Graph deviation network optimization method for robust intelligent fusion of multi-source navigation information

Abstract

The invention discloses a graph deviation network optimization method for robust intelligent fusion of multi-source navigation information. Comprising the following steps: establishing a navigation system state quantity, and proposing a multi-sensor fusion architecture based on an abstract sensor factor node construction method, the method comprises the following steps: generating concrete factor nodes according to specific auxiliary sensor measurement information, and updating state optimal fusion of multi-source navigation information based on a factor graph architecture, a graph deviation network GDN autonomous decision algorithm and a robustness graph optimization algorithm in a multi-source navigation information fusion architecture in real time through a plug-and-play mode. According to the method, innovation in multiple aspects such as effective modeling of non-Gaussian noise, multi-source information autonomous decision, system robust fusion and the like is realized; the method provides a new thought and solution for multi-source information fusion in the navigation field, and has a wide application prospect.

Description

Technical Field

[0001] The invention relates to a graph deviation network optimization method for robust intelligent fusion of multi-source navigation information, belonging to the technical field of multi-source information navigation. Background Art

[0002] At present, inertial navigation system (INS) and global navigation satellite system (GNSS) are widely used in various navigation fields. GNSS can assist INS to suppress accumulated or even divergent navigation errors, thereby improving navigation accuracy. However, GNSS signals are easily interfered by obstacles or human factors, and the positioning accuracy is poor. At this time, the measurement noise in the combined navigation system often has non-Gaussian characteristics. If the noise is still modeled as a Gaussian distribution, the error will increase and the accuracy will decrease significantly. The rapid development of sensor technology has made multi-sensor fusion navigation an important way of high-precision navigation and positioning. However, in the fusion process, different sensors have sudden and slow-changing faults due to external environment or internal hardware problems. How to quickly diagnose faults in real time and autonomously make decisions to fuse available navigation information has become an urgent problem to be solved.

[0003] At present, the processing of non-Gaussian noise problems in the field of navigation is often completed in the Kalman filter framework. For example, for the non-Gaussian noise problem caused by signal pollution in the INS / GPS integrated navigation system, the existing processing method is to model it with a Gaussian mixture model and then design an M estimator with the Huber principle. Although the existing Kalman filter and its series of nonlinear filters can be used in many scenarios, it is difficult to fuse heterogeneous asynchronous measurement results from different sensors, and for the problem of high state vector dimension in the nonlinear measurement model, each iteration of the above solution requires a long time for linearization. To address this problem, the factor graph method for navigation state estimation came into being. The factor graph fusion method uses a Bayesian data structure to store the relationship between known measurement values ​​and unknown variables. Its essence is a fusion method based on a probabilistic graphical model. This method can process the measurement information of different sensors with different frequencies and delays in a simple way, and has plug-and-play functionality, which can easily add and delete sensors.

[0004] Recently, the probabilistic graph model in multi-sensor fusion has been widely used in the field of navigation because it can improve the accuracy of navigation positioning. However, the problem that still exists is that if the failure of different sensors is not diagnosed in time, the positioning error will diverge rapidly. The traditional residual chi-square detection method that is widely used at present has a low diagnosis rate for slowly varying faults and may waste normal measurement information. The nonlinear big data feature processing capability of neural networks can solve the problems of single fault diagnosis type and low diagnosis rate in traditional fault diagnosis methods. Therefore, the present invention further optimizes fault detection and autonomous decision-making of multi-source fusion navigation systems through a graph deviation network (Graph Deviation Network, GDN).

[0005] Although the existing factor graph fusion method of multi-source navigation system has a certain degree of robustness, it has not fully considered the impact of non-Gaussian measurement noise brought by the denied environment. In addition, the detection methods for different sensor faults are mostly through models, which have the problems of untimely diagnosis and poor accuracy. Summary of the invention

[0006] The purpose of the present invention is to provide a graph deviation network optimization method for robust intelligent fusion of multi-source navigation information to solve the problems of non-Gaussian characteristic noise and multiple multi-type faults in the navigation information fusion process of inertial navigation systems and other auxiliary navigation sensors, such as the Global Navigation Satellite System (GNSS), radar, odometer, magnetometer, etc.

