CN116167412A

CN116167412A - Exercise heart rate prediction method based on time convolution network

Info

Publication number: CN116167412A
Application number: CN202310161397.1A
Authority: CN
Inventors: 杨良怀; 孙甜甜
Original assignee: Zhejiang University of Technology ZJUT
Current assignee: Zhejiang University of Technology ZJUT
Priority date: 2023-02-24
Filing date: 2023-02-24
Publication date: 2023-05-26

Abstract

The invention discloses a motion heart rate prediction method based on a time convolution network, which comprises the following steps: step 1: and collecting historical data of heart rate and longitude and latitude in the exercise process of the students, converting exercise speed and acceleration data, processing abnormal values, and carrying out normalization processing and data set division on the data. Step 2: and constructing a time convolution network model. And training a network model by using the divided training data set, and selecting and storing a model with the best prediction result by using the test data set. Step 3: and collecting heart rate and longitude and latitude data of recent student exercise, preprocessing, inputting a trained neural network model, and carrying out inverse normalization on the model output value to obtain an exercise heart rate prediction result. The invention uses the time convolution network to receive the multi-sequence speed and acceleration input, avoids the problem of gradient explosion or disappearance which often occurs in the cyclic neural network, and has the parallel characteristics that the processing efficiency is higher and the processing speed is faster.

Description

Exercise heart rate prediction method based on time convolution network

Technical Field

The invention belongs to the field of sports health, and particularly relates to a sports heart rate prediction method based on a time convolution network.

Background

In recent years, students in colleges and universities have poor overall physical quality and tend to slip down year by year. Students in colleges and universities use the student as a main backup force for social construction, and a good physical quality is a precondition for active learning or work. Although colleges and universities can adopt physical health test and other modes to urge students to exercise, the physical health test and the exercise can only be used as feedback of the physical quality of the students, and the problems faced by the physical quality of the students can not be reflected and monitored well in real time and corresponding measures can be taken. On the other hand, researches show that the heart rate (the number of times of heart beating in one minute) in the exercise process can accurately reflect the exercise intensity and exercise load of the sporter, the exercise process can be reasonably guided according to the level of the real-time heart rate and the personal condition of the sporter, the injury risk caused by unreasonable exercise is reduced, and the physical quality of the sporter is improved.

With the development of wearable technology, people have been able to measure and collect their own heart rate by mobile devices (e.g., exercise bracelets). The large-scale popularization of the exercise bracelet brings great convenience for recording the exercise process of students and counting the exercise data of the students, saves a great deal of manpower compared with the traditional manual recording, and solves the defect that the traditional method cannot know the specific situation in the exercise of the students. The school can monitor the daily exercise condition of students by checking statistical data or prescribing corresponding exercise times, can analyze heart rate data collected in the exercise process of the students, and can conduct personalized guidance and monitor the long-term change of the constitution of the students aiming at the individual constitution problem of the students.

There are still problems with using a sports wristband to monitor heart rate during actual sports. On the one hand, due to various reasons such as limitation of the exercise bracelet, temperature influence and the like, the heart rate measured and displayed by the exercise bracelet sometimes has a certain deviation from the actual heart rate of an athlete, and the exercise artifact phenomenon occurs, so that the actual situation of the exercise of the student can not be reflected well. On the other hand, the price of the sport wristband is approximately hundreds of different, the sport wristband is relatively expensive, the problem of insufficient budget can exist in independent purchase of students or in mass collective purchase of schools, and certain difficulty is brought to the use of mass high coverage of campuses. For such cases, only speed data, not heart rate data, may be acquired while the student is in motion.

Disclosure of Invention

Based on the problems to be solved mentioned in the background art, the invention provides the exercise heart rate prediction method, which can predict the heart rate time sequence through the speed and acceleration data of the exercise process and correct the true exercise heart rate to a certain extent. The invention adopts the following technical scheme:

an exercise heart rate prediction method based on a time convolution network comprises the following steps:

step 1: and collecting historical data of heart rate and longitude and latitude in the exercise process of the students, converting exercise speed and acceleration data, processing abnormal values, and carrying out normalization processing and data set division on the data.

Step 2: constructing a time convolution network model, training the network model by using the divided training data set, and selecting a model with the best prediction result by using the test data set for storage.

