CN106875630B - A kind of wearable fall detection method and system based on hierarchical classification - Google Patents

Abstract

The invention relates to a wearable fall detection method and system based on hierarchical classification. The fall detection method includes: collecting user's daily behavior data; performing synthesis, filtering and other processing on the daily behavior data to generate original data; using a sliding window mechanism to extract The time-frequency domain characteristics of the original data, generate samples, and combine the samples into a sample set; use the first layer one-class classification model to identify each sample in the sample set, and send the identified results to the second layer The weighted two-category classification model of the second layer; the weighted two-class classification model of the second layer is responsible for the weighted distribution processing, generates weighted fall samples, and sends them to the regular two-class classification model of the third layer; the regular two-class classification model of the third layer The model judges whether the user has a fall behavior according to whether the weighted fall sample meets the fall rule. The present invention realizes the accurate judgment of the user's fall behavior through the above method.

Description

A wearable fall detection method and system based on hierarchical classification

technical field

The invention relates to the fields of pervasive computing and health monitoring, in particular to a wearable fall detection method and system based on hierarchical classification.

Background technique

On January 22, 2016, Li Zhong, spokesman of the Ministry of Human Resources and Social Security, pointed out that by 2014, China’s elderly population over 60 years old had reached 210 million, accounting for 15.5% of the total population, while the United Nations standard, the elderly population over 60 years old reached 10%. It is regarded as an aging society. As the age increases, the physiological function of the elderly gradually declines, the response to accidents becomes slower and slower, and falls are more likely to occur. Falls have become the first cause of injury death among the elderly, with a high incidence and serious injuries, which bring great burdens to individuals, families and society, and have gradually become a social public issue related to the health of the elderly. Therefore, how to detect falls in real time and accurately has become a very important social issue.

Numerous wearable devices are increasingly pouring into people's daily life, and are widely used in health monitoring, sports health care and other fields. Wearable devices use embedded micro-sensors to collect data, which can effectively mine users' daily behaviors. At the same time, wearable devices have the advantages of low price, simple configuration, and easy portability, which are of great practical significance for coping with the challenges of social aging. Therefore, the present invention employs wearable devices as a tool for studying fall detection.

As a typical abnormal behavior, falling has its own characteristics. As shown in Figure 1, a fall usually includes three states: weightlessness, impact, and rest. At the beginning of the fall, the person's feet will gradually leave the ground and fall freely downward under the action of gravity. At this time, the person is in a state of weightlessness to a certain extent. When the impact action occurs, the downward velocity of the body reaches the maximum value. At this time, the body suddenly collides with the ground or other objects, making the resultant acceleration reach the maximum value instantly. For a certain period of time after the impact, regardless of the severity of the fall, the person remains relatively still. In addition, falls are often accompanied by changes in human orientation and time constraints between states. For example, the change of the orientation of the human body means that the orientation of the human body after the impact action will be different from that before the impact.

Fall detection methods based on wearable devices can be divided into threshold method, machine learning method and combination method of threshold method and machine learning method. The threshold method compares the relationship between one or several features and the corresponding threshold to identify whether the current behavior is in a certain state of falling, and then judges whether the current behavior is a fall. For example, patent CN201610058318.4 monitors human body activity status information in real time through a three-axis sensor, calculates the vector sum, and compares it with a preset threshold to determine whether a fall has occurred; patent CN201610062316.2 adopts a method based on feature quantity thresholds , calculate the resultant acceleration characteristic quantity A, synthetic angular velocity characteristic quantity W and similarity characteristic quantity S, compare with the threshold value obtained after signal vector modular integration for judgment. The machine learning method regards fall detection as a typical classification problem, and learns a classification model based on training data and uses it for fall detection. For example, patent CN201610152570.1 is based on Kalman filter and KNearest Neighbor (KNearestNeighbor, KNN) algorithm to classify and model the state of human activity, identify the type of human motion, and determine whether the module is a "fall" decision; patent CN201610083726.5 A classifier was constructed using the Support Vector Machine (SVM) algorithm; fall samples and daily activity behavior samples were obtained to form a training set, and the classifier was trained. The fall detection based on the combination method often combines the threshold method and the machine learning method to judge the fall behavior. For example, the patent CN201010285585.8 uses a class of support vector machine to make a second judgment after the threshold judgment, so as to judge whether it is a fall.

