CN115530800A - Falling detection system based on bracelet - Google Patents

Abstract

The invention discloses a fall detection system based on a bracelet, which includes: firstly, collecting fall data through an acceleration sensor built in the bracelet; then, dividing the collected fall data according to a set time window, and pre-set The processed data is filtered and analyzed in the time-frequency domain, time domain and instantaneous frequency to extract features; finally, the extracted features are used for model training, and the trained model is deployed for real-time fall detection deal with. The present invention adopts the implementation scheme of separately analyzing the time domain, time-frequency domain and instantaneous frequency of the collected data. Compared with other existing technical schemes, the technical scheme provided by the present invention considers more comprehensively, and the robustness of the system Even better, fall detection processing can be implemented quickly and accurately.

Description

A wristband-based fall detection system

technical field

The invention relates to the technical field of intelligent electronics, in particular to a fall detection system based on a bracelet.

Background technique

With the continuous advancement of science and technology and the improvement of living standards, people's awareness of health protection has also been continuously improved. At the same time, the aging of the population has gradually attracted the attention of the society. Moreover, many developed countries in the world are already in the aging society, and the proportion of the elderly over 60 years old in Germany, Italy and Japan has exceeded 20%. The aging of the population will cause many real social problems, among which the medical care of the elderly needs to be solved urgently. How to effectively solve the medical care of the elderly has attracted more and more attention from the international community, and the problem of falls needs special attention.

A fall is defined as a sudden and unexpected fall to the ground. This phenomenon can occur at any age, but it is more common in the elderly. Compared with the harm to the young, the harm of the elderly after a fall is greater and serious. Threat to the health of the elderly. Especially for the empty-nest elderly who are not with their children, they may not be able to call for help after a fall, miss the golden treatment time, and fail to receive timely treatment, which may cause more serious injuries, such as disability or even death. Therefore, it is very necessary for the detection and alarm of falls.

Based on the above requirements, how to design an accurate and real-time fall detection and alarm system has become a research focus of researchers at home and abroad. At present, according to the different means of obtaining fall feature information data, fall detection technologies are mainly divided into the following three categories:

(1) User-initiated alarm system

The user-initiated alarm system means that after the user falls, he can independently report to his family and medical institutions through the system, so as to receive timely treatment. Usually, the alarm is triggered by a button, and the user only needs to press the button to complete the alarm quickly. This system is often placed on watches, hanging ornaments, or places where people are prone to falls such as bathrooms. It has the advantages of low price and simple operation, but the corresponding disadvantages are also very obvious. The user must independently press the button to activate the alarm system, which means that the user must have a clear consciousness to operate the system. However, the elderly who have fallen are most likely to be unable to call the police autonomously due to fainting or fainting. In addition, if the user suffers from Alzheimer's or other mental illnesses, the system cannot be used normally. Due to the above two reasons, the scope of use of the system is restricted.

(2) Fall detection system based on video device

A fall detection system based on a video device refers to a system that monitors the user's body posture through intelligent video monitoring technology, and effectively recognizes and alarms in time when a fall occurs. The system is usually placed in the area where the user frequently moves, and does not need to be carried around, and does not affect the user's daily life. At the same time, the system does not require user operation after detecting a fall, and can complete the alarm in time even after the user loses consciousness. Although the research progress of the fall detection system based on the video device is considerable, it has also achieved good application results. But the biggest limitation of this system is that the video device must be placed in the area where the user frequently moves, so that the falling behavior in other areas cannot be captured. In addition, the system needs to collect video information in the placement area for a long time, which may infringe the user's privacy. The above-mentioned reasons restrict the application effect of the system.

(3) Fall detection system based on wearable devices

A wearable device-based fall detection system refers to a wearable device embedded with micro sensors to detect whether a user has fallen, such as a smart bracelet, smart watch, etc. The system can monitor human activities in real time for a long time, collect data on human activities, and use corresponding algorithms to judge whether it has fallen. Compared with the previous two systems, this system can be applied to the scene where the user loses consciousness after falling, and it can also protect the elderly with dementia or other mental diseases. At the same time, the system has no restrictions on the active area, which maximizes the protection of user privacy. Therefore, this system is an ideal fall detection system.

