Sensor data complementation device, method, and program

Abstract

In one aspect of the present invention, first, in a training phase, a machine learning model is trained such that: first sensor data that does not include a loss on a time series, second sensor data obtained by causing a loss in the first sensor data at random timing on the time series, and third sensor data that has a predetermined covariance relationship with the first sensor data on the time series are acquired; the first, second, and third sensor data are input to the machine learning model in a state in which the time positions on the time series are synchronized; and the loss of the second sensor data is complemented on the basis of the first sensor data and the covariance relationship between the first sensor data and the third sensor data. Then, in an operation phase, sensor data to be analyzed including a loss is input to the trained machine learning model, and the sensor data to be analyzed in which the loss has been complemented is acquired from the machine learning model.

Description

Sensor data complementation device, method, and program

[0001] One aspect of the present invention relates to a sensor data complementation device, method, and program used in the sensing field to complement missing sensor data.

[0002] In recent years, attempts have been made to estimate human states based on biological signals and to utilize the results. For example, pupil diameter is known to reflect a person's attention state in a similar manner to electroencephalograms, while being easier to measure than electroencephalograms. Therefore, it has been proposed to treat pupil diameter as an indicator that objectively and quantitatively reflects a person's attention state and utilize it. Using this method, for example, by using the size and fluctuations of pupil diameter as an indicator of a driver's attention state, it becomes possible to detect a decline in a person's attention state in advance and alert them, thereby reducing the risk of an accident (see, for example, Non-Patent Document 1).

[0003] However, when measuring biological signals, sensor data may be missing due to sensor or human behavior. For example, when a subject blinks, the eyelid between the pupil and the camera closes during the blink, blocking light, resulting in a loss of data during this blocking period. To address this data loss, conventional methods have proposed, for example, a method of using only the time interval excluding the time interval in which sensor data loss occurred as valid data for use in the determination, or a method of interpolating the missing interval by fitting a linear, quadratic, or cubic function.

[0004] Junpei Nishiyama, Koji Tanida, Masashi Kusumi, Yutaka Hirata, "Evaluation of Arousal State Using Pupil Fluctuation as an Index," Journal of Biomedical Engineering, Vol. 46, No. 2, pp. 212-217, April 2008, Internet <URL: https: / / www.jstage.jst.go.jp / article / jsmbe / 46 / 2 / 46_2_212 / _pdf>

[0005] However, with a method that uses only sections of sensor data excluding missing sections as valid data, it is possible to calculate trends in pupil diameter size at any time period and rough trends of increase / decrease over time, but it is not possible to perform frequency analysis over time.Furthermore, with a method that complements missing sections by fitting a linear, quadratic, or cubic function, it is possible to analyze trends in pupil diameter size, rough trends of increase / decrease, and time series features of frequencies larger than the width of the missing section, but it is difficult to analyze time series features of frequencies smaller than the width of the missing section.

[0006] The present invention has been made in light of the above circumstances, and aims to provide a technique that enables missing sections occurring in sensor data to be complemented with high accuracy.

[0007] In order to solve the above problem, one aspect of a sensor data complementation device or method according to the present invention complements missing data in sensor data using a machine learning model. In a learning phase, first, first sensor data that does not include missing data in a time series, second sensor data in which missing data is generated at random times in the time series relative to the first sensor data, and third sensor data that has a predetermined covariance relationship with the first sensor data in the time series are acquired, the first sensor data, the second sensor data, and the third sensor data are input to the machine learning model with their time positions synchronized in the time series, and the machine learning model is trained to complement the missing data in the second sensor data based on the first sensor data and the covariance relationship between the first sensor data and the third sensor data. Then, in an operation phase, sensor data to be analyzed that includes missing data is input to the trained machine learning model, and the sensor data to be analyzed with the missing data complemented is acquired from the machine learning model.

[0008] According to one aspect of the present invention, it is possible to complement missing sensor data to be analyzed by utilizing the sensor data when the missing sensor data does not exist and a predetermined covariance relationship that exists between the missing sensor data and other sensor data. This makes it possible to perform more accurate complementation than when complementing missing sensor data using a prepared linear, quadratic, or cubic function.

[0009] That is, according to one aspect of the present invention, it is possible to provide a technique that enables missing sections occurring in sensor data to be complemented with high accuracy.

