the adaptive error information setting unit 1304, the adaptive error information setting unit 1504 generates adaptive error information in which the label for the magnitude of the error for each data point is set to a value of 0.5 or more (for example, "1") to indicate that the magnitude of the error is large.

[0160] Furthermore, the adaptive error information setting unit 1504 judges the adaptive error information data using the anomaly detection model generated by the anomaly detection model learning unit 1507, and deletes adaptive error information judged to be normal from the original adaptive error information. That is, the adaptive error information setting unit 1504 judges whether or not a grid point corresponding to the adaptive error information exists in the space where the target data 303 is distributed, and if so, deletes this adaptive error information. As a result, the adaptive error information is no longer used in the space where the target data 303 is distributed.

[0161] The adaptive error model learning unit 1505 learns the relationship between the magnitude of the adaptive model inference error 606 and the target data 303. During this learning, the adaptive error model learning unit 1505 adds the adaptive error information data, from which some data has been deleted, to the learning data.

[0162] As a result, when the adaptive error model learning unit 1505 learns the adaptive error model 1506, it learns the relationship between the target data 303 and the adaptive model inference error 606 in the spatial region where the target data 303 is distributed.

[0163] Furthermore, when training the adaptive error model 1506, the adaptive error model training unit 1505 trains adaptive error information in a space where the target data 303 does not exist. That is, when training the adaptive error model 1506, the adaptive error model training unit 1505 learns information that the magnitude of the error (a value of 0.5 or more) set in the adaptive error information is large in a space where the target data 303 does not exist. In other words, the adaptive error model training unit 1505 calculates the adaptive error model 1506 by learning that in a space where the target data 303 does not exist, the value of the adaptive model inference error 606 corresponding to the target data 303 is a large value corresponding to the value (a value of 0.5 or more) set in the adaptive error information.

[0164] Even when error model generation unit 1500 learns base error model 1503 and adaptive error model 1506 in this manner, the configuration of inference device 1400 is the same as the configuration described in Fig. 14. In the second embodiment, inference device 1400 uses base error model 1503 and adaptive error model 1506 instead of base error model 1303 and adaptive error model 1306, respectively.

[0165] There are various anomaly detection methods, but in order to identify the spatial region in which data is distributed, the anomaly detection model learning unit 1507 preferably uses a technique equivalent to novelty detection, which finds data points that are far from the spatial region in which the learned data is distributed. Examples of anomaly detection methods used by the anomaly detection model learning unit 1507 include the Mahalanobis-Taguchi System (MT method), the K-Nearest Neighbor algorithm (K-NN), the Local Outlier Factor (LOF), a One-Class Support Vector Machine (SVM), a Gaussian Mixture Model (GMM), and a method that uses an autoencoder.

[0166] As described above, in the second embodiment, the error model generating unit 1500 including the anomaly detection model learning unit 1507 generates the base error model 1503 and the adaptive error model 1506, thereby enabling the inference device 1400 to obtain highly accurate inference results. That is, the inference device 1400 can obtain highly accurate inference results in the target device 103 compared to the cases of performing inference using only the base model 403, performing inference using only the adaptive model 503, and integrating the inference results of the base model 403 and the adaptive model 503 using a preset weight (fixed value).

[0167] Furthermore, the learning device 100 equipped with the error model generation unit 1500 and the inference device 1400 can improve the inference accuracy of a learning model when the learning model learned from data collected in other similar industrial systems is used in the target industrial system.

[0168] As described above, according to the second embodiment, when the data of the base error information is normal, the anomaly detection model deletes the base error information. Furthermore, when the data of the adaptive error information is normal, the anomaly detection model deletes the adaptive error information. As a result, when training the base error model 1503, the base error model training unit 1502 learns the relationship between the target data 303 and the base model inference error 603 in the spatial region where the target data 303 is distributed. Furthermore, when training the adaptive error model 1506, the adaptive error model training unit 1505 learns the relationship between the target data 303 and the adaptive model inference error 606 in the spatial region where the target data 303 is distributed.

[0169] Furthermore, when inferring the base error model inference result Eb1, the inference device 1400 performs inference based on the relationship between the target data 303 and the base model inference error 603 in the spatial region where the target data 303 is distributed. Furthermore, when inferring the adaptive error model inference result Ea1, the inference device 1400 performs inference based on the relationship between the target data 303 and the adaptive model inference error 606 in the spatial region where the target data 303 is distributed. Therefore, the learning device 100 can generate a learning model that can generate appropriate inference results with high inference accuracy even for a new environment different from the environment where data was previously acquired. Furthermore, the inference device 1400 can calculate appropriate inference results with high inference accuracy even for a new environment different from the environment where data was previously acquired.

[0170] Embodiment 3 Next, embodiment 3 will be described using Figures 16 and 17. In embodiment 3, by setting a prior distribution before learning, it is learned that in a space without data points, the base model inference error 603 is small and the adaptive model inference error 606 is large. In addition, the base model inference result Br and the adaptive model inference result Ar are weighted and integrated using the standard deviation (standard deviation Sdb, described later) of the base error model inference result (base error model inference result Eb2, described later) and the standard deviation (standard deviation Sda, described later) of the adaptive error model inference result (adaptive error model inference result Ea2, described later).

[0171] 16 is a diagram illustrating an example of the configuration of an error model generating unit 1600 according to the third embodiment. In FIG. 16, the configuration of the error model generating unit 1600, which is an example of the error model generating unit 108, is illustrated.

[0172] The error model generation unit 1600 includes a base prior information setting unit 1601 , a base error model learning unit 1602 , an adaptive prior information setting unit 1604 , and an adaptive error model learning unit 1605 .

[0173] Similar to the error model generation units 1300 and 1500, the error model generation unit 1600 receives as input the target data 303, the base model inference error 603, and the adaptive model inference error 606. The error model generation unit 1600 performs learning based on the target data 303, the base model inference error 603, and the adaptive model inference error 606, thereby generating the base error model 1603 and the adaptive error model 1606 as the error model 701.

[0174] The base error model 1603 and adaptive error model 1606 generated by the error model generation unit 1600 are Gaussian Process Regression (GPR) models. In a Gaussian Process Regression model, a prior distribution can be set before learning to adjust the output value (data properties) in a space where there are no data points. The prior distribution is knowledge that is set in advance. Specifically, the prior distribution is information corresponding to the base error information and adaptive error information, and the properties of data in a data-free area are set.

[0175] A prior distribution (base prior distribution) that specifies that the magnitude of the inference error by the base model 403 for the new target data 801 used during inference will be smaller than the magnitude of the inference error by the adapted model 503 is set in advance in the base prior information setting unit 1601 by the user. Note that a default value set in advance in the base prior information setting unit 1601 may be used as the base prior distribution.

[0176] A prior distribution (adaptive prior distribution) is set in advance by the user in the adaptive prior information setting unit 1604. The prior distribution specifies that the magnitude of the inference error by the adapted model 503 for the new target data 801 used during inference will be larger than the magnitude of the inference error by the base model 403. Note that a default value set in advance in the adaptive prior information setting unit 1604 may be used as the adaptive prior distribution.

[0177] The base prior information setting unit 1601 sets the base prior distribution set by the user in the base error model learning unit 1602. That is, the base prior information setting unit 1601 sets in the base error model learning unit 1602 base error information indicating that inference using the new target data 801, inference by the base model 403 is more accurate than inference by the adapted model 503.

[0178] For example, before the base error model learning unit 1602 learns the base error model 1603, the base prior information setting unit 1601 sets a prior distribution with a small average value, such as "0.1," in the Gaussian process regression model that serves as the basis for the base error model 1603. This reduces the error in inference using the base error model 1603. Here, the base prior information setting unit 1601 sets the average value of the prior distribution so that the magnitude of the inference error by the base model 403 is smaller than the magnitude of the inference error by the adapted model 503.

[0179] For example, the average value of the prior distribution set by the base prior information setting unit 1601 may be a value smaller than half the range of the magnitude of the inference error by the base model 403 or the adapted model 503. For example, if the range of the label is 0 to 1, the base prior information setting unit 1601 can set a value less than 0.5 as the average value of the prior distribution.

[0180] The base error model learning unit 1602 generates a base error model 1603 by learning the relationship between the magnitude of the base model inference error 603 and the target data 303 using a Gaussian process regression model for which a prior distribution has been set by the base prior information setting unit 1601.

[0181] As a result, when learning the base error model 1603, the base error model learning unit 1602 prioritizes learning the relationship between the target data 303 and the base model inference error 603 in the region of space where the target data 303 is distributed. The base error model learning unit 1602 also learns information that the magnitude of the error (a value less than 0.5 set in the prior distribution) is small in a space where the target data 303 does not exist. In other words, the base error model learning unit 1602 calculates the base error model 1603 by learning that in a space where the target data 303 does not exist, the value of the base model inference error 603 corresponding to the target data 303 is a small value corresponding to the value (a value less than 0.5) set in the prior distribution.

[0182] For example, before the adaptive error model learning unit 1605 learns the adaptive error model 1606, the adaptive prior information setting unit 1604 sets a prior distribution with a large average value, such as "0.9," in the Gaussian process regression model that forms the basis of the adaptive error model 1606. This increases the error in inference using the adaptive error model 1606. Here, the adaptive prior information setting unit 1604 sets the average value of the prior distribution so that the magnitude of the inference error by the adaptive model 503 is larger than the magnitude of the inference error by the base model 403.

[0183] For example, the average value of the prior distribution set by the adaptive prior information setting unit 1604 may be a value greater than half the range of the magnitude of the inference error by the base model 403 or the adapted model 503. For example, if the range of the label is 0 to 1, the adaptive prior information setting unit 1604 can set a value of 0.5 or greater as the average value of the prior distribution.

[0184] The adaptive error model learning unit 1605 generates an adaptive error model 1606 by learning the relationship between the magnitude of the adaptive model inference error 606 and the target data 303 using a Gaussian process regression model for which a prior distribution has been set by the adaptive prior information setting unit 1604.