[0007] To achieve the above object, the present invention adopts the following technical solution:

[0008] A graph deviation network optimization method for robust intelligent fusion of multi-source navigation information, the method comprising the following steps:

[0009] Step 1: Analyze the multi-source navigation system and establish i The state variable x of the multi-source navigation system at i ;

[0010] Step 2: Construct a multi-source navigation system fusion architecture based on the abstract auxiliary sensor factor node;

[0011] Step 3: Propose the graph deviation network (GDN) autonomous decision-making algorithm in the multi-source navigation system fusion architecture;

[0012] Step 4: Propose a robust graph optimization algorithm in the multi-source navigation information fusion architecture, design a robust M-estimator to adaptively adjust the fusion weights, and integrate the graph deviation network GDN autonomous decision-making algorithm to achieve robust, stable and efficient fusion of multi-source navigation information.

[0013] Preferably, the step 1 is:

[0014] Considering the local northeastern sky coordinate system, established at time t i The state variable x of the multi-source navigation system at i as follows:

[0015]

[0016] in represents the attitude angle vector, and are the position and velocity vectors, Represents the gyroscope bias parameter ε used to correct the inertial measurement unit IMU g and accelerometer bias parameters ▽ a .

[0017] Preferably, step 2 comprises the following steps:

[0018] Step 2.1: Establishment and concrete access of abstract auxiliary sensor factor nodes:

[0019] Abstract factor nodes have horizontal positions Vertical Position Horizontal speed Vertical speed Roll Angle Heading Angle Pitch Angle There are seven types of measurement information, and their measurement equations are described as follows:

[0020]

[0021] in is the measurement noise that follows Gaussian distribution corresponding to different measurement information, H * Indicates measurement information and the multi-source navigation system state variable x k The measurement matrix of the relationship between them is:

[0022]

[0023] Get the abstract auxiliary sensor factor node f with unary factor properties * (x k )as follows:

[0024]

[0025] It only connects the state variable x at time k k , where d(·) represents the loss function;

[0026] In order to realize the plug-and-play function of the multi-source fusion navigation system, the abstract auxiliary sensor factor node is assembled according to the actual sensor measurement information to complete the access of the actual sensor measurement information:

[0027] First, when the actual sensor is connected, the multi-source navigation system first receives the first row of sensor data containing the measurement information attributes, so as to determine the measurement information contained in the connected sensor; and selects the corresponding abstract auxiliary sensor factor node according to the type of measurement information received;

[0028] Step 2.2: Establishment of other factor nodes:

[0029] First, define the prior factor with the properties of a unary factor node as follows:

[0030]

[0031] Where d(·) represents the loss function, x0 represents the state variable x of the multi-source navigation system i The initial value of f x INS and f α bias They represent the initial value x0 and the initial deviation estimate α0 respectively;

[0032] Then the multi-source navigation system state variable x i It is further expressed by the following nonlinear discrete equation:

[0033]

[0034] where x k is the state variable of the multi-source navigation system at time k, is a given containing the gyroscope f b and accelerometer ω b Output of IMU measurement, h IMU (·) represents the nonlinear measurement function, is different from x k+1 The estimated value of the multi-source navigation system state; the correction parameter α of the inertial measurement unit IMU k It is described by the following nonlinear equation:

[0035]

[0036] in represents the predicted value of the correction parameter, and g(·) represents the nonlinear functional relationship of the deviation;

[0037] Based on the multi-source navigation system state variable x i Nonlinear discrete equation and correction parameter α kNonlinear equation, INS factor with binary factor properties is as follows:

[0038]

[0039] This formula represents the estimated value x k+1 With the predicted value The error function between

[0040] Similar to the definition of the INS factor, the deviation factor, which also has the characteristics of a binary factor, is defined as follows:

[0041]