Fall detection is a cost-sensitive problem, that is, the consequences of missed detection of falls will be very serious, and a low false alarm rate is required for the model. On the other hand, frequent false alarms will arouse the disgust of users and reduce their trust in the detection system, which is not conducive to the practical application of the method, and the false alarm rate of the model is required to be as low as possible. Although there are many methods for fall detection, the existing methods are difficult to meet the requirements of low false alarm rate and low false alarm rate at the same time. There are three main reasons for this problem: 1. The existing methods do not comprehensively consider the false alarm rate and false alarm rate of the model, but use a single evaluation standard (such as accuracy); 2. Using conventional machine learning classification methods, there is no Consider the particularity of the abnormal behavior of falling; 3. Due to the high similarity between the instantaneous process of some daily behaviors (such as running, going down the stairs, etc.) and the falling behavior, the impact of noise on the data reduces the accuracy of model detection Rate.

Contents of the invention

As an important guarantee of health and safety, the consequences of fall detection are often fatal, and frequent false alarms will also cause users to resent the system. In order to effectively reduce the false alarm rate and false alarm rate of fall detection, increase the ability of fall detection methods to distinguish fall behavior, and filter the influence of noise data on the model, this invention proposes a wearable fall detection method based on hierarchical classification. Wherein the fall detection method includes:

Step 1, use the wearable motion sensor to collect the user's daily behavior data;

Step 2, performing synthesis and filtering operations on the collected daily behavior data to generate original data;

Step 3, using the sliding window mechanism to extract the time-frequency domain features of the original data, generating samples, and combining the samples into a sample set;

Step 4: use the first-level class classification model to identify each sample in the sample set, combine the fall samples whose recognition result is "fall" into a fall sample set, and send the fall sample set to the second layer The weighted two-class classification model of ;

Step 5, the weighted two-class classification model of the second layer is responsible for performing weighted distribution processing on all the fall samples in the fall sample set, generating weighted fall samples, and sending the weighted fall samples to the regular two-class classification model of the third layer;

Step 6, the rule-based classification model of the third layer judges whether the user has a fall behavior according to whether the weighted fall sample conforms to the fall rule. If it is judged to be a fall behavior, go to step 7, otherwise go to step 1;

Step 7, trigger the corresponding alarm mechanism, if you need to continue detection, go to step 1, otherwise end.

The wearable fall detection method based on hierarchical classification, in which

Establishing the one-class classification model by learning the pre-given sample set; and using the one-class classification model to identify each sample in the pre-given sample set to generate a pre-given fall sample set;

Establishing the weighted two-category classification model by learning the pre-given fall sample set;

By learning the pre-given falling rules, a two-class classification model of the rules is established.

In the wearable fall detection method based on hierarchical classification, the type of classification model is a support vector data description model, and the support vector data description model generates a hypersphere according to the predetermined sample set, and judges whether the sample is located in Within the hypersphere, if the sample is located within the hypersphere, the sample is identified as a fall sample.

In the wearable fall detection method based on hierarchical classification, the weighted binary classification model is a weighted extreme learning machine model to assign different weights to fall samples and non-fall samples in the fall sample set.

The wearable fall detection method based on hierarchical classification, wherein the fall rule is specifically,

a) Weightlessness, in the process of weightlessness, the value of the combined acceleration gradually decreases from the acceleration of gravity and tends to zero;

b) Impact. Before the impact action occurs, the downward velocity of the body has reached the maximum value. At this time, when it collides with the ground or other objects suddenly, the resultant acceleration instantly reaches a peak value exceeding twice the acceleration of gravity. value, the speed suddenly drops to zero at this time;

c) At rest, the X, Y, Z axis readings of the accelerometer and the synthetic acceleration readings are all in a stable state.

