阅读说明：本技术 基于足底压力传感器监测体态变化的装置和方法 (Device and method for monitoring posture change based on plantar pressure sensor ) 是由 王珊珊 梁同乐 孙利玲 刘文杰 翚京华 于 2019-12-16 设计创作，主要内容包括：本发明公开了一种基于足底压力传感器监测体态变化的装置和方法,涉及计算机技术领域,针对用户体态调整提醒存在滞后性的问题,提供了以下技术方案,包括数据采集模块、数据处理模块、模型构建模块、模型训练模块、模型更新模块和更新策略优化模块,通过实时采集用户在不同行为下的体态数据,利用主成分分析、聚类分析、机器学习方法等一系列研究手段,经过抽象建模,构建健康人群体态模型；并通过DNN模型迁移算法,对不同用户实现模型迁移；同时针对用户体态发展变化等特点,面向AIoT,基于模型一致性检验理论,考虑兼顾用户体验与模型整体表现的模型更新策略,制定出符合用户要求及体征的美化方案,达到为用户提供当前状态下体态信息的目的。(The invention discloses a device and a method for monitoring body state change based on a plantar pressure sensor, relates to the technical field of computers, and provides the following technical scheme aiming at the problem that hysteresis exists in user body state adjustment reminding; model migration is realized for different users through a DNN model migration algorithm; and aiming at the characteristics of the user body state development change and the like, the method is oriented to AIoT, based on a model consistency inspection theory, a model updating strategy which gives consideration to user experience and model integral expression is considered, and a beautifying scheme which meets the user requirements and physical signs is worked out, so that the purpose of providing body state information under the current state for the user is achieved.)

1. Device based on plantar pressure sensor monitoring attitude changes, its characterized in that includes following module:

the data acquisition module acquires an original data sample of the user posture in real time through the sole pressure sensor;

the data processing module is used for carrying out filtering and denoising processing on the original data sample to obtain user posture data;

the model building module is used for extracting characteristics of the user posture data and building a user posture model;

the model training module is used for training the constructed user posture model by utilizing the user posture data;

the model updating module is used for continuously updating the user posture model through model migration, constructing a user personalized posture model and storing the user personalized posture model to a cloud end or an edge end;

and the updating strategy optimizing module is used for optimizing the updating decision of the user posture model in the model updating module through the consistency check of the user posture data.

2. The device for monitoring the posture change based on the plantar pressure sensors according to claim 1, wherein the plantar pressure sensors are integrated dot matrix pressure distribution sensors, each plantar pressure sensor comprises a plurality of sensing points, the diameters of the sensing points are 5mm, and the distances between every two adjacent sensing points in the horizontal direction or the vertical direction are 7 mm.

3. The plantar pressure sensor based device for monitoring the body state change according to claim 1, wherein the data processing module further comprises a human body gait cycle analysis module, and the human body gait cycle analysis module is used for performing single step segmentation on the raw data samples.

4. The device for monitoring posture changes based on plantar pressure sensors according to claim 1, wherein an online consistency check model is arranged in the update strategy optimization module, and the consistency check model is used for inputting a decision whether to optimize a current user posture model with the user posture data to the model training module after comparing and evaluating the user posture data and the current user posture model.

5. The plantar pressure sensor-based posture change monitoring device according to claim 1, wherein the model construction module comprises a user posture model constructed through a Deep Neural Network (DNN) algorithm.

6. A method for monitoring posture change based on a plantar pressure sensor is characterized by comprising the following steps:

step (1): obtaining an original data sample of the user posture information by taking a sole pressure sensor as a data acquisition module;

step (2): filtering and denoising an original data sample in a data processing module, and carrying out gait analysis processing to obtain user posture data;

and (3): establishing a user posture model based on the user posture data;

and (4): optimizing the user posture model through a migration algorithm, thereby completing the construction of the user personalized posture model;

and (5): synchronously transmitting the user posture data transmitted in real time to an updating strategy optimization module for consistency check, and making a decision for updating the user personalized posture model according to a check result;

and (6): and updating the user personalized posture model through transfer learning according to the instruction of the updating strategy optimization module, and performing comprehensive judgment and analysis.