[0042] It is connected to the correction parameter α k and the predicted amount of the modified parameters Based on the INS factors and deviation factors, the INS factor graph is established;

[0043] Step 2.3: Establishment of multi-source navigation system fusion architecture based on factor graph model:

[0044] The fusion architecture of the multi-source navigation system based on the factor graph model is established as follows:

[0045]

[0046] The factors are described as follows:

[0047]

[0048] in Represents the error between different sensor measurements and states, δr i ∈δr k Represents the corresponding new information, which is defined as follows:

[0049]

[0050] where h i (·) represents the nonlinear functional relationship between different sensor measurements and states;

[0051] The factor graph model can be further converted into a least squares problem as follows:

[0052]

[0053] Therefore, the multi-source information fusion navigation problem is converted into a nonlinear least squares problem.

[0054] Preferably, step 3 comprises the following steps:

[0055] Step 3.1: Data preprocessing when abstract auxiliary sensor factors are concretized:

[0056] In step 2.1, we get the concrete abstract auxiliary sensor factor node f * (x k ) and define the new information δr k =f * (x k ); Then the new information is normalized to eliminate the adverse effects caused by the singular sample data:

[0057]

[0058] Where d is the normalized new information, μ is the mean of the new information of the abstract auxiliary sensor factor node, and σ is the standard deviation of the new information of the abstract auxiliary sensor factor node;

[0059] Step 3.2: Calculation of the correlation degree of abstract auxiliary sensor factor nodes with nonlinear coupling characteristics and state transfer of graph deviation network GDN:

[0060] The graph deviation network GDN represents the relationship between the abstract auxiliary sensor factor nodes with an embedding vector V i i∈{1,…,7}, the degree of association v between the abstract auxiliary sensor factor nodes is further calculated by the embedding vector ij as follows,

[0061]

[0062] Then, the state transfer of the abstract auxiliary sensor factor node is calculated, and the hidden layer state calculation at time t is defined as follows:

[0063] z t =sigmod(W z D t +U z h t-1 )

[0064] r t =sigmod(W r D t +U r h t-1 )

[0065]

[0066] Among them, D t is the input of the graph deviation network GDN at time t, which is composed of the normalized new information d in step 3.1, and the update gate z t represents the update gate, r t Represents the reset gate; W (·) ,U (·) represents the corresponding vector matrix, Represents the activation state h of the candidate activation map bias network GDN t Activated by Candidate and the previous activation state h t-1 Linear interpolation representation of ;

[0067] The prediction layer consists of two fully connected layers stacked on top of each other, and is calculated as follows:

[0068]

[0069] Where W0 and W1 are parameter matrices, b0 and b1 are vectors, and thus the state of the graph deviation network GDN is transferred;

[0070] Step 3.3: Training and validation of the graph bias network for calculating the nonlinear coupling between multi-sensor factors:

[0071] Training data D Tr By train The new information generated by the multi-source sensors in time is composed of The kth time point Indicated by the new information δr k Obtained through data preprocessing in step 3.1;

[0072] The processed and combined training input data D Tr-in Denoted as:

[0073]

[0074] Where w represents the sliding window size of fixed length;

[0075] Training output data D Tr-out is the sensor data of the corresponding training data set at the next moment, recorded as:

[0076]

[0077] The graph deviation network GDN makes a one-step prediction of the new information, and the predicted output data of the GDN network is compared with the training output data D Tr-out The mean square error (MSE) is used to evaluate the training effect of the graph deviation network (GDN);

[0078] The performance of the graph deviation network GDN needs to be verified before it is connected to the multi-source navigation system fusion architecture: take the time t0 to t va is the verification time period. In the verification phase, the difference between the pre-verification output value of the graph deviation network GDN at time t and the sensor's new information is calculated as:

[0079]

[0080] Among them, dt is the new information of all abstract auxiliary sensor factor nodes at time t, is the new information predicted by the graph deviation network GDN; further, for each dimension e of the above difference vector i (t) are standardized as follows:

[0081]

