阅读说明：本技术 一种基于心冲击信号的多通道隐蔽性测谎方法 (Multi-channel hidden lie detection method based on ballistocardiogram signal ) 是由 唐劲天 李小龙 张砚冬 于 2019-11-14 设计创作，主要内容包括：本发明提供一种基于心冲击信号的多通道隐蔽性测谎方法。所述基于心冲击信号的多通道隐蔽性测谎方法,包括以下步骤：S100：放大BCG信号与GSR信号并对其和其余信号进行提取；通过聚偏氟乙烯压电薄膜传感器和seeed studio公司生产的GSR信号传感器同步实时获取受试者的BCG信号与GSR信号,并经过放大电路进行同步放大处理,同时还采集受试者的语音信号与视频信号。本发明采用的基于心冲击信号的多通道隐蔽性测谎系统设计合理,所用设备结构简单且技术成熟,并且便于携带和开发成本较低；可以极大囊括受试者在生理与行为上表现的差异,同时多路信号抗干扰能力增强,即使在一种信号受到干扰时,还有其他信号在准确的采集相关信息。(The invention provides a multi-channel hidden lie detection method based on a ballistocardiogram signal. The multichannel hidden lie detection method based on the ballistocardiogram signal comprises the following steps: s100: amplifying the BCG signal and the GSR signal and extracting the signals and the rest signals; BCG signals and GSR signals of a testee are synchronously acquired in real time through a polyvinylidene fluoride piezoelectric film sensor and a GSR signal sensor produced by a green studio company, synchronous amplification processing is carried out through an amplifying circuit, and meanwhile voice signals and video signals of the testee are also acquired. The multichannel hidden lie detection system based on the ballistocardiogram signal is reasonable in design, simple in structure and mature in technology of used equipment, convenient to carry and low in development cost; the method can greatly cover the difference of the physical and behavioral performances of the testee, simultaneously the anti-interference capability of the multi-channel signals is enhanced, and even if one signal is interfered, other signals can accurately acquire related information.)

1. A multi-channel hidden lie detection method based on a ballistocardiogram signal is characterized in that: the method comprises the following steps:

s100: amplifying the BCG signal and the GSR signal and extracting the signals and the rest signals;

synchronously acquiring BCG signals and GSR signals of a subject in real time, synchronously amplifying the signals by an amplifying circuit, and simultaneously acquiring voice signals and video signals of the subject;

s200: digitizing the signal and receiving and storing the signal by an upper computer;

the amplified BCG signal and GSR signal are sampled by using an ADC serial port of an STM32L series low-power-consumption singlechip, data are transmitted to PC side software for receiving through WiFi and Bluetooth, meanwhile, voice and video signals are transmitted to application software through USB, and all signals are independently stored in a pre-established storage unit for subsequent signal processing;

s300: preprocessing BCG signals and GSR signals;

preprocessing each segment of BCG and GSR signals stored in an experiment of each subject, wherein the preprocessing of the BCG signals comprises baseline drift removal, high-frequency interference filtering, BCG signal extraction and respiratory signal extraction, and the preprocessing of the GSR signals comprises high-frequency interference removal and signal curve smoothing;

s400: extracting signal characteristics;

extracting relevant characteristics of the preprocessed BCG signal, the preprocessed GSR signal, the preprocessed video signal and the preprocessed audio signal;

s500: reducing dimension of the features;

the dimensionality reduction is carried out on the Merr cepstrum coefficient with overlarge dimensionality by a dimensionality reduction method, so that the fusion of multi-channel signal characteristics is facilitated, and the calculation complexity of a final algorithm is reduced;

s600: standardizing the characteristics;

and standardizing the fused features after dimension reduction, and eliminating errors of the features expressed according to individual differences. The normalization adopts the operation of mean value removal and variance normalization;

s700: predicting the result;

and (3) carrying out classification network training and prediction on the standardized features, and carrying out weighted fusion prediction by using three machine learning methods by using a classifier to give prediction probability.

2. The multi-channel covert lie detection method based on the ballistocardiogram signal as claimed in claim 1, wherein in step S100, important physiological signals of the human body, such as BCG signals and GSR signals, are collected in real time in a non-invasive and non-inductive manner through a polyvinylidene fluoride piezoelectric film sensor and a GSR signal sensor manufactured by the company green studio, and are amplified through an amplifying circuit.

