KR102450044B1 - Apparatus for estimating impact acceleration magnitude when falling based on inertial information, method thereof and impact protection system having the same - Google Patents

Abstract

The shock acceleration magnitude estimation apparatus according to the present invention is wearable on a user's body and includes an inertial measurement module that collects the user's inertia information in real time, and an artificial neural network, and uses the inertial information as input data of the artificial neural network. and a microcontroller for deriving an estimate of the magnitude of the impact acceleration during a fall using the microcontroller, and a communication module for transmitting the estimated value of the magnitude of the impact acceleration to the outside.

Description

Apparatus for estimating the magnitude of impact acceleration during a fall based on inertial information, a method for estimating the same, and an impact protection system including the same

The present invention relates to an apparatus for estimating the magnitude of impact acceleration during a fall based on inertia information, a method for estimating the same, and an impact protection system including the same, and more particularly, to estimate the magnitude of the impact acceleration prior to a fall impact using inertia information and an artificial neural network. It relates to an impact protection system including an apparatus for estimating the magnitude of an impact acceleration, a method for estimating the magnitude of an impact acceleration, and an apparatus for estimating the magnitude of an impact acceleration.

Evaluating balance ability through postural control in the movement mechanism of the human body is one of the main tasks of biomechanics and is being studied by many clinicians and rehabilitation scientists.

A typical situation that can occur due to loss of balance while walking is a fall, which refers to an injury in which the human body's center of gravity collapses due to carelessness and rapidly falls to the ground, causing the segment of the human body to hit the floor. Falls are more likely to occur when going through an irregular ground or sloped passageway, or when there are obstacles.

In general, falls are known to occur due to a decrease in basic physical strength, postural control, and balance maintenance ability, and the incidence is particularly high in the elderly, in which lower extremity muscle strength deteriorates and joint range of motion is limited due to aging.

According to the World Health Organization (WHO), 32% of people over the age of 70 fall each year, and in the United States, over 30% of people over the age of 65 experience at least one fall per year. .

Injuries to the hip joint due to a fall accident can cause great pain and leave sequelae due to long-term hospitalization. In particular, the 'long-lie' state of staying on the floor for a long time after a fall causes dehydration, bleeding, menstrual and psychological anxiety, and it was found that half of the patients who experienced this had a very high probability of dying within 6 months.

Recently, a wearable protective device has been actively studied to prevent an injury to the human body due to a fall impact. For example, in the case of an air cushion worn on the buttocks, it is possible to protect the buttocks and the hip joint by absorbing the shock when falling. However, the air cushion may inflict a secondary impact on the human body due to recoil after contacting the human body segments when falling. That is, there is a possibility that the wearer of the protective device may be injured by the protective device rather than in a fall accident.

In order for the protective device to provide an optimal protection effect, the absorption performance of the protective device should be controlled according to different fall impacts in various fall accidents. For example, if the air cushion protection device detects the impact acceleration received while a pedestrian falls to the ground and adjusts the inflation rate of the air cushion in advance, the secondary impact suppression effect can be obtained with the optimum shock absorption rate.

However, currently, studies are mainly conducted to predict in advance only whether or not a pedestrian will fall by using the movement of the pedestrian's center of gravity and acceleration and angular velocity data. On the other hand, there are no studies to find out the magnitude of the impact acceleration during a fall, which is an important factor that can be reflected in the design of the protection device. In order to improve the performance of the protective device and reduce injuries caused by falls, it is necessary to study methods and devices for estimating the magnitude of each impact acceleration in various fall accidents.

It is an object of the present invention to provide an apparatus for estimating the magnitude of impact acceleration and a method for estimating the magnitude of the impact that can solve the inconvenience of wearing an inertial sensor on several segments, time-series synchronization between inertial sensors, unit cost, etc. .

It is an object of the present invention to provide an apparatus for estimating the magnitude of an impact acceleration, and a method for estimating the magnitude of the impact acceleration, which can accurately estimate the magnitude of the impact acceleration in advance before the impact of a fall by using inertia information collected in a fall situation.

The technical problem to be achieved by the present invention is to provide an impact acceleration magnitude estimation device capable of estimating the magnitude of the impact acceleration prior to a fall impact, and reflecting this in the impact protection device to increase the impact mitigation effect, and an impact protection system including the same .