Step 3: and collecting heart rate and longitude and latitude data of recent student exercise, preprocessing, inputting a trained neural network model, and carrying out inverse normalization on the model output value to obtain an exercise heart rate prediction result.

Further, the student movement history data collected in the step 1 is of a time series data type, and the collection frequency is once per second. During exercise, the students wear and start a bracelet heart rate acquisition function to acquire exercise real-time heart rate data; and recording the motion through intelligent application, and collecting the real-time longitude and latitude position data of the motion. The specific steps of the step 1 are as follows:

1) In order to solve the problem of missing or abnormal acquired data and remove data noise, the invention adopts the following preprocessing. Deducing a corresponding speed and acceleration sequence by using the acquired longitude and latitude information; carrying out Kalman filtering on the obtained speed and acceleration sequence; and judging whether the data is abnormal according to the maximum fluctuation threshold value of the heart rate per second for the heart rate data, and deleting the heart rate at the time point if the data is abnormal. Whereas interpolation is used to fill in data that is missing or that requires multi-dimensional alignment.

2) Unifying time steps of the preprocessed data sets, performing zero filling operation on data which are smaller than the set time steps, and performing cutting operation on data which are larger than the set time steps.

3) And (5) carrying out normalization processing on the time sequence of the speed, the acceleration and the heart rate in the data set.

4) According to 3: the 1 scale randomly divides the data set into a training set and a test set.

The data obtained after step 1 can be expressed as

Time series data representing the speed from the start time to the T time,/->

Time series data representing acceleration from start time to T time, y _1:T Representing heart rate sequence data from the start time to the T time.

Further, the neural network model in the step 2 has an overall structure including: and the superposition and output layers of the front-stage and rear-stage time convolution networks comprise a plurality of residual blocks. The specific steps of the step 2 are as follows:

1) Velocity data

Acceleration data->

Time information from start time to T time +.>

As three dimensions of input data, sequentially passing through an expansion convolution layer and a batchThe normalization layer, the ReLU activation layer, and the discard layer, the dilation convolution can be expressed by the following formula:

where d represents the expansion ratio, f represents a filter of size k, x represents a one-dimensional sequence input, and s represents a sequence element.

2) And performing layer jump connection in a residual block formed by two layers of convolution structures with the same expansion rate, and combining the output results of the 1X 1 convolution and the expansion convolution layers by using the formula:

o＝Activation(X+F(X))#(2)

where X represents the input and F (X) represents the output of the convolution module. The step eliminates the influence of network gradient elimination and explosion to a certain extent, and avoids the condition of different input and output widths, and the Activation is an Activation function.

3) And after a plurality of residual blocks, obtaining the output of the front-stage time convolution network, taking the output as the input of the rear-stage time convolution network with the same structure, and finally obtaining a prediction result through a full connection layer.

4) And calculating the predicted result and the loss of the normalized real heart rate data by using an average absolute error loss function, and updating the parameters of the time convolution network model by an Adam algorithm.

5) And stopping model iteration when the loss function value is smaller and tends to be stable, adjusting super parameters such as the learning rate, the discarding rate and the like of the model, and observing the performance change of the model, so as to acquire the optimal parameters of the model and save the optimal parameters, namely the trained time convolution network model.

Further, the specific steps of the step 3 are as follows:

1) And (3) selecting longitude and latitude data to be predicted, and performing preprocessing in the step (1) to obtain a speed and acceleration time sequence.

2) And (3) inputting the preprocessed data into the time convolution neural network model trained in the step (2) to obtain model output.

3) And (5) carrying out inverse normalization on the model output to obtain a predicted heart rate result.

Compared with the prior art, the technical scheme of the invention is characterized in that:

the heart rate prediction method based on the improved time convolution network model is used for receiving multi-sequence input, so that the problem of gradient explosion or disappearance frequently occurring in a cyclic neural network is avoided, and the parallel characteristics enable the processing efficiency to be higher and the processing speed to be higher. The trained time convolution network model can capture short-term and long-term modes in a time sequence, and can play a role in predicting the heart rate with high precision for long-time movements.