3. The multi-channel covert lie detection method based on the ballistocardiogram signal according to claim 1, wherein in the step S200, the multi-channel signal acquisition device sets the sampling frequency to be 1000Hz, the acquired signal is digitized and transmitted to the PC terminal through Bluetooth, Wifi and USB serial ports, and the upper computer software integrates and stores the signals independently.

4. The multi-channel covert lie detection method based on ballistocardiogram signals as claimed in claim 1, wherein in step S300, only the BCG signals and the GSR signals are preprocessed, and the original BCG signal components are shown as follows:

raw(N)＝bcg(N)+res(N)+n(N)

bcg (n) body fluctuation signal components mainly caused by heart beat and arterial blood circulation;

res (n) is a respiratory signal component;

n (N) is other noise components;

the BCG signal preprocessing steps are as follows:

①, the original BCG signal is passed through a low-pass filter without phase shift, the cut-off frequency is set to 30Hz, and the high-frequency interference signal is filtered;

② passing the signal with high frequency interference through a 4-order bandpass Butterworth filter with lower cut-off frequency of 4Hz and upper cut-off frequency of 20Hz to obtain a BCG signal without respiratory component;

③ passing the signal with high frequency interference removed through a 4-order band-pass Butterworth filter with lower cut-off frequency of 0.3Hz and upper cut-off frequency of 1Hz to separate the waveform of the respiratory signal;

the GSR signal is preprocessed as follows:

①, firstly, filtering the interference of high frequency of GSR signals by a low-pass filter with 40Hz cut-off frequency and without phase shift;

② adding a shift window with width of 0.2 times of sampling frequency to the processed GSR signal, adding and averaging, removing burr, and smoothing signal waveform.

5. The multi-channel covert lie detection method based on ballistocardiogram signals according to claim 1, wherein in the step S400, the extraction operation is as follows:

extracting IJ interval and amplitude, JK interval and amplitude, HJ interval and amplitude, JL interval and amplitude, signal energy, heart rate, respiration amplitude, relative cardiac output rate, Poincare image characteristics and the like from the preprocessed BCG signal;

extracting amplitude mean value, standard deviation, maximum value, minimum value and absolute difference of the maximum value from the preprocessed GSR signal;

extracting the blink rate of the subject from the video signal;

merr cepstral coefficients are extracted for the audio signal.

6. The multi-channel covert lie detection method based on ballistocardiogram signals according to claim 1, wherein in the step S700, the three machine learning methods are a lightGBM method, a Random Forest method and an XGBoost method, respectively.

Technical Field

The invention relates to the technical field of non-sensing lie detection, in particular to a multi-channel hidden lie detection method based on a ballistocardiogram signal.

Background

So called lie, attempts to mislead others using words that contradict reality. The rapid development of modern science and technology, the intellectualization of the case-making tool and the concealment of the case-making means make the case evidence-taking difficulty continuously increase, so that the oral supply for verification and identification becomes the key of case breakthrough. Therefore, a practical lie detection method which accords with the scientific principle is particularly important, the case handling efficiency can be accelerated, and manpower and material resources can be saved.

The traditional lie detection technology mainly uses a skin electrical activity level index, a heart rate mean value and a heart rate variation rate index as physiological measures, and when the skin electrical conductivity level rises obviously larger than a baseline in the lie detection period or when the heart rate mean value (or the heart rate variation rate) is obviously larger than the baseline, the psychological state abnormity is judged. The two traditional indexes have the defects that when the tested person sweats, the relevant reaction of the skin electric events is not obvious; normal physical activities such as posture changes and speaking can also cause significant increase in the skin electrical level, heart rate mean or heart rate variability, so that these traditional indicators cannot correctly indicate abnormal psychological state in a lie detection environment. The current technology for studying the fire-related event-related potential is a lie detection technology, which is mainly embodied in that P300 and CNV (computerized negative variation) can reflect the cognitive level of a subject on lie information from different angles, the familiar memory can induce more positive P300 brain waves than strange information, and strong lying motivation and unknown lying consequences can induce negatively deflected CNV brain waves.

The lie detection technologies or methods all have an inevitable defect, and the device such as a multi-channel electrode or a finger clip needs to be worn, so that inconvenience is brought to the test process, data acquisition in the whole test process is invalid due to the problem of device contact, and the non-concealed invasive lie detection device or technology brings physiological and psychological intangible pressure to a subject and finally has a great influence on the test result.

