Wearable quasi-periodic motion frequency statistical method capable of online learning, program, equipment and storage medium

Abstract

The invention belongs to the technical field of wearable motion frequency statistics, and particularly relates to a wearable quasi-periodic motion frequency statistics method capable of online learning, a program, equipment and a storage medium. The method comprises the following steps: selecting data of a counted action in a continuous motion period, constructing an analysis matrix, constructing a template function by using a singular vector and a singular value of the analysis matrix, performing template function filtering on collected real-time data by using the template function, and counting a wave crest and wave trough combination number meeting a threshold requirement in the filtered data; and recording the action occurrence frequency. According to the method, the influence of other motions on actions needing to be counted can be reduced by utilizing template function filtering, the occurrence frequency of discontinuous actions can be accurately recorded, and the accuracy of action statistics is improved; the template function is obtained through online data training, individual differences of wearers can be eliminated, the statistical result of the number of actions can be evaluated regularly or according to a user instruction, and if the statistical result is lower than an expected value, the template function can be retrained.

Description

Technical Field

[0001] The present invention belongs to the technical field of wearable action frequency statistics, and in particular relates to a wearable quasi-periodic motion frequency statistics method, program, device and storage medium capable of online learning. Background Art

[0002] Wearable motion counting devices based on micro-electromechanical systems (MEMS) (mainly accelerometers and gyroscopes) are widely used in human or animal motion analysis, indoor navigation, health monitoring and management, etc. Generally, the counted motions are quasi-periodic when in continuous motion (since the motion amplitude and motion span are difficult to be completely consistent, they are not strictly periodic changes), such as walking and running step count.

[0003] The data acquisition device equipped with an accelerometer and a gyroscope is worn on the part that can be affected by the action. When the wearer performs the above-mentioned continuous action, the acceleration and gyroscope data values ​​have obvious quasi-periodic changes over time (because the action amplitude and action span are difficult to be completely consistent, the acceleration and gyroscope data are not strictly periodic changes). This quasi-periodic change is caused by continuous rhythmic repetitive actions. The existing action count statistics method is mainly developed based on the above characteristics. For example, autocorrelation analysis method, finite state machine method, zero speed detection method, and peak detection method. Generally, in order to eliminate the false positive data caused by other actions of the wearer or environmental vibration from being counted in the action count, the above method uses the method of judging whether the number of peaks (action counts) in the observation window reaches the set threshold number to determine, and if it does not reach it, it will not be counted in the action count. Although this method eliminates the influence of other actions or environmental vibrations, it also excludes the number of scattered movements from the count, that is, if the number of actual actions in the observation window is less than the threshold number, it will be missed. This will not be conducive to obtaining more accurate action count statistics, and thus affect the deviation of indoor navigation, health status and other information. Although more complex pattern recognition methods can be used to first perform action recognition and then count the number of actions, the computational complexity is high and it is difficult to deploy on wearable devices. In view of the above problems, it is necessary to propose a more accurate method for counting the number of actions. Summary of the invention

[0004] The purpose of the present invention is to provide a wearable quasi-periodic motion counting method, program, device and storage medium that can be learned online to address the problems that the statistics of quasi-periodic actions (such as walking, running steps, chewing times) collected by wearable devices are easily disturbed by noise and the statistics of scattered actions are not accurate.

[0005] A wearable quasi-periodic motion frequency counting method capable of online learning comprises the following steps:

[0006] Step 1: When wearing the wearable device for the first time, the wearer completes the specified action within the specified time, collects data through the wearable device, and performs preprocessing and zero-mean processing on the data; through the initial template function stored in the wearable device, the data collected during the wearer's continuous motion state is identified;

[0007] Step 2: Select the data of the statistical action after preprocessing and zero mean processing during the continuous motion, construct the selected data into an analysis matrix, solve the singular values ​​and singular vectors of the analysis matrix, and subtract the first left singular vector from its mean as the updated template function; determine the detection threshold according to the updated template function;

[0008] Step 3: When the wearer is exercising, the wearable device collects data in real time and performs preprocessing and zero-mean processing on the data; for the processed data within the calculation cycle, it is filtered through the updated template function, and the number of peaks and troughs in the filtered data that meet the detection threshold is counted and recorded as the number of action occurrences.

