Uncertainty learning facility, uncertainty learning program and uncertainty learning system

Abstract

Included are: a training data acquisition unit (51) that acquires training data based on operational data obtained from a machine unit (1); a noise addition unit (52) that generates noise-added training data where noise is added to the training data acquired by the training data acquisition unit (51); an outlier detection unit (53) that calculates an outlier score from the training data acquired by the training data acquisition unit (51) and the noise-added training data generated by the noise addition unit (52); and a model learning unit (54) that calculates a weighted loss function based on the outlier score calculated by the outlier detection unit (53) and trains a machine learning model based on the training data and the noise-added training data.

Description

TECHNICAL AREA

[0001] The present disclosure relates to an uncertainty learning device, an uncertainty learning program and an uncertainty learning system that create a learned model (hereinafter referred to as the “machine learning model”) that outputs a predictive value and uncertainty for input data. STATE OF THE ART

[0002] In recent years, a technique for solving various tasks using a machine learning model in different processes has become known.

[0003] When considering the application of a machine learning model to a machine device, such as an automated guided vehicle, a device used in life-threatening situations like medical applications, or a factory automation (FA) device, the machine learning model will output some data even if no data is input. It is conceivable that the machine learning model might sometimes produce an output that would be unexpected to a human, i.e., an output with low reliability. If such a low-reliability output from the machine learning model is used to control a machine device, there is a possibility that an unexpected situation could arise.

[0004] Therefore, a technique is usually known for deriving or inferring how much uncertainty is inherent in the output of a machine learning model itself. Here, the uncertainty is the reliability.

[0005] Nonpatent literature 1, for example, discloses a technique that expresses how much uncertainty is contained in an output of a machine learning model itself by the size of a width, where the width of the uncertainty output by the machine learning model for unlearned data (extrapolation data) increases by adding noise to training data of the machine learning model. REFERENCE LIST NON-PATENT LITERATURE

[0006] Non-patent literature 1: Danijar Hafner, Noise Contrastive Priors for Functional Uncertainty, July 1, 2019 SUMMARY OF THE INVENTIONAL PROBLEM

[0007] With a conventional technique that outputs uncertainty, as described in the non-patent literature 1, the problem is that, since a machine learning model is trained by adding noise to all training data, even when input data is close to the training data, and a good predictive value is output at the time of inference using the machine learning model, the output uncertainty may increase along with the predictive value.

[0008] The present disclosure was made to solve the problem, and one goal of the present disclosure is to provide a machine learning model that is able to infer a predictive value and uncertainty, and that is able to infer uncertainty more accurately, compared with a machine learning model created by a conventional method of learning a machine learning model by adding noise to all training data. TECHNICAL SOLUTION

[0009] An uncertainty learning device according to the present disclosure is an uncertainty learning device that creates a machine learning model which receives as input data based on operational data relating to an operational result of a machine device and outputs a predicted value corresponding to the operational data and the uncertainty of the predicted value, wherein the uncertainty learning device comprises: a training data acquisition unit that acquires training data that is created based on the operational data obtained from the machine device; a noise addition unit that creates training data with added noise, in which noise is added to the training data acquired by the training data acquisition unit;an outlier detection unit that calculates an outlier score from the training data procured by the training data acquisition unit and the training data with added noise created by the noise addition unit; and a model learning unit that computes a weighted loss function based on the outlier score calculated by the outlier detection unit and trains the machine learning model based on the training data and the training data with added noise. ADVANTAGEOUS EFFECTS OF THE INVENTION

[0010] According to the present disclosure, it is possible with the above configuration to provide a machine learning model that is able to infer uncertainty more accurately than a machine learning model created by a conventional method of training a machine learning model by adding noise to all training data. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1A, Fig. 1B and Fig. 1C are diagrams describing types of uncertainty inferred by a machine learning model, and an example of an inference of ideal uncertainty performed by the machine learning model, where Fig. 1A is a diagram used to describe the uncertainty of a problem, Fig. 1B is a diagram to describe the uncertainty of learning and Fig. 1C is a diagram for describing ideal uncertainty. Fig. Figure 2 is a diagram for a more detailed description of a problem of a conventional technique that is to be solved by an uncertainty learning device according to a first embodiment. Fig. Figure 3 is a representation showing a configuration example of an uncertainty learning system including the uncertainty learning device according to the first embodiment. Fig. Figure 4 is a flowchart describing a learning time operation of the uncertainty learning device according to the first embodiment. Fig. 5 is a flowchart describing details of the processing of step ST3 in Fig. 4. Fig. Figure 6 is a flowchart describing an inference-time operation of the uncertainty learning device according to the first embodiment. Fig. 7A and Fig. Figures 7B are illustrations showing an example of a hardware configuration of the uncertainty learning device according to the first embodiment. Fig. Figure 8 is a representation showing a configuration example of an uncertainty learning system including an uncertainty learning device according to a second embodiment. Fig. Figure 9 is a flowchart describing a learning time operation of the uncertainty learning device according to the second embodiment. Fig. Figure 10 is a flowchart describing an inference-time operation of the uncertainty learning device according to the second embodiment. Fig. Figure 11 is a representation showing a configuration example of an uncertainty learning system including an uncertainty learning device according to a third embodiment. Fig. 12A, Fig. 12B and Fig. 12C are representations to describe an example of transfer learning, where Fig. 12A is a representation that shows an example of fine-tuning, Fig. 12B is a representation that shows an example of feature extraction, and Fig. 12C is a representation that shows an example of joint training. Fig. Figure 13 is a flowchart describing a learning time operation of the uncertainty learning device according to the third embodiment. Fig. 14 is a flowchart describing details of the processing of step ST3a in Fig. 13. Fig. Figure 15 is a representation showing a configuration example of an uncertainty learning system including an uncertainty learning device according to a fourth embodiment. Fig. Figure 16 is a flowchart describing the operation of the uncertainty learning device according to the fourth embodiment. DESCRIPTION OF THE EXECUTION FORMS

[0011] The following describes embodiments for realizing the present disclosure with reference to the accompanying drawings. First embodiment.

[0012] An uncertainty learning device according to a first embodiment creates a learned model (hereinafter referred to as the “machine learning model”).

[0013] Based on input data, the machine learning model outputs a predictive value (referred to below as the "predictive value") and an uncertainty. The uncertainty is an index indicating the degree of confidence in the prediction, i.e., how much confidence the machine learning model has in the predictive value. High uncertainty means the machine learning model has no confidence in the predictive value; in other words, the predictive value is unreliable. Conversely, low uncertainty suggests the machine learning model has high confidence in the predictive value; in other words, the predictive value is highly reliable. It is important for a user or other party to be aware of the uncertainty of the predictive value output by the machine learning model in order to process the information appropriately.By using information about an uncertainty, the user or similar can evaluate the risk of the prediction made by the machine learning model and make a more careful decision.

[0014] Fig. Figure 1 is a diagram describing types of uncertainty inferred by the machine learning model, and an example of an inference of ideal uncertainty performed by the machine learning model.

[0015] Examples of uncertainty in the predicted value include uncertainty due to fluctuations in data (hereinafter referred to as "interpolation data") present in the training data and caused by measurement noise or disturbances (see Fig. 1A), and the uncertainty in data (hereinafter referred to as “extrapolation data”) that is not present in the training data (see Fig. 1B).

[0016] It is necessary to construct a machine learning model that is capable of accurately outputting both uncertainty due to fluctuations in the interpolation data, i.e., uncertainty in a problem, and uncertainty in the extrapolation data, i.e., uncertainty in the learning process (see Fig. 1C).

[0017] Therefore, for example, in the conventional technique, as described in the non-patent literature 1 above, noise is added to the training data and learning is carried out in such a way that uncertainty increases, in other words, the reliability of the training data to which the noise has been added decreases.

[0018] However, since the machine learning model in conventional techniques is trained by adding noise, the machine learning model can output an uncertainty of the learning with some accuracy, but the problem is that an uncertainty of a problem cannot be output exactly.

[0019] Fig. Figure 2 is a diagram to describe in more detail a problem of conventional engineering that is to be solved by the uncertainty learning device according to the first embodiment.

[0020] For example, if the machine learning model is trained to output uncertainty by expressing fluctuations in the interpolation data as variances using a KL divergence method, the uncertainty of a problem can be output with a degree of accuracy. However, when training using this method, it is difficult to adequately infer uncertainty from the extrapolation data (see the left diagram in [reference]). Fig. 2) The reason for this is that the extrapolation data lies outside the range of the training data, and thus a large deviation from the true distribution can be generated. Since the machine learning model is trained on the basis of the training data, the problem is that it is difficult to infer an accurate uncertainty for data (extrapolation data) outside the range of the training data.

[0021] On the other hand, in the conventional technique, as described in the non-patent literature 1 above, learning is carried out by adding noise to the training data in such a way that the uncertainty for the training data to which the noise was added increases (see middle diagram in Fig. 2) Therefore, the conventional technique can solve the problem that it is difficult to infer an accurate uncertainty for the extrapolation data, as described above.

[0022] However, in the conventional technique, a machine learning model is trained by adding noise to all parts of the training data. Therefore, even if the input data closely matches the training data, and a good prediction is generated at the time of inference using the machine learning model, there is a possibility that the uncertainty, which is supposed to be output along with the prediction, will increase. In other words, with the conventional technique, there is a possibility that the uncertainty of a problem cannot be accurately represented.

[0023] The uncertainty learning device according to the first embodiment solves the problem of the conventional technique and provides a machine learning model that is able to accurately infer uncertainty for both a problem uncertainty and a learning uncertainty, i.e., for both an uncertainty of interpolation data and an uncertainty of extrapolation data.

[0024] In particular, the uncertainty learning device according to the first embodiment trains the machine learning model to infer uncertainty by using a result of inferring the deviation of the training data with added noise by an outlier detection mechanism (see the right-hand diagram in Fig. 2) Details of the configuration and operation of the uncertainty learning facility will be described later.

[0025] Fig. Figure 3 is a representation showing a configuration example of an uncertainty learning system 100 including an uncertainty learning device 2 according to the first embodiment.

[0026] The uncertainty learning system 100 comprises the uncertainty learning device 2 and a machine device 1. The uncertainty learning device 2 and the machine device 1 are connected to each other via a network.

[0027] In the first embodiment, it is assumed that the machine setup 1 is, for example, a factory automation system (FA system). The FA system comprises, for example, a servo motor, a CNC machine, a processing machine such as a sheet metal laser processing machine or an electrical discharge machining (EDM) processing machine, and a robot.

[0028] The Uncertainty Learning Facility 2 is, for example, contained in a server.

[0029] In the first embodiment, the uncertainty learning device 2 performs the "learning" of creating a machine learning model and the "inference" of inferring a predicted value and an uncertainty using the machine learning model created by the "learning". Details of the "learning" processing and details of the "inference" processing are described later.

[0030] The uncertainty learning facility 2 comprises a procurement unit 3, a preprocessing unit 4, a learning unit 5 and an inference unit 6.

[0031] Procurement unit 3 comprises an operating result procurement unit 31 and an operating result storage unit 32.

[0032] The preprocessing unit 4 comprises a preprocessing execution unit 41, a training data storage unit 42 and a test data storage unit 43.

[0033] The learning unit 5 comprises a training data acquisition unit 51, a noise addition unit 52, an outlier detection unit 53, a model learning unit 54 and a model storage unit 55.

[0034] The inference unit 6 comprises a model reading unit 61, a prediction unit 62, a test data acquisition unit 63 and an evaluation unit 64.

[0035] At the time of learning, the following components are operating in the uncertainty learning facility 2, among those mentioned above: the operating result procurement unit 31, the preprocessing execution unit 41, the training data procurement unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the test data procurement unit 63 and the evaluation unit 64.

[0036] At the time of inference, the following components are operating in the uncertainty learning facility 2, among those mentioned above: the operating result procurement unit 31, the preprocessing execution unit 41, the model reading unit 61 and the prediction unit 62.

[0037] The operating result procurement unit 31 of procurement unit 3 procures data (hereinafter referred to as "operational data") on an operating result of machine facility 1 from machine facility 1.

[0038] The operating result includes, for example, a parameter set in machine unit 1, a configuration of machine unit 1, its operating mode, data that specifies a value unique to machine unit 1, and log data operated by machine unit 1. It should be noted that the operating result procurement unit 31 obtains the operational data from machine unit 1, for example, via an encoder (not shown).

[0039] The operating result procurement unit 31 obtains the operational data relating to an operating result of an operational unit from the operational unit itself. For example, if the operational unit is a processing machine with computer numerical control (CNC), the operational data includes a command position, a command speed, a command acceleration, a feedback speed, a feedback acceleration, and a current value. Operational data can also include, for example, a value obtained by measuring a deviation of a processing position.

[0040] In the following first embodiment, the machine device 1 refers to an FA device.

[0041] The operating result procurement unit 31 procures, as operational data, for example, directly the values ​​measured by various sensors arranged in the machine setup 1. Examples of the various sensors are a speed sensor, an acceleration sensor, a gyroscope, a temperature sensor, and a humidity sensor. In addition, the operating result procurement unit 31 can, for example, procure as operational data a value that is calculated based on values ​​measured by various sensors arranged in the machine setup 1.

[0042] The operating result procurement unit 31 instructs the operating result storage unit 32 to store the procured business-related data. For example, the operating result procurement unit 31 adds a procurement date and procurement time to the business-related data and instructs the operating result storage unit 32 to store the business-related data.

[0043] The operating result storage unit 32 stores the business-related data that is procured by the operating result procurement unit 31.

[0044] It should be noted that here the operating result storage unit 32 is contained within the uncertainty learning device 2, but this is only an example. The operating result storage unit 32 can be located at a location that the uncertainty learning device 2 can access from outside the uncertainty learning device 2.

[0045] The preprocessing execution unit 41 of the preprocessing unit 4 obtains the business-related data procured by the operating result procurement unit 31 from the operating result storage unit 32 and performs preprocessing of the shaping of various data on the basis of the business-related data.

[0046] In the first embodiment, the preprocessing performed by the preprocessing execution unit 41 includes preprocessing when the uncertainty learning device 2 creates a machine learning model, and preprocessing when the uncertainty learning device 2 performs inference to obtain a predicted value and an uncertainty using the created machine learning model.

[0047] In the first embodiment, the preprocessing performed by the preprocessing execution unit 41 when the uncertainty learning unit 2 creates a machine learning model is also referred to as "learning-time preprocessing," and the preprocessing performed by the preprocessing execution unit 41 when the uncertainty learning unit 2 performs inference is also referred to as "inference-time preprocessing." It should be noted that in the uncertainty learning unit 2, the machine learning model is created by the learning unit 5 and the evaluation unit 64 of the inference unit 6, and the prediction unit 62 of the inference unit 6 performs the inference using the machine learning model. Details of the learning unit 5 and the inference unit 6 are described later.

[0048] The following describes the learning-time preprocessing and the inference-time preprocessing performed by the preprocessing execution unit 41.

[0049] First, the learning time preprocessing performed by the preprocessing execution unit 41 is described.

[0050] The preprocessing execution unit 41 obtains the business-related data procured by the business result procurement unit 31 from the business result storage unit 32 and performs preprocessing of the shaping of the training data used by the learning unit 5 and the test data used by the evaluation unit 64 of the inference unit 6, i.e. the learning time preprocessing, on the basis of the business-related data.