Drawings

FIG. 1 is a flow chart of a heart rate prediction model based on a time convolution network in the present invention;

FIG. 2 is a diagram of the overall architecture of a time-convolved neural network model of the present invention;

fig. 3 is a block diagram of a residual block in the neural network model of the present invention.

Detailed Description

The invention relates to a personalized exercise prescription recommendation method driven by exercise data, which is further described in detail below with reference to the accompanying drawings and the embodiments, and as shown in fig. 1, the method comprises the following steps:

The student movement history data collected in the step 1 is of a time series data type, and the collection frequency is 1s. The student wears and starts a bracelet heart rate acquisition function in the exercise process, and heart rate data in the exercise process are acquired; and recording the movement through intelligent application, and collecting real-time longitude and latitude position information in the movement process.

The specific steps of the step 1 are as follows:

1) The problem of data missing or data abnormality exists in the collected data, and the data needs to be preprocessed first. First, the distance d between two points is calculated by using the acquired longitude and latitude information _i The formula is:

where r represents the earth radius, arcsin is an abbreviation for arcsine function, hav is an abbreviation for semi-orthostatic function,

and->

Respectively represent the latitude, lambda of two points _i And lambda (lambda) _i-1 Representing the longitudes of the two points, respectively.

Calculating the movement speed s of the student between the two points according to the distance and the time interval _i ：

And then deducing the student exercise acceleration:

for the velocity sequence(s) ₁ ,…,s _N ) Acceleration sequence (a) ₁ ,…,a _N ) Performing Kalman filtering; for heart rate sequence (y ₁ ,…,y _N ) Judging whether the data point is abnormal or not according to the maximum fluctuation threshold value of the heart rate per second, and deleting the heart rate at the time point if the data point is abnormal. Whereas interpolation is used to fill in data that is missing or that requires multi-dimensional alignment. Taking heart rate as an example, when y _i Missing, and y _i-1 And y _i+1 All are absent, y _i The following formula can be used for calculation:

when y is _i Missing, and y _i-1 Or y _i+1 Deletion, y _i The following formula can be used for calculation:

wherein a, b is greater than or equal to 1 and y _i-a+1 ，...，y _i-1 ，y _i-1 ，...，y _i+b-1 All are missing values. When y is _i Missing, and y _i For the first value in the heart rate sequence, y _i+1 And y _i+2 All are absent, y _i The following formula can be used for calculation:

when y is _i Missing, and y _i For the last value in the heart rate sequence, y _i-1 And y _i-2 All are absent, y _i The following formula can be used for calculation:

2) Unifying the time step T of the preprocessed data set, performing zero padding operation on the data with less than the set time step (namely, the sequence length N < T), and performing cutting operation on the data with more than the set time step (namely, the sequence length N > T).

3) The method comprises the following steps of carrying out normalization processing on the time sequence of the speed, the acceleration and the heart rate in the data set, wherein the maximum and minimum normalization is adopted in the method:

wherein f and f' represent the values before and after normalization of the data points, f _min Minimum value in sequence, f _max Representing the maximum in the sequence.

The data obtained after step 1 can be expressed as

Time series data representing the speed from the start time to the T time,/->

Time series data representing acceleration from start time to T time, y _1：T Representing heart rate sequence data from the start time to the T time. An example of data is given in table 1, where the value of T is 1024.

Table 1 data example

Step 2: constructing a time convolution neural network model, training the neural network model by using the divided training data set, and selecting a model with the best prediction result by using the test data set for storage.

The neural network model in step 2, as shown in fig. 2, has the overall structure including: the superposition and output layers of the previous and subsequent time convolution networks further comprise a plurality of residual blocks, and the structure of the residual blocks can be shown in fig. 3. The specific steps of the step 2 are as follows:

1) Velocity data

Acceleration data->

Time information from start time to T time +.>

As three dimensions of input data, sequentially passing through an expansion convolution layer, a batch normalization layer, a ReLU activation layer and a discarding layerWherein the dilation convolution can be expressed by the following formula:

2) Layer jump connection is carried out in a residual block formed by two layers of convolution structures with the same expansion rate, and output results of the 1X 1 convolution and the expansion convolution layers are combined:

o＝Activation(X+F(X))#(10)

where X represents the input and F (X) represents the output of the convolution module. The step eliminates the influence of network gradient elimination and explosion to a certain extent, and avoids the condition of different input and output widths.

