Single-node wearable sensing human body behavior recognition method and system combined with state transition constraint

Abstract

The invention discloses a single-node wearable sensing human body behavior recognition method in combination with state transition constraint. The method comprises the following steps: acquiring inertial data by using single-node wearable sensing equipment, and carrying out filtering, denoising and segmentation; dynamic / static scores of the data segments are obtained through a dynamic and static segmentation module to divide dynamic and static behaviors; constructing a human body behavior recognition module based on the domain expansion features and the behavior tags; a human body behavior recognition module is used for dividing dynamic behaviors into various motion behaviors such as a transition state and a periodic state; constructing a human body state behavior transition constraint module, and training the module by using a continuous behavior sequence generated based on prior state transition constraint; and the trained human body state behavior transfer constraint module is used to realize dynamic transfer and adjustment of the behavior state. The invention further provides a single-node wearable sensing human body behavior recognition system combined with the state transition constraint. The method effectively enhances the accuracy of human behavior recognition, has the advantages of low complexity, good robustness and high flexibility, and can be applied to various wearable devices and motion monitoring scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of human behavior recognition, and more specifically to a single-node wearable sensor-based human behavior recognition method and system combined with state transition constraints. Background Art

[0002] Human behavior recognition has important application values in fields such as health monitoring, sports training, and rehabilitation medicine. Traditional technologies usually rely on multi-node sensor data fusion to obtain rich behavior information through multi-dimensional data, thereby improving the recognition accuracy. However, this method makes the system design extremely complex and significantly increases the hardware cost. At the same time, when multiple nodes of sensors work simultaneously, it makes the device extremely inconvenient to wear, greatly reducing the wearing comfort and bringing many troubles to actual use. When dealing with action recognition in wearable scenarios, the recognition methods based on multi-node sensing devices often require complex models and a large amount of computing resources, which pose higher requirements for real-time performance and the computational efficiency of the methods.

[0003] In addition, the existing methods have obvious deficiencies in dealing with the state transition relationships between behaviors. In fact, human behaviors do not exist in isolation during the actual occurrence process, but there are close correlations before and after. Traditional end-to-end behavior recognition methods are difficult to effectively model and analyze the state transitions in continuous behavior sequences. The characteristics of their model architectures determine that there are inherent defects in capturing the dynamic changes and state transition rules between behaviors and cannot fully exploit the temporal information in behavior sequences. Although data-driven deep learning methods have powerful feature extraction and generalization capabilities, they have extremely high requirements for the scale and quality of training data. Usually, it is extremely difficult to obtain a behavior dataset without interference, continuous, and capable of comprehensively covering all state transition conditions. This makes it difficult for such methods to achieve ideal recognition effects due to the lack of sufficient and adaptable data as support when only knowing limited state transition condition constraints.

[0004] Therefore, there is an urgent need for an innovative method and system that can not only maintain high recognition accuracy while significantly improving real-time performance and computational efficiency, but also effectively handle the state transition constraint problems between behaviors to meet the requirements in practical applications. Summary of the Invention

[0005] In order to overcome the deficiencies of existing human behavior recognition methods in the application of single-node wearable sensing devices, such as low accuracy, poor adaptability, and poor overall recognition performance caused by excessive method complexity, the present invention provides a single-node wearable sensing human behavior recognition method and system combined with state transition constraints, which can effectively enhance the accuracy of human behavior recognition when dealing with time-series correlation constraints and transitional features between behaviors, and at the same time has the advantages of small complexity, good robustness, and high flexibility, and can be applied to various wearable devices and motion monitoring scenarios.

