Multi-sensor data fusion method based on improved DS evidence theory

Abstract

This application relates to a multi-sensor data fusion method based on an improved D-S evidence theory. The method is used to realize data fusion of multiple inertial sensors in a micro inertial measurement unit; the method is implemented by Jousselme Distance is used to measure the similarity or difference between pieces of evidence, and the cosine of the angle between pieces of evidence is used to characterize the similarity between pieces of evidence, thereby achieving the correction of evidence. By introducing the membership function in fuzzy theory, the credibility of each inertial sensor is calculated, and then the average credibility is calculated from the credibility of all inertial sensors. Inertial sensors with credibility lower than the average credibility are eliminated. Finally, the basic probability distribution of each inertial sensor is calculated using the updated membership matrix and weight coefficients, realizing an improved data fusion combination strategy while correcting the evidence, thereby effectively alleviating the problems caused by evidence conflicts and improving the measurement accuracy of the micro inertial measurement unit.

Description

Technical Field

[0001] The present application relates to the technical field of data fusion, and in particular to a multi-sensor data fusion method based on an improved DS evidence theory. Background Art

[0002] In recent years, with the rapid development of microelectronics and MEMS (microelectromechanical systems) technologies, micro inertial measurement units (MIMUs), with their miniaturized packaging, low power consumption, and high cost-effectiveness, have found widespread and in-depth application in numerous fields, including wearable devices, autonomous driving, robotics, and the Internet of Things. Utilizing integrated circuit (IC)-compatible batch processing technology, the MIMU ingeniously combines electronic components with mechanical devices to form a unique microsystem. Primarily composed of a triaxial gyroscope and a triaxial accelerometer, this system can accurately measure the angular and displacement changes of the measured object in real time. With the continuous advancement of sensor technology, data fusion algorithms are becoming increasingly crucial in complex and ever-changing application environments.

[0003] In complex application environments, a single sensor can hardly meet detection requirements. Traditional single sensors have significant limitations in measurement accuracy and stability. In contrast, by collaborating and complementing each other, multiple sensors can comprehensively analyze and process redundant information and complementary data, resulting in more accurate decision-making. Data fusion algorithms, by integrating the complementary characteristics of MIMUs with other sensor types (such as GPS, visual sensors, and magnetometers), can effectively reduce errors and improve the accuracy of positioning, navigation, and attitude estimation. However, current data fusion systems are still limited by the lack of a universal theoretical framework. Existing algorithms are typically developed for specific application areas, and the appropriate fusion method must be selected based on the specific application scenario.

[0004] Multi-sensor data fusion is essentially a specific application of multi-source information fusion. Its core purpose is to extract characteristic information from data collected by multiple sensors and eliminate redundancy, thereby improving system accuracy and reliability. Because data from different sensors often exhibit temporal and spatial variations and uncertainties, fusion technology can be used to comprehensively process these various types of information, minimizing errors and noise.

[0005] DS evidence theory can effectively fuse evidence from multiple inertial sensors or different data sources when there is a conflict. After years of development, DS evidence theory provides a powerful framework for the fusion of multi-source data, but it also has some problems. For example, when the conflict coefficient of the combination rule leads to mutual exclusion between evidences, DS evidence theory cannot be used for fusion, which to some extent limits its wide range of practical applications. The problems are mainly divided into evidence conflict, mass Function selection, etc.

[0006] (1) Conflict of evidence

[0007] Conflict coefficient k The synthesis of evidence in DS theory presents numerous challenges, with conflicting evidence being a primary concern for researchers. When conflicting evidence is used, the resulting fusion can lead to erroneous conclusions. This situation is known as the Zadeh paradox, the core idea being that even if all sources of evidence provide high confidence and clear support, the fusion process can produce a significantly erroneous or even contradictory conclusion.

[0008] (2) mass Function selection

[0009] In DS evidence theory, the key step is mass Choice of function, different application problems, mass The choice of function is also different. If the choice is not appropriate, it may cause conflicts between evidences, thus affecting the fusion results. In order to deal with the above situation, it is usually mass The function is mapped to the interval [0,1] to facilitate subsequent processing. mass The specific function selection methods can be divided into two categories: one is to use mathematical calculations to select data with linear or nonlinear characteristics for processing; the other is to use neural networks to generate mass Function, after repeated training, adjusts the data range to ensure that it meets the fusion requirements of the [0,1] interval.