However, at present, the above-mentioned fall detection systems based on wearable devices need to collect a large amount of data for analysis and processing, which greatly increases the amount of processing and calculation of wearable devices, and the result is that the user's fall behavior cannot be quickly detected. . In addition, a large amount of data collection requirements also make the implementation of the product more complex, leading to an increase in the cost of technology implementation, which affects the promotion and application of the corresponding products.

Contents of the invention

The purpose of the present invention is to provide a fall detection system based on a wristband, which can realize rapid and accurate fall detection processing based on less data collection.

The purpose of the present invention is achieved through the following technical solutions:

A bracelet-based fall detection system comprising:

The data acquisition module collects fall data through the built-in acceleration sensor in the wristband;

The data preprocessing module divides the collected fall data according to the set time window, and the window refers to whether the fall can be identified within the time period where the window is located;

The feature extraction module performs filtering processing on the preprocessed data, and performs time-frequency domain, time domain and instantaneous frequency analysis to extract features, and the features are information that can identify and distinguish whether a fall behavior occurs;

The model training and deployment module uses the extracted features for model training, and deploys the trained model for real-time fall detection processing.

The time window used in the segmentation process by the data preprocessing module is 5 seconds.

The filtering process of the feature extraction module includes filtering out frequencies above 17Hz, and the module specifically includes:

Extract the spectral feature sub-module, and extract the spectrum through the short-time fast Fourier transform, wherein the window length of the fast Fourier transform is set to 50 sampling points, and the length of each sliding is 25 sampling points, corresponding The window function of chooses Hamming window;

Extract time-domain features sub-module, the time-domain features used for extraction include the maximum peak value during the fall, the minimum trough value during the weightlessness stage, the maximum peak duration, the minimum trough duration, the active ratio, the acceleration fluctuation after the fall and the impact, and the axle standard after the fall and impact Difference;

Extract the instantaneous frequency feature sub-module, extract 0-12Hz data, and determine its proportion in the instantaneous frequency.

The maximum peak duration is: take the maximum peak as the center point to judge whether the peak value is greater than the set first threshold to both sides, and compare the two peak points on both sides that are farthest from the maximum peak value greater than the set first threshold The length of time between them is taken as the maximum peak duration. The minimum trough duration is: judge whether the trough value is less than the set second threshold from the center point of the minimum trough value to both sides, and separate the two sides from the minimum trough value The time length between the two minimum trough points that are farthest away from the set second threshold is taken as the minimum trough duration.

The active ratio is the ratio of the active time to the total time, the active time refers to the total active time that meets the active moment conditions, and the active moment refers to the acceleration at a certain moment is outside the preset threshold range.

The active ratio is the ratio of the number of active sampling points to the total number of sampling points, and the active sampling point refers to the sampling point at the above-mentioned active moment.

The extraction of post-fall acceleration fluctuations in the time-domain feature extraction sub-module includes: extracting post-fall peak density, post-fall trough density, post-fall peak-valley pair density, and post-fall acceleration standard deviation; and/or, the The extraction of the standard deviation of the axis after the fall and impact in the sub-module of extracting time domain features includes: extracting the minimum axis standard deviation of the fluctuation after the fall and the impact and the maximum axis standard deviation of the front and rear changes after the fall and impact.

The extracted peak density after fall and impact is: the ratio of the number of peaks to the number of sampling points; the density of valleys after falling and impact is: the ratio of the number of valleys to the number of sampling points.

When the acceleration sensor in the wristband is a three-axis acceleration sensor, the data used by the feature extraction module in the process of extracting features is the resultant acceleration data determined based on the three-axis acceleration data.

During the model training process of the model training and deployment module, the features extracted by the feature extraction module are put into the SVM classifier for training, and the trained model is obtained.