[0185] As a result, when learning the adaptive error model 1606, the adaptive error model learning unit 1605 prioritizes learning of the relationship between the target data 303 and the adaptive model inference error 606 in the region of space where the target data 303 is distributed. Furthermore, the adaptive error model learning unit 1605 learns information that the magnitude of the error (a value of 0.5 or more set in the prior distribution) is large in a space where the target data 303 is not present. In other words, the adaptive error model learning unit 1605 calculates the adaptive error model 1606 by learning that in a space where the target data 303 is not present, the value of the adaptive model inference error 606 corresponding to the target data 303 is a large value corresponding to the value (a value of 0.5 or more) set in the prior distribution.

[0186] Next, we will explain the configuration of an inference device (an inference device 1700 described later) corresponding to the error model generation unit 1600. Note that the inference device 1400 can also be used as the inference device corresponding to the error model generation unit 1600.

[0187] Fig. 17 is a diagram showing an example of the configuration of an inference device 1700 according to the third embodiment. Of the components in Fig. 17, components that achieve the same functions as those in inference device 1400 shown in Fig. 14 are given the same reference numerals, and duplicated explanations will be omitted. Fig. 17 shows the configuration of inference device 1700, which is an example of inference device 800.

[0188] Compared to inference device 1400, inference device 1700 includes an error model inference unit 1701 instead of error model inference unit 1401, and an integration unit 1702 instead of integration unit 1402. Error model inference unit 1701 is an example of error model inference unit 804, and integration unit 1702 is an example of integration unit 805. Inference device 1700 also stores base error model 1603 and adaptive error model 1606, which are examples of error model 701. At least one of base error model 1603 and adaptive error model 1606 may be stored in an external device.

[0189] Error model inference unit 1701 includes base error model inference unit 1703 and adaptive error model inference unit 1704. Base error model 1603 is used in base error model inference unit 1703, and adaptive error model 1606 is used in adaptive error model inference unit 1704.

[0190] In the inference device 1700 , new target data 801 is input to a base model inference unit 802 , an adaptive model inference unit 803 , a base error model inference unit 1703 , and an adaptive error model inference unit 1704 .

[0191] The base error model inference unit 1703 inputs the new target data 801 into the base error model 1603, and calculates the base error model inference result Eb2, which is the result of inference by the base error model 1603. That is, the base error model inference unit 1703 estimates the base error model inference result Eb2 using the base error model 1603 as information on the integration weighting corresponding to the new target data 801, and thereby acquires the base error model inference result Eb2.

[0192] The adaptive error model inference unit 1704 inputs the new target data 801 into the adaptive error model 1606, and calculates the adaptive error model inference result Ea2, which is the result of inference by the adaptive error model 1606. That is, the adaptive error model inference unit 1704 estimates the adaptive error model inference result Ea2 using the adaptive error model 1606 as information on the integration weighting corresponding to the new target data 801, and thereby acquires the adaptive error model inference result Ea2.

[0193] The integrating unit 1702 calculates the inference result 806 by weighting and integrating the base model inference result Br and the adaptive model inference result Ar based on the base error model inference result Eb2 and the adaptive error model inference result Ea2. That is, the integrating unit 1702 determines the inference result 806 from the base error model inference result Eb2, the adaptive error model inference result Ea2, the base model inference result Br, and the adaptive model inference result Ar. The integrating unit 1702 outputs the inference result 806 to a display device (not shown) or the like. Note that the inference device 1700 may also include a dimensionality reduction unit 902.

[0194] The base error model 1603 and the adaptive error model 1606 are Gaussian process regression models, and during inference, a standard deviation indicating the range in which the inference values ​​are distributed can be calculated along with the inference values. The smaller the standard deviation indicating the range in which the inference values ​​are distributed, the more reliable the inference results that have been obtained.

[0195] Specifically, the base error model 1603 calculates the base error model inference result Eb2, which is an inferred value, and the standard deviation Sdb of the base error model inference result Eb2. The adaptive error model 1606 calculates the adaptive error model inference result Ea2, which is an inferred value, and the standard deviation Sda of the adaptive error model inference result Ea2. The standard deviation Sdb is the first standard deviation, and the standard deviation Sda is the second standard deviation.

[0196] Here, the base error model inference result Eb2 is the base error model inference result e b Let the base error model inference result e b The standard deviation Sdb of b In addition, the adaptive error model inference result Ea2 is set as the adaptive error model inference result e t The adaptive error model inference result e t The standard deviation Sda is the standard deviation σ t In addition, the base model inference result Br is set as the base model inference result p b and the adaptive model inference result Ar is the adaptive model inference result p t Let's say.

[0197] In this case, the base error model inference unit 1703 of the error model inference unit 1701 receives the new target data 801 as input and generates the base error model inference result e b and the base error model inference result e b Standard deviation σ b In addition, the adaptive error model inference unit 1704 of the error model inference unit 1701 receives the new target data 801 as input and outputs the adaptive error model inference result e t and the adaptive error model inference result e t Standard deviation σ t and is output.

[0198] Base error model inference result e b corresponds to a predicted value of the magnitude of the base model inference error (the inferred value of the error by the base model 403) for the new target data 801. b Standard deviation σ b is the base error model inference result e b This is a value that indicates the reliability of the

[0199] Adaptive error model inference result e t corresponds to a predicted value of the adaptive model inference error (the inferred value of the error by the adaptive model 503) for the new target data 801, and the adaptive error model inference result e t Standard deviation σ t is the adaptive error model inference result e t This is a value that indicates the reliability of the

[0200] The integration unit 1702 integrates the base model inference result p b , adaptive model inference result p t , base error model inference result e b , base error model inference result e b Standard deviation σ b , adaptive error model inference result e t , and the adaptive error model inference result e t Standard deviation σ t The inference result 806 is determined from the base error model inference result e. b , base error model inference result e b Standard deviation σ b , adaptive error model inference result e t , and the adaptive error model inference result e t Standard deviation σ t The adaptive model inference result p t and the base model inference result p b Specifically, the integration unit 1702 calculates a weighted average of the base error model inference result e b and standard deviation σ b and the adaptive error model inference result e t and standard deviation σ tThe adaptive model inference result p t and the base model inference result p b In this method, if the inference result 806 is taken as an inference result p, the integration unit 1702 determines the inference result p by, for example, the following equation (2).

[0201]

[0202] Another example of a method for determining the inference result p is the base error model inference result e b and the base error model inference result e b Standard deviation σ b The sum of these is the adaptive error model inference result e t and the adaptive error model inference result e t Standard deviation σ t If the sum of the base model inference result p b is used as the inference result 806. In this method, the base error model inference result e b and the base error model inference result e b Standard deviation σ b The sum of these is the adaptive error model inference result e t and the adaptive error model inference result e t Standard deviation σ t If the sum of the adaptive model inference result p t is the inference result 806. In addition, the base error model inference result e b and the base error model inference result e b Standard deviation σ b The sum of these is the adaptive error model inference result e t and the adaptive error model inference result e t Standard deviation σ t If the sum of the base model inference result p b and the adaptive model inference result p t The average value of these is taken as the inference result 806.

[0203] In this way, the integration unit 1702 integrates the base model inference result p b , adaptive model inference result p t For the base error model inference result e b, base error model inference result e b Standard deviation σ b , adaptive error model inference result e t , and the adaptive error model inference result e t Standard deviation σ t The integration unit 1702 determines the inference result 806 by using the base model inference result p b and the adaptive model inference result p t Among these, an inference result 806 can be determined that places emphasis on an inference result that is predicted to have a small error.

[0204] Base error model inference result e b and the adaptive error model inference result e t correspond to the magnitude of the errors learned by the base error model 1603 and the adaptive error model 1606, respectively. b Standard deviation σ b and the adaptive error model inference result e t Standard deviation σ t are the base error model inference results e b and the adaptive error model inference result e t is an indicator of how reliable it is.

[0205] Generally, in Bayesian process regression inference, the larger the standard deviation, the less reliable the predicted value. The inference device 1700 evaluates the error magnitude of the inference result and its reliability together to calculate the base model inference result p b and the adaptive model inference result p t Therefore, reasoning device 1700 can determine inference result 806 with higher accuracy than if reliability were not available.

[0206] Furthermore, when training the base error model 1603, the error model generation unit 1600 learned the relationship between the target data 303 and the base model inference error 603 in the region of space where the target data 303 is distributed, and learned information that the magnitude of the error is small in the space where the target data 303 does not exist. Furthermore, when training the adaptive error model 1606, the error model generation unit 1600 learned the relationship between the target data 303 and the adaptive model inference error 606 in the region of space where the target data 303 is distributed, and learned information that the magnitude of the error is large in the space where the target data 303 does not exist.

[0207] From these facts, if the new target data 801 is data with properties different from the target data 303, such that it deviates from the distribution of the target data 303, the base error model inference result e b The value of the adaptive error model inference result e t Therefore, in the integration unit 1702, when the new target data 801 deviates from the distribution of the target data 303, the base model inference result p b The adaptive model inference result p t The base error model inference result e takes precedence over b Standard deviation σ b and the adaptive error model inference result e t Standard deviation σ t The inference result is determined taking into account the above.

[0208] Therefore, when inferring data that the adapted model 503 has not learned, or when inferring data in which the inference of the adapted model 503 results in a larger error than the inference of the base model 403, the inference result of the base model 403, which has learned more data than the adapted model 503, is given priority.

[0209] In this way, the inference device 1700 of embodiment 3 can obtain highly accurate inference results in the target device 103 compared to any of the following cases: when inferring using only the base model 403; when inferring using only the adapted model 503; and when integrating the inference results of the base model 403 and the adapted model 503 using a pre-set weight (fixed value).

[0210] Furthermore, the learning device 100 and the inference device 1700 equipped with the error model generation unit 1600 can improve the inference accuracy of a learning model when the learning model learned from data collected in other similar industrial systems is used in the target industrial system.

[0211] Thus, according to the third embodiment, the learning device 100 sets a prior distribution before learning, thereby learning that in a space without data points, the base model inference error 603 is small and the adapted model inference error 606 is large. Then, the inference device 1700 infers that in a space without data points, the base model inference error 603 is small and the adapted model inference error 606 is large. Therefore, the inference device 1700 can generate an appropriate inference result 806 with high inference accuracy using the base model 403 for an environment (a space without data points) different from the environment in which data was previously acquired.

