Low-credibility multi-sensor information fusion method

Abstract

The invention discloses a low-credibility multi-sensor information fusion method, which comprises the following steps of: 1) acquiring observation signals of sensors for synchronously measuring parameters in a multi-sensor system so as to obtain observation data of each sensor; 2) modeling the observation signal obtained in the step 1) into Gaussian process distribution containing noise; 3) smoothing the real observation signal of the sensor at the time point t * by using the historical data set; 4) carrying out credibility evaluation on the sensor information subjected to smooth processing; and 5) carrying out weighted fusion on the sensor information after confidence evaluation. According to the method, denoising and smoothing of sensor data are achieved by introducing Gaussian process regression, fusion weights are distributed for information of each sensor through confidence evaluation, noise interference and error accumulation are effectively weakened, data reliability and environment adaptability are improved, and therefore the method has good robustness and fusion credibility.

Description

Technical Field

[0001] The present invention relates to the technical field of sensor data fusion, and in particular to a method for fusing low-credibility multi-sensor information. Background Art

[0002] With the continuous development of technology, the algorithms of multi-sensor information fusion are becoming increasingly diverse, covering theories and technologies in multiple fields, such as uncertainty theory, estimation theory, optimization theory, fuzzy mathematics, neural networks and data mining. At present, these methods can be mainly divided into two categories: artificial intelligence and random. Among the random methods, the common ones are Kalman filtering and multi-Bayesian estimation. Kalman filtering optimizes the data input by sensors by establishing linear state equations and removes noise interference, thereby providing more accurate estimates. The multi-Bayesian estimation method is based on Bayesian estimation of different sensors, comprehensively processes multiple types of information, and finally accurately describes the target characteristics through joint probability distribution and minimum likelihood estimation. In artificial intelligence methods, such as neural network methods, by using their powerful parallel computing capabilities, complex modeling steps are avoided. Their diversity and flexibility make them superior in different scenarios and data processing requirements, and they can quickly identify and process large-scale data, and are widely used in the fusion of multi-sensor information.

[0003] In low-confidence environments, sensors may experience intermittent failures, abnormal measurements, or data loss due to external interference, hardware aging, calibration errors, or the complexity of the dynamic environment. These problems increase the uncertainty of sensor data, making the measurement results full of noise, abnormal values, or missing data, thus affecting the credibility and accuracy of the data. In low-confidence environments, sensor data may also exhibit complex nonlinear characteristics due to factors such as environmental dynamics, device differences, and measurement noise. Existing multi-sensor information fusion methods still face many challenges when fusing multi-sensor system information in low-confidence acquisition environments. Stochastic methods (such as Kalman filtering and multi-Bayesian estimation) are prone to bias in fusion results and increased computational complexity when sensor weights are unevenly distributed and the internal noise of the sensor system is too large. Although artificial intelligence methods (such as neural networks) have powerful parallel processing capabilities, they may have problems such as data processing bias, gradient instability, and imbalanced fusion effects when the range of multi-dimensional data is significantly different.

[0004] At present, there are many literatures at home and abroad that have conducted research on sensor data fusion in low-credibility environments and proposed a series of effective solutions.

[0005] 1. In the article entitled "Information fusion algorithms for state estimation in multi-sensor systems with correlated missing measurements", the author R. Caballero-aguila proposed a data fusion method combining recursive linear filtering and matrix weighting to solve the problem of missing observations under low-confidence acquisition data, thereby optimizing the distributed information fusion estimator. However, when this method is used in a multi-sensor acquisition system, there is a problem that the range distribution of the collected multi-dimensional data is too different, which may cause the data collected by different sensors to be unable to be effectively matched during the fusion process, affecting the credibility of the fusion results.

