CN108593557A - Based on TE-ANN-AWF mobile pollution source telemetry errors compensation methodes - Google Patents

Abstract

The invention discloses one kind being based on TE ANN AWF mobile pollution source telemetry errors compensation methodes, the present invention carries out causal correlation analysis using TE transfer entropies to interference and measurement result, so that it is determined that uneven degree between Measuring origin and the more interference of measurement, and utilize the non-significant causal quantitative criteria of the directionality of TE transfer entropies extraction and determination method.Virtual observation method is proposed to realize the polynary destructing of unit observation sequence, the compensation of channel virtual observation sequence is disturbed by ANN error prediction model realizations single-trunk, then fusion reconstruct is carried out to the polynary virtual observation sequence after compensation using polynary adaptive weighted fusion method.For the weight convergence problem in blending algorithm, the advantages of Weight number adaptively that the good weights of TE are estimated ability and AWF by method that index forgetting is introduced in model adjusts, is combined, and improves the dynamic property of error compensation procedure.

Description

Telemetry Error Compensation Method Based on TE-ANN-AWF Mobile Pollution Source

technical field

The present invention designs a mobile pollution source remote sensing detection error compensation method based on TE-ANN-AWF, which belongs to the technical field of mobile pollution source remote sensing detection instrument error compensation, with the goal of correcting the measurement results of remote sensing detection instruments under external environmental interference According to the entropy estimation theory, adaptive fusion and neural network related theories, compensation and estimation are carried out, and then the remote sensing detection method is easily disturbed by the external environment.

Background technique

Mobile pollution sources refer to sources that do not emit air pollutants from fixed equipment, such as motor vehicles, mobile construction machinery, ships, and aircraft that emit exhaust gas during movement. Mobile source pollution, especially heavy-duty diesel trucks, commercial gasoline vehicles, old cars, and construction machinery vehicles, has become the main source of air pollution, and is an important cause of urban ultrafine particulate matter and photochemical smog pollution. In 2012, the World Health Organization has upgraded diesel exhaust from a suspected carcinogen to a definite carcinogen. In order to effectively control mobile pollution sources, especially the identification and control of high-emission motor vehicles, real-time and effective detection of emissions from mobile pollution sources is required.

At present, there are many kinds of mobile pollution source detection technologies, such as reflecting the degree of air pollution through plant growth or using chemical detection to determine the concentration of pollutants, but these methods are difficult to detect mobile pollution sources online in real time. Remote sensing detection method is a fast and effective method to detect mobile pollution sources. The method is based on precise analysis of gas optical absorption spectra, and the concentration of gas components in the environment is retrieved according to their absorption properties in the ultraviolet, visible and infrared spectral bands. In 1988, the first remote sensing monitoring system (Remote Sensing Device, RSD) using non-dispersive infrared technology (Nondispersive Infrared, NDIR) was developed by the University of Denver in the United States. In the 1990s, the University of Denver developed a remote sensing monitoring system using UltraViolet (UV) to measure NO concentration, which overcomes the problem of water vapor absorption in NDIR measurement. Afterwards, MD-LaserTech developed a NO and HC remote sensing system based on Ultra-Violet Differential Optical Absorption Spectrometry (UV-DOAS). Due to the narrow emission wavelength band of Tunable Diode Lasers (Tunable Diode Lasers, TDL), different measured gas components can be selected for measurement without interference, which meets the real-time monitoring requirements of high time resolution, high sensitivity, and high selectivity. Later, the Massachusetts Institute of Technology developed a telemetry system based on infrared laser differential absorption spectroscopy technology TDLAS (Tunable DiodeLaserAbsorption Spectrometer), which was used to detect excessive emissions from mobile pollution sources such as motor vehicles. At present, remote sensing detection technology has become the mainstream technical means for real-time online detection of gas concentration.

The remote sensing measurement method has a high degree of automation, as long as the equipment is erected beside the road, a large number of vehicles passing through the road section can be measured. At the same time, this method has less impact on traffic. Liverpool John Moores University in the United Kingdom and TSI in the United States have developed a roadside monitoring system that can detect pollutants such as CO and NOx based on TDLAS, UV-DOAS and other technologies. However, this measurement method has limited application conditions and is greatly affected by environmental interference, such as ambient temperature, humidity, air pressure, wind speed, etc., which have an important impact on the detection data. Moreover, the geographical environment of cities is mostly complex and changeable, such as: the urban canyon effect will cause sudden changes in airflow. Due to the complexity of the external environment, the German Fraunhofer Institute for Physical Measurement Technology has added the measurement of wind speed, wind direction, temperature and humidity, air pressure and other meteorological parameters on the basis of the detection of pollutant parameters, and conducted a comprehensive analysis of the pollutant detection results. Correction.