It can be seen from the above-mentioned technical solution provided by the present invention that it specifically analyzes the time domain, time-frequency domain, and instantaneous frequency according to the collected accelerometer data, and extracts effective features, which is of great significance to the behavior perception work based on commercial wristbands. big help. Moreover, the present invention also preprocesses the data to compress the multi-dimensional data collected by the sensor into one dimension, which significantly reduces the calculation amount of the system and effectively improves the fall detection efficiency. At the same time, the time domain, time-frequency domain, and instantaneous frequency of the collected data are analyzed separately, so that compared with other existing technical solutions, the technical solution provided by the present invention is considered more comprehensive and the robustness of the system is better .

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings that need to be used in the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For Those of ordinary skill in the art can also obtain other drawings based on these drawings on the premise of not paying creative efforts.

FIG. 1 is a schematic structural diagram of a system provided by an embodiment of the present invention;

Fig. 2 is a schematic diagram of the process of extracting spectral features in an embodiment of the present invention;

Fig. 3 is a schematic diagram of short-window frequency resolution difference;

Fig. 4 is the schematic diagram of long window frequency resolution difference;

Fig. 5 is a schematic diagram of a sub-band waveform of a fall action;

Fig. 6 is a schematic diagram of instantaneous frequency statistics of non-fall data;

Fig. 7 is a schematic diagram of instantaneous frequency statistics of fall data.

detailed description

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

The purpose of the present invention is to provide a fall detection system based on a commercial bracelet, which can analyze and identify whether a user has fallen by collecting data from the commercial bracelet. The system application scenario provided by the present invention can be but not limited to Mi Band 3. In practical application, the user can wear the Mi Band 3 on the left hand, and then when the user falls, the system can quickly detect the occurrence of the fall, so as to give a timely alarm.

The specific implementation framework of the system provided by the present invention is shown in Fig. 1, which is mainly divided into processing modules such as data acquisition, data preprocessing, feature extraction, classifier classification (model training and deployment).

The specific implementation of each processing module in FIG. 1 will be described in detail below, taking the embodiment of the present invention specifically applied to Mi Band 3 as an example. Obviously, the embodiments of the present invention can also be applied to other wristbands with similar data collection functions.

Referring to Fig. 1, the embodiment of the present invention may specifically include the following processing modules:

(1) Data acquisition module

In the process of implementing the present invention, data can be collected based on Xiaomi Mi Band 3, which is equipped with a three-axis acceleration sensor with a sampling rate of 25Hz and has a battery life of 20 days, so it can be used to collect data , that is, a system that collects fall data through the built-in acceleration sensor in the wristband, and performs subsequent fall detection based on the collected data.

(2). Data preprocessing module

The data preprocessing module is mainly used to segment the collected fall data according to the set time window;

Specifically, in order to identify falls more accurately, a time window can be set to segment the fall data collected by Mi Band 3. This time window refers to the length of time that the algorithm can identify whether a fall occurred within a certain period of time. In the process of determining the window length, it is found that if the time window is set too long, the amount of data to be processed will be too large, and the amount of calculation will be large. At the same time, the proportion of falling actions on the time axis is low, and the accuracy of the algorithm If the time window is set too short, it will not be enough to record the entire fall process, and the fall action will be difficult to recognize. Therefore, selecting an appropriate time window length is the basis for the excellent performance of the entire algorithm. Therefore, according to the characteristics and duration of the falling action, and after a large number of experiments and theoretical analysis, it is finally determined that the time window can be set to about 5s, so that the length of the time window can not only cover all stages of the falling action, but also can Make sure other actions are less noisy as well. Moreover, experiments have also proved that it can achieve the best results.

(3). Feature extraction module

In the embodiment of the present invention, the timing signal of the three-axis acceleration sensor is collected by using Mi Band 3, and the features are extracted by analyzing the time-frequency domain, time domain, and instantaneous frequency, wherein the feature refers to the ability to identify and distinguish whether the Information on Falling Behavior.