[0051] In the first embodiment, the learning time preprocessing performed by the preprocessing execution unit 41 includes, for example, processing the partitioning of the data contained in the operational data into an explanatory variable and a target variable, processing the transformation of the data contained in the operational data into a feature required for learning a machine learning model, and processing the partitioning of the operational data into training data and test data.

[0052] Here, the explanatory variable and the target variable are described.

[0053] The type of data used as explanatory variables and target variables is determined in advance by a user or similar entity.

[0054] At least one piece of data other than the target variable is used as the explanatory variable.

[0055] For example, if a deviation in a processing position is used as the target variable, operational data specifying the processing position or speed can be the explanatory variable.

[0056] The target variable is at least one piece of data that is to be predicted using a machine learning model. A user or similar entity can modify the target variable depending on the problem; for a regression problem, a numerical value is used as the target variable, while for a classification problem, a class to be classified is used as the target variable.

[0057] For example, if a deviation of a processing position due to friction or the like in machine equipment 1 is to be predicted, the user or the like only needs to use a value obtained by measuring the deviation of the processing position as the target variable.

[0058] Next, the transformation of the feature in the preprocessing execution unit 41 in the learn-time-learn-time preprocessing is described using an example.

[0059] Preprocessing Execution Unit 41 can transform the explanatory variable to create a learning-effective feature. For example, Preprocessing Execution Unit 41 can transform the explanatory variable by applying standardization to transform a variable scale, a linear min-max scaling transformation, or a nonlinear transformation such as Box-Cox. Alternatively, it can transform the explanatory variable by applying one-hot encoding or similar techniques if part of the data is a category variable.

[0060] Next, the division of the training data and the test data in the learning time preprocessing performed by the preprocessing execution unit 41 will be described using an example.

[0061] When learning is performed, the data contained in the operational data must be divided into training data for learning and test data for evaluating the performance of the trained machine learning model.

[0062] For example, the preprocessing execution unit 41 can randomly split the data contained in the operational data into training data and test data, or it can create training data by matching the number of pieces of data for each class as in a classification problem, and split the remaining data into test data.

[0063] It should be noted that if the operational data is time series data, if new data is learned in one direction of time, correct data will be learned for old data and data loss will occur, and therefore the preprocessing execution unit 41 must separate the training data and the test data taking a time axis into account.

[0064] The preprocessing execution unit 41 instructs the training data storage unit 42 to store the training data generated by the learning time preprocessing. The preprocessing execution unit 41 also instructs the test data storage unit 43 to store the test data generated by the learning time preprocessing.

[0065] Next, the inference-time preprocessing performed by the preprocessing execution unit 41 is described.

[0066] The preprocessing execution unit 41 obtains the operational data procured by the operational result procurement unit 31 from the operational result storage unit 32 and performs preprocessing of the shaping data used by the prediction unit 62 of the inference unit 6 (hereinafter referred to as "model input data"), i.e., inference time preprocessing, based on the operational data.

[0067] In the first embodiment, the inference-time preprocessing performed by the preprocessing execution unit 41 includes, for example, processing the use of data that is an explanatory variable among the data pieces contained in the operational data as the model input data, and processing the transformation of data contained in the operational data into a feature that is required as the model input data of a machine learning model.

[0068] The preprocessing execution unit 41 only needs to perform a transformation into a feature that is required as the model input data of the machine learning model, using the same procedure as the feature transformation procedure for transforming an explanatory variable to create a feature that is effective for learning in the learning-time preprocessing.

[0069] It should be noted that when inferring using the machine learning model, the data does not need to be divided into training data and test data.

[0070] The preprocessing execution unit 41 outputs the model input data, which is created by performing the inference time preprocessing, to the prediction unit 62.

[0071] The training data storage unit 42 of the preprocessing unit 4 stores the training data.

[0072] It should be noted that here the training data storage unit 42 is contained within the uncertainty learning device 2, but this is only an example. The training data storage unit 42 can be located in a location that the uncertainty learning device 2 can access from outside the uncertainty learning device 2.

[0073] The test data storage unit 43 of the preprocessing unit 4 stores the test data.

[0074] It should be noted that here the test data storage unit 43 is contained within the uncertainty learning device 2, but this is only an example. The test data storage unit 43 can be located in a location that the uncertainty learning device 2 can access from outside the uncertainty learning device 2.

[0075] Learning Unit 5 creates a machine learning model based on the training data stored in training data storage unit 42. One or more machine learning methods, such as a neural network, are used as the machine learning model. The neural network can be of any type, such as a hierarchical neural network, a convolutional neural network, or a recursive neural network. The neural network only needs to output at least two results to provide a prediction value and an uncertainty.

[0076] It should be noted that if the number of dimensions of input data for the machine learning model, such as the number of parameters, increases, the learning unit 5 will adjust the dimensions of the input data. Using various dimensionality reduction methods, the resulting data can be used as input for the machine learning model. Examples of these methods include principal component analysis, singular value analysis, tensor analysis, and autoencording. Since these dimensionality reduction methods are well-known, a detailed description is omitted.

[0077] The training data acquisition unit 51 of the learning unit 5 obtains the training data from the training data storage unit 42.

[0078] For the training data, N parts of the training data are stored in the training data storage unit 42. In this case, if an explanatory variable is represented by xn and a target variable by yn for the N parts of the training data, the training data can be represented by D N= {{x1,y1},...,{x n ,y n}} can be expressed. Learning unit 5 trains the machine learning model on the training data D. N It should be noted that the machine learning model is trained by model learning unit 54 of learning unit 5. Details of model learning unit 54 are described later.

[0079] The training data acquisition unit 51 outputs the acquired training data to the noise addition unit 52, the outlier detection unit 53 and the model learning unit 54.

[0080] The noise addition unit 52 of the learning unit 5 creates data (hereinafter referred to as "noise-added training data") in which noise is added to the training data procured by the training data procurement unit 51.

[0081] For example, the noise addition unit 52 creates the training data with added noise by adding a value to the explanatory variable and the target variable of the training data, which is created from a Gaussian distribution, a uniform distribution, a gamma distribution or a β distribution.

[0082] For example, in a case where the noise addition unit 52 creates the training data with added noise by adding a value from a Gaussian distribution, where an explanatory variable is represented by x and a target variable by y, An explanatory variable x̂ and a target variable ŷ are calculated after noise addition as follows. x^=x+N(μx,σy) y^=N(μy,σy)

[0083] At this point in time, N(·): Gaussian distribution, µ x , σ x: Average value and standard deviation of noise added to an explanatory variable, µ y , σ y : Average value and standard deviation of the noise added to a target variable.

[0084] It should be noted that µ y µ y = can fulfill y.

[0085] For example, if a fluctuation of the explanatory variable or the target variable is known from the characteristics of the machine device 1, the noise addition unit 52 only needs to add noise that simulates the fluctuation.

[0086] The noise addition unit 52 outputs the training data with added noise to the outlier detection unit 53 and the model learning unit 54.

[0087] It should be noted that the noise addition unit 52 can create the training data with added noise by adding noise to all parts of the training data obtained by the training data acquisition unit 51 from the training data storage unit 42, or it can create the training data with added noise by adding noise to some parts (mini-batch) of the training data obtained by the training data acquisition unit 51 from the training data storage unit 42.

[0088] The outlier detection unit 53 of the learning unit 5 calculates an outlier score from the training data procured by the training data acquisition unit 51 and the training data created by the noise addition unit 52 with added noise.

[0089] The calculation of an outlier score by the outlier detection unit 53 is described in detail.

[0090] First, the outlier detection unit 53 inputs the training data procured by the training data acquisition unit 51 into the outlier detection mechanism and learns the training data.

[0091] It should be noted that the outlier detection unit 53, for example, feeds all parts of training data procured by the training data acquisition unit 51 into the outlier detection mechanism and learns all parts of the training data.

[0092] The outlier detection unit 53 only needs to learn the training data using a known outlier detection mechanism such as Hotelling theory, the k-nearest neighbor algorithm, the local outlier factor (LOF), or the one-class support vector machine. The outlier detection unit 53 can use one or more of these known outlier detection mechanisms.

[0093] If the number of dimensions to be input is large, the outlier detection unit 53 can also reduce the dimensions of the input data using the known dimension reduction method described above and can use data with the reduced dimensions as input for the outlier detection mechanism.

[0094] The input data to be fed into the outlier detection mechanism by the outlier detection unit 53 may, for example, be all or some of the explanatory variables contained in the training data, or data obtained by adding the target variable to all or some of the explanatory variables contained in the training data.

[0095] After learning the outlier detection mechanism using all parts of the training data, the outlier detection unit 53 next feeds the training data created by the noise addition unit 52 with added noise into the learned outlier detection mechanism and calculates an outlier score.

[0096] Subsequently, the outlier detection unit 53 sets a weight γ of the loss function NCP-Loss based on the calculated outlier score. In the first embodiment, a second term in a loss function used in learning the machine learning model is referred to as "NCP-Loss". Details of the loss function are described later.

[0097] The outlier detection unit 53 sets the weighting y of NCP loss, for example, according to the size of the outlier score calculated by the outlier detection mechanism.

[0098] The outlier detection unit 53 sets the weight γ such that the weight γ is large for the training data with added noise that deviate from the training data, and sets the weight y such that the weight y is small for the training data with added noise that is close to the training data. For example, if the outlier score is calculated by the well-known LOF, which is an outlier detection mechanism where the probability of it being an outlier is greater when there is a larger difference between a local density of the training data with added noise and a local density of a nearby point, this means that the training data with added noise deviates from the training data, since the outlier score is larger.In this case, for example, the outlier detection unit 53 sets the weight γ to a value that is normalized such that a value obtained from the inverse of a positive and a negative outlier score falls within a range of [0,1].

[0099] Furthermore, the outlier detection unit 53 can, for example, set a threshold (hereafter referred to as the "outlier detection threshold") and adjust the weight γ by comparing the outlier score to the outlier detection threshold. For example, the outlier detection unit 53 can set the weight y = 1 on training data with added noise that differs from the training data where the outlier score is equal to or greater than the outlier detection threshold, and can set the weight y = 0 on training data with added noise that is close to the training data where the outlier score is less than the outlier detection threshold.

[0100] It should be noted that the example described above is only an example. If the training data with added noise differs from the training data, whether the outlier score is equal to or greater than the outlier detection threshold, or less than the outlier detection threshold, depends on the outlier detection mechanism. If the weight γ is set by comparing the outlier score to the outlier detection threshold, the outlier detection unit 53 only needs to compare the outlier score to the outlier detection threshold to set the weight y = 1 for training data with added noise that differs from the training data, and the weight y = 0 for training data with added noise that is close to the training data.

[0101] The outlier detection unit 53 outputs the set weighting γ to the model learning unit 54.

[0102] The model learning unit 54 of learning unit 5 calculates a loss function using the weight y of NCP-Loss set by the outlier detection unit 53 and trains the machine learning model.

[0103] Specifically, the model learning unit 54 calculates the loss function using the weight y set by the outlier detection unit 53 and trains the machine learning model based on the training data acquired by the training data acquisition unit 51 and the training data with added noise generated by the noise addition unit 52. As a result, the model learning unit 54 creates the machine learning model.

[0104] Model learning unit 54 uses the following loss function L(θ) when learning the machine learning model. It should be noted that the term after Epprior in the equation for calculating the following loss function L(θ) is the second term in the loss function used when training the model described above. L(θ)=−Eptrain(x,y)⌈ln pmodel(y|x,θ)⌉+γEpprior(x^)⌈DKL[pprior(y^|x^)‖pmodel(y^|x^,θ)]⌉ is used in a machine learning model.

[0105] At this point, p model(y|x,θ) a probability distribution of the output γ that is obtained when an explanatory variable x is input in a parameter θ.

[0106] Furthermore, -E ptrain(x,y) [INP model(y|x,θ) ] a maximum likelihood inference amount for the training data, and DKL⌈p(y|x)‖pmodel(y|x,θ)⌉ is fulfilled.

[0107] D KLrefers to KL divergence and is not limited to it, as long as it is a procedure for preserving an intermediate distribution distance between a distribution p (y|x) of correct data and an output distribution p model (y|x,θ) of the machine learning model. For the distribution p(y|x) of correct data, a delta function can be considered the distribution, or the distribution p(y|x) of correct data can be a normal distribution.

[0108] The second term E pprior(x̂) [D KL [P prior (ŷ|x̂)llp model (ŷ|x̂,θ)]] of the loss function is a maximum likelihood inference for an explanatory variable x̂ and a target variable ŷ after noise addition, and an inter-distribution distance between a distribution (u y , σ y ) of the target variable ŷ after noise addition and an output distribution P model (ŷ|x̂, θ) from the machine learning model is calculated by KL divergence.

[0109] Due to the weighting y of NCP loss, which is calculated from the outlier score by the outlier detection unit 53, the influence of the second term is large for training data with added noise that deviates from the training data, and the influence of the second term is small for training data with added noise that is close to the training data.

[0110] After training the machine learning model and creating the machine learning model, the model learning unit 54 of the learning unit 5 instructs the model storage unit 55 to store the created machine learning model.

[0111] The model storage unit 55 stores the machine learning model.

[0112] It should be noted that here the model storage unit 55 is contained within the uncertainty learning device 2, but this is only an example. The model storage unit 55 can be located at a location that the uncertainty learning device 2 can access from outside the uncertainty learning device 2.

[0113] The inference unit 6 performs an evaluation of the machine learning model created by the learning unit 5 and an inference, i.e. the prediction, of a predicted value and uncertainty using the machine learning model based on the actual data, here the operational data obtained by the machine facility 1.

[0114] In Inference Unit 6, the evaluation of the machine learning model created by Learning Unit 5 is performed at the time of the model's creation, i.e., at the time of training. It should be noted that the evaluation of the machine learning model is carried out by Evaluation Unit 64 of Inference Unit 6.

[0115] In inference unit 6, at the time of inference, the inference of a predicted value and its uncertainty is performed using the machine learning model based on the actual data. It should be noted that the inference using the machine learning model is performed by prediction unit 62 of inference unit 6.

[0116] The model reading unit 61 refers to the model storage unit 55 and reads the machine learning model.

[0117] The model reading unit 61 outputs the read-out machine learning model to the prediction unit 62 and the evaluation unit 64.

[0118] The test data procurement unit 63 procures test data from the test data storage unit 43.

[0119] The test data procurement unit 63 outputs the procured test data to the evaluation unit 64.

[0120] The evaluation unit 64 evaluates the machine learning model read out by the model reading unit 61, i.e. the machine learning model created by the model learning unit 54 and stored by the model storage unit 55, based on the test data procured by the test data procurement unit 63.

[0121] More specifically, evaluation unit 64 evaluates the accuracy of the machine learning model based on the inference result (prediction value and uncertainty) output by the machine learning model and the target variables of the test data procured by test data acquisition unit 63. Evaluation unit 64 only needs to use a root mean square error (RMSE), a mean absolute error (MAE), a likelihood function, or the like as an evaluation index. Evaluation unit 64 only needs to obtain the inference result output by the machine learning model from model learning unit 54. It should be noted that in Fig. 3. An arrow from the model learning unit 54 to the evaluation unit 64 is not shown.

[0122] Evaluation unit 64 only needs to evaluate the machine learning model using a known machine learning model evaluation procedure based on the test data.

[0123] Model Learning Unit 54 creates the machine learning model by inputting an explanatory variable into the model based on the training data and conducting the training so that the output inference result (prediction value and uncertainty) matches a target variable that corresponds to the explanatory variable at the time of training. Evaluation Unit 64 evaluates the accuracy of the machine learning model to determine the extent to which the accuracy of the machine learning model created by Model Learning Unit 54 has improved at the time of training.