The technical problems to be achieved by the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by those of ordinary skill in the art to which the present invention belongs from the description below. There will be.

In order to achieve the above technical problem, an apparatus for estimating the magnitude of impact acceleration according to an embodiment of the present invention includes an inertial measurement module that is wearable on a user's body and collects the user's inertial information in real time, and an artificial neural network. and a microcontroller for deriving an estimate of the magnitude of an impact acceleration during a fall by using the inertia information as input data of the artificial neural network, and a communication module for transmitting the estimated magnitude of the impact acceleration to the outside.

In order to achieve the above technical object, the method for estimating the magnitude of the impact acceleration according to an embodiment of the present invention includes the steps of collecting inertia information and an actual measurement value of the magnitude of the impact acceleration in a falling situation, the inertia information and the magnitude measurement of the impact acceleration preparing a data set using and determining the final artificial neural network according to the result, and estimating the magnitude of the shock acceleration from the artificial neural network.

In order to achieve the above technical problem, an apparatus for estimating the magnitude of an impact acceleration and an impact protection system including the same according to an embodiment of the present invention are wearable on a user's body, and the user's inertia information and artificial neural network measured in real time A shock acceleration magnitude estimation device for deriving an estimate of the magnitude of the shock acceleration before the falling impact in the user's falling situation, and generating a control signal based on the derived estimation value, and the shock absorption ability is adjusted by the control signal Includes shock protection.

According to the embodiment of the present invention, since it is possible to accurately estimate the magnitude of the impact acceleration during a fall using only one inertial sensor, time series synchronization between the inertial sensors that may occur when multiple inertial sensors are worn, discomfort of wearing, and price increase problems can be solved.

In addition, according to an embodiment of the present invention, since machine learning is performed by increasing the number of learning data from inertial information collected using various data processing methods, artificial The performance of the neural network can be sufficiently improved.

In addition, according to an embodiment of the present invention, since the magnitude of the shock acceleration can be estimated in advance before the fall shock using an artificial neural network, the optimal shock absorption rate can be obtained by reflecting this and controlling the shock absorption capacity of the shock protection device. Secondary impact by impact protection device can be reduced.

The effects of the present invention are not limited to the above effects, and it should be understood to include all effects that can be inferred from the configuration of the invention described in the detailed description or claims of the present invention.

1 is a view showing the fall situation step by step.
2 is a diagram illustrating an apparatus for estimating the magnitude of an impact acceleration when falling according to an embodiment of the present invention and an impact protection system including the same.
3 is a flowchart sequentially illustrating a method of constructing an artificial neural network of an apparatus for estimating shock acceleration magnitude according to an embodiment of the present invention.
4 is a flowchart illustrating a subdivided step of preparing the training data set of FIG. 3 .
5 is a diagram illustrating an artificial neural network according to an embodiment of the present invention.
6 is a diagram illustrating a learned artificial neural network according to an embodiment of the present invention.
7 is a view showing the results of estimating the magnitude of the shock acceleration according to the magnitude of the shock of various levels using the artificial neural network constructed according to the embodiment of the present invention.
8 is a diagram illustrating an example of mean percentage absolute errors (MPAEs) of impact acceleration magnitude estimates when each data augmentation technique is applied.
9 is a diagram illustrating an example of mean percentage absolute errors (MPAEs) of an estimate of the magnitude of an impact acceleration that varies according to the number of data sets that are increased using a data augmentation technique.

Hereinafter, the present invention will be described with reference to the accompanying drawings. However, the present invention may be embodied in several different forms, and thus is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is "connected (connected, contacted, coupled)" with another part, it is not only "directly connected" but also "indirectly connected" with another member interposed therebetween. "Including cases where In addition, when a part "includes" a certain component, this means that other components may be further provided without excluding other components unless otherwise stated.

The terms used herein are used only to describe specific embodiments, and are not intended to limit the present invention. The singular expression includes the plural expression unless the context clearly dictates otherwise. In the present specification, terms such as “comprise” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It is to be understood that this does not preclude the possibility of the presence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a view showing the fall situation step by step.

Referring to the figure, the fall situation, including both before and after the fall, has a pre-fall phase, a critical phase, a post-fall phase, and a recovery phase over time. can be divided into 4 stages.

The X-axis represents the passage of time, and the Y-axis represents the acceleration measured by an inertial sensor attached to the user's body.