7. The method for monitoring the body state change based on the plantar pressure sensors according to claim 6, wherein the body state data of the user obtained after the processing of the data processing module in the step (2) comprise plantar pressure values of the user under different behaviors, change curves of the pressure values, stress time and a change track of a three-dimensional pressure distribution center.

8. The method for monitoring posture changes based on plantar pressure sensors according to claim 6, wherein in the step (3), the user posture model is constructed on the user posture data through a deep neural network algorithm framework.

9. The method for monitoring the body state change based on the plantar pressure sensor as recited in claim 6, characterized in that a learning framework for jointly optimizing the shared weight between the source and target task models is adopted in step (4), and a reinforcement learning module based on a performance metric matrix of a target data set is used to realize the calculation of the shared weight, so as to ensure the self-adaptive output of the weight of the features in each source data set.

Technical Field

The invention relates to the technical field of computers, in particular to a device and a method for monitoring body state change based on a plantar pressure sensor.

Background

The stooping-over is an appearance problem which troubles many people, and besides, the body shape and beauty of people are influenced by various bad postures such as walking with the shape of a Chinese character 'ba', standing with oblique shoulders, and leg-cutting with Erlang. Although the user knows the importance of the beauty of the body, it is difficult to keep a tall and straight body posture continuously because of "forgetting". The common instrument of correcting the posture is clothing of getting into shape, back fixer, the shoes of correcting for having the correction function at present, and it mainly has following drawback:

(1) the comfort level of a user wearing or wearing the shaping clothes or the fixer is not good enough, and the user has obvious oppression feeling on the body;

(2) the body shaping clothes worn in summer have obvious stuffy feeling;

(3) the subjective cooperation of users is lacked, and the function of beautifying the posture is not obvious;

(4) the correcting shoes are designed uniformly and cannot be different from person to person, and part of correcting shoe manufacturers propose that users send the shoes back to a service point to analyze the wear of soles regularly, so as to remind the users how to pay attention to posture adjustment, and the 'after-the-fact' analysis method is difficult to obtain remarkable effect.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a device and a method for monitoring body state change based on a plantar pressure sensor, which have the advantage of providing body state information under the current state for a user.

In order to achieve the purpose, the invention provides the following technical scheme:

device based on plantar pressure sensor monitoring posture changes includes following module:

the data acquisition module acquires an original data sample of the user posture in real time through the sole pressure sensor;

the data processing module is used for carrying out filtering and denoising processing on the original data sample to obtain user posture data;

the model building module is used for extracting characteristics of the user posture data and building a user posture model;

the model training module is used for training the constructed user posture model by utilizing the user posture data;

the model updating module is used for continuously updating the user posture model through model migration, constructing a user personalized posture model and storing the user personalized posture model to a cloud end or an edge end;

and the updating strategy optimizing module is used for optimizing the updating decision of the user posture model in the model updating module through the consistency check of the user posture data.

By adopting the technical scheme, the data acquisition module acquires data information related to the user posture in real time through the plantar pressure sensor as an original data sample, the original data sample is transmitted to the data processing module, the sample is subjected to filtering, denoising and other processing in the data processing module to obtain user posture data, the user posture data is transmitted to the model construction module as a data set, the model construction module constructs a user posture model by using the user posture data, the model construction module outputs the constructed user posture model to the model training module, the user posture model is continuously trained by using the user posture data from the data processing module in the model training module, the user posture model is updated by using transfer learning to obtain a user personalized posture model, and an update strategy module in the device optimizes an update strategy of the user posture model, the accuracy of the model is corrected in real time according to the user posture data through model consistency check, when the user posture information data changes and continuously deviates from the original model, instruction information is sent to the model training module, posture data information is collected again, and the model is updated rapidly through transfer learning. The consistency check has the advantages that the reliability of the posture monitoring result is evaluated and guaranteed, the fault-tolerant interval of the user personalized posture model can be set by a posture trainer and a user at the same time, and the effectiveness of posture correction and the comfort level of user training are considered.

