CN115245313A

CN115245313A - Method for predicting physical consumption of daily walking and running based on portable accelerometer and gyroscope

Info

Publication number: CN115245313A
Application number: CN202110781770.4A
Authority: CN
Inventors: 杨东强; 王丕坤; 王琳; 孙倩; 刘毅; 马宏伟
Original assignee: Shandong Jianzhu University
Current assignee: Shandong Jianzhu University
Priority date: 2021-07-12
Filing date: 2021-07-12
Publication date: 2022-10-28

Abstract

The invention discloses a human body activity energy consumption calculation model. The model is constructed from accelerometer and gyroscope signals of the ankle and hip of the human body. And the prediction performance of the model on low, medium and high exercise strengths is analyzed. The experiment used the exex database as a reference database and estimated the energy consumption based on inertial data. The method adopts a SHIMMER sensor consisting of a three-axis accelerometer and a three-axis gyroscope for data acquisition. The sensors are respectively placed at the hip and ankle of the testee during data acquisition. Ten subjects were tested on running machines at three different speed levels (3.2 km/h, 4.8km/h, 6.4 km/h), respectively. The experiment extracts the raw signal on each sensor separately and preprocesses the raw signal. And then carrying out feature extraction, cross validation and regression testing. And finally establishing a model based on the artificial neural network.

Description

Method for predicting physical consumption of daily walking and running based on portable accelerometer and gyroscope

Technical Field

The invention belongs to the technical cross field of sports and computer science, and relates to a human energy consumption calculation model which is used for calculating human activity energy consumption for a portable motion sensor.

Background

Long-term research at home and abroad shows that the energy consumption of human activities is closely related to the health of people, and proper exercise is beneficial to physical and mental health. People can keep moderate exercise amount and effectively build up health. The energy consumption of human body is effectively monitored, so that the physical health of people can be monitored, and related chronic diseases can be prevented.

The method for calculating the energy consumption of the human body activity mainly comprises the following steps: double labeled water methods (double labeled water), direct calorimetric methods (direct calorimetric methods), indirect calorimetric methods (index calorimetric methods), and motion sensors (motion sensors). The double-standard water method mainly utilizes isotope labeling and element energy conservation to calculate the metabolism of human body energy, but cannot be widely applied because of high price. The direct calorimetric method is to measure the total heat emitted to the external environment in unit time by the whole body, and the direct calorimetric device is relatively complex, so the direct calorimetric method cannot be widely popularized and used. Indirect calorimetry is a method of calculating the heat energy generated by a human body by measuring the amount of oxygen consumed by the body and the amount of carbon dioxide generated and urine nitrogen excreted. This kind of instrument is expensive, and only is applicable to the laboratory and uses, and is not strong in the practicality in general daily activity aspect. The motion sensor has small volume, low price and convenient wearing, can be applied to daily activities, and is widely applied to the aspect of evaluating the energy consumption of human activities. Wherein, the accelerometer calculates the energy consumed by human body activity by measuring the acceleration and displacement information during the hot body activity, and the method is widely applied to research by scholars.

Accelerometer sensors are commonly used for the calculation of energy expenditure for human activity. Gyroscopic sensors are often used for human motion recognition. In the field of calculation of human activity energy, it is not uncommon for researchers to employ all signals of an accelerometer and a gyroscope at the same time. In order to fully mine the information in the original signal of the sensor, the algorithm adopts a deep learning method to carry out modeling, so that a more accurate human energy consumption calculation model is obtained.

Disclosure of Invention

Creativity in the experiment, a mathematical model is established by utilizing original signals of three-axis accelerometers and three-axis gyroscope sensors of the hip and the ankle. The model method is an Artificial Neural Network (ANN), and the final MSE of the model reaches 0.17. The MSE of the exercise intensity is 0.14, 0.16 and 0.21 respectively for the low exercise intensity, the medium exercise intensity and the high exercise intensity. The MSE of the model keeps basically consistent for walking and running motions with different intensities. And simultaneously, the model has good generalization capability.

Experimental data: the data used in the experiment was an exex (energy expenditure) database, and the energy consumption was estimated based on inertial data. The database contains information about 10 subjects, each of whom has a SHIMMER sensor node placed at the right hip and ankle, respectively. Each sensor node consists of a three-axis accelerometer (A1, A2, A3) and a three-axis gyroscope (G1, G2, G3). The subjects were run on a conventional treadmill (hp-cosmos model marcurgy 5.0, traunstei-n, germany) at three different speed levels (3.2 km/h, 4.8km/h, 6.4 km/h), respectively. Each speed lasted six minutes each. The data format is shown in fig. 1. Wherein A1: accelerometer axis 1, A2: accelerometer axis 2, A3: accelerometer axis 3, G1: gyroscope shaft 1, G2: gyroscope shaft 2, G3: gyroscope shaft 3, met: energy consumed in units of MET, speed: the speed level is in km/h.

The experimental scheme is as follows: the algorithm flow chart is shown in fig. 2.

Data preprocessing: the data provided by the database is the original signal and is not preprocessed. The signals collected during this period are unstable due to the running machine having a process of starting and accelerating. Therefore, according to the speed signal of the running machine, the original signal when the speed is stable is intercepted and used as experimental data. And carrying out data analysis on the original signals of the sensors. And (4) processing the null value by adopting a mean value substitution method, namely substituting the mean value of the front data and the back data of the null value. The measured MET values were then subjected to a T-test and found to be highly significant in the MET values of subject No. 1 as compared to the other 9 subjects. Therefore, the data of the first subject was discarded, and the data of the remaining 9 subjects was used for modeling.

Feature extraction: and during feature extraction, acquiring inertial data by adopting a non-overlapped sliding window. The window size is 1024. Each sliding window of inertial data corresponds to a MET value determined based on the spirometric system. To characterize the signal distribution, 10 time domain features were calculated for each axis of the sensor signal source. I.e., mean, variance/mean, maximum, minimum, 10 quantile, 25 quantile, 75 quantile, 90 quantile. The feature extraction formula is shown in table 3.1 below. And finally, normalizing the extracted features in a mean value removing and variance normalizing mode.

TABLE 3.1 common features in PAEE estimation

Selecting characteristics: some features resulting from feature extraction may have low correlation with model predictions. Too much feature quantity easily causes model overfitting. Therefore, before model training, we need to filter features. In the experiment, a recursive characteristic elimination method based on Wrapper is adopted for characteristic selection. Namely, a neural network model is used as a base model, and then a plurality of rounds of training are carried out. And removing a plurality of features with low weight coefficients after each training round, and then carrying out the next training round based on the new feature set. Finally, a feature set with the weight coefficient arranged as a first level is selected. The method is also a greedy algorithm for finding the optimal feature subset based on local search.

The experiment adopts an ANN (artificial neural network) method to establish a mathematical model. The middle hidden layer of the network is three layers, and the corresponding node number is [240,120,240]. The activation function of the network is ReLU. Because the data set of the experiment is small, the solver adopts lbfgs. Wherein the value of the L2 penalty term is 12, the learning rate is adaptive (adaptive), and the tol value is set to 0.05. The maximum iteration number is 1000, and the training is carried out in a data preheating mode, namely, the war _ start is set to True. The random state value is 925 and this parameter is used to determine the weights and to initialize the generation of the biased random numbers. The optimal parameters, namely the local optimal solution of the parameters, are found in a grid-based mode.

When the model is trained, the model is modeled in an object-based mode. I.e. the partitioning of the data set is according to the subject. To prevent the overfitting of the model, 78% (7 persons) of data were used as the training set and 22% (2 persons) of data were used as the test set, thereby ensuring that the data of both persons as the test set did not participate in the model training at all. The Cross-examination was performed in Double Cross Validation (DCV) mode during model training. DCV is mainly divided into two layers, wherein the first layer of cross validation is to select two persons from 9 subjects as a test set and 7 persons as a training set at each time, and the cross validation is carried out for 36 times. The second layer of cross validation is Leave-one-out cross validation (Leave-one-out cross validation) applied to the training set during model training. Namely, the data of one subject in the training set is used as the verification during model training.

The experimental results are as follows: the results of the model tests are shown in Table 3. FIG. 3 shows the low, medium and high intensity prediction results (with the abscissa being the predicted MET value of the model and the ordinate being the measured MET value; black line: y = x, blue line: y = x + -std (measured MET value)) based on the ANN model, respectively, and the mean-square error (MSE) index is used to determine the accuracy of the model prediction;

TABLE 1 experimental results (ALL stands for Integrated four Signal Source. ANN stands for neural network-based method. MSE-ALL, MSE-MED, MSE-HIGH represent the mean values of the MSEs of the models at motor intensities of 3.2km/h, 4.8km/h, 6.4km/h, respectively)