[0010] In summary, the DS evidence theory itself has great advantages in synthesizing uncertain information. However, in practical applications, the evidence conflict paradox still exists. The DS evidence theory suffers from convergence failure in high-conflict situations, resulting in low measurement accuracy of micro inertial measurement units (MIMUs). Summary of the Invention

[0011] Based on this, it is necessary to provide a multi-sensor data fusion method based on the improved DS evidence theory to address the above technical problems.

[0012] A multi-sensor data fusion method based on improved DS evidence theory is used to realize data fusion of multiple inertial sensors in a micro inertial measurement unit. The method includes:

[0013] Acquire the measurement data of multiple inertial sensors in a micro inertial measurement unit, build a recognition framework, and use the measurement data of each inertial sensor as a piece of evidence.

[0014] According to the evidence Jousselme The distance determines the normalized evidence support.

[0015] The cosine of the angle between evidence vectors is used to represent the similarity between evidences.

[0016] The evidence correction factor is determined based on the normalized evidence support and the similarity between the evidence.

[0017] Use evidence correction factor to adjust the mass The function is corrected to obtain the corrected evidence.

[0018] The revised evidence is differentiated using relative conflict factors to determine the credibility of the evidence source.

[0019] The membership matrix of the inertial sensor is calculated in the recognition framework by using the descending half-normal function as the membership function;

[0020] According to the membership matrix, the credibility of each inertial sensor is determined in the identification framework.

[0021] Eliminate evidence whose credibility is lower than a preset threshold in the identification framework to obtain a new identification framework;

[0022] According to the membership matrix of the inertial sensor in the new identification framework, the weight coefficient of each inertial sensor in the new identification framework is determined.

[0023] According to the membership matrix and weight coefficients under the new recognition framework, the basic probability distribution of each inertial sensor to the target under the new recognition framework is determined.

[0024] According to the basic probability distribution, the final fusion evidence is determined, and according to the final fusion evidence and the corresponding measurement data, the data fusion result of multiple inertial sensors in the micro inertial measurement unit is determined.

[0025] The multi-sensor data fusion method based on the improved DS evidence theory is used to realize the data fusion of multiple inertial sensors in a micro inertial measurement unit; the method is Jousselme Distance is used to measure the similarity or difference between pieces of evidence, and the cosine of the angle between pieces of evidence is used to characterize the similarity between pieces of evidence, thereby achieving the correction of evidence. By introducing the membership function in fuzzy theory, the credibility of each inertial sensor is calculated, and then the average credibility is calculated from the credibility of all inertial sensors. Inertial sensors with credibility lower than the average credibility are eliminated. Finally, the basic probability distribution of each inertial sensor is calculated using the updated membership matrix and weight coefficients, realizing an improved data fusion combination strategy while correcting the evidence, thereby effectively alleviating the problems caused by evidence conflicts and improving the measurement accuracy of the micro inertial measurement unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 1 is a flow chart of a multi-sensor data fusion method based on an improved DS evidence theory in one embodiment;

[0027] Figure 2 A diagram showing data collected by multiple groups of gyroscopes in another embodiment;

[0028] Figure 3 A comparison chart of the fusion results of various methods in another embodiment;

[0029] Figure 4 This is a fusion result diagram in another embodiment;

[0030] Figure 5 This is a diagram of data fusion results in motion in another embodiment;