(1) The time-frequency domain analysis part, that is, the spectral feature extraction sub-module

In time-frequency domain analysis, spectral features are very important information in time-series signals. Specifically, short-time fast Fourier transform can be used to process data. The corresponding processing flow for extracting spectral features is shown in Figure 2. Considering the calculation amount of the algorithm, it is not necessary to analyze the data of each axis, but to use the combined acceleration data obtained by processing the three-axis acceleration data. The combined acceleration formula is as follows:

Based on the above description, the specific spectral feature extraction process is shown in Figure 2, including:

(11) Filter processing

Research has found that the frequency domain of the acceleration time series signal of human activity is usually below 17Hz, so a corresponding 17Hz low-pass filter can be set to filter out the noise generated by non-human activity. In a specific implementation process, when the short-time fast Fourier transform is used, the cutoff frequency can be set to 17 Hz, that is, the part exceeding 17 Hz after the STFT (short-time fast Fourier transform) is discarded.

(12) Short-time fast Fourier transform processing

In the short-time fast Fourier transform process, the length of the window will determine the time resolution and frequency resolution of the spectrogram. The longer the window length, the longer the intercepted signal, the higher the frequency resolution after Fourier transform, and the worse the time resolution; on the contrary, the shorter the window length, the shorter the intercepted signal, the worse the frequency resolution, and the worse the time resolution. The better the rate, that is to say, in the short-time Fourier transform, the time resolution and the frequency resolution cannot be achieved at the same time, as shown in Figure 3 and Figure 4. Therefore, the length of the window will directly affect the effect of the algorithm.

For this reason, according to a large number of experiments and combined with experience, it is determined that the window length of the short-time fast Fourier transform can be set to 50 sampling points, and the length of each sliding is 25 sampling points, and the window function is selected as Hamming window.

After the above processes (11) and (12), the spectral features can be extracted from the original data of the time series signal.

(2) Extract time domain feature sub-module

In order to discuss the time-domain characteristics of the falling action more conveniently, the falling behavior is specifically divided into five stages in the present invention: stable stage, shaking stage (unstable stage), weightless stage (falling stage), impact stage and observation stage after falling . The human body acceleration waveform of the fall action has remarkable characteristics. Usually, a global minimum trough will be generated during the weightlessness stage; a larger peak will be generated when the weightlessness stage ends and the impact begins. Afterwards, due to the rebound caused by the collision with the ground, the acceleration will continue to change, and a certain number of peaks and troughs will appear in the waveform, as shown in Figure 5 for details.

By studying the physical characteristics of each stage of the fall action, observing the waveform characteristics of the fall action data, and conducting a large number of experiments and detection analysis, it is decided to select the features required by the threshold detection algorithm from the following aspects.

(21) Maximum peak value max_acc

During a fall, the human body undergoes a period of weightlessness during which the human body has a gradually increasing velocity towards the ground. When the human body hits the ground, the speed of the human body will decrease rapidly, so a large acceleration will be generated. This feature is manifested in the waveform as a higher peak during the impact phase. Based on this feature, the maximum peak can be selected as a feature in the time domain, namely:

max_acc = max(acc);

(22) The minimum valley value min_acc in the weightlessness stage

Compared with other daily actions, the falling action will produce a greater degree of weightlessness, so after the acceleration of the device coordinate system and the acceleration of gravity are canceled out, a very small acceleration value will be produced. On the oscillogram, this feature manifests itself as a smaller trough during the weightlessness phase. Based on this feature, it is determined that the minimum valley value can be selected as a feature in the time domain feature, namely:

min_acc = min(acc);