The critical stage among the four stages refers to a short section from the point in time when the center of gravity of the human body collapses and goes to the ground to the point in time when it hits the floor and receives an impact in the vertical direction.

The pre-fall phase is a period in which normal activity occurs, and the critical phase and post-fall phase are phases in which a fall event actually occurs. The fall event includes the S-stage, which is the start of the fall, the I-stage, which receives an impact from the floor, and the E-stage, which indicates the end of the fall event.

The post-fall stage refers to the stage in which the faller is motionless or lying on the floor, and the recovery stage refers to the stage in which the person stands up.

The acceleration of the human body shown on the Y-axis is the sum of the acceleration magnitude vectors of the three axes (X-axis, Y-axis, Z-axis) measured by the inertial sensor. , which indicates the peak value. The peak value of the sum of the acceleration magnitude vectors is indicated by a star.

Falls can occur while standing or walking, and can also occur in a forward, backward, left, right, or vertical downward direction toward the floor.

In addition, the magnitude of the impact acceleration during a fall impact may vary depending on the characteristics of the person who falls. For example, the impact from the floor depends on the height and weight of the fall.

The present invention classifies the direction of a fall into five categories: the front-back direction, the left-right direction, and the vertical downward direction, accurately estimating the magnitude of the impact acceleration in each direction, and utilizing it in the design and drive control of the impact protection device, thereby preventing injuries and protecting the impact. A secondary impact reduction effect can be obtained by

For example, in the case of an air cushion, which is one of the impact protection devices, by estimating the magnitude of the impact acceleration before the user hits the floor and adjusting the expansion rate of the cushion, the user absorbs the shock received from the floor and at the same time, the recoil of the air cushion Injuries can be prevented by reducing the secondary impact caused by

2 is a diagram illustrating an apparatus for estimating the magnitude of an impact acceleration when falling according to an embodiment of the present invention and an impact protection system including the same.

The impact protection system includes an impact acceleration magnitude estimation apparatus 100 and an impact protection apparatus 200 .

The shock acceleration magnitude estimation apparatus 100 includes an inertial measurement module 110 , a battery 120 , a microcontroller 130 , and a communication module 140 .

The inertia measurement module 110 is configured in a form that can be worn on a part of the human body, and can measure acceleration information and angular velocity information of the wearer in real time. The inertial measurement module 110 is a measurement module including one inertial sensor, and may be an Inertial Measurement Unit (IMU) module.

Since the inertia measurement module 110 measures acceleration information and angular velocity information for each of the X, Y, and Z axes, a total of six pieces of inertia information are collected.

The battery 120 supplies power to the inertial measurement module 110 , the microcontroller 130 , and the communication module 140 provided in the shock acceleration magnitude estimation apparatus 100 .

A microcontroller (MCU) 130 may control the inertia measurement module 110 , and may control transmission/reception between the communication module 140 and the impact protection device 200 .

Then, the microcontroller 130 processes the acceleration information and the angular velocity information using the artificial neural network built therein to estimate the magnitude of the impact acceleration in advance before the fall impact.

The microcontroller 130 inputs six pieces of inertia information (three-axis acceleration information, three-axis angular velocity information) received from the inertia measurement module 110 into the artificial neural network, and estimates the magnitude of impact acceleration for five fall directions as result values. can be obtained with The five falling directions include the device wearer's front, back, left, right, and vertical downward directions (floor direction).

Also, the microcontroller 130 may generate a control signal to control the shock absorption capability of the shock protection device 200 based on the shock acceleration magnitude estimate obtained using the artificial neural network.

The artificial neural network is preferably built in the microcontroller 130 , but may also be built in the impact protection device 200 , and a third data processing device capable of communicating with the impact acceleration magnitude estimation device 100 and the impact protection device 200 . may be built on

The communication module 140 may communicate with the impact protection device 200 by wire or wireless, and may transmit impact acceleration magnitude estimation information or a control signal to the impact protection device 200 .

The communication method applied to the communication module 140 may be any one of various methods such as BLE (Bluetooth Low Energy), Bluetooth (Bluetooth), Wi-Fi, Lora, Zigbee, etc. have.

The shock protection device 200 may be worn on the human body, and may prevent a fall accident by absorbing the shock when the human body falls by using air pressure, hydraulic pressure, or the like.