[0009] Furthermore, the wearable device carries a multi-axis accelerometer and an initial template function, and the multi-axis accelerometer is used to collect data; the initial template function is specifically:

[0010] Multi-axis accelerometer data of multiple similar wearers in known continuous motion, non-continuous motion, and no motion are collected to form a decision tree, which is updated using a machine learning process; an initial template function is used to determine whether the collected multi-axis accelerometer data is data during continuous motion under the condition of unknown motion of the new wearer.

[0011] Furthermore, the data is preprocessed and zero-mean processed, specifically:

[0012] According to the acceleration data collected by the multi-axis accelerometer, the combined acceleration modulus is calculated to obtain x1=[x1(1),x1(2),...,x1(t),...], where t is the data time domain serial number;

[0013] The zero mean value processing is performed on the combined acceleration modulus value data x1 to obtain the zero mean data x s ;

[0014] x s =x1-x′1

[0015] Among them, x′1 is obtained by mean filtering x1, and the filtering window length is not less than 4 times the duration of one action.

[0016] Furthermore, in the step 2, the data x1=[x1(1), x1(2), ..., x1(l), ..., x1(L)] of the statistical action after preprocessing and zero-mean processing during continuous motion is selected to construct an M×N dimensional analysis matrix A; wherein M is the data length corresponding to a motion cycle, N is the number of motion cycles analyzed, and N>M, L>M×N.

[0017] Furthermore, in step 2, the first left singular vector v1 of the analysis matrix A is selected minus its mean is the template function g;

[0018]

[0019] The detection threshold r is:

[0020]

[0021] Wherein, l1(g) and l2(g) are the l1 norm and l2 norm of the template function g; r0 is the ratio of the first singular value of the analysis matrix A to the sum of all singular values.

[0022] Furthermore, in step 3, for the data x after preprocessing and zero mean processing within the calculation period, s , calculate x s Convolution with the template function g gives the filtered data y = [y(1), y(2), ..., y(t), ..., y(T)]; if y(t) < -r, y(t+1) > r, the combination of y(t) and y(t+1) is recorded as the number of times an action occurs.

[0023] Furthermore, if the wearer changes, or the statistical result of the self-assessed action times of the wearable device is lower than expected, step 1 is re-executed to retrain the template function.

[0024] A computer device / equipment / system comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-mentioned wearable quasi-periodic motion frequency statistics method capable of online learning.

[0025] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned wearable quasi-periodic motion frequency statistics method that can be learned online.

[0026] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned wearable quasi-periodic motion frequency statistics method that can be learned online.

[0027] The beneficial effects of the present invention are:

[0028] The present invention selects data of the action to be counted during continuous motion and constructs it into an analysis matrix, uses the singular vectors and singular values ​​of the analysis matrix to construct a template function, uses the template function to perform template function filtering on the collected real-time data, and counts the number of combinations of peaks and troughs that meet the threshold requirements in the filtered data, which is recorded as the number of action occurrences. The present invention uses template function filtering to reduce the influence of other movements on the required statistical actions, can accurately record the number of occurrences of non-continuous actions, and improve the accuracy of action statistics; the template function is obtained through online data training, which can eliminate individual differences among wearers, and can evaluate the statistical results of the number of action times regularly or according to user instructions, and if it is lower than expected, the template function can be retrained. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the overall flow chart of the present invention.

[0030] Figure 2 This is a function diagram of a human walking step statistics template (worn on the ankle) in an embodiment of the present invention.

