Machine learning model generation method and drowsiness index value estimation method

Abstract

To provide a technique for estimating an effect of measures against sleepiness for a railroad driver.SOLUTION: A machine learning model 1 is generated on the basis of training data having, as explanatory variables, 1) whether or not to perform the pointing call, 2) whether or not to repeat from the system, 3) whether or not to pay attention to a matter noticed during driving, and 4) an elapsed time, and 5) whether or not the train is stopped at a station, regarding driving of a railroad driver. Further, the drowsiness index value is estimated by inputting each explanatory variable to the generated machine learning model 1.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a machine learning model generation method for estimating a drowsiness index value. [Background technology]

[0002] There are known techniques for assisting the driver of a vehicle to become alert. Well-known methods for assisting the driver to become alert include applying external stimuli such as sound, vibration, and light (see, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2020-82906 Summary of the Invention [Problem to be solved by the invention]

[0004] In order to effectively support driver alertness, it is necessary to verify whether the method is actually effective. For example, if alertness is supported by encouraging drivers to take active actions, it is necessary to verify how effective those actions actually are in maintaining alertness. In particular, because train driving has unique conditions that differ from car driving, it is unclear whether alertness support methods for car drivers will be effective for train drivers as they are.

[0005] The problem to be solved by the present invention is to provide a technology for estimating the effectiveness of measures to combat drowsiness for railway drivers. [Means for solving the problem]

[0006] The first invention for solving the above problem is a machine learning model generation method that generates a machine learning model based on training data with respect to the driving of a train driver, using 1) whether or not pointing and calling is performed, 2) whether or not the system repeats, 3) whether or not the train takes note of things to be aware of while driving, 4) elapsed time, and 5) whether or not the train is stopped at a station and / or the time elapsed since the train stopped at a station as explanatory variables, and a drowsiness index value as a target variable.

[0007] As another invention, a drowsiness index value estimation method may be constructed in which, with regard to the driving of a train driver, 1) whether or not pointing and calling is performed, 2) whether or not the system repeats, 3) whether or not the driver pays attention to things that he or she notices while driving, 4) elapsed time, and 5) whether or not the train is stopped at a station and / or the time that has elapsed since the train stopped at a station are used as explanatory variables, and the drowsiness index value is used as the objective variable to estimate the drowsiness index value.

[0008] According to the first invention, it is possible to estimate the drowsiness of a railway driver based on 1) whether pointing and calling was performed, 2) whether the system read back, 3) whether attention was paid to matters to be aware of while driving, 4) elapsed time, and 5) whether the train is stopped at a station and / or the time elapsed since the train stopped at a station. In other words, it is possible to estimate how to select and combine 1) to 3) to be effective as a countermeasure against drowsiness for railway drivers, based on railway-specific operating conditions such as 4) and 5).

[0009] A second invention is the machine learning model generation method according to the above invention, wherein the presence or absence of pointing and calling includes a plurality of types of presence or absence of pointing and calling depending on the frequency with which pointing and calling is performed.

[0010] According to the second invention, a machine learning model can be generated that estimates a drowsiness index value by considering whether or not pointing and calling is performed as whether or not multiple types of pointing and calling are performed according to the frequency of performing the pointing and calling. In other words, it is possible to estimate the effectiveness of drowsiness countermeasures when multiple types of pointing and calling are performed as pointing and calling. [Brief explanation of the drawings]

[0011] [Figure 1] A diagram explaining the generation of a machine learning model. [Figure 2] FIG. 10 is a diagram illustrating estimation of a drowsiness index value using a machine learning model. [Figure 3] FIG. [Figure 4] A table showing an example of the results of multiple regression analysis. [Figure 5] FIG. 10 is a diagram showing the results of a comparison between actual measured values ​​and estimated values. [Figure 6] A table showing other examples of the results of multiple regression analysis. [Figure 7] FIG. 2 is a block diagram showing an example of the functional configuration of a processing device. DETAILED DESCRIPTION OF THE INVENTION

[0012] Hereinafter, preferred embodiments of the present invention will be described with reference to the drawings. Note that the present invention is not limited to the embodiments described below, and the forms to which the present invention can be applied are not limited to the following embodiments. In addition, in the description of the drawings, the same parts are given the same reference numerals.