4) Calculating the predicted result and the loss of the normalized real heart rate data by using an average absolute error loss function:

wherein y is _i Represents the predicted value of the current value,

representing the true value and n representing the heart rate sequence length. The parameters of the time convolution network model are then updated by Adam algorithm.

5) And stopping model iteration when the loss value is small and tends to be stable, adjusting super parameters such as the learning rate, the discarding rate and the like of the network model, observing the performance change of the model, and obtaining the optimal parameters of the model for storage, namely the trained time convolution network model.

The specific steps of the step 3 are as follows:

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting thereof; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the previous embodiment can be modified or part of technical features can be replaced equivalently; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An exercise heart rate prediction method based on a time convolution network is characterized by comprising the following steps:

step 1: collecting historical data of heart rate and longitude and latitude in the exercise process of students, converting exercise speed and acceleration data, processing abnormal values, and carrying out normalization processing and data set division on the data;

step 2: constructing a time convolution neural network model, training the neural network model by using the divided training data set, and selecting a model with the best prediction result by using the test data set for storage;

2. The exercise heart rate prediction method based on a time convolution network according to claim 1, wherein the specific steps of step 1 are as follows:

1.1 The collected student movement history data is of a time sequence data type, and the collection frequency is 1s; the student wears and starts a bracelet heart rate acquisition function in the exercise process, and heart rate data in the exercise process are acquired; recording motion through intelligent application, and collecting real-time longitude and latitude position information in the motion process;

1.2 Using the acquired longitude and latitude information to deduce a corresponding speed and acceleration sequence; carrying out Kalman filtering on the obtained speed and acceleration sequence; judging whether the data is abnormal according to the maximum fluctuation threshold value of the heart rate per second for the heart rate data, and deleting the heart rate at the time point if the data is abnormal; and for missing or data requiring multi-dimensional alignment, filling by using an interpolation method;

1.3 Unifying time steps of the preprocessed data sets, performing zero filling operation on the data which are less than the set time steps, and performing cutting operation on the data which are more than the set time steps;

1.4 Normalizing the time sequence of the speed, the acceleration and the heart rate in the data set;

1.5 Randomly dividing the data set into a training set and a test set according to the ratio.

3. The exercise heart rate prediction method based on the time convolution network according to claim 1, wherein the neural network model in the step 2 has a whole structure including: the superposition and output layer of the two-stage time convolution network, the time convolution network comprises a plurality of residual blocks;

the specific steps of the step 2 are as follows:

2.1 Speed data)

Acceleration data->

Time information from start time to T time +.>

As three dimensions of the input data, sequentially pass through an expansion convolution layer, a batch normalization layer, a ReLU activation layer and a discard layer, wherein the expansion convolution is expressed by the following formula:

wherein d represents expansion rate, f represents a filter with a size of k, x represents one-dimensional sequence input, and s represents a certain sequence element;

2.2 Layer jump connection is performed in a residual block formed by two layers of convolution structures with the same expansion rate, and output results of the 1 x 1 convolution and the expansion convolution layers are combined:

o＝Activation(X+F(X))

wherein X represents input, F (X) represents output of the convolution module, and activity is an Activation function;

2.3 After a plurality of residual blocks, obtaining the output of a front-stage time convolution network as the input of a rear-stage time convolution network with the same structure, and finally obtaining a prediction result through a full connection layer;

2.4 Calculating the predicted result and the loss of the normalized real heart rate data by using an average absolute error loss function, and updating the parameters of the time convolution network model by an Adam algorithm;

2.5 Stopping model iteration when the loss value is smaller and tends to be stable, adjusting super parameters such as the learning rate, the discarding rate and the like of the time convolution network model, observing the performance change of the model, and obtaining the optimal parameters of the model for storage, namely the trained time convolution network model.

4. The exercise heart rate prediction method based on the time convolution network according to claim 1, wherein the specific steps of the step 3 are as follows:

3.1 Selecting longitude and latitude data to be predicted, and performing preprocessing in the step 1 to obtain a speed and acceleration time sequence;

3.2 Inputting the preprocessed data into the time convolution neural network model trained in the step 2 to obtain model output;

3.3 Inversely normalizing the model output to obtain a predicted heart rate result.