Therefore, the advantages of the existing lie detection technology are combined, the defects of the existing technology are improved, a multichannel concealed lie detection method based on the ballistocardiogram is provided, the Ballistocardiogram (BCG), the skin resistance signal (GSR), the voice signal and the video signal are synchronously collected in real time through the non-inductive collection technology, the physiology, the behavior and the voice signal of a subject under different states are observed, and the lie detection identification is carried out by extracting the effective characteristics of each path of signal.

Therefore, as can be seen from the comparative discussion above, the inventive advantages and features of the multi-channel covert lie detection method based on ballistocardiogram signals are:

(1) the difference of the physical and behavioral performance of the testee can be greatly covered, the anti-interference capability of the multi-channel signals is enhanced, and even if one signal is interfered, other signals can accurately acquire related information;

(2) the hidden lie detection is realized, and any one path of signal is not required to be worn by the testee, so that the vigilance of the testee can be reduced, and the abnormal situation can be accurately monitored;

(3) one new promotion of covert lie detection technology is that its price is lower and signal processing complexity is more dominant than infrared thermal imaging technology.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention aims to provide a multi-channel concealed lie detection method based on a cardioblast signal, which can make up the defects that the existing lie detection technology is single in signal, can collect signals in an intrusive mode and cannot conceal the signals, meanwhile, commonly used signal characteristics of the lie detection technology, such as physiological characteristics of the cardioblast signal, a skin resistance signal and the like, are kept, and the aim of truly and reliably acquiring information of a testee in different states on the premise of not influencing the normal physiological and psychological reactions of the testee is fulfilled.

In order to solve the technical problem, the invention provides a multi-channel hidden lie detection method based on a ballistocardiogram signal, which comprises the following steps:

s100: amplifying the BCG signal and the GSR signal and extracting the signals and the rest signals;

synchronously acquiring BCG signals and GSR signals of a subject in real time through a polyvinylidene fluoride piezoelectric film sensor and a GSR signal sensor produced by a green studio company, synchronously amplifying the signals through an amplifying circuit, and simultaneously acquiring voice signals and video signals of the subject;

s200: digitizing the signal and receiving and storing the signal by an upper computer;

the amplified BCG signal and GSR signal are sampled by using an ADC serial port of an STM32L series low-power-consumption singlechip, data are transmitted to PC side software for receiving through WiFi and Bluetooth, meanwhile, voice and video signals are transmitted to application software through USB, and all signals are independently stored in a pre-established storage unit for subsequent signal processing;

s300: preprocessing BCG signals and GSR signals;

preprocessing each segment of BCG and GSR signals stored in an experiment of each subject, wherein the preprocessing of the BCG signals comprises baseline drift removal, high-frequency interference filtering, BCG signal extraction and respiratory signal extraction, and the preprocessing of the GSR signals comprises high-frequency interference removal and signal curve smoothing;

s400: extracting signal characteristics;

extracting relevant characteristics of the preprocessed BCG signal, the preprocessed GSR signal, the preprocessed video signal and the preprocessed audio signal;

s500: reducing dimension of the features;

the dimensionality reduction is carried out on the Merr cepstrum coefficient with overlarge dimensionality through a dimensionality reduction method, the fusion of multi-channel signal characteristics is facilitated, and the calculation complexity of a final algorithm is reduced;

s600: standardizing the characteristics;

by carrying out mean value removing and variance normalizing operation on each feature of each person, the feature data error caused by individual difference can be eliminated, the effectiveness of data and the accuracy of classification are improved, the fused features after dimension reduction are standardized, the error of the features expressed by the individual difference is eliminated, and the mean value removing and variance normalizing operation is adopted for standardization;

s700: predicting the result;

the classifier performs classification network training and prediction on the standardized features, the classifier performs weighted fusion prediction by adopting three machine learning methods to give prediction probability, the generalization capability and the accuracy of the lie detection algorithm model can be further improved, meanwhile, upper computer software used by the lie detection method can process data in real time to give real-time physiological data and a lie detection state index of a subject, and the physiological condition and the psychological stress state of the subject can be observed while the lie is detected.

Preferably, in step S100, important physiological signals of the human body, such as BCG signals and GSR signals, are collected in a non-invasive and non-inductive manner in real time through the polyvinylidene fluoride piezoelectric film sensor and the GSR signal sensor manufactured by the sweet studio company, and are amplified through the amplifying circuit.

