A fast self-built behavioral mode recognition system

Abstract

The present invention provides a fast self-built behavior modal recognition system, which belongs to the field of behavior modal recognition. The system comprises: a behavior modal data acquisition and processing module, which is used to acquire human behavior modal data by using a mobile device, and perform feature engineering processing on the acquired data, thereby obtaining behavior modal features; a behavior modal recognition model establishment module, which establishes a behavior modal recognition model based on the behavior modal features by using a fast self-built method; and a behavior modal recognition module, which inputs the behavior modal features to be tested into the behavior modal recognition model, thereby obtaining the category of the behavior modal features to be tested. The present invention makes up for the problems of the existing behavior modal recognition system in terms of slow construction speed, low recognition accuracy and insufficient lightweight.

Description

Technical Field

[0001] The present invention relates to the field of human behavior modality recognition, and in particular to a fast self-constructed behavior modality recognition system. Background Art

[0002] Human behavior modality recognition as a biometric authentication technology has been widely studied and developed. Among them, human behavior modality recognition based on mobile devices is the most popular research direction. This is mainly due to the rapid development of the Internet of Things and microelectronic computing. At present, this type of research has been applied to many healthcare fields such as fall detection for the elderly and patient physical function recovery testing. However, due to the particularity of mobile devices themselves, they are limited in storage and computing power. Therefore, it is of great practical significance to study a behavioral modality recognition system with fast modeling speed, high accuracy and lightweight. Summary of the invention

[0003] Purpose of the invention: The technical problem to be solved by the present invention is to provide a fast self-built behavioral mode recognition system in view of the shortcomings of the prior art, which mainly makes up for the shortcomings of the existing methods in terms of modeling speed, recognition accuracy and lightweight. The present invention specifically includes the following steps:

[0004] Step 1: Establish a behavioral modality data acquisition and processing module, which is used to perform the following steps:

[0005] Step 1-1, obtain the acceleration and gyroscope sensor data of six behavioral modes, namely, running, walking, going downstairs, going upstairs, lying down, and sitting, through a mobile device, and combine the acceleration and gyroscope sensor data of the behavioral mode and the behavioral mode category into a behavioral mode data set D = {x, f}, where x represents the acceleration and gyroscope sensor data of the behavioral mode, and f represents the behavioral mode category label;

[0006] Step 1-2: Use Gaussian filtering and maximum and minimum normalization techniques to denoise and normalize the behavioral modality dataset D to obtain the dataset D. * ={x * ,f},x * Represents normalized acceleration and gyroscope sensor data of the behavioral mode;

[0007] Steps 1-3, use the sliding window technique to * Perform data segmentation to obtain data sets Acceleration and gyroscope sensor data representing segmented behavioral modalities;

[0008] Steps 1-4, based on Table 1 from the segmented modality dataset The time domain and frequency domain behavior features such as maximum value, minimum value, signal amplitude and vector angle are extracted to form a behavior mode feature set R = {r, f}, where r represents the time domain and frequency domain features of the behavior mode, and R is divided into a training feature set A = {a, f} and a test feature set C = {c, f} in a ratio of 7:3, where a and c represent the time domain and frequency domain features of the behavior mode in the training feature set and the time domain and frequency domain features of the behavior mode in the test feature set, respectively;

[0009] Step 2: Establish a behavior mode recognition model module, where the behavior mode recognition model module is used to train the training feature set in step 1 using a fast self-building algorithm to obtain a behavior mode recognition model;

[0010] Step 3: Establish a behavior modality recognition module, which is used to input the behavior features to be tested in the test feature set into the behavior modality recognition model to obtain the category label of the behavior feature data to be tested.