[0010] FIG. 1 is a block diagram showing an example of the functional configuration of a sensor data complementation device according to an embodiment of the present invention. FIG. 2 is a block diagram showing an example of the hardware configuration of the sensor data complementation device shown in FIG. 1. FIG. 3 is a diagram showing an example of the functional configuration of a machine learning model and its learning processing unit provided in the sensor data complementation device shown in FIG. 1. FIG. 4 is a diagram showing an example of the architecture configuration of a neural network that constitutes the machine learning model shown in FIG. 3. FIG. 5 is a flowchart showing the processing procedure and the first half of the processing content of the learning processing performed by the learning processing unit of the sensor data complementation device shown in FIG. 1. FIG. 6 is a flowchart showing the processing procedure and the second half of the processing content of the learning processing performed by the learning processing unit of the sensor data complementation device shown in FIG. 1.

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0012] [One Embodiment] (Configuration Example) FIG. 1 is a block diagram showing an example of the functional configuration of a sensor data complementing device CS according to one embodiment of the present invention.

[0013] The sensor data complementation device CS according to one embodiment is, for example, a personal computer, and includes, as functions related to the present invention, a learning model 10, a learning processing unit 20, a complementation processing unit 30, an input processing unit 40, and an output processing unit 50.

[0014] The input processing unit 40 is connected to a pupil diameter sensor SS1, a gaze direction sensor SS2, and an electroencephalogram sensor SS3.

[0015] The pupil diameter sensor SS1 measures the pupil diameter of the left and right eyes in time series based on facial image data of the person being analyzed captured by a camera, for example, and outputs two-dimensional (two series) pupil diameter data SD1 corresponding to the measured left and right eyes.

[0016] The gaze direction sensor SS2 outputs, as gaze direction data SD2, four-dimensional (four-series) time-series data represented by X and Y coordinates for each of the left and right eyes based on facial image data of the subject captured by, for example, a camera. Note that the gaze direction can also be estimated by electrooculography, etc. Furthermore, when a camera is used, the camera used for the gaze direction sensor SS2 can be shared with the camera used for the pupil diameter sensor SS1.

[0017] The brain wave sensor SS3 is, for example, an electroencephalograph (EEG), which measures electrical activity generated in the brain in time series using multiple electrodes attached to the scalp and outputs the measured multiple series of electrical activity signals as brain wave data SD3. Note that near-infrared spectroscopy (NIRS) and magnetoencephalography (MEG) can also be used as the brain wave sensor SS3.

[0018] The input processing unit 40 receives the pupil diameter data SD1, gaze direction data SD2, and brain wave data SD3 output from the pupil diameter sensor SS1, gaze direction sensor SS2, and brain wave sensor SS3, respectively, and downsamples the received pupil diameter data SD1, gaze direction data SD2, and brain wave data SD3 to obtain data with the same time resolution.

[0019] The learning model 10 is configured, for example as shown in FIG. 3, by a generative adversarial network using two discriminatory neural networks called Discriminator A and B and one generative neural network called Generator.

[0020] In the learning phase, the learning processing unit 20 receives the pupil diameter data SD1, gaze direction data SD2, and electroencephalogram data SD3 from the input processing unit 40, and uses the sensor data SD1, SD2, SD3 and the pupil diameter data SD1x, gaze direction data SD2, and electroencephalogram data SD3 that have been intentionally missing for learning purposes to train the learning model 10. An example of the learning process will be described below in the operation example.

[0021] In the operation phase, the complementation processing unit 30 imports pupil diameter data SD1 of the subject of analysis from the input processing unit 40. The complementation processing unit 30 then inputs the imported pupil diameter data SD1 into the learning model 10 as an explanatory variable, obtains complemented pupil diameter data SD1z output from the trained learning model as a target variable, and sends it to the output processing unit 50.

[0022] The output processing unit 50 outputs the pupil diameter data SD1z, with missing parts completed, output from the completion processing unit 30 to an analysis processing unit (not shown) or an external analysis device, for example, to analyze the attention state of the person being analyzed.

[0023] FIG. 2 is a block diagram showing an example of the hardware configuration of the sensor data complementing device CS.