[0124] By obtaining model input data based on the operational data from machine unit 1 and entering the model input data into the machine learning model, the prediction unit 62 receives a prediction value and an uncertainty, thereby inferring the prediction value and the uncertainty.

[0125] As described above, the machine learning model receives input data (here model input data based on the operational data) and outputs a predicted value and an uncertainty.

[0126] It is important to note that the model input data fed into the machine learning model by Predictor Unit 62 is unknown data. Predictor Unit 62 feeds this unknown data into the machine learning model and receives a predicted value that corresponds to a target variable (position error, control value, or similar) of the unknown data and the uncertainty of the predicted value. The predicted value could be, for example, a numerical value in the case of a regression problem and a probability for each class in the case of a classification problem. The uncertainty could be, for example, a variance or a standard deviation, indicating a variation in the data.

[0127] It should be noted that, as described above, the prediction unit 62 procures the model input data, which is to be an input to the machine learning model, from the machine setup 1 via the operating result procurement unit 31 of the procurement unit 3 and the preprocessing execution unit 41 of the preprocessing unit 4.

[0128] The prediction unit 62 outputs the received prediction value and the uncertainty to the machine device 1.

[0129] Upon obtaining the forecast value and the uncertainty from the uncertainty learning device 2, the machine device 1 performs a control operation based on the obtained forecast value and the uncertainty.

[0130] Using a concrete example, the control based on the predicted value and the uncertainty in machine unit 1 is described here. As an example, it is assumed that a correction amount of a control command for a processing position, based on data such as a CAD model and its uncertainty, is output by the uncertainty learning unit 2 as the predicted value and the uncertainty when a workpiece is processed (extracted) into a certain shape in machine unit 1.

[0131] Machine unit 1 corrects the control command based on the correction amount and uncertainty output by uncertainty learning unit 2. For example, if the uncertainty σ output by uncertainty learning unit 2 is equal to or greater than a preset threshold (hereinafter referred to as the "uncertainty determination threshold") th, in short, if the reliability of the predicted value is low, machine unit 1 determines the control command using the following equation. It should be noted that in the following equation, a control command U, an original command value u, and a correction amount Δt are used. U={u+f if σ≤thu if σ>th trag f can be used.

[0132] It should be noted that the example described above is only an example and, for instance, machine device 1 cannot correct the control command if the uncertainty is equal to or greater than the uncertainty threshold, and can correct the control command if the uncertainty is less than the uncertainty threshold.

[0133] Furthermore, in the uncertainty learning unit 2, for example, the prediction unit 62 can determine whether the obtained prediction value and the uncertainty should be output to the machine unit 1 or not. For example, the prediction unit 62 can output the prediction value and the uncertainty to the machine unit 1 if the obtained uncertainty is below the uncertainty determination threshold, and cannot output the prediction value and the uncertainty to the machine unit 1 if the obtained uncertainty is equal to or greater than the uncertainty determination threshold.

[0134] If the uncertainty obtained is high, the prediction unit 62 can also collect the model input data from which the uncertainty is obtained as training data, i.e., the model input data that is fed into the machine learning model through an active learning procedure.

[0135] The operation of the uncertainty learning device 2 according to the first embodiment is described.

[0136] The operation of the uncertainty learning facility 2 is described below separately for learning time operation and inference time operation.

[0137] First, the learning time operation of the uncertainty learning facility 2 is described.

[0138] Fig. Figure 4 is a flowchart describing the learning time operation of the uncertainty learning device 2 according to the first embodiment.

[0139] The Uncertainty Learning Facility 2 conducts the learning time operation, as shown in the flowchart of Fig. Figure 4 illustrates this, for example, based on a user instruction. The user operates an input device (not shown) and enters an operation start instruction. Upon receiving the operation start instruction, a control unit (not shown) of the Uncertainty Learning Device 2 causes the Operation Result Procurement Unit 31, the Preprocessing Execution Unit 41, the Training Data Procurement Unit 51, the Noise Addition Unit 52, the Outlier Detection Unit 53, the Model Learning Unit 54, the Model Reading Unit 61, the Test Data Procurement Unit 63, and the Evaluation Unit 64 to start operation. For example, the user can enter the operation start instruction by inputting operation-related data. The Uncertainty Learning Device 2 repeats the learning-time operation as shown in the flowchart of Figure 4. Fig. 4 shown, for example, until the machine learning model has been evaluated to a certain extent or until the control unit receives a shutdown instruction from the user.

[0140] The operating result procurement unit 31 procures operational data from machine facility 1 (step ST1).

[0141] The operating result procurement unit 31 instructs the operating result storage unit 32 to store the procured business-related data.

[0142] The preprocessing execution unit 41 obtains the business-related data procured by the business result procurement unit 31 from the business result storage unit 32 in step ST1 and performs preprocessing of the shaping of the training data used by the learning unit 5 and the test data used by the evaluation unit 64 of the inference unit 6, i.e. the learning time preprocessing, on the basis of the business-related data (step ST2).

[0143] The preprocessing execution unit 41 instructs the training data storage unit 42 to store the training data generated by the learning time preprocessing. The preprocessing execution unit 41 also instructs the test data storage unit 43 to store the test data generated by the learning time preprocessing.

[0144] Learning Unit 5 performs learning to create a machine learning model based on the training data stored in the training data storage unit 42 and the training data created on the basis of the training data with added noise (Step ST3).

[0145] Learning Unit 5 causes Model Storage Unit 55 to save the created machine learning model.

[0146] The evaluation unit 64 of the inference unit 6 evaluates the machine learning model created by the learning unit 5 in step ST3 (step ST4).

[0147] The test data acquisition unit 63 specifically acquires test data from the test data storage unit 43. Subsequently, the evaluation unit 64 evaluates the machine learning model read by the model reading unit 61 using the test data acquired by the test data acquisition unit 63.

[0148] Fig. 5 is a flowchart describing details of the processing of step ST3 in Fig. 4.

[0149] The training data procurement unit 51 procures training data from the training data storage unit 42 (step ST11).

[0150] The training data procurement unit 51 procures, more specifically, all parts of the training data from the training data storage unit 42.

[0151] The training data acquisition unit 51 outputs the acquired training data to the noise addition unit 52, the outlier detection unit 53 and the model learning unit 54.

[0152] In step ST11, the outlier detection unit 53 inputs all parts of the training data procured by the training data acquisition unit 51 into the outlier detection mechanism and learns all parts of the training data (step ST12).

[0153] The noise addition unit 52 procures the training data, which is procured by the training data procurement unit 51 in step ST11 (step ST13).

[0154] The training data procured by the noise addition unit 52 can be all parts of the training data procured by the training data procurement unit 51, or some parts (mini-batch) of the training data procured by the training data procurement unit 51.

[0155] The noise addition unit 52 adds noise to the training data obtained in step ST13 and creates training data with added noise (step ST14).

[0156] The noise addition unit 52 outputs the created training data with added noise to the outlier detection unit 53 and the model learning unit 54.

[0157] The outlier detection unit 53 calculates an outlier score from the training data acquired in step ST13 and the training data with added noise created by the noise addition unit 52 in step ST14 (step ST15). Specifically, the outlier detection unit 53 feeds the training data with added noise created by the noise addition unit 52 into the outlier detection mechanism learned in step ST12 and calculates an outlier score.

[0158] Subsequently, the outlier detection unit 53 sets a weighting γ of the loss function NCP - Loss based on the calculated outlier score (step ST16).

[0159] The outlier detection unit 53 outputs the graded weighting y model learning unit 54.

[0160] The model learning unit 54 calculates a loss function using the weight v of NCP loss, which is set by the outlier detection unit 53 in step ST16, and trains the machine learning model based on the training data obtained in step ST13 and the training data created in step ST14 with added noise (step ST17).

[0161] Model learning unit 54 determines whether a training session has been performed the intended number of times (step ST18).

[0162] It should be noted that the frequency with which the Model Learning Unit 54 performs training is determined in advance by a user or the like.

[0163] If the model learning unit 54 has not performed the training the intended number of times (“NO” in step ST18), the operation of the learning unit 5 proceeds to the processing of step ST19, and the noise addition unit 52 procures the training data, which is procured by the training data procurement unit 51 (step ST19).

[0164] The training data procured by the noise addition unit 52 can be all parts of the training data procured by the training data procurement unit 51, or some parts (mini-batch) of the training data procured by the training data procurement unit 51.

[0165] The operation of learning unit 5 then proceeds to process step ST14. It should be noted that in this case, in step ST17, the model learning unit 54 calculates a loss function using the weight v NCP-Loss, which is set by the outlier detection unit 53 in step ST16, and trains the machine learning model based on the training data acquired in step ST19 and the training data created in step ST14 with added noise.

[0166] On the other hand, when the model learning unit 54 has performed the training the intended number of times (“YES” in step ST18), the model learning unit 54 causes the model storage unit 55 to store the machine learning model, and the operation of the learning unit 5 terminates the processing, as shown in the flowchart of Fig. 5 shown.

[0167] Next, the inference time operation of the uncertainty learning facility 2 will be described.

[0168] Fig. Figure 6 is a flowchart describing the inference time operation of the uncertainty learning device 2 according to the first embodiment.

[0169] The Uncertainty Learning Facility 2 performs the inference time operation, as shown in the flowchart of Fig. Figure 6 illustrates this, for example, based on a user instruction. The user operates an input device and enters an operation start instruction. Upon receiving the operation start instruction, a control unit of the Uncertainty Learning Unit 2 causes the Operational Result Procurement Unit 31, the Preprocessing Execution Unit 41, the Model Reading Unit 61, and the Prediction Unit 62 to start operation. The Uncertainty Learning Unit 2 repeats the inference-time operation as shown in the flowchart of Figure 6. Fig. 6 is shown, for example, until the control unit receives a shutdown instruction from the user.

[0170] It should be noted that the inference time operation of the uncertainty learning facility 2, which uses the flowchart of Fig. 6 is described, based on the assumption that the learning time operation of the uncertainty learning facility 2, which is described using the flowchart of Fig. 4 is described, before the inference time operation is executed.

[0171] The operating result procurement unit 31 procures operational data from machine equipment 1 (step ST10).

[0172] The operating result procurement unit 31 instructs the operating result storage unit 32 to store the procured business-related data.

[0173] The preprocessing execution unit 41 obtains the operational data procured by the operational result procurement unit 31 in step ST10 from the operational result storage unit 32 and performs preprocessing of the shaping model input data used by the prediction unit 62 of the inference unit 6 on the basis of the operational data, i.e. an inference time preprocessing (step ST20).

[0174] The preprocessing execution unit 41 outputs the model input data, which is created by performing the inference time preprocessing, to the prediction unit 62.

[0175] The model reading unit 61 refers to the model storage unit 55 and reads the machine learning model (step ST30).

[0176] The model reading unit 61 outputs the read-out machine learning model to the prediction unit 62.

[0177] By obtaining model input data based on the operational data from the machine setup 1 via the operating result procurement unit 31 and the preprocessing execution unit 41 and entering the model input data into the machine learning model, the prediction unit 62 receives a prediction value and an uncertainty, from which the prediction value and the uncertainty are inferred (step ST40).

[0178] Then the prediction unit 62 outputs the received prediction value and the uncertainty to the machine device 1 (step ST50).

[0179] In step ST50, the prediction unit 62 can determine whether the obtained prediction value and the uncertainty should be output to the machine device 1 or not, and can output the obtained prediction value and the uncertainty to the machine device 1 if it is determined to output the prediction value and the uncertainty to the machine device 1.

[0180] It should be noted that in the operation shown in the flowchart of Fig. Figure 6 shows that processing is performed in the sequence of steps ST10, ST20, and ST30, but this is only an example. For instance, step ST30 can be processed before step ST10, or it can be processed in parallel with steps ST10 and ST20.

[0181] As described above, the Uncertainty Learning Unit 2 adds noise to the training data, which is generated based on the operating result obtained from the Machine Unit 1, and calculates an outlier score from the training data and the training data with added noise after the noise has been applied. Subsequently, the Uncertainty Learning Unit 2 calculates a weighted loss function based on the calculated outlier score and trains the machine learning model using the training data and the training data with added noise.

[0182] During the calculation of the loss function and the training of the machine learning model, the Uncertainty Learning Unit 2 trains the machine learning model such that the uncertainty is high for extrapolated data that are far from the training data and low for data that are close to the training data, due to the weighting γ set based on the outlier score. As a result, the Uncertainty Learning Unit 2 can provide a machine learning model that is capable of inferring the uncertainty more accurately.

[0183] In the conventional technique described above, training is performed in such a way that the uncertainty is high for extrapolated data that are far removed from the training data. This is achieved by learning the second term of the loss function using data with added noise. However, there is a section where the uncertainty can also be high for data close to the training data or even for the training data itself. On the other hand, Uncertainty Learning Facility 2 can provide a machine learning model that further improves the inferential accuracy of the uncertainty.

[0184] This means that the Uncertainty Learning Facility 2 can provide a machine learning model capable of inferring the uncertainty more accurately for both the extrapolation data and the data close to the training data by controlling the data used for the training data through the outlier detection technique, instead of training all parts of the data to which noise has been added in a single machine learning model capable of inferring both a prediction value and the uncertainty.

[0185] It should be noted that a technique for inferring a prediction value and uncertainty, as disclosed for example in the following reference literature, is known in which a prediction value and uncertainty are obtained by iteratively applying Monte Carlo dropout to a neural network, and calibration is performed with a fitness function that evaluates the accuracy of the prediction value and uncertainty.

[0186] (Reference literature) JP 2018-200677 A

[0187] However, if the Monte Carlo dropout is applied iteratively, as in the technique described above, the calculation time for the prediction is long, and therefore the application to one with a control period in units of milliseconds, such as an FA facility, is difficult.

[0188] The uncertainty learning device 2 according to the first embodiment can provide a machine learning model that is able to output a prediction value and an uncertainty that are applicable for the control of the machine device 1, which has a control period in units of milliseconds, such as an FA device.

[0189] In the first embodiment described above, the uncertainty learning device 2 comprises the procurement unit 3, the preprocessing unit 4 and the inference unit 6, but this is only an example.

[0190] The Uncertainty Learning Facility 2 only needs to have at least the Learning Unit 5, and, for example, the Procurement Unit 3, the Preprocessing Unit 4, and the Inference Unit 6 can be located at a location that the Uncertainty Learning Facility 2 can access outside of the Uncertainty Learning Facility 2.

[0191] If the uncertainty learning facility 2 does not have the procurement unit 3, the preprocessing unit 4, and the inference unit 6, the processing of steps ST1, ST2, and ST4 in the operation of the uncertainty learning facility 2, which is described with reference to the flowchart of Fig. Section 4, as described, can be omitted. Furthermore, the uncertainty learning facility 2 can be used in the flowchart of Fig. Omit operation 6 as described.

[0192] Furthermore, in the first embodiment described above, the uncertainty learning device 2 is integrated into the server, but this is just one example.

[0193] The uncertainty learning device 2 can, for example, be contained in the machine device 1.

[0194] Furthermore, for example, in the Uncertainty Learning Facility 2, some or all of the Operating Result Procurement Unit 31, the Preprocessing Execution Unit 41, the Training Data Procurement Unit 51, the Noise Addition Unit 52, the Outlier Detection Unit 53, the Model Learning Unit 54, the Model Reading Unit 61, the Prediction Unit 62, the Test Data Procurement Unit 63, and the Evaluation Unit 64 may be contained in a facility or the like outside the server.