[0031] Figure 6 2 is a comparison diagram before and after fusion under motion state in another embodiment, wherein (a) is a comparison diagram before and after fusion under the first motion state (4000 to 4500 groups of data during the acceleration process), and (b) is a comparison diagram before and after fusion under the second motion state (13000 to 13500 groups of data during the acceleration process). DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0033] After years of development, DS evidence theory provides a powerful framework for the integration of multi-source data and has undergone many improvements. However, it still has some problems, such as when the conflict coefficient of the combination rule is When the evidences are mutually exclusive, the DS evidence theory cannot be used for fusion, which limits the breadth of practical applications. The problems are mainly divided into evidence conflicts, mass Problems such as function selection. Conflict coefficient k There are many problems when synthesizing evidence in DS theory, and evidence conflict is the most concerned issue for researchers. When using conflicting evidence, the fusion will lead to wrong conclusions. The above situation is also called Zadeh paradox. The core idea is: even if all the evidence sources provide high confidence and clear support, after fusion, the results may be large or opposite. In DS theory, the key step is mass Choice of function, different application problems, mass The choice of function is also different. If the choice is not appropriate, it may cause conflicts between evidences, thus affecting the fusion results. In order to deal with the above situation, it is usually mass Functions are mapped to interval, which is convenient for subsequent processing. massThe specific function selection methods can be divided into two categories: one is to use mathematical calculations to select data with linear or nonlinear characteristics for processing; the other is to use neural networks to generate mass Function, after repeated training, adjust the data range to ensure compliance Interval fusion requirements.

[0034] The problems existing in DS evidence theory and the advantages and disadvantages of existing improvement methods are explained. For example, Murphy uses a weighted average strategy to deal with evidence conflicts during DS rule synthesis. However, when there are large conflicts between different evidence sources, simple weighted averaging cannot effectively resolve the conflicts, and excessive averaging may lose the characteristics of the original evidence. Yager can effectively alleviate the problem of high conflict in evidence by introducing a conflict redistribution mechanism. However, as the amount of evidence increases, the confidence of the combined result decreases towards the full set. This will cause the , so this method still faces certain challenges.

[0035] This application aims to improve the measurement accuracy of the micro inertial measurement unit (MIMU) and conducts research around the multi-sensor data fusion technology. In order to solve the problem of insufficient accuracy and reliability of a single sensor, DS evidence theory is used to fuse multi-sensor data. In view of the limitations of DS evidence theory in dealing with conflicting evidence, an improved algorithm that takes into account both evidence correction and combination rule modification is proposed. Subsequently, the membership degree in fuzzy theory is introduced, and the weight coefficient is obtained after eliminating the faulty sensor using the average credibility to complete the evidence conversion. Finally, a dynamic and static comparison test is designed: 5 groups of gyroscope data are collected in static and moving states for fusion. The fusion results show that the fusion accuracy based on this algorithm is high.

[0036] In one embodiment, Figure 1 As shown, a multi-sensor data fusion method based on an improved DS evidence theory is provided to realize data fusion of multiple inertial sensors in a micro inertial measurement unit; the method comprises the following steps:

[0037] Step 100: Obtain measurement data from multiple inertial sensors in a micro inertial measurement unit, build a recognition framework, and use the measurement data of each inertial sensor as evidence.

[0038] Step 102: Based on the evidence Jousselme The distance determines the normalized evidence support; the cosine of the angle between evidence vectors is used to characterize the similarity between evidences; and the evidence correction factor is determined based on the normalized evidence support and the similarity between evidences.

[0039] Specifically, through Jousselme Distance measures the similarity or difference between evidence. Jousselme The greater the distance, the greater the difference between the two pieces of evidence, that is, the more serious the conflict between the evidence, and vice versa. Jousselme The smaller the distance, the higher the similarity between the evidences. The cosine of the angle between the evidences is used to represent the similarity between the evidences to achieve the correction of the evidence.

[0040] use Jousselme The distance is used to judge the difference between evidences. The larger the distance, the greater the difference, and vice versa. Jousselme The distance measure is the similarity between two BPAs; the measurement matrix P is constructed using the Jaccard similarity between fuzzy sets, and the core is to obtain the Euclidean distance to be measured. Jousselme The specific expression of distance is as follows:

[0041] ;

[0042] Where: Between two pieces of evidence Jousselme distance; It refers to the i The basic probability distribution vector (BPA) of the evidence, It refers to the j The basic probability distribution vector (BPA) of the evidence, It is i Evidence of the incident The probability distribution of occurrence, It is i Evidence of the incident The probability distribution of occurrence, It is i Evidence of the incident The probability distribution of occurrence, It is i Evidence of the incident The probability distribution of occurrence, the superscript T is the transposition operation, P is a fuzzy metric matrix, P For one The positive definite matrix of is the fuzzy metric matrix i Rank j Elements of the column, , 、 The events and events Subsets in the recognition framework (for computing similarity).

