Objective detection method of mental workload level based on pulse wave Gaussian model characteristics

Abstract

The present invention discloses an objective detection method for mental workload level based on pulse wave Gaussian model characteristics, and the steps are as follows: collecting pulse wave signals for preprocessing and waveform quality screening; decomposing the screened pulse wave signals cycle by cycle to obtain Gaussian model parameter sequences; calculating the time domain characteristics and frequency domain characteristics of the parameter sequences, and integrating them into sample characteristics of the entire pulse wave signal; forming a training set with sample characteristics of several segments of pulse wave signals, and inputting them into an LDA-SVM model for training; extracting the pulse wave sample characteristics of the object to be tested, and inputting them into the trained LDA-SVM model to obtain the mental workload level detection result. The present invention detects the mental workload level based on the physiological characteristics of the photoelectric volumetric pulse wave, and the signal acquisition method is simple, and the detection results are objective and accurate, which is suitable for mobile wearable devices to monitor personal mental workload capacity in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to an objective detection method for mental workload level based on photoplethysmography Gaussian model characteristics. Background Art

[0002] Mental workload is usually defined as the amount of mental resources or the proportion of cognitive resources that people need to provide when performing related tasks. At present, there are two methods for detecting mental workload: subjective scale method and objective measurement method. The subjective scale method is mainly based on surveys and questionnaires, and there are subjective tendencies and time lags. Objective measurement methods are mainly divided into three categories: neurophysiology, physiology, and behavior. Neurophysiology revolves around EEG signals and functional near-infrared spectroscopy technology; physiology mainly includes EEG signals, ECG signals, respiration, skin electricity, and eye movement signals; behavior mainly indirectly detects people's mental workload by measuring their memory ability and reaction speed under multi-tasking conditions.

[0003] Objective measurement methods for mental workload based on physiological signal detection are mainly based on EEG signals and ECG signals. Although EEG signals have high accuracy, the complexity of EEG equipment means that this method is difficult to promote to daily life detection; ECG signals are more convenient to collect than EEG signals, but because they require multiple leads to be attached near the heart, they cause greater discomfort.

[0004] Photoplethysmography (Pulse Wave) signal is an emerging physiological signal that is highly homologous to ECG signal and contains rich cardiovascular information. It has been proven that there are significant differences in the characteristics of pulse waves under different cognitive load conditions. Summary of the invention

[0005] The purpose of the present invention is to solve the subjectivity problem, discomfort problem and complexity problem of detection equipment in the existing technology of mental workload and mental health detection methods, and provide an objective detection method of mental workload level based on the characteristics of pulse wave Gaussian model. The present invention is based on the photoplethysmography theory, and realizes multi-level real-time classification detection of human mental workload from the perspective of objective physiological data, overcoming the shortcomings of the existing objective measurement method of mental workload based on physiological signal detection, such as inconvenient signal collection and strong discomfort in collection.

[0006] The purpose of the present invention can be achieved by adopting the following technical solutions:

[0007] A method for detecting mental workload level based on pulse wave Gaussian model characteristics, the method comprising the following steps:

[0008] S1. Collect pulse wave signals for preprocessing and waveform quality screening: collect pulse wave signals for bandpass filtering and baseline drift removal, extract a stable reference waveform from the preprocessed pulse wave signals, and then calculate the correlation coefficient between each cycle of the pulse wave signal and the reference waveform. The correlation coefficient is used to evaluate the waveform quality, screen out periodic waveforms interfered by motion artifacts, and finally obtain the screened pulse wave signal;

[0009] Since the collected pulse wave signal usually contains interference such as DC components and power frequency noise, and hardware factors such as power ripple and temperature change of the transmitting source will cause the pulse wave signal to have baseline drift, which will affect the subsequent extraction of pulse wave signal features. Therefore, it is necessary to perform bandpass filtering and baseline drift removal on the pulse wave signal.

[0010] Bandpass filtering and baseline drift removal can remove some noise introduced by natural factors, but they are powerless against noise introduced by weak muscle movement or human action. Therefore, it is necessary to evaluate the waveform quality by calculating the correlation coefficient between each cycle of the pulse wave signal and the reference waveform, and detect whether the pulse wave signal of each cycle has severe deformation. When the correlation coefficient between a certain cycle and the reference waveform is too small, it means that the similarity between the cycle signal and the reference signal is too low, and it can be considered that the pulse wave signal of the cycle has been severely deformed due to external factors. By screening out all pulse wave cycle signals that have been severely deformed due to motion artifacts, it is ensured that each cycle has a good shape.

[0011] S2, decomposing the screened pulse wave signal 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 multiple cycles extracted from the screened pulse wave signal in the time domain;

[0012] According to the Gaussian model theory of pulse wave, a pulse wave signal of one cycle is formed by the superposition of multiple waveforms similar to Gaussian waves and a residual wave; the characteristic of the pulse wave Gaussian model is that it generalizes the traditional pulse wave morphological characteristics, can well overcome the influence of individual differences on the inconsistency of pulse wave morphology, and improve the stability of feature extraction;

[0013] Before extracting the characteristics of each component wave, it is necessary to first decompose the pulse wave signal into the fundamental wave of the Gaussian wave. The decomposition effect can be referred to Figure 2The decomposed Gaussian wave is described by three features: the maximum fundamental amplitude (A1A2A3), the fundamental area (S1S2S3) and the fundamental position (T1T2T3). Since there is a certain correlation between the amplitude and area of ​​each fundamental wave and the position of the fundamental wave, for example, when the amplitude A1 of the main wave increases, A2 and A3 will also tend to increase. Therefore, in order to minimize the correlation between features, the Gaussian model parameters to be extracted are improved to the following 11 features:

[0014] The height of the main wave is A i,1 , the height ratio of the first reflected wave to the main wave Height ratio of the second reflected wave to the main wave The area of ​​the pulse wave S i , Main wave area ratio First reflected wave area ratio Second reflected wave area ratio Pulse wave cycle length T i , relative position of the main wave peak Relative position of the first reflected wave Relative position of the second reflected wave 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 i-th cycle;

[0015] Integrating these 11 features into a Gaussian parameter vector with 11 dimensions, the Gaussian parameter vector of this period is expressed as:

[0016]

[0017] The T in the upper right corner of the matrix represents the transpose of the matrix, which will not be described in detail later;

[0018] Since each pulse wave signal has an indefinite number of cycles and each cycle appears in a certain order, the set of Gaussian parameter vectors extracted from each cycle of the entire pulse wave signal is a time-related parameter sequence; in order to more formally describe this parameter sequence, it is necessary to locate the time point of each parameter vector; since the maximum slope point is the point in the pulse wave signal that is least susceptible to interference, this method locates the position n′ of the maximum slope point of each cycle of the pulse wave signal M,i Set to the position of the periodic signal parameter vector; expressed mathematically as follows:

[0019]

[0020] Where I represents the number of qualified cycles of the whole pulse wave signal after screening, δ(·) is the Kronecherdelta function, and n′ M,i Indicates the position of the maximum slope point of the ith cycle; in particular, since the length of each cycle of the pulse wave signal is not a fixed value, and the entire pulse wave signal may be truncated in some cycles after waveform screening, F u (n) is a non-uniformly sampled sequence, labeled with the subscript u;

[0021] S3, calculating the time domain features and frequency domain features of the Gaussian model parameter sequence as the sample features of the entire pulse wave signal, wherein the time domain features refer to the mean, variance, root mean square of the difference, kurtosis and skewness of the Gaussian model parameters; the frequency domain features refer to the normalized low frequency component, normalized high frequency component and the ratio of the low frequency component to the high frequency component of the parameters;

[0022] Get the Gaussian parameter sequence F u (n), the Gaussian parameter sequence is further extracted to obtain the time domain features and frequency domain features of the entire pulse wave signal as the sample features of the pulse wave signal; for the Gaussian parameter vector F in formula (1), i For each feature in the time domain, this method extracts a total of 8 time domain features and frequency domain features: mean AVE, variance VAR, root mean square difference 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 Therefore, for the entire Gaussian parameter vector, each time (frequency) domain feature extracted has 11 dimensions, which are recorded as mean AVE, variance VAR, root mean square difference 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 ;

[0023] S4, forming a training set with sample features of pulse wave signals collected from several subjects under known different mental workload levels, and inputting the samples into the LDA-SVM model for training;

[0024] LDA is a supervised dimensionality reduction algorithm that requires the input to have corresponding training sets and category labels, where the category label refers to the mental workload level of a certain pulse wave signal when it is collected; its core idea is to minimize the intra-class variance and maximize the inter-class mean difference, which helps the classifier to better classify; SVM is the most commonly used classifier in machine learning. It calculates the cutting hyperplane based on the input features. The hyperplane can cut features of different categories, thereby forming a decision function and realizing the classification detection of features;

[0025] The LDA-SVM model is a feature dimensionality reduction classification model, which is composed of the LDA algorithm as the front-end feature processing module and the SVM as the back-end decision module. The LDA algorithm is responsible for feature reprocessing, converting the high-dimensional training set into a low-dimensional training set and inputting it into the SVM classifier. It can not only solve the classifier overfitting problem caused by too high a dimension, but also maximize the mean difference between classes and improve the classification accuracy of the SVM classifier.

[0026] S5. Extract the pulse wave sample features of the subject to be tested, input them into the trained LDA-SVM model, and obtain the output result of the mental workload level.

[0027] Furthermore, the process of step S1 is as follows:

[0028] S101, collecting a number of pulse wave signals of more than 60 seconds, and performing bandpass filtering on the pulse wave signals to obtain a filtered signal x(n);

[0029] The low frequency cut-off frequency is 0.2Hz, and the high frequency cut-off frequency is 5Hz; this step is mainly used to eliminate interference such as power frequency noise;

[0030] S102, extracting the base point sequence x of the filtered pulse wave signal x(n) B (n), wherein the base point refers to the starting point of each cycle of the pulse wave signal, and the extraction process of the base point sequence is as follows:

[0031] S1021, calculate the maximum slope point position sequence of the pulse wave signal [n M,1 ,n M,2 ,…,n M,K ], where n M,k represents the kth maximum slope point of the entire pulse wave signal, k∈[1,2,…,K], K represents the number of maximum slope points on the entire pulse wave signal;

[0032] The specific process is as follows: calculate the first-order difference of the pulse wave signal to obtain all the maximum points of the first-order difference sequence x′(n); remove the smaller maximum points by soft threshold screening; perform FFT transformation on x(n) to obtain the reciprocal of the frequency corresponding to the maximum value in the frequency domain, recorded as Remove time interval difference less than The maximum point of the new maximum point sequence is obtained. The time position of these new maximum points becomes the position of the maximum slope point, so the maximum slope point position sequence [n M,1 ,n M,2 ,…,n M,K ];

[0033] S1022, find the position sequence of the base point [n B,1,n B,2 ,…,n B,K ], where n B,k Indicates that the entire pulse wave signal passes through the kth maximum slope point n M,k The specific process of locating the base point is as follows:

[0034] The maximum slope point position sequence [n M,1 ,n M,2 ,…,n M,K ] is used as the starting point to reversely traverse the first-order differential signal x′(n) of the pulse wave signal x(n) to find the first value closest to λ·x′(n M,k ) point, and the position corresponding to this point is the base point position, so as to obtain the base point position sequence [n B,1 ,n B,2 ,…,n B,K ], where λ is an adjustable coefficient with a value range of [0, 0.1], and the optimal value in the experiment is 0.05; this step can generalize the base point position, weaken the influence of motion artifacts or the tail signal of the pulse wave on the base point positioning, and enhance the robustness of the base point positioning;

[0035] S1023, in the base point position sequence [n B,1 ,n B,2 ,…,n B,K ] to obtain the base point sequence x by taking the value of x(n) B (n),n∈[n B,1 ,n B,2 ,…,n B,K ];

[0036] S103, perform baseline drift removal processing on the filtered pulse wave signal x(n): first perform baseline drift removal processing on the base point sequence x B (n) Perform cubic spline fitting to obtain the baseline drift curve x base (n), then subtract the baseline drift curve x from the pulse wave signal x(n) base (n) Get the pulse wave signal x after removing the baseline rb (n);

[0037] S104, extracting the pulse wave signal x after removing the baseline rb (n) a reference waveform, wherein the reference waveform is x rb The most stable periodic waveform in (n) is extracted as follows: first, the base point sequence x extracted in step S102 is used to extract B (n) For the pulse wave signal x after removing the baseline rb (n) Perform fine cycle segmentation; then normalize the amplitude of each pulse wave signal with a complete cycle; then calculate the correlation coefficient of two adjacent cycles; when the pulse wave signal x rb(n) When a segment has a large correlation coefficient for multiple consecutive cycles, any cycle waveform in this segment is selected as the reference waveform; the threshold is usually set to be above 0.9;

[0038] S105, filter out pulse wave signal x rb (n) Poor quality periodic waveform: Calculate the reference waveform and x rb (n) The correlation coefficient of other cycles in the pulse wave signal; when the correlation coefficient of a pulse wave signal of a certain cycle with the reference waveform is less than a certain threshold, the pulse wave signal of this cycle is screened out. The threshold here is usually taken as less than 0.8; after screening, the number of cycles of the entire pulse wave signal changes from K to I.