Contents of the invention

The present invention aims at the problem that the remote sensing detection method of mobile pollution sources is easily disturbed by the external environment. In this paper, a new error compensation model TE-ANN- AWF. Among them, the TE transfer entropy is used in the model to quantitatively analyze the causal correlation between interference and measurement, and a method for judging the non-significant causal relationship is derived. The idea of virtual observation is proposed in the model to realize the multivariate deconstruction of the unit observation sequence, and then the multivariate virtual observation sequence is reconstructed by the multivariate fusion method. The method of introducing exponential forgetting into the model combines the advantages of TE's good weight estimation and AWF's weight adaptive adjustment, which improves the dynamic performance of the error compensation process.

Technical solution of the present invention:

Step 1: Obtain measurement samples under different interference effects through environmental simulation experiments; based on experimental measurement samples, conduct interference correlation analysis through TE transfer entropy, so as to determine the source of measurement error and measure the imbalance between multiple interferences; and use TE The directionality of the transfer entropy leads to the quantitative standard and judgment method of the non-significant causal relationship;

Step 2: adopt the environment simulation smog box experimental platform to obtain the training sample set of single interference channel, and establish the measurement error prediction model of each interference by the neural network ANN method;

Step 3: Realize the multivariate deconstruction of the unit observation sequence through the virtual observation method, perform error compensation on the deconstructed multivariate virtual observation sequence through the ANN error prediction model under different disturbances, and then use the multivariate adaptive weighted fusion method to correct the compensated multivariate Fusion and reconstruction of virtual observation sequence;

Aiming at the weight convergence problem in the fusion algorithm, the exponential forgetting method is introduced in the model to combine the weight estimation ability of TE with the advantages of adaptive weight adjustment of the multivariate adaptive weighted fusion method, which improves the error compensation process. dynamic performance.

In said step one, for the detectable characteristics of external environment interference, introduce TE transfer entropy to carry out correlation causality analysis to remote sensing measurement interference, utilize the directionality of transfer entropy to draw non-significant causal relationship quantification standard and judgment method;

Assuming that the interferences are independent to a certain extent, the measurement sequence under the control of single interference changes is obtained through the simulation experiment platform, and the transfer entropy TE _T->CO of the temperature interference to the measured value is calculated, and the transfer entropy TE of the humidity interference to the measured value is calculated. _H->CO , transfer entropy TE _P->CO from air pressure disturbance to measured value, transfer entropy TE _W->CO from wind speed disturbance factor to measured value, among them take the largest reverse transfer entropy TE ₀ as the measure of non-causality standard;

TE ₀ =max{TE _CO->T , TE _CO->H , TE _CO->P , TE _CO->W } (1).

In the third step, the multivariate adaptive weighted fusion method is used to fuse and reconstruct the compensated multivariate virtual observation sequence. Specifically, the multivariate observation values of each interference channel are adaptively fused under the criterion of the minimum mean square error; Through the virtual observation method and the exponential forgetting mechanism, the three methods of TE transfer entropy, ANN artificial neural network and AWF adaptive weighted fusion are closely combined with each other.

In the third step, the exponential forgetting method is introduced into the model to combine the advantages of TE’s good weight prediction ability and AWF’s weight adaptive adjustment, specifically as

In the TE-ANN model, although the transfer entropy has a good weight estimation ability, the weight cannot be adjusted according to the error so as to gradually converge, resulting in the error distribution also fluctuating and unable to converge; while AWF adaptive weighting The fusion algorithm has the advantage that the weights can gradually converge under the criterion of the minimum mean square error without any prior knowledge; in order to combine the advantages of TE’s good weight estimation and AWF’s weight adaptive adjustment, the introduction forgetting mechanism;

Combine the obtained confidence weight K of transfer entropy estimation and the optimal weight factor W ^* according to the minimum mean square error criterion, select the weight coefficient {β _n }, and fuse K and W to obtain As shown in the following formula;

Among them, n represents the fusion of the nth observation value, Represents the weighting coefficient of the confidence weight K, K _n represents the confidence weight of the nth observation value for transfer entropy prediction, Indicates the weighting coefficient of the optimal weighting factor W, and W _n ^* indicates the weight of the AWF after the nth observation is fused.