[0031] Figure 3 The waveform and detection threshold diagram of the processed three-axis acceleration sensor data corresponding to continuous walking and sporadic steps in the embodiment of the present invention (worn on the ankle). DETAILED DESCRIPTION

[0032] The present invention is further described below in conjunction with the accompanying drawings.

[0033] The present invention aims to solve the problem that the statistics of quasi-periodic actions (such as walking, running steps, chewing times) collected by wearable devices are easily disturbed by noise and the statistics of scattered actions are not accurate enough. A wearable quasi-periodic motion statistics method that can be learned online is proposed.

[0034] The present invention uses a template function to perform template function filtering on a multi-axis accelerometer, and counts the number of combinations of peaks and troughs that meet the threshold requirements in the filtered data, which is recorded as the number of action occurrences, wherein the template function is obtained through online data training. The present invention selects multi-axis accelerometer data during the continuous motion of the action to be counted and constructs it into an analysis matrix, and then uses the singular vectors and singular values ​​of the analysis matrix to construct the template function, and can evaluate the statistical results of the number of actions regularly or according to user instructions. If it meets the expected value and does not change the template function, and the evaluation effect is lower than expected, the template function is retrained.

[0035] A wearable quasi-periodic motion frequency counting method capable of online learning comprises the following steps:

[0036] Step 1: When wearing the wearable device for the first time, the wearer completes the specified action within the specified time, collects data through the wearable device, and performs preprocessing and zero-mean processing on the data; through the initial template function stored in the wearable device, the data collected during the wearer's continuous motion state is identified;

[0037] The wearable device carries a multi-axis accelerometer and an initial template function, and the multi-axis accelerometer is used to collect data;

[0038] The initial template function is as follows:

[0039] Collect multi-axis accelerometer data of multiple similar wearers in known continuous motion, non-continuous motion, and no motion conditions to form a decision tree, which is updated using a machine learning process; determine whether the collected multi-axis accelerometer data is data during continuous motion under the condition of unknown motion conditions of the new wearer through an initial template function;

[0040] Perform preprocessing and zero mean processing on the data, specifically:

[0041] According to the acceleration data collected by the multi-axis accelerometer, the combined acceleration modulus is calculated to obtain x1=[x1(1),x1(2),...,x1(t),...], where t is the data time domain serial number;

[0042] The zero mean value processing is performed on the combined acceleration modulus value data x1 to obtain the zero mean data x s ;

[0043] x s =x1-x′1

[0044] Among them, x′1 is obtained by mean filtering x1, and the filtering window length is not less than 4 times the duration of one action;

[0045] Step 2: Select the data x1=[x1(1),x1(2),...,x1(l),...,x1(L)] of the statistical action after preprocessing and zero mean processing during the continuous motion, and construct an M×N dimensional analysis matrix A; where M is the data length corresponding to one motion cycle, N is the number of motion cycles analyzed, and N>M, L>M×N;

[0046] Solve the singular values ​​and singular vectors of the analysis matrix A, and subtract its mean from the first left singular vector v1 of the analysis matrix A. For the template function g:

[0047]

[0048] According to the updated template function, the detection threshold r is determined as:

[0049]

[0050] Wherein, l1(g) and l2(g) are the l1 norm and l2 norm of the template function g; r0 is the ratio of the first singular value of the analysis matrix A to the sum of all singular values.

[0051] Step 3: When the wearer is exercising, the wearable device collects data in real time and performs preprocessing and zero-mean processing on the data; for the data x after preprocessing and zero-mean processing within the calculation period s , calculate x s Convolution with the template function g obtains the filtered data y = [y(1), y(2), ..., y(t), ..., y(T)]; if y(t) < -r, y(t+1) > r, then the combination of y(t) and y(t+1) is recorded as the number of action occurrences; the number of combinations of peaks and troughs that meet the detection threshold in the filtered data is counted and recorded as the number of action occurrences.