[0024] As shown in FIG. 2, the sensor data complementation device CS includes a control unit 1, to which a storage unit having a program storage unit 2 and a data storage unit 3, a sensor interface (hereinafter, interface will be abbreviated as I / F) unit 4, and an input / output I / F unit 5 are connected via a bus 6.

[0025] The control unit 1 uses a hardware processor such as a central processing unit (CPU), and executes the processes performed by the learning processing unit 20, the complementing processing unit 30, the input processing unit 40, and the output processing unit 50.

[0026] The program memory unit 2 is, for example, a combination of a non-volatile memory such as a HDD (Hard Disk Drive) or SSD (Solid State Drive) as a storage medium that can be written to and read at any time, and a non-volatile memory such as a ROM (Read Only Memory), and stores application programs necessary for the control unit 1 to execute the processing of the learning processing unit 20, the completion processing unit 30, the input processing unit 40, and the output processing unit 50, in addition to middleware such as an OS (Operating System).

[0027] The data storage unit 3 is, for example, a combination of a non-volatile memory such as an HDD or SSD as a storage medium that can be written to and read from at any time, and a volatile memory such as a RAM (Random Access Memory), and its storage area includes an area for storing the learning model 10, an area used by the input processing unit 40 to perform input processing on each of the sensor data SD1, SD2, and SD3, and an area used by the learning processing unit 20 and the completion processing unit 30 to perform learning processing and completion processing, respectively.

[0028] The pupil diameter sensor SS1, gaze direction sensor SS2, and brain wave sensor SS3 are connected to the sensor I / F unit 4, and the sensor I / F unit 4 receives the sensor data SD1, SD2, and SD3 output from the sensors SS1, SS2, and SS3, respectively.

[0029] An input device IN such as a keyboard or a mouse and an output device OU are connected to the input / output I / F unit 5. The input / output I / F unit 5 receives control commands and control information input by an operator via the input device IN, and outputs display data and the like generated by the control unit 1 to the output device OU.

[0030] The input / output I / F unit 5 is provided with a communication interface and also has a function of transmitting the interpolated pupil diameter data to an external analysis device.

[0031] (Example of Operation) Next, an example of operation of the sensor data complementing device CS configured as above will be described.

[0032] (1) Learning Phase Prior to operation, the sensor data complementing device CS causes the learning model 10 to perform the following learning under the control of the learning processing unit 20.

[0033] (1-1) Generation of Learning Data FIG. 5 is a flowchart showing an example of the processing procedure and processing content of the learning data generation processing executed by the input processing unit 40 and the learning processing unit 20.

[0034] For example, when an operator operates the input device IN to input an instruction to execute learning, the control unit 1 of the sensor data complementation device CS receives the pupil diameter data SD1, gaze direction data SD2, and brain wave data SD3 output from the pupil diameter sensor SS1, gaze direction sensor SS2, and brain wave sensor SS3, respectively, via the sensor I / F unit 4. Then, in step S10, the input processing unit 40 first downsamples the pupil diameter data SD1, gaze direction data SD2, and brain wave data SD3 at the same predetermined time resolution.

[0035] Next, in step S11, the control unit 1 of the sensor data complementing device CS divides the sampled pupil diameter data SD1 into a plurality of slices for each eye-opening duration (e.g., several seconds), which is the interval between typical human blinks, under the control of the learning processing unit 20. For example, if the pupil diameter data SD1 consists of two series (two variables) of data for the left eye and the right eye, these two series of pupil diameter data are sliced ​​for each of the eye-opening duration to generate two-dimensional data of a slice size for a time interval that does not include any missing data.

[0036] The learning processing unit 20 also divides the gaze direction data SD2 and the electroencephalogram data SD3 into a plurality of slices for each eye-opening duration, as with the pupil diameter data SD1. For example, for the gaze direction data SD2, four series of time-series data represented by X and Y coordinates for each of the left and right eyes are sliced ​​for each eye-opening duration to generate four-dimensional data of each slice size. For the electroencephalogram data SD3, n series of time-series data, the number of which is equal to the number of electrodes, are sliced ​​for each eye-opening duration to generate n-dimensional data of each slice size.

