CN114947850A - Mental load grade objective detection method based on pulse Bouss model characteristics - Google Patents

Abstract

The invention discloses a mental load grade objective detection method based on pulse Gaussian model characteristics, which comprises the following steps: collecting pulse wave signals to carry out pretreatment and waveform quality screening; decomposing the screened pulse wave signals cycle by cycle to obtain a Gaussian model parameter sequence; calculating time domain characteristics and frequency domain characteristics of the parameter sequence, and integrating the time domain characteristics and the frequency domain characteristics into sample characteristics of the whole pulse wave signal; forming a training set by using the sample characteristics of a plurality of sections of pulse wave signals, and inputting the training set into an LDA-SVM model for training; and extracting the pulse wave sample characteristics of the object to be detected, and inputting the pulse wave sample characteristics into the trained LDA-SVM model to obtain a mental load grade detection result. The invention detects the mental load grade based on the physiological characteristics of the photoelectric volume pulse wave, has simple and convenient signal acquisition mode and objective and accurate detection result, and is suitable for the mobile wearable equipment to monitor the mental load capacity of the individual in real time.

Description

Mental load grade objective detection method based on pulse Bouss model characteristics

Technical Field

The invention relates to the technical field of signal processing, in particular to an objective mental load grade detection method based on characteristics of a photoplethysmography model.

Background

Mental load is generally defined as the amount of mental or cognitive resources a person needs to provide while performing a relevant task. Currently, there are subjective scale methods and objective measurement methods for detecting mental load. The subjective scale method is mainly based on the form expansion of surveys and questionnaires, and has subjective tendency and time lag. Objective measurements are mainly classified into three major categories, neurophysiology, physiology and behaviourology. The neurophysiology is developed around electroencephalogram signals and a functional near infrared spectrum technology; the physiology mainly comprises electroencephalogram signals, electrocardiosignals, respiration, skin electricity, eye movement signals and the like; behaviours mainly indirectly detect the mental load of a person by measuring the memory capacity, reaction speed and the like of the person under multitask conditions.

The mental load objective measurement method based on physiological signal detection is mainly developed by electroencephalogram signals and electrocardiosignals. Although the electroencephalogram signals have higher accuracy, the complexity of electroencephalogram equipment means that the method is difficult to popularize for detection in daily life; compared with electroencephalogram signal acquisition, electrocardiosignals are more convenient, but because a plurality of leads are needed to be pasted near the heart, great discomfort is caused to people.

The photoplethysmography (pulse wave for short) signal, as a new physiological signal, has high homology with the electrocardio signal and also contains rich cardiovascular information, and the significant difference of the pulse wave characteristics under different cognitive load conditions is proved at present.

Disclosure of Invention

The invention aims to solve the problems of subjectivity, discomfort of a detection method and complexity of detection equipment in a mental load and mental health detection method in the prior art, and provides an objective mental load grade detection method based on pulse wave Gaussian model characteristics. The invention is based on the theory of photoplethysmography, realizes multi-level real-time classification detection on the mental load of people from the perspective of objective physiological data, and overcomes the defects that the traditional mental load objective measurement method based on physiological signal detection is inconvenient in signal acquisition and strong in acquisition discomfort.

The purpose of the invention can be achieved by adopting the following technical scheme:

a mental load grade detection method based on pulse Bouss model features comprises the following steps:

s1, acquiring pulse wave signals for preprocessing and waveform quality screening: acquiring pulse wave signals to carry out band-pass filtering processing and baseline drift removal processing, extracting a stable reference waveform from the preprocessed pulse wave signals, calculating a correlation coefficient between each period of the pulse wave signals and the reference waveform, wherein the correlation coefficient is used for evaluating waveform quality, screening out periodic waveforms interfered by motion artifacts, and finally obtaining the screened pulse wave signals;

because the acquired pulse wave signals usually have interferences such as direct current components, power frequency noise and the like, and the pulse wave signals have baseline drift due to hardware factors such as power supply ripples, emission source temperature changes and the like, which all affect the subsequent extraction of pulse wave signal characteristics, the pulse wave signals need to be subjected to band-pass filtering and baseline drift removal processing;

the band-pass filtering and baseline wandering removing process can remove some noises introduced by natural factors, but cannot remove noises introduced by weak muscle movement or human actions; therefore, waveform quality evaluation needs to be performed by calculating a correlation coefficient between each period of the pulse wave signal and a reference waveform, and whether the pulse wave signal in each period has severe deformation is detected; when the correlation coefficient of a certain period and the reference waveform is too small, which represents that the similarity between the periodic signal and the reference signal is too low, the pulse wave signal in the period can be considered to be seriously deformed due to an external force factor; all pulse wave periodic signals which are seriously deformed due to motion artifacts are screened out, so that each period can have a good form;

s2, decomposing the screened pulse wave signals cycle by cycle to obtain a Gaussian model parameter sequence, wherein the Gaussian model parameter sequence refers to the arrangement of Gaussian model parameters of a plurality of cycles extracted from the screened pulse wave signals on a time domain;

according to the Gaussian model theory of the pulse wave, the pulse wave signal of one period is formed by overlapping a plurality of Gaussian wave-shaped waveforms and a residual wave; the pulse wave Gaussian model is characterized in that the traditional pulse wave morphological characteristics are generalized, the influence of pulse wave morphological inconsistency caused by individual differences can be well overcome, and the stability of characteristic extraction is improved;

before extracting the features of each component wave, the fundamental wave decomposition of the gaussian wave needs to be performed on the pulse wave signal, and the decomposition effect can refer to fig. 2; the decomposed Gaussian waves adopt maximum fundamental wave amplitude (A) respectively ₁ A ₂ A ₃ ) Fundamental wave area (S) ₁ S ₂ S ₃ ) And the position (T) of the fundamental wave ₁ T ₂ T ₃ ) Three characteristic descriptions; because there is a certain correlation between the amplitude and the area of each fundamental wave and the position of the fundamental wave, for example, the amplitude A of the main wave ₁ At the time of enlargement, A ₂ And A ₃ Also tends to increase; therefore, in order to reduce the correlation between features to the maximum extent, the gaussian model parameters to be extracted are improved to the following 11 features:

height A of the main wave _i,1 Height ratio of the first reflected wave to the main wave

Height ratio of the second reflected wave to the main wave

Area S of pulse wave _i Ratio of main wave area

Area ratio of the first reflected wave

Second reflected wave area ratio

Pulse wave period length T _i Relative position of main wave peak

Relative position of the first reflected wave

Relative position to the second reflected wave

Wherein the area of the pulse wave is defined as S _i ＝S _i,1 +S _i,2 +S _i,3 I represents the pulse wave signal of the ith cycle;

by integrating these 11 features as a gaussian parameter vector with 11 dimensions, the periodic gaussian parameter vector is expressed as:

wherein, the T at the upper right corner of the matrix represents the transposition of the matrix, which is not repeated;

because each section of pulse wave signal has an indefinite number of cycles and the occurrence time of each cycle is sequential, the Gaussian parameter vector set extracted from each cycle of the whole section of pulse wave signal is a time-related parameter sequence; to describe this parameter sequence more canonically, the time point of each parameter vector needs to be located; since the maximum slope point is the point which is least susceptible to interference in the pulse wave signal, the method uses the position n 'of the maximum slope point of each period of the pulse wave signal' _M,i Setting the position of the parameter vector of the periodic signal; expressed by a mathematical expression:

wherein I represents the qualified cycle number of the whole pulse wave signal after screening, and delta (·) is a Kronecherdelta function n' _M,i Represents the position of the maximum slope point of the ith cycle; in particular, the pulse wave signals are each periodThe length of the period is not a fixed value, and the whole pulse wave signal may have certain periods cut off after the waveform screening, so F _u (n) is a non-uniformly sampled sequence, denoted by subscript u;

s3, calculating time domain characteristics and frequency domain characteristics of the Gaussian model parameter sequence to serve as sample characteristics of the whole pulse wave signal, wherein the time domain characteristics refer to mean values, variances, difference root-mean-square, kurtosis and skewness of the Gaussian model parameters; the frequency domain characteristics refer to normalized low-frequency components, normalized high-frequency components and the ratio of the low-frequency components to the high-frequency components of the parameters;

obtaining a Gaussian parameter sequence F _u (n), further extracting the time domain characteristic and the frequency domain characteristic of the whole section of pulse wave signal from the Gaussian parameter sequence to obtain a sample characteristic of the section of pulse wave signal; for the Gaussian parameter vector F in equation (1) _i For each feature in the method, the total number of time domain features and frequency domain features extracted by the method is 8: mean AVE, variance VAR, root mean square RMSSD, kurtosis KURT, skewness SKEW, low frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF (ii) a Therefore, for the whole gaussian parameter vector, each extracted time (frequency) domain feature has 11 dimensions, which are respectively marked as a mean AVE, a variance VAR, a difference root mean square RMSSD, a kurtosis KURT, a skewness SKEW, and a low-frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF ；

S4, forming a training set by sample characteristics of pulse wave signals acquired by a plurality of subjects under different known mental load grades, and inputting the training set into an LDA-SVM model for training;

LDA is a supervised dimension reduction algorithm, and a training set and a class label respectively correspond to each other when input is required, wherein the class label refers to the mental load grade of a certain pulse wave signal during acquisition; the core idea is to minimize the intra-class variance and maximize the inter-class mean difference, thereby being beneficial to better classification of the classifier; the SVM is the most common classifier in machine learning, and is a cutting hyperplane obtained by calculation according to input features, and the hyperplane can cut different types of features to form a decision function and realize classification detection of the features;

the LDA-SVM model is a feature dimension reduction classification model, and is formed by cascading an LDA algorithm as a preceding stage feature processing module and an SVM as a subsequent stage decision module; the LDA algorithm is responsible for feature reprocessing, a high-dimensional training set is converted into a low-dimensional training set and is input into the SVM classifier, the problem of overfitting of the classifier caused by overhigh dimension can be solved, the mean value difference among classes can be maximized, and the classification accuracy of the SVM classifier is improved;

and S5, extracting the pulse wave sample characteristics of the object to be tested, inputting the pulse wave sample characteristics into the trained LDA-SVM model, and obtaining the mental load grade output result.

Further, the process of step S1 is as follows:

s101, collecting a plurality of pulse wave signals of more than 60 seconds, and carrying out band-pass filtering on the pulse wave signals to obtain filtered signals x (n);

wherein the low-frequency cut-off frequency is 0.2Hz, and the high-frequency cut-off frequency is 5 Hz; the step is mainly used for eliminating interference such as power frequency noise and the like;

s102, extracting base point sequence x of filtered pulse wave signals x (n) _B (n), wherein the base point refers to the starting point of each period of the pulse wave signal, and the extraction process of the base point sequence is as follows:

s1021, calculating a position sequence [ n ] of the maximum slope point of the pulse wave signal _M,1 ,n _M,2 ,…，n _M,K ]Wherein n is _M,k Represents the kth maximum slope point of the whole pulse wave signal, K is the [1,2, …, K ∈]K represents the number of the maximum slope points on the whole pulse wave signal;

the specific process is as follows: solving the first order difference of the pulse wave signals to obtain all maximum value points of a first order difference sequence x' (n); removing smaller maximum value points through soft threshold screening; the reciprocal of the frequency corresponding to the maximum value of the frequency domain obtained by FFT of x (n) is recorded as

Difference of removal time interval is less than

Obtaining a new maximum point sequence, the time position of the new maximum points becomes the position of the maximum slope point, thus obtaining the position sequence [ n ] of the maximum slope point _M,1 ,n _M,2 ,…，n _M,K ]；

S1022, calculating the position sequence [ n ] of the base point _B,1 ,n _B,2 ,…，n _B,K ]Wherein n is _B,k Represents the whole pulse wave signal passing through the kth maximum slope point n _M,k The specific process of the positioned base point position is as follows:

position sequence [ n ] with maximum slope point _M,1 ,n _M,2 ,…，n _M,K ]Is used as a starting point, the first-order difference signal x '(n) of the pulse wave signal x (n) is reversely traversed to find the first closest to lambda x' (n) _M,k ) The point corresponding to the point is taken as a base point position, thereby obtaining a position sequence [ n ] of the base point _B,1 ,n _B,2 ,…，n _B,K ]Wherein λ is an adjustable coefficient and has a value range of [0, 0.1%]The experimental optimal value is 0.05; the step can generalize the position of the base point, weaken the influence of motion artifacts or pulse wave tail signals on the positioning of the base point, and enhance the robustness of the positioning of the base point;

s1023, a sequence of positions [ n ] at the base point _B,1 ,n _B,2 ,…，n _B,K ]Taking the values of x (n) to obtain the base point sequence x _B (n),n∈[n _B,1 ,n _B,2 ,…,n _B,K ]；

S103, performing baseline wander removing processing on the filtered pulse wave signals x (n): firstly, base point sequence x is aligned _B (n) performing cubic spline fitting to obtain a baseline drift curve x _base (n) subtracting the baseline shift curve x from the pulse wave signal x (n) _base (n) obtaining the pulse wave signal x after baseline removal _rb (n)；

S104, extracting the pulse wave signal x after baseline removal _rb (n), wherein the reference waveform is x _rb (n) extracting the waveform with the most stable form in one period as follows: firstly, the method is provided by step S102Sequence of base points x _B (n) for the baseline-removed pulse wave signal x _rb (n) performing periodic fine segmentation; then normalizing the amplitude of each divided pulse wave signal with a complete period; then calculating the correlation coefficients of two adjacent periods; when the pulse wave signal x _rb (n) when a certain segment continuously appears a plurality of adjacent periods and the correlation coefficient is large, selecting any one period waveform in the segment as a reference waveform; usually the threshold is above 0.9;

s105, screening out the pulse wave signal x _rb (n) poor quality periodic waveform: calculating the reference waveform and x _rb (n) correlation coefficients for other periods in the sequence; and when the correlation coefficient of the pulse wave signal in a certain period and the reference waveform is less than a certain threshold value, screening out the pulse wave signal in the period. The threshold value here is usually 0.8 or less; the cycle number of the whole section of pulse wave signals after being screened is changed from K to I.