In the impact protection device 200 , air pressure or hydraulic pressure may be controlled by a control signal of the impact acceleration magnitude estimation device 100 , so that the impact absorption ability may be controlled. Alternatively, the controller may be provided by itself to receive shock acceleration magnitude estimation information from the shock acceleration magnitude estimation apparatus 100 and generate a control signal by itself to control the shock absorption ability.

Various devices such as an air cushion, an air pad, an air jacket, and an exoskeleton robot may be applied to the impact protection device 200 .

In the embodiment of the present invention, only one shock acceleration magnitude estimation apparatus 100 is worn on the human body. Accordingly, the wearer's acceleration information and angular velocity information are collected with only one inertia measurement module 110 .

According to an embodiment of the present invention, a high-accuracy estimate of the magnitude of impact acceleration can be derived by using an artificial neural network only with the acceleration information and angular velocity information collected by one inertial measurement module 110 .

3 is a flowchart sequentially illustrating a method of constructing an artificial neural network of an apparatus for estimating shock acceleration magnitude according to an embodiment of the present invention.

The artificial neural network may be provided in the microcontroller of the apparatus for estimating the magnitude of the impact acceleration, or may be configured as a separate module.

First, in order to build an artificial neural network, data to be used for machine learning of the artificial neural network is collected through an experiment involving many volunteers.

Experimental protocols for data collection should be constructed to mimic frequently occurring realistic drop patterns. The number of activities that subjects participating in the experiment will perform are 10 types of falls. There are 10 types: fall forward, fall backward, fall side to side, fall vertically, fall forward while walking, fall backward while walking, stumble forward and fall backward, slip forward and slide backward. include

The drop data collected by the experiment is classified into five directions. That is, data may be classified into forward, backward, left, right, and vertical downward directions (ground direction) according to the falling direction.

Hereinafter, a method of constructing an artificial neural network will be described in detail with reference to FIG. 3 .

In step S100, inertia information and actual measured values of the impact acceleration are collected. Acceleration information and angular velocity information for each of the X-axis, Y-axis, and Z-axis are collected according to the subjects' activities, and the measured value of the impact acceleration is collected accordingly.

In step S110, a data set is prepared using the collected inertia information and the measured value of the impact acceleration.

A data set consists of a plurality of examples. An example is one row consisting of one or more feature vector values that are input variables used for estimating the magnitude of the impact acceleration and a label that is a result value. That is, an example consists of one or more input variables and output values.

Here, the characteristic vector may be three-axis acceleration information and angular velocity information, and the label indicates an actual measured value of the magnitude of the impact acceleration according to the acceleration information and the angular velocity information.

A more detailed data set preparation process will be described with reference to FIG. 4 .

In step S120, the generated data set is distributed as training data and test data. In this case, the distribution ratio may be set to 1:1, but may be set differently according to circumstances.

Next, a training data set is added using data augmentation technology in step S130. When machine learning is performed using time-series data with a small number of samples, an overfitting problem may occur. Therefore, a training data set may be additionally generated using date augmentation technology.

In an embodiment of the present invention, as a data augmentation technique, a jittering technique that adds mechanical noise to data, a scaling technique that changes the data size by multiplying an arbitrary scale value, and a reduction or stretching of data by reducing or stretching at an arbitrary distortion ratio A time-warping technique for changing the temporal characteristics of the training data by changing the time interval may be used. Here, the jittering technique can increase the robustness of the training model by applying white Gaussian noise to the training data set.

Next, machine learning is performed using the training data in step S140 .

The artificial neural network model constructed in the present invention may be implemented as a Long Short Term Memory (LSTM) architecture. The LSTM architecture operates as a deep learning method that learns data that changes over time, and is an artificial neural network model with high accuracy in predicting time series data as one of the recurrent neural network models that receive information from the previous stage and process it repeatedly. .

The artificial neural network model to which the LSTM architecture is applied will be described later in detail with reference to FIG. 5 .

Next, in step S150, the trained artificial neural network model is evaluated using the test data.

Since the test data includes various acceleration information, angular velocity information, and actual measured values of the magnitude of the shock acceleration, input the acceleration information and angular velocity information to the trained artificial neural network to obtain an estimate of the magnitude of the shock acceleration and the measured values included in the test data set By comparing with , it is possible to evaluate the learning result of the artificial neural network.