[0013] [About generating machine learning models] This embodiment is intended to estimate the effectiveness of drowsiness countermeasures for railway drivers. FIG. 1 is a diagram illustrating the generation of a machine learning model 1 for estimation. A plurality of training data 3 is used to generate the machine learning model 1. In this embodiment, training data 3 is prepared in which data related to drowsiness countermeasures while driving a railway and their effectiveness, and data on railway-specific operating conditions are used as explanatory variables, and a drowsiness index value is used as a target variable, and the machine learning model 1 is generated by multivariate analysis such as regression analysis or machine learning such as a neural network.

[0014] When driving, train drivers point at objects and call out aloud for safety confirmation and other purposes. This pointing and calling, along with the accompanying conversation (speech), is thought to be effective in maintaining wakefulness. Furthermore, when considering measures to combat drowsiness, it is believed that focusing on inducing proactivity and intrinsic motivation is likely to lead to maintaining wakefulness. Therefore, encouraging drivers to pay attention to anything that bothers or bothers them while driving and to report any that they notice is also thought to be an effective way to combat drowsiness. Furthermore, a tendency unique to train driving is that the scenery while driving is often monotonous, which tends to increase drowsiness over time, but also tends to wake up when the train stops at a station.

[0015] Therefore, the explanatory variables are 1) whether pointing and calling is performed, 2) whether the system reads back, 3) whether attention is paid to things to be aware of while driving, 4) elapsed time, and 5) whether the train is stopped at a station. Explanatory variables 1) to 3) are proposed measures to combat drowsiness for train drivers. 4) and 5) are explanatory variables related to the operating conditions specific to railways. 1) whether pointing and calling is performed can include multiple types of whether pointing and calling is performed depending on the frequency with which pointing and calling is performed. In addition, 2) the system reads back the pointing and calling performed by the train driver, for example, by the system in response to the pointing and calling made by the train driver. Since there is a response to the pointing and calling made by the driver himself, it is expected to have the effect of maintaining wakefulness.

[0016] [Estimation of sleepiness index values] Figure 2 is a diagram explaining the estimation of a drowsiness index value using machine learning model 1. As shown in Figure 2, when estimating a drowsiness index value, the following factors regarding the driving of the train driver to be estimated are input into machine learning model 1 as explanatory variables to estimate the drowsiness index value: 1) whether or not pointing and calling is performed, 2) whether or not the system reads back, 3) whether or not attention is paid to matters to be aware of while driving, 4) elapsed time, and 5) whether or not the train is stopped at a station.

[0017] [Specific examples] A specific example of the machine learning model will now be described. In this example, a regression equation (prediction equation) for drowsiness index values ​​is calculated as a machine learning model by multiple regression analysis.

[0018] (Example 1) 1. Survey Method The survey involved 32 train drivers, who were asked to drive using a driving simulator. Each subject drove for 19 minutes under each of the six test conditions shown in Figure 3 on a monotonous section with six-minute intervals between stations. During this time, subjects were also asked to subjectively rate their drowsiness every minute. For example, subjects were asked to rate their drowsiness over the previous minute on a nine-point scale from 1 to 9, with 1 being "very wide awake" and 9 being "very sleepy / fighting drowsiness," and this was used as a subjective assessment.

