Method for estimating dynamic characteristic of emotion

Abstract

To perform emotion estimation with high accuracy in a short period of time while using machine learning.SOLUTION: A method for estimating dynamic characteristics of emotions includes the steps of: acquiring biological information of a user; and calculating the biological information of the user by the following formula 1, wherein Bp is arterial blood pressure, hr is a heart rate, P is one-time cardiac output, Rt is a total peripheral resistance, and f is dynamic characteristics of a blood pressure system, and deriving arterial blood pressure of the user. Estimation accuracy can be improved, processing time is shortened and preprocessing of data in a method for estimating emotion characteristics can be performed by limiting parameters representing the dynamic characteristics.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to a method for estimating dynamic characteristics of emotions. [Background technology]

[0002] In recent years, with the aging of the population, maintaining physical health has become a social issue. In addition, in modern society, there are many factors that cause stress, such as work and interpersonal relationships, so maintaining mental health has also become an issue.

[0003] For example, there is a circumplex diagram (see non-patent document 1) that shows the internal mental state related to emotions. In this diagram, the horizontal axis shows pleasure-displeasure and the vertical axis shows wakefulness-sleepiness, with the second and third quadrants showing negative emotions and the first and fourth quadrants showing positive emotions.

[0004] Furthermore, for example, Patent Document 1 discloses a device that acquires from a user biodata including a biosignal of the user and motion data including a motion signal related to the user's motion, and estimates the user's emotions. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2023-89729 A [Non-patent literature]

[0006] [Non-Patent Document 1] JARussell, “A circumplex model of affect”, Journal of Personality and Social Psychology, Vol. 36, pp. 1161-1178 Summary of the Invention [Problem to be solved by the invention]

[0007] However, when it comes to internal states such as human emotions, it is difficult to accurately grasp what emotions a person is currently in and what emotions they will transition to in the future.

[0008] There is a method to measure a person's physiological state using an electroencephalogram sensor or an electric potential sensor to identify the internal state related to emotional changes on the mental side. However, this method focuses on correlating the sensor signal with a subjective scale based on a questionnaire, and does not elucidate the internal mechanism.

[0009] On the other hand, there are also constructive studies that use differential equations to represent the autonomic nervous system and cardiovascular system, which are closely related to mental emotions, and use these models to analyze their characteristics. However, these constructive models have many parameters that represent dynamic characteristics, and because parameter matching is difficult, the theoretical analysis is often qualitative. Therefore, it is difficult to estimate emotions accurately in a short time using methods that use constructive models.

[0010] The present disclosure provides a method for estimating dynamic characteristics of emotions that utilizes machine learning to perform emotion estimation with high accuracy in a short period of time. [Means for solving the problem]

[0011] The method for estimating dynamic characteristics of emotions according to the present disclosure includes a step of acquiring biometric information of a user,

number

[0012] According to the present disclosure, it is possible to provide a method for estimating dynamic characteristics of emotions that utilizes machine learning to perform emotion estimation with high accuracy in a short period of time. [Brief description of the drawings]

[0013] [Figure 1] 1 is a block diagram showing a configuration of an estimation device according to a first embodiment. [Diagram 2] FIG. 2 is a block diagram showing an example of a hierarchical hybrid model according to the first embodiment. [Diagram 3] 6 is a diagram showing the results of LF vs. HF based on real RRI and estimated RRI calculated from heart rates according to the first embodiment. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] First embodiment Hereinafter, an estimation device for dynamic characteristics of emotion according to the present embodiment will be described with reference to the drawings. As shown in Fig. 1, the estimation device for dynamic characteristics of emotion 1 includes a data acquisition unit 11 and a calculation unit 12. The estimation device for dynamic characteristics of emotion 1 is a device that estimates dynamic characteristics of emotion based on the concept of predictive coding, and enables internal inference of emotion by a hierarchical hybrid model that combines a constructive model and a data-dependent model.

[0015] Figure 2 shows an example of a hierarchical hybrid model. In this hierarchical hybrid model, an estimated signal, which represents the internal state of the brain, is calculated from a sensory signal in response to an event trigger that represents an external stimulus. The sensory signal is the instantaneous heart rate (hereafter referred to as heart rate) calculated from the cardiac potential, and the estimated signal is its estimated value. The lower layers of the hierarchical structure are composed of constructive models that represent the autonomic nervous system, cardiovascular system, etc., and a feedback converter, which is an element that transmits from the sensory estimation error to a sympathetic nerve activity correction value.

[0016] On the other hand, the upper layer of the hierarchical structure is composed of the Echo State Network, which is a type of reservoir computing, the FORCE learning, which is a sequential learner of weight parameters, and the feedforward converter, which takes the event trigger and the perceptual estimation error as input and outputs an estimate of the sympathetic nerve activity correction value. However, in the FORCE learning, the sympathetic nerve activity correction value is used as the target signal and learning is performed online using the sequential least squares method. The feedforward converter is a transmission element from the external stimulus to the bias term of the reservoir computing.

[0017] <Constructive model> Here, we explain the constructive model. In the respiratory system, the respiratory muscles are driven by neural activity from the respiratory center, and intrathoracic pressure changes in response to exercise load and mental stress. In parallel, in the cardiovascular system, the sinoatrial node of the heart is driven by activity of the sympathetic and vagus nerves from the cardiovascular center, forming the heartbeat. This determines the arterial blood flow rate and peripheral resistance of the blood vessels, and this blood pressure is stabilized by two loops, the baroreceptors and vasomotor receptors, while the heart rate and blood pressure are optimally controlled in response to exercise load and mental stress.

[0018] The constructive model used in the emotion dynamics estimation device 1 is based on a compartment-type lumped parameter model, but the nonlinear static and dynamic characteristics of the sinus node and the dynamic characteristics of the blood pressure system can be modified.

[0019] Specifically, in the constructive model, when sympathetic nerve activity fs and vagus nerve activity fv are input and heart rate hr is output, the sinus node model has a nonlinear static characteristic f sta-sinus (f s ,f v ) and the dynamics f due to the Laplace operator s dyn-sinus (s), i.e.,

number

[0020] Here, the static and dynamic characteristics are as follows:

number

[0021] In the arterial blood pressure system, the stroke volume P is calculated by the heart rate hr calculated by the formula (1). 10 By multiplying The total peripheral resistance of the vascular system Rt and the dynamic characteristics of the blood pressure system f dyn-bp Multiply by (s) This is used to derive arterial blood pressure (bp).

number

[0022] In addition, in the hierarchical hybrid model, the lower and upper layers are connected by sympathetic nerve activity. First, the transfer element from the perceptual estimation error to the sympathetic nerve activity correction value is represented by a feedback converter, and the latent variable sympathetic nerve activity is obtained. The perceptual estimation error eh regarding the heart rate hr is expressed as

number

number

number

[0023] <Reservoir Computing> The Echo State Network can be used for reservoir computing. This network is a type of recurrent neural network, but the weight parameters W in ,Wh is a fixed initial random value, and the weight parameter W out The feature of this method is that the computational load is light because only the weight parameter W in ,W h is set as a uniform random number so that the spectral radius is less than or equal to 1. The network connection between the input layer and the hidden layer is expressed as a state equation:

number

number

[0024] In addition, the network dynamics is smoothed by a filter that introduces a leak rate α.

number

[0025] The network connection of the output layer is the input u u (k) and the filtered state x u (k) is expressed by the following equation of state.

number

number

Number

Number

[0026] <FORCE Learning> In related research, the FORCE learning method has been proposed as an online learning based on the recursive least squares method. Here, by introducing the forgetting factor λ, it becomes possible to consider the experience and familiarity in predictive coding. First, the input variables for FORCE learning are converted as follows.

Number

Number

Number

Number

Number

[0027] From equations (14) and (15), the error correction term e c is as follows:

number

number

[0028] Based on the above method, the RRI is calculated from the subject's cardiac potential, converted into heart rate, input into the model to calculate the RRI time series data, and input into the hierarchical hybrid model to estimate the dynamic characteristics of emotions.

[0029] Here, the evaluation of the estimation by the estimation device 1 for estimating dynamic characteristics of emotions is described. For example, the estimation by the estimation device 1 can be evaluated by giving a calculation task or an image stimulation task to the subject. Here, a simple explanation is given of an example of evaluating the estimation by having the subject add pairs of two-digit integers every three seconds as a calculation task.

[0030] In the estimation device 1 for the dynamic characteristics of emotions, the heart rate, which is an external perception, is input to the hierarchical hybrid model to obtain the estimated RRI, which is an internal inference, and learning can be realized by setting the forgetting factor λ of the FORCE learning device. Here, for example, by setting λ to 0.9990, strong learning can be realized.

[0031] Furthermore, the emotion dynamics estimation device 1 infers emotional behavior from this RRI estimate. Here, the LF component (integrated power value of 0.05-0.15 Hz) and HF component (integrated power value of 0.15-0.4 Hz) of the power spectrum obtained from the RRI are associated with comfort / discomfort and wakefulness / drowsiness, respectively, and it can be determined that a large LF component indicates high discomfort, and a large HF component indicates high drowsiness.

[0032] Figure 3(a) shows the results of LF vs. HF based on the actual RRI calculated from the heart rate, which indicates the emotional behavior in response to external perception. On the other hand, Figure 3(b) shows the results of LF vs. HF based on the estimated RRI, which indicates the emotional behavior in response to internal inference. Note that HF ​​is processed using a one-dimensional low-pass filter with a time constant of 20 (s).

[0033] In both figures, black indicates the results during rest, and gray indicates the results during task execution. In both figures, the horizontal axis is shifted to the right during rest (i.e., pleasant), and shifted to the left during task execution (i.e., unpleasant), confirming that emotions change in the pleasant / unpleasant direction before and after the calculation task. Meanwhile, on the vertical axis, the estimated RRI results are spread out downward compared to the actual RRI results, indicating the high resolution of emotional granularity and the possibility of expressing things like drowsiness due to habituation.

[0034] In this way, the emotion dynamic characteristics estimation device 1 can perform internal inference of emotions using a hierarchical hybrid model that combines a constructive model and a data-dependent model. In particular, when performing estimation using the emotion dynamic characteristics estimation device 1, as preprocessing, calculations can be performed to perform preprocessing using arterial blood pressure, heart rate, stroke volume, total peripheral resistance, and dynamic characteristics of the blood pressure system as inputs. In this way, the emotion dynamic characteristics estimation device 1 can improve estimation accuracy and shorten processing time by limiting parameters that represent dynamic characteristics.

[0035] The present invention is not limited to the above-described embodiment, and can be modified as appropriate without departing from the spirit of the present invention. In other words, the above description has been omitted or simplified as appropriate for the purpose of clarity, and a person skilled in the art can easily modify, add, or convert each element of the embodiment within the scope of the present invention. [Explanation of symbols]

[0036] 1 Estimation device 11 Data Acquisition Section 12 Arithmetic section

Claims

[Claim 1] acquiring biometric information of a user; The biometric information of the user [0010] where Bp is arterial blood pressure, hr is heart rate, P is stroke volume, Rt is total peripheral resistance, and f is the dynamic characteristic of the blood pressure system. and deriving the user's arterial blood pressure by calculating A method for estimating the dynamic characteristics of emotions.

Images