Information Processing Systems

Abstract

This information processing system predicts, by using a trained model, the state of a device from time-series data acquired from the device. During prediction, the information processing system causes the trained model to: extract a feature depending on a permutation from a plurality of partial time-series data items obtained by dividing time-series data into a plurality of data items along a time axis; generate a plurality of first vectors corresponding one-on-one to the partial time-series data items in each of which the extracted feature is embedded; generate a second vector in which the plurality of first vectors are embedded; extract a feature depending on a permutation from the plurality of first vectors; generate a third vector in which the extracted feature is embedded; generate a fourth vector in which the second vector and the third vector are embedded; and convert the fourth vector to a first value indicating the state of the device.

Description

[Technical Field]

[0001] The present invention relates to an information processing system, an information processing method, and a recording medium for predicting the state of a device from time-series data acquired from the device. [Background technology]

[0002] To achieve predictive maintenance, it is necessary to quantitatively predict the degree of equipment deterioration. The quantitatively predicted degree of deterioration is called the remaining useful life (RUL). A method has been proposed that uses machine learning to predict the remaining useful life from measurement data, which is time-series data.

[0003] For example, a first related technology related to the present invention is a technology in which time series data acquired from equipment by a sensor is divided into multiple partial time series data along the time axis, and the feature quantities extracted from each partial time series data are input into a recurrent neural network to predict the remaining life of the equipment (e.g., Patent Document 1).

[0004] Furthermore, a second related technology related to the present invention is a technology for obtaining various statistical quantities such as effective value, maximum value, crest factor, kurtosis, and skewness from time series data of a predetermined length acquired from equipment by a sensor for each time series data, generating a feature vector, and predicting the remaining life of the equipment from this feature vector (e.g., Patent Document 2). [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2020-198081 [Patent Document 2] Patent Publication No. 2021-056153 [Patent Document 3] US2021 / 0232917A1 [Non-patent literature]

[0006] [Non-Patent Document 1] Masanao Natsumeda, Haifeng Chen, “RULENet:End-to-end Learning With the Dual-estimator for Remaining Useful Life Estimation”, 2020 IEEE International Conference on Prognostics and Health Management (ICPHM), June 8-10, 2020. Summary of the Invention [Problem to be solved by the invention]

[0007] A feature indicating the remaining life span may appear in various forms in time series data acquired by a sensor. For example, a feature indicating the remaining life span may appear as a gradual trend over a long period of time in the time series data. Alternatively, a feature indicating the remaining life span may appear as a short-term change in the time series data. Therefore, it may be difficult to identify the information necessary to predict the remaining life span. As a result, the first and second related techniques related to the present invention may not be able to accurately predict the remaining life span. Similar issues may also arise when predicting conditions other than the remaining life span of a device (e.g., the presence or absence of anomalies, fault diagnosis, degradation state, etc.).

[0008] An object of the present invention is to provide an information processing system, an information processing method, and a recording medium that solve the above-mentioned problems. [Means for solving the problem]

[0009] An information processing system according to an embodiment of the present invention includes: a learning unit that generates a trained model that predicts a state of a device from time-series data acquired from the device; The trained model is a first component that extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis, and generates a plurality of first vectors that embed the extracted features and correspond one-to-one to the partial time series data; a second component that generates a second vector in which the plurality of first vectors are embedded; a third component that extracts permutation-dependent features from the plurality of first vectors and generates a third vector in which the extracted features are embedded; a fourth component that generates a fourth vector by embedding the second vector and the third vector; a fifth component that converts the fourth vector into a first value that represents a state of the device. It is structured as follows.

[0010] An information processing system according to another aspect of the present invention includes: a prediction unit that predicts a state of the device from time-series data acquired from the device using a trained model; The trained model is a first component that extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis, and generates a plurality of first vectors that embed the extracted features and correspond one-to-one to the partial time series data; a second component that generates a second vector in which the plurality of first vectors are embedded; a third component that extracts permutation-dependent features from the plurality of first vectors and generates a third vector in which the extracted features are embedded; a fourth component that generates a fourth vector by embedding the second vector and the third vector; a fifth component that converts the fourth vector into a first value that represents a state of the device. It is structured as follows.

[0011] An information processing method according to another aspect of the present invention includes: Using the trained model, predict the state of the device from time series data acquired from the device; In the prediction, the trained model extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; It is structured as follows.

[0012] An information processing method according to another aspect of the present invention includes: generating a trained model that predicts the state of a device from time-series data acquired from the device; In the generation, the trained model is extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; It is structured as follows.

[0013] A computer-readable recording medium according to another aspect of the present invention includes: A program for causing a computer to perform a process of generating a trained model that predicts a state of a device from time series data acquired from the device, In the generation, the trained model is extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; It is configured to record a program.

[0014] A computer-readable recording medium according to another aspect of the present invention includes: A program for causing a computer to perform a process of predicting a state of a device from time series data acquired from the device using a trained model, In the prediction, the trained model extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; It is configured to record a program. [Effects of the Invention]

[0015] By having the above-described configuration, the present invention can accurately predict the state of a device from time-series data acquired from the device. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a block diagram of an information processing apparatus according to a first embodiment of the present invention. [Figure 2] 5 is a flowchart showing an example of an operation in a learning phase of the information processing device according to the first embodiment of the present invention. [Figure 3] 5 is a flowchart showing an example of an operation in a prediction phase of the information processing device according to the first embodiment of the present invention. [Figure 4] FIG. 1 is a configuration diagram showing an example of a model used in a first embodiment of the present invention. [Figure 5] FIG. 3 is a diagram showing an example of a function for calculating a weighted sum and a function for giving weights used in the first embodiment of the present invention. [Figure 6] 4 is a flowchart showing details of a process for generating a trained model using multivariate time-series data for training in the first embodiment of the present invention. [Figure 7] 10 is a flowchart showing details of a process for predicting the remaining life of equipment using a trained model in the first embodiment of the present invention. [Figure 8] FIG. 10 is a configuration diagram showing an example of a model used in a second embodiment of the present invention. [Figure 9] FIG. 10 is a configuration diagram showing an example of a model used in a third embodiment of the present invention. [Figure 10] FIG. 13 is a block diagram showing an example of a double estimation model used in a fourth embodiment of the present invention. [Figure 11] FIG. 13 is a block diagram of an information processing system according to a seventh embodiment of the present invention. [Figure 12] FIG. 13 is a block diagram of an information processing system according to an eighth embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] Next, an embodiment of the present invention will be described in detail with reference to the drawings. [First embodiment] 1 is a block diagram of an information processing device 10 according to a first embodiment of the present invention. This information processing device 10 is a device that predicts the remaining life of a device 17 based on multiple pieces of time-series data collected from the device. However, the present invention may also be configured to predict the remaining life of the device 17 based on a single piece of time-series data collected from the device.

[0018] Referring to FIG. 1, an information processing device 10 includes a device I / F (interface) unit 11, a communication I / F unit 12, an operation input unit 13, a screen display unit 14, a storage unit 15, and an arithmetic processing unit 16.