In step S160, it is determined whether the artificial neural network model needs to be modified.

If the preset error value is exceeded, it is determined that correction is necessary, so that the artificial neural network is corrected (S170) and machine learning can be performed again (S140). For example, the weight of the hidden layer of the artificial neural network may be modified so that the error between the measured value and the estimated value is reduced.

On the other hand, if the error value between the estimated value and the measured value is less than or equal to the preset error value, the flow advances to step S180 to determine the current artificial neural network as the final artificial neural network, and estimate the magnitude of the shock acceleration therefrom.

4 is a flowchart illustrating a subdivided step of preparing the training data set of FIG. 3 .

Referring to the drawing, in step S111, the received inertia information is normalized by noise filtering. For example, if a signal of a specific cutoff frequency or higher is filtered from the acceleration information and the angular velocity information, a periodic pattern of raw data is maintained while noise of a high frequency component is removed.

In step S112, an additional feature vector is extracted from the inertia information.

More specifically, in order to improve the performance of the impact protection system, four characteristic vectors capable of differentiating the activity of the device wearer are extracted from the triaxial acceleration information and the triaxial angular velocity information.

Since the shock acceleration magnitude estimation device basically acquires 3-axis acceleration information and 3-axis angular velocity information, that is, 6 pieces of information using an inertial sensor, by adding 4 more characteristic vectors, 10 data is the input value of the artificial neural network. do.

The artificial neural network can use 10 input data in the machine learning process to calculate a more accurate estimate of the shock acceleration magnitude through a causal relationship between the data.

The added four characteristic vectors include a sum vector of the three-axis acceleration, a sum vector of the horizontal plane of the accelerometer, a root mean square of the sum vector, and the resulting angular velocity. Here, the acceleration sum vector indicates the square root of the sum of the magnitudes of the three-axis acceleration.

Next, the frame in which the drop impact occurs in step S113 is data-labeled together with the actual measured value of the impact acceleration peak value.

At this time, the three-axis acceleration information, the three-axis angular velocity information, and the added four characteristic vectors become 10 input variables, and data labeling is performed including the measured values of the impact acceleration magnitudes.

Next, in step S114, the data generated so far is divided.

The divided data is divided into training data for machine learning of the artificial neural network and test data for performance evaluation of the trained artificial neural network.

Next, with reference to FIG. 5 , the step S140 of FIG. 3 in which machine learning is performed will be described in detail.

5 is a diagram illustrating an artificial neural network according to an embodiment of the present invention.

The artificial neural network is implemented as a multi-layer perceptron in which at least one hidden layer exists between one input layer and one output layer. Each of the input layer, the hidden layer, and the output layer is connected by a plurality of weighted links, and the weights of the weighted links may be adjusted through learning.

, 망각 게이트(forget gate)

, 셀(cell)

, 출력 게이트(output gate)

로 구성된다.A Long Short Term Memory (LSTM) architecture is used in the present invention to capture and retain an activity sequence from gates and memory cells. The LSTM cell has an input gate as defined in Equation 1 below.

, forget gate

, cell

, output gate

is composed of

,

,

,

는 각각 제품, 시그모이드(sigmoid) 기능,

와

사이의 가중치 행렬(weight matrix) 및

가 있는

의 편향을 나타낸다.here,

,

,

,

is a product, a sigmoid function,

Wow

a weight matrix between and

with

represents the bias of

Referring to the drawing, the artificial neural network is composed of an input layer 310 , a hidden layer 320 , and an output layer 330 .

The input layer 310 receives a time-series collected data set.

The hidden layer 320 includes a bi-directional LSTM layer 322 , a fully connected layer 324 , and a regression layer 326 .

A fully connected layer 324 is where each node is connected to all nodes in subsequent layers.

The regression layer 326 regressively analyzes the input data to derive a functional relationship between the data. Regression analysis is an analysis method for statistically estimating a relationship between input data using a parametric model.

To train the bidirectional LSTM layer 322, the data set may be divided into training data and test data. In this case, the separation ratio may be 1:1, but may be set to another ratio.

A softmax layer may be added to the hidden layer 320 , and a dropout layer may be added to reduce overfitting of the bidirectional LSTM layer. The dropout layer can be connected behind the fully connected layer.

