An aircraft cluster cooperative positioning method based on multi-source sensor data fusion

Abstract

The application belongs to the technical field of unmanned aerial vehicle cluster cooperative navigation, and discloses an aircraft cluster cooperative positioning method based on multi-source sensor data fusion. The application is characterized in that a double-stage fusion structure is formed by an adaptive Kalman filter based on an observation noise adjustment mechanism and a residual compensation network built based on an improved multi-layer full connection structure neural network. Firstly, the adaptive Kalman filter is used to fuse the inertial measurement unit (IMU) and radio equipment measurement information carried by the unmanned aerial vehicle to obtain relatively accurate fusion positioning information. Then, the IMU information, radio equipment measurement information and fusion positioning information are further input into the residual compensation network built based on the improved multi-layer full connection structure neural network for accuracy compensation, so that the cooperative positioning accuracy of the unmanned aerial vehicle in the global navigation satellite system denial area is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle cluster cooperative navigation, and particularly relates to a method for cooperative positioning of an aircraft cluster based on multi-source sensor data fusion in a satellite denial situation. BACKGROUND

[0002] An unmanned aerial vehicle is a flying vehicle controlled by a remote control or autonomous system, widely used in military reconnaissance, environmental monitoring, logistics transportation and other fields. With the development of technology, unmanned aerial vehicle clusters, as a form of multiple unmanned aerial vehicles cooperating to perform tasks, have stronger task adaptability, robustness and spatial coverage. In a satellite denial environment where satellite signals are unavailable or interfered, traditional navigation methods relying on global navigation satellite systems will face serious challenges. Therefore, unmanned aerial vehicle clusters integrate multi-source sensors (such as inertial navigation, radio systems, ultra-wideband, etc.) and cooperative navigation algorithms to achieve information sharing and mutual positioning among members, thereby maintaining overall navigation accuracy and task continuity. This cooperative navigation technology not only enhances the autonomous ability of unmanned systems in complex environments, but also lays a foundation for the development of swarm intelligence navigation systems.

[0003] In the optimization algorithm of unmanned aerial vehicle cluster cooperative navigation, Kalman filtering is one of the most commonly used algorithms. Kalman filter combines system model and measurement data to achieve more accurate state estimation. However, traditional Kalman filtering is difficult to adapt to fluctuations in sensor errors in complex environments, which may lead to decreased filtering performance or even estimation divergence. Therefore, scholars have begun to study adaptive Kalman filtering to automatically adjust filter parameters in unknown or dynamically changing noise environments, maintain stable estimation and high precision, especially suitable for multi-source heterogeneous data fusion positioning tasks in complex systems such as unmanned aerial vehicles.

[0004] In addition, in recent years, with the rapid development of artificial intelligence technology, machine learning has shown great potential in the field of unmanned aerial vehicle cluster cooperative navigation. Traditional cooperative navigation methods rely mainly on model-driven filtering algorithms, but when faced with complex, nonlinear, high-dynamic environments, model accuracy and parameter adjustment have bottlenecks. Machine learning methods have the ability to automatically learn system dynamics and environmental characteristics from massive historical or real-time data, providing new ideas for solving modeling errors, sensor drift and multi-source information fusion in unmanned aerial vehicle cluster cooperative navigation. When the global navigation satellite system fails or weakens, multi-source cooperative unmanned aerial vehicle cluster navigation can still provide reliable navigation services. SUMMARY

[0005] The purpose of the present application is to propose a kind of aircraft cluster cooperative positioning method based on multi-source sensor data fusion, which firstly fuses the inertial measurement unit (IMU) and radio equipment measurement information carried by unmanned aerial vehicle through the proposed adaptive Kalman filter to obtain a more accurate fusion positioning information, then inputs the IMU information, radio equipment measurement information and fusion positioning information as multi-source sensor data into the improved neural network model for accuracy compensation, and finally improves the cooperative positioning accuracy of unmanned aerial vehicle in global navigation satellite system denial zone.

[0006] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0007] A kind of aircraft cluster cooperative positioning method based on multi-source sensor data fusion, comprising the following steps:

[0008] Step 1. Firstly, for the aircraft cluster composed of an auxiliary aircraft and multiple reference aircraft, the three-dimensional coordinate estimate value of the auxiliary aircraft is obtained by radio ranging method, which is used as the radio equipment measurement information in step 2.

[0009] Step 2. Secondly, the inertial measurement unit (IMU) information and radio equipment measurement information carried by the reference aircraft are adaptively fused by using the improved adaptive Kalman filter method to obtain the fusion positioning information.

[0010] The improved adaptive Kalman filter method introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of observation residual on the basis of the traditional Kalman filter, and the adjustment process of the adaptive observation noise adjustment mechanism is as follows:

[0011] By calculating the Euclidean norm of the current observation residual at each time, and dynamically adjusting the observation covariance matrix based on the size of the Euclidean norm, the Kalman filter gain is dynamically adjusted, so as to realize the adaptive update of state estimation.

[0012] Step 3. Build a residual compensation network based on the improved multi-layer fully connected structure neural network, and input the fusion positioning information, IMU information and radio equipment measurement information as multi-source sensor data into the residual compensation network.