(23) Maximum peak duration peak_duration

In the case of non-fall, although the human body may produce a peak acceleration similar to that of the impact, this may meet the previously set condition (21). However, users in a normal state will subconsciously maintain body balance when their bodies are unstable, so the duration of this peak will be relatively short, so the maximum peak duration can be added to the feature as a supplement to reduce the selection of the maximum peak as a feature The resulting false alarm rate. In the specific implementation process, the maximum peak value can be used as the center point to judge whether the maximum peak duration is greater than the preset threshold. The boundary point on the left side meeting the conditions is ps, and the boundary point meeting the conditions on the right side of the peak is pe, then the maximum peak duration is:

peak_duration = pe – ps;

That is to say, the maximum peak duration is as follows: take the maximum peak as the center point to judge whether the peak value is greater than the first set threshold on both sides, and determine whether the peak on both sides is farthest from the maximum peak is greater than the first set threshold. The time length between two peak points is taken as the maximum peak duration;

(24) Minimum trough duration trough_duration

Similar to (23), in the case of non-falling, although the human body may also produce an acceleration trough similar to that of weightlessness, the user in a normal state will subconsciously maintain body balance when the body is unstable, so this The duration of the trough will be relatively short. Therefore, the minimum trough duration is added to the feature as a supplement to reduce the false alarm rate caused by selecting the minimum trough value as a feature. In the specific process, the minimum trough can be used as the center point to judge whether it is smaller than the set threshold on both sides. For example, the set threshold can be the minimum trough*1.5 in (22), assuming that the left side of the trough meets the conditions The boundary point is ts, and the qualified boundary point on the right side of the trough is te, then the minimum trough duration is:

trough_duration = te – ts;

That is to say, the minimum trough duration is: judge whether the trough value is less than the set second threshold value from the center point of the minimum trough value to both sides, and determine whether the trough value on both sides is farthest from the minimum trough value is less than the set second threshold value. The time length between the two minimum trough value points of the second threshold is taken as the minimum trough duration;

(25) Active rate activity_rate

Each of the above characteristics takes into account the characteristics of the peaks and troughs of the data, and the characteristics of the peaks and troughs of the fall data described above are also very easy to occur when the user is exercising vigorously. Taking running as an example, the acceleration of the human body changes greatly all the time, and it is likely to be recognized as a fall by the above rules. Based on this, after research and analysis, it is determined that a new feature to be extracted is introduced, that is, the activity ratio. The active ratio is the ratio of the active time to the total time, the active time refers to the total active time that meets the active moment conditions, and the active moment refers to the acceleration at a certain moment is outside the preset threshold range.

Specifically, a threshold interval [low_threshold, high_threshold] can be set. For example, the specific value of this interval can be set to [max_acc*0.2, max_acc*0.6]. When the acceleration is outside the threshold interval at a certain moment, it is not in the interval , it is called the active moment. Active ratio is defined as active time/total time. In particular, since the lengths of the falling phases of different data are inconsistent, in order to measure uniformly, the active ratio can be set as the number of active sampling points/total number of sampling points, namely:

activity_rate=activity_number/length(acc);

Wherein, the active sampling point refers to the sampling point at the above-mentioned active moment;

(26) Acceleration fluctuates after a fall and impact

When a fall occurs, the human body will receive a reaction force after hitting the ground, bounce off the ground, and then hit the ground again. This situation may usually be repeated several times, and the magnitude gradually decreases. Judging from the characteristics of the acceleration, this phenomenon is reflected in the constant fluctuation of the acceleration. Generally speaking, the acceleration is constantly changing, and the degree of dispersion of the data is relatively large. In probability statistics, standard deviation is usually used to reflect the degree of dispersion of data. In the present invention, the standard deviation of acceleration after a fall and impact is selected as a feature.

However, judging from the characteristics of the waveform image shown in Figure 5, acceleration fluctuations appear as multiple peaks and troughs after the impact. In order to reflect this feature, the number of crests and troughs after a fall and impact can be calculated. Since the length of the falling stage of different data is different, in order to unify the measurement, the density rather than the number can be used in the specific operation. Specifically, the following features can be selected:

Peak density after fall and impact, that is, the number of peaks/number of sampling points;

The trough density after the fall and impact, that is, the number of troughs/the number of sampling points;

The density of peak and trough pairs after a fall and impact, that is, the number of peak and trough pairs/number of sampling points.