[0037] Next, in step S12, the learning processing unit 20 acquires pre-stored statistics of typical human blink durations (eye closure durations during which the eyelids are closed by blinking (e.g., several hundred milliseconds)) from the data storage unit 3. The statistics of eye closure durations are calculated, for example, as the average or variance of the eye closure durations. Then, in step S13, the learning processing unit 20 randomly masks (e.g., shown as MC in FIG. 3 ) a portion of the slice of the two-dimensional data of the pupil diameter data SD1 in accordance with the statistics of the blink durations, thereby generating pupil diameter data SD1x with defects.

[0038] (1-2) Learning Using Learning Data Once the learning data has been generated as described above, the learning processing unit 20 uses this learning data to cause the learning model 10 to perform learning as follows.

[0039] FIG. 6 is a flowchart showing an example of the procedure and content of the learning process executed by the learning processing unit 20.

[0040] Fig. 3 shows the functional configuration of the learning model 10 and the input and output of learning data. As shown in Fig. 3, the learning model 10 is composed of a generative adversarial network that combines one neural network 11 for data generation called the Generator and two neural networks 12 and 13 for judgment called Discriminators A and B.

[0041] First, in step S21, the learning processing unit 20 inputs the above-mentioned pupil diameter data SD1x of a slice size with a partial defect (MC), the gaze direction data SD2 of the slice size, and the brain wave data SD3 to the Generator 11 of the learning model 10.

[0042] Here, the two variables of pupil diameter data SD1x have a strong covariance relationship between them. That is, when the left pupil diameter is large, the right pupil diameter also tends to be large, and conversely, when the left pupil diameter is small, the right pupil diameter also tends to be small. Similarly, the gaze direction data SD2 and the electroencephalogram data SD3 also have a strong covariance relationship between them. On the other hand, there is only a weak covariance relationship between the two variables of pupil diameter data and the variables of gaze direction and electroencephalogram data SD3.

[0043] The Generator 11 has an architecture that takes into account the covariance relationships between the sensor data described above. Specifically, as shown in FIG. 4 , the Generator 11 includes a convolutional layer L1, a fully connected layer L2, and a deconvolutional layer L3. First, in the convolutional layer L1, the Generator 11 separately convolves the input pupil diameter data SD1x, slice-sized gaze direction data SD2, and electroencephalogram data SD3. Next, in the fully connected layer L2, the Generator 11 combines the convolved pupil diameter data SD1x with the slice-sized gaze direction data SD2 and electroencephalogram data SD3. Finally, in the deconvolutional layer L3, the Generator 11 deconvolves the combined data to generate pupil diameter data SD1z in which the missing data has been complemented.

[0044] In step S22, the learning processing unit 20 inputs the pupil diameter data SD1z, gaze direction data SD2, and electroencephalogram data SD3 generated by the generator 11 to the discriminators A and B 12 and 13. At this time, the learning processing unit 20 inputs the entire slice intervals of the pupil diameter data SD1z, gaze direction data SD2, and electroencephalogram data SD3 to the discriminator A 12, and inputs the complementary slice intervals of the pupil diameter data SD1z, gaze direction data SD2, and electroencephalogram data SD3 to the discriminator B 13, as shown in FIG.

[0045] That is, the pupil diameter data SD1z, the gaze direction data SD2, and the electroencephalogram data SD3 are divided into the entire slice interval and the masking interval MC to be complemented, and are input to the discriminators A and B 12 and 13 in a divided manner.

[0046] In step S23, the learning processing unit 20 trains the discriminators A and B 12 and 13 based on the interpolated pupil diameter data SD1z, gaze direction data SD2, and electroencephalogram data SD3 generated by the generator 11 so as to output a generated data judgment value (for example, a binary "0") when this generated data is input.

[0047] Each of the discriminators A, B 12, 13 is composed of a convolutional layer and a fully connected layer. The discriminators A, B 12, 13 convolve the input pupil diameter data SD1z, gaze direction data SD2, and electroencephalogram data SD3 in the convolutional layer, and then combine the processed data in the fully connected layer. That is, the discriminators A, B 12, 13 perform learning using the backpropagation method.

[0048] Furthermore, in step S24, the learning processing unit 20 inputs the pupil diameter data SD1, gaze direction data SD2, and electroencephalogram data SD3 generated by the input processing unit 40, i.e., the unmasked actual data, to the Discriminators A and B 12 and 13. At this time, similar to the case of the generated data described above, the learning processing unit 20 divides the pupil diameter data SD1, gaze direction data SD2, and electroencephalogram data SD3 into the entire slice interval and the masked interval MC to be complemented, and inputs them to the Discriminators A and B 12 and 13.