[0082] in, From t0 to t k The normalized median and quartile of the new information in the i-th row at time;

[0083] Take the maximum value of the deviation score as the threshold λ for fault detection of the corresponding dimension i :

[0084] λ i =max{s i (t0),…,s i (t),…,s i (t va )},t∈{t0,…,t va}

[0085] Step 3.4: Access the prediction results of the map deviation network in the multi-source navigation system fusion architecture:

[0086] In actual application, the prediction input set D Te-in Input graph deviation network GDN to get the predicted output deviation score s i (t), and further compare it with the threshold λ obtained in step 3.3 i contrast:

[0087]

[0088] When the deviation score crosses the threshold, α i =0, then it is determined that the dimension has a fault, and the system decides not to access the measurement information of this dimension. Otherwise, it is determined that the measurement value of this dimension is normal. The prediction result completes the intervention judgment when the multi-sensor factor is concretized.

[0089] Preferably, step 4 comprises the following steps:

[0090] Step 4.1: M-estimation design for robust multi-sensor factor nodes:

[0091] First, it is universal to consider that non-Gaussian noise is modeled as Gaussian mixture noise, which has the following form:

[0092]

[0093] where τ is the error model weight satisfying 0≤τ1, μ and i∈{1,2} represents the mean and variance of Gaussian distribution respectively;

[0094] For the modeling problem of non-Gaussian noise, M estimation is used as the maximum a posteriori probability estimation, which is different from the least squares estimation and is defined as follows:

[0095] minρ(m k ),

[0096] in,

[0097]

[0098] represents the Mahalanobis distance of the measurement error, h(·) represents the nonlinear functional relationship between the measurement and state of different sensors, Σ G is the noise covariance matrix, ρ(·) is the robustness loss function, which is a positive definite convex function with a unique minimum value; based on the iterative weight redistribution algorithm, the above formula can be transformed into:

[0099]

[0100] The adaptive weight ω(m k ) is defined as follows:

[0101]

[0102] Where c is the positive parameter to be designed; at this time, the sensor factor in step 2.3 can be further robust to ω(m k )f i (x k );

[0103] Step 4.2: Robust graph bias network optimization method in multi-source navigation system fusion architecture:

[0104] For the least squares problem mentioned in step 2, after obtaining the fusion architecture of the concrete multi-source navigation system based on the autonomous decision weight of the graph deviation network in step 3 and the M estimation in step 4.1, we get:

[0105]

[0106] Therefore, the least squares problem mentioned above is solved incrementally in real time to obtain the multi-source information fusion navigation result.

[0107] Beneficial effects: The graph deviation network optimization method for robust intelligent fusion of multi-source navigation information provided by the present invention has the following advantages over the prior art: First, starting from reality, the present invention takes the inertial navigation system as the core and the Bayesian tree as the architecture to construct abstract auxiliary sensor factor nodes, and performs intelligent fusion of sensor navigation information in a plug-and-play manner; Second, considering the non-Gaussian characteristics of the sensor measurement model, a robust M estimator is designed, which can effectively reduce the modeling error and thus greatly improve the accuracy of navigation parameter estimation; Third, the graph deviation network with deep neural network characteristics of the present invention models the nonlinear relationship between different sensors, aiming to autonomously decide whether to access a certain dimension of sensor measurement information when different sensors have sudden or even slow-changing faults, so as to improve the stability of the navigation system; Finally, the incremental sliding update algorithm with robustness and autonomous decision-making weights provided by the present invention can further improve the efficiency, robustness and stability of the existing factor graph algorithm, and enrich the application of factor graph-based navigation fusion algorithm in navigation multi-source information fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Figure 1 This is a flow chart of the technical route adopted by the present invention;

[0109] Figure 2 It is a track simulation comparison chart of different algorithms, where FG represents the existing algorithm and RFG represents the algorithm of the present invention;

[0110] Figure 3 A schematic diagram of the magnetometer sensor yaw angle fault diagnosis and autonomous decision-making method of the present invention;