[0039] Furthermore, the process of step S2 is as follows:

[0040] S201, decomposing each cycle waveform of the filtered pulse wave signal into three Gaussian waves, the decomposition steps are: firstly taking a pulse wave signal of a certain cycle as input, calculating the starting point and peak point of the first Gaussian wave; taking the vertical axis at the location of the peak point as the symmetry axis, mirroring the waveform from the starting point to the peak point, and extracting the first Gaussian wave; then taking the pulse wave signal of the cycle minus the first Gaussian wave as input, and re-executing the above steps, iterating three times to extract the three Gaussian waves of the cycle: the main wave, the first reflected wave, and the second reflected wave; the specific calculation process is as follows:

[0041] S2011, determine the starting position n of the Gaussian wave onset and the peak point position n peak ;

[0042] Assume that the starting point of the periodic signal of the i-th pulse wave is the starting point of each Gaussian wave, so the relative position of the starting point is n onset is 0; the peak point position of the Gaussian wave is n peak The search method is as follows: traverse the input signal point by point in sequence, n now Indicates the position of the current traversal point; when n now The traversal stops when any of the following conditions are met:

[0043] Condition 1: r(n now )=max{r(n)}

[0044] Condition 2: r′(n now )<0∧r(n now )>0.4·max{r(n)}

[0045] Condition 3: r′(n now -1)<min{r′(n now -1),r′(n now )}∧r(nnow )>0.4·max{r(n)}

[0046] At this time, return n now As the peak point position n peak The value of r(n now )>0.4·max{r(n)} is to avoid the problem of erroneous peak point location caused by the discontinuity of the starting segment of each residual wave; in actual scenarios, this condition can be modified, but it is usually necessary to achieve the same purpose in other ways to enhance the reliability of peak point location; in the condition, the value range of n is the length of the periodic signal, r(n) is the input item, and when extracting the main wave, r(n) is the i-th pulse wave periodic signal x i (n), when extracting the first and second reflected waves, r(n) is the residual term after subtracting the Gaussian main wave from the pulse wave periodic signal The residual term after subtracting the Gaussian main wave and the first reflected wave from the pulse wave period signal r′(n) is the first-order difference result of r(n);

[0047] S2012, extracting the main wave, the first reflected wave and the second reflected wave;

[0048] Determine the starting position n of the Gaussian wave onset and the peak point position n peak Then, the Gaussian wave is extracted by formula (3):

[0049]

[0050] where x i,j (n) represents the jth Gaussian wave of the i-th pulse wave cycle signal, The pulse wave is extracted i,j-1 The residual wave after (n), that is, In particular, n onset,j and n peak,j They represent the starting point and peak point of the jth Gaussian wave respectively; here j = 1, 2, 3; corresponding to the main wave, the first reflected wave and the second reflected wave respectively;

[0051] S202, extract the Gaussian wave parameters of each period to form a Gaussian wave parameter vector, wherein there are 11 Gaussian wave parameters extracted in each period, namely: the height A of the main wave i,1 , the height ratio of the first reflected wave to the main wave The height ratio of the second reflected wave to the main wave, the area S of the pulse wave i , Main wave area ratio The first reflected wave area ratio, the second reflected wave area ratio, and the pulse wave cycle length T i, relative position of the main wave peak Relative position of the first reflected wave Relative position of the second reflected wave The area of ​​the pulse wave is defined as S i =S i,1 +S i,2 +S i,3 ; Integrate these 11 features as the Gaussian parameter vector of this period, then the Gaussian parameter vector F of this period i As described in formula (1);

[0052] S203, constructing Gaussian model parameter sequence F u (n);

[0053] As mentioned above, since the maximum slope point is the point in the pulse wave signal that is least susceptible to interference, this 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 periodic signal; since the location of the maximum slope point is before the waveform quality screening, the length N of the maximum slope point sequence must be greater than the number K of waveforms that pass the screening; therefore, it is necessary to first eliminate the maximum slope point position [n M,1 ,n M,2 ,…,n M,K ], the points corresponding to the periods that are screened out in step S105, thereby obtaining a new maximum slope point position sequence [n′ M,1 ,n′ M,2 ,…,n′ M,I ], as the time of the corresponding periodic Gaussian wave parameter vector, the Gaussian model parameter sequence F is constructed based on formula (2) u (n).

[0054] Furthermore, the process of step S3 is as follows:

[0055] S301, calculate Gaussian model parameter sequence F u (n) time domain features, wherein the time domain features include mean AVE, variance VAR, root mean square difference RMSSD, kurtosis KURT, and skewness SKEW; the calculation formula is as follows:

[0056]

[0057]

[0058]

[0059]

[0060] S302, calculate Gaussian model parameter sequence F u(n) frequency domain characteristics, since the fast Fourier transform (FFT) requires the sequence to be a uniformly sampled sequence when performing power spectrum transformation, it is necessary to first u (n) performing an interpolation resampling; wherein the frequency domain features include 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:

[0061] S3021, Gaussian model parameter sequence F u (n) interpolation resampling is performed to obtain a uniform sampling sequence F(n); the resampling frequency is 2 Hz;

[0062] S3022, performing FFT transformation on the uniform sampling sequence F(n) to obtain a power spectrum sequence P(jf);

[0063] S3023, calculate the power spectrum sequence P(jf) in sections to obtain the total power P total , ultra-low frequency power P VLF , low frequency power P LF , high frequency power P HF ; The frequency domain is first divided into very low frequency band (VLF), low frequency band (LF) and high frequency band (HF), with frequency ranges of 0-0.04Hz, 0.04-0.15Hz, and 0.15-0.4Hz respectively; then the total power P is obtained by accumulating the entire power spectrum total , the ultra-low frequency band is accumulated to obtain the ultra-low frequency power P VLF , and so on to get the low-frequency power P LF , high frequency power P HF ;

[0064] S3024, based on total power P total , ultra-low frequency power P VLF , low frequency power P LF , high frequency power P HF Get the low frequency normalized power LF nu , high frequency normalized power HF nu and low high frequency power ratio R LF / HF ; The calculation formula is as follows:

[0065]

[0066]

[0067]

[0068] S303, integrating the time domain features and the frequency domain features as the sample features Z of the entire pulse wave signal. The integration method is shown in formula (11):

[0069] Z=[AVE T VAR T RMSSD T KURT T SKEW T LF nu T HF nu T R LF / HF T ] T Formula (11)

[0070] Among them, each time (frequency) domain feature has 11 dimensions, and the entire sample feature has 5 time domain features and 3 frequency domain features. Therefore, the sample feature of a pulse wave signal sample has 88 dimensions, that is, the sample feature is an 88×1 matrix.

[0071] Furthermore, the process of step S4 is as follows:

[0072] S401, collecting r segments of pulse wave signals under a known mental workload, and the sample features of each segment of pulse wave signal obtained through step S3 are recorded as Z1, Z2, ..., Z r , construct the model training set [Z1,Z2,…,Z r ];

[0073] The training set is a common concept in machine learning and is the basic resource for classifier learning and decision-making. Generally, the higher the quality of the training set, the more helpful it is for detection. The quality of the training set directly determines the scientificity and accuracy of the classifier's decision function. Generally, the number of samples in the training set should not be too low, that is, the value of r should be large enough.