[0013] The residual compensation network learns the mapping relationship from the input multi-source sensor data information to the trajectory residual, finally outputs the residual prediction value, and compensates the accuracy of the fusion positioning information of the adaptive Kalman filter based on the residual prediction value.

[0014] In addition, on the basis of the above-mentioned aircraft cluster cooperative positioning method based on multi-source sensor data fusion, the present application also proposes a kind of aircraft cluster cooperative positioning system based on multi-source sensor data fusion corresponding thereto, and its scheme is as follows:

[0015] A multi-source sensor data fusion-based aircraft cluster cooperative positioning system comprises the following modules:

[0016] A preprocessing module is configured to obtain a three-dimensional coordinate estimation value of an auxiliary aircraft in a cluster of aircrafts composed of the auxiliary aircraft and a plurality of reference aircrafts by using a radio ranging method as radio device measurement information.

[0017] A Kalman filter module is configured to adaptively fuse inertial measurement unit (IMU) information and the radio device measurement information mounted on the reference aircrafts by using an improved adaptive Kalman filter method to obtain fused positioning information.

[0018] The improved adaptive Kalman filter method introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of observation residuals on the basis of a traditional Kalman filter, and the adjustment process of the adaptive observation noise adjustment mechanism is as follows:

[0019] The Euclidean norm of the current observation residuals is calculated at each time, and the observation covariance matrix is dynamically adjusted based on the size of the Euclidean norm, and then the Kalman filter gain is dynamically adjusted, so as to realize adaptive updating of the state estimation.

[0020] An accuracy compensation module is configured to build a residual compensation network based on an improved multi-layer full-connection structure neural network, and input the fused positioning information, the IMU information and the radio device measurement information into the residual compensation network as multi-source sensor data.

[0021] The residual compensation network learns a mapping relationship from the input multi-source sensor data information to the trajectory residual, finally outputs a residual prediction value, and compensates the fused positioning information of the adaptive Kalman filter based on the residual prediction value.

[0022] In addition, on the basis of the multi-source sensor data fusion-based aircraft cluster cooperative positioning method described above, the present application further proposes a computer device comprising a memory and one or more processors.

[0023] The executable code is stored in the memory. When the processor executes the executable code, the steps of the multi-source sensor data fusion-based aircraft cluster cooperative positioning method described above are implemented.

[0024] In addition, on the basis of the multi-source sensor data fusion-based aircraft cluster cooperative positioning method described above, the present application further proposes a computer readable storage medium having a program stored thereon, which is executed by a processor to implement the steps of the multi-source sensor data fusion-based aircraft cluster cooperative positioning method described above.

[0025] The present application has the following advantages:

[0026] As described above, the present application proposes a multi-source sensor data fusion based aircraft cluster cooperative positioning method, which is composed of a two-stage fusion structure through the proposed adaptive Kalman filter based on observation noise adjustment mechanism and the residual compensation network built based on the improved multi-layer full connection structure neural network. First, the inertial measurement unit (IMU) and radio measurement information carried by the unmanned aerial vehicle (UAV) are fused to obtain a more accurate fusion positioning information through the proposed adaptive Kalman filter. Then, the IMU information, radio measurement information and fusion positioning information are input into the improved neural network model as inputs for accuracy compensation. Finally, it is beneficial to improve the cooperative positioning accuracy of the UAV in the global navigation satellite system denial area. The present application realizes a unified solution for range noise suppression, IMU drift suppression and nonlinear error compensation. Compared with the traditional method which only relies on filter or machine learning, the present application has the characteristics of high robustness, high real-time and high accuracy. In the complex environment of satellite navigation denial, serious shielding and obvious multipath interference, the present application can still maintain stable trajectory estimation and significantly reduce the three-dimensional position error, which is suitable for unmanned aerial vehicle formation, crowd intelligence cooperation and high reliable positioning application scenarios. The present application method greatly improves the positioning accuracy of the auxiliary aircraft in the aircraft cluster. In addition, the algorithm principle of the present application is simple, the process design is simple and clear, and it is especially suitable for popularization in the environment of satellite navigation denial for unmanned aerial vehicle cooperative navigation. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The flowchart of the multi-source sensor data fusion based aircraft cluster cooperative positioning method in the embodiment of the present application is shown in Figure 1.

[0028] Figure 2 The aircraft position relationship diagram in the embodiment of the present application is shown in Figure 2.

[0029] Figure 3 The overall processing flowchart of the improved adaptive Kalman filter and neural network combination in the embodiment of the present application is shown in Figure 3.

[0030] Figure 4 The flight trajectory diagram of multiple unmanned aerial vehicles in the specific example of the present application is shown in Figure 4.

[0031] Figure 5 The error diagram of the unmanned aerial vehicle in the X, Y and Z directions in the specific example of the present application is shown in Figure 5; wherein Figure 5 (a), (b) and (c) in Figure 5 are respectively the error curves of the unmanned aerial vehicle in the X direction, the Y direction and the Z direction with respect to time. DETAILED DESCRIPTION

[0032] The present application will be further described in detail in conjunction with the drawings and specific embodiments.