In view of the characteristics of stable estimation of transfer entropy and the adaptive adjustment and convergence of weights of AWF, the estimation value of the initial dynamic process of weights should focus on K _n , while the steady-state process should focus on W _n ^* and gradually converge in W _n ^* ; in order to reflect the above characteristics, the weighting coefficient {β _n } needs to meet the following characteristics:

i represents a natural number;

In order to meet the above conditions, construct the following function;

d _n =(1-b)/(1-a·b ⁿ ), n=1, 2, 3L (4)

Among them, b is the forgetting factor, a is the attenuation factor, 0<b<1<a; the weighting coefficient is obtained by formula (3):

From this, the weight of each channel is obtained, as shown in the formula:

Advantages that the present invention exists compared with prior art:

(1) The present invention combines three methods of TE transfer entropy, ANN artificial neural network, and AWF adaptive weighted fusion to propose a new error compensation model TE-ANN-AWF. Compared with the traditional error compensation method, it does not need to know the multi-interference measurement samples a priori, it can effectively compensate the measurement error caused by the external environment interference, and improve the applicability and anti-interference ability of remote sensing measurement.

(2) The present invention is aimed at the detectable characteristics of external interference, introduces TE transfer entropy to carry out correlation causality analysis to remote sensing measurement interference, utilizes the directionality of transfer entropy to draw the quantitative standard and judgment method of non-significant causality.

(3) The present invention introduces an adaptive weighted fusion algorithm AWF, which performs adaptive fusion on the observed values of each interference channel under the criterion of minimum mean square error, thereby ensuring the long-term stability of error compensation.

(4) In order to further improve the dynamic performance of the error compensation process, the present invention combines the good weight estimation advantages of TE and the weight adaptive adjustment advantages of AWF by introducing a forgetting mechanism.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention;

Fig. 2 is the comparative figure of transfer entropy of different interference factors of the present invention;

Fig. 3 is the telemetry error neural network prediction model diagram under the single interference of the present invention;

Fig. 4 is the CO measurement error distribution diagram under the temperature-concentration of the present invention;

Fig. 5 is the distribution diagram of CO measurement error under the wind speed-concentration of the present invention;

Fig. 6 is the distribution diagram of CO measurement error under air pressure-concentration of the present invention;

Fig. 7 is an explanatory diagram of reconstruction of the virtual observation observation sequence of the present invention;

Fig. 8 is a TE-ANN-AWF error compensation model diagram of the present invention;

Detailed ways

For the technical innovation point that the present invention realizes is easy to understand, below in conjunction with Fig. 1, the realization mode of the present invention is described in further detail, and concrete steps are as follows:

Step 1: Obtain measurement samples under different interference effects through environmental simulation experiments, and then preprocess and normalize the measurement samples. Based on the experimental measurement samples, the interference correlation analysis is carried out through the TE transfer entropy, so as to determine the source of the measurement error and measure the imbalance between multiple interferences. And use the directionality of TE transfer entropy to elicit the quantitative standard and judgment method of non-significant causal relationship.

Suppose Xn and Yn are two environmental disturbance change sequences and remote sensing measurement observation sequences with discrete states of xn and yn at time n, and Xn and Yn can be approximated as k-order and l-order steady-state Markov processes respectively, then from The transfer entropy definitions of Yn and Xn are as follows:

Among them, T _Y→X represents the transfer entropy (Transfer Entroy) from Y to X, u _n =(x _n+1 ,x _n ,y _n ^(l) ), p(u _n ) represents the state state x _n+1 and The probability that the sequence x _n(k) and y _n ^(l) appear at the same time; p(x _n+1 |x _n ^(k) , y _n ^(l) ) means that at time n, it is known that x _n ^(k) , y Under the premise of _n ^(l) , the conditional probability of x _n+1 ; p(x _n+1 |x _n ^(k) ) means the conditional probability of x _n+1 under the premise of known x _n ^(k) , when x When the state of _n at a certain moment is completely determined by its own historical state, the transfer entropy is zero.