[0074] S402, for the training set [Z1, Z2, ..., Z r ] is standardized to obtain [Z′1, Z′2,…, Z′ r ]where Z′1, Z′2,…,Z′ r They respectively represent the sample characteristics of each segment of the pulse wave signal after standardization;

[0075] The purpose of training set standardization is to de-dimensionalize. Since the 88 dimensional features of sample features have different dimensions and magnitudes, standardization can reduce the impact of different feature dimensions and magnitudes on feature weights, and also reduce the complexity of SVM hyperplane calculation.

[0076] The standardized calculation process is as follows:

[0077]

[0078] Among them, Z l is the sample feature of the lth sample in the sample set, Z′ l is the standardized sample feature, μ is the mean vector of the sample feature, and σ is the standard deviation vector of the sample feature;

[0079] S403, the standardized training set [Z′1, Z′2, …, Z′ r ] perform LDA dimensionality reduction to obtain the projection matrix ω and [Z″1,Z″2,…,Z″ r ], where Z″1, Z″2,…,Z″ r They respectively represent the sample features of each segment of the pulse wave signal after LDA dimensionality reduction processing;

[0080] The LDA algorithm is a supervised dimensionality reduction algorithm that requires that the training set and the category label correspond to each other when input. The category label refers to the mental workload level of a certain pulse wave signal when it is collected. Its idea is to minimize the intra-class variance and maximize the inter-class mean difference. Its calculation process is to convert the inter-class scatter matrix S of the training set into b Maximize the intra-class scatter matrix S ω Minimize, calculate the matrix The largest d eigenvalues ​​and the corresponding d eigenvectors form the matrix [ω1 ω2ω3 … ω d ], denoted as ω; the original sample feature training set is multiplied by the projection matrix to achieve the transformation of the feature space, that is, a low-dimensional training set is obtained; in particular, d represents the dimension after dimensionality reduction. In general, the dimension after dimensionality reduction is not higher than the number of sample label types, that is, the dimension after dimensionality reduction is not higher than the number of divided mental workload levels;

[0081] S404, input the low-dimensional training set [Z″1, Z″2,…, Z″ r ]Train the SVM model;

[0082] Each sample in the training set needs to have a sample label, that is, each sample is marked as a sample under a certain mental workload level;

[0083] The SVM classifier has two basic parameters: the penalty coefficient C and the kernel function coefficient gamma; these two parameters have a very important influence on the classification effect of SVM; the size of C has a key impact on the classification effect of SVM. The larger C is, the easier it is to cause overfitting, and if C is too small, the SVM will be underfitting; the gamma coefficient implicitly determines the distribution of the data after mapping to the new feature space. The gamma coefficient is usually used as an auxiliary adjustment parameter of C to jointly optimize the fitting effect of SVM; therefore, optimizing the SVM parameters is a necessary step; the optimization method of the penalty coefficient C and the kernel function coefficient gamma is usually based on the training and test results of the training set, and the optimal parameters of the SVM are obtained through a grid search algorithm and cross-validation.

[0084] Furthermore, the process of step S5 is as follows:

[0085] S501, collecting the pulse wave of the object to be tested, and processing according to steps S1 to S3 to obtain the characteristics of the sample to be tested;

[0086] S502, standardizing the characteristics of the sample to be tested based on the mean μ and standard deviation σ of the training set samples;

[0087] S503, performing dimensionality reduction processing on the standardized features of the sample to be tested based on the projection matrix ω;

[0088] S504, input the features of the sample to be tested after dimension reduction into the trained SVM classifier to obtain the output result of the mental workload level.

[0089] Compared with the prior art, the present invention has the following advantages and effects:

[0090] 1) The method for locating the pulse wave base point based on the position of the maximum slope point proposed in the present invention (see step S102) is to locate the position of the pulse wave base point through the most stable point (maximum slope point) in the pulse wave morphology. Compared with the previous method for locating the base point based on the minimum value of the pulse wave, the base point locating method proposed in the present invention can still achieve a good base point positioning effect when the tail morphology of the pulse wave is distorted, and has higher robustness;

[0091] 2) The present invention proposes a method for detecting mental workload based on the Gaussian model features of the pulse wave (see step S2). The pulse wave features extracted based on this method generalize the traditional pulse wave morphological features to a certain extent, and can effectively overcome the problem of difficulty in extracting pulse wave features caused by individual differences; at the same time, the features extracted based on the Gaussian model can not only obtain traditional morphological features, but also extract some features that cannot be calculated by traditional morphological methods;

[0092] 3) The present invention extracts the time domain features and frequency domain features of the pulse wave Gaussian model parameters through step S3, and the time window length of the feature extraction is 60 seconds. Compared with the extraction method based on pulse rate variability and heart rate variability features, the time window length is shortened by at least 120 seconds; therefore, the method for detecting mental workload proposed by the present invention has higher time resolution and is more sensitive to changes in mental workload. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0094] Figure 1 It is a flow chart of a method for detecting mental workload level based on the photoplethysmographic Gaussian model characteristics in Embodiment 1 of the present invention;

[0095] Figure 2 This is a diagram showing the effect of periodic waveform screening in steps S104-S105 in Example 1 of the present invention;

[0096] Figure 3 is a flow chart of extracting the peak position of the pulse wave Gaussian wave in step S201 in embodiment 1 of the present invention;

[0097] Figure 4 This is a diagram showing the effect of extracting the pulse wave Gaussian wave in step S201 in Example 1 of the present invention;

[0098] Figure 5 is a logo diagram of the pulse wave Gaussian wave feature extracted in Example 1 of the present invention;

[0099] Figure 6 is a training and detection flow chart of the LDA-SVM model in Example 1 of the present invention;

[0100] Figure 7 It is a bar chart of the accuracy of the trained LDA-SVM model in the test in Example 1 of the present invention. DETAILED DESCRIPTION

[0101] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0102] Example 1

[0103] This embodiment discloses a method for detecting mental workload level based on the characteristics of the photoplethysmogram Gaussian model, the steps of which are as follows: Figure 1 The operation process is as follows:

[0104] S1. Collect pulse wave signals for preprocessing and waveform quality screening. The pulse wave signals collected in this embodiment have interferences such as DC components and power frequency noise, and there is also a baseline drift phenomenon, which affects the subsequent extraction of pulse wave signal features. Therefore, it is necessary to perform bandpass filtering and baseline drift removal on the pulse wave signals. The specific operation steps are as follows:

[0105] S101, collecting a number of pulse wave signals of more than 60 seconds, and performing bandpass filtering on the pulse wave signals to obtain a filtered signal x(n); wherein the low-frequency cutoff frequency is 0.2 Hz, and the high-frequency cutoff frequency is 5 Hz; the bandpass filtering filters out interference such as DC information and power frequency noise;

[0106] S102, extracting the base point sequence x of the filtered pulse wave signal x(n) B (n), the steps for extracting the base point sequence are as follows:

[0107] S1021, calculate the maximum slope point position sequence of the pulse wave signal [n M,1 ,n M,2 ,…,n M,K ], where n M,k represents the kth maximum slope point of the entire pulse wave signal, k∈[1,2,…,K], K represents the number of maximum slope points on the entire pulse wave signal;

[0108] Perform FFT transformation on the entire pulse wave signal to obtain the inverse of the frequency corresponding to the maximum value of the pulse wave signal in the frequency domain, recorded as The pulse wave signal is first-order differentiated to obtain x′(n). The first-order differential signal x′(n) is processed using the find_peaks function in the Python scipy library. The function parameter distance is set to Determine the maximum slope point sequence of the pulse wave signal, thereby obtaining the maximum slope point position sequence [n M,1 ,n M,2 ,…,n M,K ];

[0109] S1022, find the position sequence of the base point [n B,1 ,n B,2 ,…,n B,K ]; where n B,k Indicates that the entire pulse wave signal passes through the kth maximum slope point n M,k The specific process of locating the base point is as follows:

[0110] The maximum slope point position sequence [n M,1 ,n M,2 ,…,n M,K ] is used as the starting point to reversely traverse the first-order differential signal x′(n) of x(n) to find the first value closest to λ·x′(n M,k ), and the position corresponding to the point is set as the base point position, so as to obtain the base point position sequence [n B,1 ,n B,2 ,…,n B,K ]; where λ is set to 0.05;

[0111] S1023, the base point position sequence [n B,1 ,n B,2 ,…,n B,K ] Take the value of x(n) to get the base point sequence x B (n),n∈[n B,1 ,n B,2 ,…,n B,K ];

[0112] S103, performing baseline drift removal processing on the filtered pulse wave signal x(n);

[0113] First, the base point sequence x B (n) Perform cubic spline fitting to obtain the baseline drift curve x base (n), then subtract the baseline drift curve x from the pulse wave signal x(n) base (n) Get the pulse wave signal x after removing the baseline rb (n);

[0114] The preprocessing step can only preliminarily filter out noise that is irrelevant to the pulse wave characteristics. The preprocessing step cannot solve the impact of the unconscious movement of human muscles on the pulse wave signal during acquisition. Therefore, it is necessary to filter out the pulse wave cycles that are severely deformed due to motion artifacts; the screening steps are as follows:

[0115] S104, extracting the pulse wave signal x after removing the baseline rb (n) the reference waveform;

[0116] The pulse wave signal is segmented into periods using the base points determined in step S102, and the amplitudes of the segmented complete periodic signals are normalized one by one; each normalized pulse wave periodic signal is interpolated and resampled into a sequence of length L=100 to ensure that the correlation coefficient can be calculated in the following manner:

[0117]

[0118] Wherein, y(n) and z(n) represent two pulse wave cycle sequences; when the correlation coefficient of 20 consecutive cycles is greater than 0.99, the pulse wave is highly stable and the acquisition object maintains a good detection state. The waveform of any cycle can be selected as the reference waveform. In this embodiment, the last waveform of this pulse wave band is selected as the reference waveform;

[0119] S105, filter out x rb (n) The periodic waveform is of poor quality;

[0120] The correlation coefficient between the reference waveform and the remaining cycles of the pulse wave signal is calculated according to formula (13), wherein: when the correlation coefficient between a certain periodic signal and the reference waveform is less than a certain threshold, it is considered that the waveform is destroyed by motion artifacts, and the periodic signal is screened out; in this embodiment, the threshold is set to 0.8; Figure 2 The figure shows the difference between the waveforms before and after the waveforms are screened out in this embodiment; the black solid line waveform is the reference waveform selected in step S103, the gray shadow part is all the good normalized periodic waveforms of the 60-second pulse wave signal, and the black dotted line is the filtered out disturbed periodic pulse wave. It can be seen that the disturbed pulse wave signal no longer has the outline of the original pulse wave;

[0121] S2. Decompose the filtered pulse wave signal cycle by cycle to obtain a Gaussian model parameter sequence. The specific steps are as follows:

[0122] S201, decomposing each cycle waveform of the filtered pulse wave signal into three Gaussian waves;

[0123] The decomposition steps are as follows: first, take the pulse wave signal of a certain cycle as input, calculate the starting point and peak point of the first Gaussian wave; take the vertical axis where the peak point is located as the symmetry axis, mirror the waveform from the starting point to the peak point, and extract the first Gaussian wave; then take the pulse wave signal of the cycle minus the residual term remaining after the first Gaussian wave is subtracted as input and re-execute the above steps. After three iterations, the three Gaussian waves of the cycle can be extracted: the main wave, the first reflected wave, and the second reflected wave; the specific calculation process is as follows:

[0124] S2011, determine the starting position n of the Gaussian wave onset and the peak point position n peak ;

[0125] Assume that the starting point of the periodic signal of the i-th pulse wave is the starting point of each Gaussian wave, so the relative position of the starting point is n onset is 0; the peak point position of the Gaussian wave is n peak The search method is as follows: traverse the input signal point by point in sequence, n now Indicates the position of the current traversal point; when n now The traversal stops when any of the following conditions are met:

[0126] Condition 1: r(n now )=max{r(n)}

[0127] Condition 2: r′(n now )<0∧r(n now )>0.4·max{r(n)}

[0128] Condition 3: r′(n now -1)<min{r′(n now -1),r′(n now )}∧r(n now )>0.4·max{r(n)}

[0129] At this time, return n now As the peak point position n peak In the condition, the value range of n is the length of the periodic signal, r(n) is the input item, and when extracting the main wave, r(n) is the i-th pulse wave periodic signal x i (n), when extracting the first and second reflected waves, r(n) is the residual term after subtracting the Gaussian main wave from the pulse wave periodic signal The residual term after subtracting the Gaussian main wave and the first reflected wave from the pulse wave period signal r′(n) is the first-order difference result of r(n); the program flow chart of this embodiment is as follows Figure 3 ;

[0130] S2012, extracting the main wave, the first reflected wave and the second reflected wave;

[0131] Determine the starting position n of the Gaussian wave onset and the peak point position n peak Then, the Gaussian wave expression x is extracted through formula (3): i,1 (n), x i,2 (n), x i,3 (n); The Gaussian wave extraction process of this embodiment is shown in Figure 4 ; Figure A is the main wave extracted during the first decomposition and the remaining residual wave; Figure B is the first reflected wave extracted during the second decomposition and the remaining residual wave, where the input is the residual wave of Figure A; Figure C is the second reflected wave extracted during the third decomposition and the remaining residual wave, where the input is the residual wave of Figure B; Figure D is the relative positions of the shapes of the three decomposed waves in the pulse wave;