Further, the process of step S2 is as follows:

s201, decomposing each period waveform of the screened pulse wave signals into three Gaussian waves, wherein the decomposition steps are as follows: firstly, taking a pulse wave signal of a certain period as input, and calculating a starting point and a peak point of a first Gaussian wave; taking a longitudinal axis of the position of the peak point as a symmetry axis, carrying out mirror symmetry on the waveforms from the starting point to the peak point, and extracting a first Gaussian wave; and then, the residual items left after the first Gaussian wave is subtracted from the pulse wave signal of the period are used as input to re-execute the steps, and three iterations are carried out to extract three Gaussian waves of the period: respectively a main wave, a first reflected wave and a second reflected wave; the specific calculation process is as follows:

s2011, determining the starting position n of the Gaussian wave _onset And peak point position n _peak ；

Let the starting point of the periodic signal of the ith pulse wave be the starting point of each Gaussian wave, so the starting point is at the relative position n _onset Is 0; peak point position n of Gaussian wave _peak The search method of (1) is as follows: sequentially traversing the input signal point by point, n _now Representing the position of the current traversal point; when n is _now Satisfy any one of the followingAnd (3) stopping traversing when the workpiece is processed:

condition 1: r (n) _now )＝max{r(n)}

Condition 2: r' (n) _now )＜0∧r(n _now )＞0.4·max{r(n)}

Condition 3: r' (n) _now -1)＜min{r′(n _now -1),r′(n _now )}∧r(n _now )＞0.4·max{r(n)}

At this time, n is returned _now As peak point position n _peak A value of (d); specifically, r (n) in conditions 2 and 3 _now ) The > 0.4. max { r (n) } is to avoid the problem that the discontinuity of each residual wave start segment causes the error positioning of the peak point; in a practical scenario, this condition may be modified, but it is often necessary to achieve the same goal in other ways, enhancing the reliability of peak point positioning; in the condition, the value range of n is the length of the periodic signal, r (n) is an input item, and when the main wave is extracted, r (n) is the ith pulse wave periodic signal x _i (n), extracting the residual items when the first and second reflected waves are r (n) which are respectively the pulse wave periodic signal minus the Gaussian main wave

Subtracting the residual term of the Gaussian main wave and the first reflected wave from the sum pulse wave periodic signal

r' (n) is the first order difference result of r (n);

s2012, extracting the main wave, the first reflected wave and the second reflected wave;

determining the starting position n of a Gaussian wave _onset And peak point position n _peak Then, a gaussian wave is extracted by formula (3):

wherein x _i,j (n) a j-th Gaussian wave representing the i-th pulse wave periodic signal,

extracted x representing pulse wave _i,j-1 Residual wave after (n), i.e.

In particular, it is possible to use, for example,

n _onset,j and n _peak,j Respectively representing the starting point and the peak point of the jth Gaussian wave; where j is 1,2, 3; respectively corresponding to the main wave, the first reflected wave and the second reflected wave;

s202, extracting Gaussian wave parameters of each period to form a Gaussian wave parameter vector, wherein 11 Gaussian wave parameters extracted in each period are respectively as follows: height A of the main wave _i,1 Height ratio of the first reflected wave to the main wave

Height ratio of second reflected wave to main wave, and area S of pulse wave _i Ratio of main wave area

The area ratio of the first reflected wave to the second reflected wave, and the period length T of the pulse wave _i Relative position of main wave peak

Relative position of the first reflected wave

Relative position to the second reflected wave

Wherein the area of the pulse wave is defined as S _i ＝S _i,1 +S _i,2 +S _i,3 (ii) a These 11 features are integrated as the Gaussian parameter vector of the period, and the Gaussian parameter vector F of the period _i Is as described in formula (1);

s203, constructing Gaussian model parametersSequence F _u (n)；

As mentioned above, since the maximum slope point is the point of the pulse wave signal which is least susceptible to interference, the method sets the position of the maximum slope point of each period of the pulse wave signal as the position of the parameter vector of the period signal; because the maximum slope point is positioned before the waveform quality screening, the length N of the maximum slope point sequence is larger than the number K of waveforms passing the screening; therefore, the position of the maximum slope point n needs to be eliminated first _M,1 ,n _M,2 ,…,n _M,K ]In step S105, the points corresponding to the periods are screened out, so as to obtain a new maximum slope point position sequence [ n' _M,1 ,n′ _M,2 ,…,n′ _M,I ]Constructing and obtaining a Gaussian model parameter sequence F based on a formula (2) as the time of a corresponding periodic Gaussian wave parameter vector _u (n)。

Further, the step S3 process is as follows:

s301, calculating a Gaussian model parameter sequence F _u (n), wherein the time domain features comprise a mean AVE, a variance VAR, a difference root mean square RMSSD, a kurtosis KURT, a skewness SKEW; the calculation formula is as follows:

s302, calculating a Gaussian model parameter sequence F _u (n) frequency domain characteristics, since Fast Fourier Transform (FFT) requires a uniform sequence for power spectrum transformationSampling sequence, therefore, it is necessary to first align the Gaussian model parameter sequence F _u (n) performing interpolation resampling for one time; wherein the frequency domain features include a low frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF The calculation process is as follows:

s3021, Gaussian model parameter sequence F _u (n) carrying out interpolation resampling to obtain a uniform sampling sequence F (n); the resampling frequency is 2 Hz;

s3022, performing FFT (fast Fourier transform) on the uniform sampling sequence F (n) to obtain a power spectrum sequence P (jf);

s3023, calculating the power spectrum sequence P (jf) in a segmented manner to obtain the total power P _total Ultra low frequency power P _VLF Low frequency power P _LF High frequency power P _HF (ii) a Wherein, the frequency domain is divided into a super low frequency band (VLF), a low frequency band (LF) and a high frequency band (HF), and the frequency ranges are respectively: 0 to 0.04Hz, 0.04 to 0.15Hz, 0.15 to 0.4 Hz; then accumulating the whole power spectrum to obtain the total power P _total Accumulating the ultra-low frequency band to obtain ultra-low frequency power P _VLF And so on to obtain low-frequency power P _LF High frequency power P _HF ；

S3024, based on the total power P _total Ultra low frequency power P _VLF Low frequency power P _LF High frequency power P _HF Calculating the low frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF (ii) a The calculation formula is as follows:

and S303, integrating the time domain characteristics and the frequency domain characteristics to form sample characteristics Z of the whole pulse wave signal. The integration is shown in formula (11):

Z＝[AVE ^T VAR ^T RMSSD ^T KURT ^T SKEW ^T LF _nu ^T HF _nu ^T R _LF/HF ^T ] ^T formula (11)

Wherein, each time (frequency) domain feature has 11 dimensions, and the whole sample feature has 5 time domain features and 3 frequency domain features, so that the sample feature of one pulse wave signal sample has 88 dimensions, i.e. the sample feature is an 88 × 1 matrix.