[0006] 2. In the article titled "An efficient intelligent data fusion algorithm for wireless sensor network", the author Haitao Wang proposed an intelligent data fusion algorithm called Gapeobp, which optimizes the data processing process of the sensor network by integrating BP neural network and particle swarm optimization algorithm, reduces low-quality data transmission, and improves fusion accuracy. However, when this method is used in a low-reliability acquisition environment, there will be a lot of noise and errors in the collected raw data, which will increase the uncertainty in the data fusion process and affect the reliability and stability of the fusion results. Summary of the invention

[0007] In view of this, the purpose of the present invention is to provide a low-credibility multi-sensor information fusion method to solve the technical problem of insufficient information fusion credibility in existing multi-sensor information fusion methods for fusing information collected by multi-sensor systems in low-credibility collection environments.

[0008] The low-reliability multi-sensor information fusion method of the present invention is characterized by comprising the following steps:

[0009] 1) Collect the observation signals of n sensors that synchronously measure the same parameters in the multi-sensor system, and obtain the observation signals of each sensor in the time series t=t1,t2,…,t T The observed signal in:

[0010] X i (t1),X i (t2),…,X i (t T )(1);

[0011] 2) Model the observed signal obtained in step 1) as a Gaussian process distribution containing noise:

[0012] Assume that the sensor's observation signal is generated by a Gaussian process in the input time domain, that is:

[0013]

[0014] Wherein, m(t) represents the mean function; k(t′, t″) is the kernel function, which is used to describe the correlation between any time points t′ and t″; the kernel function is expressed as:

[0015]

[0016] in is the variance scale, is the length scale of the kernel function, which controls the amplitude and correlation speed of the signal respectively; then the observed signal X i (t) is modeled as a real signal f i (t) and noise ∈ t The superposition of:

[0017]

[0018] Among them, ∈ t is independent noise of Gaussian distribution, is the noise variance;

[0019] Based on the historical time point t=(t1, t2,…, t T ) and the corresponding observation signal X i =(X i (t1),X i (t2),…,X i (t T ), the Gaussian process assumes that the real signal f i The prior distribution of (t) is:

[0020]

[0021] Where m is a mean vector, each element of which consists of the value of the mean function m(t) at the corresponding time point, that is, m = [m(t1), m(t2), …, m(t T )] T ; The covariance matrix K is a T×T matrix, which is composed of the kernel function k(t′,t″), that is:

[0022] K t′t″ =k(t′,t″) (6)

[0023] After considering the noise, the distribution of the observed signal is:

[0024]

[0025] Where I is the identity matrix;

[0026] 3) Using the data set (t, X i ) for the sensor at time t * The real observed signal is smoothed, including:

[0027] 1) Establish time point t * The real observed signal f i (t * ) and known data (t, X i ), the joint distribution is in the form of:

[0028]

[0029] Among them, k * is the covariance vector calculated by the kernel function, representing the time point t * Correlation with historical time points:

[0030] k * =[k(t * ,t1),k(t * ,t2),...,k(t * ,t T )] T (9)

[0031] 2) Based on the observation data X of the sensor at the historical time point i and the sensor at time t * The real observed signal f i (t * ), derive f i (t * )’s posterior distribution:

[0032]

[0033] in:

[0034]

[0035] in, is the conditional mean, indicating that the i-th sensor is at time point t * The smoothed estimate of the observed signal, Var(f i (t * )) is the conditional variance, which indicates the uncertainty of the model prediction;

[0036] Finally, take the conditional mean As the i-th sensor at t* The smoothed value of the observed signal at time , denoted as

[0037]

[0038] 4) Conduct credibility assessment on the smoothed sensor information, including:

[0039] The smoothed value x of the observation signal of sensor No. i i and the smoothed value x of the observation signal of sensor j j , define the trust function b ij To quantify the trustworthiness between sensor data, the trust function takes the following form:

[0040]

[0041] Among them, b ij Represents the smoothed value x i For the smoothed value x j When the data difference is small, the trust level b ij The higher the value, the lower the trust value. On the contrary, when the data difference is large, the trust value decreases, and when it exceeds the threshold value M, the trust value directly becomes zero.