[0212] Furthermore, the inference device 1700 calculates the base error model inference result e b Standard deviation σ b and the adaptive error model inference result e t Standard deviation σ t The inference device 1700 weights and integrates the base model inference result Br and the adaptive model inference result Ar using the above. b and the adaptive error model inference result e t Reliability (standard deviation σ b and standard deviation σ t ), the base model inference result Br and the adapted model inference result Ar can be appropriately integrated.

[0213] 18 to 20, a fourth embodiment will be described. In the fourth embodiment, a model (an adaptation domain model 1804, described later) is generated by learning a domain in which the magnitude of the adapted model inference error 606 is smaller than the magnitude of the base model inference error 603, and the adaptation domain model 1804 is used to weight and integrate the base model inference result Br and the adapted model inference result Ar.

[0214] Fig. 18 is a diagram illustrating an example of the configuration of an error model generating unit 1800 according to the fourth embodiment. Fig. 18 illustrates the configuration of the error model generating unit 1800, which is an example of the error model generating unit 108.

[0215] The error model generation unit 1800 includes a data extraction unit 1801 and an adaptive domain learning unit 1803. The error model generation unit 1800 receives as input the target data 303, the base model inference error 603, and the adaptive model inference error 606. The error model generation unit 1800 performs learning based on the target data 303, the base model inference error 603, and the adaptive model inference error 606, thereby generating an adaptive domain model 1804 as the error model 701.

[0216] The data extraction unit 1801 compares the magnitude of the base model inference error 603 with the magnitude of the adapted model inference error 606 for each data. The data extraction unit 1801 extracts sensor data corresponding to data in which the magnitude of the adapted model inference error 606 is smaller than the magnitude of the base model inference error 603 from the target data 303, and sets the extracted sensor data as successful adaptation data 1802.

[0217] The adaptation domain learning unit 1803 learns the domain (space on the data) in which the successful adaptation data 1802 is distributed (exists), thereby generating an adaptation domain model 1804. Therefore, the adaptation domain model 1804 is a model that has learned the distribution of sensor data that the adapted model 503 was able to infer more accurately than the base model 403. In other words, the adaptation domain model 1804 is a model that can infer a domain in which the magnitude of the adapted model inference error 606 is predicted to be smaller than the magnitude of the base model inference error 603.

[0218] Here, the adaptive domain model 1804 may be a general outlier detection model, and among outlier detection models, it is preferable that it be a novelty detection model that evaluates and determines whether newly given data belongs to a domain. The adaptive domain model 1804 does not detect outliers in the adaptive model inference error 606, but rather uses an outlier detection method using a general outlier detection model to infer a domain where the magnitude of the adaptive model inference error 606 is predicted to be smaller than the magnitude of the base model inference error 603.

[0219] Next, an inference device 1900 corresponding to the error model generating unit 1800 will be described. Fig. 19 is a diagram showing an example of the configuration of the inference device 1900 according to the fourth embodiment. Of the components in Fig. 19, components that achieve the same functions as those in the inference device 800 shown in Fig. 8 are given the same reference numerals, and duplicated explanations will be omitted. Fig. 19 shows the configuration of the inference device 1900, which is an example of the inference device 800.

[0220] Like inference device 800, inference device 1900 includes a base model inference unit 802 and an adapted model inference unit 803. Also, like inference device 800, inference device 1900 stores a base model 403 and an adapted model 503.

[0221] The inference device 1900 also includes an adaptation domain model inference unit 1901, which is an example of the error model inference unit 804, and an integration unit 1902 instead of the integration unit 805. The inference device 1900 also stores an adaptation domain model 1804, which is an example of the error model 701. The adaptation domain model 1804 is used by the adaptation domain model inference unit 1901. At least one of the base model 403, the adapted model 503, and the adaptation domain model 1804 may be stored in an external device.

[0222] In the inference device 1900 , new object data 801 is input to a base model inference unit 802 , an adapted model inference unit 803 , and an adapted domain model inference unit 1901 .

[0223] The adaptation domain model inference unit 1901 inputs the new target data 801 into the adaptation domain model 1804, and calculates an adaptation domain model inference result Aa, which is the result of inference by the adaptation domain model 1804. That is, the adaptation domain model inference unit 1901 estimates the adaptation domain model inference result Aa corresponding to the new target data 801 using the adaptation domain model 1804, and thereby acquires the adaptation domain model inference result Aa. The adaptation domain model inference result Aa is information indicating whether the new target data 801 is within the adaptation domain, which is a domain where the magnitude of the adaptation model inference error 606 is smaller than the magnitude of the base model inference error 603.

[0224] As described above, the adaptation domain model 1804 is a model that has learned sensor data (successful adaptation data 1802) corresponding to data in which the adaptation model inference error 606 is smaller than the base model inference error 603. Therefore, when the adaptation domain model inference unit 1901 infers an error corresponding to the input new target data 801, the adaptation domain model inference result Aa serves as an index of whether the data is likely to be more accurately judged by the adaptation model 503 than by the base model 403.

[0225] The integrating unit 1902 integrates the base model inference result Br and the adapted model inference result Ar based on the adaptation domain model inference result Aa, thereby outputting the inference result 806. For example, the integrating unit 1902 determines a weight when integrating the base model inference result Br and the adapted model inference result Ar based on the adaptation domain model inference result Aa, and integrates the base model inference result Br and the adapted model inference result Ar using this weight.

[0226] Here, the adaptive domain model inference result Aa is expressed as the adaptive domain model inference result p a The weight is γ. The base model inference result Br is expressed as the base model inference result p b and the adaptive model inference result Ar is the adaptive model inference result p t In this case, if the inference result 806 is the inference result p, the integration unit 1902, for example, a The weight γ is determined based on the above, and the inference result p is determined by the following equation (3).

[0227]

[0228] In this way, the integration unit 1902 calculates the base model inference result p b and the adaptive model inference result p t The inference result p obtained by integrating these is set as an inference result 806. The method for determining the weight γ will be described later.

[0229] As another example of a method for determining the inference result p, the integration unit 1902 may a Based on this, the adaptive model inference result p t or base model inference result p b is used as the inference result 806. In this method, the integration unit 1902 determines whether the new target data 801 belongs to the area where the adaptation success data 1802 is distributed. If the new target data 801 belongs to the area where the adaptation success data 1802 is distributed, the integration unit 1902 sets the adapted model inference result p t is the inference result 806, and if the new target data 801 does not belong to the distribution area of ​​the adaptation success data 1802, the base model inference result p b is the inference result 806.

[0230] Moreover, the error model generating unit 1800 may further include an adaptive domain model inferring unit 2001 and a domain threshold determining unit 2002 shown in Fig. 20. Fig. 20 is a diagram showing an example of the configuration of the adaptive domain model inferring unit 2001 and the domain threshold determining unit 2002 included in the error model generating unit 1800 according to the fourth embodiment.

[0231] The adaptation domain model inference unit 2001 inputs the successful adaptation data 1802 into the adaptation domain model 1804 learned by the adaptation domain learning unit 1803, and calculates a domain score (domain prediction result) using the adaptation domain model 1804. That is, the adaptation domain model inference unit 2001 acquires, as the domain score, the output value obtained by inputting the successful adaptation data 1802 into the adaptation domain model 1804. The domain score is data that numerically indicates whether or not the successful adaptation data 1802 belongs to the adaptation domain.

[0232] Here, as an example, a case will be described in which the higher the region score, the more likely the new target data 801 is to belong to the region in which the successful adaptation data 1802 is distributed. The region threshold determination unit 2002 sorts the region scores in descending order and determines the region score at a pre-specified position as the region threshold. In this way, the region threshold is determined excluding some data with low region scores.

[0233] The region threshold determination unit 2002 can determine, for example, a value lower than the median of the region score (e.g., the 20th percentile value) as the region threshold. This process is equivalent to specifying a ratio of the successful adaptation data 1802 to the number of data, sorting the region scores in descending order, and determining the region score value at the specified ratio as the region threshold. When the median of the region score is determined as the region threshold, successful adaptation data 1802 lower than the median of the region score is excluded from the successful adaptation data 1802. Furthermore, when the 20th percentile value of the region score is determined as the region threshold, successful adaptation data 1802 with a region score lower than the 20th percentile value is excluded from the successful adaptation data 1802.

[0234] The region threshold is the boundary between successful adaptation data 1802 and unsuccessful adaptation data. In other words, the region threshold is a threshold for determining whether the data is successful adaptation data 1802 or unsuccessful adaptation data. Unsuccessful adaptation data is sensor data in which the magnitude of the adapted model inference error 606 is greater than the magnitude of the base model inference error 603. The region threshold determination unit 2002 outputs the region threshold to the inference device 1900. This allows the inference device 1900 to calculate weighting using the region threshold. The inference device 1900 calculates, for example, the base model inference result p b and the adaptive model inference result p t In order to determine the weight γ when integrating the base model inference result p b is transformed using the domain threshold. b The method for converting is described later.

[0235] As described above, the adaptation domain model 1804 is an outlier detection model that has learned the domain in which the successful adaptation data 1802 is distributed. The adaptation domain model 1804 is used to determine whether the new target data 801 is included in the training data. In this adaptation domain model 1804, depending on the type of model or the model parameters used, outliers may be included in the successful adaptation data 1802, which may prevent the domain score from being calculated appropriately. Therefore, the error model generation unit 1800 determines the domain threshold by excluding some data with low domain scores, thereby preventing the inference accuracy of the domain score from being degraded by outliers.

[0236] Here, it is expected that the adapted model 503 will be able to perform inference more accurately than the base model 403 in the vicinity of the distribution of the successful adaptation data 1802. On the other hand, depending on the type of model or algorithm used in the adaptation region model 1804, the region threshold may be determined at the boundary where the successful adaptation data 1802 is distributed, and outliers may not be properly excluded. The region inside the boundary where the successful adaptation data 1802 is distributed is the successful adaptation data 1802, and the region outside is the unsuccessful adaptation data.

[0237] Therefore, the error model generation unit 1800 can determine a region threshold around the distribution of the successful adaptation data 1802 by using a region threshold expansion amount (parameter) set by the user as a parameter for expanding the range of the region in which the successful adaptation data 1802 is distributed. In other words, the error model generation unit 1800 can expand the boundary for excluding outliers by using the region threshold expansion amount. This allows the integrating unit 1902 to prioritize the inference result of the adapted model 503 over the base model 403, thereby improving the accuracy of the inference result 806.