[0033] Embodiment 1

[0034] This embodiment 1 describes a method for cooperative positioning of aircraft cluster based on multi-source sensor data fusion, which proposes an improved adaptive Kalman filter and an improved neural network combined cooperative navigation optimization algorithm. First, the information of the inertial measurement unit (IMU) and radio equipment carried by the unmanned aerial vehicle is fused by the improved adaptive Kalman filter to obtain a more accurate positioning information, and then the IMU information, radio equipment measurement information and fused positioning information obtained by Kalman filter are input as multi-source sensor data into the improved neural network model for further optimization, which finally improves the cooperative positioning accuracy of the unmanned aerial vehicle in the global navigation satellite system denial area.

[0035] As shown in Figure 1 , the method for cooperative positioning of aircraft cluster based on multi-source sensor data fusion comprises the following steps:

[0036] Step 1. First, for the aircraft cluster composed of an auxiliary aircraft and multiple reference aircrafts, the three-dimensional coordinate estimate value of the auxiliary aircraft is obtained by radio ranging method as the radio equipment measurement information in step 2.

[0037] In this embodiment, the aircraft cluster, for example, includes one auxiliary aircraft with low positioning accuracy and multiple reference aircrafts with high positioning accuracy, as shown in Figure 2 , wherein the position of the auxiliary aircraft is . In order to improve the navigation accuracy of the auxiliary aircraft, the cooperative navigation model is established by obtaining the sensor information between the auxiliary aircraft and the reference aircrafts.

[0038] The distance between the ith reference aircraft and the auxiliary aircraft is defined as . Radio ranging is mainly to estimate the distance between two points through the propagation characteristics of radio signals, and is widely used in unmanned aerial vehicle positioning and navigation system.

[0039] Radio ranging usually uses TOA (Time of Arrival) method, and the TOA distance measurement formula is:

[0040] .

[0041] Wherein is the measured distance, is the speed of light, is the signal sending time, is the signal receiving time. The position of the auxiliary aircraft is , and the known coordinates of the reference unmanned aerial vehicle are The distance between the assisting UAV and the reference UAV satisfies:

[0042] .

[0043] Based on the first equation, subtract the other three equations to construct a linear form:

[0044] .

[0045] In matrix form:

[0046] ;

[0047] Where:

[0048] ;

[0049] ;

[0050] ;

[0051] Where is the coefficient matrix, is the constant vector. Finally, by least squares method:

[0052] ;

[0053] The three-dimensional coordinate estimate of the assisting aircraft can be obtained, which is the position information of the radio equipment measurement information.

[0054] Next, through the adaptive Kalman filter based on the observation noise adjustment mechanism proposed in this paper, and the residual compensation network constructed based on the improved multi-layer full connection structure neural network, a two-stage fusion structure is formed, as shown in Figure 2 .

[0055] The system first initializes the state, and uses the position and velocity information provided by the radio ranging equipment and the inertial measurement unit to perform state prediction. On this basis, by calculating the observation residual and dynamically adjusting the observation noise covariance, the adaptive update of the state estimation is realized, and the fused position information is obtained, as shown in the left side of the block diagram in Figure 2 .

[0056] Figure 2 The model structure of the residual compensation network of the improved multi-layer full connection structure neural network is shown in the right side of the block diagram in

[0057] This neural network can be used as a compensator for the filter output, improving the system's robustness to IMU drift or measurement errors.

[0058] Step 2. Using an improved adaptive Kalman filter method, the inertial measurement unit (IMU) information and radio equipment measurement information mounted on the reference aircraft are adaptively fused to obtain fused positioning information.

[0059] The improved adaptive Kalman filtering method introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of the observation residuals, building upon the traditional Kalman filter. The adjustment process of this adaptive observation noise adjustment mechanism is as follows:

[0060] By calculating the Euclidean norm of the current observation residual at each time step and dynamically adjusting the observation covariance matrix based on the magnitude of the Euclidean norm, the Kalman filter gain is dynamically adjusted, thereby achieving adaptive updating of the state estimate.

[0061] This invention utilizes an improved adaptive Kalman filter to adaptively fuse IMU data and TOA calculation results to improve positioning accuracy. System Status The model defines a six-dimensional state vector, including position and velocity:

[0062] .

[0063] in, It's about location. , It's speed. .

[0064] This invention introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of the observation residuals, building upon the traditional Kalman filter. At each time step, the Euclidean norm of the current observation residuals is calculated, and the observation covariance matrix is ​​dynamically adjusted based on its magnitude. This is achieved by introducing a nonlinear function to regulate the weights (parameters). This allows the filter to maintain a high level of confidence when the observation residuals are small, while automatically increasing the observation covariance matrix when the observation error changes abruptly or outliers occur, thereby reducing the impact of observations on state estimation and improving the system's robustness to noise changes and abnormal measurements.