[0019] The device I / F unit 11 is connected to the device 17 via a wired or wireless connection. The device 17 is an industrial device whose remaining life is to be predicted. The type of the device 17 is not limited. The device 17 is provided with one or more sensors 18. The type and number of sensors 18 are not limited. For example, the sensor 18 may be a sensor that measures vibrations generated in response to the operation of the device 17. Alternatively, the sensor 18 may be a sensor that measures the temperature of the device 17. Alternatively, the sensor 18 may be a sensor of a type other than those described above, such as a humidity sensor, a pressure sensor, a flow sensor, an acceleration sensor, a displacement sensor, a power sensor, a current sensor, or an acoustic sensor. Measurement by the sensor 18 is not limited to continuous measurement, and may also be measurement at regular time intervals. The device I / F unit 11 acquires a time series of measurement values measured continuously or periodically at the same timing by one or more sensors 18 and transmits the time series to the calculation processing unit 16.

[0020] The communication I / F unit 12 is composed of a data communication circuit and performs data communication with an external device (not shown) via wired or wireless communication. The operation input unit 13 is composed of operation input devices such as a keyboard and a mouse, and detects operator operations and outputs them to the calculation processing unit 16. The screen display unit 14 is composed of a screen display device such as an LCD (Liquid Crystal Display), and displays various information such as prediction results in response to instructions from the calculation processing unit 16.

[0021] The storage unit 15 is composed of one or more storage devices such as a hard disk or memory, and stores processing information and a program 151 required for various processes in the arithmetic processing unit 16. The program 151 is a program that is read into the arithmetic processing unit 16 and executed to realize various processing units, and is read in advance from an external device or recording medium (not shown) via a data input / output function such as the communication I / F unit 12 and stored in the storage unit 15. The main processing information stored in the storage unit 15 includes multivariate time-series data for learning 152-1, multivariate time-series data for prediction 152-2, a model before learning 153-1, a trained model 153-2, and prediction result information 154.

[0022] The learning multivariate time series data 152-1 and the prediction multivariate time series data 152-2 include time series data of measurement values for each sensor acquired from one or more devices 17. Hereinafter, multivariate time series data is assumed to be composed of n pieces of time series data (n is a positive integer equal to or greater than 2). The learning multivariate time series data 152-1 is created in advance based on data from a healthy state to the time when a failure occurs for a large number of devices 17 (also called run-to-failure data). The learning multivariate time series data 152-1 may be data from a healthy state to an unhealthy state for a large number of devices 17, or may be data from immediately after maintenance to immediately before maintenance. There are generally many pieces of learning multivariate time series data 152-1. Each learning multivariate time series data 152-1 further includes ground truth data. The correct answer data is data indicating the correct result of the remaining lifespan prediction using the multivariate time-series data for learning 152-1, and is the actual measured value of the remaining lifespan. On the other hand, the multivariate time-series data for prediction 152-2 is data from the healthy state of the equipment 17 to be predicted up to the time of prediction.

[0023] Both the pre-learning model 153-1 and the trained model 153-2 are machine learning models. The pre-learning model 153-1 uses multivariate time series data for learning 152-1 to learn parameters such as weights so as to predict the remaining lifespan of the equipment 17 from the multivariate time series data. Once the parameters of the pre-learning model 153-1 have been learned, the model is saved as trained model 153-2. The trained model 153-2 is used to predict the remaining lifespan of the equipment 17 to be predicted using the multivariate time series data for prediction 152-2.

[0024] The prediction result information 154 is information representing the result of prediction from the prediction multivariate time series data 152 using the trained model 153-2. The prediction result information 154 includes the remaining lifespan of the device 17. The remaining lifespan represents the remaining lifespan of the device 17 at the end time of the input multivariate time series data.

[0025] The arithmetic processing unit 16 has one or more processors such as an MPU and its peripheral circuits, and by reading and executing a program 151 from the storage unit 15, the hardware and the program 151 work together to realize various processing units. The main processing units realized by the arithmetic processing unit 16 include an acquisition unit 161, a learning unit 162, a prediction unit 163, and an output unit 164.

[0026] The acquisition unit 161 acquires time series data of measurement values of multiple sensors 18 attached to one or more devices 17 via the device I / F unit 11 and / or the communication I / F unit 12, and stores the data in the memory unit 15 as multivariate time series data for learning 152-1 or multivariate time series data for prediction 152-2.

[0027] The learning unit 162 uses the learning multivariate time series data 152-1 to perform machine learning on the pre-learning model 153-1 so that the pre-learning model 153-1 predicts the remaining lifespan of the equipment from the multivariate time series data. Then, the learning unit 162 stores the machine-learned model 153-1 as a trained model 153-2 in the storage unit 15. That is, the learning unit 162 generates a trained model 153-2 for predicting the remaining lifespan of the equipment 17 from the prediction multivariate time series data 152-2.

[0028] The prediction unit 163 uses the trained model 153-2 to predict the remaining lifespan of the device 17 from the multivariate time-series data for prediction 152-2 acquired from the device 17. The prediction unit 163 stores prediction result information 154 including the predicted remaining lifespan of the device 17 in the storage unit 15.

[0029] The output unit 164 reads out the prediction result information 154 including the remaining lifespan of the equipment 17 predicted by the prediction unit 163 from the memory unit 15, displays it on the screen display unit 14, and / or transmits it to an external device via the communication I / F unit 12.

[0030] Next, the operation of the information processing device 10 will be described. The operation of the information processing device 10 is roughly divided into a learning phase and a prediction phase. The learning phase is a phase in which machine learning is performed on the pre-learning model 153-1 to generate a trained model 153-2. The prediction phase is a phase in which the trained model 153-2 is used to predict the remaining life of the device 17 and output the result.

[0031] 2 is a flowchart showing an example of the operation of the learning phase. Referring to FIG. 2, first, the acquisition unit 161 acquires learning multivariate time-series data 152-1 from an external device, for example, via the communication I / F 12, and stores it in the storage unit 15 (step S1). Next, the learning unit 162 performs machine learning on the pre-learning model 153-1 using the learning multivariate time-series data 152-1, thereby generating a trained model 153-2 (step S2). Next, the learning unit 162 stores the trained model 153-2 in the storage unit 15 (step S3).

[0032] FIG. 3 is a flowchart showing an example of the operation of the prediction phase. Referring to FIG. 3, first, the prediction unit 163 reads the trained model 153-2 from the storage unit 15 (step S11). Next, the acquisition unit 161 acquires the multivariate time-series data for prediction 152-2 from the device 17 to be predicted, for example, via the device I / F unit 11, and stores the data in the storage unit 15 (step S12). Next, the prediction unit 163 predicts the remaining lifespan of the device 17 from the multivariate time-series data for prediction 152-2 using the trained model 153-2, and stores prediction result information 154 including the remaining lifespan in the storage unit 15 (step S13). Next, the output unit 164 reads the prediction result information 154 from the storage unit 15 and determines whether the remaining lifespan has fallen below a preset threshold (step S14). Next, if the remaining life is below the threshold, the output unit 164 displays an alarm and a predetermined countermeasure on the screen display unit 14 and / or transmits the same to an external device via the communication I / F unit 12 (step S15). The predetermined countermeasure may be, for example, an instruction to maintain or replace the device 17.

[0033] Next, the configuration and operation of each part of the information processing device 10 will be described in detail.

[0034] First, a detailed description will be given of configuration examples of the pre-learning model 153-1 and the trained model 153-2. Hereinafter, when there is no particular distinction between the pre-learning model 153-1 and the trained model 153-2, they will simply be referred to as models 153. Furthermore, when there is no particular distinction between the training multivariate time series data 152-1 and the prediction multivariate time series data 152-2, they will simply be referred to as multivariate time series data 152.