[0238] If the threshold extension amount is a negative value, the error model generation unit 1800 narrows the boundary for excluding outliers. In this way, the error model generation unit 1800 adjusts the region threshold based on the threshold extension amount. That is, the error model generation unit 1800 adjusts the position of the boundary defined by the region threshold based on the threshold extension amount.

[0239] The region threshold determination unit 2002 determines the final region threshold by subtracting a constant defined as the region threshold expansion amount from the region threshold determined from the adaptation region model 1804 and the adaptation success data 1802 as described above, for example.

[0240] As another example of a method for determining the domain threshold, the integrating unit 1902 may a The threshold expansion amount of the region is added to the sum (expanded adaptive region model inference result p a ) may be compared to a region threshold.

[0241] In this way, the region threshold determination unit 2002 adjusts the region threshold using the threshold expansion amount in advance, thereby improving the adaptive region model inference result p a may be expanded, or the integration unit 1902 may expand the adaptation domain model inference result p a By adding the threshold expansion amount to the adaptive domain model inference result p a The integration unit 1902 adjusts the threshold expansion amount to expand the base model inference result p b and the adaptive model inference result p t It is possible to adjust the degree to which priority is given to either of the above.

[0242] In addition, depending on the type of outlier detection model used as the adaptive domain model 1804, the adaptive model inference result p t Therefore, the integration unit 1902 integrates the base model inference result p b and the adaptive model inference result p t In order to determine the weight γ when integrating the base model inference result p b There are cases where it is necessary to convert the domain score. A method of converting the domain score is to use a monotonic saturation function. Specifically, the adaptive domain model 1804 converts the base model inference result p b As a method for converting, a saturated linear function, a sigmoid function, a hyperbolic tangent function (tanh function), etc. can be used.

[0243] Below, the base model inference result p b An example of a method for converting the base model inference result p bThe function for converting the value of d into the weight γ is defined as the score conversion function f. Here, the new target data 801 given at the time of inference is d new Let p be the adaptive domain model inference result of a certain data D. a P a (D), and the area threshold is b th The threshold expansion amount is b ex The threshold expansion amount b ex If you do not use b ex = 0. In this case, the weight γ is expressed by, for example, the following equation (4).

[0244]

[0245] When a sigmoid function is used as the score conversion function f, the score conversion function f is expressed by the following formula (5): sig are adjustment parameters that are determined in advance according to the characteristics of the model to be used or the characteristics of the data to be learned. The characteristics of the data include the distribution, density, etc. of the preliminary data 203 and the target data 303.

[0246]

[0247] Furthermore, when a saturated linear function is used for the score conversion function f, the score conversion function f is expressed by the following formula (6): lin is an adjustment parameter that is determined in advance according to the characteristics of the model to be used or the properties of the data to be learned.

[0248]

[0249] In this way, depending on the score transformation function f to be used, additional parameters exist, and the values ​​of these parameters are set in advance according to the characteristics of the model to be used or the nature of the data to be learned.

[0250] The error model generating unit 1800 calculates the threshold expansion amount b ex By using the threshold expansion amount b, it is possible to widen or narrow the area of ​​the data space in which the adapted model 503 is prioritized during inference. exWhen b is a positive value, the region of the data space in which the adapted model 503 is prioritized is widened, and when b is a negative value, the region of the data space in which the adapted model 503 is prioritized is narrowed. ex By adjusting the above, it is possible to allow the inference device 1900 to perform accurate inference.

[0251] For example, when a large amount of target data 303 is distributed in a space where almost no preliminary data 203 has been acquired, the error model generating unit 1800 ex By increasing the value of , the area determined by the adapted model 503 can be increased.

[0252] Furthermore, the error model generating unit 1800 uses the threshold expansion amount b for a space where the preliminary data 203 and the target data 303 are hardly acquired (a space with a small amount of data). ex By adjusting the above, it is possible to adjust the region determined by the base model 403 and the region determined by the adapted model 503 for the data-small space.

[0253] The error model generating unit 1800 generates a threshold expansion amount b ex In this case, if the inference device 1900 cannot perform accurate inference, the error model generating unit 1800 may, for example, increase the threshold expansion amount b for the data-poor space in accordance with further instructions from the user. ex By reducing the value, the area determined by the adapted model 503 is reduced.

[0254] Furthermore, it is inferred that the base model 403, which has been trained using a larger number of pieces of preliminary data 203 than the target data 303, will obtain accurate inference results for a wider range of new target data 801 than the adapted model 503, which has been trained using a smaller number of pieces of target data 303 than the preliminary data 203. From this, the inference device 1900 determines the adapted model inference result p t The base model inference result p bIt is desirable to give priority to the following.

[0255] In this way, the inference device 1900 of embodiment 4 can obtain highly accurate inference results in the target device 103 compared to any of the following cases: when inferring using only the base model 403, when inferring using only the adapted model 503, and when integrating the inference results of the base model 403 and the adapted model 503 using a pre-set weight (fixed value).

[0256] Furthermore, the inference device 1900 can obtain more accurate inference results than when the successful adaptation data 1802 is not extracted from the target data 303 and the adaptation domain model 1804 is trained on the domain in which the target data 303 is distributed.

[0257] Furthermore, the learning device 100 and the inference device 1900 equipped with the error model generation unit 1800 can improve the inference accuracy of a learning model when the learning model learned from data collected in other similar industrial systems is used in the target industrial system.

[0258] Thus, according to the fourth embodiment, the learning device 100 generates an adaptation domain model 1804 that learns a domain in which successful adaptation data 1802 in which the magnitude of the adaptive model inference error 606 is smaller than the magnitude of the base model inference error 603 is distributed. Then, using the adaptation domain model 1804, the inference device 1900 weights and integrates the base model inference result Br and the adaptive model inference result Ar based on the adaptation domain model inference result Aa, which indicates whether the new target data 801 is within the adaptation domain, which is a domain in which the magnitude of the adaptive model inference error 606 is smaller than the magnitude of the base model inference error 603. Therefore, the inference device 1900 can appropriately integrate the base model inference result Br and the adaptive model inference result Ar.

[0259] Fifth Embodiment Next, a fifth embodiment will be described using Figures 21 to 23. In the fifth embodiment, a specific example of the configuration of an industrial device to which the learning device 100 and the inference devices 800, 1000, 1400, 1700, and 1900 described in the first to fourth embodiments are applied will be described. The industrial system described in the fifth embodiment includes at least one of the learning device 100 and the inference devices 800, 1000, 1400, 1700, and 1900 described in the first to fourth embodiments, and an industrial device.

[0260] The configuration of the industrial device described below is an example of a device configuration that utilizes the learning device 100 and inference devices 800, 1000, 1400, 1700, and 1900 described in Embodiments 1 to 4. The example configuration of the industrial device described below does not limit other usage forms, and can be used in a form that matches the characteristics of the industrial system to which Embodiments 1 to 4 are applied or the purpose of using the learning model.

[0261] Specific examples of industrial devices included in industrial systems include servo control devices, machine tools, laser processing devices, electric discharge processing devices, robot devices, transport devices, etc. The inference results of the inference devices 800, 1000, 1400, 1700, and 1900 of the first to fourth embodiments are used in devices associated with these industrial devices, such as control devices, PCs (Personal Computers) that monitor or record the status of the devices, and servers.

[0262] As described below, the learning device 100 and the inference devices 800, 1000, 1400, 1700, and 1900 of the first to fourth embodiments are devices used to measure the state of each industrial device or the state of a workpiece in an industrial system with a sensor and perform learning and inference based on the measured data, and can be used in industrial devices. Note that, although the following describes a case in which the inference device 800 is applied to an industrial system, the inference devices 1000, 1400, 1700, and 1900, or the learning device 100, may also be applied to the industrial system.

[0263] The state of the learning device 100 in embodiment 5 is information that expresses the state of the learning device 100 in data, and information that measures the state of the learning device 100. Examples of the state of the learning device 100 include information about the drive unit, generated sound or light, information about specific parts, the degree of wear on the tool, and images of all or part of the learning device 100. Information about the drive unit includes whether the drive unit is abnormal or normal, the magnitude of the current flowing through the drive unit, the magnitude of the voltage applied to the drive unit, etc. Information about the specific part includes the temperature, position, speed, acceleration, angle, angular velocity, angular acceleration, pressure, etc. of the specific part.

[0264] The state of the processing object is a value measured to determine the state of the processing object, such as the acceleration or angular velocity obtained by an acceleration sensor or gyro sensor attached to the processing object. Examples of the state of the processing object include the intensity or wavelength of light if the processing involves light emission, the temperature of the processing object if the processing involves heat generation, the sound volume or frequency if the processing involves sound generation, and an image of the entire or partial processing object.

[0265] The items or values ​​measured by the measurement unit are usually selected to be related to the items to be predicted in the model output. Therefore, the type of sensor in the measurement unit must be devised depending on the items to be predicted by the model or the type of processing equipment. In addition, the sensors used in the measurement unit may be a combination of multiple sensors depending on the purpose.

[0266] Next, a specific example of a servo control device will be described. In a servo control device, machine learning is used to diagnose the deterioration or failure state of drive systems such as motors, amplifiers, gears, and ball screws, and to determine the quality of servo operation results. Furthermore, when a servo control device is installed in a machining device, machine learning is used to determine the quality of machining, diagnose tool deterioration, and so on.

[0267] Here, the deterioration diagnosis of a ball screw will be explained as an example. Fig. 21 is a diagram showing the configuration of a servo control device 2100, which is a first example of an industrial device according to the fifth embodiment. The servo control device 2100 of the industrial system includes a control unit 2101, a driver 2102, a current sensor 2103, an encoder 2104, a motor 2105, a coupling 2106, and a ball screw 2107.

[0268] Ball screw 2107 is attached to motor 2105 via coupling 2106, and converts the rotational motion of motor 2105 into linear motion. Also, a load 2108 is attached to ball screw 2107. Servo control device 2100 controls the position of load 2108.