The present invention also provides a wearable fall detection system based on hierarchical classification, wherein the fall detection system includes:

The data acquisition module collects the user's daily behavior data with wearable motion sensors;

The data processing module is used to synthesize and filter the collected daily behavior data to generate raw data;

The sample generation module uses the sliding window mechanism to extract the time-frequency domain characteristics of the original data, generates samples, and combines the samples into a sample set;

The first layer of recognition module, the first layer of recognition module includes a class of classification model, used to identify each sample in the sample set, the recognition result is "fall" fall samples are combined into a fall sample set, and the The fall sample set is sent to the second layer weighting module;

The second layer weighting module, the second layer weighting module includes a weighted two-class classification model, which is used to perform weighted distribution processing on all the fall samples in the fall sample set, generate weighted fall samples, and send the weighted fall samples to the third layer judgment module;

The third layer of judging module, the third layer of judging module includes a rule-two classification model, which is used to judge whether the user has a fall behavior according to whether the weighted fall sample conforms to the fall rule, if it is judged to be a fall behavior, go to step 7, otherwise Go back to the data collection module and continue to collect the user's daily behavior data;

The alarm trigger module is used to trigger the corresponding alarm mechanism. If it is necessary to continue detection, it returns to the data acquisition module and continues to collect the user's daily behavior data, otherwise it ends.

The wearable fall detection system based on hierarchical classification, where

Establishing the one-class classification model by learning the pre-given sample set; and using the one-class classification model to identify each sample in the pre-given sample set to generate a pre-given fall sample set;

Establishing the weighted two-category classification model by learning the pre-given fall sample set;

By learning the pre-given falling rules, a two-class classification model of the rules is established.

In the wearable fall detection system based on hierarchical classification, the type of classification model is a support vector data description model, and the support vector data description model generates a hypersphere according to the predetermined sample set, and judges whether the sample is located in Within the hypersphere, if the sample is located within the hypersphere, the sample is identified as a fall sample.

In the wearable fall detection system based on hierarchical classification, the weighted binary classification model is a weighted extreme learning machine model to assign different weights to fall samples and non-fall samples in the fall sample set.

The wearable fall detection system based on hierarchical classification, wherein the fall rule is specifically,

a) Weightlessness, in the process of weightlessness, the value of the combined acceleration gradually decreases from the acceleration of gravity and tends to zero;

b) Impact. Before the impact action occurs, the downward velocity of the body has reached the maximum value. At this time, when it collides with the ground or other objects suddenly, the resultant acceleration instantly reaches a peak value exceeding twice the acceleration of gravity. value, the speed suddenly drops to zero at this time;

c) At rest, the X, Y, Z axis readings of the accelerometer and the synthetic acceleration readings are all in a stable state.

It can be seen from the above scheme that the advantage of the present invention is that, compared with the prior art, the hierarchical classification fall detection method provided by the present invention uses wearable sensor data as the target object, and uses a layered architecture to gradually solve various problems in the fall detection process. question. By resolving the multi-objective (high detection rate, low false alarm rate) complex problem of fall detection into a simple problem that can be solved independently by several single objects (the smallest enclosing sphere, the highest F-Score, and the lowest false alarm rate), then Using the divide-and-conquer strategy, the built models are finally related to each other in a layered architecture, and finally meet the overall requirements of the fall detection task.

Description of drawings

Figure 1 is a schematic diagram of the three-stage acceleration of the falling behavior;

Fig. 2 is a frame diagram of the wearable fall detection method based on hierarchical classification of the present invention;

Fig. 3 is the flow chart of the present invention's wearable fall detection based on hierarchical classification;

Fig. 4 is the schematic diagram of SVDD model of the present invention;

Fig. 5 is a schematic diagram of the DT model of the present invention.

Detailed ways

In order to make the above-mentioned features and effects of the present invention more clear and understandable, the following specific examples are given together with the accompanying drawings for detailed description as follows.

As an important guarantee of health and safety, the consequences of fall detection are often fatal, and frequent false alarms will also cause users to resent the system. In order to effectively reduce the false alarm rate and false alarm rate of fall detection, increase the ability of fall detection methods to distinguish fall behaviors, and at the same time filter the influence of noise data on the model, this invention proposes a framework for wearable fall detection methods based on hierarchical classification , as shown in Figure 2, the first layer constructs the smallest hypersphere that includes all fall samples, and accurately locks the target domain (the distribution space of fall samples); for the imbalance degree of the distribution of fall and non-fall samples in the hypersphere, by giving different Different weights are set for the samples of each category, and the second layer builds a weighted classification model to improve the overall recognition ability of the two categories; for the samples whose recognition result is "fall" in the hypersphere, the third layer checks whether the "fall" sample occurs before and after Meet the rules related to falls, thereby reducing the false alarm rate of falls.

Firstly, the user’s daily behavior data is collected through the wearable motion sensor, and the daily behavior data is processed into a sample set through pre-processing, feature extraction and other pre-processing. The sample set at this stage includes fall samples and non-fall samples. The fall samples include confirmed fall samples and suspected fall samples, and each sample in the sample set is identified by a classification model of the first layer, and the fall samples whose recognition result is "fall" are sent to the weighted two class classification model for further processing. The establishment method of this type of classification model is to learn from a pre-given sample set. Each sample in the pre-given sample set is identified with the built one-class classification model, and a pre-given fall sample set is generated for learning by the second-layer weighted two-class classification model. The pre-specified sample set can be a sample set under big data statistics, and the same is true for the establishment of the weighted two-class classification model and the regular two-class classification model. The fall data sample, as the first layer of the whole wearable fall detection method framework, aims to reduce the false alarm rate of falls;

Secondly, the weighted two-category classification model of the second layer receives the fall sample, and performs weighted distribution processing on all the fall samples to generate weighted fall samples, the purpose of which is to make all the fall samples have a higher identification ability, and This weighted fall sample is sent to the regular binary classification model in the third layer for further processing. The method for establishing the weighted two-category classification model is to learn the preset fall sample set. The weighted two-class classification model is used as the second layer of the whole wearable fall detection method framework, and the purpose is to improve the overall recognition accuracy;

Finally, the rule-two classification model of the third layer analyzes and judges whether the user has a fall behavior based on the weighted fall samples and fall rules, and triggers the corresponding alarm mechanism if it is judged to be a fall. Wherein, the establishment method of the rule-two classification model is to learn the pre-given falling rules. As the third layer of the framework of the whole wearable fall detection method, the rule-based two-class classification model aims to filter and smooth the detection results of falls, so as to further reduce the false alarm rate of falls. The analysis and judgment process is to judge whether the adjacent samples in the weighted fall sample satisfy the fall rule (for example: the human body is usually in a flat state after the impact, the orientation of the human body usually changes before and after the impact, and within a period of time after the fall The human body is in a relatively static state), if it is satisfied, it is judged that the user has fallen, and if it is not satisfied, it is judged that the user has not fallen.

Further, the flow chart of wearable fall detection based on hierarchical classification is shown in Figure 3, which mainly includes the following steps:

Step 1, using wearable motion sensors such as accelerometers and gyroscopes to collect user's daily behavior data;

Step 2, performing processing operations such as synthesis and filtering on the collected daily behavior data to generate original data;

Step 3, use the sliding window mechanism to extract the time-frequency domain features of the original data, generate samples, and combine the samples into a sample set. It should be noted that the samples generated in this step include fall samples and non-fall samples. The fall samples Can be used to train models or test;

Step 4: use the first-level class classification model to identify each sample in the sample set, combine the fall samples whose recognition result is "fall" into a fall sample set, and send the fall sample set to the second layer The weighted two-category classification model, if the recognition result is "fall", go to step 5, otherwise, go to step 1, the construction process of this one-class classification model can be, based on the fall samples in the previous sample set, carry out Learning and training, constructing a class classification model at the first layer;

Step 5, for the fall samples whose identification result is "fall" in the first layer (step 4), the weighted two-class classification model of the second layer receives the fall sample set, and conducts all the fall samples in the fall sample set Weighted distribution processing, generating weighted fall samples, and sending the weighted fall samples to the regular two-class classification model of the third layer for further processing;

Step 6, for the weighted fall sample in the sound field in the second layer (step 5), the rule-based classification model of the third layer analyzes and judges whether the user has a fall behavior according to whether the weighted fall sample conforms to the fall rule, and if it is judged to be a fall Then go to step 7, otherwise go to step 1;

Step 7, trigger the corresponding alarm mechanism, if you need to continue detection, go to step 1, otherwise end.

Wherein the class classification model can be a support vector data description model, the support vector data description model generates a hypersphere according to the predetermined sample set, and by judging whether the sample is located in the hypersphere, if the sample is located in the hypersphere Within the hypersphere, the sample is identified as a fall sample; the weighted two-class classification model is a weighted extreme learning machine model to assign different weights to the fall samples and non-fall samples in the fall sample set.

In the fall detection framework based on hierarchical classification, the following three models are involved:

1. A classification model. The present invention mainly constructs a support vector data description (Support Vector Domain Description, SVDD) model. That is to find a hyperball with center a and radius R, while including as many fall training samples as possible, the radius of the ball is required to be as small as possible;

2. Weighted two-class classification model. The present invention mainly constructs a weighted extreme learning machine (Weighted ELM) model. Based on the unbalanced training set of fall and non-fall data, a single-layer feedforward neural network model is learned by assigning different weights to samples of different categories, so that the F-Score value of the model is as high as possible;

3. Rule-based two-class classification model. Aiming at the particularity of the abnormal behavior of falling, the present invention constructs a rule set by judging that the human body is usually in a flat state after the impact, the orientation of the human body usually changes before and after the impact, and that the human body is in a relatively static state for a period of time after the fall. filter out as many suspected falls (pseudo-falls) as possible;

The present invention is not limited to above-mentioned method, and the SVDD of first layer can adopt any kind of classification algorithm; The WeightedELM of second layer can adopt any weighted two kinds of classification algorithms; Tree, DT), Random Forest (Random Forest, RF) or other rule-based classification algorithms; the F-Score of the second layer can use ROC, G-Means and other metrics to measure the overall recognition ability of the two types of data sets.

The three-layer fall detection method consists of three models from top to bottom: one-class classification model SVDD, weighted two-class classification model WeightedELM, and two-class classification model DT based on fall rules. These three models are introduced in detail below.

1. One class classification model SVDD

The goal of SVDD is to accurately obtain the distribution area of fall samples. Its basic idea is to learn a hypersphere to wrap as many fall samples as possible, and the radius of the ball will not be too large. The schematic diagram of the model of SVDD in two-dimensional space is shown in Fig. 4. The circles in Fig. 4 represent fall samples, and the crosses represent non-fall samples.

Given a training set {x _i ∈ R ^d |i=1,...,N ₁ } consisting of N ₁ fall samples, where _xi represents the i-th training sample, and the feature dimension is d. The process of solving SVDD is to find a sphere whose center is a and radius is R. The optimization problem of SVDD is as follows:

Among them, ξ _i represents the slack variable, which measures the degree of misclassification of the sample x _i during the training process; ξ represents the N _{1-dimensional vector composed of N 1 ξ i} ; T represents the matrix transpose; C is the penalty parameter, which is used to weigh the experience There are two goals of minimizing risk and maximizing generalization ability. In the prediction stage of SVDD, to judge whether a sample z ∈ R ^d is a fall, that is, to see whether the sample z is located within the hypersphere, that is, to judge whether (za) ^T (za) ≤ R ² is true, if true, the recognition of SVDD A result of "fall" identifies the sample as a fall sample; otherwise, "non-fall" identifies the sample as a non-fall sample. All fall samples judged as "fall" by SVDD will enter WeightedELM for secondary judgment.

Second, the weighted two-class classification model WeightedELM

After SVDD, the target domain (the distribution area of the fall samples) has been reduced to a relatively small sub-area. During the training process of the hierarchical classification model, only those training samples identified as "falling" by the SVDD of the first layer can enter the second layer and participate in the training of the weighted two-class classification model WeightedELM. However, the "fall" here is the predicted category of the sample by SVDD, not the true category of the sample. Since the training process of SVDD requires as many falling samples as possible, some non-falling samples are also surrounded by hyperspheres (see Figure 4). The training process of WeightedELM is to use all the fall and non-fall samples in the hyperball as the training set. For the unbalanced set consisting of all fall samples and a small number of non-fall samples contained in the target domain, first assign different weights to samples of different categories, and construct WeightedELM based on the weighted training sample set, so that the G-Means of the model can be as far as possible High, which guarantees a high discrimination ability for samples of both categories (fall and non-fall) in the target domain.

Assuming that there are a total of N ₂ fall and non-fall samples x _j in the target domain, the composed set is represented by the set {(x _j , t _j )∈R ^d ×R ² |j=1,...,N ₂ }, where t _j represents the category label of the j-th sample x _j , and the value is (1,-1) ^T or (-1,1) ^T , indicating that the sample x _j is a fall or a non-fall sample, respectively. The optimization problem of WeightedELM is as follows:

Among them, β is the independent variable to be solved; h(x) is the mapping function that maps the sample x from the original d-dimensional space to a certain high-dimensional space; the explanation of the penalty parameter can be found in the "one-class classification model SVDD"section; W is a diagonal weight matrix of N ₂ ×N ₂ , and the diagonal element W _jj of W represents the weight of sample x _j . Compared with the training samples of the large class, the training samples of the small class are usually assigned a relatively large weight according to certain rules.

3. The two-class classification model DT based on the fall rule

Using SVDD can effectively obtain the distribution area of fall samples; using Weighted ELM can accurately distinguish fall and non-fall samples in the target domain; however, due to the high similarity between the instantaneous process of individual normal behaviors (such as running, going down stairs, etc.) and falls , plus the influence of noise data, the false alarm rate of fall detection may still be high. Therefore, DT is used to re-filter the samples whose recognition result is "fall" by Weighted ELM to reduce the false alarm rate.

As a special abnormal behavior, falling has some specific rules. According to several stages of falling behavior (as shown in Figure 1), it can be seen that there are usually the following rules before and after the occurrence of conventional falling behavior: the human body is usually in a flat state after the impact, the orientation of the human body usually changes before and after the impact, and the human body after the impact usually changes. The human body is in a relatively static state for a period of time, etc. The measurement means (fall rules) adopted by the present invention is to use the motion sensor worn on the body to collect data and perceive the motion of the human body, and analyze the data according to the unique rules of the fall behavior The final decision is whether a fall occurred. Taking the three-axis accelerometer worn on the waist as an example, the commonly used indicators are the raw readings of the accelerometer's X, Y, and Z axes, and the synthetic acceleration (root mean square of the readings of the X, Y, and Z axes) readings. Specifically, the present invention utilizes one or more motion sensors worn on the user's body to effectively sense the user's motion, thereby detecting whether the user has fallen. We use waist acceleration readings as an example to describe the typical characteristics of the abnormal behavior of falling. a) Weightlessness. At the beginning of the fall, the person's feet will gradually leave the ground and fall freely downward under the action of gravity. At this time, the person is in a state of weightlessness to a certain extent. Due to the force of gravity alone, the downward velocity of the body will gradually increase before hitting the ground. In the process of weightlessness, the value of the combined acceleration gradually decreases from the acceleration of gravity (1g) and tends to zero (see the dotted line of "root mean square" in the state of weightlessness in Figure 1). b) Impact. Before the impact action occurs, the downward velocity of the body has reached the maximum value. At this time, when it suddenly collides with the ground or other objects, the resultant acceleration momentarily reaches a peak value exceeding 2g (see "Impact state" in Figure 1 Root mean square" dotted line) the highest value, at which point the speed suddenly drops to zero. c) During a certain period of time after the impact, regardless of the severity of the fall, the person will be in a relatively static state, specifically shown in the static state in Figure 1, the X, Y, and Z axes of the accelerometer The readings, as well as the synthetic acceleration readings, are in a relatively flat state. In addition, falls are often accompanied by changes in human orientation and time constraints between states. For example, the change of human body orientation means that the orientation of the human body after the impact action will be different from that before the impact (see the X, Y, Z axis readings in Figure 1 for the sign changes before and after the impact). Use DT to build a dedicated rule set to filter out samples of normal behavior and noise that are suspected of falling, and ultimately reduce the false alarm rate. A schematic diagram of the DT model is shown in Figure 5.

The following are system embodiments corresponding to the foregoing method embodiments, and this implementation manner may be implemented in cooperation with the foregoing implementation manners. The relevant technical details mentioned in the foregoing implementation manners are still valid in this implementation manner, and will not be repeated here in order to reduce repetition. Correspondingly, the relevant technical details mentioned in this implementation manner may also be applied in the foregoing implementation manners.