[0042] Based on b ij , construct the trust matrix B, whose dimension is n×n, which represents the mutual trust degree between n sensors:

[0043]

[0044] By calculating the sum of the elements in each row of the matrix To evaluate the comprehensive credibility of the observation signal of the i-th sensor; if the sum is large, it means that the observation signal of the i-th sensor is trusted by most other sensors;

[0045] 5) Weighted fusion of sensor information after confidence evaluation:

[0046] The weighted fusion model is expressed as:

[0047]

[0048] Among them, w i For each set of smoothed data x i Assign weights, the weights w i It is used to reflect the comprehensive credibility of this group of smoothed data in all data, and the weight w i satisfy The weight w is calculated as follows i :

[0049] First, let the maximum eigenvalue of the trust matrix B be λ, and the corresponding eigenvector be A = [a1, a2, …, a n ] T , the eigenvalues ​​and eigenvectors satisfy the following relationship:

[0050] λA=BA (16)

[0051] By extracting the components a1, a2, ..., a in the eigenvector n , and obtain the initial weight of each smoothed data; then, these weight coefficients are normalized to calculate the weight w of each smoothed data i :

[0052]

[0053] Finally, the weight w i Substitute into the weighted fusion model and calculate the fusion result of sensor information:

[0054]

[0055] 5) Optimize the weight w in the weighted fusion model by using an improved genetic algorithm i ,include:

[0056] (1) Coding:

[0057] Using decimal encoding, the weights of n sensor data are combined into a random number sequence w1, w2,::::, w n , taking random number sequences as chromosomes, each random number sequence corresponds to an individual in the population;

[0058] (2) Set the initial population of the genetic algorithm:

[0059] Use the improved circle algorithm to set the initial population of the genetic algorithm, that is, for the initial circle,

[0060] C=π1…π u-1 π u π u+1 …π v π v+1 …π n ,1≤u≤v≤n,π u ≤π v ≤n; (19)

[0061] Swap the order of u and v, and the new path is:

[0062] π1…π u-1 π v π u+1 …π v-1 π u πv+1 …π n ,(20)

[0063] Recorded as:

[0064]

[0065] (3) Fitness evaluation:

[0066] The real value With fusion value The absolute value of the error between is defined as the objective function, that is:

[0067]

[0068] The objective function is used as the fitness function of the genetic algorithm;

[0069] (4) Improve the crossover operation of the genetic algorithm: First, the parent individuals are sorted according to the fitness function value, the individuals with small fitness values ​​are paired with individuals with small fitness values, and the individuals with large fitness values ​​are paired with individuals with large fitness values. Then, the Logistic chaotic sequence is used to determine the location of the intersection and perform the crossover operation.

[0070] x(n+1)=4x(n)(1-x(n)) (23)

[0071] The crossover operation steps include: taking a random initial value (0, 1), iterating once using formula 23), generating a chaotic value on (0, 1), and saving the chaotic value, using the value as the initial value of the chaotic iteration to generate the next generation of crossover terms, and then multiplying the initial value by the number of sensor nodes n, and finally rounding off to get the crossover operator;

[0072] (5) Improve the mutation operation of genetic algorithm:

[0073] According to the given mutation rate, an integer between 1 and n is randomly selected, and the gene mutation is performed at the position corresponding to these two numbers; the current gene value is used as the initial value, and the Logistic chaotic sequence is used for iteration to obtain the new gene value after mutation, thereby obtaining a new chromosome;

[0074] (6) Perform selection operation:

[0075] The number of individuals in each generation is set to be equal, and the individuals are arranged according to the size of the fitness. The roulette wheel selection method is used for selection. The probability of each individual being selected is proportional to the value of its fitness function. After the selection is completed, the above steps (4)-(6) are repeated until the number of iterations is reached, the evolution is terminated, and finally the optimal weight of each sensor data is obtained.

[0076] 6) Bring the optimal weight into the weighted fusion model to calculate the optimal fusion result of the sensor information.

[0077] Beneficial effects of the present invention:

[0078] 1. In a low-confidence environment, sensors may experience intermittent failures, abnormal measurements, or data loss due to external interference, hardware aging, calibration errors, or the complexity of the dynamic environment. These problems increase the uncertainty of sensor data, making the measurement results full of noise, abnormal values, or missing data, thus affecting the credibility and accuracy of the data.