[0035] Information features suggesting remaining life that appear in time-series data of sensor measurements can be broadly divided into the following two types. One type is a feature that appears as a gradual trend over a long period of time in the time-series data. Such features are referred to herein as long-term features. For example, an upward or downward trend in the measurements of a specific sensor (e.g., a temperature sensor) is an example of a long-term feature. The other type is a feature that appears as a short-term change in the time-series data. Such features are referred to herein as short-term features. For example, short-term, intense fluctuations in the sensor measurement data or a sudden drop or rise in the measurement data over a short period of time are examples of short-term features. Model 153 is trained to extract such long-term and short-term features separately from the time-series data and predict remaining life based on them.

[0036] 4 is a configuration diagram showing an example of the model 153. The model 153 in this example is made up of five components 21 to 25.

[0037] Component 21 inputs m pieces of partial multivariate time series data 171 to 17m obtained by dividing multivariate time series data 152 into m pieces (m is a positive integer of 2 or more) along the time axis from outside model 153. Multivariate time series data 152 includes n pieces of time series data. Therefore, one piece of partial multivariate time series data 15i (i=1, . . . , m) includes n pieces of partial time series data.

[0038] The multivariate time series data 152 may be divided by any of the following methods. (a) Divide into equal intervals. (b) If the device 17 has a repetitive operation, the device 17 is divided into one cycle of the repetitive operation. Alternatively, if the device 17 has a repetitive operation, the section of one cycle cut out may be further divided into a plurality of sections at equal intervals. (c) The lengths of all the divided partial multivariate time series data 171 to 17m do not have to be the same (partial multivariate time series data of different lengths may be included). (d) After division, there may be overlaps among the multiple partial multivariate time series data 171 to 17m. However, since the component 23 at the rear stage of the model 153 learns the temporal dependency between the column vectors, it is preferable that there are no overlaps in order to reduce the amount of calculation. (e) A plurality of partial multivariate time series data 171 to 17m may be extracted from only a portion of the multivariate time series data 152. For example, when dividing the data into equal intervals for convenience, it is possible to discard the remainder. (f) Multivariate time series data 152 that has been padded may be used as the division target. For example, when dividing at equal intervals for convenience, it is possible to add values immediately before or after the multivariate time series data 152 before division so as to avoid any remainders. (g) A plurality of partial multivariate time series data 171 to 17m obtained by dividing in advance may be input. (h) The multivariate time series data 152 may be signal-processed. For example, not only the time domain information, which is the measurement value in the multivariate time series data 152, but also the frequency domain information obtained by Fourier transforming the time domain information, or the quefrency domain information obtained by further Fourier transforming the frequency domain information may be subject to division.

[0039] The component 21 extracts various features that depend on the permutation (order) of the data constituting each of the m pieces of partial multivariate time series data 171 to 17m. Examples of features that depend on the permutation include, but are not limited to, the following: (a) Statistical values of the difference between previous and next data (average, variance, maximum value, minimum value, etc.) (b) Calculus values (slope, area, etc.) (c) Peak detection values (number of peaks, peak height, etc.) (d) Frequency components obtained by Fourier transform, etc. (e) Pattern as a waveform

[0040] Permutation-dependent features may become different when time series data is rearranged. For example, permutation-dependent features may be different between time series data (D1, D2, D3) and the rearranged time series data (D2, D1, D3).

[0041] However, the features extracted by the component 21 are not limited to permutation-dependent features. The component 21 may further extract permutation-independent features from each of the m partial time-series data 171 to 17m. Examples of permutation-independent features include, but are not limited to, the following: (f) Statistics (mean, variance, maximum, minimum, etc.)

[0042] The features extracted by component 21 are short-term features because they are extracted from individual partial multivariate time series data 171-17m obtained by dividing multivariate time series data 152 along the time axis. The same type of short-term features extracted from different partial multivariate time series data are managed in association with the time information of the partial multivariate time series data from which they were extracted.

[0043] Furthermore, the component 21 generates column vectors 181-18m from the extracted short-term features. That is, the component 21 generates column vectors 181-18m into which the extracted short-term features are embedded. The column vectors 181-18m correspond one-to-one to the partial multivariate time series data 171-17m. For example, the component 21 generates one column vector in which each short-term feature extracted from one partial multivariate time series data is used as each vector element.

[0044] The component 21 having the above functions may be realized by training a neural network such as a recurrent neural network (RNN, LSTM, GRU, etc.), a CNN, or a Transformer. The trained component 21 inputs m pieces of partial multivariate time series data 171-17m constituting the multivariate time series data 152 and extracts short-term features effective for the task (remaining life expectancy prediction in this example) from each piece of partial multivariate time series data. The component 21 then generates and outputs column vectors 181-18m in which the extracted short-term features are embedded. A neural network extracts features by performing nonlinear transformation on the input data. Therefore, it can be said that the component 21, which is configured as a neural network, extracts short-term features by performing nonlinear transformation depending on the permutation on the input time series data.

[0045] Component 22 receives column vectors 181 to 18m from component 21, and generates and outputs intermediate vector 191 in which all of column vectors 181 to 18m are embedded. For example, component 22 may generate intermediate vector 191 as a weighted sum of column vectors 181 to 18m.

[0046] Component 22 is the weighted sum v of column vectors 181 to 18m (k) An example of a function for calculating is shown in Equation 1 in FIG. 5. In Equation 1, k is the number of the original time series data 152, J is the number of input column vectors, and v (k,j) is a column vector, a (k,j) is a column vector v (k,j) The weight of a (k,j) is a column vector v (k,j) Any function whose value is determined depending on

[0047] Weight a (k,j) An example of a function that gives l is shown in Equation 2 in Figure 5. In Equation 2, l kis the number of input column vectors, * is element-wise multiplication, and sigm() is the sigmoid function. Also, W, P, and Q are parameters optimized by learning, W is a vector, and P and Q are matrices. The dimensions of W, P, and Q are the scalar values of a (k,j) It is decided to give it as:

[0048] However, the intermediate vector 191 is not limited to the weighted sum of the column vectors 181 to 18m, but may be the sum, weighted average, or average of the column vectors 181 to 18m.

[0049] In this way, all of the column vectors 181 to 18m are embedded in the intermediate vector 191. As described above, short-term features extracted from the corresponding partial multivariate time series data are embedded in each column vector. Therefore, various short-term features extracted from the original multivariate time series data are embedded in the intermediate vector 191. The weighted sum described above is a linear transformation. Therefore, it can be said that the component 22 generates the intermediate vector 191 by applying a linear transformation that does not depend on the permutation to the multiple column vectors 181 to 18m.

[0050] Component 23 receives column vectors 181-18m from component 21, generates intermediate vector 192, and outputs it. Specifically, component 23 rearranges the received column vectors 181-18m according to the acquisition times of the corresponding partial multivariate time series data 171-17m. Next, component 23 extracts features that depend on the permutation (order) of the column vectors from the rearranged column vectors 181-18m. As described above, short-term features corresponding to the respective acquisition times are embedded in the column vectors 181-18m. Therefore, the features extracted by component 23 are features that appear as gradual trends of short-term features over a long period of time, i.e., long-term features.

[0051] Furthermore, the component 23 generates an intermediate vector 192 from the extracted long-term features. That is, the component 23 generates an intermediate vector 192 into which the extracted long-term features are embedded.