[0269] The encoder 2104 is a sensor that detects the rotation angle of the motor 2105. The rotation angle detected by the encoder 2104 is read by the control unit 2101. The control unit 2101 calculates a command value for the current to be output to the motor 2105 based on the target position of the transported object 2108 and the detected rotation angle, and outputs this command value to the driver 2102.

[0270] Driver 2102 outputs a current to motor 2105 in accordance with the current command value to rotate motor 2105 and control the position of transported object 2108. Current sensor 2103 measures the current output by driver 2102. Here, the current value of the current measured by current sensor 2103 can also be replaced with the current command value from control unit 2101 to driver 2102. Current sensor 2103 outputs the measured current value as current detection data to an external device such as learning device 100 or inference device 800.

[0271] The ball screw 2107 is configured to include a rotating screw portion, a nut portion attached to the screw portion, and a ball portion arranged between the screw portion and the nut portion. The ball portion is made up of multiple spherical balls. The ball portion is sandwiched between the nut portion and the screw portion and is arranged so that the nut portion moves smoothly when the screw portion rotates. The transported object 2108 is attached to the nut portion and moves due to the rotation of the screw portion.

[0272] When the ball screw 2107 is repeatedly operated, the ball portion may become damaged, or the lubricant injected between the ball portion and the nut portion or between the ball portion and the screw portion may become insufficient. In such cases, the ball screw 2107 cannot operate smoothly, causing the transported object 2108 to vibrate, noise to be generated, and the position of the transported object 2108 to be unable to be accurately controlled. In such cases, it becomes necessary to inject lubricant or replace the ball screw 2107. However, even if an operator observes the operation of the ball screw 2107, it is difficult to estimate the degree of deterioration. Furthermore, the drive unit may be covered for safety reasons, making it impossible to directly observe the ball screw 2107. Therefore, the learning device 100 learns a learning model based on the current value (current detection data) acquired by the current sensor 2103, and the inference device 800 estimates the degree of deterioration of the ball screw 2107.

[0273] In this method, the servo control device 2100 is made to perform various operations. Then, under various conditions that reproduce normal and abnormal states of the servo control device 2100, the learning device 100 collects current detection data measured by the current sensor 2103 and performs machine learning to associate the measured current detection data with the degree of deterioration of the ball screw 2107. In this way, the learning device 100 can create a learning model that estimates the degree of deterioration of the ball screw 2107 from the current detection data measured by the current sensor 2103.

[0274] Here, an example of using the learning device 100 and the inference device 800 will be described using two servo control devices 2100 as an example. To distinguish between the first and second servo control devices 2100, the first servo control device will be referred to as servo control device 2100A, and the second servo control device will be referred to as servo control device 2100B. In addition, the reference numerals of the components included in the first servo control device 2100A will have an A suffix, and the reference numerals of the components included in the second servo control device 2100B will have a B suffix.

[0275] Consider a case where a learning model that has learned data acquired from servo control device 2100A estimates the degree of deterioration of ball screw 2107B based on data (current detection data) measured by current sensor 2103B of servo control device 2100B. In this case, between servo control device 2100A and servo control device 2100B, there are individual differences in motors 2105A, 2105B, ball screws 2107A, 2107B, and transported objects 2108A, 2108B, as well as installation errors of transported objects 2108A, 2108B. For this reason, a difference occurs between the current detection data measured by current sensor 2103A and the current detection data measured by current sensor 2103B, degrading the judgment accuracy of the learning model.

[0276] Next, the correspondence between the learning device 100, servo control device 2100A, and servo control device 2100B of the first to fourth embodiments will be described. With regard to the standby device 101 and target device 103 applied to the learning device 100 of the first to fourth embodiments, the standby device 101 corresponds to the servo control device 2100A, and the target device 103 corresponds to the servo control device 2100B. Furthermore, the standby sensor unit 102 corresponds to the current sensor 2103A, and the target sensor unit 104 corresponds to the current sensor 2103B.

[0277] The servo control device 2100A, which is the standby device 101, reproduces various states of the ball screw 2107A. In this case, the ball screw 2107A is operated and current detection data collected by the current sensor 2103A becomes the standby data 203. Then, a label indicating the state of the ball screw 2107A corresponding to the collected standby data 203 becomes the standby label 401. Then, the base model learning unit 105 generates a base model 403 that has learned the relationship between the standby data 203 and the standby label 401.

[0278] Furthermore, the servo control device 2100B, which is the target device 103, reproduces various states within a range in which the ball screw 2107B does not break down. In this case, the ball screw 2107B is operated and current detection data collected by the current sensor 2103B becomes the target data 303. Then, a label indicating the state of the ball screw 2107B corresponding to the collected target data 303 becomes the target label 501. Then, the adaptive model learning unit 106 generates an adapted model 503 that has learned the relationship between the target data 303 and the target label 501.

[0279] The inference error acquisition unit 107 infers an object label by inputting the object data 303 into the base model 403, and acquires the base model inference result Br. The inference error acquisition unit 107 calculates the error between the base model inference result Br and the object label 501, thereby acquiring the base model inference error 603.

[0280] The inference error acquiring unit 107 infers a target label by inputting the target data 303 into the adapted model 503, and acquires the adapted model inference result Ar. The inference error acquiring unit 107 calculates the error between the adapted model inference result Ar and the target label 501, thereby acquiring the adapted model inference error 606.

[0281] The error model generation unit 108 learns the error model 701 based on the target data 303 , the base model inference error 603 , and the adapted model inference error 606 .

[0282] The servo control device 2100B, which is the target device 103, only needs to reproduce various states within the range in which the ball screw 2107B does not break down and acquire data, so data can be collected without breaking down the target device 103.

[0283] Next, the correspondence between the inference device 800 and the servo control device 2100B will be explained. The inference device 800 is used when the servo control device 2100B is operating. In response to the operation of the servo control device 2100B, the current detection data acquired by the current sensor 2103B becomes new target data 801.

[0284] A base model inference unit 802 of an inference device 800 inputs new target data 801 to the base model 403 to obtain a base model inference result Br, which is the inference result. Furthermore, an adapted model inference unit 803 inputs the new target data 801 to the adapted model 503 to obtain an adapted model inference result Ar, which is the inference result. An error model inference unit 804 inputs the new target data 801 to the error model 701 to obtain an error model inference result E1, which is the inference result. An integration unit 805 then integrates the base model inference result Br and the adapted model inference result Ar based on the error model inference result E1 to obtain an inference result 806.

[0285] In this way, the learning device 100 and the inference device 800 can be applied to the servo control device 2100. According to the learning device 100 and the inference device 800, when a learning model that has learned data collected by the servo control device 2100A is used in the servo control device 2100B, the inference accuracy of the learning model can be improved.

[0286] Next, a specific example of a laser processing device will be described. Laser processing devices include laser cutting devices, laser drilling devices, additional processing devices, laser marking devices, and laser welding devices. In laser processing devices, machine learning is used to determine whether the processing results are good or bad, adjust processing conditions, diagnose the deterioration state of the drive unit, and diagnose the malfunctioning state of the drive unit.

[0287] The following description will be given taking the pass / fail judgment in a laser cutting apparatus as an example. Fig. 22 is a diagram showing the configuration of a laser cutting apparatus 2200 which is a second example of the industrial apparatus according to the fifth embodiment.

[0288] The laser cutting apparatus 2200 included in the industrial system is an apparatus that uses a laser (laser light) to cut an object to be processed (workpiece) 2207. The laser cutting apparatus 2200 includes a control device 2201, a laser oscillator 2202, an optical path 2203, a processing head 2204, a drive unit 2205, and an optical sensor 2206.

[0289] The control device 2201 controls the laser oscillator 2202 and the drive unit 2205 based on specified processing conditions. The laser emitted from the laser oscillator 2202 is transmitted to the processing head 2204 via an optical path 2203. Here, the optical path 2203 may be configured as an optical system using mirrors or an optical fiber, depending on the type and characteristics of the laser used.

[0290] The processing head 2204 is equipped with an optical system that irradiates the transmitted laser onto the processing object 2207. The processing head 2204 also sprays processing gas from a nozzle toward a processing point on the processing object 2207. A drive unit 2205 changes the relative positions of the processing head 2204 and the processing object 2207.

[0291] The control device 2201 controls the relative positions of the processing head 2204 and the processing object 2207 using the drive unit 2205, moves the processing head 2204 so as to trace a path specified in the processing program, and commands the laser oscillator 2202 to irradiate the laser at an appropriate timing. The processing head 2204 cuts the processing object 2207 by irradiating it with a laser that is focused near the surface of the processing object 2207.

[0292] When irradiated with a laser, the workpiece 2207 is heated and emits light, so the state of the processing can be obtained by using time-series data (time-series data of light detection values) from the optical sensor 2206 that captures the light emitted from the workpiece 2207. The results of laser cutting include normal processing and abnormal processing. Abnormal processing includes defective processing such as dross, where the molten material adheres to the back surface of the workpiece 2207, scratches that cause unevenness on the cut surface, and burning, where the molten material blows up without penetrating the cut portion.

[0293] In laser cutting, it is not easy to check the results of the cutting process until the process is completed and the cut material is removed. Therefore, even if a processing defect occurs, the process may continue, resulting in waste of the material to be processed 2207. Furthermore, if a processing defect occurs, such as burning, causing the molten material to blow up, the molten material may adhere to the nozzle, preventing the processing gas from flowing normally, or the molten material may adhere to the optical system inside the processing head 2204, preventing the laser from being transmitted.

[0294] When collecting learning data, it is desirable to change the processing parameters in various ways so that the optical sensor 2206 acquires time-series data corresponding to various processing results. Then, the type of processing result (e.g., dross) is attached as a label to each piece of time-series data. When a label is attached to the time-series data of the optical sensor 2206, the processing result is associated with time based on the time-series data of the processing position and the processing result for the processing position, and then the time of the optical sensor 2206 is associated with the time of the processing result, thereby associating the processing result, i.e., the label, with the sensor information, i.e., the time-series data of the optical sensor 2206. In this association, time-series data having a certain time width is extracted from the time-series data, and a data pair in which the extracted time-series data and the label are associated is obtained. The learning model can learn the relationship between this data pair, i.e., the relationship between the extracted time-series data and the label, and thereby learn the relationship between the time-series data of the optical sensor 2206 and the processing result.