[0111] Figure 4 This is a comparison chart of navigation position errors of the method provided by the present invention, the chi-square detection method and the LSTM detection method after injection of sudden and slow-changing faults, where GDN represents the method of the present invention. DETAILED DESCRIPTION

[0112] The technical solution of the present invention is further described in detail below in conjunction with the accompanying drawings. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention. Figure 1 , a method based on the present invention comprises the following steps:

[0113] Step 1: Taking the unmanned navigation system as the background, considering the local northeast sky coordinate system, establish the i The state quantity x of the multi-source navigation system at i as follows:

[0114]

[0115] in represents the attitude angle vector, and are position and velocity vectors respectively. Similar to the common integrated navigation system, represents the gyroscope ε used to correct the inertial measurement unit (IMU) g and accelerometer bias ▽ a Due to the advantages of high frequency, strong robustness and good stability of the inertial navigation system (INS), INS is often considered as the main system in the integrated navigation system, and other systems such as the global satellite navigation system (GNSS), radar, odometer, magnetometer, etc. are auxiliary navigation sensor systems. Therefore, it is necessary to set the correction parameter α of the inertial navigation i Expanded to the state quantity of the navigation system.

[0116] It is worth mentioning that the patent of the present invention considers directly reflecting the state parameters of the navigation system rather than its corresponding errors. This is because the "plug and play" function of the factor graph is taken into consideration, which enables different sensors to be directly inserted into the navigation system without prior data preprocessing, and has the advantage of reduced system nonlinearity.

[0117] Step 2, for the multi-source fusion navigation system, consider a multi-sensor fusion architecture based on an abstract sensor factor node construction method, which can generate concrete factor nodes according to specific auxiliary sensor measurement information and update the state optimal fusion of multi-source navigation information based on the factor graph architecture in real time through plug-and-play. Step 2 includes the following steps:

[0118] Step 2.1: Establishment and concrete access of abstract auxiliary sensor factor nodes

[0119] In order to meet the plug-and-play functional requirements of the multi-source fusion navigation system, this step designs an abstract sensor factor node, which aims to flexibly select factor nodes and insert them into the factor graph when specific sensors are connected. Based on the navigation parameter information, the abstract factor node has horizontal position Vertical Position Horizontal speed Vertical speed Roll Angle Heading Angle Pitch Angle There are seven types of measurement information, and their measurement equations are described as follows:

[0120]

[0121] in is the measurement noise that follows Gaussian distribution corresponding to different measurement information, H * Indicates the measured value And the multi-source navigation state x in formula (1)k The measurement matrix of the relationship between them is:

[0122]

[0123] From the above measurement equation, we can get the abstract auxiliary sensor factor node f with unary factor properties: * (x k )as follows:

[0124]

[0125] It only connects the state variable x at time k k , where d(·) represents the loss function.

[0126] In order to realize the plug-and-play function of the multi-source fusion navigation system, this step assembles abstract factor nodes according to the actual access sensor measurement information to complete the access of the actual sensor measurement information.

[0127] First, when the actual sensor is connected, the navigation system first receives the first row of sensor data containing the measurement information attributes, so as to determine what measurement information the connected sensor contains.

[0128] Furthermore, the corresponding abstract auxiliary sensor factor node is selected according to the type of measurement information received. For example, when accessing a GNSS that provides three-dimensional position and velocity information, there is We choose to access f P1 (x k ),f P2 (x k ),f V1 (x k ),f V2 (x k ) These four factor nodes, among which the measurement information That is, it is concretely represented by the horizontal position, vertical position, horizontal speed and vertical speed information of GNSS.

[0129] Step 2.2, establishment of other factor nodes

[0130] In order to make the state estimation of the navigation information fusion algorithm based on factor graph more accurate, it is necessary to provide the establishment of factor nodes of prior states. The prior factors with the properties of unary factor nodes are defined as follows:

[0131]

[0132] Where d(·) represents the loss function, x0 represents the state variable x of the multi-source navigation system i In the present invention, further f x INSand f α bias They represent the initial value x0 and the initial bias estimate α0 respectively.

