CN103617347A

CN103617347A - Method for dynamically monitoring abnormal states of users of treadmill

Info

Publication number: CN103617347A
Application number: CN201310568078.9A
Authority: CN
Inventors: 郭春生; 沈佳; 徐俊
Original assignee: Hangzhou Dianzi University
Current assignee: Hangzhou Dianzi University
Priority date: 2013-11-14
Filing date: 2013-11-14
Publication date: 2014-03-05

Abstract

The invention discloses a treadmill-based dynamic monitoring method for abnormal states of users. First, collect the training data of treadmill users, then preprocess the data, and then use LDA to extract the semantic features of the processed data, then use this semantic feature as an observation to accumulate normal state knowledge, that is, HMM training, and finally use the obtained The model and semantic features are used to judge the abnormal state. The present invention utilizes a dynamic monitoring algorithm based on the treadmill user's abnormal state to realize rapid and accurate monitoring, achieves the purpose of accumulating knowledge of normal state while monitoring while training, and better solves the monitoring problem when the user's health state is unknown, thereby Make monitoring smarter.

Description

Treadmill user abnormal state monitoring method

技术领域technical field

本发明具体涉及一种基于跑步机用户异常状态动态监测的方法，按国际专利分类表（IPC）划分属于物理部，仪器分部，计算；推算，计数大类，电数字数据处理小类，特别用于特定应用的数字计算或数据处理设备或数据处理方法大组的技术领域。The invention specifically relates to a dynamic monitoring method based on the abnormal state of treadmill users. According to the International Patent Classification (IPC), it belongs to the Department of Physics, Instrument Division, Calculation; Calculation, Counting, and Electronic Digital Data Processing, especially The technical field of a large group of digital computing or data processing devices or data processing methods for specific applications.

背景技术Background technique

国内外学者一直在探索对于人体平衡功能障碍的检测手段,发明了许多种检测方法。传统的平衡功能检测方法是主观观察法：临床最早建立的Romberg氏检查法，又名闭目站立检查法，他是用肉眼观察受检者在两足并拢直立情况下，睁眼、闭眼时身体摇晃情况。1966年，Graybie改进了Romberg试验中单纯的双足并拢直立试验，他在临床上开始使用单腿直立实验法及强化的Romberg试验法，具体方法就是前者要求受检者在30s内单腿直立，先睁眼，后闭眼；后者则是在60s内使受检者两足一前一后，足尖接足跟的直立方式，因前者负重面积及支撑面小，其难度大于后者，而后者所需的技巧比原有的Romberg试验法高。以上这些目测法只能定性地进行评定，很难作定量分析。后来逐渐改善了这种检测方法，进一步引入了量表评定法：目前，在国外临床上常用的平衡表主要有Berg平衡表、Tinetti量表、“起立-行走”计时测试、功能性前伸及跌倒危险指数等。这些检测方法尽管带有定量性质，但仍属功能的综合评估，带有主观性，缺乏对平衡障碍的摇摆特点深入细致的分析，所以其应用价值有限。平衡功能测试仪检查法：1976年Terekhov首先应用压力平板即固定平板评定平衡功能，记录人体压力中心在平台上变化的轨迹，从而反映人体的重心变化，为我们研究平衡功能提供了一个新的手段，在临床上也可用于对骨科、神经科及其它学科疾病所致的平衡障碍进行检查和诊断。随着计算机技术的发展，实现平衡功能监测的方法多种多样，平衡功能监测也逐渐转向用于医疗和保健，而同时随着跑步机的问世，使平衡功能监测用作日常保健成为了现实，不再使平衡功能监测单一的成为医疗诊断的一种方式。跑步机用户可以在日常的锻炼过程中同时监测自己的健康状态，完全不必刻意的去进行健康状态的监测，减轻用户的经济和心理双重负担，因此，这种健康监测方式在未来应有更为广阔的使用空间。Scholars at home and abroad have been exploring the detection methods for human balance dysfunction, and invented many detection methods. The traditional balance function testing method is subjective observation: Romberg’s test, also known as standing test with eyes closed, is the earliest clinically established test. Body shaking situation. In 1966, Graybie improved the simple upright test with the feet close together in the Romberg test. He began to use the single-leg upright test method and the enhanced Romberg test method in clinical practice. The specific method is that the former requires the subject to stand upright on one leg within 30 seconds. Open the eyes first, then close the eyes; the latter is to put the subject's two feet one in front of the other within 60 seconds, and the toes connect to the heels. The former is more difficult than the latter because of its small load-bearing area and support surface. The latter requires more skill than the original Romberg test. The above visual inspection methods can only be evaluated qualitatively, and it is difficult to make quantitative analysis. Later, this detection method was gradually improved, and the scale evaluation method was further introduced: at present, the commonly used balance scales in clinical practice abroad mainly include Berg balance scale, Tinetti scale, "stand-to-walk" timing test, functional reach and fall hazard index, etc. Although these detection methods are quantitative in nature, they are still comprehensive evaluations of functions, which are subjective and lack in-depth and detailed analysis of the swing characteristics of balance disorders, so their application value is limited. Balance function tester inspection method: In 1976, Terekhov first applied the pressure plate, that is, the fixed plate, to evaluate the balance function, and recorded the track of the change of the pressure center of the human body on the platform, so as to reflect the change of the center of gravity of the human body, providing a new means for us to study the balance function It can also be used clinically to examine and diagnose balance disorders caused by diseases in orthopedics, neurology and other disciplines. With the development of computer technology, there are many ways to realize balance function monitoring, and balance function monitoring has gradually turned to medical and health care. At the same time, with the advent of treadmills, balance function monitoring has become a reality for daily health care. It is no longer a single way to make balance function monitoring a medical diagnosis. Treadmill users can monitor their health status at the same time in the daily exercise process, and there is no need to deliberately monitor the health status, reducing the user's economic and psychological burden. Therefore, this health monitoring method should be more important in the future. Wide use space.

发明内容Contents of the invention

为了能更加简单、快速、准确的检测出用户健康状态的变化情况，本发明提出了基于跑步机用户异常状态监测方法。该方法具体思路如下:首先采集用户每次使用跑步机时所产生的训练数据；然后对数据进行预处理得到更加准确且能够进行LDA处理的数据形式，通过LDA提取数据的语义特征，将得到的语义特征视为HMM的观测量；最后训练出数据的先验知识库即HMM模型，同时计算该观测量在上次保留的HMM监测模型下的数据产生概率值，并计算该概率值与前一次获取的数据产生概率值之差，若差值的绝对值小于更新因子的绝对值，则视为用户的健康未发生变化，并将该数据产生概率值与前面保留的数据产生概率值进行相加，以便用来更新更新因子，同时将本次数据训练出的HMM模型作为下次用户产生数据的监测模型和训练的基础模型。反之，则代表着用户的健康状态发生了变化，可以记录本次异常，同时舍弃本次的数据产生概率值和HMM模型。不断的重复这个过程，就会将用户的健康信息进行累积。若某天用户产生数据的概率值减去上一次所获的数据产生概率值之差的严重偏离了更新因子的绝对值，则表明该用户的健康发生了变化，这时可以提醒用户到医院做进一步的专门检查。In order to detect the change of the user's health state more simply, quickly and accurately, the present invention proposes a method for monitoring the abnormal state of the user based on the treadmill. The specific idea of this method is as follows: first, collect the training data generated when the user uses the treadmill each time; then preprocess the data to obtain a more accurate data form that can be processed by LDA, extract the semantic features of the data through LDA, and obtain the Semantic features are regarded as observations of HMM; finally, the prior knowledge base of the data is trained, that is, the HMM model, and the probability value of the observation under the last retained HMM monitoring model is calculated at the same time, and the difference between the probability value and the previous time is calculated. The difference between the obtained data generation probability values, if the absolute value of the difference is less than the absolute value of the update factor, it is considered that the user's health has not changed, and the data generation probability value is added to the previously reserved data generation probability value , so as to update the update factor, and at the same time, use the HMM model trained by this data as the monitoring model and the basic model for training of the next user-generated data. On the contrary, it means that the user's health status has changed, and this abnormality can be recorded, and the probability value and HMM model generated by this data will be discarded at the same time. Repeating this process continuously will accumulate the user's health information. If the difference between the probability value of data generated by the user on a certain day minus the probability value of the data generated last time seriously deviates from the absolute value of the update factor, it indicates that the user's health has changed, and the user can be reminded to go to the hospital for treatment. Further specialized inspections.

为了方便描述本发明的内容，首先作以下术语定义：In order to describe content of the present invention conveniently, at first do following term definition:

定义1词汇Definition 1 Vocabulary

词汇一般定义是一篇文档或者语言里所有的词和固定短语的总和，本发明定义是将用户产生的平衡数据经过处理后得到的数据形式视为词汇。The general definition of vocabulary is the sum of all words and fixed phrases in a document or language. The definition of the present invention is to regard the data form obtained after processing the balance data generated by users as vocabulary.