[0052] If the wearer changes, or the statistical result of the self-assessed action count of the wearable device is lower than expected, re-execute step 1 and re-train the template function.

[0053] Embodiment 1:

[0054] This embodiment takes the walking step counting method worn on the ankle as an example to illustrate the process, and the data sampling rate is 150Hz.

[0055] 1. Data preprocessing

[0056] Collect three-axis acceleration data and calculate the combined acceleration modulus value.

[0057]

[0058] Among them, a x (t),a y (t),a z (t) are the data of the x-axis, y-axis and z-axis of the three-axis accelerometer at time t, and t represents the time domain sampling sequence number. Perform mean filtering on x1 with a filter window length of 256 data points to obtain x′1, and then calculate the zero-mean data x s =x1-x′1.

[0059] 2. Collect pre-processed data from 20 wearers of the same type with different heights, body shapes, and movement frequencies under known continuous walking, non-continuous walking, and no-movement conditions. Collect 100 sets of data for each condition and use machine learning methods to form a decision tree. The decision tree can determine whether the collected multi-axis accelerometer data is continuous walking data under the condition of unknown movement conditions of the new wearer.

[0060] 3. For initial wearing or when the template function needs to be updated, perform template function training.

[0061] Using the configured decision tree, select the data x1=[x1(1),x1(2),...,x1(l),...,x1(L)] during continuous walking and construct an M×N dimensional analysis matrix A; where M is the data length corresponding to one movement cycle, N is the number of movement cycles analyzed, and N>M, L>M×N;

[0062] Solve the singular values ​​and singular vectors of the analysis matrix A, and subtract its mean from the first left singular vector v1 of the analysis matrix A. For the template function g:

[0063]

[0064] Calculate the ratio r0 of the first singular value of the analysis matrix A to the sum of all singular values ​​of the analysis matrix A, and calculate the detection threshold r as:

[0065]

[0066] 4. Step counting operation

[0067] After the template function training is completed, the wearable device collects data in real time during the wearer's exercise, and performs preprocessing and zero-mean processing on the data; for the data x after preprocessing and zero-mean processing within the calculation cycle s , calculate x s Convolution with the template function g obtains the filtered data y = [y(1), y(2), ..., y(t), ..., y(T)]; if y(t) < -r, y(t+1) > r, then the combination of y(t) and y(t+1) is recorded as the number of action occurrences; the number of combinations of peaks and troughs that meet the detection threshold in the filtered data is counted and recorded as the number of action occurrences.

[0068] like Figure 3 The figure shows the processed waveform and detection threshold diagram of the three-axis acceleration sensor data during continuous walking and sporadic stepping. It can be observed that there is an obvious “trough-peak-trough” in the continuous walking and sporadic stepping data segments.

[0069] 5. Statistical result evaluation and template function update

[0070] Replace the wearer with a significantly different body shape and activate the step counting result evaluation function.

[0071] The default setting period of the device, the user's device button operation and the wireless command setting are evaluated against the action count statistics.

[0072] The decision tree is used to select data segments in which the wearer walks continuously for more than 100 steps. The conventional movement frequency statistical methods (for example, peak-trough detection method, autocorrelation analysis method) are used to calculate the movement frequency of the selected data segment, and the results are compared with the calculation results of the above-mentioned "step count statistics operation". The result difference is 11%. The template function training module is automatically started and re-evaluated until the difference between the two results is less than 5%.