[0043] Step 104: Use the evidence correction factor to adjust the massThe function is modified to obtain the modified evidence; the modified evidence is differentiated using the relative conflict factor to determine the credibility of the evidence source.

[0044] Step 106: Using the descending half normal function as the membership function, the membership matrix of the inertial sensor is calculated in the recognition framework; and according to the membership matrix, the credibility of each inertial sensor is determined in the recognition framework.

[0045] Specifically, the credibility of each inertial sensor is calculated by introducing the membership function in fuzzy theory.

[0046] The membership function is:

[0047] ;

[0048] in, For the i and j The membership function of an inertial sensor is For the j and i The membership function of an inertial sensor is 、 Respectively i 、 j The measurements of the inertial sensors, ;

[0049] The membership function is essentially a Gaussian function of Euclidean distance. The exponential function takes a maximum value of 1 when the error is greater than 1, and decays rapidly as the difference between the two increases. Sensor measurement errors typically follow a normal distribution, and the exponential function is part of the probability density function (excluding the coefficient) in the standard normal distribution, which can well simulate the decay characteristics of the error distribution.

[0050] Will As the matrix i Rank j The elements of the column are used to obtain the membership matrix of each inertial sensor measurement value in the identification framework. The membership matrix expression of each inertial sensor measurement value in the identification framework is:

[0051] ;

[0052] in, U is the membership matrix of each inertial sensor measurement value in the identification framework, m To identify the amount of evidence in the framework.

[0053] Assume there is m An inertial sensor detects a certain attribute, and the measurement value of each inertial sensor is , the recognition frame is Since the measured values ​​of the inertial sensor include true values ​​and noise, and within the normal deviation range, the measured values ​​are highly consistent with the true values, especially the error part presents a normal distribution, the descending half-normal function is selected as the membership function in this section. The formula is shown in the above membership function expression.

[0054] Membership value The size of is related to the difference between the two inertial sensor measurements. When the difference between the measurements is small, the membership value is larger; otherwise, the membership value is smaller. In the above, the membership matrix of each inertial sensor to the monitoring value is shown in the above-mentioned membership matrix expression of each inertial sensor measurement value in the identification framework.

[0055] Step 108: Eliminate evidence with a credibility lower than a preset threshold in the recognition framework to obtain a new recognition framework; and determine the weight coefficient of each inertial sensor in the new recognition framework based on the membership matrix of the inertial sensor in the new recognition framework.

[0056] Specifically, the average credibility is calculated from the credibility of all inertial sensors, and the inertial sensors with credibility lower than the average credibility are eliminated to obtain a new recognition framework.

[0057] Step 110: Determine the basic probability distribution of each inertial sensor to the target in the new recognition framework based on the membership matrix and weight coefficients in the new recognition framework.

[0058] Specifically, the updated membership matrix and weight coefficients are used to calculate the basic probability distribution of each inertial sensor, so as to improve the combined strategy of data fusion while correcting the evidence, thereby effectively alleviating the problems caused by evidence conflicts.

[0059] Step 112: Determine final fusion evidence based on the basic probability distribution, and determine the data fusion result of multiple inertial sensors in the micro inertial measurement unit based on the final fusion evidence and the corresponding measurement data.

[0060] In the above-mentioned multi-sensor data fusion method based on the improved DS evidence theory, the method is used to realize the data fusion of multiple inertial sensors in a micro inertial measurement unit; the method is achieved by JousselmeDistance measures the similarity or difference between evidence, and the cosine of the angle between evidence is used to characterize the similarity between evidence, so as to achieve the correction of evidence. By introducing the membership function in fuzzy theory, the credibility of each inertial sensor is calculated, and then the average credibility is calculated from the credibility of all inertial sensors, and inertial sensors with credibility lower than the average credibility are eliminated. Finally, the basic probability distribution of each inertial sensor is calculated using the updated membership matrix and weight coefficient, realizing an improved data fusion combination strategy while correcting the evidence, thereby effectively alleviating the problems caused by evidence conflicts.