[0019] As shown in Figure 3, the survey conditions were six: "control condition," "three-times pointing and calling condition," "five-times pointing and calling condition," "three-times pointing and calling + repeat condition," "five-times pointing and calling + repeat condition," and "free condition." The control condition was a condition in which no pointing and calling was performed, no repeat from the system, and no attention was paid to matters to be aware of while driving. In the three-times pointing and calling condition, the survey subject performed pointing and calling approximately every 150 seconds, without repeat from the system or attention to matters to be aware of while driving. In the five-times pointing and calling condition, the survey subject performed pointing and calling approximately every 75 seconds, without repeat from the system or attention to matters to be aware of while driving. Under each of the three-times pointing and calling condition and the five-times pointing and calling condition, the survey subject drove while pointing and calling at the corresponding frequency. The three-times pointing and calling + repeat condition is a condition in which the survey subject points and calls approximately every 150 seconds, the system repeats the pointing and calling, and no attention is paid to matters to be aware of while driving. The five-times pointing and calling + repeat condition is a condition in which the survey subject points and calls approximately every 75 seconds, the system repeats the pointing and calling, and no attention is paid to matters to be aware of while driving. Under each of the three-times pointing and calling + repeat condition and the five-times pointing and calling + repeat condition, the survey subject points and calls at the corresponding frequency and drives while listening to the system's repeat. The free condition is a condition in which the survey subject receives a pointing and calling and the system repeats the pointing and calling if a pointing and calling is made, and also a condition in which audio recordings are received at any time and the system responds if an audio recording is found. Receiving an audio recording means that attention is paid to matters to be aware of while driving. The content of the system's response here is not particularly limited. Under these conditions, the subjects of the survey drove while pointing and calling out as appropriate, and speaking out loud any points they noticed, with the audio being recorded.

[0020] 2.Analysis method Based on the survey results, a data set of explanatory variables and target variables is prepared. In this example, the explanatory variables are six dummy variables set according to the survey conditions: a 150-second call dummy, a 75-second call dummy, a voice recording dummy, a repeat dummy, elapsed time, and a station stop dummy.

[0021] The "every 150 seconds" dummy indicates whether or not there is a pointing call approximately every 150 seconds, the "every 75 seconds" dummy indicates whether or not there is a pointing call approximately every 75 seconds, the "repeat" dummy indicates whether or not there is a repeat from the system in response to the pointing call, and the "voice recording" dummy indicates whether or not there is a voice recording (whether or not attention was paid to things noticed while driving). For example, in the case of driving under the control conditions in Figure 3, all variables are set to "0." On the other hand, in the case of driving under the "five pointing calls + repeat" condition, the "every 75 seconds" dummy and the "repeat" dummy are set to "1," and all other variables are set to "0." The elapsed time indicates the time elapsed since the start of driving (duration of driving). The "stop at station" dummy indicates whether or not the train is stopping at a station, or whether or not it is one minute after departing from the station. In this example, the distance between stations is 6 minutes, so the station stopping dummy variable is set to "1" if it is 6, 7, 12, 13, 18, or 19 minutes after departure from the first station, and "0" otherwise. The target variable is the average value of the subjective evaluations given every minute by 32 survey subjects under each survey condition during the survey (the average value of the subjective evaluations obtained under the same survey conditions for each time period from the start of operation). No variable selection was performed, and the statistical significance level was set at 5%.

[0022] This resulted in 114 sets of data (six survey conditions x 19 minutes) of explanatory variables and target variables. A portion of the obtained data set was used for verification to verify the estimation accuracy of the machine learning model. To avoid bias in elapsed time and station stop dummies, 24 sets were extracted from the data set for each survey condition, taking into account 4, 8, 12, and 16 minutes after departure, and used for verification. The remaining 90 sets were then used for training, and multiple regression analysis was performed using each of the training data sets as training data.

[0023] Figure 4 shows the results of the multiple regression analysis in this example. The adjusted coefficient of determination is R 2 =0.89, which was significant (p<0.05). The p-value shows that all variables are statistically significant. The partial regression coefficients can be summarized as follows, with other variables held constant:

[0024] First, when the system calls out every 150 seconds, the drowsiness index value is 0.32 lower than when it is not called out. Second, when the system calls out every 75 seconds, the drowsiness index value is 0.50 lower than when it is not called out. Third, when voice recording is performed, the drowsiness index value is 1.36 lower than when it is not recorded. Fourth, when the system reads back, the drowsiness index value is 0.30 lower than when it is not read back. Fifth, when the elapsed time is one minute longer, the drowsiness index value is 0.06 higher. Sixth, when the train is stopped at a station, the drowsiness index value is 0.31 lower than when it is moving. Furthermore, the standard partial regression coefficients showed that of the six explanatory variables, voice recording (attention to things to be aware of while driving) had the greatest influence. The regression equation for the obtained drowsiness index value is shown in Equation (1).

number

[0025] Furthermore, to verify the accuracy of the estimation of drowsiness index values ​​using the generated machine learning model, the objective variable of each dataset used for verification (the average value of the subjective evaluations performed by each survey subject) was used as the actual measured value of the drowsiness index value, and the drowsiness index value was compared with the estimated (predicted) value calculated using regression equation (1) based on the explanatory variables of the corresponding dataset. Figure 5 shows the comparison results, with the horizontal axis representing the actual measured value and the vertical axis representing the estimated value. As shown in Figure 5, the actual measured and estimated drowsiness index values ​​matched well, confirming that estimation was performed with high accuracy. The estimation error was 0.14 on average and 0.36 at most.

[0026] (Example 2) In the above-described specific example 1, two dummy variables (a dummy for calling every 150 seconds and a dummy for calling every 75 seconds) were used as explanatory variables indicating whether pointing and calling was performed, indicating whether two types of dummy variables were performed depending on the frequency of pointing and calling. Alternatively, a single continuous variable may be set for the frequency of pointing and calling. In this case, the explanatory variables are the continuous variable, the calling frequency, the voice recording dummy, the recitation dummy, the elapsed time, and the station stop dummy. The dependent variable is the drowsiness index value. In this example, a survey was conducted in the same manner as in specific example 1, and a multiple regression analysis was performed using the obtained dataset as training data to generate a machine learning model. FIG. 6 is a table showing the results of the multiple regression analysis in this example. This example makes it possible to estimate the alertness effect of pointing and calling when performed at any frequency.

[0027] [Function Configuration] 7 is a block diagram showing an example of the functional configuration of a processing device 10 for generating a machine learning model and estimating a drowsiness index value. As shown in FIG. 7, the processing device 10 includes an operation unit 11, a display unit 12, a communication unit 13, a processing unit 15, and a storage unit 17, and is configured as a type of computer.

[0028] The operation unit 11 is realized by an input device such as a keyboard, mouse, or touch panel, and outputs an operation signal according to an operation input to the processing unit 15. The display unit 12 is realized by a display device such as an LCD (Liquid Crystal Display) or touch panel, and performs various displays according to a display signal from the processing unit 15. The communication unit 13 is realized by a wired or wireless communication device, and communicates with external devices.

[0029] The processing unit 15 is realized by an arithmetic circuit such as a CPU (Central Processing Unit) or a control board including the arithmetic circuit, and performs various arithmetic processes based on programs, data, etc. stored in the storage unit 17 to control the operation of the processing device 10. In this embodiment, the processing unit 15 includes a generation unit 151 and an estimation unit 153.

[0030] The generation unit 151 generates a machine learning model based on training data regarding the driving of a train driver, using 1) whether or not pointing and calling is performed, 2) whether or not the system repeats, 3) whether or not attention is paid to matters of concern while driving, 4) elapsed time, and 5) whether or not the train is stopped at a station as explanatory variables, and using the drowsiness index value as the objective variable (see Figure 1).

[0031] The estimation unit 153 estimates a drowsiness index value by inputting the following explanatory variables into the machine learning model generated by the generation unit 151 regarding the driving of the train driver to be estimated: 1) whether or not pointing and calling was performed, 2) whether or not the system repeated the information, 3) whether or not the driver paid attention to things that he or she noticed while driving, 4) the elapsed time, and 5) whether or not the train was stopped at a station (see Figure 2).

[0032] The storage unit 17 is realized by a storage medium such as an IC memory or a hard disk. The storage unit 17 stores in advance or temporarily stores each time processing is performed programs for operating the processing device 10 and realizing various functions of the processing device 10, data used during execution of the programs, etc. In this embodiment, the storage unit 17 stores a processing program 171 for causing the processing unit 15 to function as the generation unit 151 and the estimation unit 153, multiple data sets 173 of explanatory variables and target variables prepared by a prior investigation, and model data 175 of the machine learning model generated by the generation unit 151.

[0033] As described above, according to this embodiment, it is possible to estimate the drowsiness level of a railway driver based on 1) whether pointing and calling was performed, 2) whether the system reads back, 3) whether attention was paid to matters to be aware of while driving, 4) elapsed time, and 5) whether the train is stopped at a station. In other words, it is possible to estimate how to select and combine 1) to 3) to be effective as a countermeasure against drowsiness for railway drivers, based on railway-specific operating conditions such as 4) and 5).

[0034] In the above embodiment, five explanatory variables are used: 1) whether pointing and calling is performed, 2) whether the system reads back, 3) whether attention is paid to matters of concern during driving, 4) elapsed time, and 5) whether the train is stopped at a station. As the explanatory variable for 5), a dummy variable (stop-at-station dummy) indicating whether the train is stopped at a station or whether it is one minute after departing from the station is used as an example. However, for 5), instead of the dummy variable shown as an example, a continuous variable indicating the elapsed time since the train stopped at a station may be used. Alternatively, the explanatory variables may include both a dummy variable indicating whether the train is stopped at a station and a continuous variable indicating the elapsed time since the train stopped at a station. The elapsed time since the train stopped at a station indicates the elapsed time since the last time the train stopped at a station. This is because a driving tendency specific to railways is that drowsiness increases when the distance between stations is long. This makes it possible to estimate the alertness effect according to the elapsed time since the train stopped at a station.

[0035] Although an example of an embodiment to which the present invention is applied has been described, the forms to which the present invention can be applied are not limited to the above embodiment. For example, in the above embodiment, an example has been described in which a machine learning model is obtained as a regression equation by multiple regression analysis, but as long as it is machine learning by so-called supervised learning, it may be applied to other machine learning such as a neural network to generate a machine learning model, and the generated machine learning model may be used to estimate a drowsiness index value. [Explanation of symbols]

[0036] 1. Machine learning model 10...Processing equipment 15...Processing section 151...Generation section 153...Estimation part 17...Storage section 171...Processing program 173...Dataset 175...Machine learning model data

Claims

1. A machine learning model generation method for generating a machine learning model based on training data in which the following explanatory variables are used regarding the driving of a train driver: 1) whether or not pointing and calling is performed, 2) whether or not the system repeats, 3) whether or not the driver pays attention to things that he or she notices while driving, 4) elapsed time, and 5) whether or not the train is stopped at a station and / or the time that has elapsed since the train stopped at a station, and the drowsiness index value is the objective variable.

2. The presence or absence of pointing and calling includes a plurality of types of presence or absence of pointing and calling depending on the frequency of pointing and calling. The machine learning model generation method according to claim 1 .

3. A drowsiness index value estimation method for estimating a drowsiness index value by inputting the explanatory variables, which are the following, regarding the driving of a train driver: 1) whether or not pointing and calling out is performed, 2) whether or not the system repeats, 3) whether or not the driver pays attention to things that he or she notices while driving, 4) elapsed time, and 5) whether or not the train is stopped at a station and / or the time that has elapsed since the train stopped at a station, into a machine learning model generated based on teacher data in which the drowsiness index value is the objective variable.

Images