[0133] In formula (1), the state variable x of the multi-source navigation system is i It can be further expressed by the following nonlinear discrete equation:

[0134]

[0135] where x k is the state variable of the multi-source navigation system at time k, is a given containing the gyroscope f b and accelerometer ω b Output of IMU measurement, h IMU (·) represents the nonlinear measurement function, is different from t k+1 x at time k+1 The estimated value of the multi-source navigation system state. In order to minimize the impact of random walk noise, the correction parameter α of the IMU k It is described by the following nonlinear equation:

[0136]

[0137] in represents the predicted value of the correction parameter, and g(·) represents the nonlinear functional relationship of the deviation;

[0138] Based on the above two formulas, the INS factor with binary factor properties is obtained as follows:

[0139]

[0140] This formula represents the estimated value x k+1 With the predicted value The error function between them connects the state variables between adjacent moments.

[0141] Similar to the above definition, the deviation factor, which also has the properties of a binary factor, is defined as follows:

[0142]

[0143] It is also connected to the correction amount α k and

[0144] Based on equations (8) and (9), the factor graph with the inertial navigation system as the core is established.

[0145] Step 2.3: Establishment of multi-source navigation system fusion architecture based on factor graph model

[0146] Based on equations (5), (8), and (9), a multi-source information fusion architecture with the inertial navigation system as the core and other navigation sensors as auxiliary is established:

[0147]

[0148] The factors can be described as follows:

[0149]

[0150] in Represents the error between different sensor measurements and states, δr i ∈δr k Represents the corresponding new information, which is defined as follows.

[0151]

[0152] where h i (·) represents the nonlinear functional relationship between different sensor measurements and states;

[0153] The factor graph model can be further converted into a least squares problem as follows:

[0154]

[0155] Therefore, the multi-source information fusion navigation problem is converted into a nonlinear least squares problem.

[0156] Step 3, in order to solve the problem that the navigation accuracy of the multi-source information fusion navigation system is often reduced due to external or hardware failures in actual applications, this step proposes a Graph Deviation Network (GDN) autonomous decision-making algorithm in the multi-source navigation information fusion architecture, which performs sensor measurement information fault detection when assisting the multi-sensor factor node concretization and autonomously decides whether to access the sensor measurement information. Step 3 includes the following steps:

[0157] Step 3.1: Data preprocessing when auxiliary sensor factors are concretized

[0158] Since different abstract auxiliary sensor factor nodes will produce quantitative measurements on the states of each dimension during the navigation process, the new information δr generated by step 2.1 k There are often different dimensions and dimensional units between them, which will affect the results of data analysis and further affect the training effect of the GDN network. Therefore, it is necessary to normalize the new information to eliminate the adverse effects caused by singular sample data:

[0159]

[0160] Wherein, d is the normalized new information, μ is the mean of the new information of the abstract auxiliary sensor factor node, and σ is the standard deviation of the new information of the abstract auxiliary sensor factor node.

[0161] Step 3.2: Calculation of the correlation degree of abstract auxiliary sensor factor nodes with nonlinear coupling characteristics and network state transfer of graph deviation

[0162] There are nonlinear connections between different sensor measurements. GDN represents the relationship between the measurement information of abstract auxiliary sensor factor nodes using an embedding vector V i i∈{1,…,7}, for the embedding vector V i It is randomly initialized and updated during the GDN training process to represent the association between the new information of the abstract auxiliary sensor factor nodes.

[0163] The correlation degree v between the abstract auxiliary sensor factor nodes is further calculated by the embedding vector ij as follows,

[0164]

[0165] The different abstract auxiliary sensor factor nodes with strong correlation in measurement data also have a high degree of association. Therefore, when a certain axis of the sensor fails, it can be determined whether the channel has a fault by associating the new information state of the abstract auxiliary sensor factor node.