[0065] The adaptive observation covariance adjustment mechanism based on the Euclidean norm of observation residuals can dynamically adjust the filter gain according to the real-time measurement reliability. The adjustment process of the adaptive observation noise adjustment mechanism is as follows:

[0066] When the residual decreases, the observation covariance decreases accordingly, the Kalman filter gain increases, and the adaptive Kalman filter makes full use of stable and reliable radio ranging information to correct IMU drift, thereby improving positioning accuracy.

[0067] When the residual increases, the observation covariance increases, and the Kalman filter gain decreases. The adaptive Kalman filter reduces the dependence on abnormal measurements, so as to suppress the influence of sudden error caused by multipath interference and signal blocking on state estimation.

[0068] The adaptive observation noise adjustment mechanism significantly enhances the robustness of the system in processing non-Gaussian noise and sudden abnormalities, avoids oscillation or divergence of state estimation, and improves the stability and reliability of the cooperative positioning system in the GPS denial environment.

[0069] The processing flow of the improved adaptive Kalman filter method in this embodiment is as follows:

[0070] Step 2.1. State initialization, set the initial state and covariance.

[0071] The initial state is as follows:

[0072] .

[0073] The initial covariance is , and the initial velocity is .

[0074] Step 2.2. Construct the state transition matrix :

[0075] .

[0076] Where is the transformation matrix, represents the time difference, represents a 3x3 zero matrix.

[0077] Step 2.3. State prediction, the formula is as follows:

[0078] .

[0079] Where represents the current state update, represents the previous state.

[0080] Step 2.4. The initial state velocity is replaced by IMU correction, the formula is as follows:

[0081] .

[0082] Where represents the current velocity, represents the velocity measured by IMU.

[0083] Step 2.5. Prediction covariance update, formula as follows:

[0084] ;

[0085] .

[0086] where is the process noise covariance matrix, is the variance of the process noise, is the transformation matrix, i.e., the identity matrix, denotes the covariance matrix at the current time, denotes the covariance matrix before transition at the last time.

[0087] Step 2.6. Dynamically adjust Kalman filter gain based on adaptive observation noise adjustment mechanism , the specific process as follows:

[0088] First, give the expression of the observation matrix , formula as follows:

[0089] .

[0090] where is the transformation matrix transformation matrix, i.e., the 3 × 3 identity matrix, denotes the 3 × 3 0 matrix.

[0091] Then calculate the observation residual , formula as follows:

[0092] .

[0093] where denotes the position of the auxiliary aircraft, denotes the observation matrix, denotes the current state update.

[0094] Then calculate the Euclidean norm of the residual , formula as follows:

[0095] .

[0096] Next, give the expression of the adaptive observation noise covariance matrix , formula as follows:

[0097] .

[0098] where is the basic variance of the observation noise, is the adjustment coefficient is a certain value, represents the weight, used to adjust , This represents the norm of the observed residuals using the hypertangent function. Perform nonlinear mapping.

[0099] Finally, dynamically adjust the Kalman filter gain. The formula is expressed as follows:

[0100] ;in This represents the covariance matrix at the current time.

[0101] The adaptive observation noise adjustment mechanism proposed in this invention can dynamically adjust the noise matrix according to the real-time observation quality of the sensor, thereby automatically balancing the weights of observation information and prediction models, and avoiding the instability and error amplification problems caused by fixed noise assumptions.

[0102] This adaptive observation noise adjustment mechanism can effectively suppress abnormal measurements and outliers, significantly improve the robustness and accuracy of filtering in complex environments, and at the same time reduce the need for manual parameter tuning. This invention can intelligently adapt to the characteristics of noise changing with the environment, and is particularly suitable for UAV navigation scenarios where noise is unpredictable, such as satellite denial and multipath interference.

[0103] Step 2.7. Update Status:

[0104] .

[0105] in Indicates the current state. This indicates an update to the current state; To observe the residuals.

[0106] Step 2.8. Update covariance:

[0107] .

[0108] in This represents the updated covariance.

[0109] Step 2.9. Output estimated location :

[0110] .

[0111] in , , They represent Coordinates in the X, Y, and Z directions.

[0112] The estimated location That is, the improved adaptive Kalman filter is used to adaptively fuse the information of the inertial measurement unit (IMU) and the radio equipment measurement information carried by the reference reference aircraft, and the obtained fusion positioning information.

[0113] Next, a feedforward neural network is used in step 3 to learn the mapping relationship from multi-source sensor information to trajectory residual, which is finally used for precision compensation of the fusion result of the adaptive Kalman filter.

[0114] Step 3. Build a residual compensation network based on an improved multi-layer fully connected structure neural network, and input the fusion positioning information, IMU information and radio equipment measurement information into the residual compensation network as multi-source sensor data.

[0115] The residual compensation network learns the mapping relationship from the input multi-source sensor data information to the trajectory residual, finally outputs the residual prediction value, and performs precision compensation on the fusion positioning information of the adaptive Kalman filter based on the residual prediction value.

[0116] Input of the residual compensation network The formula is expressed as follows:

[0117] ;

[0118] represents the position information of the IMU;

[0119] wherein X, Y and Z are the coordinates in three directions obtained by the IMU;

[0120] represents the radio equipment measurement information;

[0121] wherein X, Y and Z represent the coordinates in three directions measured by the radio equipment;

[0122] represents the fusion positioning information;

[0123] wherein , , represents the fusion positioning information X, Y and Z are the coordinates in three directions;

[0124] represents the speed information of the IMU;

[0125] wherein X, Y and Z represent the speed in three directions obtained by the IMU.

[0126] The input Z-score standardization is performed, and the formula is as follows:

[0127] ;

[0128] wherein is the standardization result, is the mean value of each feature, is the standard deviation of each feature.

[0129] The output target of the residual compensation network is the residual prediction value , and the formula is as follows:

[0130] ;

[0131] wherein , , respectively represent the correction value of the residual prediction value in the X, Y, and Z directions.

[0132] Multi-source sensor data fusion refers to the process of comprehensively processing data from multiple different sensors according to certain mathematical methods and strategies, so as to obtain more accurate, stable, and reliable information than a single sensor.

[0133] The processing flow of the residual compensation network based on the improved multi-layer full connection structure neural network in the embodiment is as follows:

[0134] Step 3.1. Input feature extraction, obtain multi-source sensor information.

[0135] Step 3.2. Linear transformation, affine transformation is performed to generate intermediate layer output, and the formula is expressed as follows:

[0136] ;

[0137] wherein represents the full connection layer layer full connection operation, is the weight matrix, is the output or input of the previous layer, is the bias vector.

[0138] Step 3.3. Batch normalization, standardize the features of each mini-batch:

[0139] , ;

[0140] wherein represents the average value of all samples in the current mini-batch, is the number of samples in the current mini-batch, is the activation value of the th sample, represents the variance of the activation values in the current mini-batch.

[0141] The present invention introduces Batch Normalization. Batch Normalization is a method of standardizing the input features or intermediate activation values of each layer in a neural network. Its goal is to solve the problems of gradient vanishing / explosion, slow training convergence, and activation distribution drift between different layers that occur during the training process of a neural network.

[0142] Step 3.4. Normalize the features, standardize to mean 0 and variance 1, and introduce a small regularization term :

[0143] .

[0144] where is the normalized feature of the activation value of the th sample.

[0145] is a very small positive number to prevent the denominator from being 0 and to improve numerical stability.

[0146] Step 3.5. Introduce Dropout, randomly mask some neurons to prevent overfitting. Dropout is a regularization method that randomly discards some neurons (i.e., sets their activation values to 0) during training to prevent overfitting.

[0147] For the normalized output , generate a mask vector of the same dimension. is a Bernoulli distribution, where: is the Dropout mask, is the retention probability.

[0148] Step 3.6. Perform scaling and translation to restore the distribution with trainable parameters, .

[0149] where is the trainable scaling factor that controls the amplitude of the normalized output, is the trainable offset factor that controls the bias of the normalized output, and the final output is used as input to the next layer.

[0150] Step 3.7. Add the activation function ReLU:

[0151] ;

[0152] wherein represents an activation function.

[0153] Multiple full connection layers add multiplication can be finally simplified as a linear function, and the model expression ability will be weak.

[0154] Step 3.7 introduces the activation function ReLU, retains the positive value, discards the negative value, introduces the nonlinearity, so that the neural network can fit more complex relationship. And the method is very simple, only once maximum value comparison is made to the numerical value, so it is calculated very fast in the forward and backward propagation. The gradient of ReLU is 1 (for positive numbers), and the gradient will not become very small far away from 0.

[0155] Step 3.8. Repeat steps 3.2 to 3.7 to perform multi-layer deep feature learning.

[0156] Step 3.9. The output layer outputs the residual prediction value :

[0157] ;

[0158] wherein represents the output layer of the neural network, represents the output of the third layer, represents the bias vector of the output.

[0159] Step 3.10. Correct the predicted position using the residual prediction value , output the corrected trajectory estimation result :

[0160] ;

[0161] wherein represents the estimated position output by the improved adaptive Kalman filtering method.

[0162] In the actual cooperative navigation system, the process noise and observation noise of the traditional Kalman filter often have uncertainty or dynamic change characteristics. The traditional Kalman filter relies on a static preset covariance matrix, and it is difficult to adapt to the fluctuations of sensor errors in complex environments, which may lead to a decline in filtering performance and even estimation bias. Therefore, the present application proposes an improved adaptive Kalman filter (AKF), which can automatically adjust the parameters and maintain stable estimation and high precision in unknown or dynamically changing noise environments, especially suitable for multi-source heterogeneous data fusion positioning tasks in complex systems such as unmanned aerial vehicles.

[0163] In addition, the application of machine learning in UAV cooperative navigation is becoming increasingly critical. Although it requires a large amount of training data and has weak generalization ability, it can significantly improve the positioning accuracy, robustness and intelligence level in dynamic environments, and optimize autonomous decision-making and multi-source data fusion. Neural Network (NN) is essentially an information processing and learning system that simulates the operation of human brain neurons. This system uses a multi-layer structure and contains a large number of "neurons" that extract input features step by step to ultimately complete prediction, classification, regression and other tasks. Neural networks can not only learn the non-linear relationships between complex multi-source data, but also automatically extract important feature combinations, and have strong expansion capabilities. It can integrate time series information or physical models to ultimately approximate a residual model that is closer to the true trajectory. Therefore, neural networks are very suitable for solving cooperative navigation-related problems.