Since remote sensing detection is based on the principle of optical absorption to complete the detection of mobile pollution source sewage gas, and optical signals have problems such as absorption, scattering, beam deflection, and beam diffusion. Most of the detection is carried out in an outdoor environment, and the measurement process is easily affected by factors such as temperature, humidity, air pressure, wind speed, wind direction, and dust. It is a complex nonlinear dynamic problem that is affected by multiple factors. Transfer entropy can quantitatively measure the linear and nonlinear relationship between system variables, and has good anti-noise ability, so it is often used to describe the dynamic nonlinear characteristics of complex systems. Therefore, the transfer entropy is selected in the compensation model to conduct correlation analysis on the measurement data, which helps to trace the source of the error and compensate the error more reasonably and accurately.

In order to verify the effectiveness of the compensation model, CO is selected as the target detection object, and the measurement environment of temperature, humidity, air pressure and wind speed changes is simulated through the environmental simulation experiment platform, and the target object is measured by telemetry equipment. In the compensation model, transfer entropy is used to measure the degree of imbalance among multiple disturbances. Assuming that the disturbances are independent of each other to a certain extent, the measurement sequence under the control of the change of single disturbance is obtained through the simulation experiment platform, so as to calculate the transfer entropy TE _T -> CO from the temperature disturbance to the measured value, and the transfer entropy TE from the humidity disturbance to the measured value _H -> CO, transfer entropy from air pressure disturbance to measured value TE _P -> CO, transfer entropy from wind speed disturbance factor to measured value TE _W -> CO, where the largest reverse transfer entropy TE0 is taken as the measure of non-causality .

TE ₀ =max{TE _CO->T , TE _CO->H , TE _CO->P , TE _CO->W }

It can be seen from Figure 2 that the transfer entropy TE _W->CO , TE _T->CO , and TE _{P->CO measured from wind speed, temperature, and air pressure to CO} are obviously greater than the transfer entropy TE _H-> from humidity to CO measurement _CO , while the transfer entropy TE _H->CO and the reverse transfer entropy TE ₀ from humidity to CO measurements are not much different. From the perspective of information theory, the information contained in the measurement sequence can be significantly explained from the interference sequence of wind speed, temperature, and air pressure, while a small part can be explained from the humidity sequence. Therefore, wind speed, temperature, and air pressure have a significant causal relationship to the measurement results, that is, the three environmental factors have a greater interference with the measurement, and the proportion of the three in the measurement error is relatively large. It can be seen that the transfer entropy from humidity to measurement TE _H->CO is very close to the reverse transfer entropy TE ₀ , reflecting that humidity has no obvious causal relationship with CO measurement results. Therefore, the influence of humidity can not be considered in the measurement compensation of CO.

Step 2: adopt the environment simulation smog box experimental platform to obtain the training sample set of single interference channel, and establish the measurement error prediction model of each interference by the neural network ANN method;

As shown in Figure 3, it is the neural network prediction model of CO gas telemetry error under temperature interference. Because the remote measurement error is not only affected by interference factors, but also the real concentration of the gas to be measured will also affect the absolute value of the error. Therefore, the input data of neural network is standard gas concentration n, temperature t, and the output data is measurement error e. According to the selection method of the hidden layer points of the forward network and the effect of the actual data processing, the number of hidden layers is selected to be 7. Through the smog box environment simulation platform, other factors are fixed, and the temperature in the box is regulated to obtain training samples under single temperature interference. Similarly, neural network prediction models under wind speed disturbance and air pressure disturbance can also be established. Figures 4-6 are the error distribution of the test samples under the interference of single temperature, single air pressure and single wind speed respectively.

After the three prediction models of temperature, wind speed and air pressure are established, the error prediction can be performed on the measured values under certain temperature, wind speed and air pressure respectively, and then the three prediction errors can be summed to obtain the final compensation value.

Step 3: According to the TE-ANN-AWF model structure diagram shown in Figure 8, the transfer entropy TE, neural network ANN, and adaptive weighted fusion AWF are combined for the final measurement value with disturbance through virtual observation deconstruction and exponential forgetting mechanism. estimate.

1) First, the multivariate deconstruction of the unit observation sequence is realized through the virtual observation method, and the corresponding error compensation is performed on the deconstructed multivariate virtual observation sequence through the ANN error prediction model under different disturbances, and then the multivariate adaptive weighted fusion method AWF is used to The compensated multivariate virtual observation sequence is fused and reconstructed.

For the ANN error prediction model, the real concentration of the gas to be measured has a direct impact on the error distribution. In fact, the true concentration of the gas to be measured cannot be known, and can only be roughly estimated through historical sequences, which also leads to inaccurate measurement errors in prediction. Therefore, the effect of directly adding the error compensation values of each channel is not obvious. According to Figure 7, in order to apply the data fusion method to the error compensation model, the observation sequence needs to be deconstructed. To this end, the concept of virtual observation is firstly introduced.