[0166] Compared with LSTM network, GDN has a similar network structure, but the transmission of its internal hidden layer state is slightly different. It does not require additional storage units and has fewer corresponding parameters. Its definition is that the hidden layer state calculation at time t is as follows:

[0167] z t =sigmod(W z D t +U z h t-1 ) (16)

[0168] r t =sigmod(W r X t +U r h t-1 ) (17)

[0169]

[0170] Among them, D t is the input of GDN at time t, which is composed of the normalized information data d in step 3.1, and the update gate z t Determines the degree to which the unit updates its activation, resetting gate r tThe calculation of W is similar to that of (·) ,U (·) represents the corresponding vector matrix. In addition, the candidate activation It exhibits the same function as a recurrent unit. The activation state of GDN h t The candidate activation and the previous activation state h t-1 The linear interpolation representation of .

[0171] The prediction layer consists of two fully connected layers stacked on top of each other. One layer is the degree of association between sensors as described in step 3.2, and the prediction is based on the last layer of the GDN output, which is calculated as follows:

[0172]

[0173] Where W0 and W1 are parameter matrices, b0 and b1 are vectors, and thus the GDN network state is transferred.

[0174] Step 3.3: Training and validation of graph bias network for calculating nonlinear coupling between multi-sensor factors

[0175] Training data D Tr By train The new information of multi-source sensors generated in time is composed of the new information used for training obtained in step 2. At every point in time Represents the new information δr k The data preprocessing in step 3.1 is used to establish and update the correlation degree of sensor data. The data used in the training set are all normal data without faults.

[0176] The processed and combined training input data D Tr-in Denoted as:

[0177]

[0178] Where w represents the fixed-length sliding window size.

[0179] Training output data D Tr-out is the sensor data of the corresponding training data set at the next moment, recorded as:

[0180]

[0181] The GDN network makes a one-step prediction of the new information value, and the predicted output of the GDN network is compared with the training output D Tr-out The mean square error (MSE) is used to evaluate the training effect of GDN.

[0182] After the GDN network training is completed, the deviation scoring mechanism and threshold need to be determined through the validation set data, that is, the threshold of each channel is determined based on the sensor information of each channel, and the validation set data is used to calculate the threshold. va is the verification time period. In the verification phase, the difference between the pre-verification output value of the GDN network and the sensor information at time t is calculated as:

[0183]

[0184] Among them, d t is the new information of all auxiliary sensors at time t, is the new information predicted by the GDN network. Further, for each dimension e of the above difference vector i (t) are standardized as follows:

[0185]

[0186] in, From t0 to t k The normalized median and quartile of the i-th row of new information at time.

[0187] The standardized difference s i (t) is the deviation score, The value is updated in each prediction step, so as to adaptively adjust the deviation score and threshold to adapt to the latest motion state, and take the maximum deviation score as the threshold λ of fault detection of the corresponding dimension i :

[0188] λ i =max{s i (t0),…,s i (t),…,s i (t va )},t∈{t0,…,t va} (25)

[0189] The setting of the threshold affects the fault detection capability of the system. If the threshold is too large or too small, it will affect the system's robustness. Therefore, in this step, each channel uses a differentiated threshold, and updates the median and quartile of the currently collected data when calculating the deviation score, and adaptively adjusts the deviation scoring mechanism. When the deviation score of each channel exceeds the corresponding threshold, it can be determined that the channel has a fault, thereby realizing multi-channel fault detection and autonomous decision-making.

[0190] Step 3.4: Access the prediction results of the graph deviation network in the multi-sensor fusion architecture

[0191] For the GDN network that has been trained and verified in step 3.3, in the actual application process, an autonomous decision-making scheme is designed, that is, the result prediction of the graph deviation network. First, the new information generated in real time is normalized to obtain the network's prediction input set D Te-in . It is further input into the GDN network to obtain the predicted output Calculate the difference between the GDN network prediction value and the sensor new information as e(t), and then get the deviation score s i (t), the corresponding calculation method is the same as step 3.3.