The output layer 330 outputs an estimate of the magnitude of the impact acceleration in the five falling directions derived from the hidden layer 320 .

Table 1 shows hyper parameters that can be set in the bidirectional LSTM architecture.

LSTM architecture training options Num Type of parameters Range of parameters One. Number of hidden units [100,50] 2. Maximum epochs 125 3. Mini-batch size 64 4. Weight initializer function Glorot 5. Solver Adam 6. Dropout rate 0.2 7. Initial learning rate 0.01 8. gradient threshold 2 9. gradient threshold method Global-l2norm 10. L2Regularization 1e^-5

A mean absolute percentage error (MAPE) may be adopted for performance evaluation of the trained artificial neural network. MAPE is used to measure the relative error between the measured value and the estimated value, and can be calculated by Equation 2 below.

는 충격 가속도 크기 실측값을 가리키며,

는 충격 가속도 크기 추정치를 가리킨다. 실측값과 추정치 사이의 관계를 평가하기 위해 상대 오차 결과는 회귀 분석을 이용하여 분석될 수 있다.

denotes the measured value of the impact acceleration magnitude,

is an estimate of the magnitude of the impact acceleration. Relative error results can be analyzed using regression analysis to evaluate the relationship between the measured and estimated values.

6 is a diagram illustrating a learned artificial neural network according to an embodiment of the present invention.

Referring to the figure, the trained artificial neural network is composed of an input layer, a hidden layer, and an output layer.

When the inertia information (3-axis acceleration information and 3-axis angular velocity information) collected from the inertia measurement module is input to the input layer, the hidden layer estimates the impact acceleration magnitudes in five directions based on the learned content using the inertia information and the output layer output through

Hereinafter, with reference to FIGS. 7 to 9 , the performance improvement results of an artificial neural network that can be obtained when machine learning is performed after processing training data by applying data augmentation technology will be described.

7 is a view showing the results of estimating the magnitude of the shock acceleration according to the magnitude of the shock of various levels using the artificial neural network constructed according to the embodiment of the present invention.

The impact acceleration magnitude estimate that can be obtained using an artificial neural network varies depending on the characteristics of the falling person. For example, depending on a person's height and weight, or walking and standing, the time of the critical phase and the magnitude of the impact acceleration vary.

However, the artificial neural network trained according to an embodiment of the present invention can derive an estimate of the magnitude of the shock acceleration close to the actual value even in various human characteristics and in various situations.

Looking at a) to c), when falling, various impacts from the floor are artificially applied to the human body.

In neural network learning, each impact acceleration magnitude estimate with and without data augmentation technology is shown by comparing it with the measured value.

In order to estimate the magnitude of the shock acceleration, time series data (indicated by a thick solid line) near the time when a person loses balance is used as input data for the artificial neural network.

Looking at the graph, in the case where the artificial neural network was trained by applying the data augmentation technology to the input data, the artificial neural network produced an estimate closer to the actual value than the case where the data augmentation technology was not applied.

Here, 4-fold means that the training data set is increased by 4 times using data augmentation technology.

8 is a diagram illustrating an example of mean percentage absolute errors (MPAEs) of impact acceleration magnitude estimates when each data augmentation technique is applied.

Referring to the figure, it can be seen that the average percentage absolute error (MPAE) result between the measured value and the estimated value of the impact acceleration magnitude is different depending on the case where the learning data used for machine learning is processed by data augmentation technology and the case where it is not processed.

Specifically, when data augmentation technology is not applied, MPAE shows a value exceeding 25%, but when machine learning is performed by processing training data using jittering, scaling, or time distortion techniques, MPAE is 20% The following values are shown.

In addition, when the training data is processed by applying all of jittering, scaling, and time distortion techniques, and machine learning is performed using the training data, the performance of the artificial neural network is further improved, resulting in an MPAE of about 10%.

Since a smaller MPAE means that the estimate is closer to the actual value, a smaller MPAE means higher performance of the artificial neural network.

9 is a diagram illustrating an example of the mean percentage absolute errors (MPAEs) of an estimate of the magnitude of an impact acceleration that varies according to the number of data sets that are increased using a data augmentation technique.

Referring to the figure, as the number of training data sets is increased from 1-fold to 4-fold by using data augmentation technology, the performance of the trained artificial neural network also improves and the MPAE becomes smaller. Able to know.