Preferably, in step S200, the multichannel signal acquisition device sets the sampling frequency to 1000Hz, digitizes the acquired signal, transmits the digitized signal to the PC terminal through bluetooth, Wifi and USB serial ports, and the upper computer software integrates and stores each path of signal independently.

Preferably, in step S300, only the BCG signal and the GSR signal are preprocessed, and the original BCG signal component is shown as the following formula:

raw(N)＝bcg(N)+res(N)+n(N)

bcg (n) body fluctuation signal components mainly caused by heart beat and arterial blood circulation;

res (n) is a respiratory signal component;

n (N) is other noise components;

the BCG signal preprocessing steps are as follows:

① the original BCG signal is passed through a low-pass filter without phase shift, the cut-off frequency is set to 30Hz, so as to realize the filtering of high-frequency interference signals (such as 50Hz AC power frequency interference);

② passing the signal with high frequency interference through a 4-order bandpass Butterworth filter with lower cut-off frequency of 4Hz and upper cut-off frequency of 20Hz to obtain a BCG signal without respiratory component;

③ passing the signal with high frequency interference removed through a 4-order band-pass Butterworth filter with lower cut-off frequency of 0.3Hz and upper cut-off frequency of 1Hz to separate the waveform of the respiratory signal;

the GSR signal is preprocessed as follows:

①, firstly, filtering the interference of high frequency of GSR signals by a low-pass filter with 40Hz cut-off frequency and without phase shift;

② adding a shift window with width of 0.2 times of sampling frequency to the processed GSR signal, adding and averaging, removing burr, and smoothing signal waveform.

Preferably, in step S400, the extracting operation is as follows:

extracting IJ interval and amplitude, JK interval and amplitude, HJ interval and amplitude, JL interval and amplitude, signal energy, heart rate, respiration amplitude, relative cardiac output rate, Poincare image characteristics and the like from the preprocessed BCG signal;

extracting amplitude mean value, standard deviation, maximum value, minimum value and absolute difference of the maximum value from the preprocessed GSR signal;

extracting the blink rate of the subject from the video signal;

merr cepstral coefficients are extracted for the audio signal.

Preferably, in step S700, the three machine learning methods are a lightGBM method, a RandomForest method, and an XGBoost method, respectively.

Compared with the related technology, the multichannel hidden lie detection method based on the ballistocardiogram signal has the following beneficial effects:

1. the multichannel hidden lie detection system based on the ballistocardiogram signal is reasonable in design, simple in structure and mature in technology of used equipment, convenient to carry and low in development cost.

2. The method can greatly cover the difference of the physical and behavioral performances of the testee, simultaneously the anti-interference capability of the multi-channel signals is enhanced, and even if one signal is interfered, other signals can accurately acquire related information.

3. The hidden lie detection is realized, and any one path of signal is not required to be worn by the testee, so that the vigilance of the testee can be reduced, and the abnormal situation can be accurately monitored.

4. Compared with the infrared thermal imaging technology, the hidden lie detection technology is lower in price and more dominant in signal processing complexity, the system applies the cardiac shock signals to lie detection for the first time, the technology can realize non-sensing acquisition of physiological signals, and can also perform signal processing to separate respiratory signals and realize the function of multiple signals of a single sensor.

Drawings

FIG. 1 is a flow chart of a multi-channel concealed lie detection method based on a ballistocardiogram signal according to the invention;

FIG. 2 is a schematic diagram of a BCG signal and GSR signal acquisition device system;

FIG. 3 is a view showing a model of a chair according to the present invention;

FIG. 4 illustrates waveforms of speech signals collected according to the present invention;

FIG. 5 is a waveform of a BCG signal acquired by the present invention;

fig. 6 is a waveform of a GSR signal acquired by the present invention.

Detailed Description

The invention is further described with reference to the following figures and embodiments.

Please refer to fig. 1, fig. 2, fig. 3, fig. 4, fig. 5 and fig. 6 in combination, in which fig. 1 is a flowchart of a multi-channel concealed lie detection method based on a ballistocardiogram signal according to the present invention; FIG. 2 is a schematic diagram of a BCG signal and GSR signal acquisition device system; FIG. 3 is a view showing a model of a chair according to the present invention; FIG. 4 illustrates waveforms of speech signals collected according to the present invention; FIG. 5 is a waveform of a BCG signal acquired by the present invention; fig. 6 illustrates the waveform of the GSR signal collected by the present invention. A multi-channel hidden lie detection method based on a ballistocardiogram signal comprises the following steps:

s100: amplifying the BCG (ballistocardiogram) signal and the GSR (Galvanic Skin response) signal and extracting the same and the rest of signals;

referring to the attached figure 3 of the specification, the chair armrest is provided with a GSR electrode for detecting skin resistance signals of a testee, a polyvinylidene fluoride piezoelectric film is embedded into a chair cushion for noninductive acquisition of BCG signals, and the specific experimental process is as follows:

① A quiet small room is selected to avoid interference of other noises to voice signal acquisition, and the examinee is put on a lie detection chair and is provided with a skin resistance finger stall or a skin resistance electrode which is directly pressed on the armrest of the chair by the palm;

②, enabling the testee to relax, adjusting for 1-2 minutes, waiting for the bioelectricity signal to be stabilized in a range value by observing the waveform on the visual software, and simultaneously adjusting the psychological shock signal, the video signal and the voice signal;

③ adjusting each signal, starting experiment, the experiment operator extracting a playing card, the testee seeing the playing card, then the testee telling the seen playing card, at this time, the answer of the testee can explain the fact, also can say lie, another experiment operator records the true or false answer of the testee as the data label;

④ the time from the beginning of the subject response to the end of the data acquisition is 10s for each round, i.e. the data acquisition time is 10s for each round;

⑤ repeat step 4. each subject received 30 cycles;

⑥ at the end of the experiment, subjects were asked to fill out a questionnaire.

As shown in the experimental flow, data of 10 seconds of signal from each time the subject starts answering the question is collected and stored in a preset unit, and is calculated by using the next data, including 10 seconds of BCG signal, GSR signal, voice signal and video signal.

S200: digitizing the signal and receiving and storing the signal by an upper computer;

referring to the attached figure 2 of the specification, after signals are amplified, an ADC acquisition circuit in an STM32L series chip is used for sampling and digitizing BCG signals and GSR signals processed by a filter circuit and a signal amplification circuit, the sampling frequency can be set according to the requirements of a use scene, the digitized signals are uploaded to an upper computer at a PC end through WIFI and Bluetooth modules for data relevant processing, the uploading rate can be set through a timer unit, and all the signals are independently stored in a storage unit.

S300: preprocessing BCG signals and GSR signals;

the acquired signals need to be visualized dynamically in real time, so as to enable a user to determine whether the acquired signals are effective and observe the change condition of the data of the subject; the upper computer part of the invention adopts a QT library to perform real-time visualization operation on signals, and the specific signal waveforms are shown in the attached figure 4, the attached figure 5 and the attached figure 6 of the specification, wherein the waveforms shown in the attached figure 5 and the attached figure 6 of the specification are preprocessed signal waveforms, and the specific processing method is as follows:

the original BCG signal composition is shown below:

raw(N)＝bcg(N)+res(N)+n(N)

bcg (n) body fluctuation signal components mainly caused by heart beat and arterial blood circulation;

res (n) is a respiratory signal component;

n (N) is other noise components.

The BCG signal preprocessing steps are as follows:

① the original BCG signal is passed through a low-pass filter without phase shift, the cut-off frequency is set to 30Hz, so as to realize the filtering of high-frequency interference signals (such as 50Hz AC power frequency interference);

② passing the signal with high frequency interference through a 4-order bandpass Butterworth filter with lower cut-off frequency of 4Hz and upper cut-off frequency of 20Hz to obtain a BCG signal without respiratory component;

③ the waveform of the respiratory signal can be separated by passing the high frequency interference-filtered signal through a 4-order bandpass Butterworth filter with a lower cut-off frequency of 0.3Hz and an upper cut-off frequency of 1 Hz.

Considering that the main components of the human skin resistance signal are concentrated below 30Hz, the GSR signal preprocessing steps are as follows:

①, firstly, filtering the interference of high frequency of GSR signals by a low-pass filter with 40Hz cut-off frequency and without phase shift;

② adding a shift window with width of 0.2 times of sampling frequency to the processed GSR signal, adding and averaging, removing burr, and smoothing signal waveform.

S400: extracting signal characteristics;

taking the feature point positioning of the BCG signal as an example, the specific process of extracting the BCG features is described, and the key point positioning algorithm of the BCG signal H, I, J, K, L is as follows:

① the preprocessed BCG signal is subjected to a windowing translation energy calculation as shown in the following equation:

② further sets a window of 0.5 times the sampling rate to find EN (N) the maximum point in the window, labeled as M_NN represents the position of the data point, and the window is then slid backwards by one unit, e.g. the maximum value M_NThe position is not changed, and the sliding is continued; if changed, and is compared with the last M_NThe point is in the same rectangular window, and the current point is substituted for the last M_NPoint;

(3) will M_NFinding the maximum value in the corresponding position in BCG (N) in the list within +/-200 ms, and marking the maximum value as a J peak; in the range of J peak to 200ms, searching a minimum value, and marking the minimum extreme value as I; in the range of J peak +200ms, searching a minimum value, and marking the minimum extreme value as K; similarly, in the range of point I to 200ms, a maximum value is searched, and the minimum extreme value is marked as H; in the range of K point-200 ms, a minimum value is found, and the minimum extreme value is marked as L.

Based on the located signal key points, 32 statistical characteristics of the extracted BCG and the extracted respiratory signal exist.

The calculation formulas of features SD1 and SD2 extracted based on Poincare diagram analysis are as follows:

JJ_ithe spacing between adjacent J peaks is such that,

is the mean of the intervals.

The extracted partial features are shown in table 1:

TABLE 1

Wherein the meaning of the parameter names in table 1 is as follows:

t1: the interval of BCG feature points i to j;

t2: the interval from BCG feature point j to k;

t3: the interval from BCG feature points i to k;

t4: the interval from the BCG feature points h to j;

t5: the interval of BCG feature points j to l;

t6: the interval from the BCG characteristic point h to the point l;

h1: the absolute difference in amplitude between point i and point J;

h2: the absolute difference in amplitude of point k and point J;

h3: the magnitude of point i;

h4: the magnitude of point j;

h5: the magnitude of point k;

SD 1: fitting the minor axis of the ellipse in the poincare map;

SD 2: fitting the long axis of the ellipse in the poincare map;

SD1/SD 2: the ratio of minor axis to major axis;

resp _ rate: a respiration rate;

heart _ rate: heart rate;

bcg _ energy: BCG waveform energy;

GSR _ std: skin resistance signal amplitude standard deviation;

GSR _ mean: average value of amplitude of skin resistance signal;

GSR _ max: skin resistance signal amplitude maximum;

GSR _ min: skin resistance signal amplitude minimum;

GSR _ abs: absolute difference between maximum and minimum values of amplitude of skin resistance signal;

speech _ mel: a mel cepstrum of the speech signal;

blink _ rate: extracting a blink rate according to the video signal;

blink _ count: the number of blinks extracted.

S500: reducing dimension of the features;

the feature dimensionality reduction is mainly used for processing a voice signal, and when a Mel cepstrum coefficient of the voice signal is extracted, the feature dimensionality is too large, so that the feature dimensionality reduction is firstly carried out for the purpose of subsequent feature fusion of each model and reduction of algorithm calculation complexity.

The self-encoder adopts a BP neural network, neurons of a first layer to a fourth layer of an encoding layer are respectively 256, 128, 64 and 32, neurons of a corresponding decoding layer are 64, 128 and 256, and a layer with the number of 32 of the extracted middle hidden neurons is taken as a data dimension after dimension reduction.

S600: standardizing the characteristics;

the feature normalization is performed by means of mean-removing and variance normalization, and the formula is shown as follows:

x is the original feature matrix, mu is the mean value of each feature, and delta is the standard deviation of each feature.

S700: predicting the result;

the data after signal preprocessing, feature extraction, feature dimensionality reduction and normalization are transmitted to three algorithm models, namely lightGBM, RandomForest and XGboost for training, then test data are transmitted to a trained network for testing, and the test result of each algorithm model is subjected to linear weighted average to serve as a final prediction result, so that the generalization capability of the model can be well improved.

It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.

The preferred embodiments of the invention disclosed above are intended to be illustrative only. The preferred embodiments are not intended to be exhaustive or to limit the invention to the precise embodiments disclosed. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the invention and the practical application, to thereby enable others skilled in the art to best utilize the invention. The invention is limited only by the claims and their full scope and equivalents.