[0049] Then, in step S25, the learning processing unit 20 trains the discriminators A and B 12 and 13 to output an actual data judgment value (for example, a binary "1") based on the pupil diameter data SD1, gaze direction data SD2, and electroencephalogram data SD3 when these actual data are input.

[0050] Next, in step S26, the learning processing unit 20 inputs the pupil diameter data SD1x with the defects, the gaze direction data SD2, and the electroencephalogram data SD3 to the Generator 11. Then, in step S27, the learning processing unit 20 divides the generated data of the pupil diameter data SD1z, the gaze direction data SD2, and the electroencephalogram data SD3 that have been interpolated by the Generator 11 into the entire slice section and the masking section MC to be interpolated, as described above, and assigns and inputs them to the Discriminators A and B 12 and 13.

[0051] Next, in step S28, the learning processing unit 20 causes the Generator 11 to learn by the error backpropagation method so that when the Discriminators A, B 12, 13 receive the interpolated pupil diameter data SD1z, gaze direction data SD2, and electroencephalogram data SD3, the Generator 11 outputs an actual data judgment value (for example, a binary value of "1").

[0052] Finally, in step S29, the learning processing unit 20 calculates the learning error function of the learning model 10 and proceeds to step S20. In step S20, the learning processing unit 20 compares the learning error function with a threshold value and determines whether the learning error function is equal to or less than the threshold value. If the result of this determination is that the learning error function is not equal to or less than the threshold value, the learning processing sequence of steps S21 to S29 described above is executed again.

[0053] Thereafter, the learning processing unit 20 repeatedly executes the series of learning processes in steps S21 to S29 until the learning error function becomes equal to or less than the threshold value. When the learning error function becomes equal to or less than the threshold value, the parameters of the learned learning model are stored and the learning process ends.

[0054] (2) Operation Phase For example, when analyzing the attention state of the subject while driving, the control unit 1 of the sensor data complementation device CS receives the pupil diameter data SD1 of the subject output from the pupil diameter sensor SS1 via the sensor I / F unit 4. The pupil diameter data SD1 includes data missing due to blinking of the subject.

[0055] The control unit 1 of the sensor data complementation device CS downsamples the acquired pupil diameter data SD1 in the input processing unit 40. Then, the control unit 1 divides the sampled pupil diameter data into slices in the complementation processing unit 30 in the same way as in the learning phase, and inputs the slices as explanatory variables to the trained learning model 10.

[0056] As a result, data in which missing data contained in the pupil diameter data SD1 has been complemented is output from the trained learning model 10. This complemented pupil diameter data takes into account the covariance relationship between gaze direction and electroencephalograms during the learning phase. Therefore, the validity of the complemented pupil diameter of the subject is guaranteed.

[0057] (Effects) As described above, in one embodiment, in the learning phase, a set of actual data of pupil diameter data SD1, gaze direction data SD2, and electroencephalogram data SD3 measured during the same period, and a data set of pupil diameter data SD1x in which a defect corresponding to the blink duration has been randomly introduced into the pupil diameter data SD1, and the gaze direction data SD2 and electroencephalogram data SD3 are used to cause learning model 10, which is a neural network combining Generator 11 and Discriminators A and B 12, 13, to perform learning by backpropagation, thereby creating a trained learning model 10 that can output pupil diameter data SD1z in which the defect in the pupil diameter data SD1x has been complemented, using the pupil diameter data SD1 in combination with the gaze direction data SD2 and electroencephalogram data SD3 that have a covariant relationship with the pupil diameter data SD1.

[0058] Then, in the operational phase, the pupil diameter data obtained from the subject of analysis is preprocessed and then input into the trained learning model, and the pupil diameter data output from the learning model after interpolation processing is provided to an analysis device for analysis of time series features of frequencies correlated with the attention state of the subject of analysis.