The size and the distribution of the plantar pressure can effectively reflect the change of the body state of a human body, so that certain physiological or pathological information of the human body can be obtained, namely data reflecting the body state change of a user can be obtained through a plantar pressure sensor, the body state model is stored at the cloud end and the edge end according to the information of the body state of the current user fed back from the data extracted through processing of all modules in the device, and the AIoT (AIoT) technology is combined, so that the body state model is collected in real time, support is provided for comprehensively analyzing the body state change process, the user can rapidly obtain the body state information of the current person through the device, and adjustment is made according to the body state information.

Preferably, plantar pressure sensor is integrated dot matrix pressure distribution sensor, plantar pressure sensor includes a plurality of induction points, and is a plurality of the diameter of induction point is 5mm, and the distance between the adjacent induction point is 7mm in horizontal direction or vertical direction.

By adopting the technical scheme, the quantity of the original data samples when the human body stands, sits and walks is increased by arranging the dense dot-matrix sensing points, so that the acquired data can accurately reflect the body state information of the current user, and the accuracy of subsequent modeling is improved.

Preferably, the data processing module further comprises a human body gait cycle analysis module, and the human body gait cycle analysis module is used for performing single-step segmentation on the original data sample.

By adopting the technical scheme, from the analysis of human gait cycles, one gait cycle is called a stride, which is the time from the landing of one heel to the landing of the heel of the same leg. A stride comprises two phases: a two-legged support phase and a single-legged support phase. For a leg, a complete gait cycle includes a stance phase and a swing phase. When the human body walks on the flat ground, the following relationship is approximately found in each period:

D_sz＝0.248D_ci+0.143

D_z＝0.752D_ci-0.143

D_dz＝0.752D_ci-0.143

wherein: d_ciOne gait cycle; d_szIn the single-foot support period; d_dzThe two feet supporting period; d_zIs the support period of a certain leg in a walking cycle. The single-step segmentation is carried out on the collected original data sample by using the gait cycle, the pressure characteristics of the left/right foot supporting phases are collected, and the characteristic dimension reduction is carried out by using a principal component analysis method.

Preferably, an online consistency check model is arranged in the update strategy optimization module, and the consistency check model is used for comparing and evaluating the user posture data and the current user posture model, and then inputting a decision whether to optimize the current user posture model with the user posture data to the model training module.

By adopting the technical scheme, when the user posture information data changes and continuously deviates from the original model, the original data sample of the user posture needs to be collected again, and the current user posture model is rapidly updated by utilizing transfer learning. The online consistency check model refers to: and after the evaluation of the single posture is finished, the posture information and the posture label thereof are used as a new sample to be added into the original sample set to generate a new sample set, and then the evaluation is carried out on the next object based on the new sample set. In the course of prediction, the sample set is constantly being updated. When the user posture is continuously changed and the accumulated user posture data enables the error proportion of the current user posture model to be larger than the preset significance level value, the fact that the current user posture model needs to be optimized and updated is shown.

Preferably, the model construction module comprises a user posture model constructed by a deep neural network DNN algorithm.

By adopting the technical scheme, a model frame based on a deep neural network DNN algorithm is preset in the model construction module, and the user posture data in the data processing module is substituted into the frame, so that a user posture model is obtained.