[0132] S202, extracting Gaussian wave parameters of each period to form a Gaussian wave parameter vector;

[0133] After decomposition is completed, extract the height A of the main wave i,1 , the height ratio of the first reflected wave to the main wave The height ratio of the second reflected wave to the main wave, the area S of the pulse wave i , Main wave area ratio The first reflected wave area ratio, the second reflected wave area ratio, and the pulse wave cycle length T i , relative position of the main wave peak Relative position of the first reflected wave Relative position of the second reflected wave The area of ​​the pulse wave is defined as S i =S i,1 +S i,2 +S i,3 ; The mark represented by each feature in the pulse wave cycle is shown in Figure 5 ; Integrate these 11 features into a Gaussian parameter vector with 11 dimensions. The parameter vector F of this period i As described in formula (1);

[0134] S203, constructing Gaussian model parameter sequence F u (n);

[0135] Since the maximum slope point is located before the waveform quality screening, the length K of the maximum slope point sequence must be greater than the number of waveforms that pass the screening I; therefore, it is necessary to first remove the maximum slope point position sequence [n M,1 ,n M,2 ,…,n M,K ], thereby obtaining a new maximum slope point position sequence [n′ M,1 ,n′ M,2 ,…,n′ M,I ], as the time of the corresponding periodic Gaussian wave parameter vector, the Gaussian model parameter sequence F is constructed based on formula (2) u (n);

[0136] S3. Calculate parameter sequence F u The time domain features and frequency domain features of (n) are used as sample features of the entire pulse wave signal; the specific steps are:

[0137] S301, calculate Gaussian model parameter sequence F u (n) time domain characteristics;

[0138] Including mean AVE, variance VAR, root mean square difference RMSSD, kurtosis KURT and skewness SKEW, the calculation process refers to formulas (4) to (7);

[0139] S302, calculate Gaussian model parameter sequence F u (n) frequency domain characteristics;

[0140] S3021, parameter sequence Fu (n) interpolation resampling is performed with a resampling frequency of 2 Hz to obtain a uniformly sampled parameter sequence F(n);

[0141] S3022, performing FFT transformation on F(n) to convert it into a power spectrum sequence P(jf), dividing the frequency domain into a very low frequency band (VLF), a low frequency band (LF), and a high frequency band (HF), with frequency ranges of 0 to 0.04 Hz, 0.04 to 0.15 Hz, and 0.15 to 0.4 Hz, respectively;

[0142] S3023, accumulate the entire power spectrum to obtain the total power P total , the ultra-low frequency band is accumulated to obtain the ultra-low frequency power P VLF , and so on to get the low-frequency power P LF , high frequency power P HF ;

[0143] S3024, based on total power P total , ultra-low frequency power P VLF , low frequency power P LF , high frequency power P HF Find the low frequency normalized power LF nu , high frequency normalized power HF nu and low high frequency power ratio R LF / HF , the calculation method refers to formula (8) to formula (10);

[0144] S303, integrating the time domain features and the frequency domain features according to formula (11) as the sample features Z of the entire pulse wave signal;

[0145] S4, forming a training set with sample features of pulse wave signals collected from several subjects under known different mental workload levels, and inputting the samples into the LDA-SVM model for training;

[0146] S401, collecting r segments of pulse wave signals under a known mental workload, and the sample features of each segment of pulse wave signal obtained through step S3 are recorded as Z1, Z2, ..., Z r , construct the model training set [Z1,Z2,…,Z r ];

[0147] This embodiment collects 660 pulse wave sample data from 22 different individuals. The sample duration of each pulse wave is 60 seconds. The distribution of data samples is shown in Table 1:

[0148] Table 1. Sample data distribution table of Example 1

[0149]

[0150] The 1-Back state and the 2-Back state in the acquisition state are two memory experiments of different difficulty in the retrospective experiment; the main purpose of the memory experiment is to put the subject in different mental load states, among which the 2-Back experiment is more difficult than the 1-Back experiment, and the corresponding mental load is greater; this embodiment divides the subject's state into a low-load state, a medium-load state and a high-load state according to two objective indicators of task difficulty and subject task score and the subject's subjective feelings recorded after the task; among which the rest state corresponds to the low-load state, the 1-Back state corresponds to the medium-load state, and the 2-Back state corresponds to the high-load state;

[0151] The detection model of this embodiment is a two-level classification model, that is, the model has two levels for detecting mental workload; in order to test the detection effect of the proposed method for different types of two-level classification, three sample groups are constructed by mixing samples in pairs, namely, a medium-high load sample group, a low-high load sample group, and a medium-low load sample group, each sample group contains 440 samples; the three sample groups are used as inputs of the model to perform three different mental workload level detection experiments on the model; the implementation steps for the three sample groups are completely the same, and the subsequent steps are described by taking only one of the sample groups as an example;

[0152] In order to test the accuracy of model detection so as to better illustrate the characteristics of this method, this embodiment randomly divides the samples into two parts, a training set and a test set, in a ratio of 4:1. Therefore, the training set contains 352 samples and the test set contains 88 samples. The mental workload levels of the training set and the test are known. The mental workload level of the training set is input into the model together with the features of the training set to train the model. During the test, only the features of the test set are input into the trained model without the corresponding mental workload level. The model inputs the predicted value of the mental workload level according to the decision function, and compares the predicted value with the actual mental workload level to obtain the detection accuracy of the model. The specific training and detection process is as follows: Figure 6 The specific operation process of this embodiment is as follows:

[0153] S402, for the training set [Z1, Z2, ..., Z r ] is standardized to obtain [Z′1, Z′2,…, Z′ r ], where Z′1, Z′2,…, Z′ r They respectively represent the sample characteristics of each segment of the pulse wave signal after standardization processing; the standardization processing method refers to formula (12);

[0154] S403, the standardized training set [Z′1, Z′2, …, Z′ r ] perform LDA dimensionality reduction to obtain the projection matrix ω and [Z″1,Z″2,…,Z″ r], where Z″1, Z″2,…,Z″ r They respectively represent the sample features of each segment of the pulse wave signal after LDA dimensionality reduction processing;

[0155] The LDA algorithm uses the API in the scikit-learn library in Python, and sets the API parameter n_components to 1; that is, the LDA algorithm reduces the dimension of the 88-dimensional features; since the mental workload levels of the three sample groups in this embodiment are all 2 categories, the dimension after dimensionality reduction is not higher than the number of mental workload levels divided, so after LDA dimensionality reduction, the original 88-dimensional sample features are transformed into 1-dimensional sample features, which greatly reduces the computational complexity of SVM; in particular, LDA is a supervised dimensionality reduction algorithm, so the sample feature set and sample label need to be input together when inputting, that is, each training sample in the training set is labeled with the mental workload level;

[0156] S404, input the low-dimensional training set [Z″1, Z″2,…, Z″ r ]Train the SVM model;

[0157] The low-dimensional training set is input into the SVM classifier with Gaussian kernel for training; the penalty coefficient C and the kernel function coefficient gamma are optimized through the grid search algorithm and 5-fold cross validation to maximize the classification accuracy, thereby obtaining the decision function;

[0158] At this point, the training of the detection model for mental workload level is completed; the next step is actually to use and test the method;

[0159] S5. Input the test set into the trained LDA-SVM model to obtain the corresponding mental workload level output result, and compare it with the mental workload level corresponding to the known test set; since the test set has been processed in the above part, this embodiment skips step S501, and the rest of the process is as follows:

[0160] S502, standardize the test set based on the mean μ and standard deviation σ of the training set samples to remove the dimension effect of the test data;

[0161] S503, performing a feature dimensionality reduction transformation based on the projection matrix ω on the standardized test set, projecting it into a new feature space that can achieve optimal classification, and obtaining a 1-dimensional vector test set after dimensionality reduction;

[0162] S504, input the reduced dimension test set features into the trained SVM classifier to obtain the output result of the mental workload level, which is a level evaluation of the input data based on the existing mental workload level; finally, compare the evaluation result of the test set with the actual result, and the accuracy is as follows: Figure 7 shown.

[0163] When detecting that the mental load level is high load or low load, this embodiment can achieve an accuracy of 96%; when detecting that the mental load level is medium load and low load, this embodiment can achieve an accuracy of 90%. In other words, the proposed method for detecting the mental load level based on photoplethysmography has a high accuracy rate for detecting mental load, averaging 93%. When detecting that the mental load level is high load or medium load, the accuracy rate of this embodiment is 75%. Combining the above test results, this embodiment can reflect that the proposed detection method has a high detection accuracy for medium and high loads, while the accuracy of distinguishing medium and high loads is slightly lower, but this is consistent with the demand scenarios of actual life: for most people, what they need to know is often whether they are in a medium and high load state. Once they enter a medium and high load state, it often means that the current mental state is no longer optimal and requires proper rest and adjustment.

[0164] Embodiment 2:

[0165] The difference between this embodiment and embodiment 1 is that the resolution of the mental workload level is increased from two levels to three levels, that is, the detection results of the model include: high load, medium load and low load. Therefore, the samples of the three mental workload levels in embodiment 1 are used as a sample group for model training and testing. The specific steps are as follows:

[0166] S1. Refer to the corresponding steps in Example 1 and implement them, which will not be repeated here;

[0167] S2. Refer to the corresponding steps in Example 1 and implement them, which will not be repeated here;

[0168] S3, refer to the corresponding steps in Example 1, and no further details will be given here;

[0169] S4, a training set is composed of sample features of several pulse wave signals collected under a known mental workload level, and is input into the LDA-SVM model for training;

[0170] After feature extraction, the LDA-SVM model needs to be trained first. The optimal parameters and decision plane are obtained by training the LDA-SVM classifier, and then the input features of the mental workload level to be tested are detected and classified. This embodiment is based on the feature samples in Example 1, but does not separate the samples. All of them are taken as a sample group, so the sample set contains 660 samples.

[0171] Consistent with Example 1, in order to test the accuracy of the model in detecting three levels of mental workload, this example divides the feature set into a training set and a test set in a ratio of 4:1. Therefore, the training set contains 528 samples and the test set contains 132 samples. The mental workload levels of the training set and the test are both known.

[0172] S402, for the training set [Z1, Z2, ..., Z r ] is standardized to obtain [Z′1, Z′2,…, Z′ r ], where Z′1, Z′2,…, Z′ r They respectively represent the sample characteristics of each segment of the pulse wave signal after standardization;

[0173] S403, the standardized training set [Z′1, Z′2, …, Z′ r ] perform LDA dimensionality reduction to obtain the projection matrix ω and [Z″1,Z″2,…,Z″ r ], where Z″1, Z″2,…,Z″ r They respectively represent the sample features of each segment of the pulse wave signal after LDA dimensionality reduction processing;

[0174] The LDA algorithm uses the API in the scikit-learn library in Python, and sets the API parameter n_components to 2, so that the LDA algorithm can reduce the dimension of the sample features of 88 dimensions; since the mental workload level of the sample group in this embodiment is 3 categories, the dimension after dimension reduction is not higher than the number of mental workload levels, so after LDA dimension reduction, it is transformed into a sample feature of 2 dimensions;

[0175] S404, input the low-dimensional training set [Z″1, Z″2,…, Z″ r ]Train the SVM model;

[0176] The training set is input into the SVM classifier with Gaussian kernel for training; the penalty coefficient C and the kernel function coefficient gamma are optimized by grid search algorithm and 5-fold cross validation to maximize the classification accuracy, thereby obtaining the decision function; in particular, for different features and different level classifications, that is, when the scenes are inconsistent, the optimization results of C and gamma are different; after the training is completed, enter the testing step;

[0177] S502, standardize the test set based on the mean μ and standard deviation σ of the training set samples to remove the dimension effect of the test data;

[0178] S503, performing a feature dimensionality reduction transformation based on the projection matrix ω on the standardized test set, projecting it into a new feature space that can achieve optimal classification, and obtaining a 2D vector test set after dimensionality reduction;

[0179] S504, input the transformed test set features into the trained SVM classifier to obtain the output result of the mental workload level.

[0180] The results show that when the test results show that the resolution of the mental workload level is three levels, the accuracy of this method is 73%, which is 40% higher than the random prediction probability of 33%.