[0006] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0007] A single-node wearable sensing human behavior recognition method combined with state transition constraints includes the following steps:

[0008] Step 1), using the inertial sensors of the single-node wearable sensing device to collect triaxial acceleration and triaxial angular velocity data where i is the data entry index in the collected data set, and the number of sample points of each inertial sensing signal is L i = f s ·T i , f s is the sampling frequency, T i is the sampling duration of the i-th data entry, d is the number of data channel dimensions, the collected behavior category set is M = {m1, m2,..., m C}, C is the predefined number of behavior categories, and perform preprocessing operations of filtering, denoising, and data segmentation on the signals of each collected channel to obtain d-channel behavior data of length N

[0009] Step 2), for the d-channel data in each segmentation window, calculate the dynamic / static score of a single window through a dynamic / static segmentation module based on fuzzy approximate entropy and multi-feature fusion, and divide dynamic and static behaviors based on a decision threshold;

[0010] Step 3), for the data segments of dynamic behaviors, further perform domain expansion on the collected acceleration and angular velocity data to obtain single-axis time-domain data, single-axis frequency-domain data, and time-domain and frequency-domain data of the synthetic modulus value, and obtain a domain expansion vector where DFT represents the discrete Fourier transform operation, and SMV is the synthetic modulus value of triaxial angular velocity and triaxial acceleration. Subsequently, calculate statistical features for the domain-expanded data v where k is the number of features;

[0011] Step 4), based on the training samples of data feature F and their corresponding true behavior labels y true∈M, train the human behavior recognition module D, which constructs a hyperplane or a set of hyperplanes in a high-dimensional space. Through training, the distance between these hyperplanes and the nearest feature vectors in any category is maximized, and finally the parameters are found and such that the prediction result of y pred = sign(w T φ(F)+b) is as consistent as possible with the true behavior label, where φ(·) is used to implicitly map the feature vector into a higher-dimensional space;

[0012] Step 5), input the data segment of the dynamic behavior into the human behavior recognition module D trained in Step 4), and subdivide the motion behavior into a transition state and P periodic states;

[0013] Step 6), construct the human state behavior transition constraint module R, and generate a continuous behavior sequence A = [a1, a2, … a , a ∈ M, where H is a single constraint condition for state transition, H is the length of the continuous behavior sequence, the probability of a1 selecting a certain behavior is inversely proportional to the probability under the stationary distribution of the behavior, and train the transition constraint module R to learn to fuse the state information of the history and the current moment;

[0014] Step 7), input the historical behavior sequence information and the current prior behavior state into the human state behavior transition constraint module trained in Step 6) to achieve the dynamic transition and adjustment of the current behavior state.

[0015] Furthermore, in the above-mentioned Step 1), the single-node wearable sensing device includes two built-in sensors, namely an accelerometer and a gyroscope, which are used to collect 6-channel behavior data, including three-axis acceleration and three-axis angular acceleration.

[0016] Still further, in the above-mentioned Step 1), use the single-node wearable sensing device to collect the action data of various human behaviors,

[0017] Define three basic behaviors: static behavior, transition state behavior, and periodic state behavior. Static behavior refers to the behavior in which the subject remains relatively stationary, and its inertial sensing signal is mainly affected by gravity, with a small signal change rate and relatively stable. Periodic state behavior refers to the behavior with repeatability and following a specific periodic pattern, and its signal characteristics are manifested as periodic repeated changes, usually lasting for a long time, and can be divided into various more detailed motion behaviors. Transition state behavior refers to the short-term behavior that occurs between two static / periodic state behaviors, and its signal characteristics are manifested as a large change rate and a short duration;

[0018] ​For each human behavior, a single-node wearable sensing device is worn on the user's body to collect inertial sensing data of specified multiple motion behaviors, and the categories of the inertial sensing data of the collected human behaviors are labeled to obtain a set of sensing data of human behaviors S = {s i , y true}.

[0019] Preferably, in the above step 1), common filtering methods such as FIR low-pass filters are used for filtering and denoising to remove high-frequency noise and ensure the stability and reliability of the data.

[0020] Furthermore, the above step 2) includes the following sub-steps:

[0021] Step (2-1), calculate the combined modulus values of the three-axis angular velocity and the three-axis acceleration Use the fuzzy approximate entropy algorithm to obtain the fuzzy approximate entropy fApEn of the SMV, calculate the standard deviation std and the peak-to-peak value pp of the angular velocity, and clamp them to 0-1;

[0022] Step (2-2), based on step (2-1), perform weighted fusion of the angular velocity standard deviation, the peak-to-peak value, and the fuzzy approximate entropy to obtain the dynamic / static score DS = (std × std_factor + pp × p2p_factor + fApEn) × scale_factor, where std_factor, p2p_factor, and scale_factor are all hyperparameters, and the dynamic / static score characterizes the comprehensive characteristics of dynamic and static behaviors;

[0023] Step (2-3), set the empirical threshold parameter as the decision boundary, and divide the mode in the interval higher than the given decision boundary into dynamic, and vice versa into static.