[0079] The low-credibility multi-sensor information fusion method of the present invention realizes denoising and smoothing of sensor data by introducing Gaussian process regression, and assigns fusion weights to the information of each sensor through confidence evaluation, which can effectively reduce noise interference and error accumulation in a low-credibility acquisition environment, and improve the reliability of data and environmental adaptability. Therefore, the method of the present invention has good robustness and fusion credibility when fusing sensor information collected in a low-credibility environment.

[0080] 2. In a low-trust environment, sensor data may exhibit complex nonlinear characteristics due to factors such as environmental dynamics, device differences, and measurement noise. The weights calculated by the trust matrix proposed by the method of the present invention alone may not fully consider these complexities, resulting in unreasonable weight distribution. Therefore, the method of the present invention further adjusts the weight coefficient globally by introducing an improved genetic algorithm based on the obtained weights.

[0081] In the complex multi-dimensional data fusion process, the uncertainty within the multi-sensor system and the interaction between the collected data become important factors affecting the results. Although the mutation operation in the traditional genetic algorithm introduces external random interference such as changes in Gaussian distribution, it fails to effectively reflect the uncertainty within the system and lacks a direct connection with the multi-dimensional interaction in data fusion. The present invention adopts an improved genetic algorithm to improve the operation of the genetic algorithm so that it can dynamically adapt to the uncertainty and interdependence between the data in the fusion, improve the optimization effect, make the fusion result of the method of the present invention more in line with the actual situation, and further improve the robustness of the information fusion method of the present invention in a low-credibility environment and the credibility of data fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 Schematic diagram of the data layer fusion process.

[0083] Figure 2 Flowchart for optimizing weights for improved genetic algorithm. DETAILED DESCRIPTION

[0084] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0085] The low-reliability multi-sensor information fusion method in this embodiment includes the following steps:

[0086] 1) Collect the observation signals of n sensors that synchronously measure the same parameters in the multi-sensor system, and obtain the observation signals of each sensor in the time series t=t1,t2,…,t T The observed signal in:

[0087] X i (t1),X i (t2),…,X i (t T )(1).

[0088] 2) Model the observed signal obtained in step 1) as a Gaussian process distribution containing noise:

[0089] Given any set of time points t=(t1, t2,…, t r ), the Gaussian process assumes that the function values ​​at these time points are f(t) = (f(t1), f(t2), …, f(t r )) satisfies the multivariate Gaussian distribution. In this embodiment, it is assumed that the observation signal of the sensor is generated by a Gaussian process in the input time domain, that is:

[0090]

[0091] Wherein, m(t) represents the mean function; k(t′, t″) is the kernel function, which is used to describe the correlation between any time points t′ and t″; the kernel function is expressed as:

[0092]

[0093] in is the variance scale, is the length scale of the kernel function, which controls the amplitude and correlation speed of the signal respectively; then the observed signal X i (t) is modeled as a real signal f i (t) and noise ∈ t The superposition of:

[0094]

[0095] Among them, ∈ t is independent noise of Gaussian distribution, is the noise variance.

[0096] Based on the historical time point t=(t1, t2,…, t T ) and the corresponding observation signal X i =(X i (t1),Xi (t2),…,X i (t T ), the Gaussian process assumes that the real signal f i The prior distribution of (t) is:

[0097]

[0098] Where m is a mean vector, each element of which consists of the value of the mean function m(t) at the corresponding time point, that is, m = [m(t1), m(t2), …, m(t T )] T ; The covariance matrix K is a T×T matrix, which is composed of the kernel function k(t′,t″), that is:

[0099] K t′t″ =k(t′,t″) (6)

[0100] After considering the noise, the distribution of the observed signal is:

[0101]

[0102] Where I is the identity matrix.

[0103] 3) Using the data set (t, X i ) for the sensor at time t * The real observed signal is smoothed, including:

[0104] 1) Establish time point t * The real observed signal f i (t * ) and known data (t, X i ), the joint distribution is in the form of:

[0105]

[0106] Among them, k * is the covariance vector calculated by the kernel function, representing the time point t * Correlation with historical time points:

[0107] k * =[k(t * ,t1),k(t * ,t2),...,k(t * ,t T )] T (9).