[0052] Component 23 having the above functions may be realized by training a neural network such as a recurrent neural network (RNN, LSTM, GRU, etc.), CNN, or Transformer. Trained component 23 inputs column vectors 181 to 18m and extracts long-term features effective for the task (remaining life prediction in this example) from them. Component 23 then generates and outputs intermediate vector 192 in which the extracted long-term features are embedded. It can be said that component 23, which is configured as a neural network, extracts long-term features by performing a nonlinear transformation that depends on the permutation of the time series of input column vectors.

[0053] Component 24 receives intermediate vector 191 from component 22 and intermediate vector 192 from component 23. Component 24 then generates and outputs feature vector 193, in which intermediate vector 191 and intermediate vector 192 are embedded. For example, component 24 may use a single vector obtained by concatenating intermediate vector 191 and intermediate vector 192 as feature vector 193. Alternatively, component 24 may use a sum of intermediate vector 191 and intermediate vector 192 or a weighted sum calculated in the same manner as component 22 as feature vector 193.

[0054] As described above, various short-term features extracted from multivariate time series data are embedded in intermediate vector 191. Also, as described above, various long-term features extracted from multivariate time series data are embedded in intermediate vector 192. Therefore, various short-term and long-term features are embedded in feature vector 193, into which intermediate vectors 191 and 192 are embedded.

[0055] Component 25 receives feature vector 193 from component 24 and outputs scalar value 194 indicating remaining lifespan. Component 25 may be implemented by training a neural network (e.g., a multi-layer percepton) that converts vectors into scalar values. After training, component 25 receives feature vector 193 and outputs scalar value 194 indicating remaining lifespan. Various short-term and long-term features are embedded in feature vector 193. Therefore, component 25 outputs scalar value 194 indicating remaining lifespan based on the various short-term and long-term features.

[0056] Next, details of step S2 in FIG. 2 executed by the learning unit 162, that is, details of the process of generating the trained model 153-2 using the multivariate time-series data for training 152-1, will be described.

[0057] FIG. 6 is a flowchart showing an example of detailed processing of step S2. Referring to FIG. 6, the learning unit 162 focuses on one piece of learning multivariate time series data 152-1 (step S21). Next, the learning unit 162 uses component 21 to extract short-term features from the multiple pieces of partial multivariate time series data 171-17m that make up the learning multivariate time series data under consideration, and generates multiple column vectors 181-18m into which the short-term features are embedded (step S22). Next, the learning unit 162 uses component 22 to generate one intermediate vector 191 into which all of the short-term features embedded in the multiple column vectors 181-18m are embedded (step S23). Next, the learning unit 162 uses component 23 to extract long-term features that depend on permutations from the multiple column vectors 181-18m, and generates one intermediate vector 192 into which the extracted long-term features are embedded (step S24). Next, the learning unit 162 uses the component 24 to generate a feature vector 193 in which all of the short-term features and long-term features embedded in the two intermediate vectors 191 and 192 are embedded (step S25). Next, the learning unit 162 uses the component 25 to convert the feature vector 193 into a scalar value 194 representing the remaining lifespan of the equipment (step S26). Next, the learning unit 162 adjusts the parameters of the model 153 so as to minimize the difference between the predicted value of the remaining lifespan and the actual measured value of the remaining lifespan given by the ground truth data included in the learning multivariate time-series data 152-1 of interest (step S27).

[0058] Next, the learning unit 162 shifts its attention to the next piece of multivariate time-series data for learning 152-1 (steps S28 and S29), returns to step S22, and repeats the same process as described above using the newly focused multivariate time-series data for learning 152-1. Then, when the learning unit 162 has finished focusing on all pieces of multivariate time-series data for learning 152-1 (Yes in step S29), it ends the process of FIG.

[0059] Next, details of step S13 in FIG. 3 executed by the prediction unit 163, that is, details of the process of predicting the remaining life of the equipment from the prediction multivariate time series data 152-2 using the trained model 153-2, will be described.

[0060] FIG. 7 is a flowchart showing an example of detailed processing of step S13. Referring to FIG. 7, the prediction unit 163 uses the component 21 to extract short-term features from the multiple partial multivariate time series data 171-17m constituting the prediction-use multivariate time series data 152-2, and generates multiple column vectors 181-18m into which the short-term features are embedded (step S31). Next, the prediction unit 163 uses the component 22 to generate a single intermediate vector 191 into which all of the short-term features embedded in the multiple column vectors 181-18m are embedded (step S32). Next, the prediction unit 163 uses the component 23 to extract permutation-dependent long-term features from the multiple column vectors 181-18m, and generates a single intermediate vector 192 into which the extracted long-term features are embedded (step S33). Next, the prediction unit 163 uses the component 24 to generate a feature vector 193 into which all of the short-term features and long-term features embedded in the two intermediate vectors 191 and 192 are embedded (step S34). Next, the prediction unit 163 uses the component 25 to convert the feature vector 193 into a scalar value 194 representing the remaining life of the equipment (step S35).

[0061] As described above, the information processing device 10 according to this embodiment includes a learning unit 162 that generates a trained model 153-2 that predicts the remaining lifespan of the device 17 from multivariate time-series data acquired from the device 17. The trained model 153-2 includes components 21 to 25. Component 21 extracts short-term features, which are permutation-dependent features, from each of a plurality of partial multivariate time-series data 171 to 17m obtained by dividing the multivariate time-series data 152 along the time axis, and generates a plurality of column vectors 181 to 18m into which the short-term features are embedded. Component 22 generates a single intermediate vector 191 into which all of the short-term features embedded in the plurality of column vectors 181 to 18m are embedded. Component 23 extracts long-term features, which are permutation-dependent features, from the plurality of column vectors 181 to 18m and generates a single intermediate vector 192 into which the long-term features are embedded. Component 24 generates feature vector 193 in which all short-term features and long-term features embedded in intermediate vector 191 and intermediate vector 192 are embedded. Component 25 generates and outputs scalar value 194 indicating the remaining lifespan of device 17 from feature vector 193. Therefore, information processing device 10 can predict the remaining lifespan of device 17 with higher accuracy than when only long-term features or only short-term features are used.

[0062] Furthermore, the paths from the input to the output of model 153 include a first path via component 22 and a second path via component 23. The first path is a path when short-term features are important for predicting the remaining lifespan. On the other hand, the second path is a path when long-term features are important for predicting the remaining lifespan. As such, there are paths for efficient learning depending on the features, so learning progresses efficiently, and features that are useful for predicting the remaining lifespan can be efficiently learned. As a result, the accuracy of predicting the remaining lifespan of device 17 can be improved.

[0063] Next, another embodiment of the present invention will be described. [Second embodiment] This embodiment differs from the first embodiment in the configuration of the model 153, but is otherwise the same as the first embodiment.

[0064] Fig. 8 is a configuration diagram showing an example of model 153 used in this embodiment. Model 153 in this example differs from model 153 shown in Fig. 4 in that it further includes components 26 to 28, but is otherwise the same as model 153 shown in Fig. 4.

[0065] Component 26 receives column vectors 181-18m from component 21, generates and outputs difference vectors 201-20m that correspond one-to-one to column vectors 181-18m. The difference vectors are generated as follows: Component 26 first calculates the average vector of all column vectors 181-18m. As described above, short-term features extracted from partial multivariate time series data 171-17m are embedded in column vectors 181-18m. Therefore, the average vector can be said to be the average of the short-term features. Next, component 26 calculates the difference between each column vector and the average vector to generate a difference vector. Therefore, the difference vector represents the difference from the average of each short-term feature extracted from partial multivariate time series data 171-17m.