[0295] After the learning model has completed, the laser cutting device 2200 can obtain a prediction of the processing result as an output during processing by inputting time-series data from the optical sensor 2206 that detects the light emitted from the processing object 2207 into the learning model. Furthermore, if the predicted processing result is not good, the processing can be stopped or the processing parameters can be changed.

[0296] Here, an example of using the learning device 100 and the inference device 800 will be described using two laser cutting devices 2200 as an example. To distinguish between the first and second laser cutting devices 2200, the first laser cutting device will be referred to as laser cutting device 2200A, and the second laser cutting device will be referred to as laser cutting device 2200B. Furthermore, the reference numerals of the components included in the first laser cutting device 2200A will have an A suffix added to the end, and the reference numerals of the components included in the second laser cutting device 2200B will have a B suffix added to the end.

[0297] Consider a case where a learning model that has learned data acquired from the laser cutting apparatus 2200A predicts the processing result of the workpiece 2207B based on data (light detection values) measured by the optical sensor 2206B of the laser cutting apparatus 2200B. In this case, there are individual differences between the laser cutting apparatus 2200A and the laser cutting apparatus 2200B in the laser oscillators 2202A and 2202B, the optical systems of the processing heads 2204A and 2204B, the optical sensors 2206A and 2206B, the workpieces 2207A and 2207B, etc. As a result, there is a difference in the data acquired by the optical sensor 2206A and the optical sensor 2206B, and the judgment accuracy of the learning model deteriorates.

[0298] Next, the correspondence between the learning device 100, the laser cutting device 2200A, and the laser cutting device 2200B of the first to fourth embodiments will be described. With regard to the spare device 101 and the target device 103 applied to the learning device 100 of the first to fourth embodiments, the spare device 101 corresponds to the laser cutting device 2200A, and the target device 103 corresponds to the laser cutting device 2200B. Furthermore, the spare sensor unit 102 corresponds to the optical sensor 2206A, and the target sensor unit 104 corresponds to the optical sensor 2206B.

[0299] The laser cutting apparatus 2200A, which is the standby apparatus 101, processes the object 2207A under various changing processing conditions to reproduce processing states corresponding to various processing results. In this case, the data collected by the optical sensor 2206A becomes the preliminary data 203. Then, a label indicating the processing result of the object 2207A corresponding to the collected preliminary data 203 becomes the preliminary label 401. Then, the base model learning unit 105 generates a base model 403 that has learned the relationship between the preliminary data 203 and the preliminary label 401.

[0300] Furthermore, the laser cutting apparatus 2200B, which is the target device 103, processes the processing object 2207B under various processing conditions, thereby reproducing processing states corresponding to various processing results. In this case, the data collected by the optical sensor 2206B becomes the target data 303. At this time, to prevent breakdowns in the laser cutting apparatus 2200B, which is the target device 103, data is acquired within a range that does not cause processing defects that could cause breakdowns or malfunctions of the laser cutting apparatus 2200B. Then, a label indicating the processing result of the processing object 2207B corresponding to the collected target data 303 becomes the target label 501. The adapted model learning unit 106 then generates an adapted model 503 that has learned the relationship between the target data 303 and the target label 501.

[0301] Below, the process by which the inference error acquisition unit 107 acquires the base model inference error 603 and the adaptive model inference error 606, and the process by which the error model generation unit 108 learns the error model 701 are the same as when the inference device 800 is applied to the servo control device 2100, so their explanation will be omitted.

[0302] The laser cutting processing device 2200B, which is the target device 103, only needs to acquire data within a range that does not cause processing defects that could lead to breakdowns or malfunctions, so data can be collected without causing breakdowns in the target device 103.

[0303] Next, the correspondence between the inference device 800 and the laser cutting processing device 2200B will be explained. The inference device 800 is used when the laser cutting processing device 2200B is processing. The light detection value acquired by the optical sensor 2206B when the laser cutting processing device 2200B is processing becomes new target data 801.

[0304] Hereinafter, the process by which the inference device 800 calculates the inference result 806 from the new target data 801 is the same as when the inference device 800 is applied to the servo control device 2100, and therefore a description thereof will be omitted.

[0305] In this way, the learning device 100 and the inference device 800 can be applied to the laser cutting processing device 2200. According to the learning device 100 and the inference device 800, when a learning model that has learned data collected by the laser cutting processing device 2200A is used by the laser cutting processing device 2200B, the inference accuracy of the learning model can be improved.

[0306] Furthermore, in laser cutting, differences occur not only in the individual devices or sensors but also in the individual workpiece 2207, which can cause differences in the data acquired by the optical sensor 2206A, and this can degrade the inference accuracy of the learning model. Therefore, the learning device 100 and the inference device 800 can also be applied when the laser cutting device 2200A that collected the preliminary data 203 processes a workpiece that has similar properties to the workpiece 2207A that collected the preliminary data 203, but is different from the workpiece 2207A. Here, properties similar to those of the workpiece 2207A mean that the main components are the same but the proportion of additives is different, the surface condition is different, or the thickness differs by several to 20 percent.

[0307] Next, a specific example of a robot device will be described. There are various types of robot devices, such as vertical articulated type, horizontal articulated type, Cartesian coordinate type, and parallel link type. In robot devices, machine learning is used to diagnose the deterioration state of the drive system of the robot device, diagnose the malfunction of the drive system, determine whether the work performed by the robot device is good or bad, and determine collisions during robot operation. Furthermore, when a robot device is used for machining operations, machine learning is used to determine whether the machining is good or bad, diagnose tool deterioration, and the like.

[0308] The following description will be given taking as an example the determination of the quality of an assembly operation by a vertical articulated robot device. Fig. 23 is a diagram showing the configuration of a robot device 2300 as a third example of an industrial device according to the fifth embodiment.

[0309] The robot device 2300 included in the industrial system is a device that changes the position and angle of a part 2313 placed on a workbench 2312, tightens screws, and assembles the part 2313 with other parts.

[0310] The robot device 2300 includes a control unit 2301, a base 2302, a first axis 2303, a first arm 2304, a second axis 2305, a second arm 2306, a third axis 2307, a third arm 2308, a fourth axis 2309, a hand unit 2310, a force sensor 2311, and a workbench 2312.

[0311] In the robot device 2300, a first shaft 2303 that rotates about a vertical axis is attached to a fixed base 2302. A first arm 2304 is attached to the rotating part of the first shaft 2303, and a second shaft 2305 that rotates about a horizontal axis is attached to the tip of the first arm 2304.

[0312] A second arm 2306 is attached to the rotating part of the second shaft 2305, and a third shaft 2307 that rotates around a horizontal axis is attached to the tip of the second arm 2306. A third arm 2308 is attached to the rotating part of the third shaft 2307, and a fourth shaft 2309 that rotates around an axis in the extension direction of the third arm 2308 is attached to the tip of the third arm 2308. A hand unit 2310 is attached to the rotating part of the fourth shaft 2309, and a force sensor 2311 is attached to the hand unit 2310.

[0313] The hand unit 2310 is equipped with a mechanism for clamping an object, a mechanism for attaching a tool, etc. The force sensor 2311 measures the magnitude of the force or moment acting on the hand unit 2310. The control unit 2301 controls the rotation angles of the first axis 2303, the second axis 2305, the third axis 2307, and the fourth axis 2309, as well as the operation of the hand unit 2310, based on the measurement results (force detection data) from the force sensor 2311. The force detection data is data indicating the magnitude of the force or moment.

[0314] Consider the case where a pass / fail judgment is made on the fastening of a screw during the assembly work of a part 2313. In the fastening work of a screw, the robot device 2300 attaches a tool to the hand unit 2310, presses the tool against a screw at a specified position on the part 2313, and rotates the tool. At this time, if the tool is not in an appropriate position relative to the screw and is not pressed against the screw with an appropriate force, the screw may not rotate properly or may not be tightened with the specified torque, which may result in failure to fasten the screw.

[0315] Therefore, the learning model performs machine learning of the force or moment applied to the hand unit 2310 measured by the force sensor 2311 when tightening a screw. At this time, screw tightening is repeatedly performed while changing the position or pressing force of the hand unit 2310, the degree of tool deterioration, etc., and data on successful cases and unsuccessful cases is collected for machine learning.

[0316] When the robotic device 2300 tightens a screw, the learning model determines whether the screw has been tightened successfully. By having the learning model determine whether the operation has succeeded or failed, in the event of failure, the operation can be redone or the part 2313 can be automatically discarded, thereby making the robotic device 2300 autonomous.

[0317] Here, an example of using the learning device 100 and the inference device 800 will be described using two robot devices 2300 as an example. To distinguish between the first and second robot devices 2300, the first robot device 2300 will be referred to as robot device 2300A, and the second robot device 2300 will be referred to as robot device 2300B. Furthermore, the reference numerals of the components included in the first robot device 2300A will have an A suffix, and the reference numerals of the components included in the second robot device 2300B will have a B suffix.

[0318] Consider a case where a learning model that has learned data acquired from a robot device 2300A predicts the outcome of a screw tightening operation based on data (force detection data) measured by a force sensor 2311B of a robot device 2300B. In this case, there are individual differences between the robot devices 2300A and 2300B in the bases 2302A and 2302B, each axis, tool installation errors, hand units 2310A and 2310B, tools, force sensors 2311A and 2311B, etc. For this reason, a difference occurs between the force detection data measured by the force sensor 2311A and the force detection data measured by the force sensor 2311B, degrading the judgment accuracy of the learning model.

[0319] Next, the correspondence between the learning device 100, the robot device 2300A, and the robot device 2300B according to embodiments 1 to 4 will be described. With regard to the spare device 101 and the target device 103 applied to the learning device 100 according to embodiments 1 to 4, the spare device 101 corresponds to the robot device 2300A, and the target device 103 corresponds to the robot device 2300B. Furthermore, the spare sensor unit 102 corresponds to the force sensor 2311A, and the target sensor unit 104 corresponds to the force sensor 2311B.