Further, the step S4 process is as follows:

s401, collecting r sections of pulse wave signals under the known mental load state, and respectively recording the sample characteristics of each section of pulse wave signal obtained in the step S3 as Z ₁ 、Z ₂ 、…、Z _r Construction model training set [ Z ] ₁ ,Z ₂ ,…,Z _r ]；

The training set is a common concept in machine learning and is a basic resource for a classifier learning decision; generally, the higher the quality of the training set, the greater the assistance to the detection; the scientificity and accuracy of a decision function of the classifier are directly determined by the quality of the training set; in general, the number of samples in the training set should not be too low, i.e. the value of r should be large enough;

s402, training set [ Z ] ₁ ,Z ₂ ,…，Z _r ]Standardized to obtain [ Z' ₁ ,Z′ ₂ ,…，Z′ _r ]Wherein Z' ₁ 、Z′ ₂ 、…、Z′ _r Respectively representing the sample characteristics of each pulse wave signal after the normalization processing;

the purpose of training set normalization is to de-dimensionalize; because 88 dimensional features of the sample features have different dimensions and magnitudes, the weight influence of the difference of the dimensions and magnitudes of the features on the features can be reduced after standardization, and meanwhile, the complexity of SVM for computing the hyperplane is also reduced;

the normalized calculation procedure is as follows:

wherein Z is _l Sample feature, Z 'of the first sample in the set of samples' _l Mu is a mean vector of the sample characteristics and sigma is a standard deviation vector of the sample characteristics;

s403, training set [ Z 'after standardization' ₁ ,Z′ ₂ ,…，Z′ _r ]LDA dimension reduction processing is carried out to obtain projection matrixes omega and [ Z ₁ ,Z″ ₂ ,…，Z″ _r ]Wherein Z ″) ₁ 、Z″ ₂ 、…、Z″ _r Respectively representing the sample characteristics of each section of pulse wave signal subjected to LDA dimension reduction processing;

the LDA algorithm is a supervised dimension reduction algorithm, and a training set and a class label are respectively corresponding when input is required, wherein the class label refers to the mental load grade of a certain pulse wave signal during acquisition; the idea is to minimize the intra-class variance and maximize the inter-class mean difference; the calculation process is to use the interspecies divergence matrix S of the training set _b Maximum and intra-class divergence matrix S _ω Minimizing, computing the matrix

The largest d eigenvalues and the corresponding d eigenvectors form a matrix [ omega ] ₁ ω ₂ ω ₃ … ω _d ]Marked as ω; multiplying the original sample feature training set by a projection matrix to realize the transformation of a feature space, and obtaining a low-dimensional training set; in particular, d represents the dimensionality after dimensionality reduction, and generally, the dimensionality after dimensionality reduction is not higher than the number of sample label types, namely the dimensionality after dimensionality reduction is not higher than the number of divided mental load grade series;

s404, inputting the low-dimensional training set [ Z "", after dimension reduction ₁ ,Z″ ₂ ,…，Z″ _r ]Training an SVM model;

each sample in the training set needs to be provided with a sample label, namely each sample is marked as a sample in a certain mental load grade state;

the SVM classifier has two basic parameters: a penalty coefficient C and a kernel function coefficient gamma; the two parameters have very important influence on the classification effect of the SVM; the size of C has a critical influence on the classification effect of the SVM, the larger the C is, the overfitting is easy to cause, and the smaller the C is, the under-fitting is easy to cause; the gamma coefficient implicitly determines the distribution of the data after mapping to a new feature space, and the gamma coefficient is usually used as an auxiliary adjustment parameter of C to jointly optimize the fitting effect of the SVM; therefore, optimization of SVM parameters is a necessary step; the penalty coefficient C and the kernel function coefficient gamma are optimized by finding the optimal parameters of the SVM through a grid search algorithm and cross validation based on the training test results of the training set.

Further, the step S5 process is as follows:

s501, collecting pulse waves of the object to be detected, and processing according to the steps S1-S3 to obtain characteristics of the sample to be detected;

s502, standardizing the characteristics of the sample to be tested based on the mean value mu and the standard deviation sigma of the sample of the training set;

s503, performing dimension reduction processing on the standardized characteristics of the sample to be detected based on the projection matrix omega;

s504, inputting the characteristics of the sample to be tested after dimensionality reduction into the trained SVM classifier to obtain a mental load grade output result.

Compared with the prior art, the invention has the following advantages and effects:

1) the method for positioning the pulse wave base point based on the maximum slope point position (see step S102) positions the position of the pulse wave base point through the most stable point (maximum slope point) of the pulse wave form, and compared with the past method for positioning the base point based on the minimum value of the pulse wave, the base point positioning method provided by the invention can still realize the good positioning effect of the base point when the tail form of the pulse wave is distorted, and has higher robustness;

2) the invention provides a mental load detection mode which is realized based on Gaussian model characteristics of pulse waves (see step S2), the pulse wave characteristics extracted based on the mode generalize the morphological characteristics of the traditional pulse waves to a certain extent, and the problem of difficult extraction of the pulse wave characteristics caused by individual differences can be well solved; meanwhile, the features extracted based on the Gaussian model not only can acquire the traditional morphological features, but also can extract the features which cannot be obtained by calculation in part of the traditional morphological ways;

3) according to the method, the time domain characteristics and the frequency domain characteristics of the pulse wave Gaussian model parameters are extracted through the step S3, the time window length of characteristic extraction is 60 seconds, and compared with an extraction mode based on pulse rate variability and heart rate variability characteristics, the time window length is at least shortened by more than 120 seconds; therefore, the method for detecting the mental load has higher time resolution and is more sensitive to the change response of the mental load.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the invention without limiting the invention. In the drawings:

fig. 1 is a flowchart of a method for detecting mental load level based on features of a photoplethysmography model in embodiment 1 of the present invention;

FIG. 2 is a graph showing the effect of periodic waveform screening in steps S104-S105 in example 1 of the present invention;

fig. 3 is a flowchart of extracting the position of the pulse bogaussian wave peak in step S201 in embodiment 1 of the present invention;

fig. 4 is a diagram illustrating the effect of pulse wave gaussian wave extraction in step S201 according to embodiment 1 of the present invention;

FIG. 5 is a diagram of identification of pulse Boussian wave features extracted in example 1 of the present invention;

FIG. 6 is a flow chart of the training and detection of the LDA-SVM model in embodiment 1 of the present invention;

fig. 7 is a bar graph of the accuracy of the trained LDA-SVM model in example 1 of the present invention in the test.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

Example 1

The embodiment discloses a mental load level detection method based on photoplethysmography model characteristics, and the steps are shown in fig. 1. The operation flow is as follows:

s1, collecting pulse wave signals to carry out preprocessing and waveform quality screening; the pulse wave signals acquired in the embodiment have interference such as direct current components, power frequency noise and the like, and simultaneously have a baseline drift phenomenon, which all affect the extraction of the subsequent pulse wave signal characteristics, so that the pulse wave signals need to be subjected to band-pass filtering and baseline drift removal processing; the specific operation steps are as follows:

s101, collecting a plurality of pulse wave signals of more than 60 seconds, and carrying out band-pass filtering on the pulse wave signals to obtain filtered signals x (n); wherein the low-frequency cut-off frequency is 0.2Hz, and the high-frequency cut-off frequency is 5 Hz; the band-pass filtering filters the interference such as direct current information, power frequency noise and the like;

s102, extracting base point sequence x of the filtered pulse wave signals x (n) _B (n), the extraction step of the base point sequence is as follows:

s1021, calculating a position sequence [ n ] of the maximum slope point of the pulse wave signal _M,1 ,n _M,2 ,…，n _M,K ]Wherein n is _M,k Represents the kth maximum slope point of the whole pulse wave signal, K is the [1,2, …, K ∈]K represents the number of the maximum slope points on the whole section of pulse wave signal;

the reciprocal of the frequency corresponding to the maximum value of the frequency domain of the pulse wave signal obtained by FFT conversion of the whole pulse wave signal is recorded as

Making a first order difference on the pulse wave signalDividing to obtain x '(n), processing the first-order difference signal x' (n) by using find _ peaks function in python scipy library, and setting function parameter distance as

Determining the maximum slope point sequence of the pulse wave signal so as to obtain the position sequence [ n ] of the maximum slope point _M,1 ,n _M,2 ,…，n _M,K ]；

S1022, a base point position sequence [ n ] is obtained _B,1 ,n _B,2 ,…，n _B,K ](ii) a Wherein n is _B,k Represents the whole pulse wave signal passing through the kth maximum slope point n _M,k The specific process of the positioned base point is as follows:

sequence [ n ] in maximum slope point position _M,1 ,n _M,2 ,…，n _M,K ]Is used as a starting point, the first-order difference signal x '(n) of x (n) is reversely traversed to find the first closest to lambda x' (n) _M,k ) The point corresponding to the point is determined as a base point position, thereby obtaining a base point position sequence [ n ] _B,1 ,n _B,2 ,…，n _B,K ](ii) a Wherein λ is set to 0.05;

s1023, base point position sequence [ n ] _B,1 ,n _B,2 ,…，n _B,K ]Obtaining the base point sequence x by taking the values of x (n) _B (n),n∈[n _B,1 ,n _B,2 ,…，n _B,K ]；

S103, performing baseline wander removing processing on the filtered pulse wave signals x (n);

firstly, base point sequence x is aligned _B (n) performing cubic spline fitting to obtain a baseline drift curve x _base (n), subtracting the baseline drift curve x from the pulse wave signals x (n) _base (n) obtaining the pulse wave signal x after baseline removal _rb (n)；

The preprocessing step can only primarily filter noise irrelevant to pulse wave characteristics, and cannot solve the problem of influence of involuntary actions of human muscles on pulse wave signals during acquisition, so that the pulse wave periods with serious deformation caused by motion artifacts need to be screened; the screening steps are as follows:

s104, extracting and removing basePost-line pulse wave signal x _rb The reference waveform of (n);

carrying out period segmentation on the pulse wave signals by using the base points determined in the step S102, and carrying out amplitude normalization processing on the segmented complete periodic signals one by one; and carrying out interpolation resampling on each normalized pulse wave periodic signal to obtain a sequence with the length L being 100 so as to ensure that the correlation coefficient can be calculated in the following way:

wherein y (n) and z (n) represent two pulse wave period sequences; when the correlation coefficient of 20 continuous adjacent periods is greater than 0.99, the pulse wave height is stable, the collected object keeps a good detection state, the waveform of any period can be selected as a reference waveform, and the last waveform of the pulse wave band is selected as the reference waveform in the embodiment;

s105, screening out x _rb (n) a periodic waveform of poor quality;

calculating a correlation coefficient between the reference waveform and the rest periods of the pulse wave signal according to a formula (13), wherein; when the correlation coefficient of a certain periodic signal and a reference waveform is smaller than a certain threshold, the waveform is considered to be damaged due to motion artifacts, and the periodic signal is screened out; in the present embodiment, the threshold is set to 0.8; FIG. 2 shows the differences before and after the waveforms are eliminated in this example; wherein, the waveform of the black solid line is the reference waveform selected in the step S103, the gray shaded part is all good normalized periodic waveforms of the 60 second pulse wave signals, the black dotted line is the screened interfered periodic pulse wave, and it can be seen that the interfered pulse wave signals do not have the original pulse wave contour;

s2, decomposing the screened pulse wave signals cycle by cycle to obtain a Gaussian model parameter sequence; the method comprises the following specific steps:

s201, decomposing each period waveform of the screened pulse wave signals into three Gaussian waves;

the decomposition steps are as follows: firstly, taking a pulse wave signal of a certain period as input, and calculating a starting point and a peak point of a first Gaussian wave; taking a longitudinal axis of the position of the peak point as a symmetry axis, carrying out mirror symmetry on the waveform from the starting point to the peak point, and extracting a first Gaussian wave; and then, the residual items left after the first Gaussian wave is subtracted from the pulse wave signal of the period are used as input to re-execute the steps, and three iterations are carried out to extract three Gaussian waves of the period: respectively as a main wave, a first reflected wave and a second reflected wave; the specific calculation process is as follows:

s2011, determining the starting position n of the Gaussian wave _onset And peak point position n _peak ；

Let the starting point of the periodic signal of the ith pulse wave be the starting point of each Gaussian wave, so the starting point is at the relative position n _onset Is 0; peak point position n of Gaussian wave _peak The search method of (1) is as follows: sequentially traversing the input signal point by point, n _now Representing the position of the current traversal point; when n is _now Stopping traversing when any one of the following conditions is met:

condition 1: r (n) _now )＝max{r(n)}

Condition 2: r' (n) _now )＜0∧r(n _now )＞0.4·max{r(n)}

Condition 3: r' (n) _now -1)＜min{r′(n _now -1),r′(n _now )}∧r(n _now )＞0.4·max{r(n)}

At this time, n is returned _now As peak point position n _peak A value of (d); in the condition, the value range of n is the length of the periodic signal, r (n) is an input item, and when the main wave is extracted, r (n) is the ith pulse wave periodic signal x _i (n), extracting the residual items when the first and second reflected waves are r (n) which are respectively the pulse wave periodic signal minus the Gaussian main wave

Subtracting the residual term of the Gaussian main wave and the first reflected wave from the sum pulse wave periodic signal

r' (n) is the first order of r (n)Differentiating the result; the program flow chart of the present embodiment is shown in fig. 3;

s2012, extracting the main wave, the first reflected wave and the second reflected wave;

determining the starting position n of a Gaussian wave _onset And peak point position n _peak Then, extracting the expression x of the Gaussian wave by the formula (3) _i,1 (n)，x _i,2 (n)，x _i,3 (n); the gaussian wave extraction process of this example is shown in fig. 4; wherein, A is the main wave and the residual wave extracted in the first decomposition; b is a first reflected wave and a residual wave extracted at the time of second decomposition, wherein the residual wave of A is input; graph C is the second reflected wave and the remaining residual wave extracted at the third decomposition, the input being the residual wave of graph B; FIG. D is the relative position of the decomposed three wave shapes in the pulse wave;

s202, extracting Gaussian wave parameters of each period to form a Gaussian wave parameter vector;

after the decomposition is completed, the height A of the dominant wave is extracted _i,1 Height ratio of the first reflected wave to the main wave

Height ratio of second reflected wave to main wave, and area S of pulse wave _i Ratio of main wave area

The area ratio of the first reflected wave to the second reflected wave, and the period length T of the pulse wave _i Relative position of main wave peak

Relative position of the first reflected wave

Relative position to the second reflected wave

Wherein the area of the pulse wave is defined as S _i ＝S _i,1 +S _i,2 +S _i,3 (ii) a The identity of the individual features represented in the pulse wave cycle is shown in FIG. 5; these 11 features are integrated as a gaussian parameter vector with 11 dimensions, the parameter vector F of the period _i Is as described in formula (1);

s203, constructing a Gaussian model parameter sequence F _u (n)；

Because the maximum slope point is positioned before the waveform quality screening, the length K of the maximum slope point sequence is surely larger than the waveform quantity I passing the screening; therefore, the maximum slope point position sequence [ n ] needs to be eliminated first _M,1 ,n _M,2 ,…，n _M,K ]The points corresponding to the period are screened out, so that a new maximum slope point position sequence [ n' _M,1 ,n′ _M,2 ,…，n′ _M,I ]Constructing and obtaining a Gaussian model parameter sequence F based on a formula (2) as the time of a corresponding periodic Gaussian wave parameter vector _u (n)；

S3 calculation parameter sequence F _u (n) the time domain feature and the frequency domain feature are used as the sample feature of the whole pulse wave signal; the method comprises the following specific steps:

s301, calculating a Gaussian model parameter sequence F _u (n) a time domain feature;

the method comprises the steps of averaging AVE, variance VAR, difference root mean square RMSSD, kurtosis KURT and skewness SKEW, and the calculation process refers to formulas (4) to (7);

s302, calculating a Gaussian model parameter sequence F _u (n) frequency domain characteristics;

s3021, parameter pair sequence F _u (n) carrying out interpolation resampling, wherein the resampling frequency is 2Hz, and obtaining a parameter sequence F (n) of uniform sampling;

s3022, performing FFT (fast Fourier transform) on the F (n) to convert the F (n) into a power spectrum sequence P (jf), and dividing a frequency domain into an ultra-low frequency band (VLF), a low frequency band (LF) and a high frequency band (HF), wherein the frequency ranges are 0-0.04 Hz, 0.04-0.15 Hz and 0.15-0.4 Hz respectively;

s3023, accumulating the whole power spectrum to obtain the total power P _total Accumulating the ultra-low frequency band to obtain ultra-low frequency power P _VLF And so on to obtain low-frequency power P _LF High frequency power P _HF ；

S3024, based on the total power P _total Ultra low frequency power P _VLF Low frequency power P _LF High frequency power P _HF Calculating the low frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF The calculation mode refers to the formula (8) to the formula (10);

s303, integrating the time domain characteristics and the frequency domain characteristics according to a formula (11) to obtain sample characteristics Z of the whole pulse wave signal;

s4, forming a training set by sample characteristics of pulse wave signals acquired by a plurality of subjects under different known mental load grade states, and inputting the training set into an LDA-SVM model for training;

s401, collecting r sections of pulse wave signals under the known mental load state, and respectively recording the sample characteristics of each section of pulse wave signal obtained in the step S3 as Z ₁ 、Z ₂ 、…、Z _r Construction model training set [ Z ] ₁ ,Z ₂ ,…,Z _r ]；

In this embodiment, 660 pulse wave sample data from 22 different individuals are collected, the sample duration of each pulse wave is 60 seconds, and the data sample distribution is shown in table 1:

TABLE 1 sample data distribution Table of example 1

The 1-Back state and the 2-Back state in the acquisition state are two memory experiments with different difficulties of a backtracking experiment; the memory experiment mainly aims at enabling a subject to be in different mental load states, wherein the 2-Back experiment is more difficult than the 1-Back experiment, and the corresponding mental load is larger; according to the embodiment, subjective feelings of the testee are recorded after a task according to two objective indexes of task difficulty and task score of the testee, and the states of the testee are divided into a low-load state, a medium-load state and a high-load state; wherein, the rest state is a low load state correspondingly, the 1-Back state is a medium load state correspondingly, and the 2-Back state is a high load state correspondingly;

the detection model of the embodiment is a two-level classification model, namely the model has two levels for detecting the mental load; in order to test the detection effect of different types of classification grades, the method respectively constructs three sample groups for the samples in a pairwise mixed manner, namely a medium-high load sample group, a low-high load sample group and a medium-low load sample group, wherein each sample group comprises 440 samples; respectively taking the three sample groups as input of the model to carry out three different mental load level detection experiments on the model; the implementation steps of the three sample groups are completely consistent, and the subsequent steps are only described by taking one sample group as an example;

in order to test the accuracy of the model detection so as to better explain the characteristics of the method, the embodiment performs the following steps on the sample according to 4: 1 into a training set and a test set, so that the training set comprises 352 samples, the test set comprises 88 samples, and the mental load levels of the training set and the test set are known; the mental load level of the training set and the characteristics of the training set are input into the model together for training the model; during testing, only inputting the characteristics of the test set into the trained model without inputting the corresponding mental load grade, inputting the predicted value of the mental load grade by the model according to a decision function, and comparing the predicted value with the actual mental load level to obtain the detection accuracy of the model; the specific training and detection process is shown in fig. 6; the specific operation flow of this embodiment is as follows:

s402, training set [ Z ] ₁ ,Z ₂ ,…，Z _r ]Standardized to obtain [ Z' ₁ ,Z′ ₂ ,…，Z′ _r ]Wherein Z' ₁ 、Z′ ₂ 、…、Z′ _r Respectively representing the sample characteristics of each pulse wave signal after the normalization processing; the normalization processing mode refers to formula (12);

s403, training set [ Z 'after standardization' ₁ ,Z′ ₂ ,…，Z′ _r ]LDA dimension reduction processing is carried out to obtain projection matrixes omega and [ Z ₁ ,Z″ ₂ ,…，Z″ _r ]Wherein Z ″) ₁ 、Z″ ₂ 、…、Z″ _r Respectively represent the components after LDA dimension reduction treatmentSample characteristics of each section of pulse wave signal;

the LDA algorithm adopts an API in a scinit-spare library in python, and sets an API parameter n _ components to be 1; namely, the LDA algorithm reduces the dimension of the 88-dimension characteristics; in the embodiment, the mental load grades of the three sample groups are all 2 types, and the dimensionality after dimension reduction is not higher than the divided mental load grade series, so that original 88-dimensional sample characteristics are converted into 1-dimensional sample characteristics after LDA dimension reduction, and the calculation complexity of the SVM is greatly reduced; particularly, LDA is a supervised dimension reduction algorithm, so that a sample feature set and a sample label need to be input together during input, that is, each training sample of a training set is provided with a label of mental load grade;

s404, inputting the low-dimensional training set [ Z "", after dimension reduction ₁ ,Z″ ₂ ,…，Z″ _r ]Training an SVM model;

inputting the low-dimensional training set into an SVM classifier adopting a Gaussian kernel for training; optimizing a penalty coefficient C and a kernel function coefficient gamma through a grid search algorithm and 5-fold cross validation to maximize the classification accuracy, thereby obtaining a decision function;

completing training of the detection model of the mental load grade; the next step is actually the use and testing of the method;

s5, inputting the test set into the trained LDA-SVM model to obtain a corresponding mental load grade output result, and comparing the mental load grade output result with the mental load grade corresponding to the known test set; since the test set has already been processed in the foregoing part, the present embodiment skips step S501, and the rest of the process is as follows:

s502, standardizing the test set based on the mean value mu and the standard deviation sigma of the training set sample, and removing the dimensional influence of the test data;

s503, performing feature dimension reduction transformation based on the projection matrix omega on the standardized test set, projecting the normalized test set to a new feature space capable of realizing optimal classification, and obtaining a dimension-reduced 1-dimensional vector test set;

s504, inputting the feature of the test set subjected to dimensionality reduction into the trained SVM classifier to obtain a mental load grade output result, wherein the mental load grade output result is obtained by carrying out grade evaluation on input data based on the existing mental load grade; finally, the evaluation result of the test set is compared with the actual result, and the accuracy is shown in fig. 7.

When the brain load grade is detected to be high load or low load, the accuracy of 96% can be achieved; when detecting the mental load grade as medium load and low load, the accuracy rate of 90 percent can be achieved by the embodiment. In other words, the proposed method for detecting the mental load grade based on the photoplethysmography has high detection accuracy of the mental load, and the average rate is 93%. When the mental load level is detected to be high load or medium load, the accuracy rate of the embodiment is 75%. By combining the above test results, the present embodiment can reflect that the detection accuracy of the proposed detection method for medium and high loads is higher, and the distinguishing accuracy for medium and high loads is slightly lower, but this is in accordance with the demand scenario of real life: for most people, whether the people are in a medium-high load state or not is often required to know, and once the people enter the medium-high load state, the mental state is not optimal at present, and appropriate rest adjustment is required.

Example 2:

the difference between this embodiment and embodiment 1 is mainly that the resolution of brain load level is improved from two levels to three levels, that is, the detection results of the model are: high load, medium load and low load. Therefore, the training and testing of the model were performed using the samples of the three mental stress levels in example 1 as a sample group. The method comprises the following specific steps:

s1, refer to the corresponding steps in embodiment 1, which are not described herein again;

s2, refer to the corresponding steps in embodiment 1, which are not described herein again;

s3, refer to the corresponding steps in embodiment 1, which are not described herein again;

s4, forming a training set by sample characteristics of a plurality of sections of pulse wave signals acquired under the known mental load level state, and inputting the training set into an LDA-SVM model for training;

after the feature extraction is finished, firstly training an LDA-SVM model, obtaining optimal parameters and a decision plane by training an LDA-SVM classifier, and then detecting and classifying the input features of the mental load grade to be detected; in this embodiment, the characteristic samples in embodiment 1 are used as a sample group without sample separation, and therefore the sample set includes 660 samples;

consistent with example 1, in order to test the accuracy of the model for detecting the mental load level of three levels, the present embodiment performs the feature set detection according to 4: the scale of 1 is divided into two parts, namely a training set and a test set, so that the training set comprises 528 samples, the test set comprises 132 samples, and the mental load levels of the training set and the test set are known;

s402, training set [ Z ] ₁ ,Z ₂ ,…，Z _r ]Standardized to obtain [ Z' ₁ ,Z′ ₂ ,…，Z′ _r ]Wherein Z' ₁ 、Z′ ₂ 、…、Z′ _r Respectively representing the sample characteristics of each pulse wave signal after the normalization processing;

s403, training set [ Z 'after standardization' ₁ ,Z′ ₂ ,…，Z′ _r ]LDA dimension reduction processing is carried out to obtain projection matrixes omega and [ Z ₁ ,Z″ ₂ ,…，Z″ _r ]Wherein Z ″) ₁ 、Z″ ₂ 、…、Z″ _r Respectively representing the sample characteristics of each section of pulse wave signal subjected to LDA dimension reduction processing;

the LDA algorithm adopts an API in a scinit-spare library in python, and sets an API parameter n _ components to be 2, so that the LDA algorithm can reduce the dimension of 88-dimensional sample features; in the embodiment, the mental load grades of the sample groups are 3 types, and the dimensionality after dimensionality reduction is not higher than the dimensionality of the divided mental load grades, so that the mental load grades are converted into sample characteristics with 2 dimensionalities after the dimensionality reduction by LDA;

s404, inputting the low-dimensional training set [ Z "", after dimension reduction ₁ ,Z″ ₂ ,…，Z″ _r ]Training an SVM model;

inputting the training set into an SVM classifier adopting a Gaussian kernel for training; optimizing a penalty coefficient C and a kernel function coefficient gamma through a grid search algorithm and 5-fold cross validation to maximize the classification accuracy, so as to obtain a decision function; particularly, the optimization results of C and gamma are different when different characteristics and different grade classifications are used, namely, scenes are inconsistent; after training is finished, entering a testing step;

s502, standardizing the test set based on the mean value mu and the standard deviation sigma of the training set sample, and removing the dimensional influence of the test data;

s503, performing feature dimension reduction transformation based on a projection matrix omega on the standardized test set, projecting the normalized test set to a new feature space capable of realizing optimal classification, and obtaining a 2-dimensional vector test set after dimension reduction;

and S504, inputting the transformed test set characteristics into the trained SVM classifier to obtain a mental load grade output result.

The results show that: according to the test result, when the resolution of the mental load grade is three grades, the accuracy of the method is 73%, and compared with 33% of the random prediction probability, the accuracy of the method is improved by 40%.

The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims (6)

1. An objective detection method for mental load grade based on pulse Bouss model features is used for objectively detecting the state of mental load of a human body, and is characterized by comprising the following steps:

s1, acquiring pulse wave signals for preprocessing and waveform quality screening: acquiring pulse wave signals, carrying out band-pass filtering processing and baseline drift removal processing, extracting a stable reference waveform from the preprocessed pulse wave signals, calculating a correlation coefficient between each period of the pulse wave signals and the reference waveform, wherein the correlation coefficient is used for evaluating waveform quality, screening out periodic waveforms interfered by motion artifacts, and finally obtaining the screened pulse wave signals;

s2, decomposing the screened pulse wave signals cycle by cycle to obtain a Gaussian model parameter sequence, wherein the Gaussian model parameter sequence refers to the arrangement of Gaussian model parameters of a plurality of cycles extracted from the screened pulse wave signals on a time domain;

s3, calculating time domain characteristics and frequency domain characteristics of the Gaussian model parameter sequence to be used as sample characteristics of the whole pulse wave signal, wherein the time domain characteristics refer to the mean value, the variance, the root mean square of the difference value, the kurtosis and the skewness of the Gaussian model parameters; the frequency domain characteristics refer to normalized low-frequency components, normalized high-frequency components and the ratio of the low-frequency components to the high-frequency components of the parameters;

s4, forming a training set by sample characteristics of pulse wave signals acquired by a plurality of subjects under different known mental load grades, and inputting the training set into an LDA-SVM model for training;

and S5, extracting the pulse wave sample characteristics of the object to be tested, inputting the pulse wave sample characteristics into the trained LDA-SVM model, and obtaining the mental load grade output result.

2. The method for objectively detecting mental load based on pulse gaussian model characteristics according to claim 1, wherein in step S1, the process of collecting pulse wave signals for preprocessing and waveform quality screening is as follows:

s101, collecting a plurality of pulse wave signals of more than 60 seconds, and carrying out band-pass filtering on the pulse wave signals to obtain filtered signals x (n);

s102, extracting base point sequence x of filtered pulse wave signals x (n) _B (n), wherein the base point refers to the starting point of each period of the pulse wave signal, and the extraction process of the base point sequence is as follows:

s1021, calculating a position sequence [ n ] of the maximum slope point of the pulse wave signal _M,1 ,n _M,2 ,…，n _M,K ]Wherein n is _M,k Represents the kth maximum slope point of the whole pulse wave signal, K is the [1,2, …, K ∈]K represents the number of the maximum slope points on the whole pulse wave signal;

s1022, calculating the position sequence [ n ] of the base point _B,1 ,n _B,2 ,…，n _B,K ]Wherein n is _B,k Represents the whole pulse wave signal passing through the kth maximum slope point n _M,k The base point position is positioned; position sequence [ n ] with maximum slope point _M,1 ,n _M,2 ,…，n _M,K ]Reversely traverse the first order difference signal x '(n) of the pulse wave signal x (n) with each value as the starting point to find the first closest to λ · x' (n) _M,k ) The point corresponding to which the position is the base point position n _B,k Thereby obtaining the position sequence [ n ] of the base point _B,1 ,n _B,2 ,…，n _B,K ]λ is an adjustable coefficient;

s1023, a sequence of positions [ n ] at the base point _B,1 ,n _B,2 ,…，n _B,K ]Taking the values of x (n) to obtain the base point sequence x _B (n),n∈[n _B,1 ,n _B,2 ,…,n _B,K ]；

S103, performing baseline wander removing processing on the filtered pulse wave signals x (n): firstly, base point sequence x is aligned _B (n) performing cubic spline fitting to obtain a baseline drift curve x _base (n), subtracting the baseline drift curve x from the pulse wave signals x (n) _base (n) obtaining the pulse wave signal x after baseline removal _rb (n)；

S104, extracting the pulse wave signal x after baseline removal _rb (n), wherein the reference waveform is x _rb (n) extracting the waveform with the most stable form in one period as follows: the base point sequence x extracted in step S102 is used first _B (n) for the baseline-removed pulse wave signal x _rb (n) performing periodic fine segmentation; then normalizing the amplitude of each divided pulse wave signal with a complete period; then calculating the correlation coefficients of two adjacent periods; when the pulse wave signal x _rb (n) when a certain segment continuously appears a plurality of adjacent periods and the correlation coefficient is large, selecting any one period waveform in the segment as a reference waveform;

s105, screening out the pulse wave signal x _rb (n) poor quality periodic waveform: calculating the reference waveform and x _rb (n) correlation coefficients for other periods in the sequence; when the correlation coefficient between the pulse wave signal of a certain period and the reference waveform is less than a certain threshold valueAnd screening out the pulse wave signals of the period.

3. The method for objectively detecting a mental load level based on pulse gaussian model characteristics according to claim 2, wherein in step S2, the filtered pulse wave signals are decomposed cycle by cycle to obtain a gaussian model parameter sequence, which comprises the following steps:

s201, decomposing each period waveform of the screened pulse wave signals into three Gaussian waves, wherein the decomposition steps are as follows: firstly, taking a pulse wave signal of a certain period as input, and calculating a starting point and a peak point of a first Gaussian wave; taking a longitudinal axis of the position of the peak point as a symmetry axis, carrying out mirror symmetry on the waveform from the starting point to the peak point, and extracting a first Gaussian wave; and then, the residual items left after the first Gaussian wave is subtracted from the pulse wave signal of the period are used as input to re-execute the steps, and three iterations are carried out to extract three Gaussian waves of the period: respectively as a main wave, a first reflected wave and a second reflected wave;

s202, extracting Gaussian wave parameters of each period to form a Gaussian wave parameter vector, wherein the number of the Gaussian wave parameters extracted in each period is 11, the number of the Gaussian wave parameters is respectively the height of a main wave, the height ratio of a first reflected wave to the main wave, the height ratio of a second reflected wave to the main wave, the area of a pulse wave, the area ratio of the main wave, the area ratio of the first reflected wave, the area ratio of the second reflected wave, the period length of the pulse wave, the relative position of the peak value of the main wave, the relative position of the first reflected wave and the relative position of the second reflected wave, and the Gaussian wave parameters extracted in each period form the Gaussian wave parameter vector of the period;

s203, constructing a Gaussian model parameter sequence F _u (n) the construction steps are: culling maximum slope point position sequence [ n ] _M,1 ,n _M,2 ,…，n _M,K ]In step S105, the points corresponding to the period are screened out to obtain a new maximum slope point position sequence [ n' _M,1 ,n′ _M,2 ,…，n′ _M,I ]Representing the time of the corresponding periodic Gaussian wave parameter vector to obtain a Gaussian model parameter sequence F _u (n), the subscript u indicates that the sequence is a non-uniformly sampled sequence.

4. The method for objectively detecting a mental load level based on pulse gaussian model characteristics according to claim 1, wherein in step S3, the time domain characteristics and the frequency domain characteristics of the gaussian model parameter sequence are calculated as the sample characteristics of the whole pulse wave signal, and the process is as follows:

s301, calculating a Gaussian model parameter sequence F _u (n), wherein the time domain features comprise a mean AVE, a variance VAR, a difference root mean square RMSSD, a kurtosis KURT, a skewness SKEW;

s302, calculating a Gaussian model parameter sequence F _u (n), wherein the frequency domain characteristics include a low frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF The calculation process is as follows:

s3021, Gaussian model parameter sequence F _u (n) carrying out interpolation resampling to obtain a uniform sampling sequence F (n);

s3022, performing FFT (fast Fourier transform) on the uniform sampling sequence F (n) to obtain a power spectrum sequence P (jf);

s3023, calculating the power spectrum sequence P (jf) in a segmented manner to obtain the total power P _total Ultra low frequency power P _VLF Low frequency power P _LF High frequency power P _HF ；

S3024, based on the total power P _total Ultra low frequency power P _VLF Low frequency power P _LF High frequency power P _HF Calculating the low frequency normalized power LF _nu High frequency normalized power HF _nu And low high frequency power ratio R _LF/HF ；

And S303, integrating the time domain characteristic and the frequency domain characteristic to be used as a sample characteristic Z of the whole pulse wave signal.

5. The method for objectively detecting the mental load level based on the pulse gaussian model characteristics of claim 1, wherein in step S4, the sample characteristics of the pulse wave signals collected by a plurality of subjects under different known mental load levels are combined into a training set, and input into the LDA-SVM model for training, the process is as follows:

s401, collecting the sample characteristics of r sections of pulse wave signals, and respectively recording the sample characteristics of each section of pulse wave signal obtained in the step S3 as Z ₁ 、Z ₂ 、…、Z _r Construction model training set [ Z ] ₁ ,Z ₂ ,…,Z _r ]；

S402, training set [ Z ] ₁ ,Z ₂ ,…，Z _r ]Standardized to obtain [ Z' ₁ ,Z′ ₂ ,…，Z′ _r ]Wherein Z' ₁ 、Z′ ₂ 、…、Z′ _r Respectively representing the sample characteristics of each pulse wave signal after the normalization processing;

s403, training set [ Z 'after standardization' ₁ ,Z′ ₂ ,…，Z′ _r ]LDA dimension reduction processing is carried out to obtain projection matrixes omega and [ Z ₁ ,Z″ ₂ ,…，Z″ _r ]Wherein Z ″) ₁ 、Z″ ₂ 、…、Z″ _r Respectively representing the sample characteristics of each section of pulse wave signal subjected to LDA dimension reduction processing;

s404, inputting the low-dimensional training set [ Z "", after dimension reduction ₁ ,Z″ ₂ ,…，Z″ _r ]And training the SVM model.

6. The method for objectively detecting the mental load level based on the pulse gaussian model characteristics of claim 1, wherein in the step S5, the pulse wave sample characteristics of the object to be detected are extracted and input into the trained LDA-SVM model to obtain the mental load level output result, and the process is as follows:

s501, collecting pulse waves of the object to be detected, and processing according to the steps S1-S3 to obtain characteristics of the sample to be detected;

s502, standardizing the characteristics of the sample to be tested based on the mean value mu and the standard deviation sigma of the sample of the training set;

s503, performing dimension reduction processing on the standardized characteristics of the sample to be detected based on the projection matrix omega;

s504, inputting the characteristics of the sample to be tested after dimensionality reduction into the trained SVM classifier to obtain a mental load grade output result.

Images