[0066] Component 27 receives difference vectors 201-20m from component 26, generates intermediate vector 195 in which all of these difference vectors are embedded, and outputs the intermediate vector 195. For example, component 27 may generate a weighted sum of difference vectors 201-20m as intermediate vector 195. The method by which component 27 generates a weighted sum of multiple vectors may be the same as the method by which component 22 generates a weighted sum of multiple vectors. As described above, difference vectors 201-20m represent the differences from the average of each of the short-term features extracted from partial multivariate time series data 171-17m. Therefore, the differences from the average of each of the short-term features extracted from the original multivariate time series data 152 are embedded in intermediate vector 195.

[0067] The component 28 receives the difference vectors 201-20m from the component 26, generates the intermediate vector 196, and outputs it. Specifically, the component 28 rearranges the input difference vectors 201-20m according to the acquisition times of the corresponding partial multivariate time series data 171-17m. Next, the component 28 extracts features dependent on the permutation (order) of the difference vectors 201-20m from the rearranged difference vectors 201-20m. As described above, the difference vectors 201-20m contain embedded differences from the average of short-term features corresponding to their respective acquisition times. Therefore, the features extracted by the component 28 are long-term features, i.e., features that appear as gradual trends over a long period of time in the differences from the average of the short-term features. The component 28 also generates the intermediate vector 196 from the extracted long-term features. That is, the component 28 generates the intermediate vector 196 into which the extracted long-term features are embedded. The component 28 having the above-described functions may be realized by training a neural network such as a recurrent neural network (RNN, LSTM, GRU, etc.), a CNN, or a Transformer.

[0068] Component 24 receives intermediate vectors 191, 192, 195, and 196 from components 22, 23, 27, and 28. Component 24 then generates and outputs feature vector 193, into which intermediate vectors 191, 192, 195, and 196 are embedded. For example, component 24 may use a single vector obtained by concatenating intermediate vectors 191, 192, 195, and 196 as feature vector 193. Alternatively, component 24 may use the sum of intermediate vectors 191, 192, 195, and 196 or a weighted sum calculated in the same manner as component 22 as feature vector 193. Component 25 receives feature vector 193 from component 24 and outputs a scalar value 194 indicating remaining lifespan.

[0069] As described above, the model 153 of this embodiment generates a feature vector 193 in which a feature (a type of short-term feature) representing the difference from the average of short-term features extracted from each of the individual partial multivariate time series data constituting the multivariate time series data is further embedded. The model 153 of this embodiment also generates a feature vector 193 in which a feature (a type of long-term feature) appearing as a gradual trend over a long period of time of the short-term feature representing the difference from the average is further embedded. Therefore, the model 153 of this embodiment can extract a greater variety of short-term and long-term features from multivariate time series data compared to the model 153 of the first embodiment. As a result, this embodiment can more reliably acquire features suggesting the remaining lifespan of equipment that appear in various forms in time series data, thereby improving the accuracy of remaining lifespan prediction.

[0070] As with the first embodiment, various methods for dividing the multivariate time series data 152 in this embodiment are possible, and are not particularly limited. However, in this embodiment, if the device 17 performs a repetitive operation, it is preferable to divide the data into each cycle of the repetitive operation. By dividing the multivariate time series data 152 in this manner, the short-term features generated by the component 26 of the model 153 and embedded in the intermediate vector 195 by the component 27 become short-term features that represent the difference from the average of the short-term features for each repetitive operation of the device 17. Therefore, such short-term features can be acquired as information suggesting the remaining life of the device 17. Furthermore, the long-term features generated by the component 28 and embedded in the intermediate vector 196 become long-term features that appear as a gradual trend over a long period of time of the short-term features that represent the difference from the average of the short-term features for each repetition of the device 17. Therefore, such long-term features can be acquired as information suggesting the remaining life of the device 17.

[0071] [Third embodiment] This embodiment differs from the first embodiment in the configuration of the model 153, but is otherwise the same as the first embodiment.

[0072] Fig. 9 is a configuration diagram showing an example of model 153 used in this embodiment. Model 153 in this example differs from model 153 shown in Fig. 4 in that it further includes components 29 to 31, but is otherwise the same as model 153 shown in Fig. 4.

[0073] Component 29 receives column vectors 181-18m from component 21 and generates and outputs difference vectors 211-21m that correspond one-to-one to column vectors 181-18m. The difference vectors are generated as follows: Component 29 first classifies column vectors 181-18m into a group of odd-numbered column vectors 181, 183, ..., 18m-1 and a group of even-numbered column vectors 182, 184, ..., 18m when sorting the column vectors 181-18m in the time order of the corresponding partial multivariate time series data 171-17m. Next, component 29 calculates the average vector of all column vectors belonging to each group. As described above, short-term features extracted from the partial multivariate time series data 171-17m are embedded in column vectors 181-18m. Therefore, the average vector for each group can be considered the average of the short-term features for each group. Next, component 29 calculates the difference between the column vectors belonging to each group and the average vector, generating a difference vector for each group. Thus, component 29 generates difference vectors 211, 213, . . . , 21m-1 representing the differences between the column vectors belonging to odd-numbered groups and the average vector, and difference vectors 212, 214, . . . , 21m representing the differences between the column vectors belonging to even-numbered groups and the average vector.

[0074] Component 30 receives the difference vectors 211-21m of the odd-numbered and even-numbered groups from component 29, generates an intermediate vector 197 in which all of these difference vectors are embedded, and outputs the intermediate vector 197. For example, component 30 may generate the weighted sum of the difference vectors 211-21m as intermediate vector 197. The method by which component 30 generates the weighted sum of multiple vectors may be the same as the method by which component 22 generates the weighted sum of multiple vectors. As described above, the odd-numbered and even-numbered difference vectors represent the differences from the average of the odd-numbered and even-numbered short-term features extracted from the odd-numbered and even-numbered partial multivariate time series data 171-17m. Therefore, features (which are also short-term features) representing the differences from the average of the odd-numbered and even-numbered short-term features are embedded in intermediate vector 197.

[0075] Component 31 receives the difference vectors of each group from component 29 and generates and outputs intermediate vectors 198. Specifically, component 31 rearranges the received difference vectors 211-21m according to the acquisition times of the partial multivariate time series data 171-17m corresponding to each group. Next, component 31 extracts features dependent on the permutation (order) of the difference vectors for each group from the rearranged difference vectors. As described above, the difference vectors of each group contain embedded differences from the average of the short-term features of each group according to their respective acquisition times. Therefore, the features extracted by component 31 are long-term features, i.e., features that appear as gradual trends over a long period of time in the differences from the average of the short-term features of each group. Furthermore, component 31 generates intermediate vectors 198 from the extracted long-term features of each group. That is, component 31 generates intermediate vectors 198 into which the extracted long-term features of each group are embedded. The component 31 having the above-described functions may be realized by training a neural network such as a recurrent neural network (RNN, LSTM, GRU, etc.), a CNN, or a Transformer.