[0024] Preferably, in the above step (2-1), the calculation process of the fuzzy approximate entropy is as follows:

[0025] Step (2-1-1), parameter initialization: set the parameter values of the mode dimension m and the similarity tolerance r;

[0026] Step (2-1-2), data window extraction: use a time window with a length of m, with a step size of 1, to process a signal s with a length of N to obtain N - m + 1 new signal vectors with a dimension of m, where the i-th new signal vector is expressed as shown in Equation (1):

[0027]

[0028] Step (2-1-3), data window mean calculation: define u m (i) as the mean of:

[0029]

[0030] Step (2-1-4), calculation of distance between data: Calculate the distance between new signals

[0031]

[0032] Step (2-1-5), calculation of similarity between data: Introduce a fuzzy membership function and calculate according to Equation (4) of the similarity

[0033]

[0034] Step (2-1-6), calculation of similarity between a single data and a data set: According to Equation (5), count the average similarity degree of each m-dimensional new signal vector and other new signal vectors

[0035]

[0036] Step (2-1-7), calculation of average similarity degree of the data set: According to Equation (6), count the average value Φ of the logarithms of all average similarity degrees m :

[0037]

[0038] Step (2-1-8), dimension promotion and calculation of fuzzy approximate entropy: Change the pattern dimension to m+1 dimension, and repeat steps (2-1-2) to (2-1-7) to obtain Φ m+1 , and calculate the fuzzy approximate entropy according to Equation (7):

[0039] En(m) = Φ m - Φ m+1 (7)

[0040] When the pattern dimension m = 2 and the similarity tolerance r = 0.1 - 0.25, the dependence degree of the approximate entropy on the sequence length N is the smallest. Therefore, the above parameters are taken as m = 2 and r = 0.15. The value range of the fuzzy approximate entropy is 0 - 1, and the larger its value indicates the higher the complexity of the time series and the greater the probability of generating new patterns.

[0041] Furthermore, in step 6), the human state behavior transfer constraint module R(W,B) is a state transfer constraint mechanism based on iterative optimization. This module takes the historical behavior sequence information and the current prior behavior state as inputs, and uses the historical behavior state formed by combining the current module with the first H-1 upstream behavior recognition outputs at the current moment and the behavior state recognized at the current moment to perform constraint optimization on the behavior state at the current moment, thereby achieving a more accurate transfer of the human behavior state;

[0042] Let the continuous behavior sequence A = [a1, a2, … a H , a ∈ M, where the current behavior state is a H , and the historical behavior sequence is [a1, a2, … a H-1 , and the constraint condition is Then the model is constructed as:

[0043]

[0044] where W a , W h , W o , B1 and B2 are the trainable parameters of the model, ψ and σ are two non-linear functions, is the posterior behavior state, and the optimization objective is:

[0045]

[0046] Through multiple iterations of optimization, the model gradually adjusts the parameters to minimize the objective function J, thereby learning the state information that fuses history and the current moment and achieving dynamic transfer and adjustment of the current behavior state.

[0047] A single-node wearable sensor-based human behavior recognition system combined with state transfer constraints includes a single-node wearable sensor signal acquisition module, one or more processors, a memory, and one or more applications. The one or more applications are stored in the memory and configured to be executed by the one or more processors. The one or more programs are configured to execute the single-node wearable sensor-based human behavior recognition method combined with state transfer constraints as described above.