[0195] Furthermore, in the first embodiment described above, the machine device 1 is an FA device, but this is only one example. The machine device 1 could, for instance, be any of several different devices that solve various tasks using a machine learning model, such as a control device that performs automatic driving control of a mobile object, or a medical device used at a medical site.

[0196] Fig. 7A and Fig. Figures 7B are illustrations showing an example of a hardware configuration of the uncertainty learning device 2 according to the first embodiment.

[0197] In the first embodiment, the functions of the operating result acquisition unit 31, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, and the control unit (not shown) are implemented by a processing circuit 1001. That is, the uncertainty learning unit 2 includes the processing circuit 1001 for performing control to create a machine learning model capable of inferring uncertainty more accurately by training the model such that the uncertainty is high for extrapolation data and low for data close to the training data, due to weighting set based on an outlier score.

[0198] The processing circuit 1001 may be dedicated hardware, as in Fig. 7A is shown, or to represent a processor 1004 executing a program stored in main memory, as in Fig. 7B is shown.

[0199] If the processing circuit 1001 is dedicated hardware, then a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or a combination thereof corresponds to the processing circuit 1001.

[0200] If the processing circuit is the processor 1004, the functions of the operating result acquisition unit 31, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, and the control unit (not shown) are implemented by software, firmware, or a combination of software and firmware. The software or firmware is described as a program and stored in a working memory 1005.When reading and executing the program stored in memory 1005, processor 1004 performs the functions of the operating result acquisition unit 31, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, and the control unit (not shown). That is, the uncertainty learning unit 2 includes memory 1005 for storing a program that causes the steps ST1 to ST4 described above to be executed. Fig. 4 are shown, or the steps ST10 to ST50 described above, which are in Fig. The results shown in Figure 6 are executed when the program is run by processor 1004. Furthermore, it can also be determined that the program stored in memory 1005 causes a computer to execute a processing sequence or procedure performed by the operating result acquisition unit 31, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, and the control unit (not shown).Here, for example, the working memory 1005 corresponds to a non-volatile or volatile semiconductor memory such as RAM, read-only memory (ROM), flash memory, erasable programmable read-only memory (EPROM) or electrically erasable programmable read-only memory (EEPROM) (registered trademark not described below), a magnetic disk, a flexible disk, an optical disk, a compact disc, a mini-disc or a digital versatile disc (DVD).

[0201] It should be noted that the functions of the operating result procurement unit 31, the preprocessing execution unit 41, the training data procurement unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data procurement unit 63, the evaluation unit 64 and the control unit (not shown) are partly implemented by dedicated hardware and partly by software or firmware.For example, the functions of the operating result procurement unit 31 and the preprocessing execution unit 41 can be implemented by the processing circuit 1001 as dedicated hardware, and the functions of the training data procurement unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data procurement unit 63, the evaluation unit 64 and the control unit (not shown) can be implemented by the processor 1004, which reads and executes a program stored in the main memory 1005.

[0202] The operating result storage unit 32, the training data storage unit 42, the test data storage unit 43 and the model storage unit 55 each consist, for example, of a hard disk drive (HDD, not shown) or a solid state drive (SSD, not shown).

[0203] Furthermore, the uncertainty learning device 2 includes an input interface device 1002 and an output interface device 1003, which perform wired or wireless communication with a device such as the machine device 1.

[0204] As explained above, the uncertainty learning device 2 according to the first embodiment comprises: the training data acquisition unit 51, which acquires training data based on operational data obtained from the machine device 1; the noise addition unit 52, which creates training data with added noise, where noise is added to the training data acquired by the training data acquisition unit 51; the outlier detection unit 53, which calculates an outlier score from the training data acquired by the training data acquisition unit 51 and the training data with added noise created by the noise addition unit 52;and the model learning unit 54, which calculates a weighted loss function based on the outlier score calculated by the outlier detection unit 53 and trains a machine learning model based on the training data and the training data with added noise.

[0205] Thus, the Uncertainty Learning Facility 2 can provide a machine learning model that is able to infer uncertainty more accurately than a machine learning model created by a conventional method of training a machine learning model by adding noise to all parts of the training data. Second embodiment.

[0206] Even if a multitude of pieces of operational data obtained from a machine have the same value, the conditions under which these pieces of operational data are obtained may differ. For example, if sheets of two different materials, A and B, are processed by a machine tool, the machine tool feeds the sheets at the same speed, but their behavior may differ due to the different materials. If, for example, a machine learning model trained on operational data obtained from processing material A is used to infer a predictive value and uncertainty from operational data obtained from processing material B, the inferred uncertainty will be small.This is because the operational data obtained from processing material A and the operational data obtained from processing material B are each operational data obtained by processing material A or material B at the same rate, and therefore a machine learning model incorrectly recognizes that it is the same data, i.e., it incorrectly recognizes that it is learned data and infers low uncertainty.

[0207] This was not taken into account in the first embodiment.

[0208] A second embodiment describes an embodiment of creating a machine learning model that is able to infer a high degree of uncertainty using data obtained by an external sensor that cannot be obtained by a machine device, even if certain explanatory variables have the same value because parts of the data obtained by the external sensor are different.

[0209] Fig. Figure 8 is a representation showing a configuration example of an uncertainty learning system 100a including an uncertainty learning device 2a according to the second embodiment.

[0210] The uncertainty learning system 100a comprises the uncertainty learning device 2a, a machine device 1 and an external sensor 7. The uncertainty learning device 2a is connected to the machine device 1 and the external sensor 7 via a network.

[0211] In the second embodiment, it is assumed that the machine device 1 is, for example, an FA device.

[0212] The external sensor 7 is one of several sensors, such as a temperature sensor, a humidity sensor and a vibration sensor, which are located outside the machine unit 1 independently of the machine unit 1.

[0213] The uncertainty learning device 2a according to the second embodiment is, for example, contained in a server.

[0214] In a configuration example of the uncertainty learning device 2a according to the second embodiment, similar components to those of the one described with reference to Fig. In the configuration example described in section 3 of the uncertainty learning device 2 according to the first embodiment, the same reference numerals are used and a redundant description is omitted.

[0215] The configuration example of the uncertainty learning device 2a differs from the configuration example of the uncertainty learning device 2 according to the first embodiment in that a procurement unit 3a comprises, in addition to an operating result procurement unit 31 and an operating result storage unit 32, a sensor data procurement unit 33 and a sensor data storage unit 34.

[0216] The sensor data acquisition unit 33 functions at the time of training and at the time of inference.

[0217] It should be noted that in the second embodiment, similar to the uncertainty learning device 2 according to the first embodiment, the uncertainty learning device 2a performs the “learning” of creating a machine learning model and the “inference” of inferring a predicted value and an uncertainty using the machine learning model created by the “learning”.

[0218] The sensor data procurement unit 33 procures data (hereinafter referred to as "sensor data") that is procured by the external sensor 7, from the external sensor 7.

[0219] The sensor data obtained by the sensor data acquisition unit 33 from the external sensor 7 include, for example, at least one of the following: temperature data of a location where the machine device 1 is situated; humidity data of the location where the machine device 1 is situated; vibration data of the machine device 1; data relating to a state of the machine device 1; and data relating to a task of the machine device 1. In the case of a processing machine, examples of the sensor data include data indicating a temperature of the machine device 1, a vibration of the machine device 1, and the shape of a workpiece, which affects the uniformity of a processing form or a processing speed.

[0220] The sensor data acquisition unit 33 instructs the sensor data storage unit 34 to store the acquired sensor data. For example, the sensor data acquisition unit 33 adds a acquisition date and acquisition time to the sensor data and instructs the sensor data storage unit 34 to store the sensor data.

[0221] The sensor data storage unit 34 stores the sensor data that is procured by the sensor data acquisition unit 33.

[0222] It should be noted that here the sensor data storage unit 34 is contained within the uncertainty learning device 2a, but this is only an example. The sensor data storage unit 34 can be located at a location that the uncertainty learning device 2a can access outside of the uncertainty learning device 2.

[0223] In the second embodiment, a preprocessing execution unit 41 of a preprocessing unit 4 obtains operational data, which is obtained by the operating result procurement unit 31, from the operating result storage unit 32 and obtains the sensor data, which is obtained by the sensor data procurement unit 33, from the sensor data storage unit 34.

[0224] Subsequently, the preprocessing execution unit 41 performs a learning-time preprocessing of shaping training data used by a learning unit 5 and test data used by an evaluation unit 64 or an inference unit 6, based on the operational data and the sensor data.

[0225] More specifically, the preprocessing execution unit 41 adds the sensor data received from the external sensor 7 to an explanatory variable.

[0226] Furthermore, the preprocessing execution unit 41 obtains the operational data procured by the operational result procurement unit 31 from the operational result storage unit 32, procures the sensor data procured by the sensor data procurement unit 33 from the sensor data storage unit 34, and performs preprocessing of the shaping model input data used by a prediction unit 62 of the inference unit 6, based on the operational data and the sensor data, i.e., inference time preprocessing.

[0227] Learning Unit 5 creates a machine learning model based on training data that includes the explanatory variable, based on sensor data.

[0228] Since a specific procedure for creating a machine learning model by the learning unit 5 is similar to the specific creation procedure described in the first embodiment, a redundant description is omitted.

[0229] Learning Unit 5 creates a single machine learning model based on the training data created by the preprocessing execution unit 41, using operational data and sensor data.

[0230] It should be noted that this is only an example and if, for example, some parts of sensor data with different content were procured by the sensor data procurement unit 33 at the time of operation of machine equipment 1, the learning unit 5 can create the machine learning model for each piece of sensor data.

[0231] In this case, for example, preprocessing execution unit 41 creates training data and test data, grouped based on the contents of the sensor data during the learning-time preprocessing. Learning unit 5 then creates a machine learning model based on the training data for each group. It is important to note that when creating the machine learning model for each piece of sensor data, preprocessing execution unit 41 only needs to use the sensor data for grouping the training and test data based on the operational data; it does not need to add the sensor data to the explanatory variable.

[0232] As a concrete example, it is assumed that the sensor data acquired specifies different types (e.g., a circle or a straight line) of processing shapes. In this case, the preprocessing execution unit 41 creates training and test data based on the operational data for each type of processing shape and assigns sensor data specifying the type of processing shape to the created training and test data.

[0233] Learning Unit 5 creates the machine learning model for each processing type. For example, if an operating pattern (operating mode) in machine setup 1 changes depending on the processing type, the uncertainty learning unit 2a can provide a machine learning model that corresponds to the operating pattern of machine setup 1 by creating the machine learning model for each processing type.

[0234] It should be noted that in learning unit 5, a model learning unit 54 causes a model storage unit 55 to store the created machine learning model and the sensor data, in the example described above, the data that specify the type of processing mode, together.

[0235] In the inference unit 6, a model reading unit 61 reads a machine learning model that corresponds to the test data, which is grouped based on the content of the sensor data. For example, the model reading unit 61 only needs to obtain test data from a test data acquisition unit 63 and only needs to specify one machine learning model to be read from the sensor data associated with the test data. It should be noted that in Fig. 8 an arrow from the test data procurement unit 63 to the model reading unit 61 is not shown.

[0236] The evaluation unit 64 of the inference unit 6 evaluates the machine learning model that corresponds to the sensor data associated with the test data, using the test data procured by the test data procurement unit 63.

[0237] By feeding the model input data, generated by the preprocessing execution unit 41 in the inference-time preprocessing, into the machine learning model corresponding to the sensor data associated with the model input data, the prediction unit 62 of the inference unit 6 receives a prediction value and an uncertainty, from which the prediction value and uncertainty are inferred. It should be noted that when the machine learning model is created for each piece of sensor data, the preprocessing execution unit 41 generates model input data for each piece of sensor data in the inference-time preprocessing and outputs the generated model input data and the sensor data, in conjunction with each other, to the prediction unit 62.

[0238] The operation of the uncertainty learning device 2a according to the second embodiment is described.

[0239] First, a learning time operation of the uncertainty learning device 2a according to the second embodiment is described.

[0240] Fig. Figure 9 is a flowchart describing the learning time operation of the uncertainty learning device 2a according to the second embodiment.

[0241] It should be noted that the in Fig. 9. The learning time operation of the uncertainty learning facility 2a shown is an operation in which the learning unit 5 creates a machine learning model in the uncertainty learning facility 2a.

[0242] The uncertainty learning facility 2a conducts the learning time operation as shown in the flowchart of Fig. Figure 9 illustrates this, for example, based on a user instruction. The user operates an input device (not shown) and enters an operation start instruction. Upon receiving the operation start instruction, a control unit (not shown) of the uncertainty learning device 2a causes the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the test data acquisition unit 63, and the evaluation unit 64 to start operation. For example, the user can enter the operation start instruction by inputting operation-related data. The uncertainty learning device 2a repeats the learning-time operation as shown in the flowchart of Figure 9. Fig. 9 shown, for example, until the machine learning model has been evaluated to a certain extent or until the control unit receives a shutdown instruction from the user.

[0243] Since the specific content of the processing of steps ST3 and ST4 in Fig. 9 those of the processing of steps ST3 and ST4 in Fig. Since steps 4, which are described in the first embodiment, are similar, redundant descriptions are omitted. It should be noted that the processing of step ST3 in Fig. 9 in the uncertainty learning facility 2a, specifically the processing is shown in the flowchart of Fig. 5 is shown. Since the operation of the flowchart of Fig. Since the uncertainty learning device 2a shown in section 5, more precisely the model learning unit 54, has already been described in the first embodiment, a redundant description is omitted.

[0244] The operating result procurement unit 31 procures operational data from machine facility 1 (step ST1a).

[0245] The operating result procurement unit 31 instructs the operating result storage unit 32 to store the procured business-related data.

[0246] The sensor data procurement unit 33 procures sensor data from the external sensor 7 (step ST1b).

[0247] The sensor data acquisition unit 33 instructs the sensor data storage unit 34 to store the acquired sensor data.

[0248] The preprocessing execution unit 41 procures operational data, which is procured in step ST1a by the operating result procurement unit 31, from the operating result storage unit 32 and procures the sensor data, which is procured in step ST1b by the sensor data procurement unit 33, from the sensor data storage unit 34.

[0249] The preprocessing execution unit 41 performs preprocessing of the shaping of the training data used by a learning unit 5 and the test data used by the inference unit 6, based on the operational data and the sensor data, i.e. the learning time preprocessing (step ST2a).

[0250] The preprocessing execution unit 41 instructs the training data storage unit 42 to store the training data generated by the learning time preprocessing. The preprocessing execution unit 41 also instructs the test data storage unit 43 to store the test data generated by the learning time preprocessing.

[0251] It should be noted that, for example, when learning unit 5, more precisely model learning unit 54 of learning unit 5, creates a machine learning model for each piece of sensor data, preprocessing execution unit 41 creates training data and test data, which are grouped based on the type of sensor data in the learning time preprocessing in step ST2a. In step ST3, learning unit 5 creates a machine learning model for each piece of sensor data. In step ST4, model reading unit 61 of inference unit 6 reads a machine learning model corresponding to the sensor data, and evaluation unit 64 evaluates the machine learning model corresponding to the sensor data associated with the test data, using the test data acquired by test data acquisition unit 63.

[0252] Next, an inference-time operation of the uncertainty learning device 2a according to the second embodiment is described.

[0253] Fig. Figure 10 is a flowchart describing the inference time operation of the uncertainty learning device 2a according to the second embodiment.