[0076] Component 24 receives intermediate vectors 191, 192, 197, and 198 from components 22, 23, 30, and 31. Component 24 then generates and outputs feature vector 193, into which intermediate vectors 191, 192, 197, and 198 are embedded. For example, component 24 may use a single vector obtained by concatenating intermediate vectors 191, 192, 197, and 198 as feature vector 193. Alternatively, component 24 may use the sum of intermediate vectors 191, 192, 197, and 198 or a weighted sum calculated in the same manner as component 22 as feature vector 193. Component 25 receives feature vector 193 from component 24 and outputs a scalar value 194 indicating remaining lifespan.

[0077] As described above, the model 153 of this embodiment classifies partial multivariate time series data into odd-numbered and even-numbered groups, and generates, for each group, a feature vector 193 in which a feature (a type of short-term feature) representing a difference from the average of short-term features extracted from each of the partial multivariate time series data belonging to that group is further embedded. The model 153 of this embodiment also generates, for each group, a feature vector 193 in which a feature (a type of long-term feature) appearing as a gradual trend over a long period of time of the short-term feature representing the difference from the average is further embedded. Therefore, the model 153 of this embodiment can extract a greater variety of short-term and long-term features from multivariate time series data than the model 153 of the first embodiment. As a result, this embodiment can more reliably obtain information suggesting the remaining lifespan of equipment that appears in various forms in time series data, thereby improving the accuracy of remaining lifespan prediction.

[0078] As in the first embodiment, various methods for dividing the multivariate time series data 152 in this embodiment are possible, and are not particularly limited. However, in this embodiment, if the device 17 performs a repetitive operation, it is preferable to divide the repetitive operation into cycles and then divide each cycle into a first half and a second half. By dividing the multivariate time series data 152 in this manner, the short-term features generated by the component 29 of the model 153 and embedded in the intermediate vector 197 by the component 30 represent the difference from the average of the short-term features of the first and second half of each repetitive operation of the device 17. Therefore, such short-term features can be acquired as information suggesting the remaining life of the device 17. Furthermore, the long-term features generated by the component 31 and embedded in the intermediate vector 198 represent the long-term gradual trend over a long period of the short-term features representing the difference from the average of the short-term features of the first and second half of each repetition of the device 17. Therefore, such long-term features can be acquired as information suggesting the remaining life of the device 17.

[0079] In the above, the column vectors 181 to 18m are classified into two groups, but they may be classified into three or more groups. For example, if one cycle of operation of the device 17 consists of four steps, i.e., step 1, step 2, step 3, and step 4, the multivariate time series data 152 obtained from the device 17 may be divided into partial multivariate time series data corresponding one-to-one to each step, and the column vectors 181 to 18m may be classified into four groups corresponding to each step.

[0080] [Fourth embodiment] In this embodiment, the present invention is applied to the dual estimation model that performs RUL estimation, which is described in Patent Document 3 and Non-Patent Document 1.

[0081] 10 is a block diagram illustrating the configuration of a dual estimation model to which the present invention is applied. This dual estimation model 300 includes five components 301 to 305. All components 301 to 305 are used in the learning phase, and components 301 and 302 are used in the prediction phase. In Patent Document 3 and Non-Patent Document 1, component 301 is called Tss2Vec, component 302 is called Vec2HI, component 303 is called Tss2Mat, component 304 is called Mat2HIch, and component 305 is called HIch2HI.

[0082] X (k) Let X be the kth example from K run-to-failure data. (k,j) Let be the jth observation in the example, and X (k,jk) Let be the observation at the fault. Here, j denotes the time index of the data up to the fault occurrence, with smaller values indicating older records. lk denotes the length of the time series and the time index at the time of the fault. X (k,lk) is a vector of length A, where A denotes the number of attributes, such as sensors. X (k) A subseries of can be x. (k,j) , 1 st X starting at the time index and ending at the jth time index (k) Let v be a time series of (k,j) is its feature representation, and V (k) =[v (k,1) ,v (k,2) ,···,v (k,jk) ] to X (k) Let x be the overall feature representation. (k,j) begins with an X (k) may be at the beginning of a

[0083] x (k,j) When input, the component 301 converts the feature representation v (k,j) and the component 302 outputs that v (k,j) Based on the remaining life H1 at j (k,j) Output the kth execution data X (k)When input, component 303 outputs V (k) and component 304 outputs V (k) The change point H of HI (health index) ch (k) Finally, component 305 converts it into H ch (k) Based on this, the remaining life at j is H2 (k,j) Output.

[0084] Components 301 and 303 are identical except for their inputs and outputs. Component 301 receives a partial time series x (k,j) Enter the vector v (k,j) Component 303 repeats the process for all j and concatenates all vectors to form a matrix. For example, components 301 and 303 may be configured with components 21 to 24 in FIG. 4. Alternatively, components 301 and 303 may be configured with components 21 to 24 and 26 to 28 in FIG. 8. Alternatively, components 301 and 303 may be configured with components 21 to 24 and 29 to 31 in FIG. 9. Component 302 may be the same as component 25 in FIG. 4.

[0085] Component 304 may be configured with any neural network that converts a group of column vectors into a scalar value when there is a large amount of data. When there is a small amount of data, component 304 is preferably configured to use, for example, a weighted sum of remaining life spans using an attention mechanism. Component 305 calculates the change point H of the health index. ch The remaining life H2 may be calculated from the correct value of the remaining life H1 (estimated by the component 304). For example, the component 305 may be configured with a leaky truncated RUL function or a piece-wise RUL function.

[0086] In the learning phase, the dual estimation model 300 receives as input the data up to the failure X (k) and its subsequence x (k,j)Enter the example and calculate the two RUL estimates H1 (k,j) , H2 (k,j) Then, in the learning phase, the two RUL estimates H1 are used under the condition that the change point of the health index is as large as possible. (k,j) , H2 (k,j) The weights of the parameters of each component of the dual estimation model 300 are adjusted so as to minimize the difference between the values of the parameters and the values of the remaining lifespan at the end of the multivariate time series data. In the prediction phase, the dual estimation model 300 receives the multivariate time series data for prediction as input and outputs the remaining lifespan at the end of the multivariate time series data.

[0087] [Fifth embodiment] In the above embodiment, the information processing device predicts the remaining life of the device 17 based on time-series data of sensor measurement values acquired from the device 17. However, the information processing device may predict the remaining life of the device based on the history of events (time-series data) such as failures and maintenance, instead of or in addition to the time-series data of sensor measurement values. In this case, for example, as shown in FIG. 1 , the device 17 is provided with a storage unit 19 that stores time-series data of events that have occurred in the device 17. The type and number of events are not limited. For example, the event may be related to a failure or maintenance. For example, the time-series data of a failure event includes the event type, the date and time when the failure occurred, the date and time when recovery was achieved, etc. Furthermore, the time-series data of a maintenance event includes the event type, the date and time when maintenance was performed, the maintenance content, etc. The device I / F unit 11 is configured to read the time-series data of one or more events from the storage unit 19 and transmit the data to the calculation processing unit 16.

[0088] [Sixth embodiment] In the above embodiment, the remaining life of a device is predicted. However, the present invention may also be applied to other applications besides remaining life prediction, such as two-class classification for device anomaly detection, and multi-class classification for fault diagnosis and degradation state estimation (discretized remaining life estimation). In this case, one possible implementation method is to set the number of components 25 (corresponding to nodes in the final layer) to the same number as the number of classes, and train a model to minimize cross-entropy as the objective function (loss function). Alternatively, instead of using the components 25 configured as described above, feature vectors may be discriminated using embedding or k-NN within a distance learning framework. In the case of discretized remaining life estimation, a weighted average value using a k-NN average value or kernel may be output.