The number of sampling points after the fall is length(acc_after_peak), and the numbers below are the qualified sampling points after the fall (that is, after the maximum peak).

A qualified peak is defined as: it needs to be larger than the previous sampling point and larger than the next sampling point; meanwhile, it needs to be larger than a preset threshold, for example, the preset threshold can be 12.

The qualified valley is defined as: it needs to be smaller than the previous sampling point and smaller than the next sampling point; meanwhile, it also needs to be smaller than a preset threshold, for example, the preset threshold is 10.

Based on this, the above selected feature expressions are as follows:

peak_density=peak_number/length(acc_after_peak);

trough_density=trough_number/length(acc_after_peak);

peak_trough_density=peak_trough_number/length(acc_after_peak);

(27) There is little fluctuation in the acceleration of one axis after a fall and impact

Usually, a human body will temporarily lose mobility after a fall, resulting in a small change in the acceleration data of at least one coordinate axis after the fall. Therefore, this phenomenon can be reflected by calculating the standard deviation of the axis with the least fluctuation after a fall, called the standard deviation of the axis with the least fluctuation after fall and impact, and using it as a time-domain feature. At the same time, the standard deviation of the acceleration after the fall can also be used as a feature, the two are:

min_std = min(x_std, y_std, z_std);

acc_std = std(acc_after_peak);

Among them, x_std, y_std, and z_std represent the standard deviation of the data after the fall on the x, y, and z axes, respectively;

(28) There is a large change in the acceleration of one axis before and after the fall and impact

Because a fall will cause a change in the human body posture, this change is reflected in the data of the three-axis acceleration sensor, that is, the data of one axis has a relatively large change. In terms of specific data processing, a representative point can be selected for comparison before and after the fall. For example, the representative points can be selected to be the first peak delta_before on the left side of the largest peak and the first peak on the right side of the largest peak delta_after, to find the acceleration change of the coordinate axis with the largest acceleration change, as a time domain feature, namely:

max_delta=max(x _{delta_after-} x _{delta_before} , y _{delta_after} -y _{delta_before} , z _{delta_after} -z _{delta_before} ).

It can be seen from the above description that a total of 11 features are selected as time-domain features in the embodiment of the present invention, and the summary is shown in Table 1 below:

Table 1

(3) Extract the instantaneous frequency feature sub-module

Specifically, in the embodiment of the present invention, the Hilbert transform can be used to estimate the instantaneous frequency, so as to reduce the false alarm rate of the model on non-fall data. After the Hilbert transform is performed on the data, the instantaneous frequency can be obtained, and then rounded, and the proportion of different frequencies in the instantaneous frequency of the non-fall data and the fall data is counted, and it can be found that the proportion of 0Hz and 12Hz is higher than that of the fall data and the non-fall data. The difference in the fall data is obvious, as shown in Figure 6 and Figure 7 for details, so the proportion of 0Hz and 12Hz in the instantaneous frequency can be used as the instantaneous frequency feature.

(4) Model training and deployment module

After completing the extraction of the above features, the extracted features can be input into the SVM (Support Vector Machine) classifier for training, and the corresponding trained model can be obtained. After the trained model can be deployed to the bracelet China and Israel establish a real-time fall detection system, so that real-time fall monitoring and processing can be realized.

After the deployment of the model is completed, the fall monitoring process can be performed, and the monitoring process results can be output through the output part, for example, an alarm can be set in a set way, and so on.

The above-mentioned specific application examples provided by the present invention are implemented based on low-cost commercial wristbands, but they can still achieve effective fall detection under the condition of data collected by low-sampling rate equipment, and can ensure that the detection effect is better than other detection schemes , that is, a high-precision fall detection function can be realized, so that the fall detection can be more universal and convenient.