[0073] The present invention selects data of the action to be counted during continuous motion and constructs it into an analysis matrix, then uses the singular vectors and singular values ​​of the analysis matrix to construct a template function, uses the template function to perform template function filtering on the collected real-time data, and counts the number of combinations of peaks and troughs that meet the threshold requirements in the filtered data, which is recorded as the number of action occurrences. The present invention uses template function filtering to reduce the influence of other movements on the action to be counted, and can accurately record the number of occurrences of non-continuous actions, thereby improving the accuracy of action statistics; the template function is obtained through online data training, which can eliminate the differences in individual wearers (such as height, body shape, action frequency, etc.), and can evaluate the action number statistics results regularly or according to user instructions. If the evaluation effect is lower than expected, the template function can be retrained to eliminate the influence of reduced adaptability of the original template function caused by changes in habits and physical fitness.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A wearable quasi-periodic motion frequency counting method capable of online learning, characterized in that: The following steps are involved: Step 1: When wearing the wearable device for the first time, the wearer completes the specified action within the specified time, collects data through the wearable device, and performs preprocessing and zero-mean processing on the data; Using an initial template function stored in the wearable device, identifying data collected during a period when the wearer is in a state of continuous motion; Step 2: Select the data of the statistical action after preprocessing and zero mean processing during the continuous motion, construct the selected data into an analysis matrix, solve the singular values ​​and singular vectors of the analysis matrix, and subtract the first left singular vector from its mean as the updated template function; determine the detection threshold according to the updated template function; Step 3: When the wearer is exercising, the wearable device collects data in real time and performs preprocessing and zero-mean processing on the data; for the processed data within the calculation cycle, it is filtered through the updated template function, and the number of peaks and troughs in the filtered data that meet the detection threshold is counted and recorded as the number of action occurrences.

2. The wearable quasi-periodic motion frequency counting method capable of online learning according to claim 1 is characterized in that: The wearable device carries a multi-axis accelerometer and an initial template function, and the multi-axis accelerometer is used to collect data; the initial template function is specifically: Multi-axis accelerometer data of multiple similar wearers in known continuous motion, non-continuous motion, and no motion are collected to form a decision tree, which is updated using a machine learning process; an initial template function is used to determine whether the collected multi-axis accelerometer data is data during continuous motion under the condition of unknown motion of the new wearer.

3. The wearable quasi-periodic motion frequency counting method capable of online learning according to claim 2 is characterized in that: The data is preprocessed and zero-mean processed, specifically: According to the acceleration data collected by the multi-axis accelerometer, the combined acceleration modulus is calculated to obtain x1=[x1(1),x1(2),...,x1(t),...], where t is the data time domain serial number; The zero mean value processing is performed on the combined acceleration modulus value data x1 to obtain the zero mean data x s ; x s =x1-x1′ Among them, x1′ is obtained by mean filtering x1, and the filtering window length is not less than 4 times the duration of one action.

4. The wearable quasi-periodic motion frequency counting method capable of online learning according to claim 1 is characterized in that: In the step 2, the data x1=[x1(1), x1(2), ..., x1(l), ..., x1(L)] of the statistical action after preprocessing and zero-mean processing during continuous motion is selected to construct an M×N dimensional analysis matrix A; wherein M is the data length corresponding to one motion cycle, N is the number of motion cycles analyzed, and N>M, L>M×N.

5. The wearable quasi-periodic motion frequency counting method capable of online learning according to claim 4 is characterized in that: In step 2, the first left singular vector v1 of the analysis matrix A is selected minus its mean me v1 is the template function g; The detection threshold r is: Wherein, l1(g) and l2(g) are the l1 norm and l2 norm of the template function g; r0 is the ratio of the first singular value of the analysis matrix A to the sum of all singular values.

6. The wearable quasi-periodic motion frequency counting method capable of online learning according to claim 5 is characterized in that: In step 3, the data x after preprocessing and zero mean processing in the calculation cycle is s , calculate x s Convolution with the template function g gives the filtered data y = [y(1), y(2), ..., y(t), ..., y(T)]; if y(t) < -r, y(t+1) > r, the combination of y(t) and y(t+1) is recorded as the number of times an action occurs.

7. The wearable quasi-periodic motion frequency counting method capable of online learning according to claim 1 is characterized in that: If the wearer changes, or the statistical result of the self-assessed action count of the wearable device is lower than expected, re-execute step 1 and re-train the template function.

8. A computer device / equipment / system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images