CN115245313A - A prediction method of daily walking and running physical energy consumption based on portable accelerometer and gyroscope - Google Patents

Abstract

The invention discloses a human activity energy consumption calculation model. The model is constructed from accelerometer and gyroscope signals at the human ankle and hip. And the prediction performance of the model for low, medium and high exercise intensity was analyzed. The experiment uses the EnEx (Energy Expenditure) database as the benchmark database, and estimates the energy consumption based on inertial data. It uses a SHIMMER sensor composed of a three-axis accelerometer and a three-axis gyroscope for data acquisition. Sensors were placed on the subjects' hips and ankles during data collection. Ten subjects were tested on treadmills at three different speed levels (3.2km/h, 4.8km/h, 6.4km/h). The experiment extracts the original signal on each sensor separately and preprocesses the original signal. Then perform feature extraction, cross-validation, and regression testing. Finally, the model is established based on artificial neural network.

Description

A prediction of daily walking and running physical energy consumption based on portable accelerometer and gyroscope method

technical field

The invention belongs to the cross field of sports science and computer science and technology, and relates to a human body energy consumption calculation model, which provides a portable motion sensor to calculate the human body activity energy consumption.

Background technique

Long-term studies at home and abroad have shown that the energy consumption of human activities is closely related to people's health, and moderate exercise is beneficial to physical and mental health. People maintain a moderate amount of exercise, which can effectively enhance their physical fitness. Effective monitoring of human energy consumption can monitor people's physical health on the one hand, and prevent some related chronic diseases on the other hand.

The methods for calculating the energy consumption of human activities mainly include: double labeled water method, direct calorimetry, indirect calorimetry, and motion sensors. The double-labeled water method mainly uses isotope labeling and element energy conservation to calculate the metabolism of human energy, but it has not been widely used because of its high price. Direct calorimetry is to measure the total heat dissipated from the whole body to the external environment in unit time. Because the direct calorimetry device is relatively complicated, it has not been widely used. Indirect calorimetry is a method of calculating the heat energy generated by the human body by measuring the amount of oxygen consumed by the human body, the amount of carbon dioxide generated, and the amount of urine nitrogen excreted. Such instruments are expensive and only suitable for laboratory applications, and are not practical for general daily activities. Motion sensors are small in size, inexpensive, easy to wear, and can be used in daily activities, and are widely used in evaluating human activity energy consumption. Among them, the accelerometer calculates the energy consumed by human activities by measuring the acceleration and displacement information of the hot body activity. This method is widely used in research by scholars.

Accelerometer sensors are often used for the calculation of energy expenditure of human activities. Gyroscope sensors are often used for human motion recognition. In the field of calculating human activity energy, it is rare for researchers to use all the signals of the accelerometer and the gyroscope at the same time. In order to fully mine the information contained in the original sensor signal, the algorithm will use the deep learning method to model, so as to obtain a more accurate calculation model of human energy consumption.

SUMMARY OF THE INVENTION

Innovation: This experiment uses the raw signals of the three-axis accelerometer and three-axis gyroscope sensors of the hip and ankle to establish a mathematical model. The model method is artificial neural network (ANN), and the final MSE of the model reaches 0.17. For low, medium, and high exercise intensities, the MSEs were 0.14, 0.16, and 0.21, respectively. The MSE of the model remains basically the same for different intensities of walking and running. At the same time, it shows that the model has good generalization ability.

Experimental data: The data used in the experiment is the EnEx (Energy Expenditure) database, which estimates energy consumption based on inertial data. The database contains information about 10 subjects, each with a SHIMMER sensor node placed on the right hip and right ankle. Each sensor node consists of a three-axis accelerometer (A1, A2, A3) and a three-axis gyroscope (G1, G2, G3). The subjects performed on a conventional treadmill (hp-cosmos model mercury med 5.0, Traunstei-n, Germany) at three different speed levels (3.2 km/h, 4.8 km/h, 6.4 km/h). Each speed lasts six minutes. The data format is shown in Figure 1. where A1: accelerometer axis 1, A2: accelerometer axis 2, A3: accelerometer axis 3, G1: gyroscope axis 1, G2: gyroscope axis 2, G3: gyroscope axis 3, MET: consumed energy in MET Speed in units: Speed level in km/h.