SOURCE	MSE-ALL	MSE-LOW	MSE-MEDIUM	MSE-HIGH
					ANN	0.17±0.006	0.14±0.003	0.16±0.010	0.21±0.011

Note: results retained form: mean + -var

From the test results, it can be seen that the overall MSE for the model is 0.17. For walking and running motions with different intensities of low, medium and high, the mean square error value of the model is 0.14, 0.16 and 0.21. As shown in fig. 3.3, which is a test result randomly drawn from 36 out-cross-validation of MLP model. The test set is the total data for

subjects

2 and 3. Wherein the black dot data indicates that the MET value predicted by the model is within the standard deviation range of the measured MET value, and can be regarded as effective prediction data. And red star data represent the interval of standard deviation of the MET value predicted by the model exceeding the measured MET value and are regarded as invalid data. The model shows good prediction performance for walking and running motions with three different strengths.

From the test results, it can be seen that the integrated MSE of the model is 0.17. For walking and running motions with different intensities of low, medium and high, the mean square error value of the model is 0.14, 0.16 and 0.21. As shown in fig. 3, which is a test result randomly drawn from 36 out-cross-validation of MLP model. The test set is the total data for

subjects

2 and 3. The black dot data indicate that the MET value predicted by the model is within the standard deviation range of the measured MET value, and can be regarded as effective prediction data. The black data represents the interval of standard deviation of the predicted MET value of the model exceeding the measured MET value, and is regarded as invalid data. The model shows good prediction performance for walking and running motions with three different strengths.

Claims

1. A daily walking and running physical consumption calculation method based on a portable accelerometer and a gyroscope is characterized by comprising the following steps:

a user in daily life wears a motion sensor on each of the hip and the ankle to acquire signals of an accelerometer and a gyroscope of the motion sensor;

the system mainly comprises signals of X, Y, Z three axes of an accelerometer and signals of X, Y, Z three axes of a gyroscope, wherein the accelerometer is mainly responsible for recording displacement information of a user, the gyroscope is mainly responsible for recording rotation angular velocity information of the user, and the signals and user data are bound and stored when the signals of a motion sensor of the user are obtained each time; then, calculating the characteristics required by the model according to the sensor signals, wherein the characteristics comprise the average value, the variance/mean value, the maximum value, the minimum value, the 10 quantile, the 25 quantile, the 75 quantile and the 90 quantile of the signals; and then, transmitting the signal characteristics into the trained model, outputting the physical ability consumption data of the user, and storing the physical ability consumption data.

2. The method of claim 1, wherein the obtaining physical consumption data of the user in real time comprises: the worn device sends a data transmission signal to at least one base station, and the base station is in a networking state.

3. The method of claim 1, wherein computing a time domain signature of the sensor signal comprises: the two sensors have four signal sources in total, and each signal source comprises 3 axes;

calculating the time domain characteristics of each axis of the signal source, which are respectively: mean, variance/mean, maximum, minimum, 10 quantile, 25 quantile, 75 quantile, 90 quantile of the signal.

4. The method of claim 1, wherein the computed time domain features are propagated into a trained model, comprising:

the model is an Artificial Neural Network (ANN), the middle hidden layer of the network is three layers, the corresponding node numbers are [240,120,240] respectively, the activation function of the network is ReLU, because the data set of the experiment is small, the solver adopts lbfgs, the value of an L2 penalty term is 12, the learning rate is self-adaptive (adaptive), the tol value is set to 0.05, the maximum iteration frequency is 1000, the model is trained in a data preheating mode, namely, the term _ start is set to True, the random state value is 925, the parameter is used for determining the weight and generating random numbers of initialization bias, and the setting of the parameters is realized by searching for the optimal parameter in a grid-based mode, namely, the local optimal solution of the parameter;

during model training, modeling is carried out in an object-based (subject) mode, namely, a data set is divided according to subjects, in order to prevent an overfitting phenomenon of the model, 78% (7 persons) of data are used as a training set, 22% (2 persons) of data are used as a test set, so that the data of two persons used as the test set are ensured not to participate in model training at all, cross Validation is carried out in a Double Cross Validation (DCV) mode during model training, the DCV is mainly divided into two layers, the first layer of Cross Validation is that two persons are selected from 9 subjects as the test set each time, 7 persons are used as the training set, and Cross Validation is carried out for 36 times in total, and the second layer of Cross Validation is that one-out Cross Validation (Leave-one-out Cross Validation) is carried out on the training set during model training; namely, the data of one subject in the training set is used as the verification during model training.