[0061] In one embodiment, based on the evidence Jousselme The distance determines the normalized evidence support, including: the distance between the current evidence and other different evidence Jousselme The opposite of the distance is added to 1, and the distance between the current evidence and all other different evidences is calculated. Jousselme The distances are summed to obtain evidence support; and the evidence support is normalized to obtain normalized evidence support.

[0062] Specifically, the expression of normalized evidence support is:

[0063] ;

[0064] ;

[0065] in, is the normalized evidence support, Between two pieces of evidence Jousselme distance, For the evidence support, For the j The evidential support of the evidence, .

[0066] The support of evidence is negatively correlated with the distance between the evidences. The greater the distance, the lower the support.

[0067] In one embodiment, determining an evidence correction factor based on normalized evidence support and similarity between pieces of evidence includes: multiplying the normalized evidence support and similarity between pieces of evidence, and normalizing the result to obtain the evidence correction factor.

[0068] Specifically, the cosine of the angle between the evidence vectors is used to quantify the spatial relationship of the evidence vectors, and the sum of the cosine values ​​between the evidence vectors is used to represent the similarity. From mathematical knowledge, we know that when the cosine value between two pieces of evidence is When the cosine value is , the two pieces of evidence are completely consistent and there is no conflict. When the two pieces of evidence are perpendicular to each other, there is no correlation; when the cosine value When the two pieces of evidence are in completely opposite directions, the information is completely contradictory.

[0069] The similarity parameter and normalization formula are as follows:

[0070] ;

[0071] in, is the angle between two evidence vectors, is the cosine of the angle between the evidence vectors, is the similarity between evidences before normalization, is the similarity between evidences, 、 Evidence vectors are i and evidence vector j .

[0072] The support and similarity between evidences can be obtained through the similarity parameter and normalization processing formula. The product of the two is used as the evidence correction factor of this method. The specific formula and normalization processing are as follows:

[0073] ;

[0074] in, is the evidence correction factor (normalized evidence correction factor), is the evidence correction factor before normalization, For the i The reliability of a sensor.

[0075] In one embodiment, an evidence correction factor is used to adjust the value of each piece of evidence. mass The function is corrected and the corrected evidence is:

[0076] ;

[0077] in, For the revised evidence, For evidence of mass function, is the evidence correction factor (normalized evidence correction factor), is the corrected probability mass of the entire framework, represents the probability mass assigned to the entire framework, To identify the framework.

[0078] In one embodiment, the corrected evidence is differentiated using a relative conflict factor to determine the credibility of the evidence source:

[0079] ;

[0080] in, Represents the basic probability distribution after fusion, that is, the proposition The final probability mass of To identify the framework, As a source of evidence i Focus Element The quality distribution of As a source of evidence i Focus Element The quality distribution, is a conflict modifier used to add quality when evidence conflicts. As a source of evidence j Focus Element The quality distribution of As a source of evidence j Focus Element A The quality distribution, This is the standard requirement of BPA (the mass of the empty set is zero).

[0081] Specifically, after obtaining the corrected evidence, in order to further improve the fusion accuracy, this method proposes the concept of relative conflict factor, which uses differentiated processing to more flexibly determine the credibility of the evidence source, thereby minimizing the negative impact of low-credibility evidence on the final fusion result.

[0082] In one embodiment, determining the credibility of each inertial sensor in the identification framework based on the membership matrix includes: performing arithmetic averaging on all elements in each row of the membership matrix of the inertial sensor measurement values ​​in the identification framework to obtain the credibility of each inertial sensor in the identification framework.

[0083] Specifically, the credibility expression of the inertial sensor is:

[0084] ;

[0085] Where: For the i The reliability of the inertial sensors, m To identify the amount of evidence in a frame, For the i and j The membership function of an inertial sensor.