After the embodiment of the present invention is implemented in the Mi Band 3, the user wears the Mi Band 3 with his left hand. After the user simulates the fall behavior, the fall behavior can be detected in time, and the alarm will be triggered quickly after the fall behavior is detected.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person familiar with the technical field can easily conceive of changes or changes within the technical scope disclosed in the present invention. Replacement should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the protection scope of the claims.

Claims (10)

1. A bracelet-based fall detection system, comprising:

the data acquisition module acquires falling data through an acceleration sensor arranged in the bracelet;

the data preprocessing module is used for dividing the acquired falling data according to a set time window, wherein the window is used for identifying whether the falling data fall or not in a time period of the window;

the feature extraction module is used for filtering the preprocessed data, analyzing a time-frequency domain, a time domain and an instantaneous frequency, and extracting features, wherein the features are information capable of identifying whether a falling behavior occurs or not;

and the model training and deployment module is used for performing model training by using the extracted features and deploying the model obtained by training so as to perform real-time fall detection processing.

2. The system of claim 1, wherein the time window used in the segmentation by the data pre-processing module is 5 seconds.

3. The system according to claim 1 or 2, wherein the filtering process of the feature extraction module includes filtering out frequencies above 17Hz, and the module specifically includes:

the frequency spectrum characteristic extraction sub-module extracts a frequency spectrum through short-time fast Fourier transform, wherein the window length of the fast Fourier transform is set to be 50 sampling points, the sliding length of each time is 25 sampling points, and a Hamming window is selected as a corresponding window function;

the time domain characteristic extraction submodule is used for extracting time domain characteristics, wherein the time domain characteristics comprise a maximum peak value during falling, a minimum valley value in a weightlessness stage, maximum peak value duration time, minimum valley duration time, an activity ratio, acceleration fluctuation after falling impact and a standard deviation of a shaft after falling impact;

and extracting an instantaneous frequency characteristic submodule, extracting 0-12Hz data and determining the proportion of the data in the instantaneous frequency.

4. The system of claim 3, wherein the maximum peak duration is: judging whether the peak value is larger than a set first threshold value or not from two sides by taking the maximum peak value as a central point, and taking the time length between two peak value points which are farthest away from the maximum peak value and are larger than the set first threshold value at two sides as the maximum peak value duration time; the minimum trough duration is: and judging whether the valley value is smaller than a set second threshold value or not from two sides by taking the minimum valley value as a central point, and taking the time length between two minimum valley value points which are farthest away from the minimum valley value and are smaller than the set second threshold value at two sides as the minimum valley duration.

5. The system according to claim 3, wherein the active ratio is a ratio of an active time length to a total time length, the active time length is a total active time length meeting an active time condition, and the active time is an acceleration at a certain time which is outside a preset threshold interval.

6. The system of claim 5, wherein the active ratio is a ratio of the number of active sampling points to the total number of sampling points, and the active sampling points are the sampling points at the active time.

7. The system of claim 3, wherein the extracting temporal features submodule extracts the post-fall impact acceleration fluctuations comprising: extracting the peak density after falling impact, the trough density after falling impact, the peak and trough pair density after falling impact and the acceleration standard deviation after falling impact; and/or the step of extracting the standard deviation of the falling impact rear axle in the time domain feature extraction submodule comprises the following steps: and extracting the fluctuation minimum axis standard deviation after falling impact and the front-back change maximum axis standard deviation after falling impact.

8. The system of claim 7, wherein the extracted post-fall impact peak density is: the ratio of the number of wave crests to the number of sampling points; the trough density after falling and impacting is: the ratio of the number of the wave troughs to the number of the sampling points.

9. The system of claim 3, wherein when the acceleration sensors in the bracelet are triaxial acceleration sensors, the data used by the feature extraction module in extracting features is resultant acceleration data determined based on the triaxial acceleration data.

10. The system according to claim 3, wherein the model training and deployment module performs a model training process, specifically, the features extracted by the feature extraction module are placed into an SVM classifier for training, and a trained model is obtained.

Images