定义2语义特征Definition 2 Semantic Features

语义特征是一篇文档中能够描述这篇文档主题分布的参数。本发明定义为能够最佳代表每个用户平衡能力信息数据的量。Semantic features are parameters in a document that can describe the topic distribution of this document. The present invention is defined as the amount of information data that can best represent each user's balance capability.

定义3数据产生概率值Definition 3 Probability value of data generation

假设用户第n-1天产生的数据经过LDA提取出语义特征为O_n-1且由它训练出的HMM模型为λ_n-1，同时上次获取的HMM模型λ_n-2被保留下来，那么用户第n-1天产生的训练数据经过LDA提取语义特征O_n-1在λ_n-2下产生的概率P(O_n-1|λ_n-2)被定义为语义特征O_n-1在模型λ_n-2下的数据产生概率值，并规定在用户第一次产生的训练数据经过LDA提取语义特征训练出的模型下，该语义特征产生的概率P(O₀|λ₀)为初始数据产生概率值。Assuming that the data generated by the user on day n-1 is extracted by LDA, the semantic feature is On _-1 and the HMM model trained by it is λ _n-1 , and the HMM model λ _n-2 obtained last time is retained. Then the training data generated by the user on the n-1 day is extracted by LDA, and the probability P(O _n-1 |λ _n-2 ) of the semantic feature O _n- 1 under λ _n-2 is defined as the semantic feature O _n-1 The probability value of the data under the model λ _n-2 is generated, and it is stipulated that the probability P(O ₀ |λ ₀ ) generated by the semantic feature is: The initial data yields probability values.

定义4更新因子Definition 4 update factor

假设用户N天的使用中，有m项数据产生概率值被保留了下来，设这m项数据产生概率值为:P(O₁|λ₀),P(O₄|λ₃),...,P(O_n|λ_n-1)，相加后取m项数据产生概率值的均值为：Assume that during the user's N days of use, there are m items of data generation probability values that are retained, and the m items of data generation probability values are: P(O ₁ |λ ₀ ),P(O ₄ |λ ₃ ),.. .,P(O _n |λ _n-1 ), the average value of the probability value of the m items of data after addition is:

$mean mean = = \frac{P P (({O o}_{11} | | {λ λ}_{00})) + + P P (({O o}_{44} | | {λ λ}_{33})) + + . . . . . . + + P P (({O o}_{n no} | | {λ λ}_{n no - - 11}))}{m m}$

设更新因子为L,则L＝mean-P(O₀|λ₀)。Let the update factor be L, then L=mean-P(O ₀ |λ ₀ ).

定义5前向算法Definition 5 Forward Algorithm

前向算法是用来计算给定隐马尔可夫模型（HMM）后一个观察序列的概率，给定这种算法，可以直接用来确定对于已知的一个观察序列，在一些隐马尔科夫模型（HMM）中哪一个HMM最好的描述了它--先用前向算法评估每一个（HMM），再选取其中概率最高的一个。The forward algorithm is used to calculate the probability of an observation sequence after a given hidden Markov model (HMM). Given this algorithm, it can be directly used to determine that for a known observation sequence, in some hidden Markov model Which of the (HMMs) best describes it - Evaluate each (HMM) with a forward algorithm and pick the one with the highest probability.

定义6前向-后向算法Definition 6 Forward-Backward Algorithm

前向-后向算法又称之为Baum-Welch算法，是EM算法的一个特例。由前向算法和后向算法组合而成，是解决HMM三个基本问题中的学习问题，给定一系列的观察序列，通过不断迭代寻求最佳的HMM参数λ。The forward-backward algorithm, also known as the Baum-Welch algorithm, is a special case of the EM algorithm. Combining the forward algorithm and the backward algorithm, it solves the learning problem among the three basic problems of HMM. Given a series of observation sequences, iteratively seeks the best HMM parameter λ.

本发明提出了基于跑步机的用户异常状态动态监测方法，该方法在跑步机的基础上对用户健康状况进行监测，包括基于LDA的数据语义特征提取和HMM的动态评估和学习两项关键技术。具体处理步骤如下：The invention proposes a treadmill-based user abnormal state dynamic monitoring method, which monitors the user's health status on the basis of the treadmill, including two key technologies: LDA-based data semantic feature extraction and HMM dynamic evaluation and learning. The specific processing steps are as follows:

步骤一：平衡数据采集Step 1: Balance Data Acquisition

跑步机按照驱动动力可分为机械跑步机和电动跑步机，目前家用跑步机多为电动跑步机，因此本发明是基于家用的电动跑步机。将柔性阵列压力传感器安装在跑步机前轮滚筒两个轴承的上部和前部，以及后轮滚筒两个轴承的上部和后部，四个轴承的压力传感器感受到压力的作用会产生八路电压信号，将这八路电压信号经过计算可得出用户压力重心在跑步机的投影位置，八个传感器的位置坐标为A(x₁,y₁),B(x₂,y₂),C(x₃,y₃),D(x₄,y₄),E(x₅,y₅),F(x₆,y₆),G(x₇,y₇),H(x₈,y₈)，为得到大量的训练数据，将相邻两个传感器产生数据进行差分处理，得到AB两个位置坐标的差分数据D_ab(Δx₁₂,Δy₁₂)，BC位置坐标的差分数据D_bc(Δx₂₃,Δy₂₃)，CD位置坐标的差分数据D_cd(Δx₃₄,Δy₃₄)，AD位置坐标的差分数据D_ad(Δx₁₄,Δy₁₄)，同理可以得到D_ef(Δx₅₆,Δy₅₆)，D_fg(Δx₆₇,Δy₆₇),D_gh(Δx₇₈,Δy₇₈),D_eh(Δx₅₈,Δy₅₈)，Δx₁₂表示A位置的x减去B位置的x的差值，Δy₁₂表示A位置的y减去B位置的y的差值，Δx₅₆表示E位置的x减去F位置的x的差值，Δy₅₆表示E位置的y减去F位置的y的差值，以此类推Δx₅₈,Δy₅₈等差值也具有相同含义。同时将获取的差分数据进行均方差处理：Treadmills can be divided into mechanical treadmills and electric treadmills according to driving power, and most household treadmills are electric treadmills at present, so the present invention is based on household electric treadmills. Install the flexible array pressure sensor on the upper part and the front part of the two bearings of the front wheel roller of the treadmill, and the upper part and the rear part of the two bearings of the rear wheel roller. The pressure sensors of the four bearings will generate eight voltage signals when they feel the pressure , the eight-way voltage signals can be calculated to obtain the projected position of the user's pressure center of gravity on the treadmill, and the position coordinates of the eight sensors are A(x ₁ ,y ₁ ), B(x ₂ ,y ₂ ), C(x ₃ ,y ₃ ),D(x ₄ ,y ₄ ),E(x ₅ ,y ₅ ),F(x ₆ ,y ₆ ),G(x ₇ ,y ₇ ),H(x ₈ ,y ₈ ), In order to obtain a large amount of training data, the data generated by two adjacent sensors are differentially processed to obtain the differential data D _ab (Δx ₁₂ ,Δy ₁₂ ) of the two position coordinates of AB and the differential data D _bc (Δx ₂₃ , Δy 12 ) of the BC position coordinates. Δy ₂₃ ), CD position coordinate difference data D _cd (Δx ₃₄ , Δy ₃₄ ), AD position coordinate difference data D _ad (Δx ₁₄ , Δy ₁₄ ), similarly, De _ef (Δx ₅₆ , Δy ₅₆ ), D _fg (Δx ₆₇ , Δy ₆₇ ), D _gh (Δx ₇₈ , Δy ₇₈ ), D _eh (Δx ₅₈ , Δy ₅₈ ), Δx ₁₂ represents the difference between x at position A minus x at position B, and Δy ₁₂ represents The difference between y at position A minus y at position B, Δx ₅₆ represents the difference between x at position E minus x at position F, and Δy ₅₆ represents the difference between y at position E minus y at position F, so that By analogy, Δx ₅₈ , Δy ₅₈ and other differences also have the same meaning. At the same time, the obtained differential data is subjected to mean square error processing:

${μ μ}_{Δx Δx} = = \frac{11}{44} (({Δx Δx}_{1212} + + {Δx Δx}_{23 twenty three} + + {Δx Δx}_{3434} + + {Δx Δx}_{1414}))$

${μ μ}_{Δy Δy} = = \frac{11}{44} (({Δy Δy}_{1212} + + {Δy Δy}_{23 twenty three} + + {Δy Δy}_{3434} + + {Δy Δy}_{1414}))$