[0164] When the Global Navigation Satellite System fails or the signal attenuates, the present application combines AKF with NN to realize UAV cooperative positioning. In the state prediction stage, the velocity information estimated by the inertial measurement unit is introduced to replace the model derivation of the traditional velocity state component, thereby significantly improving the real-time performance and physical consistency of state prediction. By dynamically adjusting the observation covariance matrix according to the observation residual norm, AKF can adaptively handle measurement anomalies and environmental uncertainties, effectively enhancing the robustness and estimation accuracy of the filter. The neural network uses multi-source heterogeneous information fusion as input, and its network architecture uses a multi-layer fully connected design, combined with batch normalization and random deactivation mechanism, and introduces the ReLU activation function to support sparse response, making the improved neural network converge more stably, faster in training speed and stronger in generalization ability. The method proposed in the present application has high positioning accuracy. In addition, compared with existing methods, the method of the present application also has the advantages of lower computational complexity and resource consumption.

[0165] The present application uses the residual compensation network based on the improved multi-layer fully connected structure neural network to predict and correct, which not only improves the system's modeling ability for nonlinear errors, but also enhances the robustness and accuracy of the final trajectory results.

[0166] In addition, in order to verify the effectiveness of the method proposed in the present application, the following experimental process is given.

[0167] As Figure 4 The flight paths of four UAVs are shown, which are performing a cooperative task. Among them, the navigation device of aircraft 1 has low positioning accuracy, resulting in a large position error, so it is designated as an auxiliary aircraft, and the remaining aircraft (i.e. aircraft 2, 3, 4) have high onboard navigation performance and are designated as reference aircraft.

[0168] To improve the positioning accuracy of the auxiliary aircraft, the application uses a reference aircraft to measure and solve the trajectory of the auxiliary aircraft. The sensor information exchange between the reference aircraft and the auxiliary aircraft is used to establish a mathematical relationship, and then the flight trajectory of the auxiliary aircraft is calculated, and finally the cooperative positioning method proposed in the application is used to reduce the positioning error.

[0169] Table 1 Average error

[0170]

[0171] The effect comparison of the method of the application is shown in Figure 5 and Table 1, wherein Figure 5 (a), (b), (c) in the figure are respectively the error curves of the unmanned aerial vehicle in the X direction, the Y direction and the Z direction with time.

[0172] As can be seen from Table 1, after the auxiliary unmanned aerial vehicle is calculated by the method of the application, the X-axis error is about 0.0873 meters, the Y-axis error is about 0.0628 meters, the Z-axis error is about 0.0758 meters, and the position error is about 0.1316 meters. It can be seen that the errors are in a very low range, indicating that the method of the application has greatly improved the positioning accuracy in the satellite denial environment.

[0173] Example 2

[0174] This embodiment 2 describes a kind of aircraft cluster cooperative positioning system based on multi-source sensor data fusion, which is based on the same inventive concept as the aircraft cluster cooperative positioning method based on multi-source sensor data fusion in above-mentioned embodiment 1.

[0175] The system of aircraft cluster cooperative positioning based on multi-source sensor data fusion in this embodiment includes the following modules:

[0176] The preprocessing module is used for obtaining the three-dimensional coordinate estimate value of the auxiliary aircraft by radio ranging method for the aircraft cluster composed of the auxiliary aircraft and multiple reference aircrafts, as radio device measurement information;

[0177] The Kalman filter module is used for adaptive fusion of the inertial measurement unit (IMU) information and the radio device measurement information mounted on the reference aircraft by using the improved adaptive Kalman filter method to obtain the fusion positioning information;

[0178] The improved adaptive Kalman filter method introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of observation residual on the basis of the traditional Kalman filter, and the adjustment process of the adaptive observation noise adjustment mechanism is as follows:

[0179] The adaptive updating of the state estimation is realized by calculating the Euclidean norm of the current observation residual at each moment, and dynamically adjusting the observation covariance matrix based on the size of the Euclidean norm, and then dynamically adjusting the Kalman filter gain;

[0180] and the precision compensation module is configured to build a residual compensation network based on the improved multi-layer full connection structure neural network, and input the integrated positioning information, IMU information and radio equipment measurement information into the residual compensation network as multi-source sensor data.

[0181] The residual compensation network learns the mapping relationship from the input multi-source sensor data information to the trajectory residual, finally outputs the residual prediction value, and compensates the precision of the integrated positioning information of the adaptive Kalman filter based on the residual prediction value.

[0182] It should be noted that the contents not mentioned in the above-mentioned various functional modules in the system described in this embodiment can be described with reference to the steps of the corresponding method in Embodiment 1 described above, and will not be described in detail here.

[0183] Embodiment 3

[0184] This embodiment 3 describes a computer device, which includes a memory and one or more processors. The executable code is stored in the memory. When the processor executes the executable code, the steps of the aircraft cluster cooperative positioning method based on multi-source sensor data fusion in Embodiment 1 described above are implemented.

[0185] The computer device in this embodiment is any device or apparatus with data processing capability, which will not be described here.

[0186] Embodiment 4

[0187] This embodiment 4 describes a computer readable storage medium, which stores a program. When the program is executed by a processor, the steps of the aircraft cluster cooperative positioning method based on multi-source sensor data fusion in Embodiment 1 described above are implemented.

[0188] The computer readable storage medium can be an internal storage unit of any device or apparatus with data processing capability, such as a hard disk or memory, or an external storage device of any device or apparatus with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc.

[0189] Of course, the above description is only the preferred embodiment of the present application, and the present application is not limited to the above-described embodiments. It should be noted that any skilled person in the art can make all equivalent substitutions and obvious modifications under the teaching of the present application, and all such substitutions and modifications shall fall within the scope of the present application, and shall be protected by the present application.

Claims

1. A collaborative localization method for aircraft swarms based on multi-source sensor data fusion, characterized in that, Includes the following steps: Step 1. First, for the aircraft cluster consisting of an auxiliary aircraft and multiple reference aircraft, obtain the estimated three-dimensional coordinates of the auxiliary aircraft using radio ranging, which will be used as the measurement information for the radio equipment in Step 2; Step 2. Next, using the improved adaptive Kalman filtering method, the inertial measurement unit (IMU) information and the radio equipment measurement information mounted on the reference aircraft are adaptively fused to obtain fused positioning information; The improved adaptive Kalman filtering method introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of the observation residuals, building upon the traditional Kalman filter. The adjustment process of this adaptive observation noise adjustment mechanism is as follows: By calculating the Euclidean norm of the current observation residual at each time step and dynamically adjusting the observation covariance matrix based on the magnitude of the Euclidean norm, the Kalman filter gain is dynamically adjusted, thereby achieving adaptive updating of the state estimate. Step 3. Construct a residual compensation network based on an improved multi-layer fully connected neural network structure, and input the fused positioning information, IMU information and radio equipment measurement information as multi-source sensor data into the residual compensation network; The residual compensation network learns the mapping relationship between the input multi-source sensor data and the trajectory residual, and finally outputs the residual prediction value. Based on the residual prediction value, it performs accuracy compensation on the fused positioning information of the adaptive Kalman filter.

2. The collaborative localization method for aircraft swarms based on multi-source sensor data fusion according to claim 1, characterized in that, In step 2, the adjustment process of the adaptive observation noise adjustment mechanism is as follows: When the residual decreases, the observation covariance decreases accordingly, the Kalman filter gain increases, and the adaptive Kalman filter makes full use of stable and reliable radio ranging information to correct IMU drift, thereby improving positioning accuracy. As the residual increases, the observation covariance also increases, the Kalman filter gain decreases, and the adaptive Kalman filter reduces its dependence on anomalous measurements in order to suppress the impact of abrupt errors caused by multipath interference and signal occlusion on state estimation.

3. The collaborative localization method for aircraft swarms based on multi-source sensor data fusion according to claim 1 or 2, characterized in that, In step 2, the adaptive observation noise adjustment mechanism dynamically adjusts the Kalman filter gain as follows: First, the observation matrix is ​​given. The expression, formula is as follows: ； in It is the transformation matrix, i.e., the 3×3 identity matrix. Represents a 3×3 zero matrix; Then calculate the observation residuals. The formula is as follows: ； in Indicates the location of the auxiliary aircraft. Represents the observation matrix. This indicates an update to the current state; Then calculate the residual Euclidean norm. The formula is as follows: ； The adaptive observation noise covariance matrix is ​​then given. The expression, formula is as follows: ； in It is the fundamental variance of the observation noise. The adjustment coefficient is a fixed value, used for adjustment. , This represents the norm of the observed residuals using the hypertangent function. Perform nonlinear mapping; Finally, dynamically adjust the Kalman filter gain. The formula is expressed as follows: ;in This represents the covariance matrix at the current time.

4. The collaborative localization method for aircraft swarms based on multi-source sensor data fusion according to claim 3, characterized in that, The processing flow of the improved adaptive Kalman filter method is as follows: Step 2.

1. State initialization: Set the initial state and covariance; initial state The formula is as follows: ； The initial covariance is The initial velocity is ; Step 2.

2. Construct the state transition matrix : ； in It is the transformation matrix, i.e., the 3×3 identity matrix. Indicates time difference, Represents a 3×3 zero matrix; Step 2.

3. Perform state prediction, using the following formula: ； in This indicates an update to the current state. Indicates the state at the previous moment; Step 2.

4. The initial state velocity is corrected using IMU substitution, as shown in the following formula: ； in Indicates the current speed. This indicates the speed measured by the IMU; Step 2.

5. Predict the covariance update, the formula is as follows: ； ； in It is the process noise covariance matrix. It is the variance of process noise. It is the transformation matrix, i.e., a 6×6 identity matrix. Let the covariance matrix at the current time be denoted as . This represents the covariance matrix before the transition at the previous time step; Step 2.

6. Dynamically adjust the Kalman filter gain based on an adaptive observation noise adjustment mechanism. ; Step 2.

7. Update Status: ； in Indicates the current state. This indicates an update to the current state; To observe residuals; Step 2.

8. Update covariance: ； in Indicates the updated covariance; Step 2.

9. Output estimated location The formula is expressed as follows: ； in , , They represent Coordinates in the X, Y, and Z directions.

5. The collaborative localization method for aircraft swarms based on multi-source sensor data fusion according to claim 1, characterized in that, In step 3, the input of the residual compensation network The formula is expressed as follows: ； This indicates the location information of the IMU; in The coordinates obtained for the IMU in the X, Y, and Z directions; Indicates measurement information from wireless equipment; in These represent the coordinates measured by the wireless equipment in the X, Y, and Z directions, respectively. This indicates the fusion of location information; in , , Indicates fusion of location information Coordinates in the X, Y, and Z directions; This indicates the speed information of the IMU; in These represent the velocities in the X, Y, and Z directions acquired by the IMU, respectively.

6. The collaborative localization method for aircraft swarms based on multi-source sensor data fusion according to claim 1, characterized in that, The processing flow of the residual compensation network is as follows: Step 3.

1. Input feature extraction to obtain multi-source sensor information; Step 3.

2. Linear transformation: Perform an affine transformation to generate the intermediate layer output, expressed by the following formula: ;in Indicates the fully connected layer number 1 Fully connected layer operations, It is a weight matrix. It is the output or input of the previous layer. It is the bias vector; Step 3.

3. Batch normalization: Standardize the features of each mini-batch. ， ； in This represents the average value of all samples in the current mini-batch. This is the number of samples in the current mini-batch. It is the first The activation value of each sample. This represents the variance of the activation values ​​in the current mini-batch of the neural network; Step 3.

4. Normalize the features, standardizing them to a mean of 0 and a variance of 1, and introduce a regularization term. : ； in Indicates the first Normalized features of activation values ​​for each sample; Step 3.

5. Introduce Dropout to randomly disable some neurons to prevent overfitting; Step 3.

6. Perform scaling and translation to reconstruct a trainable distribution. ; in It is a trainable scaling factor that controls the amplitude after standardization. It is a trainable offset factor that controls the standardized bias, and the final output. Used to input the next level; Step 3.

7. Add the ReLU activation function: ； in Indicates the activation function; Step 3.

8. Repeat steps 3.2 to 3.7 to perform multi-layer deep feature learning; Step 3.

9. Output the predicted residual values ​​from the output layer. : ； in This represents the output layer of the neural network. This indicates the output of the third layer. This represents the output bias vector; Step 3.

10. Utilize residual prediction values Correct the predicted position and output the corrected trajectory estimation result. : ； in This represents the estimated position output using the improved adaptive Kalman filter method.

7. The collaborative localization method for aircraft swarms based on multi-source sensor data fusion according to claim 1, characterized in that, The aforementioned aircraft swarm cooperative positioning method is applicable to cooperative positioning in areas denied by global navigation satellite systems. There are at least three reference aircraft, and the positioning accuracy of the auxiliary aircraft is lower than that of the reference aircraft.

8. A collaborative positioning system for aircraft swarms based on multi-source sensor data fusion, characterized in that, Includes the following modules: The preprocessing module is used to obtain the estimated three-dimensional coordinates of the auxiliary aircraft using radio ranging for a cluster of aircraft consisting of auxiliary aircraft and multiple reference aircraft, as measurement information for radio equipment. The Kalman filter module is used to adaptively fuse the inertial measurement unit (IMU) information and radio equipment measurement information mounted on the reference aircraft using an improved adaptive Kalman filter method to obtain fused positioning information. The improved adaptive Kalman filtering method introduces an adaptive observation noise adjustment mechanism based on the Euclidean norm of the observation residuals, building upon the traditional Kalman filter. The adjustment process of this adaptive observation noise adjustment mechanism is as follows: By calculating the Euclidean norm of the current observation residual at each time step and dynamically adjusting the observation covariance matrix based on the magnitude of the Euclidean norm, the Kalman filter gain is dynamically adjusted, thereby achieving adaptive updating of the state estimate. And a precision compensation module, used to build a residual compensation network based on an improved multi-layer fully connected neural network structure, which integrates positioning information, IMU information and wireless equipment measurement information as multi-source sensor data input to the residual compensation network; The residual compensation network learns the mapping relationship between the input multi-source sensor data and the trajectory residual, and finally outputs the residual prediction value. Based on the residual prediction value, it performs accuracy compensation on the fused positioning information of the adaptive Kalman filter.

9. A computer device, comprising a memory and one or more processors; characterized in that, The memory stores executable code, which, when executed by the processor, is used to implement the steps of the collaborative localization method for aircraft swarms based on multi-source sensor data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a program stored thereon; characterized in that, When executed by the processor, the program is used to implement the steps of the collaborative localization method for aircraft swarms based on multi-source sensor data fusion as described in any one of claims 1 to 7.

Images