For example, the remote sensing detection of CO is interfered by temperature, air pressure, and wind speed in the environment. Assuming that the mutual coupling of the influence of the three interference factors on the measurement results is small, the composition of the error can be considered as additive noise. Therefore, the measured value can be regarded as expressed by the following formula, where Y is the measured value, yr is the real value, Wt, Ww, Wp are the environmental interference noise, and θ is the measurement random noise.

Y＝y _r +W _t +W _w +W _p +θ

Among them, Wt, Ww, and Wp are the measurement errors of single interference, which can be predicted by the single interference ANN model.

Assuming that the measured value under temperature single interference is Yt, it can be considered as the value after excluding both Ww and Wp, as shown in the following formula.

Y _t ＝y _r +W _t +θ＝YW _w -W _p

Obviously, the measured value Yt under single disturbance of temperature does not exist in practice. In fact, it is a reconstruction of the original observation value, so it is called "virtual observation". In the same way, the measured value Yp under the single interference of air pressure and the measured value Yw under the single interference of wind speed can be obtained, as shown in the following formula.

Y _w ＝y _r +W _w +θ＝YW _t -W _p

Y _p ＝y _r +W _p +θ＝YW _t -W _w

After deconstructing observations through the concept of virtual observations, reconstruction is required. Compared with temperature and air pressure, wind speed has a greater impact on the measurement results, especially when the wind speed is greater than 5m/s. Obviously, the reliability of the measured value at this time should be the smallest. In order to express the size of this credibility, transfer entropy is introduced to express the credibility of the three virtual observations. From the analysis of transfer entropy, it can be seen that if the interference transfer entropy is larger, the causal relationship of interference measurement is stronger, and the interference of interference on the measurement results is stronger. Therefore, the credibility of the virtual observation is inversely proportional to the transfer entropy. At the same time, in order to keep the sum of the weights at 1, it is necessary to calculate the weights of the credibility of different virtual observations in the three overall, and the credibility of the three virtual observations can be obtained as follows.

Among them, TE _T , TE _P , and TE _W can be estimated according to the principle of transfer entropy, and K _T , K _P , and K _W satisfy the following formula.

K _t +K _p +K _w =1

Use the estimated confidence of different virtual observations and use them as the weights of virtual observations for reconstruction. The reconstructed measurement results are shown in the following formula.

2) Secondly, aiming at the weight convergence problem in the fusion algorithm, the exponential forgetting method is introduced to combine the advantages of TE's good weight prediction ability and AWF's weight adaptive adjustment, which improves the dynamic performance of the error compensation process.

In the TE-ANN model, although the transfer entropy has a good ability to predict the weight value, the weight value cannot be adjusted according to the error so as to gradually converge, resulting in the error distribution also fluctuating and unable to converge. The AWF adaptive weighted fusion algorithm has the advantage that the weights can gradually converge under the criterion of the minimum mean square error without any prior knowledge. In order to combine the advantages of TE's good weight estimation and AWF's weight adaptive adjustment, a forgetting mechanism is introduced.

Based on the confidence weight K estimated by the transfer entropy and the optimal weight factor W based on the minimum mean square error criterion, the weight coefficient {βn} is selected, and K and W are fused, as shown in the following formula.

Among them, n represents the fusion of the nth observation value.

In view of the stable prediction characteristics of the transfer entropy and the weight adaptive adjustment and convergence characteristics of AWF, the initial dynamic process of the weight value should focus on Kn, while the steady-state process should focus on Wn*, and gradually converge to Wn *. In order to reflect the above characteristics, the weighting coefficient {βn} needs to meet the following characteristics:

In order to satisfy the above conditions, construct the following function.

d _n =(1-b)/(1-a·b ⁿ ), n=1, 2, 3L

Among them, b is the forgetting factor, a is the attenuation factor, 0<b<1<a. The weighting coefficient can be obtained from condition 28:

β _n ^K = d _n , β _n ^W = 1-d _n

From this, the weight of each channel can be obtained, as shown in the formula:

Among them, Wn^ satisfies the following formula, and the obtained new weight Wn^ is used as a new weight to replace the original Wn*.