[0011] In step 2, the fast self-building algorithm specifically includes:

[0012] Step 2-1, use the training feature set A = {a, f} as the input of the fast self-building algorithm, and set the parameters required for the fast self-building algorithm to establish behavioral mode recognition, including the maximum number of hidden layer nodes L max , current number of nodes L, maximum number of random parameters T max , hidden layer parameter allocation interval limits ±λ and expected error l;

[0013] Step 2-2, randomly generate T from the allocated interval max For the hidden layer node parameters, the hidden layer node parameters include the input weight w L and bias b L , the most appropriate hidden layer node parameters are selected according to the following interpretability space geometric angle maximization constraints:

[0014]

[0015] In the formula, σ is an arbitrarily small positive number, usually 0.0005, θ L Indicates e L-1 and g L The angle between L ) max =(arcos∠(e L -1,gL)) max ∈(0,π / 2),(θ L ) max represents Tmaxθ L The largest θ in L , represents the activation function of the Lth hidden layer node, eL-1 =ff L-1 , e L-1 represents the residual of the behavioral mode recognition model with L-1 nodes, f L-1 represents the predicted class label of the behavioral modality recognition model with L-1 nodes;

[0016] Step 2-3, calculate the model output weight β:

[0017]

[0018] In the formula, represents the intermediate variable,

[0019] T stands for transpose; H L =[H L-1 h L ] refers to the hidden layer output matrix of the behavioral mode recognition model when the number of hidden layer nodes is L, H L-1 =[h 1 ,...,h L-1 ] represents the hidden layer output matrix of the behavioral mode recognition model when the number of hidden layer nodes is L-1, where w ii and b ii is selected by maximizing the geometric angle constraint in the interpretability space in step 2-2, and I represents the L-1 A L-1 The identity matrix of the same type, E is the same as h L υ is the unit matrix of the same type, C is a constant, usually 0.05.

[0020] Step 2-4, calculate the residual e of the current behavioral mode recognition model L :

[0021] e L =fH L β;

[0022] Step 2-5, if the residual e L If the expected tolerance l is met, the modeling of the behavioral modality recognition model is terminated. Otherwise, return to step 2-2 and continue to add new nodes to the hidden layer of the behavioral modality recognition model, that is, L = L + 1; continue the modeling process of the above behavioral modality recognition model until the residual meets the set expected tolerance l or the number of hidden layer nodes reaches L max until.

[0023] In step 3, the behavior feature c to be tested in the test feature set C = {c, f} is input into the behavior mode recognition model to obtain the category label of the behavior feature data to be tested

[0024] Beneficial effects: It can be seen from the above scheme that the present invention has the following advantages and effects compared with the prior art:

[0025] The present invention uses spatial geometric angle maximization constraints to select the most appropriate hidden layer node parameters. This constraint not only improves the quality of node parameters but also improves the compactness of the network structure of the model built by the fast self-building method. Therefore, the present invention has the advantages of fast modeling speed, high recognition accuracy and lightweight system size. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0027] Figure 1 This is the flow chart of the human behavior modality recognition system.

[0028] Figure 2 This is a demonstration diagram of the human behavior modal data collection process.

[0029] Figure 3 Schematic diagram of the network structure of the fast self-built behavioral mode recognition algorithm in the present invention. DETAILED DESCRIPTION

[0030] The schematic diagram of the process of the present invention is as follows Figure 1 As shown, human behavior modal data is obtained by using a mobile device and transmitted to a computer terminal, and then feature engineering technology is used to extract a data set for establishing a behavior modal recognition model.

[0031] The network structure of the fast self-built behavior mode recognition model used in the present invention is as follows: Figure 3 As shown in the figure, it consists of three layers: input layer, hidden layer and output layer.

[0032] The present invention provides a fast self-constructed behavior mode recognition model, which is as follows:

[0033] 1. Behavioral modality data collection and processing module. The present invention uses a mobile device to collect acceleration and gyroscope sensor data of six behavioral modalities such as running, walking, going downstairs, going upstairs, lying down, and sitting at a sampling frequency of 50.0 Hz and a time span of 10 seconds. The data and the behavioral modality category are combined into a behavioral modality data set D = {x, f}, where x represents the sensor data of the behavioral modality and f represents the behavioral modality category label. The acquired behavioral modality data set D is then uploaded to the terminal in the form of a data stream. Figure 2The figure shows a demonstration diagram of the human behavior modality data collection process. The pictures in the figure represent the behavior modes of going upstairs, going downstairs, standing, jogging, lying down and sitting from left to right and from top to bottom respectively. Then, the feature engineering technology is used on this terminal to process the data set D to obtain the final data set A, as follows:

[0034] 1.1 Collect acceleration and gyroscope sensor data of six behavioral modes: running, walking, going downstairs, going upstairs, lying down, and sitting, and obtain the behavioral mode data set D = {x, f}, where x represents the sensor data of the behavioral mode and f represents the behavioral mode category label;

[0035] 1.2 Use Gaussian filtering and maximum and minimum normalization techniques to denoise and normalize the behavioral modality dataset D, and then obtain the dataset D * ={x * ,f},x * Represents normalized behavioral modality data;

[0036] 1.3 Use sliding window technology to analyze the dataset D * Perform data segmentation to obtain data sets Represents the behavioral modality data after segmentation;

[0037] 1.4 Based on Table 1, the segmented modal dataset The time domain and frequency domain behavior features such as maximum value, minimum value, signal amplitude and vector angle are extracted to form a behavior mode feature set R = {r, f}, where r represents the time domain and frequency domain features of the behavior mode, and R is divided into a training feature set A = {a, f} and a test feature set C = {c, f} in a ratio of 7:3, where a and c represent the time domain and frequency domain features of the behavior mode in the training feature set and the time domain and frequency domain features of the behavior mode in the test feature set, respectively;

[0038] Table 1

[0039]

[0040]

[0041] 2. Establish a behavioral mode recognition model module. The details are as follows:

[0042] 2.1 Input the training feature set A = {a, f} into the fast self-building algorithm and set the parameters for the fast self-building algorithm to establish behavioral mode recognition: the maximum number of hidden layer nodes L max =500, L=1, maximum number of random parameters T max =20, the hidden layer parameter allocation interval bound is 1 and the expected error is 0.05.

[0043] 2.2 Randomly generate 20 pairs of hidden layer node parameters (input weights w L and bias b L ), the most appropriate hidden layer node parameters are selected according to the following interpretability space geometric angle maximization constraints:

[0044]

[0045] In the formula, σ is an arbitrarily small positive number, usually 0.0005, θ L Indicates e L-1 and g L The angle between L ) max =(arcos∠(e L-1 ,g L )) max ∈(0,π / 2) represents T max θ L The largest θ in L , represents the activation function of the Lth hidden layer node, e L-1 =ff L-1 represents the residual of the behavioral mode recognition model with L-1 nodes, f L-1 represents the predicted class label of the behavioral modality recognition model with L-1 nodes;

[0046] 2.3 Calculate the model output weights:

[0047]

[0048] In the formula, represents the intermediate variable,

[0049] T stands for transpose; H L =[H L-1 h L ] refers to the hidden layer output matrix of the behavioral mode recognition model when the number of hidden layer nodes is L, H L-1 =[h 1 ,...,h L-1 ] represents the hidden layer output matrix of the behavioral mode recognition model when the number of hidden layer nodes is L-1, where w ii and b ii is selected by maximizing the geometric angle constraint in the interpretability space in step 2.2, and I represents the L-1 A L-1 The same type of identity matrix, E is the same as h L υ is the unit matrix of the same type, C is a constant, usually 0.05.

[0050] 2.4 Calculate the residual e of the current behavioral mode recognition model L

[0051] e L =fH L β

[0052] Step 2-5, if the residual e L If the expected tolerance l is met, the modeling of the behavioral modality recognition model is terminated. Otherwise, return to step 2-2 and continue to add new nodes to the hidden layer of the behavioral modality recognition model, that is, L = L + 1; continue the modeling process of the above behavioral modality recognition model until the residual meets the set expected tolerance l or the number of hidden layer nodes reaches L max until.

[0053] 3. Behavior mode recognition module: Input the behavior feature c to be tested in the test feature set C = {c, f} into the behavior mode recognition model, and then obtain the category label of the behavior feature data to be tested For example, when the behavioral modality recognition model is built, a person wears a mobile device and performs a behavioral modality activity. At this time, the terminal can identify the behavioral modality of the wearer based on the sensor data transmitted by the mobile terminal. Then it can determine which of the six behavioral modalities the wearer is engaged in, such as running, walking, going downstairs, going upstairs, lying down, and sitting.