[0108] 2) Based on the observation data X of the sensor at the historical time point i and the sensor at time t *The real observed signal f i (t * ), derive f i (t * )’s posterior distribution:

[0109]

[0110] in:

[0111]

[0112] in, is the conditional mean, indicating that the i-th sensor is at time point t * The smoothed estimate of the observed signal, Var(f i (t * )) is the conditional variance, which indicates the uncertainty of the model prediction;

[0113] Finally, take the conditional mean As the i-th sensor at t * The smoothed value of the observed signal at time , denoted as

[0114]

[0115] 4) Conduct credibility assessment on the smoothed sensor information, including:

[0116] The smoothed value x of the observation signal of sensor No. i i and the smoothed value x of the observation signal of sensor j j , define the trust function b ij To quantify the trustworthiness between sensor data, the trust function takes the following form:

[0117]

[0118] Among them, b ij Represents the smoothed value x i For the smoothed value x j When the data difference is small, the trust level b ij On the contrary, when the data difference is large, the trust decreases, and when it exceeds the threshold M, the trust is directly zero.

[0119] Based on b ij , construct the trust matrix B, whose dimension is n×n, which represents the mutual trust degree between n sensors:

[0120]

[0121] By calculating the sum of the elements in each row of the matrix To evaluate the comprehensive credibility of the observation signal of the i-th sensor; if the sum is large, it means that the observation signal of the i-th sensor is trusted by most other sensors.

[0122] 5) Weighted fusion of sensor information after confidence evaluation:

[0123] The weighted fusion model is expressed as:

[0124]

[0125] Among them, w i For each set of smoothed data x i Assign weights, the weights w i It is used to reflect the comprehensive credibility of this group of smoothed data in all data, and the weight w i satisfy 0 <w i <1, the weight w is calculated by the following method i :

[0126] First, let the maximum eigenvalue of the trust matrix B be λ, and the corresponding eigenvector be A = [a1, a2, …, a n ] T , the eigenvalues ​​and eigenvectors satisfy the following relationship:

[0127] λA=BA (16)

[0128] By extracting the components a1, a2, ..., a in the eigenvector n , and obtain the initial weight of each smoothed data; then, these weight coefficients are normalized to calculate the weight w of each smoothed data i :

[0129]

[0130] Finally, the weight w i Substitute into the weighted fusion model and calculate the fusion result of sensor information:

[0131]

[0132] Fusion results It represents the weighted average of all sensor data with the same parameter and is the optimal estimate after comprehensively considering the credibility of each sensor data.

[0133] As an improvement to the above embodiment, the low-confidence multi-sensor information fusion method further includes the following steps:

[0134] 5) Optimizing the weight w in the weighted fusion model by using an improved genetic algorithm i ,include:

[0135] (1) Coding:

[0136] Using decimal encoding, the weights of n sensor data are combined into a random number sequence w1, w2,::::, w n , taking the random number sequence as the chromosome, where 0 <w i <1,(i=1,2,:::,n); each random number sequence corresponds to an individual in the population.

[0137] (2) Set the initial population of the genetic algorithm:

[0138] Use the improved circle algorithm to set the initial population of the genetic algorithm, that is, for the initial circle,

[0139] C=π1…π u-1 π u π u+1 …π v π v+1 …π n ,1≤u≤v≤n,π u ≤π v ≤n; (19)

[0140] Swap the order of u and v, and the new path is:

[0141] π1…π u-1 π v π u+1 …π v-1 π u π v+1 …π n ,(20)

[0142] Recorded as:

[0143]

[0144] (3) Fitness evaluation:

[0145] The real value With fusion value The absolute value of the error between is defined as the objective function, that is:

[0146]

[0147] The objective function is used as the fitness function of the genetic algorithm.