[0048] The beneficial effects of the present invention are mainly manifested in: collecting human behavior data by means of a single-node wearable sensing device, significantly reducing the complexity of system design and hardware costs, while improving the comfort and flexibility of device wearing, and being applicable to various wearable devices and motion monitoring scenarios. In the data processing stage, through the dynamic and static segmentation module based on fuzzy approximate entropy and multi-feature fusion, dynamic and static behaviors can be accurately divided. For dynamic behaviors, through domain expansion and feature calculation, the data feature dimension is enriched. The constructed human behavior recognition module can effectively recognize various motion behaviors including transitional states and multiple periodic states, improving the recognition granularity. Based on the prior state transition constraint to generate a continuous behavior sequence, the human state behavior transition constraint module fully learns and fuses the state information of the historical and current moments, realizes the dynamic transition and adjustment of the current behavior state, effectively solves the state transition problem between behaviors, and greatly improves the accuracy of human behavior recognition when dealing with the case of time-series correlation constraints and transitional characteristics between behaviors, and has the advantages of small complexity and good robustness, and has extremely high practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flowchart of a single-node wearable sensing human behavior recognition method combining state transition constraints according to the present invention;

[0050] Figure 2 is a flowchart of dynamic and static segmentation based on fuzzy approximate entropy and multi-feature fusion according to the present invention;

[0051] Figure 3 is a schematic diagram of a human behavior recognition state transition method according to the present invention;

[0052] Figure 4 is an effect diagram of dynamic and static segmentation according to the present invention;

[0053] Figure 5 is a confusion matrix of the test results of the human behavior recognition method according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] The present invention will be further described below with reference to the accompanying drawings.

[0055] Referring to Figures 1 to 5 , a single-node wearable sensing human behavior recognition method combining state transition constraints includes the following steps:

[0056] Step 1), using the inertial sensors of the single-node wearable sensing device to collect triaxial acceleration and triaxial angular velocity data where i is the data item index in the collected data set, and the number of sample points of each inertial sensing signal is L i = f s ·T i , fs is the sampling frequency, and T i is the sampling duration of the i-th data entry. The number of data channel dimensions d = 6, including three acceleration channels and three angular velocity channels; the set of collected behavior categories is M = {m1, m2,..., m C}, C is the predefined number of behavior categories. Preprocessing operations such as filtering and denoising and data segmentation are performed on the signals of each collected channel to obtain several data with a length of N The segmentation window length is N, and the step size is

[0057] In step 1), the single-node wearable sensing device includes two built-in sensors, namely an accelerometer and a gyroscope, which are used to collect 6-channel motion data, including three-axis acceleration and three-axis angular acceleration.

[0058] Referring to Figure 1 , in step 1), the action data of various human behaviors are collected by using a single-node wearable sensing device, specifically including:

[0059] Define three basic motion behaviors: static behavior, transition behavior, and periodic behavior. Among them, the periodic behavior includes 9 more detailed motion behaviors: walking, jumping jacks, long-distance running, chin-ups, push-ups, sit-ups, cycling, sprinting, and crawling. So there are a total of 11 motion behaviors as the collection targets;

[0060] For each human behavior, wear the single-node wearable sensing device on the user's body, collect the inertial sensing data of the above 11 motion behaviors, and label the category of the inertial sensing data of the collected human behavior to obtain the sensing data set S = {s i , y true}.

[0061] In step 1), common filtering methods such as FIR low-pass filters are used for filtering and denoising to remove high-frequency noise and ensure the stability and reliability of the data.

[0062] Step 2), for the d-channel data in each segmentation window, calculate the dynamic / static score of a single window through a dynamic / static segmentation module based on fuzzy approximate entropy and multi-feature fusion, and divide dynamic and static behaviors based on a decision threshold;

[0063] Referring to Figure 2 , step 2) includes the following steps:

[0064] Step (2-1), calculate the combined modulus of the three-axis angular velocity and the three-axis acceleration The fuzzy approximate entropy fApEn of the SMV is obtained by using the fuzzy approximate entropy algorithm, and the standard deviation std and the peak-to-peak value pp of the angular velocity are calculated and clamped to 0-1;

[0065] In step (2-2), based on step (2-1), the standard deviation of the angular velocity, the peak-to-peak value, and the fuzzy approximate entropy are weighted and fused to obtain the dynamic / static score DS = (std × std_factor + pp × p2p_factor + fApEn) × scale_factor, where std_factor = 0.015, p2p_factor = 0.04, and scale_factor = 0.3. The dynamic / static score characterizes the comprehensive characteristics of dynamic and static behaviors;