${μ μ}_{Δx Δx}^{' '} = = \frac{11}{44} (({Δx Δx}_{5656} + + {Δx Δx}_{6767} + + {Δx Δx}_{7878} + + {Δx Δx}_{5858}))$

${μ μ}_{Δy Δy}^{' '} = = \frac{11}{44} (({Δy Δy}_{5656} + + {Δy Δy}_{6767} + + {Δy Δy}_{7878} + + {Δy Δy}_{5858}))$

$var var (({Δ Δ}_{x x})) = = \sqrt{(({(({Δx Δx}_{1212} - - {μ μ}_{Δx Δx}))}^{22} + + {(({Δx Δx}_{23 twenty three} - - {μ μ}_{Δx Δx}))}^{22} + + {(({Δx Δx}_{3434} - - {μ μ}_{Δx Δx}))}^{22} + + {(({Δx Δx}_{1414} - - {μ μ}_{Δx Δx}))}^{22})) / / 44}$

$var var (({Δ Δ}_{y the y})) = = \sqrt{(({(({Δy Δy}_{1212} - - {μ μ}_{Δy Δy}))}^{22} + + {(({μy μy}_{23 twenty three} - - {μ μ}_{Δy Δy}))}^{22} + + {(({Δy Δy}_{3434} - - {μ μ}_{Δy Δy}))}^{22} + + {(({Δy Δy}_{1414} - - {μ μ}_{Δy Δy}))}^{22})) / / 44}$

$var var {(({Δ Δ}_{x x}))}^{' '} = = \sqrt{(({(({Δx Δx}_{5656} - - {μ μ}_{Δx Δx}^{' '}))}^{22} + + {(({Δx Δx}_{6767} - - {μ μ}_{Δx Δx}^{' '}))}^{22} + + {(({Δx Δx}_{7878} - - {μ μ}_{Δx Δx}^{' '}))}^{22})) / / 44}$

$var var {(({Δ Δ}_{y the y}))}^{' '} = = \sqrt{(({(({Δy Δy}_{5656} - - {μ μ}_{Δy Δy}^{' '}))}^{22} + + {(({Δy Δy}_{6767} - - {μ μ}_{Δy Δy}^{' '}))}^{22} + + {(({Δy Δy}_{7878} - - {μ μ}_{Δy Δy}^{' '}))}^{22} + + {(({Δy Δy}_{5858} - - {μ μ}_{Δy Δy}^{' '}))}^{22})) / / 44}$

Var(Δx,Δy)＝(var(Δ_x),var(Δ_y))Var(Δx,Δy)=(var( _Δx ),var( _Δy ))

Var(Δx,Δy)′＝(var(Δ_x)′,var(Δ_y)′)Var(Δx,Δy)'=(var( _Δx )',var( _Δy )')

其中μ_Δx,μ′_Δx分别表示ABCD、EFGH八个坐标位置x坐标差值的均值，μ_Δy,μ′_Δy分别表示ABCD、EFGH八个坐标位置y坐标差值的均值，var(Δ_x),var(Δ_x)′分别表示ABCD、EFGH八个坐标位置x坐标差值的均方差，var(Δ_y),var(Δ_y)′分别表示ABCD、EFGH八个坐标位置y坐标差值的均方差，Var(Δx,Δy),Var(Δx,Δy)′表示ABCD、EFGH八个坐标的均方差位置坐标。Among them, μ _Δx , μ′ _Δx represent the average value of the x-coordinate difference of the eight coordinate positions of ABCD and EFGH respectively, μ _Δy , μ′ _Δy represent the mean value of the y-coordinate difference of the eight coordinate positions of ABCD and EFGH respectively, var(Δ _x ) , var(Δ _x )' represent the mean square error of the x-coordinate difference of the eight coordinate positions of ABCD and EFGH respectively, var(Δ _y ), var(Δ _y )' represent the y-coordinate difference of the eight coordinate positions of ABCD and EFGH respectively The mean square error, Var(Δx, Δy), Var(Δx, Δy)' represents the mean square error position coordinates of the eight coordinates of ABCD and EFGH.

将上述获取的18组数据组成18维的训练数据data：Combine the 18 sets of data obtained above into 18-dimensional training data data:

data＝(A(x₁,y₁),B(x₂,y₂),C(x₃,y₃),D(x₄,y₄),E(x₅,y₅),F(x₆,y₆),G(x₇,y₇),H(x₈,y₈),data＝(A(x ₁ ,y ₁ ),B(x ₂ ,y ₂ ),C(x ₃ ,y ₃ ),D(x ₄ ,y ₄ ),E(x ₅ ,y ₅ ),F( x ₆ ,y ₆ ),G(x ₇ ,y ₇ ),H(x ₈ ,y ₈ ),

D_ab(Δx₁₂,Δy₁₂),D_bc(Δx₂₃,Δy₂₃),D_cd(Δx₃₄,Δy₃₄),D_ad(Δx₁₄,Δy₁₄),D_ef(Δx₅₆,Δy₅₆),D _ab (Δx ₁₂ ,Δy ₁₂ ),D _bc (Δx ₂₃ ,Δy ₂₃ ),D _cd (Δx ₃₄ ,Δy ₃₄ ),D _ad (Δx ₁₄ ,Δy ₁₄ ),D _ef (Δx ₅₆ ,Δy ₅₆ ),

D_fg(Δx₆₇,Δy₆₇),D_gh(Δx₇₈,Δy₇₈),D_eg(Δx₅₈,Δy₅₈),Var(Δx,Δy),Var(Δx,Δy)′)D _fg (Δx ₆₇ ,Δy ₆₇ ),D _gh (Δx ₇₈ ,Δy ₇₈ ),D _eg (Δx ₅₈ ,Δy ₅₈ ),Var(Δx,Δy),Var(Δx,Δy)′)

步骤二：数据的预处理Step 2: Data preprocessing

将得到训练数据data，通过数据清洗去掉噪声和无关的数据，通过数据变换将数据转化成为适合信息处理的形式，即处理为LDA可以使用的文档形式。The training data will be obtained, the noise and irrelevant data will be removed through data cleaning, and the data will be converted into a form suitable for information processing through data transformation, that is, processed into a document form that can be used by LDA.

步骤三：LDA提取数据语义特征Step 3: LDA extracts data semantic features

LDA将每个文档表示为一个主题混合，每个主题是固定词表上的一个多项式分布。LDA假设文档由一个主题混合产生，同时每个主题是在固定词表上的一个多项式分布；这些主题被集合中的所有文档共享；每个文档有一个特定的主题混合比例，其从Dirichlet分布中抽样产生。作为一种生成式文档模型，用LDA提取文档的隐含语义结构和文档表征已经成功地应用到很多文本相关的领域。LDA represents each document as a mixture of topics, where each topic is a multinomial distribution over a fixed vocabulary. LDA assumes that documents are produced by a mixture of topics, and each topic is a multinomial distribution on a fixed vocabulary; these topics are shared by all documents in the collection; each document has a specific topic mixture ratio, which is derived from the Dirichlet distribution Sampling generated. As a generative document model, using LDA to extract the hidden semantic structure and document representation of documents has been successfully applied to many text-related fields.

具体来说生成LDA生成过程如下：Specifically, the LDA generation process is as follows:

1）选择N，N服从Poisson(ξ)分布，N表示每篇文档的词汇量；1) Select N, N obeys the Poisson (ξ) distribution, and N represents the vocabulary of each document;

2）选择θ，θ服从Dirichlet(α)分布，θ是列向量，代表的是主题发生的概率，α是Dirichlet分布的参数；2) Select θ, θ obeys the Dirichlet (α) distribution, θ is a column vector, representing the probability of the topic occurring, and α is a parameter of the Dirichlet distribution;

3）选择主题参数z_mn，z_mn服从Multinomial(θ)分布；3) Select the topic parameter z _mn , z _mn obeys the Multinomial(θ) distribution;

4）选择词汇参数w_mn，w_mn服从Multinomial分布，其中

是主题词项分布矩阵。4) Select the vocabulary parameter w _mn , w _mn obeys Multinomial distribution, where

is the subject term distribution matrix.

主题模型的推理是生成文档的逆向过程，已知α和先验分布β，根据文档生成过程可以写出各种随机变量D、z和θ的联合概率，其中D＝{w₁,w₂...w_M}，表示文档集合，z表示主题，w₁,w₂,...,w_M表示各个词汇，其下标表示此词汇是第多少个词汇。The reasoning of the topic model is the reverse process of document generation. Knowing α and prior distribution β, the joint probability of various random variables D, z and θ can be written according to the document generation process, where D={w ₁ ,w ₂ . ..w _M }, represents the document collection, z represents the topic, w ₁ ,w ₂ ,...,w _M represents each vocabulary, and its subscript indicates how many words this vocabulary is.

$p p ((D D.,, z z,, θ θ | | α α,, β β)) = = {Π Π}_{m m = = 11}^{M m} [[p p (({θ θ}_{m m} | | α α)) {Π Π}_{n no = = 11}^{N N} p p (({z z}_{mn mn} | | {θ θ}_{m m})) p p (({w w}_{mn mn} | | {β β}_{{z z}_{mn mn}}))]]$

其中M表示文档数目，θ_m是第m篇文档的主题分布。Where M represents the number of documents, and θ _m is the topic distribution of the mth document.

对主题先验参数θ进行积分得到p(z,w|α,β)，则根据贝叶斯公式，Integrate the subject prior parameter θ to get p(z,w|α,β), then according to the Bayesian formula,

$p p (({z z}_{n no} | | {z z}_{- - n no},, w w,, α α,, β β)) = = \frac{p p (({z z}_{n no},, {z z}_{- - n no},, w w | | α α,, β β))}{p p (({z z}_{- - n no},, w w | | α α,, β β))}$

其中，z_n表示第m篇文档中第n篇词汇所对应的主题词，z_-n表示除第m篇文档中第n篇词汇所对应的w主题词以外，其它所有的已知文档中的词汇所对应的主题词，表示文档词汇。上式可用来完成主题参数的估计，即确定其对应的具有语义特征的关键词。Among them, z _n represents the keyword corresponding to the nth vocabulary in the mth document, and z _-n represents the w keyword in all known documents except the w keyword corresponding to the nth vocabulary in the mth document. The keyword corresponding to the vocabulary, which represents the document vocabulary. The above formula can be used to complete the estimation of topic parameters, that is, to determine the corresponding keywords with semantic features.

给定α和β情况下，主题先验参数θ、主题z以及每篇文档词汇w的联合分布可以表示为：Given α and β, the joint distribution of topic prior parameters θ, topic z, and vocabulary w of each document can be expressed as:

$p p ((w w,, z z,, θ θ | | α α,, β β)) = = p p ((θ θ | | α α)) {Π Π}_{n no = = 11}^{N N} p p (({z z}_{n no} | | θ θ)) p p (({w w}_{n no} | | {z z}_{n no},, β β))$

其中，整合θ和z，得到一个词汇的边缘分布：Among them, integrate θ and z to get the marginal distribution of a vocabulary:

$p p ((w w | | α α,, β β)) = = &Integral; &Integral; p p ((θ θ | | α α)) (({Π Π}_{n no = = 11}^{N N} \underset{{z z}_{n no}}{Σ Σ} p p (({z z}_{n no} | | θ θ)) p p (({w w}_{n no} | | {z z}_{n no},, β β)))) dθ dθ$

依据上面的边缘分布，得到M篇文档概率分布为：According to the above marginal distribution, the probability distribution of M documents is obtained as:

$p p ((D D. | | α α,, β β)) = = {Π Π}_{m m = = 11}^{M m} &Integral; &Integral; p p (({θ θ}_{m m} | | α α)) (({Π Π}_{n no = = 11}^{N N} \underset{{z z}_{mn mn}}{Σ Σ} p p (({z z}_{mn mn} | | {θ θ}_{m m})) P P (({w w}_{mn mn} | | {z z}_{mn mn},, β β)))) {dθ dθ}_{m m}$

其中，D代表M篇文档集合，m表示第m篇文档标签序号。Among them, D represents the set of M documents, and m represents the tag number of the mth document.

然后计算给定一篇文档条件下隐藏变量的后验分布，公式如下所述：Then calculate the posterior distribution of hidden variables given a document condition, the formula is as follows:

$p p ((θ θ,, z z | | w w,, α α,, β β)) = = \frac{p p ((θ θ,, z z,, w w | | α α,, β β))}{p p ((w w | | α α,, β β))}$

此后验分布采用变分EM算法进行推理可得到，取Q分布的函数

去逼近q(θ,z|w,α,β)，其中Q分布的形式如下：The posterior distribution can be obtained by reasoning with the variational EM algorithm, and the function of the Q distribution is taken as

To approximate q(θ,z|w,α,β), where the form of Q distribution is as follows:

其中γ表示狄利克雷参数，

表示多项式参数，表示多项式的第n个参数。where γ represents the Dirichlet parameter,

represents a polynomial parameter, Indicates the nth argument of the polynomial.

为了得到最佳的变分参数将变分分布

和真实分布q(θ,z|w,α,β)之间的KL散度最小化，最小值可以通过迭代的方法获得。通过推导可得到

和γ的迭代公式：In order to get the best variational parameters the variational distribution

The KL divergence between the real distribution q(θ,z|w,α,β) is minimized, and the minimum value can be obtained by an iterative method. can be obtained by derivation

and the iterative formula of γ:

其中：表示在γ条件下θ_i的条件概率期望值，Ψ是对数伽玛函数，

是条件多项式参数，α_i表示第i次迭代时的狄利克雷参数。更新Dirichlet参数α用的是Newton-Raphson方法。in: Indicates the conditional probability expectation value of θ _i under the condition of γ, Ψ is the logarithmic gamma function,

is the conditional polynomial parameter, and α _i represents the Dirichlet parameter at the ith iteration. The Newton-Raphson method is used to update the Dirichlet parameter α.

通过上述的推理，得到新参数γ的估计值，其中携带了是每次用户产生平衡数据组成文档的语义特征。这样就完成了LDA提取数据的语义特征的过程。Through the above reasoning, the estimated value of the new parameter γ is obtained, which carries the semantic features of the documents that are composed of balanced data generated by each user. This completes the process of LDA extracting the semantic features of the data.

步骤四：正常状态知识的训练Step 4: Training of normal state knowledge

HMM异常检测包含两个部分，前一个部分是对事件的评估，后一部分是参数的学习，参数学习就是模型参数未知，求最佳模型参数λ的问题。将用户产生的数据组成的数据文档的语义特征视为HMM模型的观测量O＝O₁,O₂,...,O_T。HMM anomaly detection consists of two parts. The first part is the evaluation of events, and the latter part is the learning of parameters. Parameter learning is the problem of finding the best model parameter λ for unknown model parameters. The semantic features of data documents composed of user-generated data are regarded as observations O=O ₁ , O ₂ ,...,O _T of the HMM model.

隐马尔可夫模型的第三个问题是如何根据观察序列O＝O₁,O₂,...,O_T，求得模型参数或调整模型参数λ，使得P(O|λ)最大。而第三个问题是通过前向-后向算法解决的。The third problem of the hidden Markov model is how to obtain the model parameters or adjust the model parameters λ according to the observation sequence O=O ₁ ,O ₂ ,...,O _T , so as to maximize P(O|λ). And the third problem is solved by forward-backward algorithm.

前向-后向算法forward-backward algorithm

首先定义两个变量，给定观察序列O和隐马尔可夫模型λ，定义t时刻位于隐藏状态S_i的概率变量为：γ_t(i)＝P(q_t＝s_i|O,λ)First define two variables, given the observation sequence O and the hidden Markov model λ, define the probability variable of being in the hidden state S _i at time t as: γ _t (i)=P(q _t =s _i |O,λ)

根据前向变量α_t(i)和后向变量β_t(i)的定义，将上式子用前向，后向变量表示：According to the definition of forward variable α _t (i) and backward variable β _t (i), the above formula is expressed by forward and backward variables:

${γ γ}_{t t} ((i i)) = = \frac{{α α}_{t t} ((i i)) {β β}_{t t} ((i i))}{P P ((O o | | λ λ))} = = \frac{{α α}_{t t} ((i i)) {β β}_{t t} ((i i))}{{Σ Σ}_{i i}^{N N} {α α}_{t t} ((i i)) {β β}_{i i} ((i i))}$

其中分母是保证：

where the denominator is the guarantee:

后向变量为：β_t(j)＝Σa_ijb_j(O_t+1)β_t+1(j),t＝T-1,T-2,...1，1≤i≤NThe backward variable is: β _t (j)=Σa _ij b _j (O _t+1 )β _t+1 (j), t=T-1,T-2,...1, 1≤i≤N

给定观察序列O和隐马尔可夫模型λ，定义t时刻位于隐藏状态S_i及t+1时刻位于隐藏状态S_j的概率变量为：Given the observation sequence O and the hidden Markov model λ, define the probability variable of being in the hidden state S _i at time t and in the hidden state S _j at time t+1 as:

ξ_t(i,j)＝P(q_t＝s_i,q_t+1＝s_j|O,λ)ξ _t (i,j)=P(q _t =s _i ,q _t+1 =s _j |O,λ)

同时该变量可有前向-后向变量表示：At the same time, the variable can be represented by a forward-backward variable:

${ξ ξ}_{t t} ((i i,, j j)) = = \frac{{α α}_{t t} ((i i)) {a a}_{ij ij} {b b}_{j j} (({O o}_{t t + + 11})) {β β}_{t t + + 11} ((j j))}{P P ((O o | | λ λ))} = = \frac{{α α}_{t t} ((i i)) {a a}_{ij ij} {b b}_{j j} (({O o}_{t t + + 11})) {β β}_{t t + + 11} ((j j))}{{Σ Σ}_{i i}^{N N} {Σ Σ}_{j j}^{N N} {α α}_{t t} ((i i)) {a a}_{ij ij} {b b}_{j j} (({O o}_{t t + + 11})) {β β}_{t t + + 11} ((j j))}$

上述定义的两个变量也存在着如下关系：The two variables defined above also have the following relationship:

${γ γ}_{t t} ((i i)) = = {Σ Σ}_{j j}^{N N} {ξ ξ}_{t t} ((i i,, j j))$

而且

表示观察序列O中从状态S_i出发的转移期望概率，表示观察序列中从状态S_i转移到状态S_j的转移期望概率，定义了两个变量及相应的转移期望，一种合理的重新估计隐马尔可夫模型（HMM）的参数π，A和B方法如下：and

Denotes the transition expectation probability starting from the state S _i in the observation sequence O, Denotes the transition expectation probability of transitioning from state S _i to state S _j in the observation sequence, defines two variables and the corresponding transition expectations, a reasonable re-estimation of the hidden Markov model (HMM) parameters π, A and B Methods as below:

$\overset{&OverBar; &OverBar;}{{π π}_{11}} = = {γ γ}_{11} ((i i))$

在t=1时处在状态i的期望概率。The expected probability of being in state i at t=1.

$\overset{&OverBar; &OverBar;}{{a a}_{ij ij}} = = \frac{{Σ Σ}_{t t = = 11}^{T T - - 11} {ξ ξ}_{t t} ((i i,, j j))}{{Σ Σ}_{t t = = 11}^{T T - - 11} {γ γ}_{t t} ((i i))}$

从状态i到状态j的转移期望概率除以从状态i转移出去的期望概率。The expected probability of transitioning from state i to state j divided by the expected probability of transitioning out of state i.

$\overset{&OverBar; &OverBar;}{{b b}_{j j}} ((k k)) = = \frac{{Σ Σ}_{t t = = 11}^{T T} {γ γ}_{t t} ((j j)) \times \times δ δ (({o o}_{t t},, {v v}_{k k}))}{{Σ Σ}_{t t = = 11}^{T T} {γ γ}_{t t} ((j j))}$

在状态j观察到v_k的期望概率除以处在状态j的期望概率。The expected probability of observing v _k in state j divided by the expected probability of being in state j.

其中，π表示初始状态概率，A表示状态转移矩阵，B表示混淆矩阵，表示训练后更新的初始状态概率，

表示训练后更新的由状态i转移到状态j的概率，

表示训练后更新的在状态j下观察到k的概率。通过上述的训练可以得到描述本次训练状态的模型的参数

Among them, π represents the initial state probability, A represents the state transition matrix, B represents the confusion matrix, Indicates the initial state probability updated after training,

Indicates the probability of transferring from state i to state j updated after training,

Denotes the probability of observing k at state j updated after training. Through the above training, the parameters of the model describing the training state can be obtained

步骤五：健康状态监测Step 5: Health status monitoring

此过程主要涉及到HMM三个问题中的评估问题，采取前向算法则可以解决这一问题。This process mainly involves the evaluation problem among the three problems of HMM, which can be solved by adopting the forward algorithm.

前向算法forward algorithm

定义t时刻状态j的局部概率为α_t(j)＝P_r（观察状态|隐藏状态j）×P_r(t时刻所有指向状态j的路径)，对于最后的状态，其局部概率包括了通过所有可能的路径到达这些状态的概率。Define the local probability of state j at time t as α _t (j)=P _r (observed state|hidden state j)×P _r (all paths pointing to state j at time t), for the final state, its local probability includes passing The probabilities of all possible paths to these states.

特别当t=1时，没有任何指向当前状态的路径。故t=1时位于当前状态的概率是初始概率，即P_r(state|t=1)=P(state)，因此，t=1时的局部概率α₁(i)等于当前状态的初始概率乘以相关的观察概率：Especially when t=1, there is no path to the current state. Therefore, the probability of being in the current state at t=1 is the initial probability, that is, P _r (state|t=1)=P(state), therefore, the local probability α ₁ (i) at t=1 is equal to the initial probability of the current state Multiply by the associated observation probabilities:

α₁(i)＝π(i)b_i(o₁)α ₁ (i)=π(i)b _i (o ₁ )

计算t>1时的局部概率α′sCalculate the local probability α's when t>1

t-1时刻α's给出了所有到达此t时刻状态的前一路径概率，因此，我们可以通过t-1时刻的局部概率定义t时刻的α′s:α's at time t-1 gives all the previous path probabilities to the state at time t, therefore, we can define α's at time t by the local probability at time t-1:

${α α}_{t t + + 11} ((j j)) = = {b b}_{j j} (({o o}_{t t + + 11})) \overset{N N}{Σ Σ} {α α}_{t t} ((i i)) {a a}_{ij ij}$

可以递归地计算给定隐马尔科夫模型后一个观察序列的概率，即通过t=1时刻的局部概率α′s计算t=2时刻的α′s，通过t=2时刻的α's计算t=3时刻的α′s等等直到t=T。给定隐马尔科夫模型的观察序列的概率就等于t=T时刻的局部概率之和。The probability of an observation sequence after a given hidden Markov model can be calculated recursively, that is, the local probability α's at time t=1 is used to calculate the α's at time t=2, and the α's at time t=2 is used to calculate t= α's at time 3 and so on until t=T. The probability of an observation sequence given a HMM is equal to the sum of the local probabilities at time t=T.

使用前向算法计算T时长的观察序列的概率：Use the forward algorithm to calculate the probability of an observation sequence of length T:

已知T时长观察序列：O＝O₁O₂...O_T Known T duration observation sequence: O=O ₁ O ₂ ... O _T

1）t=1时刻所有状态的局部概率α：1) Local probability α of all states at time t=1:

α₁(j)＝π(j)b_j(o₁)α ₁ (j)=π(j)b _j (o ₁ )

2）在t=2,...T时，对于每个状态的局部概率，由下式计算：2) At t=2,...T, the local probability for each state is calculated by the following formula:

${α α}_{t t + + 11} ((j j)) = = {b b}_{j j} (({o o}_{t t + + 11})) \overset{N N}{Σ Σ} {α α}_{t t} ((i i))$

3)最后，给定HMM,观察序列的概率等于T时刻所有局部概率之和：3) Finally, given the HMM, the probability of observing the sequence is equal to the sum of all local probabilities at time T:

${P P}_{r r} ((O o)) = = \overset{N N}{Σ Σ} {α α}_{T T} ((j j))$

用户每次使用跑步机都将产生训练数据，同时使用此数据训练出一个HMM模型，并计算此数据在上次获取的HMM模型下产生的概率，如果前后两次的数据产生概率值小于动态的更新因子，则保留此数据产生概率值和此次的HMM模型，不断的重复这个过程，将会获取更加精细的变化范围，缩小的变化范围将能更加准确的去判断每次使用跑步机所产生数据的异常状态，同时将获取逐渐趋近于正常事件先验知识的HMM模型，反之，则代表着用户的健康状态发生了变化，可以记录本次异常，同时舍弃本次的数据产生概率值和HMM模型。不断的重复这个过程，就到达了边训练边监测，边训练边优化的目的。Every time the user uses the treadmill, training data will be generated, and an HMM model will be trained using this data at the same time, and the probability of this data generated under the HMM model obtained last time will be calculated. To update the factor, keep the probability value of this data and the HMM model this time, and repeat this process continuously to obtain a more refined range of changes, and the narrowed range of changes will be able to more accurately judge the occurrence of each treadmill. The abnormal state of the data will at the same time obtain the HMM model that gradually approaches the prior knowledge of normal events. On the contrary, it means that the user's health status has changed, and this abnormal state can be recorded, and the probability value and HMM model. Repeating this process continuously achieves the goal of monitoring while training and optimizing while training.

本发明的有益效果在于针对传统的平衡功能健康监测方法对于用户无法进行动态、快速有效的监测，提供了一种基于LDA和HMM的动态平衡功能异常监测方法，该方法在LDA的语义特征提取的基础上，利用基于HMM动态评估和建模的算法作为平衡功能异常检测的方法，实现了动态积累正常状态信息和动态监测用户健康状态，并较好的解决了健康状态未知人群的监测问题。The beneficial effect of the present invention is to provide a dynamic balance function abnormality monitoring method based on LDA and HMM for the traditional balance function health monitoring method for the user can not be dynamically, quickly and effectively monitored, the method in the semantic feature extraction of LDA On the basis, using the algorithm based on HMM dynamic evaluation and modeling as a method of balanced function abnormality detection, it realizes the dynamic accumulation of normal state information and dynamic monitoring of user health status, and better solves the monitoring problem of people with unknown health status.

附图说明Description of drawings

图1为本发明的流程图；Fig. 1 is a flow chart of the present invention;

图2为本发明数据采集器的安装位置示意图；Fig. 2 is the installation position schematic diagram of data collector of the present invention;

图3为LDA生成模型图，其中图中ω是观测量，其他元素表示隐藏量。Figure 3 is a model diagram generated by LDA, in which ω is the observed quantity, and other elements represent the hidden quantity.

具体实施方式Detailed ways

本发明的实施流程图如图1和图2所示，具体实施步骤如下：The implementation flowchart of the present invention is as shown in Figure 1 and Figure 2, and concrete implementation steps are as follows:

步骤一：使用者通过跑步机完成各类型的跑步动作，由于电动跑步机的四个轴承都安装有柔性阵列压力传感器，人在跑步带上面运动后将会产生各种压力，四个轴承的压力传感器感受到压力的作用，产生八路电压信号，将这八路电压信号经过计算可得出用户压力重心在跑步机的投影位置，然后将传感器的位置坐标A(x₁,y₁),B(x₂,y₂),C(x₃,y₃),D(x₄,y₄),E(x₅,y₅),F(x₆,y₆),G(x₇,y₇),H(x₈,y₈),及相邻位置差分所得数据D_ab(Δx₁₂,Δy₁₂),D_bc(Δx₂₃,Δy₂₃),D_cd(Δx₃₄,Δy₃₄),D_ad(Δx₁₄,Δy₁₄),D_ef(Δx₅₆,Δy₅₆),D_fg(Δx₆₇,Δy₆₇),D_gh(Δx₇₈,Δy₇₈),D_eg(Δx₅₈,Δy₅₈)和它们的均方差数据Var(Δx,Δy),Var(Δx,Δy)′组成的18维训练数据data，通过无线网络传给主机。Step 1: The user completes various types of running actions through the treadmill. Since the four bearings of the electric treadmill are equipped with flexible array pressure sensors, people will generate various pressures after exercising on the running belt. The pressure of the four bearings The sensor feels the effect of pressure and generates eight voltage signals. After calculating these eight voltage signals, the projected position of the user's pressure center of gravity on the treadmill can be obtained, and then the sensor's position coordinates A(x ₁ ,y ₁ ), B(x ₂ ,y ₂ ),C(x ₃ ,y ₃ ),D(x ₄ ,y ₄ ),E(x ₅ ,y ₅ ),F(x ₆ ,y ₆ ),G(x ₇ ,y ₇ ) , H(x ₈ , y ₈ ), and the data obtained from the difference between adjacent positions D _ab (Δx ₁₂ , Δy ₁₂ ), D _bc (Δx ₂₃ , Δy ₂₃ ), D _cd (Δx ₃₄ , Δy ₃₄ ), D _ad ( Δx ₁₄ ,Δy ₁₄ ),D _ef (Δx ₅₆ ,Δy ₅₆ ),D _fg (Δx ₆₇ ,Δy ₆₇ ),D _gh (Δx ₇₈ ,Δy ₇₈ ),D _eg (Δx ₅₈ ,Δy ₅₈ ) and their mean The 18-dimensional training data data consisting of variance data Var(Δx, Δy), Var(Δx, Δy)' is transmitted to the host through the wireless network.

步骤二：将主机上的训练数据data进行预处理，使其成为更为优化的数据源。首先对数据进行数据的清洗，这一步主要是消除噪声和无关数据，通过填补遗漏数据，消除异常数据，平滑噪声数据，以及纠正不一致的数据，去掉数据中的噪声，填充空值、丢失值和处理不一致数据。由于原始数据的形式不适合LDA算法处理需要，还要对数据进行数据变换，将数据变换成文档形式，该文档内容代表着处理过的平衡功能数据。Step 2: Preprocess the training data on the host to make it a more optimized data source. First of all, data cleaning is performed on the data. This step is mainly to eliminate noise and irrelevant data. By filling missing data, eliminating abnormal data, smoothing noise data, and correcting inconsistent data, removing noise in data, filling null values, missing values and Handle inconsistent data. Since the form of the original data is not suitable for the processing needs of the LDA algorithm, the data needs to be transformed into a document form, and the content of the document represents the processed balance function data.

步骤三：将经过上一步处理得到的文档形式的数据，通过LDA进行语义特征提取，主题模型的推理生成文档的逆向过程，已知参数α和β，根据文档生成过程可写出各种随机变量z和θ的联合概率，其中D＝{w₁,w₂...w_M}，代表着M篇文档集合，参见图3：Step 3: The data in the form of documents processed in the previous step is used to extract semantic features through LDA, and the inference of the topic model generates the reverse process of documents. The parameters α and β are known, and various random variables can be written according to the document generation process. The joint probability of z and θ, where D={w ₁ ,w ₂ ...w _M }, represents a collection of M documents, see Figure 3:

进一步处理上式可以得到M篇文档概率分布为：By further processing the above formula, the probability distribution of M documents can be obtained as:

LDA关键的推理问题是计算给定一篇文档条件下隐藏变量的后验分布，公式如下：The key inference problem of LDA is to calculate the posterior distribution of hidden variables given a document, the formula is as follows:

通过变分推理算法，我们可以得到代表每个文档语义特征的参数估计式：Through the variational inference algorithm, we can obtain the parameter estimation formula representing the semantic features of each document:

这也就是要提取的每个文档的语义特征。This is the semantic feature of each document to be extracted.

步骤四：前一次使用跑步机产生的数据训练的并同时被保留的HMM模型被用来定义为当前的HMM模型λ＝(π,A,B)（基础模型）,利用该模型和本次使用者产生的数据，重新估计的HMM模型为

Step 4: The HMM model that was previously trained using the data generated by the treadmill and retained at the same time is used to define the current HMM model λ=(π,A,B) (basic model), using this model and this time The data generated by the former, the re-estimated HMM model is

$\overset{&OverBar; &OverBar;}{{π π}_{11}} = = {γ γ}_{11} ((i i))$

利用上述结论，即可进行模型估计，假设前一次使用跑步机产生的数据训练的并同时被保留的HMM模型被用来定义为当前的HMM模型λ＝(π,A,B)（基础模型）,利用该模型计算上面三个式子的右端，再定义重新估计的HMM模型为

那么上面三个式子的左端就是重估的HMM参数，亦即本次使用者产生数据训练出的HMM，通过不断的选取上次训练且被保留HMM模型作为下次训练的基础模型用于计算上面三个式子，由此可不断的重新估计HMM参数，由于每次基础模型的选取都是基于上次趋近于正常事件的数据训练出来的，那么在多次迭代后HMM模型将越来越趋近于正常事件的先验知识库，并且随着迭代次数的增加将与正常事件的知识库到达无限接近的程度。为下一步数据的监测提供了无限接近于正常事件先验知识库的模型，将使下一步的监测更加准确有效。Using the above conclusions, model estimation can be performed, assuming that the HMM model that was previously trained using the data generated by the treadmill and retained at the same time is used to define the current HMM model λ=(π,A,B) (basic model) , use this model to calculate the right-hand side of the above three equations, and then define the re-estimated HMM model as

Then the left end of the above three formulas is the re-evaluated HMM parameter, that is, the HMM trained by the data generated by the user this time, and the HMM model that was kept for the previous training is continuously selected as the basic model for the next training for calculation The above three formulas can continuously re-estimate the HMM parameters. Since each selection of the basic model is based on the data that is close to the normal event last time, the HMM model will become more and more accurate after multiple iterations. The closer it is to the prior knowledge base of normal events, and as the number of iterations increases, it will be infinitely close to the knowledge base of normal events. It provides a model infinitely close to the prior knowledge base of normal events for the next step of data monitoring, which will make the next step of monitoring more accurate and effective.

步骤五：使用者每次使用跑步机都将产生训练数据，经过一系列数据预处理和LDA提取数据的语义特征，同时使用此语义特征训练出一个HMM模型，并计算此语义特征在步骤四获取的HMM模型下产生的概率，如果前后两次的语义特征产生概率值小于动态的更新因子，则保留此数据产生概率值和此次的HMM模型，并将此次获取的HMM模型为下次训练数据的评估模型（监测模型）和基础模型，同时视为使用者的健康状态与之前的状况未发生变化。反之，则代表着使用者的健康状态发生了变化，可以记录本次异常并通知使用者，同时舍弃本次的数据产生概率值和HMM模型。不断的重复这个过程将不断的获取相对更加小的更新因子，亦即会获取更加小的变化范围，缩小的变化范围将能更加准确的去判断每次使用跑步机产生数据的异常状态，同时将获取逐渐趋近于正常事件先验知识的HMM模型，不断的重复这个过程，就可以边训练边监测，边训练边优化。Step 5: Every time the user uses the treadmill, the training data will be generated, after a series of data preprocessing and LDA to extract the semantic features of the data, and at the same time use this semantic feature to train an HMM model, and calculate this semantic feature to obtain in step 4 The probability generated under the HMM model, if the semantic feature generation probability value of the previous two times is less than the dynamic update factor, then keep the probability value of this data generation and the HMM model this time, and use the HMM model obtained this time as the next training The evaluation model (monitoring model) and the basic model of the data are considered to be unchanged from the previous state of the user's health status. On the contrary, it means that the health status of the user has changed, and the abnormality can be recorded and the user can be notified, and at the same time, the probability value and HMM model generated by the data of this time are discarded. Constantly repeating this process will continuously obtain a relatively smaller update factor, that is, a smaller change range will be obtained. The narrowed change range will be able to more accurately judge the abnormal state of the data generated each time the treadmill is used. At the same time, the Obtain an HMM model that gradually approaches the prior knowledge of normal events, and repeat this process continuously, so that you can monitor while training, and optimize while training.

Claims

1. treadmill user monitoring abnormal state method, is characterized in that the concrete steps of the method are:

The collection of step 1. equilibrium criterion, specifically: the top and the front portion that flexible array pressure sensor are arranged on to two bearings of treadmill front-wheel cylinder, and top and the rear portion of two bearings of trailing wheel cylinder, the effect that the pressure transducer of four bearings is experienced pressure can produce eight road voltage signals, Jiang Zhe eight road voltage signals are through can be calculated user's pressure center of gravity at the projected position of treadmill, and the position coordinates of eight sensors is A (x ₁, y ₁), B (x ₂, y ₂), C (x ₃, y ₃), D (x ₄, y ₄), E (x ₅, y ₅), F (x ₆, y ₆), G (x ₇, y ₇), H (x ₈, y ₈), for obtaining a large amount of training datas, adjacent two sensors are produced to data and carry out difference processing, obtain the differential data D of two position coordinateses of AB _ab(Δ x ₁₂, Δ y ₁₂), the differential data D of BC position coordinates _bc(Δ x ₂₃, Δ y ₂₃), the differential data D of CD position coordinates _cd(Δ x ₃₄, Δ y ₃₄), the differential data D of AD position coordinates _ad(Δ x ₁₄, Δ y ₁₄), in like manner can obtain D _ef(Δ x ₅₆, Δ y ₅₆), D _fg(Δ x ₆₇, Δ y ₆₇), D _gh(Δ x ₇₈, Δ y ₇₈), D _eh(Δ x ₅₈, Δ y ₅₈), Δ x ₁₂represent that the x of A position deducts the difference of the x of B position, Δ y ₁₂represent that the y of A position deducts the difference of the y of B position, Δ x ₅₆represent that the x of E position deducts the difference of the x of F position, Δ y ₅₆represent that the y of E position deducts the difference of the y of F position, by that analogy Δ x ₅₈, Δ y ₅₈etc. difference, also there is identical meanings; The differential data obtaining is carried out to mean square deviation processing simultaneously:

μ_{Δx} = \frac{1}{4} ({Δx}_{12} + {Δx}_{23} + {Δx}_{34} + {Δx}_{14})

μ_{Δy} = \frac{1}{4} ({Δy}_{12} + {Δy}_{23} + {Δy}_{34} + {Δy}_{14})

μ_{Δx}^{'} = \frac{1}{4} ({Δx}_{56} + {Δx}_{67} + {Δx}_{78} + {Δx}_{58})

μ_{Δy}^{'} = \frac{1}{4} ({Δy}_{56} + {Δy}_{67} + {Δy}_{78} + {Δy}_{58})

var (Δ_{x}) = \sqrt{({({Δx}_{12} - μ_{Δx})}^{2} + {({Δx}_{23} - μ_{Δx})}^{2} + {({Δx}_{34} - μ_{Δx})}^{2} + {({Δx}_{14} - μ_{Δx})}^{2}) / 4}

var (Δ_{y}) = \sqrt{({({Δy}_{12} - μ_{Δy})}^{2} + {({μy}_{23} - μ_{Δy})}^{2} + {({Δy}_{34} - μ_{Δy})}^{2} + {({Δy}_{14} - μ_{Δy})}^{2}) / 4}

var {(Δ_{x})}^{'} = \sqrt{({({Δx}_{56} - μ_{Δx}^{'})}^{2} + {({Δx}_{67} - μ_{Δx}^{'})}^{2} + {({Δx}_{78} - μ_{Δx}^{'})}^{2}) / 4}

var {(Δ_{y})}^{'} = \sqrt{({({Δy}_{56} - μ_{Δy}^{'})}^{2} + {({Δy}_{67} - μ_{Δy}^{'})}^{2} + {({Δy}_{78} - μ_{Δy}^{'})}^{2} + {({Δy}_{58} - μ_{Δy}^{'})}^{2}) / 4}

Var(Δx,Δy)＝(var(Δ _x),var(Δ _y))

Var(Δx,Δy)′＝(var(Δ _x)′,var(Δ _y)′)

μ wherein _{Δ x}, μ ' _{Δ x}the average that represents respectively ABCD, eight coordinate position x coordinate differences of EFGH, μ _{Δ y}, μ ' _{Δ y}the average that represents respectively ABCD, eight coordinate position y coordinate differences of EFGH, var (Δ _x), var (Δ _x) ' respectively represents the mean square deviation of ABCD, eight coordinate position x coordinate differences of EFGH, var (Δ _y), var (Δ _y) ' respectively represents the mean square deviation of ABCD, eight coordinate position y coordinate differences of EFGH, Var (Δ x, Δ y), and Var (Δ x, Δ y) ' represents the mean square deviation position coordinates of ABCD, eight coordinates of EFGH;

The training data data that above-mentioned 18 groups of data obtaining is formed to 18 dimensions:

data＝(A(x ₁,y ₁),B(x ₂,y ₂),C(x ₃,y ₃),D(x ₄,y ₄),E(x ₅,y ₅),F(x ₆,y ₆),G(x ₇,y ₇),H(x ₈,y ₈),

D _ab(Δx ₁₂,Δy ₁₂),D _bc(Δx ₂₃,Δy ₂₃),D _cd(Δx ₃₄,Δy ₃₄),D _ad(Δx ₁₄,Δy ₁₄),D _ef(Δx ₅₆,Δy ₅₆),

D _fg(Δx ₆₇,Δy ₆₇),D _gh(Δx ₇₈,Δy ₇₈),D _eg(Δx ₅₈,Δy ₅₈),Var(Δx,Δy),Var(Δx,Δy)′)

The pre-service of step 2. data, specifically: will obtain training data data, and by data cleansing, remove noise and irrelevant data, and by data transformation, data be transformed into the form that is applicable to information processing, and be treated to the operable document form of LDA;

Step 3.LDA extracts data semantic feature, and specifically: LDA is that a theme mixes by each document representation, each theme is a fixing multinomial distribution on vocabulary; LDA hypothesis document is mixed and is produced by a theme, and each theme is a multinomial distribution on fixing vocabulary simultaneously; These themes are by all document sharing in gathering; Each document has a specific theme blending ratio, and its sampling from Dirichlet distributes produces;

Generate specifically LDA generative process as follows:

1) select N, N obeys Poisson (ξ) and distributes, and N represents the vocabulary of every piece of document;

2) select θ, θ obeys Dirichlet (α) and distributes, and θ is column vector, representative be the probability that theme occurs, α is the parameter that Dirichlet distributes;

3) choosing a topic parameter z _mn, z _mnobeying Multinomial (θ) distributes;

4) select vocabulary parameter w _mn, w _mnobey Multinomial distribute, wherein

it is theme lexical item distribution matrix;

The reasoning of topic model is the reverse process that generates document, and known α and prior distribution β can write out the joint probability of various stochastic variable D, z and θ, wherein D={w according to document generative process ₁, w ₂... w _m, representing collection of document, z represents theme, w ₁, w ₂..., w _mrepresent each vocabulary, its subscript represents that this vocabulary is the how many vocabulary;

p (D, z, θ | α, β) = Π_{m = 1}^{M} [p (θ_{m} | α) Π_{n = 1}^{N} p (z_{mn} | θ_{m}) p (w_{mn} | β_{z_{mn}})]

Wherein M represents number of documents, θ _mthe theme that is m piece of writing document distributes;

Theme priori parameter θ is carried out to integration and obtain p (z, w| α, β), according to Bayesian formula,

p (z_{n} | z_{- n}, w, α, β) = \frac{p (z_{n}, z_{- n}, w | α, β)}{p (z_{- n}, w | α, β)}