Referring to the results of FIGS. 8 and 9 , it can be seen that the performance of the learned artificial neural network is further improved when the variety and amount of the training data set is increased by processing the training data used for machine learning with data augmentation technology.

Specifically, it can be seen that the performance of the artificial neural network improves as more data augmentation techniques are applied, and the performance of the artificial neural network improves as the training data set increases.

100: shock acceleration magnitude estimation device
110: inertial measurement module
120: battery
130: microcontroller
140: communication module
200: shock protection device

Claims (22)

In the shock acceleration magnitude estimation device that can be worn on the user's body,
an inertia measurement module that collects the user's inertia information in real time;
A microcontroller having an artificial neural network and using the inertial information as input data of the artificial neural network to derive an estimate of the magnitude of impact acceleration during a fall;
A communication module for transmitting the impact acceleration magnitude estimate to the outside,
The artificial neural network is constructed by performing machine learning on the inertia information and the feature vector extracted from the inertia information as learning data,
The extracted characteristic vector includes a sum vector of three-axis acceleration, a sum vector of a horizontal plane of an accelerometer, a root mean square of the sum vector, and a resultant angular velocity. The method of claim 1,
The inertia information is an impact acceleration magnitude estimation apparatus that is 3-axis acceleration information and 3-axis angular velocity information. The method of claim 1,
wherein the artificial neural network includes a long short-term memory (LSTM) architecture. The method of claim 1,
The impact acceleration magnitude estimation device, wherein the impact acceleration magnitude estimate is estimated for each of five falling directions. delete delete The method of claim 1,
The amount of the training data is increased by processing by at least one data augmentation technique before performing machine learning. 8. The method of claim 7,
The data augmentation technique includes at least one of a jittering technique, a scaling technique, and a time distortion technique. Collecting inertia information and impact acceleration magnitude measured values in a falling situation;
preparing a data set using the inertia information and the measured value of the impact acceleration;
distributing the data set into a training data set and a test data set;
machine learning an artificial neural network using the training data set;
Evaluating the learned artificial neural network, determining a final artificial neural network according to the result, and estimating the magnitude of the shock acceleration from the artificial neural network,
Preparing the data set includes extracting at least one feature vector to be used as input data of the artificial neural network from the inertia information,
The method for estimating the magnitude of impact acceleration, wherein the extracted characteristic vector is at least one of a sum vector of three-axis accelerations, a sum vector of a horizontal plane of an accelerometer, a root mean square of the sum vector, and a resultant angular velocity. 10. The method of claim 9,
The method for estimating the magnitude of impact acceleration, wherein the inertia information is 3-axis acceleration information and 3-axis angular velocity information. 10. The method of claim 9,
wherein the artificial neural network comprises a long short-term memory (LSTM) architecture. 10. The method of claim 9,
The impact acceleration magnitude estimation method is that the impact acceleration magnitude is estimated for each of the five falling directions. 13. The method of claim 12,
The five falling directions are forward, backward, left, right, and vertical downward directions. delete delete 10. The method of claim 9,
The method of estimating the magnitude of impact acceleration further comprising adding the training data set by applying at least one data augmentation technique before the machine learning. 17. The method of claim 16,
The data augmentation technique includes at least one of a jittering technique, a scaling technique, and a time distortion technique. It can be worn on the user's body, and by using the user's inertia information and artificial neural network measured in real time, an estimate of the impact acceleration magnitude before the fall impact in the user's fall situation is derived, and a control signal is generated based on the derived estimate. a device for estimating the magnitude of the impact acceleration,
and a shock protection device whose shock absorption ability is controlled by the control signal,
The impact acceleration magnitude estimation device is
an inertia measurement module that collects the user's inertia information in real time;
a microcontroller having the artificial neural network and using the inertial information as input data of the artificial neural network to derive an estimate of the magnitude of impact acceleration during a fall;
and a communication module that transmits an estimate of the magnitude of the impact acceleration to the impact protection device. delete 19. The method of claim 18,
The inertia information is an impact protection system that is three-axis acceleration information and three-axis angular velocity information. 19. The method of claim 18,
wherein the artificial neural network comprises a Long Short Term Memory (LSTM) architecture. 19. The method of claim 18,
and the impact acceleration magnitude estimate is derived for each of the five drop directions.

Images