[0066] In step (2-3), an empirical threshold parameter is set as the decision boundary, and the modes in the interval higher than the given decision boundary are classified as dynamic, and vice versa as static;

[0067] In the said step (2-1), the calculation process of the fuzzy approximate entropy is as follows:

[0068] In step (2-1-1), parameter initialization: set the mode dimension m = 2 and the similarity tolerance r = 0.15;

[0069] In step (2-1-2), data window extraction: use a time window with a length of m, with a step size of 1, to process a signal s with a length of N to obtain N - m + 1 new signal vectors of dimension m, where the i-th new signal vector is expressed as shown in Equation (1):

[0070]

[0071] In step (2-1-3), data window mean calculation: define u m (i) as the mean of:

[0072]

[0073] In step (2-1-4), data distance calculation: calculate the distance between new signals

[0074]

[0075] In step (2-1-5), data similarity calculation: introduce a fuzzy membership function, and calculate the similarity of according to Equation (4)

[0076]

[0077] Step (2-1-6), single data - dataset similarity calculation: According to Equation (5), calculate the average similarity degree of each m-dimensional new signal vector with other new signal vectors

[0078]

[0079] Step (2-1-7), dataset average similarity calculation: According to Equation (6), calculate the average value Φ of the logarithms of all average similarity degrees m :

[0080]

[0081] Step (2-1-8), dimension promotion and fuzzy approximate entropy calculation: Change the pattern dimension to m + 1 dimension, and repeat Steps (2-1-2) to (2-1-7) to obtain Φ m+1 , and calculate the fuzzy approximate entropy according to Equation (7):

[0082] En(m) = Φ m -Φ m+1 (7)

[0083] Step 3), for the data segments of dynamic behaviors, further perform domain expansion on the collected acceleration and angular velocity data to obtain uniaxial time-domain data, uniaxial frequency-domain data, and time-domain and frequency-domain data of the synthetic modulus value, and obtain the domain expansion vector where DFT represents the discrete Fourier transform operation, and SMV is the synthetic modulus value of the three-axis angular velocity and three-axis acceleration calculated in Step 2). Subsequently, calculate the statistical features of the data v after domain expansion where k is the number of features;

[0084] Step 4), based on the training samples with data features F and their corresponding true behavior labels y true ∈M, train the human behavior recognition module D. This module will construct a hyperplane or a set of hyperplanes in a high-dimensional space, and through training, make the distance between these hyperplanes and the nearest feature vectors in any category the largest, and finally find the parameters and such that the prediction result of y pred = sign(w T φ(F)+b) is as consistent as possible with the true behavior label, where φ(·) is used to implicitly map the feature vector to a higher-dimensional space;

[0085] Step 5), input the data segments of dynamic behaviors into the human behavior recognition module D trained in Step 4), and subdivide the motion behaviors into transition states and P periodic states;

[0086] Step 6), construct the human body state behavior transfer constraint module R, and generate a continuous behavior sequence A = [a1, a2, … a H , a ∈ M, where is a single constraint condition for state transfer, and H is the length of the continuous behavior sequence. The probability of selecting a certain behavior for a1 is inversely proportional to the probability under the stationary distribution of this behavior. Train the module R to learn to fuse the state information of history and the current moment;

[0087] Refer to Figure 3 , the human body state behavior transfer constraint module R(W, B) in Step 6) is a state transfer constraint mechanism based on iterative optimization. This module takes the historical behavior sequence information and the current prior behavior state as inputs, and uses the historical behavior state formed by combining the current moment's first H - 1 upstream behavior recognition outputs and the behavior state of the current moment's recognition output by this module to perform constraint optimization on the behavior state of the current moment, so as to achieve a more accurate human body behavior state transfer;

[0088] Specifically, let the continuous behavior sequence A = [a1, a2, … a H , a ∈ M, where the current prior behavior state is a H , and the historical behavior sequence is [a1, a2, … a H-1 , and the constraint condition is Then the model is constructed as:

[0089]

[0090] where W a , W h , W o , B1 and B2 are the trainable parameters of the model, ψ = tanh and σ = sigmoid are two non - linear functions, is the posterior behavior state. The optimization objective is:

[0091]

[0092] Through multiple iterations of optimization, the model gradually adjusts the parameters to minimize the objective function J, thereby learning to fuse the state information of history and the current moment and realizing the dynamic transfer and adjustment of the current behavior state.

[0093] Step 7), input the historical behavior sequence information and the current prior behavior state into the human body state behavior transfer constraint module trained in Step 6) to achieve the dynamic transfer and adjustment of the current behavior state.

[0094] ​A system implemented by a single-node wearable sensing human behavior recognition method combining state transition constraints includes a single-node wearable sensing signal acquisition module, one or more processors, a memory, and one or more applications, where the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to execute the single-node wearable sensing human behavior recognition method combining state transition constraints as described above.

[0095] Refer to Figure 4 , the dynamic and static segmentation effect in a typical sit-to-stand conversion process shows the segmentation effect of the dynamic and static segmentation module based on fuzzy approximate entropy and multi-feature fusion. The synthetic IMU modulus refers to the fused SMV of acceleration and angular velocity, and the dynamic and static score is the score obtained after multi-feature fusion for evaluating whether it is dynamic or static. The dashed interval in the "synthetic IMU modulus" curve is the true dynamic interval, and the dashed part in the segmentation result is the estimated dynamic interval when the decision boundary is set to 0.4.

[0096] In this example, the above system is used to collect 11 types of human behavior data from 23 subjects. The sampling rate is 50Hz, the data window length is N = 50Hz × 2s = 100, and the step is 50. A human behavior recognition dataset with 43,274 samples is constructed. The situation of the constructed dataset is shown in Table 1:

[0097] Exercise type Number of data entries Walking 5224 Jumping jacks 1314 Long-distance running 4947 Chin-up 765 Push-up 1283 Sit-up 1367 Cycling 5089 Sprint 383 Crawling 1145 Static 1546 Transition state 20211

[0098] Table 1

[0099] In this example, the above method is used to test the dataset. The detailed information of the features selected in step 3) is presented in Table 2, and the definitions of the recognition results and evaluation metrics are shown in Tables 3 and 4 respectively.

[0100]

[0101] Table 2

[0102] Recognition result Definition True positive (TP) The number of samples correctly predicted as belonging to the specified behavior category True negative (TN) The number of samples correctly predicted as not belonging to the specified behavior category False positive (FP) The number of samples incorrectly predicted as belonging to the specified behavior category False negative (FN) The number of samples incorrectly predicted as not belonging to the specified behavior category

[0103] Table 3

[0104]

[0105]

[0106] Table 4

[0107] Refer to Table 5 and Figure 5, the test results using the above method in this example show that the human behavior recognition method included in the present invention exhibits high accuracy in both dynamic and static discrimination and periodic motion recognition. For the dynamic and static discrimination problem, the accuracy reaches 95.8%. For the recognition of periodic motion behaviors (such as walking, jumping jacks, long-distance running, pull-ups, push-ups, sit-ups, cycling, sprinting, crawling, etc.), the average accuracy reaches 91.7%. The average accuracy of all motion modalities reaches 94.0%, indicating that the method proposed in the present invention realizes effective human behavior recognition of single-node sensing data.

[0108]

[0109]

[0110] Table 5

[0111] The content described in the embodiments of this specification is only a list of implementation forms of the inventive concept and is only for illustrative purposes. The protection scope of the present invention should not be regarded as limited to the specific forms stated in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be conceived by those of ordinary skill in the art based on the inventive concept of the present invention.

Claims

1. A single-node wearable sensing human behavior recognition method combined with state transition constraints, characterized in that, The method includes the following steps: Step 1), collect triaxial acceleration and triaxial angular velocity data using the inertial sensors of the single-node wearable sensing device where i is the index of the data entry in the collected data set, and the number of sample points of each inertial sensing signal is L i = f s · T i f s is the sampling frequency, T i is the sampling duration of the i-th data entry, d is the number of data channel dimensions, the set of collected behavior categories is M = {m1, m2,..., m C}, C is the predefined number of behavior categories, perform filtering, denoising and data segmentation preprocessing operations on the signals of each collected channel to obtain d-channel behavior data of length N Step 2), for the d-channel data within each segmentation window, calculate the dynamic / static score of a single window through a dynamic / static segmentation module based on fuzzy approximate entropy and multi-feature fusion, and divide dynamic and static behaviors based on a decision threshold; Step 3), for the data segments of dynamic behaviors, further perform domain expansion on the collected acceleration and angular velocity data, so as to obtain uniaxial time-domain data, uniaxial frequency-domain data, and time-domain data and frequency-domain data of the synthesized modulus value, and obtain the domain expansion vector where DFT represents the discrete Fourier transform operation, and SMV is the synthesized modulus value of the three-axis angular velocity and the three-axis acceleration. Subsequently, calculate the statistical features of the data v after domain expansion where k is the number of features; Step 4), based on the training samples of data feature F and their corresponding true behavior labels y true ∈M, train the human behavior recognition module D, which constructs a hyperplane or a set of hyperplanes in a high-dimensional space. Through training, the distance between these hyperplanes and the nearest feature vectors in any category is maximized, and finally the parameters and are found such that the prediction result of y pred = sign(w T φ(F)+b) is as consistent as possible with the true behavior label, where φ(·) is used to implicitly map the feature vector into a higher-dimensional space; Step 5), input the data segments of dynamic behaviors into the human behavior recognition module D trained in Step 4), and subdivide the motion behaviors into a transition state and P periodic states; Step 6), construct a human state behavior transition constraint module R, through the prior-based state transition constraint Generate a continuous behavior sequence A = [a1, a2, … a H , a ∈ M, where is a single constraint condition for state transition, H is the length of the continuous behavior sequence, the probability of selecting a certain behavior for a1 is inversely proportional to the probability under the stationary distribution of this behavior, and train the transition constraint module R to learn to fuse the state information of the historical and current moments; Step 7), input the historical behavior sequence information and the current prior behavior state into the human state behavior transition constraint module trained in Step 6) to achieve dynamic transfer and adjustment of the current behavior state.

2. The single-node wearable sensing human behavior recognition method combining state transition constraints according to claim 1, wherein In the said Step 1), the single-node wearable sensing device includes two built-in sensors, namely an accelerometer and a gyroscope, which are used to collect 6-channel behavior data, including three-axis acceleration and three-axis angular acceleration.

3. The single-node wearable sensing human behavior recognition method combining state transition constraints as claimed in claim 1 or 2, wherein In the said Step 1), use the single-node wearable sensing device to collect action data of various human behaviors. Define three basic behaviors: static behavior, transition state behavior, and periodic state behavior. Static behavior refers to the behavior where the subject remains in a relatively stationary state, and its inertial sensing signal is mainly affected by gravity, with a small signal change rate and relatively stable. Periodic state behavior refers to the behavior with repeatability and following a specific periodic pattern, and its signal characteristics are manifested as periodic repeated changes, usually with a long duration, and can be divided into various more detailed motion behaviors. Transition state behavior refers to the short-term behavior that occurs between two static / periodic state behaviors, and its signal characteristics are manifested as a large change rate and a short duration; For each human behavior, a single-node wearable sensing device is worn on the user's body to collect inertial sensing data of specified multiple motion behaviors, and the categories of the inertial sensing data of the collected human behaviors are labeled to obtain a sensing data set S of human behaviors = {s i , y true}.

4. The single-node wearable sensing human behavior recognition method combining state transition constraints according to claim 1 or 2, characterized in that, In the said Step 1), FIR low-pass filter is used for filtering and denoising to remove high-frequency noise.

5. The single-node wearable sensor-based human behavior recognition method incorporating state transition constraints as claimed in claim 1 or 2, characterized in that The said Step 2) includes the following sub-steps: Step (2-1), calculate the combined modulus of the three-axis angular velocity and the three-axis acceleration Use the fuzzy approximate entropy algorithm to obtain the fuzzy approximate entropy fApEn of the SMV, calculate the standard deviation std and the peak-to-peak value pp of the angular velocity, and clamp them between 0 and 1; Step (2-2), based on Step (2-1), perform weighted fusion of the standard deviation of angular velocity, peak-to-peak value, and fuzzy approximate entropy to obtain the dynamic / static score DS = (std × std_factor + pp × p2p_factor + fApEn) × scale_factor, where std_factor, p2p_factor, and scale_factor are all hyperparameters, and the dynamic / static score characterizes the comprehensive characteristics of dynamic and static behaviors; Step (2-3), set the empirical threshold parameter as the decision boundary, and divide the mode in the interval higher than the given decision boundary into dynamic, and vice versa into static.

6. The single-node wearable sensing human behavior recognition method combined with state transition constraints according to claim 5, characterized in that In the said Step (2-1), the calculation process of fuzzy approximate entropy is as follows: Step (2-1-1), parameter initialization: set the parameter values of the mode dimension m and the similarity tolerance r; Step (2-1-2), data window extraction: Using a time window with a window length of m and a step size of 1, process the signal s with a length of N to obtain N-m+1 new signal vectors with m dimensions. The expression of the i-th new signal vector is shown in Equation (1): Step (2-1-3), calculation of the data window mean: Define u according to Equation (2) m (i) is the mean value of: Step (2-1-4), calculation of data distance: Calculate the distance between new signals Step (2-1-5), calculation of data similarity: Introduce a fuzzy membership function and calculate the similarity of according to Equation (4). Step (2-1-6), single data-dataset similarity calculation: According to Equation (5), calculate the average similarity degree of each m-dimensional new signal vector with other new signal vectors Step (2-1-7), calculation of the average similarity of the data set: Calculate the average Φ of the logarithms of all similarity averages according to Equation (6). m : Step (2-1-8), dimension promotion and fuzzy approximate entropy calculation: Change the pattern dimension to m+1 dimension, repeat steps (2-1-2) to (2-1-7), and obtain Φ m+1 , and calculate the fuzzy approximate entropy according to Equation (7): En(m) = Φ m -Φ m+1 (7) When the mode dimension m = 2 and the similarity tolerance r = 0.1 - 0.25, the dependence of approximate entropy on the sequence length N is the smallest. Therefore, the above parameters are taken as m = 2 and r = 0.

15. The value range of fuzzy approximate entropy is 0 - 1, and the larger its value, the higher the complexity of the time series and the greater the probability of generating new patterns.

7. The single-node wearable sensing human behavior recognition method combining state transition constraints according to claim 1 or 2, characterized in that In step 6), the human body state behavior transfer constraint module R(W,B) is a state transfer constraint mechanism based on iterative optimization. This module takes the historical behavior sequence information and the current prior behavior state as inputs, and uses the historical behavior state formed by combining the module with the outputs of the first H-1 upstream behavior recognitions at the current moment and the behavior state recognized and output at the current moment to perform constraint optimization on the behavior state at the current moment, so as to achieve a more accurate transfer of the human body behavior state; Let the continuous behavior sequence A = [a1, a2, … a H , where a ∈ M, and the current behavior state is a H , and the historical behavior sequence is [a1, a2, … a H-1 , and the constraint condition is Then the model is constructed as follows: Among which W a 、W h 、W o 、B1 and B2 are trainable parameters of the model, ψ and σ are two non-linear functions, is the posterior behavior state, and the optimization objective is: Through multiple iterations of optimization, the model gradually adjusts the parameters to minimize the objective function J, thereby learning to fuse the state information of the historical and current moments and realizing the dynamic transfer and adjustment of the current behavior state.

8. A system for implementing the single-node wearable sensing human behavior recognition method with combined state transition constraints as described in claim 1, characterized in that, The system includes a single-node wearable sensing signal acquisition module, one or more processors, a memory, and one or more applications, where the one or more applications are stored in the memory and are configured to be executed by the one or more processors, and the one or more programs are configured to execute the single-node wearable sensing human behavior recognition method combined with state transfer constraint as described above.

Images