[0086] In one embodiment, a weight coefficient of each inertial sensor in the new identification framework is determined based on the membership matrix of the inertial sensor in the new identification framework, including: determining the credibility of each inertial sensor based on the membership matrix of the inertial sensor in the new identification framework; normalizing the credibility of the inertial sensor and using the normalized result as the weight coefficient of each inertial sensor.

[0087] In one embodiment, a basic probability distribution of each inertial sensor to a target in the new recognition framework is determined based on a membership matrix and a weight coefficient in the new recognition framework, including: in the new recognition framework, weighting the membership of the measurement value of each inertial sensor with the weight coefficient to obtain a weighted membership of the measurement value of each inertial sensor; and using the ratio of the weighted membership of the measurement value of each inertial sensor to the sum of the weighted memberships of the measurement values ​​of all inertial sensors as the basic probability distribution of the corresponding inertial sensor to the target.

[0088] Specifically, in the recognition framework Next, i The reliability of each inertial sensor is shown in the above reliability expression of the inertial sensor. After obtaining the reliability of each inertial sensor, the average reliability of all measurement units can be obtained. Since the errors of different inertial sensors are different, and even inertial sensor failure occurs, it is necessary to eliminate the inertial sensors with the above situation. The average reliability Defined as:

[0089] ;

[0090] in, is the average reliability.

[0091] New recognition framework after removing the ones below average credibility for:

[0092] ;

[0093] In the new identification framework Next, calculate the remaining n The membership matrix of the inertial sensors , and calculate the credibility of each inertial sensor based on the new membership matrix, and normalize the credibility to obtain the weight coefficient. The formula is as follows:

[0094] ;

[0095] ;

[0096] in, In the new identification framework Down n The membership matrix of the inertial sensors, n New identification framework The number of inertial sensors, W is the weight coefficient vector, For the i The weight coefficients of the inertial sensors.

[0097] According to the membership matrix and weight coefficient of the inertial sensor, by weighting the membership of the measurement values ​​of each inertial sensor, the basic probability distribution of each inertial sensor to the target can be obtained, and the conditional expression for each basic probability distribution function to meet the basic probability assignment (BPA) is as follows:

[0098] ;

[0099] in, To define sources of evidence i Proposition The quality distribution, For the i and j The membership function of an inertial sensor is for j The weight coefficient of the inertial sensor, .

[0100] In one embodiment, the cosine of the angle between the evidence vectors is used to represent the similarity between the evidences, including: calculating the cosine of the angle between each two different evidence vectors; i and every other evidence vector j The sum of the cosines of the angles between them is taken as the similarity between the evidences before normalization; , ; Normalize the similarity between evidences before normalization to obtain the similarity between evidence vectors.

[0101] The calculation formula for the similarity between evidence vectors is as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] in, is the angle between two evidence vectors, is the cosine of the angle between the evidence vectors, is the similarity between evidences before normalization, is the similarity between evidences, Evidence vectors are i and evidence vector j .

[0106] It should be understood that although Figure 1The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0107] In a verification example, this example designs five groups of gyroscopes for data acquisition in a stationary state and a moving state, and uses this method to perform data fusion to compare and verify the multi-sensor data fusion effect.

[0108] The experimental design is to keep 5 groups of gyroscopes stationary, and collect 16,000 groups for fusion experiment, such as Figure 2 True value measured at rest , we now randomly select the 2000th set of data for testing. The collected data of 5 sets of gyroscopes are as follows:

[0109] Gyroscope 1: =0.3435; Gyroscope 2: =0.1052; Gyroscope 3: =0.0549;

[0110] Gyroscope 4: =0.0464; Gyroscope 5: =0.0213;

[0111] The membership matrix is:

[0112] ;

[0113] Credibility is:

[0114] ;

[0115] Average credibility:

[0116] ;

[0117] Since the credibility of gyroscope 1 is lower than the average, it is removed. The remaining four sets of gyroscope data are used to construct a new recognition framework and update the membership matrix as shown below:

[0118] ;

[0119] ;

[0120] The credibility coefficient is updated to:

[0121] ;

[0122] Normalized weight coefficient:

[0123] ;

[0124] In the new identification framework The remaining four groups of evidence with high credibility are shown in Table 1:

[0125] Table 1 Basic probability distribution of evidence

[0126]

[0127] After obtaining the evidence, this method is used to fuse the evidence in Table 1. The final fused evidence is:

[0128] ;

[0129] Comparing the measurements with the above evidence The fusion result can be obtained by combining:

[0130] ;

[0131] In order to intuitively see the effectiveness of this method, the above fusion results are compared with the fusion results obtained by the weight coefficient method, DS evidence theory, literature 1 (Xu Sunqing, Geng Junbao, Wei Shuhuan, et al. An improved DS conflict evidence synthesis method [J]. Firepower and Command Control, 2019, 44(10): 84-88), and literature 2 (Dong Yuan, Cao Menglong, Jiang Kai. Research on an improved method of conflict evidence in DS evidence theory [J]. Electronic Measurement Technology, 2018, 41(23): 29-33). The experimental comparison results are shown in Table 2.

[0132] Table 2 Fusion results of various algorithms

[0133]

[0134] As shown in Table 2, this method eliminates low-credibility inertial sensors through data screening and improves the evidence sources and combination rules, resulting in the best fusion effect compared to the other four methods. The weight coefficient method and DS evidence theory lack a correction step, resulting in poor fusion results. The methods in Reference 1 and Reference 2 significantly improve fusion results, but Reference 1 fails to consider the bias introduced by DS evidence theory when there are significant conflicts between expert ratings. In the multi-source evidence fusion scenario, the method in Reference 2 relies solely on the conflict degree to correct the evidence, ignoring the impact of fuzzy information.

[0135] The above data is obtained based on the 2000th set of sampling values, which does not mean that this method is completely superior to other algorithms. Therefore, the first 10 sets of data are taken for comparison of fusion results. Figure 3 As shown:

[0136] Depend on Figure 3 It can be seen that the results of the fusion of the first 10 groups of continuous sampling points are the best and closer to the true value, which verifies the effectiveness of this method. In order to further verify the effect of this method in multi-sensor data fusion, the 16,000 groups of collected data are simulated, and the results are as follows: Figure 4 shown.

[0137] Figure 4 The red line in the middle is the fusion result. It can be clearly seen that this method is effective in data fusion, and the fusion result is closest to the true value.

[0138] In order to verify the fusion effect of this method in motion, the following experiment was designed: a total of 16,000 sets of data were collected, the first 3,000 sets were static stages, sets 3,001 to 6,000 were uniform acceleration stages, and sets 6,001 to 11,000 were maintained. Uniform rotation, 11001 to 14000 groups are uniformly accelerated to , and finally maintain Uniform motion, the fusion result is as follows Figure 5 As shown:

[0139] In order to more intuitively see the comparison diagram before and after the fusion of gyroscope sampling data, the 4000 to 4500 groups and 13000 to 13500 groups of data in the two acceleration processes are intercepted for plotting, as shown in the figure below. Figure 6 As shown, (a) is the comparison diagram before and after fusion under the first motion state (the 4000th to 4500th group of data during the acceleration process), and (b) is the comparison diagram before and after fusion under the second motion state (the 13000th to 13500th group of data during the acceleration process). Figure 6 The error of gyroscope 1 is too large to be observed, and it is eliminated during fusion, so it is also eliminated this time.

[0140] Through data collection and simulation experiments in static and moving states, it can be seen that this method has good stability and significant fusion effect.

[0141] To verify the effectiveness of the algorithm, fusion experiments were conducted based on five sets of gyroscope sampling data under static and moving conditions. The results show that the improved DS evidence theory performs well in multi-sensor data fusion, proving the feasibility and effectiveness of the algorithm.

[0142] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the scope of the present application, and such modifications and improvements are all within the scope of protection of the present application.