[0148] (4) Improve the crossover operation of the genetic algorithm: First, the parent individuals are sorted according to the fitness function value, the individuals with small fitness values ​​are paired with individuals with small fitness values, and the individuals with large fitness values ​​are paired with individuals with large fitness values. Then, the Logistic chaotic sequence is used to determine the location of the intersection and perform the crossover operation.

[0149] x(n+1)=4x(n)(1-x(n)) (23)

[0150] The crossover operation steps include: taking a random initial value (0, 1), iterating once using formula 23), generating a chaotic value on (0, 1), and saving the above chaotic value, using this value as the initial value of the chaotic iteration to generate the next generation of crossover terms, and then multiplying the initial value by the number of sensor nodes n, and finally rounding it off to get the crossover operator.

[0151] (5) Perform mutation operation:

[0152] According to the given mutation rate (usually the probability of mutation is small, and the mutation rate is selected as 0.01 in this embodiment), an integer between 1 and n is randomly selected, and gene mutation is performed at the positions corresponding to these two numbers; the current gene value is used as the initial value, and the Logistic chaotic sequence is used to iterate to obtain the new gene value after mutation, thereby obtaining a new chromosome.

[0153] (6) Perform selection operation:

[0154] The number of individuals in each generation is set to be equal, and the individuals are arranged according to the size of the fitness. The roulette wheel selection method is used for selection. The probability of each individual being selected is proportional to the value of its fitness function. After the selection is completed, the above steps (4)-(6) are repeated until the number of iterations is reached, the evolution is terminated, and finally the optimal weight of each sensor data is obtained.

[0155] 6) Bring the optimal weight into the weighted fusion model to calculate the optimal fusion result of the sensor information.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A low-confidence multi-sensor information fusion method, characterized in that The following steps are involved: 1) Collect the observation signals of n sensors that synchronously measure the same parameters in the multi-sensor system, and obtain the observation signals of each sensor in the time series t=t1,t2,…,t T The observed signal in: X i (t1),X i (t2),…,X i (t T )(1); 2) Model the observed signal obtained in step 1) as a Gaussian process distribution containing noise: Assume that the sensor's observation signal is generated by a Gaussian process in the input time domain, that is: Wherein, m(t) represents the mean function; k(t′, t″) is the kernel function, which is used to describe the correlation between any time points t′ and t″; the kernel function is expressed as: in is the variance scale, l is the length scale of the kernel function, which respectively control the amplitude and correlation speed of the signal; then the observed signal X i (t) is modeled as a real signal f i (t) and noise ∈ t The superposition of: Among them, ∈ t is independent noise of Gaussian distribution, is the noise variance; Based on the historical time point t=(t1, t2,…, t T ) and the corresponding observation signal X i =(X i (t1),X i (t2),…,X i (t T ), the Gaussian process assumes that the real signal f i The prior distribution of (t) is: Where m is a mean vector, each element of which consists of the value of the mean function m(t) at the corresponding time point, that is, m = [m(t1), m(t2), …, m(t T )] T ; The covariance matrix K is a T×T matrix, which is composed of the kernel function k(t′,t″), that is: K t′t″ =k(t′,t″) (6) After considering the noise, the distribution of the observed signal is: Where I is the identity matrix; 3) Using the dataset (t, X i ) for the sensor at time t * The real observed signal is smoothed, including: 1) Establish time point t * The real observed signal f i (t * ) and known data (t, X i ), the joint distribution is in the form of: Among them, k * is the covariance vector calculated by the kernel function, representing the time point t * Correlation with historical time points: k * =[k(t * ,t1),k(t * ,t2),…,k(t * ,t T )] T (9); 2) Based on the observation data X of the sensor at the historical time point i and the sensor at time t * The real observed signal f i (t * ), derive f i (t * )’s posterior distribution: in: in, is the conditional mean, indicating that the i-th sensor is at time point t * The smoothed estimate of the observed signal, Var(f i (t * )) is the conditional variance, which indicates the uncertainty of the model prediction; Finally, take the conditional mean f i (t * ) as the i-th sensor at t * The smoothed value of the observed signal at time , denoted as 4) Conduct credibility assessment on the smoothed sensor information, including: The smoothed value x of the observation signal of sensor No. i i and the smoothed value x of the observation signal of sensor j j , define the trust function b ij To quantify the trustworthiness between sensor data, the trust function takes the following form: Among them, b ij Represents the smoothed value x i For the smoothed value x j When the data difference is small, the trust level b ij The higher the value, the lower the trust value. On the contrary, when the data difference is large, the trust value decreases, and when it exceeds the threshold value M, the trust value directly becomes zero. Based on b ij , construct the trust matrix B, whose dimension is n×n, which represents the mutual trust degree between n sensors: By calculating the sum of the elements in each row of the matrix To evaluate the comprehensive credibility of the observation signal of the i-th sensor; if the sum is large, it means that the observation signal of the i-th sensor is trusted by most other sensors; 5) Weighted fusion of sensor information after confidence evaluation: The weighted fusion model is expressed as: Among them, w i For each set of smoothed data x i Assign weights, the weights w i It is used to reflect the comprehensive credibility of this group of smoothed data in all data, and the weight w i satisfy 0 <w i <1, the weight w is calculated by the following method i : First, let the maximum eigenvalue of the trust matrix B be λ, and the corresponding eigenvector be A = [a1, a2, …, a n ] T , the eigenvalues ​​and eigenvectors satisfy the following relationship: λA=BA (16) By extracting the components a1, a2, ..., a in the eigenvector n , and obtain the initial weight of each smoothed data; then, these weight coefficients are normalized to calculate the weight w of each smoothed data i : Finally, the weight w i Substitute into the weighted fusion model and calculate the fusion result of sensor information:

2. The low-confidence multi-sensor information fusion method according to claim 1 is characterized in that The following steps are also included: 5) Optimizing the weight w in the weighted fusion model by using an improved genetic algorithm i ,include: (1) Coding: Using decimal encoding, the weights of n sensor data are combined into a random number sequence w1, w2,::::, w n , taking random number sequences as chromosomes, each random number sequence corresponds to an individual in the population; (2) Set the initial population of the genetic algorithm: Use the improved circle algorithm to set the initial population of the genetic algorithm, that is, for the initial circle, C=π1…π u-1 p u p u+1 …p v p v+1 …p n ,1≤u≤v≤n,π u ≤π v ≤n; (19) Swap the order of u and v, and the new path is: π1…π u-1 π v π u+1 …π v-1 π u π v+1 …π n ,(20) Recorded as: (3) Fitness evaluation: Combine the real value Y with the fused value The absolute value of the error between is defined as the objective function, that is: The objective function is used as the fitness function of the genetic algorithm; (4) Improve the crossover operation of the genetic algorithm: First, the parent individuals are sorted according to the fitness function value, the individuals with small fitness values ​​are paired with individuals with small fitness values, and the individuals with large fitness values ​​are paired with individuals with large fitness values. Then, the Logistic chaotic sequence is used to determine the location of the intersection and perform the crossover operation. x(n+1)=4x(n)(1-x(n)) (23) The crossover operation steps include: taking a random initial value (0, 1), iterating once with formula (23), generating a chaotic value on (0, 1), and saving the chaotic value, using the value as the initial value of the chaotic iteration to generate the next generation of crossover terms, and then multiplying the initial value by the number of sensor nodes n, and finally rounding to get the crossover operator; (5) Perform mutation operation: According to the given mutation rate, an integer between 1 and n is randomly selected, and the gene mutation is performed at the position corresponding to these two numbers; the current gene value is used as the initial value, and the Logistic chaotic sequence is used for iteration to obtain the new gene value after mutation, thereby obtaining a new chromosome; (6) Perform selection operation: The number of individuals in each generation is set to be equal, and the individuals are arranged according to the size of the fitness. The roulette wheel selection method is used for selection. The probability of each individual being selected is proportional to the value of its fitness function. After the selection is completed, the above steps (4)-(6) are repeated until the number of iterations is reached, the evolution is terminated, and finally the optimal weight of each sensor data is obtained. 6) Bring the optimal weight into the weighted fusion model to calculate the optimal fusion result of the sensor information.

Images