[0181] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. An objective detection method for mental workload level based on pulse wave Gaussian model characteristics, used for objectively detecting the mental workload state of a human body, characterized in that: The detection method comprises the following steps: S1. Collect pulse wave signals for preprocessing and waveform quality screening: collect pulse wave signals for bandpass filtering and baseline drift removal, extract a stable reference waveform from the preprocessed pulse wave signals, and then calculate the correlation coefficient between each cycle of the pulse wave signal and the reference waveform. The correlation coefficient is used to evaluate the waveform quality, screen out periodic waveforms interfered by motion artifacts, and finally obtain the screened pulse wave signal; S2, decomposing the screened pulse wave signal 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 multiple cycles extracted from the screened pulse wave signal in the time domain; S3, calculating the time domain features and frequency domain features of the Gaussian model parameter sequence as the sample features of the entire pulse wave signal, wherein the time domain features refer to the mean, variance, root mean square of the difference, kurtosis and skewness of the Gaussian model parameters; the frequency domain features refer to the normalized low frequency component, normalized high frequency component and the ratio of the low frequency component to the high frequency component of the parameters; S4, forming a training set with sample features of pulse wave signals collected from several subjects under known different mental workload levels, and inputting the samples into the LDA-SVM model for training; S5, extracting the pulse wave sample features of the subject to be tested, inputting them into the trained LDA-SVM model, and obtaining the output result of the mental workload level; Among them, in step S1, the process of collecting pulse wave signals for preprocessing and waveform quality screening is as follows: S101, collecting a number of pulse wave signals of more than 60 seconds, and performing bandpass filtering on the pulse wave signals to obtain filtered signals ; S102, extracting the pulse wave signal after filtering The base point sequence , where the base point refers to the starting point of each cycle of the pulse wave signal, and the extraction process of the base point sequence is as follows: S1021. Calculate the maximum slope point position sequence of the pulse wave signal ,in, The first The maximum slope point, , Indicates the number of points with the maximum slope on the entire pulse wave signal; S1022, find the position sequence of the base point ,in, Indicates that the entire pulse wave signal passes through the Maximum slope point The located base point position; the maximum slope point position sequence Each value in is the starting point for the pulse wave signal The first-order difference signal Perform reverse traversal and find the first one closest to The point corresponding to this point is the base point position , thus obtaining the position sequence of the base points , is an adjustable coefficient; S1023, position sequence at the base point Top pair Get the base point sequence ; S103: filtering the pulse wave signal To remove the baseline drift: first, Perform cubic spline fitting to obtain the baseline drift curve , and then the pulse wave signal Subtract baseline drift curve Get the pulse wave signal after removing the baseline ; S104, extracting the pulse wave signal after removing the baseline The reference waveform is The most stable periodic waveform in the middle is extracted by the following steps: first, the base point sequence extracted in step S102 is used The pulse wave signal after baseline removal Perform fine cycle segmentation; then normalize the amplitude of each pulse wave signal with a complete cycle; then calculate the correlation coefficient of two adjacent cycles; when the pulse wave signal When a segment has a large correlation coefficient of multiple consecutive cycles, any cycle waveform in the segment is selected as the reference waveform; S105. Screening out pulse wave signals Poor quality periodic waveform: Calculate the reference waveform and When the correlation coefficient between the pulse wave signal of a certain period and the reference waveform is less than a certain threshold, the pulse wave signal of this period is screened out.

2. The objective detection method of mental workload level based on pulse wave Gaussian model characteristics according to claim 1 is characterized in that: In step S2, the pulse wave signal after screening is decomposed cycle by cycle to obtain a Gaussian model parameter sequence, and the process is as follows: S201, decomposing each cycle waveform of the filtered pulse wave signal into three Gaussian waves, the decomposition steps are: firstly taking a pulse wave signal of a certain cycle as input, calculating the starting point and peak point of the first Gaussian wave; taking the longitudinal axis at the location of the peak point as the symmetry axis, mirroring the waveform from the starting point to the peak point, and extracting the first Gaussian wave; then taking the pulse wave signal of the cycle minus the first Gaussian wave as input, and re-executing the above steps, iterating three times to extract the three Gaussian waves of the cycle: the main wave, the first reflected wave, and the second reflected wave; S202, extracting Gaussian wave parameters of each cycle to form a Gaussian wave parameter vector, wherein there are 11 Gaussian wave parameters extracted from each cycle, namely, the height of the main wave, the height ratio of the first reflected wave to the main wave, the height ratio of the second reflected wave to the main wave, the area of ​​the pulse wave, the main wave area ratio, the first reflected wave area ratio, the second reflected wave area ratio, the pulse wave cycle length, the relative position of the main wave peak, the relative position of the first reflected wave, and the relative position of the second reflected wave. The Gaussian wave parameters extracted from each cycle form the Gaussian wave parameter vector of the cycle; S203, constructing Gaussian model parameter sequence , the construction steps are: remove the maximum slope point position sequence The points corresponding to the cycles that are screened out in step S105 are used to obtain a new maximum slope point position sequence. , represents the time of the corresponding periodic Gaussian wave parameter vector, and the Gaussian model parameter sequence is obtained , the subscript u indicates that the sequence is a non-uniformly sampled sequence.

3. The objective detection method of mental workload level based on pulse wave Gaussian model characteristics according to claim 1 is characterized in that: In step S3, the time domain features and frequency domain features of the Gaussian model parameter sequence are calculated as sample features of the entire pulse wave signal, and the process is as follows: S301, calculating Gaussian model parameter sequence The time domain features include the mean ,variance , RMS , Kurtosis , skewness ; S302, calculating Gaussian model parameter sequence The frequency domain features include the low-frequency normalized power , high frequency normalized power and low high frequency power ratio , the calculation process is as follows: S3021, Gaussian model parameter sequence Interpolate and resample to obtain a uniform sampling sequence ; S3022, uniform sampling sequence Perform FFT transformation to obtain the power spectrum sequence ; S3023, power spectrum sequence Calculate the total power by segment , ultra low frequency power , low frequency power , high frequency power ; S3024, based on total power , ultra low frequency power , low frequency power , high frequency power Get the low frequency normalized power , high frequency normalized power and low high frequency power ratio ; S303: Integrate the time domain features and frequency domain features as sample features of the entire pulse wave signal .

4. The objective detection method of mental workload level based on pulse wave Gaussian model characteristics according to claim 1 is characterized in that: In step S4, the sample features of pulse wave signals collected from several subjects under known different mental workload levels are used to form a training set, which is input into the LDA-SVM model for training. The process is as follows: S401, collect sample features of r segments of pulse wave signals, and the sample features of each segment of pulse wave signal obtained through step S3 are recorded as , , …, , construct model training set ; S402, training set After standardization, we get ,in , , …, They respectively represent the sample characteristics of each segment of the pulse wave signal after standardization; S403, standardized training set Perform LDA dimensionality reduction to obtain the projection matrix and ,in , , …, They respectively represent the sample features of each segment of the pulse wave signal after LDA dimensionality reduction processing; S404. Input the low-dimensional training set after dimensionality reduction Train the SVM model.

5. The objective detection method of mental workload level based on pulse wave Gaussian model characteristics according to claim 1 is characterized in that: In step S5, the pulse wave sample features of the subject to be tested are extracted and input into the trained LDA-SVM model to obtain the output result of the mental workload level. The process is as follows: S501, collecting the pulse wave of the object to be tested, and processing according to steps S1 to S3 to obtain the characteristics of the sample to be tested; S502, based on the mean of training set samples and standard deviation Standardize the characteristics of the samples to be tested; S503, based on the projection matrix Perform dimensionality reduction on the standardized features of the samples to be tested; S504, input the features of the sample to be tested after dimension reduction into the trained SVM classifier to obtain the output result of the mental workload level.

Images