Claims

1. A multi-sensor data fusion method based on improved DS evidence theory, characterized in that: Used to realize data fusion of multiple inertial sensors in a micro inertial measurement unit; the method includes: Obtain measurement data from multiple inertial sensors in a micro-inertial measurement unit, build a recognition framework, and use the measurement data of each inertial sensor as a piece of evidence; According to the evidence Jousselme The distance determines the normalized evidence support; The cosine of the angle between evidence vectors is used to represent the similarity between evidences; Determine the evidence correction factor based on the normalized evidence support and the similarity between the evidences; Apply the evidence correction factor to each piece of evidence. mass The function is corrected to obtain the corrected evidence; The revised evidence is treated differently using relative conflict factors to determine the credibility of the evidence source; The membership matrix of the inertial sensor is calculated in the recognition framework by using the descending half-normal function as the membership function; determining the credibility of each inertial sensor in the identification framework based on the membership matrix; Eliminating evidence with a credibility lower than a preset threshold from the identification framework to obtain a new identification framework; According to the membership matrix of the inertial sensor under the new identification framework, the weight coefficient of each inertial sensor under the new identification framework is determined; According to the membership matrix and weight coefficient under the new recognition framework, the basic probability distribution of each inertial sensor to the target under the new recognition framework is determined; According to the basic probability distribution, final fusion evidence is determined, and according to the final fusion evidence and corresponding measurement data, a data fusion result of multiple inertial sensors in the micro inertial measurement unit is determined.

2. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: According to the evidence Jousselme The distance determines the normalized evidence support, including: Compare the current evidence with other different evidence Jousselme The opposite of the distance is added to 1, and the distance between the current evidence and all other different evidences is calculated. Jousselme The sum of the distances is calculated to obtain the evidence support; The evidence support is normalized to obtain a normalized evidence support.

3. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: Determine the evidence correction factor based on the normalized evidence support and the similarity between the evidence, including: The normalized evidence support and the similarity between the evidences are multiplied and the result is normalized to obtain the evidence correction factor.

4. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: Apply the evidence correction factor to each piece of evidence. mass The function is corrected and the corrected evidence is: in, For the revised evidence, For evidence of mass function, is the evidence correction factor, is the corrected probability mass of the entire framework, represents the probability mass assigned to the entire framework, To identify the framework.

5. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: The revised evidence is treated differently using the relative conflict factor to determine the credibility of the evidence source: in, Represents the basic probability distribution after fusion, that is, the proposition The final probability mass of To identify the framework, As a source of evidence i Focus Element The quality distribution of As a source of evidence i Focus Element The quality distribution, is a conflict modifier used to add quality when evidence conflicts. As a source of evidence j Focus Element The quality distribution of As a source of evidence j Focus Element A The quality distribution, This is the standard requirement of BPA.

6. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: Determining the credibility of each inertial sensor in the identification framework according to the membership matrix includes: All elements of each row in the membership matrix of each inertial sensor measurement value in the identification framework are arithmetic averaged to obtain the credibility of each inertial sensor in the identification framework.

7. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: According to the membership matrix of the inertial sensor under the new identification framework, the weight coefficient of each inertial sensor under the new identification framework is determined, including: Determine the credibility of each inertial sensor based on its membership matrix under the new identification framework; The credibility of the inertial sensor is normalized, and the normalized result is used as the weight coefficient of each inertial sensor.

8. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: According to the membership matrix and weight coefficient under the new recognition framework, the basic probability distribution of each inertial sensor to the target under the new recognition framework is determined, including: In the new recognition framework, the membership of the measurement value of each inertial sensor is weighted with the weight coefficient to obtain the weighted membership of the measurement value of each inertial sensor; the ratio of the weighted membership of the measurement value of each inertial sensor to the sum of the weighted memberships of the measurement values ​​of all inertial sensors is used as the basic probability distribution of the corresponding inertial sensor to the target.

9. The multi-sensor data fusion method based on the improved DS evidence theory according to claim 1 is characterized in that: The cosine of the angle between evidence vectors is used to characterize the similarity between evidences, including: Calculate the cosine of the angle between each two different evidence vectors; Evidence Vector i and every other evidence vector j The sum of the cosines of the angles between them is taken as the similarity between the evidences before normalization; , ; Normalize the similarity between evidences before normalization to obtain the similarity between evidence vectors.

Images