[0054] As described above, the present invention provides a fast self-built behavior mode recognition system. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A fast self-constructed behavior mode recognition method, characterized in that: The following steps are involved: Step 1: Establish a behavioral modality data acquisition and processing module, which is used to perform the following steps: Step 1-1, obtain the acceleration and gyroscope sensor data of six behavioral modes, namely, running, walking, going downstairs, going upstairs, lying down, and sitting, through a mobile device, and combine the acceleration and gyroscope sensor data of the behavioral mode and the behavioral mode category into a behavioral mode data set D = {x, f}, where x represents the acceleration and gyroscope sensor data of the behavioral mode, and f represents the behavioral mode category label; Step 1-2: Use Gaussian filtering and maximum and minimum normalization techniques to denoise and normalize the behavioral modality dataset D to obtain the dataset D. * ={x * ,f},x * Represents normalized acceleration and gyroscope sensor data of the behavioral mode; Steps 1-3, use the sliding window technique to * Perform data segmentation to obtain data sets Acceleration and gyroscope sensor data representing segmented behavioral modalities; Steps 1-4, from the segmented modality dataset Extract time domain and frequency domain behavior features to form a behavior mode feature set R = {r, f}, where r represents the time domain and frequency domain features of the behavior mode, and divide R into a training feature set A = {a, f} and a test feature set C = {c, f} according to a certain ratio, where a and c represent the time domain and frequency domain features of the behavior mode in the training feature set and the time domain and frequency domain features of the behavior mode in the test feature set, respectively; Step 2: Establish a behavior mode recognition model module, where the behavior mode recognition model module is used to train the training feature set in step 1 using a fast self-building algorithm to obtain a behavior mode recognition model; Step 3: Establish a behavior mode recognition module, which is used to input the behavior features to be tested in the test feature set into the behavior mode recognition model to obtain the category label of the behavior feature data to be tested; In step 2, the fast self-building algorithm specifically includes: Step 2-1, use the training feature set A = {a, f} as the input of the fast self-building algorithm, and set the parameters required for the fast self-building algorithm to establish behavioral mode recognition, including the maximum number of hidden layer nodes L max , current number of nodes L, maximum number of random parameters T max , hidden layer parameter allocation interval limits ±λ and expected error l; Step 2-2, randomly generate T from the allocated interval max For the hidden layer node parameters, the hidden layer node parameters include the input weight w L and bias b L , the most appropriate hidden layer node parameters are selected according to the following interpretability space geometric angle maximization constraints: In the formula, σ is an arbitrarily small positive number, θ L Indicates e L-1 and g L The angle between L ) max =(arcos∠(e L-1 ,g L )) max ∈(0,π / 2),(θ L ) max represents Tmaxθ L The largest θ in L , represents the activation function of the Lth hidden layer node, e L-1 =ff L-1 , e L-1 represents the residual of the behavioral mode recognition model with L-1 nodes, f L-1 represents the predicted class label of the behavioral modality recognition model with L-1 nodes; Step 2-3, calculate the model output weight β: In the formula, represents the intermediate variable, T stands for transpose; H L =[H L-1 h L ] refers to the hidden layer output matrix of the behavioral mode recognition model when the number of hidden layer nodes is L, H L-1 =[h1,...,h L-1 ] represents the hidden layer output matrix of the behavioral mode recognition model when the number of hidden layer nodes is L-1, where w i and b i is selected by maximizing the geometric angle constraint in the interpretability space in step 2-2, and I represents the L-1 A L-1 The same type of identity matrix, E is the same as h L υ is the identity matrix of the same type, Z is a constant; Step 2-4, calculate the residual e of the current behavioral mode recognition model L : e L =fH L b; Step 2-5, if the residual e L If the expected tolerance l is met, the modeling of the behavioral modality recognition model is terminated. Otherwise, return to step 2-2 and continue to add new nodes to the hidden layer of the behavioral modality recognition model, that is, L = L + 1; continue the modeling process of the above behavioral modality recognition model until the residual meets the set expected tolerance l or the number of hidden layer nodes reaches L max until.

Images