[0254] It should be noted that the in Fig. Figure 10 shows the inference time operation of the uncertainty learning device 2a, an operation in which the learning unit 5 creates a machine learning model in the uncertainty learning device 2a.

[0255] The uncertainty learning facility 2a conducts the learning time operation as shown in the flowchart of Fig. Figure 10 illustrates this, for example, based on a user instruction. The user operates an input device and enters an operation start instruction. Upon receiving the operation start instruction, the control unit of the uncertainty learning device 2a causes the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the model reading unit 61, and the prediction unit 62 to start operation. The uncertainty learning device 2a repeats the inference-time operation as shown in the flowchart of Figure 10. Fig. 10, for example, until the control unit receives a shutdown instruction from the user.

[0256] It should be noted that the inference time operation of the uncertainty learning facility 2a, which uses the flowchart of Fig. 10 is described, based on the assumption that the learning time operation of the uncertainty learning facility 2a, which is described using the flowchart of Fig. 9 is described, before the inference time operation is executed.

[0257] Since the specific content of the processing of steps ST30 to ST50 in Fig. 10 of those processing steps ST30 to ST50 in Fig. Since the 6 described in the first embodiment are similar, a redundant description is omitted.

[0258] The operating result procurement unit 31 procures operational data from machine equipment 1 (step ST10a).

[0259] The operating result procurement unit 31 instructs the operating result storage unit 32 to store the procured business-related data.

[0260] The sensor data procurement unit 33 procures sensor data from the external sensor 7 (step ST10b).

[0261] The sensor data acquisition unit 33 instructs the sensor data storage unit 34 to store the acquired sensor data.

[0262] The preprocessing execution unit 41 obtains the operational data, which is obtained in step ST10a by the operational result procurement unit 31, from the operational result storage unit 32, obtains the sensor data, which is obtained in step ST10b by the sensor data procurement unit 33, from the sensor data storage unit 34, and performs preprocessing of the shaping model input data, which is used by the prediction unit 62 of the inference unit 6, based on the operational data and the sensor data, i.e. an inference time preprocessing (step ST20a).

[0263] The preprocessing execution unit 41 outputs the model input data, which is created by performing the inference time preprocessing, to the prediction unit 62.

[0264] It should be noted that in the operation shown in the flowchart of Fig. Figure 10 shows that processing is carried out in the sequence of steps 10a and 10b, step ST20 and step ST30, but this is only an example. For instance, the processing of step ST30 can be carried out before the processing of steps ST10a and 10b, or it can be carried out in parallel with the processing of steps ST10a and 10b and step ST20.

[0265] If, for example, the learning unit 5, specifically the model learning unit 54 of learning unit 5, creates the machine learning model for each piece of sensor data, the preprocessing execution unit 41 creates model input data for each piece of sensor data in step ST20a and outputs the created model input data and the sensor data in conjunction with each other to the prediction unit 62. In step ST40, the model reading unit 61 of the inference unit 6 reads a machine learning model that corresponds to the sensor data, and in step ST50, the prediction unit 62 receives a prediction value and an uncertainty by inputting the model input data generated by the preprocessing execution unit 41 into the machine learning model that corresponds to the sensor data associated with the model input data, thereby inferring the prediction value and the uncertainty.

[0266] As described above, the uncertainty learning device 2a acquires the sensor data obtained from the external sensor 7, which is located independently of the machine device 1, and acquires training data based on the operational data obtained from the machine device 1 and the sensor data. As described above, the uncertainty learning device 2a adds noise to the training data created from the operational data and the sensor data and calculates an outlier score from the training data and the training data with added noise after the noise has been applied. Subsequently, the uncertainty learning device 2a calculates a weighted loss function based on the calculated outlier score and trains the machine learning model based on the training data and the training data with added noise.

[0267] By creating a machine learning model capable of detecting a difference in operational data caused by a difference in the operational outcome, in other words, a difference in the conditions under which the operational data is obtained, the Uncertainty Learning Facility 2a can provide a machine learning model in which the accuracy of uncertainty inference is further improved.

[0268] It should be noted that in the second embodiment described above, the uncertainty learning device 2a comprises the procurement unit 3a, the preprocessing unit 4 and the inference unit 6, but this is only an example.

[0269] The uncertainty learning facility 2a only needs to have at least the learning unit 5, and, for example, the procurement unit 3a, the preprocessing unit 4 and the inference unit 6 can be located at a location that the uncertainty learning facility 2a can access outside of the uncertainty learning facility 2a.

[0270] If the uncertainty learning facility 2a does not have the procurement unit 3a, the preprocessing unit 4, and the inference unit 6, the processing of steps ST1a, ST1b, ST2, and ST4 in the operation of the uncertainty learning facility 2a, which is based on the flowchart of Fig. Section 9, as described, can be omitted. Furthermore, the uncertainty learning facility 2a can be used in the flowchart of Fig. Omit the operation described in section 10.

[0271] Furthermore, the uncertainty learning device 2a is integrated into the server in the second embodiment described above, but this is just one example.

[0272] The uncertainty learning device 2a can, for example, be contained in the machine device 1.

[0273] Furthermore, for example, in the Uncertainty Learning Facility 2a, some or all of the Operating Result Procurement Unit 31, the Sensor Data Procurement Unit 33, the Preprocessing Execution Unit 41, the Training Data Procurement Unit 51, the Noise Addition Unit 52, the Outlier Detection Unit 53, the Model Learning Unit 54, the Model Reading Unit 61, the Prediction Unit 62, the Test Data Procurement Unit 63, and the Evaluation Unit 64 may be contained in a facility or the like outside the server.

[0274] Furthermore, in the second embodiment described above, the machine device 1 is an FA device, but this is only one example. The machine device 1 could, for instance, be any of several different devices that solve various tasks using a machine learning model, such as a control device that performs automatic driving control of a mobile object, or a medical device used at a medical site.

[0275] Since a hardware configuration of the uncertainty learning device 2a corresponds to the second embodiment of the configuration, which refers to the Fig. 7A and Fig. Since the first embodiment is described in section 7B, its description is omitted.

[0276] In the second embodiment, the functions of the operating result procurement unit 31, the sensor data procurement unit 33, the preprocessing execution unit 41, the training data procurement unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data procurement unit 63, the evaluation unit 64 and the control unit (not shown) are implemented by the processing circuit 1001.That is, the uncertainty learning device 2a includes the processing circuit 1001 for performing control to create a machine learning model that is able to infer uncertainty more accurately by training it such that the uncertainty for extrapolation data is high and the uncertainty for data close to the training data is low, due to the weighting γ set on the basis of an outlier score, and that is able to detect a difference in operational data caused by a difference in the conditions when the operational data are obtained.

[0277] The processing circuit 1001 may be dedicated hardware, as in Fig. 7A is shown, or to illustrate the processor 1004, which executes a program stored in the main memory, as in Fig. 7B is shown.

[0278] When reading and executing the program stored in a working memory 1005, the processing circuit 1001 performs the functions of the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, and the control unit (not shown). That is, the uncertainty learning device 2a includes the working memory 1005 for storing a program that causes the steps ST1a and ST1b to ST4 described above to be executed. Fig. 9 are shown, or the steps ST10a and ST10b to ST50 described above, which are in Fig. The programs shown in Figure 10 are executed as a result when the program is run by the processing circuit 1001. Furthermore, it can also be determined that the program stored in memory 1005 causes a computer to execute a processing sequence or procedure performed by the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, and the control unit (not shown).

[0279] The operating result storage unit 32, the sensor data storage unit 34, the training data storage unit 42, the test data storage unit 43 and the model storage unit 55 each consist, for example, of an HDD or an SSD.

[0280] Furthermore, the uncertainty learning device 2a includes the input interface device 1002 and the output interface device 1003, which perform wired or wireless communication with a device such as the machine device 1.

[0281] As described above, the uncertainty learning device 2a according to the second embodiment comprises the sensor data acquisition unit 33, which acquires sensor data obtained from the external sensor 7, which is arranged independently of the machine device 1, and the training data acquisition unit 51, which acquires training data based on operational data acquired from the machine device 1 and the sensor data acquired by the sensor data acquisition unit 33.

[0282] In the uncertainty learning facility 2a, the model learning unit 54 calculates a weighted loss function based on an outlier score calculated by the outlier detection unit 53 and trains a machine learning model based on the training data created from the operational data and sensor data, as well as the training data with added noise.

[0283] Thus, the Uncertainty Learning Facility 2a can provide a machine learning model capable of inferring uncertainty more accurately than a machine learning model created using a conventional method of training a machine learning model by adding noise to all parts of the training data. Furthermore, by creating a machine learning model capable of detecting a difference in operational data caused by a difference in the operational outcome—in other words, a difference in the conditions under which the operational data is obtained—the Uncertainty Learning Facility 2a can provide a machine learning model with further improved accuracy in inferring uncertainty. Third embodiment.

[0284] Even if the machine unit 1 is not in operation, data (hereinafter referred to as "simulation data") obtained by having a simulator perform a simulation operation in advance, or operational data obtained at an earlier time of operation of the machine unit 1, can be obtained.

[0285] In a third embodiment, an embodiment is described in which a machine learning model is created using pre-acquired simulation data or from previous operational data, and a machine learning model is created on the basis of the machine learning model by performing transfer learning on training data created from operational data obtained through the operation of a machine device 1.

[0286] In the following third embodiment, simulation data obtained in advance or previous operational data are referred to as "preliminary data", and a machine learning model created using the preliminary data is referred to as the "preliminary learning model".

[0287] Fig. Figure 11 is a representation showing a configuration example of an uncertainty learning system 100b including an uncertainty learning device 2b according to the third embodiment.

[0288] The uncertainty learning system 100b comprises the uncertainty learning device 2b and the machine device 1. The uncertainty learning device 2b is connected to the machine device 1 via a network.

[0289] In the third embodiment, it is assumed that the machine device 1 is, for example, an FA device.

[0290] The uncertainty learning device 2b according to the third embodiment is, for example, contained in a server.

[0291] In a configuration example of the uncertainty learning device 2b according to the third embodiment, similar components to those of the one described with reference to Fig. In the configuration example of the uncertainty learning device 2 described in Section 3, the following are designated with the same reference numerals according to the first embodiment, and a redundant description is omitted.

[0292] The configuration example of the uncertainty learning device 2b differs from the configuration example of the uncertainty learning device 2 according to the first embodiment in that the configuration example of the uncertainty learning device 2b comprises a pre-data acquisition unit 8 and that a learning unit 5a comprises a pre-learning unit 56 in addition to a training data acquisition unit 51, a noise addition unit 52, an outlier detection unit 53, a model learning unit 54 and a model storage unit 55.

[0293] The pre-data acquisition unit 8 and the pre-learning unit 56 are operating at the time of training.

[0294] It should be noted that in the third embodiment, similar to the uncertainty learning device 2 according to the first embodiment, the uncertainty learning device 2b performs the “learning” of creating a machine learning model and the “inference” of inferring a predicted value and an uncertainty using the machine learning model created by the “learning”.

[0295] The preliminary data procurement unit 8 procures preliminary data.

[0296] For example, simulation data obtained by inducing a simulator (not shown) to perform a simulation operation, or operational data acquired during the earlier operation of machine unit 1, are stored as pre-data in a cloud (not shown). The pre-data acquisition unit 8 checks whether the pre-data is stored in the cloud and, if so, acquires the pre-data.

[0297] It should be noted that the simulator reproduces the operation or behavior of machine unit 1. The simulation data obtained by initiating a simulation operation includes, for example, data that specifies a value unique to machine unit 1, such as a parameter set in machine unit 1, a configuration of machine unit 1, or its operating mode, as well as log data from machine unit 1.

[0298] The preliminary data acquisition unit 8 outputs the acquired preliminary data to a preprocessing execution unit 41 of a preprocessing unit 4.

[0299] In the third embodiment, when pre-data is output by the pre-data acquisition unit 8, the pre-processing execution unit 41 acquires the pre-data and, based on the pre-data, performs a learning-time pre-processing of the training data used by the pre-learning unit 56 of the learning unit 5a. Details of the pre-learning unit 56 are described later.

[0300] Furthermore, the preprocessing execution unit 41 procures operational data, which is procured by an operational result procurement unit 31, from an operational result storage unit 32 and performs a learning time preprocessing of shaping training data, which is used by the model learning unit 54 of the learning unit 5a, and test data, which is used by a model learning unit 54 of an inference unit 6, based on the operational data.

[0301] Details of the learning-time preprocessing based on the operational data, which is performed by the preprocessing execution unit 41, have already been described in the first embodiment, so a redundant description is unnecessary. Furthermore, the preprocessing execution unit 41 only needs to shape the training data based on the preliminary data according to a procedure similar to the procedure for shaping the training data based on the operational data.

[0302] The preprocessing execution unit 41 adds data to the training data, which is created by performing the learning-time preprocessing based on the pre-data, indicating that the training data is based on the pre-data, and causes a training data storage unit 42 to store the training data. In the following third embodiment, the training data based on the pre-data is also referred to as "pre-training data." When creating the training data by performing the learning-time preprocessing based on the operational data, the preprocessing execution unit 41 adds data to the created training data indicating that the training data is based on the operational data, and causes the training data storage unit 42 to store the training data.Furthermore, the preprocessing execution unit 41 instructs the test data storage unit 43 to store the test data created by performing the learning time preprocessing based on the operational data.

[0303] It should be noted that the details of the inference-time preprocessing performed by the preprocessing execution unit 41 in the third embodiment are similar to the details of the inference-time preprocessing described in the first embodiment, so that a redundant description is omitted. The preprocessing execution unit 41 does not require any preliminary data for the inference-time preprocessing.

[0304] The preprocessing execution unit 41 outputs the model input data, which is created by performing the inference time preprocessing, to the prediction unit 62.

[0305] In the third embodiment, the training data acquisition unit 51 of the learning unit 5a checks whether pre-training data is stored in the training data storage unit 42. If the pre-training data is stored, the training data acquisition unit 51 acquires the pre-training data and outputs the acquired pre-training data to the pre-learning unit 56.

[0306] Furthermore, the training data acquisition unit 51 obtains training data from the training data storage unit 42 and outputs the acquired training data to the noise addition unit 52, the outlier detection unit 53 and the model learning unit 54.

[0307] The pre-learning unit 56 of learning unit 5a causes a pre-learning model to learn a relationship between an explanatory variable and a target variable based on the pre-training data issued by the training data procurement unit 51.

[0308] It should be noted that the pre-learning model and the machine learning model trained by model learning unit 54 are preferably the same type of machine learning model.

[0309] Preliminary learning unit 56 outputs the learned preliminary learning model to model learning unit 54. Model learning unit 54 stores the preliminary learning model in an internal buffer or similar.

[0310] In the third embodiment, the model learning unit 54 of the learning unit 5a reads the pre-learning model, which is stored in the internal buffer or similar, calculates a loss function using a weight γ of NCP loss, which is set by the outlier detection unit 53, and performs transfer learning on the machine learning model.

[0311] More specifically, the model learning unit 54 reads the pre-learning model stored in the internal buffer or similar, calculates the loss function using the weight γ set by the outlier detection unit 53, and trains the machine learning model on the basis of the training data procured by the training data acquisition unit 51 and the training data with added noise created by the noise addition unit 52.

[0312] It should be noted that the calculation of the loss function using the weighting γ of NCP loss was already described in the first embodiment, so a redundant description is omitted. Furthermore, the details of the noise addition unit 52 and the outlier detection unit 53 are similar to those described in the first embodiment, so a redundant description is omitted.