Experimental scheme: The algorithm flow chart is shown in Figure 2.

Data preprocessing: The data provided by the database is the original signal without data preprocessing. Since the treadmill has a start-up acceleration process, the signals collected during this period are not stable. Therefore, according to the speed signal of the treadmill, the original signal when the speed is stable is intercepted as the experimental data. Data analysis is performed on the raw signal of the sensor. The null value is processed by the mean substitution method, that is, the mean value of the two data before and after the null value is used for substitution. Then a T test was performed on the measured MET values, and it was found that the MET value of subject 1 was significantly different from the other 9 subjects. Therefore, the data of subject No. 1 was eliminated, and the data of the remaining 9 subjects were used for modeling.

Feature extraction: During feature extraction, non-overlapping sliding windows are used to obtain inertial data. The window size is 1024. Each sliding window of inertial data corresponds to a spirometry-based MET value. To characterize the signal distribution, 10 time-domain features were calculated for each axis of the sensor signal source. That is, mean, variance, variance/mean, maximum, minimum, 10th quantile, 25th quantile, 75th quantile, 90th quantile. The feature extraction formula is shown in Table 3.1 below. Finally, the extracted features are standardized by means of de-mean and variance normalization.

Table 3.1 Common Features in PAEE Estimation

Feature extraction: Some features produced by feature extraction may have low correlation with model predictions. Too many features can easily lead to model overfitting. So before model training, we need to filter the features. This experiment adopts the recursive feature elimination method based on Wrapper for feature selection. That is, the neural network model is used as the base model, and then multiple rounds of training are performed. After each round of training, several features with low weight coefficients are removed, and then the next round of training is performed based on the new feature set. Finally, the feature set whose weight coefficient is ranked as the first level is selected. This method is also a greedy algorithm based on local search to find the optimal feature subset.

The experiment adopts the method of ANN (Artificial God General Network) to establish the mathematical model. The middle hidden layer of the network is three layers, and the corresponding number of nodes are [240, 120, 240] respectively. The activation function of the network is ReLU. Due to the small dataset in this experiment, the solver uses lbfgs. The value of the L2 penalty term is 12, the learning rate is adaptive, and the tol value is set to 0.05. The maximum number of iterations is 1000, and data preheating is used for training, that is, warm_start is set to True. The random state value is 925, this parameter is used to determine the random number generation for the weights and initialization bias. The settings of the above parameters are all based on the grid to find the optimal parameters, that is, the local optimal solution of the parameters.

When the model is trained, it is modeled in a subject-based way. That is, the division of the dataset is based on the subjects. In order to prevent the model from overfitting, 78% (7 people) of the data is used as the training set, and 22% (2 people) of the data is used as the test set, so as to ensure that the data of the two people as the test set do not participate in the model training at all . During model training, Double Cross Validation (DCV) is used for cross-checking. DCV is mainly divided into two layers. The first layer of cross-validation is to select two subjects from 9 subjects as the test set and 7 subjects as the training set, and conduct a total of 36 cross-validation times. The second layer of cross-validation is to use Leave-one-out cross-validation on the training set during model training. That is, when the model is trained, the data of one subject in the training set is used as validation.

Experimental results: The model test results are shown in Table 3. Figure 3 shows the prediction results of low, medium and high intensity based on the ANN model respectively (the abscissa is the MET value predicted by the model, and the ordinate is the measured MET value. Black line: y=x, blue line: y=x±std ( The measured MET value)) The experimental results use the mean-square error (MSE) indicator to judge the accuracy of the model prediction;

Table 1 Experimental results (ALL stands for integrating four signal sources. ANN stands for neural network based method. MSE- ALL, MSE-MED, and MSE-HIGH represent the level of MSE of the model when the exercise intensity is 3.2km/h, 4.8km/h, and 6.4km/h, respectively. mean ;) 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 are retained in the form: mean±var

As can be seen from the test results, the comprehensive MSE of the model is 0.17. For low, medium and high intensities of walking and running, the mean square error values of the model are 0.14, 0.16, and 0.21. As shown in Figure 3.3, it is a test result randomly selected from 36 outer cross-validations of the MLP model. The test set is all the data of subjects No. 2 and No. 3. The black dotted data indicates that the MET value predicted by the model is within the standard deviation interval of the measured MET value, which can be regarded as valid prediction data. Data with a red * indicates that the MET value predicted by the model exceeds the standard deviation interval of the measured MET value, and is regarded as invalid data. For three different intensities of walking and running, the model showed good prediction performance.

As can be seen from the test results, the comprehensive MSE of the model is 0.17. For low, medium and high intensities of walking and running, the mean square error values of the model are 0.14, 0.16, and 0.21. As shown in Figure 3, it is a test result randomly selected from 36 times of outer cross-validation of the MLP model. The test set is all the data of subjects No. 2 and No. 3. The black dotted data indicates that the MET value predicted by the model is within the standard deviation interval of the measured MET value, which can be regarded as valid prediction data. The black * data indicates that the MET value predicted by the model exceeds the standard deviation interval of the measured MET value, and is regarded as invalid data. For three different intensities of walking and running, the model showed good prediction performance.

Claims (4)

1. a daily walking, running physical energy consumption calculation method based on portable accelerometer and gyroscope, is characterized in that, comprises: When the user in daily life wears a motion sensor on the hip and ankle, the accelerometer and gyroscope signals of the motion sensor are obtained; It mainly includes the signals of the three axes of X, Y, and Z of the accelerometer and the three axes of X, Y, and Z of the gyroscope. The accelerometer is mainly responsible for recording the user's displacement information, and the gyroscope is mainly responsible for recording the user's rotational angular velocity information. Each time the signal of the user's motion sensor is obtained, the signal is bound and stored with the user data; then the features required by the model are calculated according to the sensor signal, including the average, variance, variance/mean, maximum value, The minimum value, the 10th quantile, the 25th quantile, the 75th quantile, and the 90th quantile; then the signal features are passed into the trained model, and the user's physical energy consumption data is output and saved. 2 . The method according to claim 1 , wherein acquiring the user's physical energy consumption data 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 networked state. 3 . 3 . The method according to claim 1 , wherein calculating the time-domain characteristics of the sensor signal comprises: a total of four signal sources for the two sensors, and each signal source includes three axes; 4 . Calculate its time domain characteristics for each axis of the signal source, namely: mean, variance, variance/mean, maximum, minimum, 10th quantile, 25th quantile, 75th quantile, 90th quantile of the signal digits. 4. The method according to claim 1, wherein the calculated time domain features are passed into the trained model, comprising: The model is an artificial neural network (ANN), the middle hidden layer of the network is three layers, the corresponding number of nodes are [240, 120, 240], and the activation function of the network is ReLU. Since the data set in this experiment is small, the solver uses lbfgs, The value of the L2 penalty item is 12, the learning rate is adaptive, the tol value is set to 0.05, the maximum number of iterations is 1000, and the training is performed by data preheating, that is, warm_start is set to True, and the random state value is 925 , this parameter is used to determine the random number generation of the weight and initialization bias. The settings of the above parameters are all based on the grid to find the optimal parameters, that is, the local optimal solution of the parameters; When the model is trained, the model is based on the subject (subject), that is, the data set is divided according to the subjects. In order to prevent the model from overfitting, the data of 78% (7 people) is used as the training set. 22 The data of % (2 people) is used as the test set to ensure that the data of the two people as the test set do not participate in the model training at all. During the model training, Double Cross Validation (DCV) is used for cross-checking. DCV is mainly divided into two layers , the first layer of cross-validation is to select two people from 9 subjects as the test set and 7 as the training set each time, a total of 36 cross-validations are performed, and the second layer of cross-validation is to use the training set when the model is trained. Leave-one-out cross-validation; that is, the data of one subject in the training set is used as validation during model training.

Images