[0059] Therefore, missing pupil diameter data caused by blinking or the like is complemented by utilizing the covariance relationship with other sensor data, such as gaze direction and electroencephalograms. This allows for more accurate complementation than when missing pupil diameter data is complemented using a linear, quadratic, or cubic function. As a result, by using the complemented pupil diameter data, it becomes possible to analyze time-series features of frequencies smaller than the width of the missing section, which serve as an index of the subject's attention state. This makes it possible to analyze the subject's attention state with high accuracy based on their pupil diameter.

[0060] [Other Embodiments] (1) In the embodiment, the functions of the sensor data complementing device CS are implemented by a personal computer. However, the functions of the sensor data complementing device CS may be implemented by a server computer located on the Web or in the cloud. Furthermore, when the sensor data complementing device CS is used for the purpose of analyzing the driver's attention while driving, the sensor data complementing device CS may be provided in an in-vehicle device such as a drive recorder.

[0061] (2) In one embodiment, pupil diameter data is used as the sensor data to be analyzed, and gaze direction data and electroencephalogram data are used as other sensor data that have a covariance relationship with the pupil diameter data. However, the other sensor data is not limited to this, and only one of the gaze direction data and the electroencephalogram data may be used.

[0062] (3) In one embodiment, a case where a human attention state is analyzed using a biological signal as sensor data has been described as an example. However, the present invention is not limited to this. Vibration data of a monitored device such as a manufacturing facility, a robot, a vehicle, a drone, or a ship may be complemented with other sensor data that has a covariance relationship with the vibration data, and the complemented sensor data may be used to detect an abnormality in the monitored device.

[0063] (4) In addition, the functional configuration of the sensor data complementation device, the control procedures and processing contents for operating the functions, the types and detection methods of each sensor, the architecture of the learning model, and the learning processing algorithms can be modified and implemented in various ways without departing from the spirit of this invention.

[0064] Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in every respect. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, when implementing the present invention, specific configurations according to the embodiments may be appropriately adopted.

[0065] In short, this invention is not limited to the above-described embodiments, and in the implementation stage, the components can be modified and embodied without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0066] CS...sensor data complementation device SS1...pupil diameter sensor SS2...gaze direction sensor SS3...brain wave sensor IN...input device OU...output device 1...control unit 2...program storage unit 3...data storage unit 4...sensor I / F unit 5...input / output I / F unit 6...bus 10...learning model 11...generator 12, 13...discriminator 20...learning processing unit 30...complementation processing unit 40...input processing unit 50...output processing unit

Claims

1. A sensor data completion device that completes missing data in sensor data using a machine learning model, comprising: an acquisition processing unit that acquires first sensor data that does not include missing data in a time series, second sensor data in which missing data is generated in the first sensor data at random times in the time series, and third sensor data that has a predetermined covariance relationship with the first sensor data in the time series; a learning processing unit that inputs the first sensor data, the second sensor data, and the third sensor data to the machine learning model with their time positions in the time series synchronized, and causes the machine learning model to train so as to complete missing data in the second sensor data based on the first sensor data and the covariance relationship between the first sensor data and the third sensor data; and an completion processing unit that inputs sensor data to be analyzed that includes missing data to the trained machine learning model, and acquires the sensor data to be analyzed with the missing data completed from the machine learning model.

2. The sensor data complementation device of claim 1, wherein the machine learning model is constructed so that a second covariance relationship between the multiple data series of the second sensor data and the multiple data series of the third sensor data is weaker than a first covariance relationship between the multiple data series of the second sensor data.

3. A sensor data imputation method for imputing missing data in sensor data using a machine learning model, the sensor data imputation method comprising: a step of acquiring first sensor data that does not include missing data in a time series; second sensor data in which missing data is introduced into the first sensor data at random timings in the time series; and third sensor data that has a predetermined covariance relationship with the first sensor data in the time series; a step of inputting the first sensor data, the second sensor data, and the third sensor data into the machine learning model with their time positions in the time series synchronized, and training the machine learning model so as to imput the missing data in the second sensor data based on the first sensor data and the covariance relationship between the first sensor data and the third sensor data; and a step of inputting sensor data of an analysis target that includes missing data to the trained machine learning model, and acquiring the sensor data of the analysis target with the missing data imputed from the machine learning model.

4. A program that causes a processor included in the sensor data complementation device to execute the processes performed by the acquisition processing unit, the learning processing unit, and the complementation processing unit included in the sensor data complementation device according to claim 1.

Images