[0313] The transfer learning carried out by model learning unit 54 is a well-known technique, but here transfer learning is only briefly described.

[0314] Fig. Figure 12 is a representation to illustrate an example of transfer learning.

[0315] In Fig. 12 shows Fig. 12A is an example of fine-tuning, Fig. 12B shows an example of feature extraction and Fig. 12C shows an example of joint training.

[0316] Transfer learning is a technique for adapting a model trained in one region to another region and is in particular an effective method for adapting a model trained in a region where a lot of data can be obtained to a region with only a small amount of data or for adapting a model trained in a simulator to a real environment.

[0317] As an example of transfer learning, as in Fig. As shown in Figure 12, there is fine-tuning, which only learns new data; feature extraction, which performs training by adding a new layer to a previously trained model; joint training, which can correspond to a variety of tasks; and the like.

[0318] It should be noted that this is only an example and that the model learning unit 54 can also perform transfer learning using a different transfer learning procedure than these procedures.

[0319] At the time of training, the model learning unit 54 can arbitrarily change the layer of a neural network to be trained and can learn all layers or can learn only one layer.

[0320] The details of the outlier detection unit 53 are similar to those described in the first embodiment. That is, when the outlier detection unit 53 inputs the training data procured by the training data acquisition unit 51—more precisely, the training data generated based on operational data—into an outlier detection mechanism, and learns from the training data and trains the outlier detection mechanism, the outlier detection unit 53 then inputs the training data with added noise, generated by the noise addition unit 52, into the learned outlier detection mechanism and calculates an outlier score.

[0321] However, this is just one example, and the outlier detection unit 53 can, for example, use the pre-training data to train the outlier detection mechanism.

[0322] In particular, the outlier detection unit 53 trains the outlier detection mechanism using data containing the training data and the pre-training data. Then, the outlier detection unit 53 feeds the training data, generated by the noise addition unit 52 and containing added noise, into the outlier detection mechanism and calculates an outlier score.

[0323] It should be noted that in this case the training data acquisition unit 51 outputs the pre-training data, which is obtained from the training data storage unit 42, to the pre-learning unit 56 and the outlier detection unit 53.

[0324] If there is a large amount of pre-processed data and a small amount of application target data, i.e., operational data obtained by machine facility 1, if the outlier detection mechanism is trained with training data based on a small amount of data, when the model learning unit 54 creates a machine learning model, a machine learning model is created in which data that are not included in the small amount of training data based on the small amount of operational data used for training, but are originally close to the training data (especially training data with added noise), are identified as deviations, and the training uncertainty is not correctly controlled.

[0325] The outlier detection unit 53 trains the outlier detection mechanism with data containing the large data set and the small data set, and calculates an outlier score for data obtained by adding noise to the small data set, thus making it possible to prevent the occurrence of an event where the uncertainty training is not correctly controlled, as described above.

[0326] The operation of the uncertainty learning device 2b according to the third embodiment is described.

[0327] First, a learning time operation of the uncertainty learning device 2b according to the third embodiment is described.

[0328] Fig. Figure 13 is a flowchart describing the learning time operation of the uncertainty learning device 2b according to the third embodiment.

[0329] The uncertainty learning facility 2b conducts the learning time operation as shown in the flowchart of Fig. Figure 13 illustrates this, for example, based on a user instruction. The user operates an input device (not shown) and enters an operation start instruction. Upon receiving the operation start instruction, a control unit (not shown) of the Uncertainty Learning Device 2b causes the Operation Result Procurement Unit 31, the Preprocessing Execution Unit 41, the Training Data Procurement Unit 51, the Noise Addition Unit 52, the Outlier Detection Unit 53, the Model Learning Unit 54, the Pre-Learning Unit 56, the Model Reading Unit 61, the Test Data Procurement Unit 63, the Evaluation Unit 64, and the Pre-Data Procurement Unit 8 to start operation. For example, the user can enter the operation start instruction by inputting operational data or pre-data. The Uncertainty Learning Device 2b repeats the learning-time operation as shown in the flowchart of Figure 13. Fig. 13, for example, until the machine learning model has been evaluated to a certain extent or until the control unit receives a shutdown instruction from the user.

[0330] Since the specific content of the processing of steps ST1, ST2 and ST4 in Fig. 13 those of the processing of steps ST1, ST2 and ST4 in Fig. Since the four embodiments described in the first embodiment are similar, a redundant description is omitted.

[0331] The preliminary data acquisition unit 8 determines whether a preliminary learning model has been created or not (step ST101).

[0332] A pre-learning model presence indicator, which specifies whether a pre-learning model has been created or not, is set, for example, at a location to which every component of Uncertainty Learning Facility 2b can refer. The pre-learning model presence indicator specifies, for example, "1: Pre-learning model is present" and "0: Pre-learning model is not present," with an initial value of "0." In Uncertainty Learning Facility 2b, Pre-learning Unit 56 sets the pre-learning model presence indicator to "1" when a pre-learning model is created.

[0333] For example, the preliminary data procurement unit 8 can use the indicator for the presence of a preliminary learning model to determine whether a preliminary learning model has been created or not.

[0334] If the pre-data acquisition unit 8 determines that a pre-learning model has been created (“YES” in step ST101), the operation of the uncertainty learning facility 2b proceeds to the processing of step ST1.

[0335] If it is determined that no pre-learning model has been created (“NO” in step ST101), the pre-data acquisition unit 8 checks whether pre-data is stored in the cloud or not (step ST102).

[0336] If the pre-data is stored in the cloud (“YES” in step ST102), the pre-data procurement unit 8 procures the pre-data (step ST103).

[0337] The preliminary data acquisition unit 8 outputs the acquired preliminary data to a preprocessing execution unit 41 of a preprocessing unit 4.

[0338] The preprocessing execution unit 41 procures the pre-data output by the pre-data procurement unit 8 in step ST103 and performs preprocessing of the shaping of pre-training data used by the pre-learning unit 56 of the learning unit 5a, i.e., a learning time preprocessing, based on the pre-data (step ST104).

[0339] The preprocessing execution unit 41 adds to the pre-training data data indicating that the pre-training data is training data based on the pre-training data, and causes a training data storage unit 42 to store the pre-training data.

[0340] The training data acquisition unit 51 procures the pre-training data that the pre-processing execution unit 41 had stored from the training data storage unit 42 in step ST104, and the pre-learning unit 56 causes the pre-learning model to learn a relationship between an explanatory variable and a target variable based on the pre-training data procured by the training data acquisition unit 51 (step ST105).

[0341] Preliminary learning unit 56 outputs the learned preliminary learning model to model learning unit 54. Model learning unit 54 stores the preliminary learning model in an internal buffer or similar.

[0342] If, however, the pre-data is not stored in the cloud (“NO” in step ST102), the pre-data acquisition unit 8 sets a pre-learning model inability indicator to “1”. The pre-learning model inability indicator is a flag indicating that the pre-learning model cannot be created and is set in a location to which every component of the uncertainty learning facility 2b can refer. For example, the pre-learning model inability indicator specifies “1: a pre-learning model cannot be created” and “0: a pre-learning model can be created”, and an initial value for the pre-learning model inability indicator is “0”. Afterward, the operation of the uncertainty learning facility 2b proceeds to process step ST1.

[0343] In step ST3a, the learning unit 5a performs transfer learning to create a machine learning model based on the training data stored in the training data storage unit 42, the training data created on the basis of the training data with added noise, and the pre-learning model created by the pre-learning unit 56 in step ST105 (step ST3a).

[0344] Learning Unit 5 causes Model Storage Unit 55 to save the created machine learning model.

[0345] Fig. 14 is a flowchart describing details of the processing of step ST3a in Fig. 13.

[0346] Since the specific content of the processing of steps ST11 to ST16, ST18 and ST19 in Fig. 14 of those processing steps ST11 to ST16, ST18 and ST19 in Fig. Since the 5 described in the first embodiment are each similar, a redundant description is omitted.

[0347] Once the pre-learning model has been created, the model learning unit 54 reads the pre-learning model stored in an internal buffer or the like, calculates a loss function using the read-out pre-learning model, the training data obtained in step ST13, the training data created in step ST14 with added noise, and the weight γ of the NCP loss set by the outlier detection unit 53 in step ST16, and performs transfer learning on the machine learning model (step ST17a).

[0348] If the pre-learning model has not been created, the model learning unit 54 calculates a loss function using the training data obtained in step ST13, the training data created in step ST14 with added noise and the weight γ of NCP loss set by the outlier detection unit 53 in step ST16, and trains the machine learning model.

[0349] Model Learning Unit 54 can, for example, use the indicator for the presence of a pre-learning model and the indicator for the inability to create a pre-learning model to determine whether the pre-learning model has been created. For instance, if the indicator for the presence of a pre-learning model is "0" and the indicator for the inability to create a pre-learning model is also "0", Model Learning Unit 54 can postpone the execution of processing step ST17a until the indicator for the presence of a pre-learning model is set to "1"—in other words, until the pre-learning model has been created.

[0350] It should be noted that, as described above, for example the outlier detection unit 53 can use the pre-training data for training the outlier detection mechanism.

[0351] In this case, in step ST12, as shown in the flowchart of Fig. In step ST14, the outlier detection unit 53 trains the outlier detection mechanism using data containing the training data and the pre-training data. Specifically, in step ST15, the outlier detection unit 53 feeds the training data created in step ST14 by the noise addition unit 52, with added noise, into the outlier detection mechanism trained in step ST12 and calculates an outlier score.

[0352] In step ST11, the training data acquisition unit 51 outputs the pre-training data, which is obtained from the training data storage unit 42, to the outlier detection unit 53.

[0353] Since the inference time operation of the uncertainty learning device 2b according to the third embodiment is similar to the inference time operation of the uncertainty learning device 2 according to the first embodiment, which refers to the flowchart of Fig. Since the first embodiment is similar to the one described in section 6, a redundant description is omitted.

[0354] As described above, the uncertainty learning device 2b acquires pre-test data based on simulation data obtained from running simulation operations on machine device 1, or on operational data relating to a previous operating result of machine device 1. It then trains the pre-test learning model using pre-test training data generated from this acquired pre-test data. Subsequently, the uncertainty learning device 2b calculates a weighted loss function based on the calculated outlier score and performs transfer learning on the pre-test learning model using both the training data and the training data with added noise.

[0355] If the training data is not prepared to a certain degree when creating the machine learning model, the machine learning model may be one in which the inference of a predicted value and uncertainty cannot be performed well, and as a result, the inference accuracy may deteriorate when the inference is actually performed based on unknown data (here, model input data based on operational data).On the other hand, by creating a pre-learning machine learning model and performing transfer learning using prior operational data obtained in advance or simulation data obtained from a simulator, the uncertainty learning facility 2b can provide a machine learning model that can well infer a predictive value and uncertainty and has high accuracy, even if there is a small amount of operational data obtained from machine facility 1, with training data for creating a machine learning model being generated from the operational data.

[0356] It should be noted that in the third embodiment described above, the uncertainty learning device 2b comprises the procurement unit 3, the preprocessing unit 4, the inference unit 6 and the pre-data procurement unit 8, but this is only an example.

[0357] The Uncertainty Learning Facility 2b only needs to have at least the Learning Unit 5a, and, for example, the Procurement Unit 3, the Preprocessing Unit 4, the Inference Unit 6, and the Pre-Data Procurement Unit 8 can be located at a location that the Uncertainty Learning Facility 2b can access from outside the Uncertainty Learning Facility 2b.

[0358] If the uncertainty learning facility 2b does not have procurement unit 3, preprocessing unit 4, inference unit 6, and pre-data procurement unit 8, the processing of steps ST101 to ST105, ST1, ST2, and ST4 to ST4 in the operation of the uncertainty learning facility 2b, which is based on the flowchart of Fig. 13, described, can be omitted. Furthermore, the uncertainty learning facility 2b can be used in the flowchart of Fig. Skip the inference time operation shown in section 6.

[0359] Furthermore, the uncertainty learning device 2b is integrated into the server in the third embodiment described above, but this is just one example.

[0360] The uncertainty learning device 2b can, for example, be included in the machine device 1.

[0361] Furthermore, for example, in the Uncertainty Learning Facility 2b, some or all of the Operating Result Procurement Unit 31, the Preprocessing Execution Unit 41, the Training Data Procurement Unit 51, the Noise Addition Unit 52, the Outlier Detection Unit 53, the Model Learning Unit 54, the Pre-Learning Unit 56, the Model Reading Unit 61, the Prediction Unit 62, the Test Data Procurement Unit 63, the Evaluation Unit 64, and the Pre-Data Procurement Unit 8 may be contained in a facility or the like outside the server.

[0362] Furthermore, in the third embodiment described above, the machine device 1 is an FA device, but this is only one example. The machine device 1 could, for instance, be any of several different devices that solve various tasks using a machine learning model, such as a control device that performs automatic driving control of a mobile object, or a medical device used at a medical site.

[0363] Since a hardware configuration of the uncertainty learning device 2b corresponds to the third embodiment of the configuration, which refers to the Fig. 7A and Fig. Since the first embodiment is described in section 7B, its description is omitted.

[0364] In the third embodiment, the functions of the operating result procurement unit 31, the preprocessing execution unit 41, the training data procurement unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the prediction unit 62, the test data procurement unit 63, the evaluation unit 64, the pre-data procurement unit 8 and the control unit (not shown) are implemented by the processing circuit 1001.That is, the uncertainty learning facility 2 includes the processing circuit 1001 for performing control to create a machine learning model that is able to infer uncertainty more accurately by creating a pre-learning model based on pre-data and performing transfer learning such that uncertainty is high for extrapolation data and low for data close to the training data, due to the weighting set on the basis of an outlier score.

[0365] The processing circuit 1001 may be dedicated hardware, as in Fig. 7A is shown, or to illustrate the processor 1004, which executes a program stored in the main memory, as in Fig. 7B is shown.

[0366] When reading and executing the program stored in the main memory 1005, the processing circuit 1001 performs the functions of the operating result acquisition unit 31, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, the pre-data acquisition unit 8, and the control unit (not shown). That is, the uncertainty learning device 2b includes the main memory 1005 for storing a program that causes the steps ST101 to ST105 and ST1 to ST4 described above to be executed. Fig. 13 are shown, or the steps ST10 to ST50 described above, which are in Fig. The processes shown in Figure 6 are executed as a result when the program is run by the processing circuit 1001. Furthermore, it can also be determined that the program stored in memory 1005 causes a computer to execute a processing sequence or procedure performed by the operating result acquisition unit 31, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, the pre-data acquisition unit 8, and the control unit (not shown).

[0367] The operating result storage unit 32, the training data storage unit 42, the test data storage unit 43 and the model storage unit 55 each consist, for example, of an HDD or an SSD.

[0368] Furthermore, the uncertainty learning device 2b includes the input interface device 1002 and the output interface device 1003, which perform wired or wireless communication with a device such as the machine device 1.

[0369] As explained above, the uncertainty learning device 2b according to the third embodiment comprises: the pre-data acquisition unit 8, which acquires pre-data based on simulation data obtained by performing a simulation operation on the machine device 1, or on operational data relating to a previous operational result of the machine device 1;and the pre-learning unit 56, which trains a pre-learning model based on pre-training data generated from pre-training data procured by the pre-training data acquisition unit 8, and the model learning unit 54, which creates a machine learning model by calculating a weighted loss function based on the outlier score calculated by the outlier detection unit 53 and by performing transfer learning on the pre-learning model trained by the pre-learning unit 56 on training data and training data with added noise.

[0370] The uncertainty learning facility 2b can provide a machine learning model that can well infer a predictive value and uncertainty and has high accuracy, even if there is a small amount of operational data obtained from machine facility 1, by creating a pre-training machine learning model using previous operational data obtained in advance or simulation data obtained from a simulator, even if there is a small amount of operational data obtained from machine facility 1, with training data for creating a machine learning model being created from the operational data. Fourth embodiment.

[0371] In the third embodiment, the uncertainty learning device creates the pre-learning model based on the pre-data and creates the machine learning model by transfer learning using the created pre-learning model.

[0372] In a fourth embodiment, a further embodiment is described in which transfer learning is performed using sensor data obtained from an external sensor that cannot be obtained from a machine device, and a machine learning model is created.

[0373] Fig. Figure 15 is a representation showing a configuration example of an uncertainty learning system 100c including an uncertainty learning device 2c according to the fourth embodiment.

[0374] The uncertainty learning system 100c comprises the uncertainty learning device 2c, a machine device 1, and an external sensor 7. The uncertainty learning device 2c is connected to the machine device 1 and the external sensor 7 via a network. Since the external sensor 7 has already been described in the second embodiment, a redundant description is omitted.

[0375] In the fourth embodiment, it is assumed that the machine device 1 is, for example, an FA device.

[0376] The uncertainty learning device 2c according to the fourth embodiment is, for example, contained in a server.

[0377] In a configuration example of the uncertainty learning device 2b according to the fourth embodiment, similar components to those of the one described with reference to Fig. In the configuration example of the uncertainty learning device 2b described in Section 11, the third embodiment is designated with the same reference numerals, and a redundant description is omitted.

[0378] The configuration example of the uncertainty learning device 2c differs from the configuration example of the uncertainty learning device 2b according to the third embodiment in that a procurement unit 3a comprises, in addition to an operating result procurement unit 31 and an operating result storage unit 32, a sensor data procurement unit 33 and a sensor data storage unit 34.

[0379] The details of procurement unit 3a are similar to the details of procurement unit 3a contained in the uncertainty learning device 2a described in the second embodiment, so that a duplicate description is unnecessary.

[0380] It should be noted that in the fourth embodiment, similar to the uncertainty learning device 2b according to the third embodiment, the uncertainty learning device 2c performs the “learning” of creating a machine learning model and the “inference” of inferring a predicted value and an uncertainty using the machine learning model created by the “learning”.

[0381] In the fourth embodiment, when pre-data is output by a pre-data procurement unit 8, the pre-processing execution unit 41 procures the pre-data and performs a learning-time pre-processing of the shaping of the pre-training data used by a pre-learning unit 56 of a learning unit 5a on the basis of the pre-data.

[0382] Furthermore, the preprocessing execution unit 41 obtains operational data, acquired by the operational result acquisition unit 31, from the operational result storage unit 32, obtains sensor data, acquired by the sensor data acquisition unit 33, from the sensor data storage unit 34, and performs a learning-time preprocessing of shaping training data, used by a model learning unit 54 of the learning unit 5, and test data, used by an evaluation unit 64 of an inference unit 6, based on the operational data and the sensor data. More specifically, the preprocessing execution unit 41 adds the sensor data obtained from the external sensor 7 to an explanatory variable.

[0383] Details of the learning-time preprocessing of shaping the pre-training data based on the pre-training data and details of the learning-time-learning preprocessing of shaping the training data and the test data based on the operational data and the sensor data, which are performed by the pre-processing execution unit 41, have already been described in the second embodiment or the third embodiment, so that a redundant description is dispensed with.

[0384] The preprocessing execution unit 41 adds to the pre-training data created by performing the learning-time preprocessing based on the pre-training data data indicating that the training data is training data based on the pre-training data, and causes a training data storage unit 42 to store the pre-training data.

[0385] When creating the training data by performing the learning-time preprocessing based on the operational data, the preprocessing execution unit 41 also adds data to the created training data indicating that the training data is based on the operational data and instructs the training data storage unit 42 to store the training data. Additionally, the preprocessing execution unit 41 instructs the test data storage unit 43 to store the test data created by performing the learning-time preprocessing.

[0386] It should be noted that the details of the inference-time preprocessing performed by the preprocessing execution unit 41 in the fourth embodiment are similar to the details of the inference-time-to-inference-time preprocessing described in the second embodiment, so that a redundant description is omitted. In the inference-time preprocessing, the preprocessing execution unit 41 does not require any preliminary data.

[0387] The preprocessing execution unit 41 outputs the model input data, which is created by performing the inference time preprocessing, to the prediction unit 62.

[0388] By reading the pre-learning model created by the pre-learning unit 56, which is stored in an internal buffer or similar, by calculating a loss function using a weight γ NCP-Loss, which is set by the outlier detection unit 53 on the basis of training data that has an explanatory variable based on sensor data, and by performing transfer learning on a machine learning model, the model learning unit 54 of the learning unit 5 creates a machine learning model.

[0389] Since details of the preliminary learning unit 56 have already been described in the third embodiment, a redundant description is omitted. Furthermore, since details of the transfer learning performed by the model learning unit 54 have already been described in the third embodiment, a redundant description is omitted.

[0390] The model learning unit 54 creates a single machine learning model based on the operational data and sensor data created by the preprocessing execution unit 41, and a single pre-learning model created by the pre-learning unit 56.

[0391] It should be noted that this is only an example and, for instance, the pre-learning unit 56 can create a pre-learning model for each piece of sensor data, and the model-learning unit 54 can create a machine learning model obtained by performing transfer learning on the pre-learning model for each piece of sensor data.

[0392] For example, if the pre-data acquisition unit 8 can acquire simulation data with the same content as the sensor data acquired by the sensor data acquisition unit 33 from the external sensor 7, the pre-learning unit 56 can create a pre-learning model for each piece of sensor data content, and the model learning unit 54 can create a machine learning model obtained by performing transfer learning on the pre-learning model for each piece of sensor data content.

[0393] In this case, for example, the preprocessing execution unit 41 creates pre-training data during the learning-time preprocessing phase. This pre-training data is grouped based on simulation data that has the same content as the sensor data and is contained within the pre-training data. This simulation data, which has the same content as the sensor data and is contained within the pre-training data, is subsequently referred to as "model classification pre-training data." The preprocessing execution unit 41 then combines the model classification pre-training data with the created pre-training data.

[0394] Furthermore, the preprocessing execution unit 41 creates, for example, training data and test data, which are grouped based on the content of the sensor data in the learning time preprocessing. The preprocessing execution unit 41 then links sensor data with the created training and test data.

[0395] Preliminary learning unit 56 creates a preliminary learning model for all parts of the preliminary training data that are grouped with the model classification preliminary data. Preliminary learning unit 56 then outputs the created preliminary learning model, along with the model classification preliminary data associated with the preliminary training data from which the preliminary learning model was created, to model learning unit 54.

[0396] The model learning unit 54 reads a pre-learning model corresponding to the sensor data linked to the training data acquired by the training data acquisition unit 51 as a pre-learning model used for transfer learning. It calculates a loss function using the read pre-learning model and a weight γ of NCP loss set by the outlier detection unit 53, and performs the transfer learning on the machine learning model. The model learning unit 54 then instructs the model storage unit 55 to store the created machine learning model and the associated sensor data.

[0397] Since, as described above, the pre-data acquisition unit 8 outputs the pre-learning model to the model learning unit 54 in conjunction with the model classification pre-data associated with the pre-training data from which the pre-learning model is created, the model learning unit 54 can specify a pre-learning model corresponding to a content of the sensor data associated with the training data, based on the model classification pre-data associated with the pre-learning model.

[0398] In the inference unit 6, a model reading unit 61 reads a machine learning model that corresponds to the test data, which is grouped based on the content of the sensor data.

[0399] The evaluation unit 64 of the inference unit 6 evaluates the machine learning model that corresponds to the sensor data associated with the test data, using the test data procured by the test data procurement unit 63.

[0400] By feeding the model input data, generated by the preprocessing execution unit 41 in the inference-time preprocessing, into the machine learning model corresponding to the sensor data associated with the model input data, the prediction unit 62 of the inference unit 6 receives a prediction value and an uncertainty, from which the prediction value and uncertainty are inferred. It should be noted that when the machine learning model is created for each type of sensor data, the preprocessing execution unit 41 generates model input data for each element of the sensor data in the inference-time preprocessing and outputs the generated model input data and the sensor data, in conjunction with each other, to the prediction unit 62.

[0401] In the fourth embodiment, the outlier detection unit 53 can also use the pre-training data for training the outlier detection mechanism, similar to the third embodiment.

[0402] The operation of the uncertainty learning device 2c according to the fourth embodiment is described.

[0403] First, a learning time operation of the uncertainty learning device 2c according to the fourth embodiment is described.

[0404] Fig. Figure 16 is a flowchart describing the operation of the uncertainty learning device 2c according to the fourth embodiment.

[0405] It should be noted that the in Fig. 16. The operation of the uncertainty learning facility 2c shown is an operation in which the learning unit 5a creates a machine learning model in the uncertainty learning facility 2c.

[0406] The uncertainty learning facility 2c conducts the learning time operation as shown in the flowchart of Fig. Figure 16 illustrates this, for example, based on a user instruction. The user operates an input device (not shown) and enters an operation start instruction. Upon receiving the operation start instruction, a control unit (not shown) of the uncertainty learning device 2a causes the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the test data acquisition unit 63, the evaluation unit 64, and the pre-data acquisition unit 8 to start operation. For example, the user can enter the operation start instruction by inputting operation-related data.The uncertainty learning facility 2c repeats the learning time operation, as shown in the flowchart of . Fig. 16 shown, for example, until the machine learning model has been evaluated to a certain extent or until the control unit receives a shutdown instruction from the user.

[0407] Since the specific content of the processing of steps ST101 to ST105 and ST3a in Fig. 16 of those involved in the processing of ST101 to ST105 and ST3a in Fig. Since steps 13, which are described in the third embodiment, are all similar, a redundant description is omitted. Furthermore, since the specific content of the processing of steps ST1a, ST1b, ST2a and ST4 in Fig. 16 of those processing steps ST1a, ST1b, ST2a and ST4 in Fig. Since steps 9, which are described in the second embodiment, are similar, redundant descriptions are omitted. It should be noted that the processing of step ST3a in Fig. 16 in the uncertainty learning facility 2c specifically the processing is that shown in the flowchart of Fig. 14 is shown. Since the operation of the flowchart of Fig. Since the uncertainty learning device 2c shown in Figure 14, more precisely the model learning unit 54, has already been described in its third embodiment, a redundant description is omitted.

[0408] It should be noted that, for example, if learning unit 5a, more specifically model learning unit 54 of learning unit 5a, creates a machine learning model for each piece of sensor data, preprocessing execution unit 41 in step ST104 creates data with the same content as the sensor data contained in the pre-processing data, i.e., pre-training data grouped based on the model classification pre-processing data in the learning-time preprocessing. Furthermore, preprocessing execution unit 41 in step ST2a creates training data and test data grouped based on the content of the sensor data in the learning-time preprocessing.

[0409] In step ST105, pre-learning unit 56 creates a pre-learning model for each part of the grouped pre-training data. Pre-learning unit 56 then outputs this created pre-learning model, along with the model classification pre-data contained in the pre-training data from which the pre-learning model was created, to model-learning unit 54.

[0410] In step ST3a, more specifically in step ST17a in Fig. 14 The model learning unit 54 reads a pre-learning model corresponding to the sensor data linked to the training data obtained by the training data acquisition unit 51 as a pre-learning model used for transfer learning, calculates a loss function using the read-out pre-learning model and a weight γ of NCP loss set by the outlier detection unit 53, and performs the transfer learning on the machine learning model.

[0411] In step ST4, the model reading unit 61 of the inference unit 6 reads a machine learning model corresponding to the sensor data, and the evaluation unit 64 evaluates the machine learning model corresponding to a content of the sensor data associated with the test data, using the test data obtained by the test data procurement unit 63.

[0412] Since the inference-time operation of the uncertainty learning device 2c according to the fourth embodiment is similar to the inference-time operation of the uncertainty learning device 2a according to the second embodiment, which is referred to in the flowchart of Fig. Since the second embodiment is similar as described in section 10, a redundant description is omitted.

[0413] As described above, the uncertainty learning device 2c obtains the pre-test data based on the operational result-related simulation data obtained by performing the simulation operation on machine device 1, or the operational data relating to a previous operational result of machine device 1, and trains the pre-test learning model based on the pre-test training data created from the pre-test data.

[0414] The uncertainty learning device 2c obtains the sensor data obtained from the external sensor 7, which is arranged independently of the machine device 1, and obtains training data created on the basis of the operational data obtained from the machine device 1 and the sensor data.

[0415] Subsequently, by calculating a weighted loss function based on the outlier score and performing transfer learning on the pre-learning model based on the training data and the training data with added noise, the uncertainty learning facility 2 creates a machine learning model.

[0416] By creating a machine learning model capable of detecting a difference in operational data caused by a difference in conditions when an operational result is obtained—in other words, when the operational data is obtained—the Uncertainty Learning Facility 2c can provide a machine learning model in which the accuracy of uncertainty inference is further improved. Furthermore, by creating a pre-learning model of a machine learning model using prior operational data obtained in advance or simulation data obtained from a simulator, the Uncertainty Learning Facility 2c can provide a machine learning model that can accurately infer a predicted value and uncertainty, even when there is a small amount of operational data obtained from Machine Facility 1.where training data for the creation of a machine learning model is generated from the operational data.

[0417] Furthermore, the uncertainty learning device 2c can group the pre-test data into groups that have the same content as the sensor data contained in the pre-test data, create the grouped pre-test training data, and create a pre-test learning model for each part of the grouped pre-test training data. Subsequently, the uncertainty learning device 2c can create a machine learning model by calculating a weighted loss function based on an outlier score using the training data created from the operational data obtained from machine device 1, the sensor data acquired from external sensor 7, and the training data with added noise, and by performing transfer learning on the pre-test learning model corresponding to the sensor data.As a result, the uncertainty learning facility 2c can select a machine learning model suitable for transfer learning and can provide a machine learning model where the uncertainty inference accuracy is further improved.

[0418] It should be noted that in the fourth embodiment described above, the uncertainty learning device 2c comprises the procurement unit 3a, the preprocessing unit 4, the inference unit 6 and the pre-data procurement unit 8, but this is only an example.

[0419] The Uncertainty Learning Facility 2c only needs to have at least the Learning Unit 5a, and, for example, the Procurement Unit 3a, the Preprocessing Unit 4, the Inference Unit 6, and the Pre-Data Procurement Unit 8 can be located at a location that the Uncertainty Learning Facility 2c can access from outside the Uncertainty Learning Facility 2c.

[0420] If the uncertainty learning facility 2c does not have the procurement unit 3a, the preprocessing unit 4, the inference unit 6, and the pre-data procurement unit 8, the processing of steps ST101 to ST105, ST1a, ST1b, ST2a, and ST4 in the operation of the uncertainty learning facility 2c, which is based on the flowchart of Fig. 16 can be described, omitted. Furthermore, the uncertainty learning facility 2a can be used in the flowchart of Fig. Omit the inference time operation described in section 10.

[0421] Furthermore, the uncertainty learning device 2c is integrated into the server in the fourth embodiment described above, but this is just one example.

[0422] The uncertainty learning device 2c can, for example, be included in the machine device 1.

[0423] Furthermore, for example, in the Uncertainty Learning Facility 2c, some or all of the Operating Result Procurement Unit 31, the Sensor Data Procurement Unit 33, the Preprocessing Execution Unit 41, the Training Data Procurement Unit 51, the Noise Addition Unit 52, the Outlier Detection Unit 53, the Model Learning Unit 54, the Pre-Learning Unit 56, the Model Reading Unit 61, the Prediction Unit 62, the Test Data Procurement Unit 63, the Evaluation Unit 64, and the Pre-Data Procurement Unit 8 may be contained in a facility or the like outside the server.

[0424] Furthermore, in the fourth embodiment described above, the machine device 1 is an FA device, but this is only one example. The machine device 1 could, for instance, be any of several different devices that solve various tasks using a machine learning model, such as a control device that performs automatic driving control of a mobile object, or a medical device used at a medical site.

[0425] Since a hardware configuration of the uncertainty learning device 2c corresponds to the fourth embodiment of the configuration, which refers to the Fig. 7A and Fig. Since the first embodiment is described in section 7B, its description is omitted.

[0426] In the fourth embodiment, the functions of the operating result procurement unit 31, the sensor data procurement unit 33, the preprocessing execution unit 41, the training data procurement unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the prediction unit 62, the test data procurement unit 63, the evaluation unit 64, the pre-data procurement unit 8 and the control unit (not shown) are implemented by a processing circuit 1001.That is, the uncertainty learning device 2c includes the processing circuit 1001 for performing control to create a machine learning model capable of inferring uncertainty more accurately by creating a pre-learning model based on pre-training data and performing transfer learning such that the uncertainty is high for extrapolation data and low for data close to the training data, due to the weighting γ set on the basis of the pre-learning model and an outlier score. It is also capable of detecting a difference in operational data caused by a difference in conditions when the operational data is obtained.

[0427] The processing circuit 1001 may be dedicated hardware, as in Fig. 7A is shown, or to illustrate the processor 1004, which executes a program stored in the main memory, as in Fig. 7B is shown.

[0428] When reading and executing the program stored in the main memory 1005, the processing circuit 1001 performs the functions of the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, the pre-data acquisition unit 8, and the control unit (not shown). That is, the uncertainty learning device 2c includes the main memory 1005 for storing a program that causes the steps ST101 to ST105, ST1a, and ST1b to ST4 described above to be executed. Fig. 16 are shown, or the steps ST10a and ST10b to ST50 described above, which are in Fig.The programs shown in Figure 10 are executed as a result when the program is run by the processing circuit 1001. Furthermore, it can also be determined that the program stored in memory 1005 causes a computer to execute a processing sequence or procedure performed by the operating result acquisition unit 31, the sensor data acquisition unit 33, the preprocessing execution unit 41, the training data acquisition unit 51, the noise addition unit 52, the outlier detection unit 53, the model learning unit 54, the pre-learning unit 56, the model reading unit 61, the prediction unit 62, the test data acquisition unit 63, the evaluation unit 64, the pre-data acquisition unit 8, and the control unit (not shown).

[0429] The operating result storage unit 32, the sensor data storage unit 34, the training data storage unit 42, the test data storage unit 43 and the model storage unit 55 each consist, for example, of an HDD or an SSD.

[0430] Furthermore, the uncertainty learning device 2c includes the input interface device 1002 and the output interface device 1003, which perform wired or wireless communication with a device such as the machine device 1.

[0431] As explained above, the uncertainty learning device 2c according to the fourth embodiment comprises: the pre-data acquisition unit 8, which acquires pre-data based on operational result-related simulation data obtained by performing a simulation operation on the machine device 1, or on operational data relating to a previous operational result of the machine device 1; the pre-learning unit 56, which trains a pre-learning model based on pre-training data generated on the basis of pre-data acquired by the pre-data acquisition unit 8;and the sensor data acquisition unit 33, which acquires sensor data obtained from the external sensor 7, which is arranged independently of the machine device 1; the training data acquisition unit 51, which acquires training data based on operational data obtained from the machine device 1 and sensor data acquired by the sensor data acquisition unit 33; and the model learning unit 54, which creates a machine learning model by calculating a weighted loss function based on the outlier score calculated by the outlier detection unit 53 and by performing transfer learning on the pre-learning model trained by the pre-learning unit 56 based on training data and training data with added noise.

[0432] By creating a machine learning model capable of detecting a difference in operational data caused by a difference in conditions when an operational result is obtained—in other words, when the operational data is obtained—the Uncertainty Learning Facility 2c can provide a machine learning model in which the accuracy of uncertainty inference is further improved. Furthermore, by creating a pre-learning model of a machine learning model using prior operational data obtained in advance or simulation data obtained from a simulator, the Uncertainty Learning Facility 2c can provide a machine learning model that can accurately infer a predicted value and uncertainty, even when there is a small amount of operational data obtained from Machine Facility 1.where training data for the creation of a machine learning model is generated from the operational data.

[0433] It should be noted that the embodiments can be freely combined, that each individual element in each embodiment can be changed, or that each individual element in each embodiment can be omitted.

[0434] The following section describes various aspects of the present revelation together as supplementary remarks. (Supplementary note 1)

[0435] An uncertainty learning device that creates a machine learning model which receives as input data based on operational data relating to an operational result of a machine device and outputs a predicted value corresponding to the operational data and the uncertainty of the predicted value, wherein the uncertainty learning device comprises: a training data procurement unit that procures training data created on the basis of operational data obtained from the machine equipment; a noise addition unit that creates training data with added noise, where noise is added to the training data procured by the training data procurement unit; an outlier detection unit that calculates an outlier score from the training data procured by the training data acquisition unit and the training data with added noise created by the noise addition unit; and a model learning unit that calculates a weighted loss function based on the outlier score calculated by the outlier detection unit, and trains the machine learning model based on the training data and the training data with added noise. (Supplementary note 2)

[0436] The uncertainty learning facility according to Supplementary Note 1 further includes: a test data procurement unit that procures test data created based on operational data obtained from the machine setup; and an evaluation unit that evaluates the machine learning model by inferring the prediction value and uncertainty based on the test data procured by the test data acquisition unit and the machine learning model created by the model learning unit. (Supplementary note 3)

[0437] The uncertainty learning facility according to Supplementary Note 1 further comprises a preprocessing execution unit which creates the training data based on the operational data, wherein The training data acquisition unit procures the training data created by the preprocessing execution unit. (Supplementary note 4)

[0438] The uncertainty learning facility according to Supplementary Note 2 further comprises a preprocessing execution unit which creates the training data and the test data based on the operational data, wherein The training data acquisition unit procures the training data created by the preprocessing execution unit, and The test data acquisition unit procures the test data created by the preprocessing execution unit. (Supplementary note 5)

[0439] The uncertainty learning facility according to one of the supplementary notes 1 to 4, in which Uncertainty is represented by a variance or a standard deviation, which indicates a fluctuation in the data. (Supplementary note 6)

[0440] The uncertainty learning facility according to one of the supplementary notes 1 to 5, further exhibiting: a preliminary data acquisition unit that acquires preliminary data based on simulation data obtained by conducting a simulation operation on the machine equipment, or on operational data relating to the previous operating results of the machine equipment; and a pre-learning unit that trains a pre-learning model based on pre-training data generated from pre-data procured by the pre-data procurement unit, wherein The model learning unit creates the machine learning model by calculating a weighted loss function based on the outlier score calculated by the outlier detection unit and by performing transfer learning on the pre-learning model, which is trained by the pre-learning unit based on the training data and the training data with added noise. (Supplementary note 7)

[0441] The uncertainty learning device according to one of Supplementary Notes 1 to 5, further comprising a sensor data acquisition unit that acquires sensor data obtained from an external sensor that is arranged independently of the machine device, wherein The training data procurement unit procures the training data, which is created on the basis of the operational data procured by the machine equipment and the sensor data procured by the sensor data procurement unit. (Supplementary note 8)

[0442] The uncertainty learning facility according to one of the supplementary notes 1 to 5, further exhibiting: a preliminary data procurement unit that procures preliminary data based on the operating result-related simulation data obtained by carrying out a simulation operation on the machine equipment, or the operating-related data relating to the previous operating result of the machine equipment; a pre-learning unit that trains a pre-learning model based on pre-training data generated from pre-data acquired by the pre-data acquisition unit; and a sensor data acquisition unit that acquires sensor data obtained from an external sensor that is arranged independently of the machine setup, wherein The training data acquisition unit procures the training data, which is created based on the operational data received from the machine equipment and the sensor data procured by the sensor data acquisition unit, and The model learning unit creates the machine learning model by calculating a weighted loss function based on the outlier score calculated by the outlier detection unit and by performing transfer learning on the pre-learning model, which is trained by the pre-learning unit based on the training data and the training data with added noise. (Supplementary note 9)

[0443] The uncertainty learning facility according to one of the supplementary notes 1 to 8, in which The machine setup is an FA setup, and the operational data includes a command position, a command speed, and a command duration. acceleration, a feedback velocity, a feedback acceleration, a current value, or a value obtained by measuring a deviation of a processing position. (Supplementary note 10)

[0444] An uncertainty learning program that creates a machine learning model which receives as input data based on operational data relating to an operational result of a machine device and outputs a predicted value corresponding to the operational data and the uncertainty of the predicted value, wherein the uncertainty learning program is used to cause a computer to operate as: a training data procurement unit that procures training data created on the basis of operational data obtained from the machine equipment; a noise addition unit that creates training data with added noise, where noise is added to the training data procured by the training data procurement unit; an outlier detection unit that calculates an outlier score from the training data procured by the training data acquisition unit and the training data with added noise created by the noise addition unit; and a model learning unit that calculates a weighted loss function based on the outlier score calculated by the outlier detection unit, and trains the machine learning model based on the training data and the training data with added noise. (Supplementary note 11)

[0445] A comprehensive uncertainty learning system: the uncertainty learning facility according to one of Supplementary Notes 1 to 9; and the machine setup. INDUSTRIAL APPLICABILITY

[0446] The uncertainty learning device according to the present disclosure can provide a machine learning model that is capable of inferring a predictive value and uncertainty, and is capable of inferring uncertainty more accurately than a machine learning model created by a conventional method of training a machine learning model by adding noise to all parts of the training data. REFERENCE MARK LIST

[0447] 1: Machine setup, 2, 2a, 2b, 2c: Uncertainty learning unit, 3, 3a: Procurement unit, 31: Operating result procurement unit, 32: Operating result storage unit, 33: Sensor data procurement unit, 34: Sensor data storage unit, 4: Preprocessing unit, 41: Preprocessing execution unit, 42: Training data storage unit, 43: Test data storage unit, 5, 5a: Learning unit, 51: Training data procurement unit, 52: Noise addition unit, 53: Outlier detection unit, 54: Model learning unit, 55: Model storage unit, 56: Pre-learning unit, 6: Inference unit, 61: Model read unit, 62: Prediction unit, 63: Test data procurement unit, 64: Evaluation unit 7: External sensor, 8: Preliminary data acquisition unit, 100, 100a, 100b, 100c: Uncertainty learning system, 1001: Processing circuit, 1002: Input interface device, 1003: Output interface device, 1004: Processor, 1005: Main memory QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] JP 2018-200677 A

[0186]

Claims

Uncertainty learning device for creating a machine learning model to receive as input data based on operational data relating to an operational result of a machine device and to output a predicted value corresponding to the operational data and the uncertainty of the predicted value, wherein the uncertainty learning device comprises: a training data acquisition unit for acquiring training data created based on the operational data obtained from the machine device; a noise addition unit for creating training data with added noise, where noise is added to the training data acquired by the training data acquisition unit;an outlier detection unit for calculating an outlier score from the training data procured by the training data acquisition unit and the training data with added noise created by the noise addition unit; and a model learning unit for calculating a weighted loss function based on the outlier score calculated by the outlier detection unit, and training the machine learning model based on the training data and the training data with added noise. Uncertainty learning device according to claim 1, further comprising: a test data acquisition unit for acquiring test data that is created on the basis of the operational data obtained from the machine device; and an evaluation unit for evaluating the machine learning model by inferring the prediction value and the uncertainty on the basis of the test data acquired by the test data acquisition unit and the machine learning model created by the model learning unit. Uncertainty learning device according to claim 1, further comprising a preprocessing execution unit for creating the training data based on the operational data, wherein the training data procurement unit procures the training data created by the preprocessing execution unit. Uncertainty learning device according to claim 2, further comprising a preprocessing execution unit for creating the training data and the test data based on the operational data, wherein the training data procurement unit procures the training data created by the preprocessing execution unit, and the test data procurement unit procures the test data created by the preprocessing execution unit. Uncertainty learning device according to one of claims 1 to 4, wherein the uncertainty is represented by a variance or a standard deviation indicating a fluctuation of the data. Uncertainty learning device according to any one of claims 1 to 5, further comprising: a pre-data acquisition unit for acquiring pre-data based on simulation data obtained by performing a simulation operation on the machine device, or the operational data relating to the previous operational result of the machine device; and a pre-learning unit for training a pre-learning model based on pre-training data generated on the basis of the pre-data acquired by the pre-data acquisition unit, wherein the model learning unit creates the machine learning model by calculating a weighted loss function based on the outlier score calculated by the outlier detection unit and by performing transfer learning on the pre-learning model trained by the pre-learning unit on the basis of the training data and the training data with added noise. Uncertainty learning device according to one of claims 1 to 5, further comprising a sensor data acquisition unit for acquiring sensor data obtained from an external sensor that is arranged independently of the machine device, wherein the training data acquisition unit acquires the training data which is created on the basis of the operational data acquired by the machine device and the sensor data acquired by the sensor data acquisition unit. Uncertainty learning device according to one of claims 1 to 5, further comprising: a preliminary data acquisition unit for acquiring preliminary data based on the operational result-related simulation data obtained by performing a simulation operation on the machine device, or the operational data relating to the previous operational result of the machine device; a preliminary learning unit for training a preliminary learning model based on preliminary training data generated on the basis of the preliminary data acquired by the preliminary data acquisition unit;and a sensor data acquisition unit for acquiring sensor data obtained from an external sensor located independently of the machine setup, wherein the training data acquisition unit acquires the training data created based on the operational data obtained from the machine setup and the sensor data acquired by the sensor data acquisition unit, and the model learning unit creates the machine learning model by calculating a weighted loss function based on the outlier score calculated by the outlier detection unit and by performing transfer learning on the pre-learning model trained by the pre-learning unit on the basis of the training data and the training data with added noise. Uncertainty learning device according to one of claims 1 to 8, wherein the machine device is an FA device and the operational data includes a command position, a command rate, a command acceleration, a feedback rate, a feedback acceleration, a current value or a value obtained by measuring a deviation of a processing position. Uncertainty learning program for creating a machine learning model to receive as input data based on operational data relating to an operational result of a machine facility and to output a predicted value corresponding to the operational data and the uncertainty of the predicted value, wherein the uncertainty learning program is used to cause a computer to operate as: a training data acquisition unit for acquiring training data based on the operational data obtained from the machine facility; a noise addition unit for creating training data with added noise, where noise is added to the training data acquired by the training data acquisition unit;an outlier detection unit for calculating an outlier score from the training data procured by the training data acquisition unit and the training data with added noise created by the noise addition unit; and a model learning unit for calculating a weighted loss function based on the outlier score calculated by the outlier detection unit, and training the machine learning model based on the training data and the training data with added noise. Uncertainty learning system, comprising: the uncertainty learning device according to any one of claims 1 to 9; and the machine device.

Images