[0320] The robot device 2300A, which is the standby device 101, repeatedly tightens screws while changing the position or pressing force of the hand unit 2310A, the degree of tool deterioration, and the like, thereby acquiring data from the force sensor 2311A corresponding to various results. In other words, the robot device 2300A reproduces machining states corresponding to various machining results. In this case, the data collected by the force sensor 2311A becomes the preliminary data 203. Then, a label indicating the screw tightening result corresponding to the collected preliminary data 203 becomes the preliminary label 401. Then, the base model learning unit 105 generates a base model 403 that has learned the relationship between the preliminary data 203 and the preliminary label 401.

[0321] Furthermore, the robot device 2300B, which is the target device 103, can acquire data from the force sensor 2311B corresponding to various results by repeatedly tightening screws while changing the position or pressing force of the hand unit 2310B, the degree of tool deterioration, and the like. In other words, the robot device 2300B reproduces machining states corresponding to various machining results. In this case, the data collected by the force sensor 2311B becomes the target data 303. At this time, to prevent the robot device 2300B from breaking down or the part 2313B from becoming unusable, data is acquired within a range that does not cause defects that would cause the robot device 2300B to break down or malfunction, or that would make the part 2313B unusable. Then, a label indicating the screw tightening result corresponding to the collected target data 303 becomes the target label 501. The adaptive model training unit 106 then generates an adapted model 503 that learns the relationship between the target data 303 and the target label 501.

[0322] Below, the process by which the inference error acquisition unit 107 acquires the base model inference error 603 and the adaptive model inference error 606, and the process by which the error model generation unit 108 learns the error model 701 are the same as when the inference device 800 is applied to the servo control device 2100, so their explanation will be omitted.

[0323] The robot device 2300B, which is the target device 103, only needs to acquire data to the extent that it does not cause any defects that could cause a breakdown or malfunction, or any defects that would make the part 2313B unusable, so data can be collected without causing a breakdown in the target device 103 and while preventing any defects that would make the part 2313B unusable.

[0324] Next, the correspondence between the inference device 800 and the robot device 2300B will be described. The inference device 800 is used when machining the robot device 2300B. The light detection value acquired by the force sensor 2311B when machining the robot device 2300B becomes new target data 801.

[0325] Hereinafter, the process by which the inference device 800 calculates the inference result 806 from the new target data 801 is the same as when the inference device 800 is applied to the servo control device 2100, and therefore a description thereof will be omitted.

[0326] In this way, the learning device 100 and the inference device 800 can be applied to the robot device 2300. According to the learning device 100 and the inference device 800, when a learning model that has learned data collected by the robot device 2300A is used by the robot device 2300B, the inference accuracy of the learning model can be improved.

[0327] As described above, according to the fifth embodiment, the learning device 100 and the inference device 800 can be applied to an industrial system, and therefore the industrial system can obtain a highly accurate inference result 806.

[0328] Next, we will explain the hardware configurations of learning device 100 and inference devices 800, 1000, 1400, 1700, and 1900. Since learning device 100 and inference devices 800, 1000, 1400, 1700, and 1900 have similar hardware configurations, we will explain the hardware configuration of inference device 800 according to the first embodiment here.

[0329] FIG. 24 is a diagram illustrating an example of a configuration in which the processing circuit 2400 included in the inference device 800 according to the first embodiment is realized by a processor 2401 and a memory 2402.

[0330] When the processing circuit 2400 is composed of a processor 2401 and a memory 2402, each function of the processing circuit 2400 of the inference device 800 is realized by software, firmware, or a combination of software and firmware. The software, firmware, etc. are written as a program and stored in the memory 2402. In the processing circuit 2400, each function is realized by the processor 2401 reading and executing the program stored in the memory 2402. In other words, the processing circuit 2400 includes the memory 2402 for storing an inference program that results in the processing of the inference device 800 being executed. This inference program can also be said to cause a computer to execute the procedures and methods of the inference device 800.

[0331] In the case of learning device 100, processing circuitry 2400 includes processor 2401 and memory 2402 for storing a learning program that results in the processing of learning device 100. This learning program can also be said to cause a computer to execute the procedures and methods of learning device 100.

[0332] Here, the processor 2401 may be a CPU (Central Processing Unit), a processing device, an arithmetic unit, a microprocessor, a microcomputer, or a DSP (Digital Signal Processor), etc. The memory 2402 may be a non-volatile or volatile semiconductor memory such as a RAM (Random Access Memory), a ROM (Read Only Memory), a flash memory, an EPROM (Erasable Programmable ROM), or an EEPROM (Electrically EPROM). The memory 2402 may also be a magnetic disk, a flexible disk, an optical disk, a compact disk, a minidisk, a DVD (Digital Versatile Disc), etc.

[0333] 25 is a diagram showing an example of the configuration of the processing circuit 2500 included in the inference device 800 according to the first embodiment, when the processing circuit 2500 is configured with dedicated hardware. When the processing circuit 2500 is configured with dedicated hardware, the processing circuit 2500 shown in FIG. 25 is, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or a combination thereof. The inference device 800 may be realized by the processing circuit 2500 for each function, or may be realized by the processing circuit 2500 for a plurality of functions collectively.

[0334] Furthermore, the processing circuitry 2500 may be partially implemented by dedicated hardware and partially implemented by software or firmware. In this manner, the processing circuitry 2500 can implement each of the above-described functions by dedicated hardware, software, firmware, or a combination thereof.

[0335] The location where the processing circuits 2400 and 2500 are installed is not limited to inside an industrial system. For example, the processing circuits 2400 and 2500 may be installed in a location separate from the industrial equipment and connected to the industrial equipment via a network. Furthermore, for example, a control device, a numerical control device, or the like provided in the industrial machine may be realized by the processing circuit.

[0336] The configurations shown in the above embodiments are merely examples, and may be combined with other known technologies, or different embodiments may be combined with each other. It is also possible to omit or modify parts of the configurations as long as they do not deviate from the gist of the invention.

[0337] 100 Learning device, 101 Spare device, 102 Spare sensor unit, 103 Target device, 104 Target sensor unit, 105 Base model learning unit, 106 Adaptive model learning unit, 107 Inference error acquisition unit, 108, 1300, 1500, 1600, 1800 Error model generation unit, 201, 301 Sensor, 202, 302 Feature extraction unit, 203 Spare data, 303 Target data, 401 Spare label, 403 Base model, 501 Target label, 503 Adaptive model, 601, 802 Base model inference unit, 602 Base model error calculation unit, 603 Base model inference error, 604, 803 Adaptive model inference unit, 605 Adaptive model error calculation unit, 606 Adaptive model inference error, 701, 901 Error model, 800, 1000, 1400, 1700, 1900 Inference device, 801 New target data, 804, 1001, 1401, 1701 Error model inference unit, 805, 1402, 1702, 1902 Integration unit, 806 Inference result, 902 Dimension reduction unit, 1301 Base error information setting unit, 1302, 1502, 1602 Base error model learning unit, 1303, 1503, 1603 Base error model, 1304, 1504 Adaptive error information setting unit, 1305, 1505, 1605 Adaptive error model learning unit, 1306, 1506, 1606 Adaptive error model, 1403, 1703 Base error model inference unit, 1404, 1704 Adaptive error model inference unit, 1501 Base error information setting unit, 1507 anomaly detection model learning unit, 1601 base prior information setting unit, 1604 adaptation prior information setting unit, 1801 data extraction unit, 1802 adaptation success data, 1803 adaptation domain learning unit, 1804 adaptation domain model, 1901 adaptation domain model inference unit, 2001 adaptation domain model inference unit, 2002 domain threshold determination unit, 2100, 2100A, 2100B servo control device, 2101, 2301 control unit, 2102 driver, 2103, 2103A, 2103B current sensor, 2104 encoder, 2105, 2105A, 2105B motor, 2106 coupling, 2107, 2107A, 2107B ball screw, 2108, 2108A, 2108B Transported object, 2200, 2200A, 2200B laser cutting processing device, 2201 control device, 2202, 2202A,2202B laser oscillator, 2203 optical path, 2204, 2204A, 2204B machining head, 2205 drive unit, 2206, 2206A, 2206B optical sensor, 2207, 2207A, 2207B machining object, 2300, 2300A, 2300B robot device, 2302, 2302A, 2302B base, 2310, 2310A, 2310B hand unit, 2311, 2311A, 2311B force sensor, 2312 work table, 2313, 2313B parts, 2400, 2500 processing circuit, 2401 processor, 2402 memory, Aa adaptation region model inference result, Ar adaptation model inference result, Br base model inference result, E1, E2 Error model inference results, Ea1, Ea2: Adaptive error model inference results, Eb1, Eb2: Base error model inference results, Sda, Sdb: Standard deviation.

Claims

1. An adaptive model learning unit that generates an adaptive model that has learned the relationship between target data, which is data measured by a target device that is the target of learning and inference when the target device is operating, and a first target label, which is the label of the target data; an inference error acquisition unit that uses a base model that has learned the relationship between preliminary data, which is data measured by a spare device that is a different device of the same type as the target device and the target of learning when the spare device is operating, and a spare label, which is the label of the preliminary data, to infer a second target label corresponding to the target data, and calculates the error between the first target label and the second target label as a base model inference error, and uses the adaptive model to infer a third target label corresponding to the target data, and calculates the error between the first target label and the third target label as an adaptive model inference error; an error model generation unit that generates an error model used in integrating a base model inference result, which is the output when new target data, which is data newly measured by the target device when the target device is newly operated, is input to the base model, and an adaptive model inference result, which is the output when the new target data is input to the adaptive model, by learning based on the target data, the base model inference error, and the adaptive model inference error; wherein the error model learns the relationship between the magnitude of the base model inference error, the magnitude of the adaptive model inference error, and the target data, and when the new target data is input, outputs an error model inference result, which is data corresponding to the weighting of the base model inference result and the adaptive model inference result during the integration.

2. The learning device according to claim 1, further comprising a base model learning unit that generates the base model based on the preliminary data and the preliminary labels.

3. The learning device according to claim 1 or 2, further comprising a dimension reduction unit that reduces the dimensions of the target data, and the error model generation unit generates the error model by learning the relationship between the dimension-reduced target data, the base model inference error, and the adaptive model inference error.