[0089] [Seventh embodiment] 11 is a block diagram of an information processing system according to a seventh embodiment of the present invention. Referring to FIG. 11, an information processing system 70 according to this embodiment includes a learning unit 72 that generates a trained model 71 that predicts the state of an appliance from time-series data acquired from the appliance.

[0090] The trained model 71 includes a first component that extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing time series data along the time axis, and generates a plurality of first vectors that embed the extracted features and correspond one-to-one to the partial time series data. The trained model 71 also includes a second component that generates a second vector into which the plurality of first vectors are embedded. The trained model 71 also includes a third component that extracts permutation-dependent features from the plurality of first vectors and generates a third vector into which the extracted features are embedded. The trained model 71 also includes a fourth component that generates a fourth vector into which the second vector and the third vector are embedded. The trained model 71 also includes a fifth component that converts the fourth vector into a first value that represents the state of the device.

[0091] The information processing system 70 configured as described above operates as follows: The learning unit 72 generates a trained model 71 that predicts the state of a device from time-series data acquired from the device. In this generation, the learning unit 72 causes the trained model 71 to extract permutation-dependent features from each of a plurality of partial time-series data obtained by dividing the time-series data along the time axis, generate a plurality of first vectors that correspond one-to-one to the partial time-series data and into which the extracted features are embedded, generate a second vector into which the plurality of first vectors are embedded, extract permutation-dependent features from the plurality of first vectors, generate a third vector into which the extracted features are embedded, generate a fourth vector into which the second vector and the third vector are embedded, and convert the fourth vector into a first value that represents the state of the device.

[0092] The information processing system 70 configured and operated as described above can acquire both short-term and long-term features that suggest the state of the equipment, such as the remaining lifespan of the equipment, which may appear in various forms in the time-series data acquired from the equipment, thereby enabling accurate prediction of the equipment state. Furthermore, since the trained model 71 has an efficient learning path based on the short-term and long-term features, learning progresses efficiently, enabling efficient acquisition of features useful for predicting the state of the equipment, such as the remaining lifespan of the equipment. As a result, the accuracy of prediction of the equipment state can be improved.

[0093] [Eighth embodiment] 12 is a block diagram of an information processing system according to an eighth embodiment of the present invention. Referring to FIG. 12, an information processing system 80 according to this embodiment includes a prediction unit 82 that predicts the state of an appliance from time-series data acquired from the appliance using a trained model 81.

[0094] The trained model 81 includes a first component that extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing time series data along the time axis, and generates a plurality of first vectors that correspond one-to-one to the partial time series data, with the extracted features embedded. The trained model 81 also includes a second component that generates a second vector with the plurality of first vectors embedded. The trained model 81 also includes a third component that extracts permutation-dependent features from the plurality of first vectors and generates a third vector with the extracted features embedded. The trained model 81 also includes a fourth component that generates a fourth vector with the second vector and the third vector embedded. The trained model 81 also includes a fifth component that converts the fourth vector into a first value that represents the state of the device.

[0095] The information processing system 80 configured as described above operates as follows: The prediction unit 82 predicts the state of a device from time-series data acquired from the device using the trained model 81. In the prediction, the prediction unit 82 causes the trained model 81 to extract permutation-dependent features from each of a plurality of partial time-series data obtained by dividing the time-series data along the time axis, generate a plurality of first vectors in one-to-one correspondence with the partial time-series data and into which the extracted features are embedded, generate a second vector into which the plurality of first vectors are embedded, extract permutation-dependent features from the plurality of first vectors, generate a third vector into which the extracted features are embedded, generate a fourth vector into which the second vector and the third vector are embedded, and convert the fourth vector into a first value representing the state of the device.

[0096] The information processing system 80 configured and operated in this manner can acquire both short-term and long-term features that suggest the state of the equipment, such as the remaining lifespan of the equipment, which may appear in various forms in the time-series data acquired from the equipment, thereby enabling accurate prediction of the equipment state. Furthermore, since the trained model 81 has an efficient learning path according to the short-term and long-term features, learning progresses efficiently, enabling efficient acquisition of features useful for predicting the state of the equipment, such as the remaining lifespan of the equipment. As a result, the accuracy of prediction of the equipment state can be improved.

[0097] Although the present invention has been described above with reference to the above-mentioned embodiments, the present invention is not limited to the above-mentioned embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention. [Industrial Applicability]

[0098] The present invention can be used in a wide range of fields, including predicting the remaining lifespan and other conditions of various types of equipment, such as machine tools, chemical plants, IT equipment, and semiconductor devices, based on time-series data of sensor measurements obtained from the equipment and time-series data of events recorded in the equipment, and performing predictive maintenance based on the prediction results.

[0099] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes. [Appendix 1] a learning unit that generates a trained model that predicts a state of a device from time-series data acquired from the device; The trained model is a first component that extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis, and generates a plurality of first vectors that embed the extracted features and correspond one-to-one to the partial time series data; a second component that generates a second vector in which the plurality of first vectors are embedded; a third component that extracts permutation-dependent features from the plurality of first vectors and generates a third vector in which the extracted features are embedded; a fourth component that generates a fourth vector by embedding the second vector and the third vector; a fifth component that converts the fourth vector into a first value that represents a state of the device. Information processing system. [Appendix 2] The trained model is a sixth component that generates a plurality of fifth vectors by calculating, for each of the plurality of first vectors, a difference between the first vector and an average vector of the plurality of first vectors; a seventh component that generates a sixth vector in which the plurality of fifth vectors are embedded; an eighth component that extracts permutation-dependent features from the plurality of fifth vectors and generates a seventh vector in which the extracted features are embedded; The fourth component generates the fourth vector in which the sixth vector and the seventh vector are further embedded. 10. The information processing system of claim 1. [Appendix 3] The trained model is a sixth component that divides the plurality of first vectors into a plurality of groups, and generates a plurality of fifth vectors for each group by calculating a difference between the first vector belonging to the group and an average vector of the plurality of first vectors belonging to the group; a seventh component that generates a sixth vector in which the plurality of fifth vectors are embedded; an eighth component that extracts, for each group, permutation-dependent features from the plurality of fifth vectors belonging to the group, and generates a seventh vector in which the extracted features are embedded; The fourth component generates the fourth vector in which the sixth vector and the seventh vector are further embedded. 10. The information processing system of claim 1. [Appendix 4] The trained model is a ninth component including the first component, the second component, the third component, and the fourth component, which receives input of a plurality of partial time series data constituting time series data representing execution data up to an observed state, and generates and outputs a plurality of the fourth vectors corresponding one-to-one to the input plurality of time series data; a tenth component that receives the fourth vectors output from the ninth component and calculates a change point in a health index; and an eleventh component that generates and outputs a second value that is a teacher of the first value based on a change point of the health index. 4. An information processing system according to any one of appendices 1 to 3. [Appendix 5] The first value is a value representing the remaining life of the device. 5. An information processing system according to any one of appendices 1 to 4. [Appendix 6] The first value is a value that indicates whether or not the device has an abnormality, whether or not the device has a malfunction, or whether or not the device is in a deteriorated state. 6. An information processing system according to any one of appendices 1 to 5. [Appendix 7] further comprising an output unit that issues an alarm in response to the first value; 7. An information processing system according to any one of appendices 1 to 6. [Appendix 8] and an output unit that executes a predefined countermeasure on the device in accordance with the first value. 8. An information processing system according to any one of appendices 1 to 7. [Appendix 9] a prediction unit that predicts a state of the device from time-series data acquired from the device using a trained model; The trained model is a first component that extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis, and generates a plurality of first vectors that embed the extracted features and correspond one-to-one to the partial time series data; a second component that generates a second vector in which the plurality of first vectors are embedded; a third component that extracts permutation-dependent features from the plurality of first vectors and generates a third vector in which the extracted features are embedded; a fourth component that generates a fourth vector by embedding the second vector and the third vector; a fifth component that converts the fourth vector into a first value that represents a state of the device. Information processing system. [Appendix 10] generating a trained model that predicts the state of a device from time-series data acquired from the device; In the generation, the trained model is extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; Information processing methods. [Appendix 11] Using the trained model, predict the state of the device from time series data acquired from the device; In the prediction, the trained model extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; Information processing methods. [Appendix 12] A program for causing a computer to perform a process of generating a trained model that predicts a state of a device from time series data acquired from the device, In the generation, the trained model is extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; A computer-readable recording medium on which a program is recorded. [Appendix 13] A program for causing a computer to perform a process of predicting a state of a device from time series data acquired from the device using a trained model, In the prediction, the trained model extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors, each having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector in which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; A computer-readable recording medium on which a program is recorded. [Explanation of symbols]