A method for monitoring posture change based on a plantar pressure sensor comprises the following steps:

step (1): obtaining an original data sample of the user posture information by taking a sole pressure sensor as a data acquisition module;

step (2): filtering and denoising an original data sample in a data processing module, and carrying out gait analysis processing to obtain user posture data;

and (3): establishing a user posture model based on the user posture data;

and (4): optimizing the user posture model through a migration algorithm, thereby completing the construction of the user personalized posture model;

and (5): synchronously transmitting the user posture data transmitted in real time to an updating strategy optimization module for consistency check, and making a decision for updating the user personalized posture model according to a check result;

and (6): and updating the user personalized posture model through transfer learning according to the instruction of the updating strategy optimization module, and performing comprehensive judgment and analysis.

Preferably, the user posture data obtained after the processing of the data processing module in the step (2) includes plantar pressure values, change curves of the pressure values, stress time and change tracks of a three-dimensional pressure distribution center of the user under different behaviors.

Preferably, in the step (3), the user posture model is constructed for the user posture data through a deep neural network algorithm framework.

Preferably, in the step (4), a learning framework for jointly optimizing the shared weight between the source and target task models is adopted, and a reinforcement learning module based on a performance metric matrix of the target data set is used for realizing calculation of the shared weight, so that the weight of the feature in each source data set is ensured to be adaptively output.

In summary, the invention has the following advantages:

(1) the plantar pressure sensor is used as a data acquisition module for monitoring the body state, and the plantar pressure sensor is wearable intelligent sensor equipment, so that the limitation of field equipment on body state monitoring is reduced, and body state beautification is integrated into daily behavior activities of a user;

(2) by connecting the device with the AIoT platform, the current posture of the user can be acquired and recorded as daily posture data of the user, and complete posture data change reference is provided for the user;

(3) in the face of the problems of high sample size and high time cost of posture model training, a transfer learning algorithm (such as a shared weight value joint optimization learning framework L2TL) is firstly utilized. The domain similarity is adaptively inferred based on the user posture data set, so that the high requirement of a deep neural network on the data sample size is reduced, and the time for data sample acquisition and user posture model training is shortened. And guarantee is provided for the accuracy and the real-time performance of the posture model.

Drawings

Fig. 1 is a schematic structural diagram of the device for monitoring the posture change based on the plantar pressure sensor.

Fig. 2 is a diagram of human plantar pressure.

Fig. 3 is a human gait cycle chart.

Fig. 4 is a schematic diagram of the L2TL test processing in the method for monitoring posture change based on plantar pressure sensors.

Detailed Description

The following is further detailed by way of specific embodiments:

a device for monitoring posture change based on a plantar pressure sensor is shown in figure 1 and comprises a data acquisition module, a data processing module, a model building module, a model training module, a model updating module and an updating strategy optimizing module. The data acquisition module acquires an original data sample of the user body state in real time through the sole pressure sensor; the output end of the data acquisition module is connected with the input end of the data processing module, the data processing module carries out filtering and denoising processing on an original data sample to obtain user posture data, and the user posture data obtained in the process comprises plantar pressure values of a user under different behaviors, change curves of the pressure values, stress time and a three-dimensional pressure distribution center change track; the model construction module is used for extracting the characteristics of the user posture data from the data processing module and constructing a user posture model; training the constructed user posture model by utilizing the user posture data transmitted in real time in the model training module; the model updating module updates the user posture model through model migration to construct a user personalized posture model; the model training module is also connected with an updating strategy optimization module, and the updating strategy optimization module is used for optimizing the updating decision of the user posture model in the model updating module through the consistency check of the user posture data.

The method for monitoring the posture change based on the plantar pressure sensor comprises the following steps:

step (1) acquiring user posture information by using a plantar pressure sensor;

the size and the distribution according to plantar pressure can reflect the change of human body attitude effectively, thereby obtain some physiology or pathological information of human body, plantar pressure sensor sets up to integrated dot matrix pressure distribution sensor, plantar pressure sensor includes a plurality of induction points, the diameter of every induction point is 5mm, the centre of a circle distance between adjacent induction point in horizontal direction or vertical direction is 7mm, gather the human body station in succession, sit, pressure information during the walking, plantar dynamics information is the internal mechanics cause of motion, plantar dynamics information has high relevance with human foot physiological structure and walking mode, the data acquisition module passes through plantar pressure sensor and gathers the original data sample under the different actions.