Wherein, z _nrepresent the corresponding descriptor of n piece of writing vocabulary in m piece of writing document, z _-nexpression is in m piece of writing document the corresponding w descriptor of n piece of writing vocabulary, and the corresponding descriptor of vocabulary in other all known document, represents document vocabulary; Above formula can be used to the estimation of topic parameter, determines the keyword with semantic feature of its correspondence;

In given α and β situation, the joint distribution of theme priori parameter θ, theme z and every piece of document vocabulary w is expressed as:

p (w, z, θ | α, β) = p (θ | α) Π_{n = 1}^{N} p (z_{n} | θ) p (w_{n} | z_{n}, β)

Wherein, integrate θ and z, obtain the marginal distribution of a vocabulary:

p (w | α, β) = &Integral; p (θ | α) (Π_{n = 1}^{N} \underset{z_{n}}{Σ} p (z_{n} | θ) p (w_{n} | z_{n}, β)) dθ

According to marginal distribution above, obtain M piece of writing document probability distribution and be:

p (D | α, β) = Π_{m = 1}^{M} &Integral; p (θ_{m} | α) (Π_{n = 1}^{N} \underset{z_{mn}}{Σ} p (z_{mn} | θ_{m}) P (w_{mn} | z_{mn}, β)) {dθ}_{m}

Wherein, D represents M piece of writing collection of document, and m represents m piece of writing document label sequence number;

Then the posteriority that calculates hidden variable under given one piece of document condition distributes, and formula is as described below:

p (θ, z | w, α, β) = \frac{p (θ, z, w | α, β)}{p (w | α, β)}

This posteriority distribution adopts variation EM algorithm to carry out reasoning and can obtain, and gets the function that Q distributes

remove to approach q (θ, z|w, α, β), the form that wherein Q distributes is as follows:

Wherein γ represents Di Li Cray parameter,

representative polynomial parameter,

n parameter of representative polynomial;

In order to obtain best variational parameter, variation is distributed

minimize with the KL divergence between true distribution q (θ, z|w, α, β), minimum value obtains by the method for iteration; By deriving, can obtain

iterative formula with γ:

Wherein:

be illustrated in θ under γ condition _iconditional probability expectation value, Ψ is logarithm gamma function,

condition polynomial parameters, α _idi Li Cray parameter while representing the i time iteration; What renewal Dirichlet parameter alpha was used is Newton-Raphson method;

By above-mentioned reasoning, obtain the estimated value of new argument γ, wherein having carried is that each user produces the semantic feature that equilibrium criterion forms document; So just completed the process of the semantic feature of LDA extraction data;

The training of step 4. normal condition knowledge, specifically: HMM abnormality detection comprises two parts, previous part is the assessment to event, and rear portion is the study of parameter, and parameter learning is exactly Unknown Parameters, asks the problem of optimum model parameter λ; The semantic feature of the data file that the data that user is produced form is considered as the observed quantity O=O of HMM model ₁, O ₂..., O _t;

How according to the observation the 3rd problem of hidden Markov model be sequence O=O ₁, O ₂..., O _t, try to achieve model parameter or adjustment model parameter lambda, make P (O| λ) maximum; And the 3rd problem solves by Forward-backward algorithm;

Forward-backward algorithm

First define two variablees, given observation sequence O and hidden Markov model λ, definition t is positioned at hidden state S constantly _iprobability variable be: γ _t(i)=P (q _t=s _i| O, λ)

According to forward variable α _tand backward variable β (i) _t(i) definition, by upper formula forward direction, backward variable represents:

γ_{t} (i) = \frac{α_{t} (i) β_{t} (i)}{P (O | λ)} = \frac{α_{t} (i) β_{t} (i)}{Σ_{i}^{N} α_{t} (i) β_{i} (i)}

Wherein denominator is to guarantee:

Backward variable is: β _t(j)=Σ a _ijb _j(O _t+1) β _t+1(j), t=T-1, T-2 ... 1,1≤i≤N

Given observation sequence O and hidden Markov model λ, definition t is positioned at hidden state S constantly _iand t+1 is positioned at hidden state S constantly _jprobability variable be:

ξ _t(i,j)＝P(q _t＝s _i,q _t+1＝s _j|O,λ)

This variable can have forward direction-backward variable to represent simultaneously:

ξ_{t} (i, j) = \frac{α_{t} (i) a_{ij} b_{j} (O_{t + 1}) β_{t + 1} (j)}{P (O | λ)} = \frac{α_{t} (i) a_{ij} b_{j} (O_{t + 1}) β_{t + 1} (j)}{Σ_{i}^{N} Σ_{j}^{N} α_{t} (i) a_{ij} b_{j} (O_{t + 1}) β_{t + 1} (j)}

Two variablees defined above also exist following relation:

γ_{t} (i) = Σ_{j}^{N} ξ_{t} (i, j)

And

(i) represent in observation sequence O from state S _ithe transfer expected probability setting out,

represent in observation sequence from state S _itransfer to state S _jtransfer expected probability, defined two variablees and corresponding shifted expectation, a kind of parameter π that reasonably reappraises hidden Markov model (HMM), A and B method are as follows:

\overset{&OverBar;}{π_{1}} = γ_{1} (i)

When t=1, be in the expected probability of state i;

\overset{&OverBar;}{a_{ij}} = \frac{Σ_{t = 1}^{T - 1} ξ_{t} (i, j)}{Σ_{t = 1}^{T - 1} γ_{t} (i)}

Transfer expected probability from state i to state j is divided by the expected probability shifting away from state i;

\overset{&OverBar;}{b_{j}} (k) = \frac{Σ_{t = 1}^{T} γ_{t} (j) \times δ (o_{t}, v_{k})}{Σ_{t = 1}^{T} γ_{t} (j)}

At state, j observes v _kexpected probability divided by the expected probability that is in state j;

Wherein, π represents original state probability, and A represents state-transition matrix, and B represents confusion matrix,

represent the original state probability upgrading after training,

what represent to upgrade after training transfers to the probability of state j by state i, what represent to upgrade after training observes the probability of k under state j; By above-mentioned training, can obtain describing the parameter of the model of this physical training condition

Step 5. health status monitoring, specifically: this process is mainly concerned with the evaluation problem in tri-problems of HMM, takes forward direction algorithm to address this problem;

Forward direction algorithm

The definition t constantly local probability of state j is α _t(j)=P _r(observation state | hidden state j) * P _r(t constantly all sensing state j paths), for last state, its local probability has comprised by all possible path and has arrived these shape probability of states;

Special when t=1, without any the path of pointing to current state; Therefore be positioned at the probability of current state during t=1, be initial probability, i.e. P _r(state|t=1)=P (state), therefore, local probability α during t=1 ₁(i) the initial probability that equals current state is multiplied by relevant observation probability:

α ₁(i)＝π(i)b _i(o ₁)

Local probability α ' s while calculating t>1

T-1 constantly α ' s has provided this t of all arrival last path probability of state constantly, therefore, can pass through t-1 local definition of probability t constantly α ' s constantly:

α_{t + 1} (j) = b_{j} (o_{t + 1}) \overset{N}{Σ} α_{t} (i) a_{ij}

Can recursively calculate the probability of an observation sequence after given Hidden Markov Model (HMM), by t=1 local probability α ' s constantly, calculate t=2 α ' s constantly, by t=2 α ' s constantly, calculate t=3 α ' s constantly etc. until t=T; The probability of the observation sequence of given Hidden Markov Model (HMM) just equals t=T local probability sum constantly;

Use forward direction algorithm to calculate the probability of the observation sequence of T duration:

Known T duration observation sequence: O=O ₁o ₂... O _t

1) the local probability α of t=1 all states of the moment:

α ₁(j)＝π(j)b _j(o ₁)

2) at t=2 ... during T, the local probability for each state, is calculated by following formula:

α_{t + 1} (j) = b_{j} (o_{t + 1}) \overset{N}{Σ} α_{t} (i)

3) last, given HMM, the probability of observation sequence equals T all local probability sums constantly:

P_{r} (O) = \overset{N}{Σ} α_{T} (j)

User uses treadmill will produce training data at every turn, use these data to train a HMM model simultaneously, and calculate the probability that these data produce under the HMM model obtaining last time, if the data that front and back are twice produce probable value and are less than the dynamic renewal factor, retain these data and produce probable value and this HMM model, constantly repeat this process, will obtain meticulousr variation range, the variation range of dwindling can go to judge the abnormality of each use data that treadmill produces more accurately, by obtaining, level off to gradually the HMM model of normal event priori simultaneously, otherwise, there is variation in the health status that is representing user, can record this abnormal, the data of simultaneously giving up this produce probable value and HMM model, constantly repeat this process, just arrived the monitoring of training limit, limit, the object that training limit, limit is optimized.