The above examples are provided only for the purpose of describing the present invention, not to limit the scope of the present invention. The scope of the invention is defined by the appended claims. Various equivalent replacements and modifications made without departing from the spirit and principle of the present invention shall fall within the scope of the present invention.

Claims (4)

1. Based on the TE-ANN-AWF mobile pollution source telemetry error compensation method, the method specifically includes the following steps: Step 1: Obtain measurement samples under different interference effects through environmental simulation experiments; based on experimental measurement samples, conduct interference correlation analysis through TE transfer entropy, so as to determine the source of measurement error and measure the imbalance between multiple interferences; and use TE The directionality of the transfer entropy leads to the quantitative standard and judgment method of the non-significant causal relationship; Step 2: Obtain a training sample set of a single interference channel by using the environmental simulation smoke chamber experimental platform, and establish a measurement error prediction model for each interference through the neural network ANN method; Step 3: Realize the multivariate deconstruction of the unit observation sequence through the virtual observation method, perform error compensation on the deconstructed multivariate virtual observation sequence through the ANN error prediction model under different disturbances, and then use the multivariate adaptive weighted fusion method to correct the compensated multivariate Fusion and reconstruction of virtual observation sequence; Aiming at the weight convergence problem in the fusion algorithm, the exponential forgetting method is introduced in the model to combine the weight estimation ability of TE with the advantages of adaptive weight adjustment of the multivariate adaptive weighted fusion method, which improves the error compensation process. dynamic performance. 2. The telemetry error compensation method based on TE-ANN-AWF mobile pollution source according to claim 1, characterized in that: in the step 1, for the detectable characteristics of external environment interference, TE transfer entropy is introduced to interfere with remote sensing measurement Carry out correlation causality analysis, and use the directionality of transfer entropy to elicit quantitative standards and judgment methods for non-significant causality; Assuming that the interferences are independent to a certain extent, the measurement sequence under the control of single interference changes is obtained through the simulation experiment platform, and the transfer entropy TE _T->CO of the temperature interference to the measured value is calculated, and the transfer entropy TE of the humidity interference to the measured value is calculated. _H->CO , transfer entropy TE _P->CO from air pressure disturbance to measured value, transfer entropy TE _W->CO from wind speed disturbance factor to measured value, among them take the largest reverse transfer entropy TE ₀ as the measure of non-causality standard; TE ₀ =max{TE _CO->T , TE _CO->H , TE _CO->P , TE _CO->W } (1). 3. The telemetry error compensation method based on TE-ANN-AWF mobile pollution sources according to claim 1, characterized in that: in the step 3, the multivariate adaptive weighted fusion method is used to fuse and recompensate the multivariate virtual observation sequence after compensation The specific structure is as follows: under the criterion of the minimum mean square error, the multivariate observation values of each interference channel are adaptively fused; through the virtual observation method and the exponential forgetting mechanism, the TE transfer entropy, the ANN artificial neural network, and the AWF adaptive weighted fusion are combined. The methods are closely integrated with each other. 4. The telemetry error compensation method based on TE-ANN-AWF mobile pollution source according to claim 1, characterized in that: in the step 3, the method of exponential forgetting is introduced into the model to combine the good weight estimation ability of TE and The advantages of AWF's weight adaptive adjustment are combined, specifically: Combine the obtained confidence weight K of transfer entropy estimation and the optimal weight factor W ^* according to the minimum mean square error criterion, select the weight coefficient {β _n }, and fuse K and W to obtain As shown in the following formula; Among them, n represents the fusion of the nth observation value, Represents the weighting coefficient of the confidence weight K, K _n represents the confidence weight of the nth observation value for transfer entropy prediction, Represents the weighting coefficient of the optimal weighting factor W, W _n ^* represents the weight of the AWF after the fusion of the nth observation value; In view of the characteristics of stable estimation of transfer entropy and the adaptive adjustment and convergence of weights of AWF, the estimation value of the initial dynamic process of weights should focus on K _n , while the steady-state process should focus on W _n ^* and gradually converge in W _n ^* ; in order to reflect the above characteristics, the weighting coefficient {β _n } needs to meet the following characteristics: i represents a natural number; In order to meet the above conditions, construct the following function; d _n =(1-b)/(1-a·b ⁿ ), n=1, 2, 3L (4) Among them, b is the forgetting factor, a is the attenuation factor, 0<b<1<a; the weighting coefficient is obtained by formula (3): From this, the weight of each channel is obtained, as shown in the formula: