CN114595748B - Data segmentation method for fall protection system - Google Patents

Abstract

The invention relates to a data segmentation method for a fall protection system, which comprises the following steps: acquiring a data set of falling actions and daily behavior actions of a user; preprocessing the acquired data set, intercepting the data of an inertial sensor worn by a user before the user falls to strike the ground, and constructing a training data set; training a training data set based on a multichannel convolutional neural network model MC-CNN to obtain feature maps of all channels and corresponding weights; and combining the obtained feature map with the corresponding weight and an importance mapping method to obtain the importance of each sequence position for data segmentation, thereby determining a specific area of the data segmentation and further being used for fall detection of a fall protection system. The method can clearly divide the triggering condition of fall detection and the intercepting length of the window, avoids the defect that the length of the data dividing window is determined empirically by the existing dividing method, and solves the problem that the fall protection system is difficult to divide the data.

Description

A data segmentation method for fall protection systems

Technical field

The present invention relates to the technical field of sensor data segmentation. Specifically, it is a data segmentation method for a fall protection system.

Background technique

Globally, falls among older adults are a major public health problem. Falls are the leading cause of death, disability and loss of independence in people over 65, as well as the leading cause of hospitalization (an older person is hospitalized due to a fall every 11 seconds). Therefore, automatic fall protection systems make a lot of sense.

The main purpose of a fall protection system is to accurately detect when a fall occurs so that the fall protection device can be activated before the body hits the ground to reduce fall injuries.

In general, fall protection systems are more suitable for wearable systems based on inertial sensors. At the heart of the fall protection system is the detection algorithm, which is responsible for constantly classifying any movement of the human body as a fall or an action of daily living. It has been widely verified that AI-based detection methods can achieve higher accuracy than threshold-based methods.

Because for fall protection systems based on artificial intelligence models, classification features can only be extracted from data segments of limited duration, the detection algorithm first segments segments from the data, i.e., divides the continuous data stream obtained from the inertial sensor into Multiple data segments. There are usually two methods of data segmentation. The first method is a sliding window of fixed duration. The sliding window duration defines the boundaries of the data segments, which are intercepted for further feature extraction and classification. Although this method is simple, it requires a large amount of calculation and is not suitable for fall protection systems that require high real-time performance. The second approach first sets up trigger conditions to detect potential fall events in the input data stream by searching for inertial sensor values that exceed a preset threshold. When a potential fall event is detected, data segments are intercepted through one or more windows to achieve classification. However, there is currently no clear method for how to determine the window duration, and most of it is based on experimental experience. If the window duration is too long, it will affect the real-time performance of fall detection and the protective device will not have enough time to open; if the window duration is too short, it will affect the accuracy of fall detection.

Contents of the invention

In order to solve the defects and shortcomings of the prior art, the purpose of the present invention is to provide a method for segmenting inertial sensor data. This data segmentation method is implemented through multi-channel convolutional neural network MC-CNN and importance mapping method. First, the existing data set is preprocessed, and the inertial sensor data before the human body hits the ground is intercepted and a training data set is formed. Secondly, the MC-CNN network is used for training, the model parameters are extracted, and combined with the importance mapping method, the method for classification is determined. The importance of each sequence position determines the clear range of data segmentation. This method can clearly divide the trigger conditions for fall detection and the interception length of the window, optimizing the shortcomings of the existing segmentation methods and solving the problem of segmentation of the fall protection system before impact. The technical pain points of the method.

In order to achieve the above objects, the present invention adopts the following technical solutions:

A data segmentation method for fall protection systems, including the following steps:

S1. Obtain the data set of user's falling actions and daily behaviors;

S2. Preprocess the acquired data set, intercept the inertial sensor data worn by the user before he fell and hit the ground, and form a training data set;

S3. Train the training data set based on the multi-channel convolutional neural network model MC-CNN, and obtain the feature map of all channels and the corresponding weights;

S4. Combine the obtained feature map and corresponding weights with the importance mapping method to obtain the importance of each sequence position for data segmentation, thereby determining the specific area for data segmentation;

S5. Set the corresponding data segmentation algorithm in the fall protection system according to the determined data segmentation area, which is used for fall detection of the fall protection system.

Preprocess the acquired data set as described in step S2. The preprocessing process is as follows:

S21. Calculate the three-axis inertial sensor signal amplitude vector Among them,/> and/> Represents the x, y and z inertial sensor values respectively;

S22. Establish a data set: For falling actions, the SMV peak point represents the moment when the body hits the ground, so the signal 1s before the maximum peak point is intercepted as a training data set; for daily behaviors such as slow walking, running, going up and down stairs, etc. In terms of actions, the data is intercepted at 1s time intervals as a training data set.

The inertial sensor in step S2 includes a three-axis accelerometer and a gyroscope.

The training data set is trained based on the multi-channel convolutional neural network model MC-CNN as described in step S3. The training process is:

First, the three-axis accelerometer and gyroscope signals are processed separately. After the convolution operation, the impact of the three-axis accelerometer and gyroscope signals on the classification is comprehensively considered. Then the feature maps of the two parts are merged to obtain the classification For more important areas, the cross-entropy loss function is finally used to judge the classification quality to complete the training of the data set;

In the above process, the convolution operation consists of four one-dimensional convolutions 1D Convolution and the global average pooling layer GAP, the linear fully connected layer FC, and the softmax logistic regression layer. The convolution kernels of the four one-dimensional convolutions 1D Convolution are respectively 8, 16, 32 and 64. The global average pooling layer GAP reduces the feature map dimension of each channel of the fourth convolutional layer from (1×100) to (1×1).

Combined with the importance mapping method described in step S4, the importance of each sequence position for data segmentation is obtained. The specific process is as follows:

After MC-CNN model training, the feature map of all channels and the corresponding weights can be obtained. For the input inertial sensor time series, S _k (x) represents the output sequence on channel k, and x represents the time on the sequence. Location, Represents the weight of each channel feature k to different classification c. The input of the softmax logistic regression layer is expressed as g _c , then there is:

Based on this, an importance mapping from the sequence to each category C is established, defined as M _c ;

M _c (x) represents the importance of position x in the time series for classifying the series into c.

In step S4, after realizing the visualization of M _c (x), the specific area of data segmentation is determined by summarizing the rules of its importance map, and the data fragments corresponding to this area are used for behavior classification based on the machine learning model. , to verify whether it is a falling action.

The beneficial effects of the present invention compared with the prior art:

The method of the present invention is implemented through the multi-channel convolutional neural network MC-CNN and the importance mapping method. Through preprocessing of existing data sets, the inertial sensor data before the human body hits the ground is intercepted and a training data set is formed. Then the MC-CNN network is used to train the data set, extract model parameters, and combine with the importance mapping method to determine The importance of each sequence position for classification determines the clear range of data segmentation. This method can clearly divide the trigger conditions for fall detection and the interception length of the window, avoiding the existing segmentation methods that solve the problem of data segmentation window length based on experience. The identified drawbacks solve the problem of difficulty in data segmentation in fall protection systems.

Description of the drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are: For some embodiments of the present invention, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

Figure 1 is a schematic diagram of the overall flow of a data segmentation method for a fall protection system according to the present invention;

Figure 2 is an analysis diagram of fall characteristics in the embodiment of the present invention;

Figure 3 is a schematic diagram of the composition of a multi-channel convolutional neural network in an embodiment of the present invention;

Figure 4 is a sample diagram of importance mapping in the embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present invention.

Example: See Figures 1-4.

As shown in Figure 1, a data segmentation method for a fall protection system of the present invention includes the following steps:

S1. Obtain the data set of user's falling actions and daily behaviors;

S2. Preprocess the acquired data set, intercept the inertial sensor data worn by the user before he fell and hit the ground, and form a training data set;

S3. Train the training data set based on the multi-channel convolutional neural network model MC-CNN, and obtain the feature map of all channels and the corresponding weights;

S4. Combine the obtained feature map and corresponding weights with the importance mapping method to obtain the importance of each sequence position for data segmentation, thereby determining the specific area for data segmentation;

S5. Set the corresponding data segmentation algorithm in the fall protection system according to the determined data segmentation area, which is used for fall detection of the fall protection system.

Preprocess the acquired data set as described in step S2. The preprocessing process is as follows:

S21. Calculate the three-axis inertial sensor signal amplitude vector Among them,/> and/> Represents the x, y and z inertial sensor values respectively;

S22. Establish a data set: As shown in Figure 2, for falling actions, the SMV peak point represents the moment when the body hits the ground, so the signal 1s before the maximum peak point is intercepted as a training data set; for slow walking, running, and climbing stairs For daily behavioral actions such as walking down the stairs and going down the stairs, the data is intercepted at 1s time intervals as a training data set.

The inertial sensor in step S2 includes a three-axis accelerometer and a gyroscope.

The training data set is trained based on the multi-channel convolutional neural network model MC-CNN as described in step S3. The training process is:

As shown in Figure 3, the three-axis accelerometer and gyroscope signals are first processed separately. After the convolution operation, the impact of the three-axis accelerometer and gyroscope signals on the classification is comprehensively considered, and then the feature maps of the two parts are combined. Merge to obtain areas that are more important for classification, and finally use the cross-entropy loss function to judge the classification quality to complete the training of the data set;

In the above process, the convolution operation consists of four one-dimensional convolutions 1D Convolution and the global average pooling layer GAP, the linear fully connected layer FC, and the softmax logistic regression layer. The convolution kernels of the four one-dimensional convolutions 1D Convolution are respectively 8, 16, 32 and 64. The global average pooling layer GAP reduces the feature map dimension of each channel of the fourth convolutional layer from (1×100) to (1×1); compared with the existing The difference between the CNN network and the present invention is that any pooling operation after the convolution layer is canceled, thereby ensuring that the feature map data length remains unchanged, so as to facilitate the determination of subsequent segmentation areas.

Combined with the importance mapping method described in step S4, the importance of each sequence position for data segmentation is obtained. The specific process is as follows:

After MC-CNN model training, the feature map of all channels and the corresponding weights can be obtained. For the input inertial sensor time series, S _k (x) represents the output sequence on channel k, and x represents the time on the sequence. Location, Represents the weight of each channel feature k to different classification c. The input of the softmax logistic regression layer is expressed as g _c , then there is:

Based on this, an importance mapping from the sequence to each category C is established, defined as M _c ;

M _c (x) represents the importance of position x in the time series for classifying the series into c.

In step S4, after realizing the visualization of M _c (x), as shown in Figure 4, the specific areas of data segmentation are determined by summarizing the rules of its importance map. The darker the color bar in the figure, the deeper the corresponding position sequence pair classification. The more important, in this embodiment, when the acceleration count value is less than 0.93 times the gravity acceleration, it is determined as a potential fall. From this moment on, the accelerometer and gyroscope data fragments are intercepted according to the 310ms window length for behavioral classification based on the machine learning model. , and finally verify whether it fell.

It should be noted that the scope of application of the data segmentation method of the present invention is not limited to fall detection. For other types of actions in life, this method can be trained and extracted according to specific circumstances, and then the corresponding data segmentation positions can be obtained.

The above descriptions are only preferred embodiments of the present invention and do not impose any formal restrictions on the structure of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention fall within the scope of the technical solution of the present invention.

Claims (4)

1. A data segmentation method for a fall protection system, comprising the steps of:

s1, acquiring a data set of falling actions and daily behavior actions of a user;

s2, preprocessing the acquired data set, intercepting the data of an inertial sensor worn by a user before the user falls down to strike the ground, and constructing a training data set;

s3, training a training data set based on a multichannel convolutional neural network model MC-CNN to obtain feature maps of all channels and corresponding weights;

s4, combining the obtained feature map with the corresponding weight, and obtaining the importance of each sequence position for data segmentation by an importance mapping method, so as to determine a specific region of the data segmentation;

after training through MC-CNN model, feature map and corresponding weight of all channels can be obtained, for the input inertial sensor time series, S is used for _k (x) Representing the output sequence on channel k, x represents the time position on the sequence,the input to the softmax logistic regression layer, representing the weight of each channel feature k to a different class c, is denoted g _c The following steps are:

from this, an importance map from the sequence to each class C is established, defined as M _c ；

M _c (x) Representing the importance of position x in the time series to classify the series as c;

in realizing the M _c (x) After the visualization of the data segmentation, determining a specific region of the data segmentation by summarizing the rule of the importance map of the region, and using the data segment corresponding to the region for behavior classification based on a machine learning model so as to verify whether the region is a falling action or not;

s5, setting a corresponding data segmentation algorithm in the fall protection system according to the determined data segmentation area, and using the data segmentation algorithm for fall detection of the fall protection system.

2. A data segmentation method for fall protection systems according to claim 1, wherein the preprocessing of the acquired data set in step S2 is as follows:

s21, calculating signal amplitude vectors of triaxial inertial sensorWherein (1)>Andrespectively representing the values of the inertial sensors of x, y and z;

s22, constructing a data set: for fall motion, the SMV peak represents the moment when the body hits the ground, so the signal 1s before the maximum peak is intercepted as the training dataset; for daily behavior actions such as slow walking, running, ascending stairs, descending stairs and the like, data are intercepted at time intervals of 1s and used as a training data set.

3. A data segmentation method for a fall protection system as claimed in claim 1, wherein the steps of

The inertial sensor in S2 includes a tri-axial accelerometer and a gyroscope.

4. A data segmentation method for fall protection systems according to claim 1, wherein in step S3, the training data set is trained based on a multi-channel convolutional neural network model MC-CNN, and the training process is:

firstly, processing triaxial accelerometer and gyroscope signals independently, comprehensively considering the influence of the triaxial accelerometer and gyroscope signals on classification after convolution operation, merging feature maps of the two parts to obtain a region with important bisection class, and finally judging classification quality by using a cross entropy loss function to finish training a data set;

in the above process, the Convolution operation is composed of four one-dimensional Convolution 1D Convolitions and a global average pooling layer GAP, a linear full-connection layer FC and a softmax logistic regression layer, the Convolution kernels of the four one-dimensional Convolution 1D Convolitions are respectively 8, 16, 32 and 64, and the global average pooling layer GAP reduces the feature mapping dimension of each channel of the fourth Convolution layer from (1×100) to (1×1).