Step (2) denoising, filtering, clustering, segmenting and other processing are carried out on the collected data;

when a user walks on a blind road, a broken stone road, a muddy road or a grassland road and other special road conditions, the original data sample acquired by the plantar pressure sensor does not have a general representative value, so that the acquired original data sample needs to be filtered and denoised by the data processing module, and the interference of external noise in the data acquisition process of the plantar pressure sensor is reduced.

And carrying out segmentation and feature dimension reduction on the original data sample in the data processing module. First, from the analysis of the sole pressure load, the pressure distribution at each part of the sole when the human body stands is shown in fig. 2, the darker the color represents that the force applied to the region is larger, that is, the pressure distribution at each part of the sole when the human body stands is different, and the foot force points in the standard posture and the deviation posture are different. Secondly, from the analysis of human gait cycle, the gait cycle is as shown in fig. 3, one gait cycle is called a stride, and refers to the time from the landing of one side leg heel to the landing of the same side leg heel again. A stride comprises two phases: a double-leg supporting phase and a single-leg supporting phase; for a leg, a complete gait cycle includes a stance phase and a swing phase. When the human body walks on the flat ground, the basic relationship of each period is as follows:

D_sz＝0.248D_ci+0.143

D_z＝0.752D_ci-0.143

D_dz＝0.752D_ci-0.143

wherein D is_ciOne gait cycle; d_szIn the single-foot support period; d_dzThe two feet supporting period; d_zIs the support period of a certain leg in a walking cycle. And performing single-step segmentation on the original data sample by using a gait cycle, collecting the pressure characteristics of the left/right foot supporting phases, and performing characteristic dimension reduction by using a principal component analysis method.

Step (3) modeling the standard posture and the user posture by utilizing a DNN algorithm;

the user posture data obtained by the processing of the data processing module comprises step frequency, step speed, step length (left/right), step width, plantar pressure distribution space-time characteristics and the like, and the user posture model is formed by modeling the posture characteristics in the model building module by utilizing the deep neural network.

Step (4) realizing user personalized body state model migration through migration learning algorithms such as a learning frame of sharing weight joint optimization;

as shown in fig. 4, in order to reduce the amount of high data samples and high time cost required for repeatedly training the user posture model, a Learning framework (Learning algorithm) for jointly optimizing the shared weight between the source and target task models (L2 TL) is adopted in the model training module (migration Learning algorithm), wherein the calculation of the shared weight is implemented by using a Reinforcement Learning module (RL) based on a performance metric matrix of the target data set, so as to ensure that the weight of the feature in each source data set is adaptively output. L2TL is based on the similarity of the self-adaptive inferred domain of the user posture (target) data set, thereby achieving the high demand of the deep neural network on the data sample size and shortening the time of data sample acquisition and user posture model training.

The goal of L2TL is to adaptively learn weight distribution, and a framework for optimizing the shared weight of the source domain and the target domain is proposed.

The pre-training objective function is as follows:

wherein, L is an objective function, and the objective function L corresponding to the source domain_SAnd an objective function L corresponding to the target domain_TAnd (4) forming. Unlike DATL, which optimizes the loss function only in the source domain (i.e., an importance domain adaptive migration learning method based on the target dataset), pre-training of L2TL is achieved by joint optimization in both the source and target domains. D_SAnd D_TRespectively representing a given source (standard posture data) and target data set (target user posture data), B_SAnd B_TDenote the source and target batch sizes, respectively, α_s[i]And α_t[i]Respectively representing the scale parameters of a source domain and a target domain during the ith round of iteration,