The present invention also provides a wearable fall detection system based on hierarchical classification, wherein the fall detection system includes:

The data acquisition module uses wearable motion sensors to collect user's daily behavior data;

The data processing module is used to synthesize and filter the collected daily behavior data to generate raw data;

The sample generation module uses the sliding window mechanism to extract the time-frequency domain characteristics of the original data, generates samples, and combines the samples into a sample set;

The first layer of recognition module, the first layer of recognition module includes a class of classification model, used to identify each sample in the sample set, the recognition result is "fall" fall samples are combined into a fall sample set, and the The fall sample set is sent to the second layer weighting module;

The second layer weighting module, the second layer weighting module includes a weighted two-class classification model, which is used to perform weighted distribution processing on all the fall samples in the fall sample set, generate weighted fall samples, and send the weighted fall samples to the third layer judgment module;

The third layer of judging module, the third layer of judging module includes a rule-two classification model, which is used to judge whether the user has a fall behavior according to whether the weighted fall sample conforms to the fall rule, if it is judged to be a fall behavior, go to step 7, otherwise Go back to the data collection module and continue to collect the user's daily behavior data;

The alarm trigger module is used to trigger the corresponding alarm mechanism. If it is necessary to continue detection, it returns to the data acquisition module and continues to collect the user's daily behavior data, otherwise it ends.

The wearable fall detection system based on hierarchical classification, where

Establishing the one-class classification model by learning the pre-given sample set; and using the one-class classification model to identify each sample in the pre-given sample set to generate a pre-given fall sample set;

Establishing the weighted two-category classification model by learning the pre-given fall sample set;

By learning the pre-given falling rules, a two-class classification model of the rules is established.

In the wearable fall detection system based on hierarchical classification, the type of classification model is a support vector data description model, and the support vector data description model generates a hypersphere according to the predetermined sample set, and judges whether the sample is located in Within the hypersphere, if the sample is located within the hypersphere, the sample is identified as a fall sample.

In the wearable fall detection system based on hierarchical classification, the weighted binary classification model is a weighted extreme learning machine model to assign different weights to fall samples and non-fall samples in the fall sample set.

The wearable fall detection system based on hierarchical classification, wherein the fall rule is specifically,

a) Weightlessness, in the process of weightlessness, the value of the combined acceleration gradually decreases from the acceleration of gravity and tends to zero;

b) Impact. Before the impact action occurs, the downward velocity of the body has reached the maximum value. At this time, when it collides with the ground or other objects suddenly, the resultant acceleration instantly reaches a peak value exceeding twice the acceleration of gravity. value, the speed suddenly drops to zero at this time;

c) At rest, the X, Y, Z axis readings of the accelerometer and the synthetic acceleration readings are all in a stable state.

Although the present invention is disclosed with the above embodiments, the specific embodiments are only used to explain the present invention, and are not intended to limit the present invention. Any person skilled in the art can make some changes without departing from the concept and scope of the present invention. and perfection, so the scope of protection of the present invention is defined by the claims.

Claims (10)

1. A wearable fall detection method based on hierarchical classification, characterized in that, the fall detection method comprises: Step 1, use the wearable motion sensor to collect the user's daily behavior data; Step 2, performing synthesis and filtering operations on the collected daily behavior data to generate original data; Step 3, using the sliding window mechanism to extract the time-frequency domain features of the original data, generating samples, and combining the samples into a sample set; Step 4: use the first-level class classification model to identify each sample in the sample set, combine the fall samples whose recognition result is "fall" into a fall sample set, and send the fall sample set to the second layer The weighted two-class classification model of ; Step 5, the weighted two-class classification model of the second layer is responsible for performing weighted distribution processing on all the fall samples in the fall sample set, generating weighted fall samples, and sending the weighted fall samples to the regular two-class classification model of the third layer; Step 6, the rule-based classification model of the third layer judges whether the user has a fall behavior according to whether the weighted fall sample conforms to the fall rule. If it is judged to be a fall behavior, go to step 7, otherwise go to step 1; Step 7, trigger the corresponding alarm mechanism, if you need to continue detection, go to step 1, otherwise end. 2. the wearable fall detection method based on hierarchical classification as claimed in claim 1, is characterized in that, Establishing the one-class classification model by learning the pre-given sample set; and using the one-class classification model to identify each sample in the pre-given sample set to generate a pre-given fall sample set; Establishing the weighted two-category classification model by learning the pre-given fall sample set; By learning the pre-given falling rules, a two-class classification model of the rules is established. 3. The wearable fall detection method based on hierarchical classification as claimed in claim 2, wherein the class classification model is a support vector data description model, and the support vector data description model generates according to the predetermined sample set A hypersphere, and determine whether the sample is located within the hypersphere, and if the sample is located within the hypersphere, identify the sample as a fall sample. 4. The wearable fall detection method based on hierarchical classification as claimed in claim 2, wherein the weighted two-class classification model is a weighted extreme learning machine model, so that the fall samples and non-fall samples in the fall sample set Assign different weights. 5. The wearable fall detection method based on hierarchical classification as claimed in claim 1, wherein the fall rule is specifically: a) Weightlessness, in the process of weightlessness, the value of the combined acceleration gradually decreases from the acceleration of gravity and tends to zero; b) Impact. Before the impact action occurs, the downward velocity of the body has reached the maximum value. At this time, when it suddenly impacts with the ground or other objects, the resultant acceleration instantly reaches a peak value exceeding twice the acceleration of gravity. , at this moment the velocity suddenly drops to zero; c) At rest, the X, Y, Z axis readings of the accelerometer and the synthetic acceleration readings are all in a stable state. 6. A wearable fall detection system based on hierarchical classification, characterized in that the fall detection system comprises: The data acquisition module collects the user's daily behavior data with wearable motion sensors; The data processing module is used to synthesize and filter the collected daily behavior data to generate raw data; The sample generation module uses the sliding window mechanism to extract the time-frequency domain characteristics of the original data, generates samples, and combines the samples into a sample set; The first layer of recognition module, the first layer of recognition module includes a class of classification model, used to identify each sample in the sample set, the recognition result is "fall" fall samples are combined into a fall sample set, and the The fall sample set is sent to the second layer weighting module; The second layer weighting module, the second layer weighting module includes a weighted two-class classification model, which is used to perform weighted distribution processing on all the fall samples in the fall sample set, generate weighted fall samples, and send the weighted fall samples to the third layer judgment module; The third layer of judging module, the third layer of judging module includes a rule two classification model, used to judge whether the user has a fall behavior according to whether the weighted fall sample conforms to the fall rule, if it is judged to be a fall behavior, it will turn to the alarm trigger module, otherwise Then call the data collection module to continue to collect the user's daily behavior data; The alarm trigger module is used to trigger the corresponding alarm mechanism. If it is necessary to continue detection, it returns to the data acquisition module and continues to collect the user's daily behavior data, otherwise it ends. 7. The wearable fall detection system based on hierarchical classification as claimed in claim 6, characterized in that, Establishing the one-class classification model by learning the pre-given sample set; and using the one-class classification model to identify each sample in the pre-given sample set to generate a pre-given fall sample set; Establishing the weighted two-category classification model by learning the pre-given fall sample set; By learning the pre-given falling rules, a two-class classification model of the rules is established. 8. The wearable fall detection system based on hierarchical classification as claimed in claim 7, wherein the class classification model is a support vector data description model, and the support vector data description model generates A hypersphere, and by judging whether the sample is located within the hypersphere, if the sample is located within the hypersphere, the sample is identified as a fall sample. 9. The wearable fall detection system based on hierarchical classification as claimed in claim 7, wherein the weighted two-class classification model is a weighted extreme learning machine model, so that the fall samples and non-fall samples in the fall sample set Assign different weights. 10. The wearable fall detection system based on hierarchical classification as claimed in claim 6, wherein the fall rule is specifically: a) Weightlessness, in the process of weightlessness, the value of the combined acceleration gradually decreases from the acceleration of gravity and tends to zero; b) Impact. Before the impact action occurs, the downward velocity of the body has reached the maximum value. At this time, when it collides with the ground or other objects suddenly, the resultant acceleration instantly reaches a peak value of more than twice the acceleration of gravity. value, the speed suddenly drops to zero at this time; c) At rest, the X, Y, Z axis readings of the accelerometer and the synthetic acceleration readings are all in a stable state.