[0100] 10. Information processing equipment 11 Device I / F section 12 Communication I / F section 13 Operation input section 14 Screen display section 15 Storage section 16 Processing unit 17 Equipment 18 Sensors 19 Memory section 21~25 Components 70 Information Processing Systems 71 trained models 72 Learning Department 80 Information Processing Systems 81 trained models 82 Prediction Department 151 Programs 152-1 Multivariate time series data for learning 152-2 Multivariate time series data for forecasting 153-1 Pre-learning model 153-2 Trained Model 154 Prediction result information 161 Acquisition Department 162 Learning Department 163 Prediction Department 164 Output section 171~17m Partial multivariate time series data 181~18m column vector 191, 192, 195, 196, 197, 198 Intermediate vectors 193 feature vectors 194 Scalar Values 201~20m differential vector 211~21m differential vector 300 Double Estimation Model 301~305 Components

Claims

1. An information processing device including a prediction unit that predicts a state of a device from time-series data acquired from the device, The prediction unit a first process of extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis, and generating a plurality of first vectors having one-to-one correspondence with the partial time series data and embedding the extracted features therein; a second process for generating a second vector in which the plurality of first vectors are embedded; a third process of extracting permutation-dependent features from the plurality of first vectors and generating a third vector in which the extracted features are embedded; a fourth process of generating a fourth vector in which the second vector and the third vector are embedded; and a fifth process of converting the fourth vector into a first value representing a state of the device. Information processing device.

2. The prediction unit a sixth process of generating a plurality of fifth vectors by calculating, for each of the plurality of first vectors, a difference between the first vector and an average vector of the plurality of first vectors; a seventh process of generating a sixth vector in which the plurality of fifth vectors are embedded; an eighth process of extracting permutation-dependent features from the plurality of fifth vectors and generating seventh vectors in which the extracted features are embedded; The fourth process is configured to generate the fourth vector in which the sixth vector and the seventh vector are further embedded.

2. The information processing device according to claim 1.

3. The prediction unit a sixth process of dividing the plurality of first vectors into a plurality of groups, and generating a plurality of fifth vectors for each group by calculating a difference between the first vector belonging to the group and an average vector of the plurality of first vectors belonging to the group; a seventh process of generating a sixth vector in which the plurality of fifth vectors are embedded; an eighth process of extracting, for each group, permutation-dependent features from the plurality of fifth vectors belonging to the group, and generating a seventh vector in which the extracted features are embedded; The fourth process is configured to generate the fourth vector in which the sixth vector and the seventh vector are further embedded. The information processing device according to claim 1 .

4. The prediction unit a ninth process of performing the first process, the second process, the third process, and the fourth process, inputting a plurality of partial time series data constituting time series data representing execution data up to an observed state, and generating and outputting a plurality of the fourth vectors corresponding one-to-one to the input plurality of time series data; a tenth process of inputting the fourth vectors output from the ninth process and calculating a change point of a health index; and (11) an eleventh process of generating and outputting a second value that serves as a teacher for the first value based on the change point of the health index.

4. The information processing device according to claim 1.

5. A method for predicting a state of a device from time-series data acquired from the device, the method comprising: In the prediction, the computer extracts permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; the computer generates a plurality of first vectors, into which the extracted features are embedded, that correspond one-to-one to the partial time-series data; the computer generates a second vector in which the plurality of first vectors are embedded; the computer extracts permutation-dependent features from the plurality of first vectors; the computer generates a third vector in which the extracted features are embedded; the computer generates a fourth vector in which the second vector and the third vector are embedded; the computer converts the fourth vector into a first value representing a state of the device. Information processing methods.

6. In the prediction, a sixth process in which the computer generates a plurality of fifth vectors by calculating, for each of the plurality of first vectors, a difference between the first vector and an average vector of the plurality of first vectors; a seventh process in which the computer generates a sixth vector in which the plurality of fifth vectors are embedded; an eighth process in which the computer extracts permutation-dependent features from the plurality of fifth vectors and generates seventh vectors in which the extracted features are embedded; In generating the fourth vector, the computer generates the fourth vector in which the sixth vector and the seventh vector are further embedded.

6. The information processing method according to claim 5.

7. In the prediction, a sixth process in which the computer divides the plurality of first vectors into a plurality of groups, and generates a plurality of fifth vectors for each group by calculating a difference between the first vector belonging to the group and an average vector of the plurality of first vectors belonging to the group; a seventh process in which the computer generates a sixth vector in which the plurality of fifth vectors are embedded; an eighth process in which the computer extracts, for each group, permutation-dependent features from the plurality of fifth vectors belonging to the group, and generates a seventh vector in which the extracted features are embedded; In generating the fourth vector, the computer generates the fourth vector in which the sixth vector and the seventh vector are further embedded. The information processing method according to claim 5 .

8. A program for causing a computer to perform a process of predicting a state of a device from time-series data acquired from the device, the program comprising: In the prediction, the computer a process of extracting permutation-dependent features from each of a plurality of partial time series data obtained by dividing the time series data into a plurality of parts along a time axis; generating a plurality of first vectors having one-to-one correspondence with the partial time-series data, into which the extracted features are embedded; generating a second vector in which the plurality of first vectors are embedded; extracting permutation-dependent features from the plurality of first vectors; generating a third vector into which the extracted features are embedded; generating a fourth vector in which the second vector and the third vector are embedded; converting the fourth vector into a first value representing a state of the device; A program that performs the following.

9. In the prediction, the computer a sixth process of generating a plurality of fifth vectors by calculating, for each of the plurality of first vectors, a difference between the first vector and an average vector of the plurality of first vectors; a seventh process of generating a sixth vector in which the plurality of fifth vectors are embedded; an eighth process of extracting permutation-dependent features from the plurality of fifth vectors and generating seventh vectors in which the extracted features are embedded; In generating the fourth vector, the computer is caused to perform a process of generating the fourth vector in which the sixth vector and the seventh vector are further embedded. The program according to claim 8.

Images