4. The error model generation unit comprises: a base error model learning unit that generates a base error model that has learned the relationship between the magnitude of the base model inference error and the target data; a base error information setting unit that sets base error information indicating that the magnitude of the base model inference error is smaller than the magnitude of the adaptive model inference error in the base error model learning unit before the base error model is learned; an adaptive error model learning unit that generates an adaptive error model that has learned the relationship between the magnitude of the adaptive model inference error and the target data; and an adaptive error information setting unit that sets adaptive error information indicating that the magnitude of the adaptive model inference error is larger than the magnitude of the base model inference error in the adaptive error model learning unit before the adaptive error model is learned; and the learning device described in any one of claims 1 to 3, characterized in that the base error model learning unit generates the base error model based on the base error information, and the adaptive error model learning unit generates the adaptive error model based on the adaptive error information.

5. The learning device described in any one of claims 1 to 3, characterized in that the error model generation unit comprises: a base error model learning unit that generates a base error model that has learned the relationship between the magnitude of the base model inference error and the target data; a base prior information setting unit that sets, in the base error model learning unit, a prior distribution indicating that the magnitude of the base model inference error is smaller than the magnitude of the adaptive model inference error as a base prior distribution before the base error model is learned; an adaptive error model learning unit that generates an adaptive error model that has learned the relationship between the magnitude of the adaptive model inference error and the target data; and an adaptive prior information setting unit that sets, in the adaptive error model learning unit, a prior distribution indicating that the magnitude of the adaptive model inference error is larger than the magnitude of the base model inference error as an adaptive prior distribution before the adaptive error model is learned; wherein the base error model and the adaptive error model are Gaussian process regression models; the base error model learning unit generates the base error model based on the base prior distribution; and the adaptive error model learning unit generates the adaptive error model based on the adaptive prior distribution.

6. The learning device described in any one of claims 1 to 3, characterized in that the error model generation unit has: a data extraction unit that compares the base model inference error with the adaptive model inference error, and extracts, from the target data, target data corresponding to data in which the magnitude of the adaptive model inference error is smaller than the magnitude of the base model inference error, as successful adaptation data; and an adaptive domain learning unit that generates, based on the successful adaptation data, an adaptive domain model that learns the data space in which the successful adaptation data is distributed, wherein, when new target data is input, the adaptive domain model is a model that outputs an adaptive domain model inference result indicating whether the new target data is within an adaptive domain, which is a domain in which the magnitude of the adaptive model inference error is smaller than the magnitude of the base model inference error.

7. The learning device described in claim 6, characterized in that the error model generation unit further has: an adaptation domain model inference unit that inputs the adaptation success data into the adaptation domain model to calculate a domain score that indicates, in numerical form, whether the adaptation success data belongs to the adaptation domain; and a domain threshold determination unit that determines a value smaller than the median of the domain score as a domain threshold that corresponds to the boundary between the adaptation success data and adaptation failure data, which is data other than the adaptation success data.

8. The learning device according to any one of claims 1 to 7, wherein the amount of the preliminary data is greater than the amount of the target data.

9. An industrial system comprising: a learning device according to any one of claims 1 to 8; and the target device.

10. An adaptive model inference unit that calculates an adaptive model inference result that is an inference result by the adaptive model by inputting new target data, which is data newly measured by the target device when the target device newly operates, into an adaptive model that has learned the relationship between target data, which is data measured by the target device that is the target of learning and inference when the target device operates, and a first target label, which is the label of the target data; and a base model inference unit that calculates a base model inference result that is an inference result by the base model by inputting the new target data into a base model that has learned the relationship between spare data, which is data measured by a spare device that is a different device of the same type as the target device and the target of learning when the spare label is the label of the spare data. an error model inference unit that calculates an error model inference result that is an inference result by the error model by inputting new target data into an error model that has learned the magnitude of a base model inference error, which is the error between the first target label and the second target label when a second target label corresponding to the target data is inferred using the base model, and the magnitude of an adaptive model inference error, which is the error between the first target label and the third target label when a third target label corresponding to the target data is inferred using the adaptive model; and an integration unit that calculates an inference result by weighting the base model inference result and the adaptive model inference result based on the error model inference result, wherein the error model inference unit inputs the new target data into the error model and calculates the error model inference result that corresponds to a predicted value of an error in the base model inference result and a predicted value of an error in the adaptive model inference result, and outputs the error model inference result to the integration unit.

11. An inference device as described in claim 10, characterized in that it has a dimension reduction unit that reduces the dimensions of the new target data, the error model is a model that has learned the relationship between the dimension-reduced target data, the base model inference error, and the adaptive model inference error, and the error model inference unit calculates the error model inference result by inputting the dimension-reduced new target data into the error model.

12. The error model includes a base error model that has learned the relationship between the magnitude of the base model inference error and the target data, and an adaptive error model that has learned the relationship between the magnitude of the adaptive model inference error and the target data, and the error model inference unit includes a base error model inference unit that calculates a base error model inference result that is an inference result by the base error model by inputting the new target data into the base error model, and an adaptive error model inference unit that calculates an adaptive error model inference result that is an inference result by the adaptive error model by inputting the new target data into the adaptive error model, and the base error model inference unit calculates the base error model inference result corresponding to a predicted value of the error of the base model inference result by inputting the new target data into the base error model, and outputs it to the integrating unit, and the adaptive error model inference unit calculates the adaptive error model inference result corresponding to a predicted value of the error of the adaptive model inference result by inputting the new target data into the adaptive error model, and outputs it to the integrating unit, and the integrating unit 12. The inference device according to claim 10, wherein the inference device calculates an integrated inference result by weighting the base model inference result and the adaptive model inference result based on the base error model inference result and the adaptive error model inference result.

13. The inference device described in claim 12, characterized in that: the base error model is a model generated based on a prior distribution indicating that the magnitude of the base model inference error is smaller than the magnitude of the adaptive model inference error; the adaptive error model is a model generated based on a prior distribution indicating that the magnitude of the adaptive model inference error is larger than the magnitude of the base model inference error; the base error model inference unit calculates a first standard deviation, which is the standard deviation of the base error model inference result, by inputting the new target data into the base error model, and outputs the first standard deviation together with the base error model inference result to the integrating unit; the adaptive error model inference unit calculates a second standard deviation, which is the standard deviation of the adaptive error model inference result, by inputting the new target data into the adaptive error model, and outputs the second standard deviation together with the adaptive error model inference result to the integrating unit; and the integrating unit calculates a weighted average of the adaptive model inference result and the base model inference result using the sum of the base error model inference result and the first standard deviation and the sum of the adaptive error model inference result and the second standard deviation.

14. The inference device described in claim 10 or 11, characterized in that: the error model is an adaptive domain model that learns a data space in which the adaptation success data is distributed based on adaptation success data, which is target data corresponding to data among the target data where the magnitude of the adaptation model inference error is smaller than the magnitude of the base model inference error; the error model inference unit is an adaptive domain model inference unit that, by inputting the new target data into the adaptive domain model, calculates an adaptation domain model inference result indicating whether the new target data is within an adaptation domain where the magnitude of the adaptation model inference error is smaller than the magnitude of the base model inference error, and outputs the result to the integrating unit; and the integrating unit calculates an integrated inference result by weighting the base model inference result and the adaptation model inference result based on the adaptation domain model inference result.

15. The inference device described in claim 14, characterized in that the integration unit calculates an integrated inference result by weighting the base model inference result and the adaptation model inference result based on a domain threshold set as a boundary between the adaptation success data and adaptation failure data, which is data other than the adaptation success data, and a value smaller than the median of the domain score, which indicates numerically whether the adaptation success data belongs to the adaptation domain.

16. The inference device described in claim 15, characterized in that the integration unit adjusts the domain threshold based on a threshold expansion amount used when expanding the domain threshold, and calculates an integrated inference result by weighting the base model inference result and the adaptive model inference result based on the adjusted domain threshold and the adaptive domain model inference result.

17. An industrial system comprising: an inference device according to any one of claims 10 to 16; and the target device.

18. An adaptive model learning step in which a learning device generates an adaptive model that has learned the relationship between target data, which is data measured by a target device that is the target of learning and inference when the target device is operating, and a first target label, which is the label of the target data; an inference error acquisition step in which the learning device infers a second target label corresponding to the target data using a base model that has learned the relationship between preliminary data, which is data measured by a spare device that is a different device of the same type as the target device and is the target of learning when the spare device is operating, and a preliminary label, which is the label of the preliminary data, and calculates the error between the first target label and the second target label as a base model inference error, and infers a third target label corresponding to the target data using the adaptive model and calculates the error between the first target label and the third target label as an adaptive model inference error; an error model generation step in which the learning device generates an error model used for integrating a base model inference result, which is the output when new target data, which is data newly measured by the target device when the target device is newly operated, is input to the base model, and an adaptive model inference result, which is the output when the new target data is input to the adaptive model, by learning based on the target data, the base model inference error, and the adaptive model inference error; wherein the error model learns the relationship between the magnitude of the base model inference error, the magnitude of the adaptive model inference error, and the target data, and when the new target data is input, outputs an error model inference result, which is data corresponding to the weighting of the base model inference result and the adaptive model inference result during the integration.

19. An adaptive model inference step in which an inference device calculates an adaptive model inference result that is an inference result by the adaptive model by inputting new target data, which is data newly measured by a target device when the target device newly operates, into an adaptive model that has learned the relationship between target data, which is data measured by a target device that is the target of learning and inference when the target device operates, and a first target label, which is the label of the target data; and a base model inference step in which the inference device calculates a base model inference result that is an inference result by the base model by inputting the new target data into a base model that has learned the relationship between spare data, which is data measured by a spare device that is a device of the same type as the target device and the target of learning when the spare device operates, and a spare label, which is the label of the spare data. an error model inference step in which the inference device calculates an error model inference result, which is an inference result obtained by the error model, by inputting the new target data into an error model that has learned the magnitude of a base model inference error, which is the error between the first target label and the second target label when a second target label corresponding to the target data is inferred using the base model, and the magnitude of an adaptive model inference error, which is the error between the first target label and the third target label when a third target label corresponding to the target data is inferred using the adaptive model; and an integration step in which the inference device calculates an inference result by weighting the base model inference result and the adaptive model inference result based on the error model inference result, wherein in the error model inference step, the inference device inputs the new target data into the error model and calculates the error model inference result corresponding to a predicted value of the error of the base model inference result and a predicted value of the error of the adaptive model inference result.

Images