weight parameter f for controlling the proportion of the source domain and the target domain in the optimization process_SAnd f_TThe source and target coding functions found using the neural network are represented separately. The training parameters of the whole objective function are omega and zeta_SAnd ζ_T. To benefit most from the source data set, as many training parameters as possible should be shared in the source and target domains. Based on the pre-training objective function, the learning objective of T2TL is to generalize from the maintain objective validation dataset into unknown samples and maximize the following evaluation metrics:

where R is an indicator of target performance, such as top-1 accuracy or area under the curve (AUC) for classification. Omega and zeta_TThe resulting parameters are optimized for pre-training. In the fine tuning process, optimization of Ns step source data set with sample weight λ (x, y) ═ 1 is considered first, and then optimization of the target data set using the pre-training weights is considered:

the first step is to optimize the loss function by using an optimization method based on gradient descent, so as to obtain the weight of a learning encoder omega and the weight ζ of a classifier_SAnd ζ_T：

At this stage, the neural network model is fixed, and the actions are sampled to determine the respective weights.

Second, optimize the strategy weight

The strategy weight is based on the encoder weight (omega, zeta) calculated in the last stage_S、ζ_T) Maximize the evaluation metric R on the target validation set:

this process is solved by using a strategy model

And α to achieve reward optimization is a reinforcement learning RL problem.

And thirdly, updating the model to complete the construction of the user personalized physical model.

Step 5, checking the accuracy of the user personalized posture model in real time aiming at the user posture data by using model consistency;

the system further optimizes the updating strategy of the user posture model. And (3) checking the accuracy of the user personalized posture model in real time aiming at the user posture data through a model consistency Check (CP). When the user posture data changes and continuously deviates from the original model, the posture data information of the user is collected again, and the current user posture model is rapidly updated by using transfer learning. The consistency check has the advantages that the reliability of the posture monitoring is evaluated and guaranteed, the fault-tolerant interval of the user personalized posture model can be set by a posture trainer and a user at the same time, and the effectiveness of posture correction and the comfort level of user training are considered.

For simple predictors, any finite sequence of samples (x)₁,y₁),...,(x_n-1,y_n-1) And a new object x_n。

And (4) predicting results:

the consistency test contains a Significance Level (Significance Level) parameter epsilon, i.e. the confidence is 1-epsilon, reflecting the confidence Level of the prediction result, any finite sample sequence (x)₁,y₁),...,(x_n-1,y_n-1) And a new object x_nThe consistency checker outputs a subset of the label space Y:

the subset satisfies:

for any epsilon_l＞ε₂. This means that for different significance levels epsilon the predicted regions are nested subsets of the label space Y. By varying the value of epsilon, the magnitude of the tolerance for body state deviation can be adjusted.

As the posture monitoring belongs to real-time data acquisition, the method has better effectiveness in an Online Mode (Online Mode) of consistency detection. The online mode of the consistency check refers to: after the evaluation of the single body state is finished, the body state information and the body state label thereof are used as a new sample to be added into the original sample set to generate a new sample set, and then the evaluation is carried out on the next object based on the new sample set. In the course of prediction, the sample set is constantly being updated. When the user posture is continuously changed and the accumulated user posture data enables the error proportion of the current user posture model to be larger than the preset significance level epsilon, the fact that the current user posture model needs to be optimally updated is indicated.

The user posture data continuously updates and changes along with the increase of time t (n)₁,n₂,…,n_t) And when the error prediction proportion exceeds a preset significance level epsilon, carrying out model adjustment on the cloud and edge user posture models by using a transfer learning method facing AIoT.

And (6) carrying out model quick updating on the posture data developed by the user through transfer learning, and then carrying out comprehensive judgment analysis.

The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments, and all technical solutions belonging to the idea of the present invention belong to the protection scope of the present invention. It should be noted that modifications and embellishments within the scope of the present invention may be made by those skilled in the art without departing from the principle of the present invention, and such modifications and embellishments should also be considered as within the scope of the present invention.