[0192] After obtaining the deviation score, further compare it with the threshold λ obtained in step 3.3 i contrast:

[0193]

[0194] When the deviation score crosses the threshold, α i =0, then it is determined that a fault has occurred in this dimension, and the system decides not to access the measurement information of this dimension. Otherwise, it is determined that the measurement value of this dimension is normal.

[0195] Step 4: In the actual navigation process, although the fault detection method in step 3 can detect and shield faults, the problem that some sensors present non-Gaussian noise characteristics due to external environment rejection cannot be isolated in time, and the existing factor graph framework for navigation information fusion algorithm often assumes that the noise model of the corresponding factor node is Gaussian, which will lead to certain errors in the navigation information fusion process under special environments. Therefore, this step proposes a graph optimization method for robust intelligent fusion. Step 4 includes the following steps:

[0196] Step 4.1, M-estimation design for robust multi-sensor factor nodes

[0197] Different sensor factor f in step 2 i (x k ) is based on the Gaussian noise assumption. However, in practice, noise often has non-Gaussian characteristics, that is, heavy-tailed noise. This phenomenon is caused by environmental rejection or severe excitement of the body. For this problem, modeling is the first priority. It is universal to consider non-Gaussian noise as Gaussian mixed noise, which has the following form:

[0198]

[0199] where τ is the error model weight satisfying 0≤τ<<1, μ and i∈{1,2} represents the mean and variance of the Gaussian distribution. This distribution is equivalent to the following probability density function:

[0200]

[0201] For the modeling problem of non-Gaussian noise, M estimation is used as the maximum a posteriori probability estimation, which is different from the least squares estimation and is defined as follows:

[0202] minρ(m k ), (29)

[0203] in,

[0204]

[0205] represents the Mahalanobis distance of the measurement error, h(·) represents the nonlinear functional relationship between the measurement and state of different sensors, Σ G is the noise covariance matrix, and ρ(·) is the robustness loss function, which is a positive definite convex function with a unique minimum.

[0206] Based on the iterative weight redistribution algorithm, the above formula can be transformed into:

[0207]

[0208] The adaptive weight ω(m k ) is defined as follows:

[0209]

[0210] ψ(m k )=dρ(m k ) / dm k is called the influence function, and Further using the Huber function, it is defined as follows:

[0211]

[0212] Where c is the positive parameter to be designed.

[0213] At this point, the sensor factor in step 2.3 can be further robust to ω(m k )f i (x k ).

[0214] Step 4.2, Robust Graph Bias Network Optimization Method in Multi-Sensor Fusion Architecture

[0215] The data stored in the factor graph model will grow nonlinearly over time, which will reduce the efficiency of estimating the navigation state. The invention patent adopts an incremental sliding optimization algorithm, which only updates the Bayesian data structure related to the newly added factors, thereby improving the update efficiency.

[0216] For the least squares problem obtained in step 2, after adding autonomous decision-making and robustness factors, we get:

[0217]

[0218] Therefore, the least squares problem mentioned above can be solved incrementally in real time to obtain the multi-source information fusion navigation result.

[0219] Example:

[0220] The embodiment is based on the simulation data of the multi-source information fusion navigation system composed of INS and GNSS, radar, odometer and magnetometer sensors to verify the effectiveness of the navigation information fusion method (RFG) of the present invention. To further illustrate the robustness of the present invention, it is further compared with the factor graph fusion algorithm (FG) without considering non-Gaussian noise. The fusion position results of different algorithms are shown in Figure 2. Figure 2 As shown, the black curve is the true value of the trajectory. It can be seen that the method of the present invention represented by the red solid line has better robustness in the GNSS failure stage. Figure 3 The figure shows a schematic diagram of the yaw angle fault diagnosis of a magnetometer sensor based on GDN. It can be seen that for slowly varying faults, compared with the chi-square and LSTM methods, the method adopted by the present invention can accurately diagnose the fault in real time and generate autonomous decision weights. Figure 4 It indicates that after the multi-source information fusion navigation system is injected with sudden and slow-changing faults, the method of the present invention has lower navigation position error than the other two methods, and the method of the present invention is effective.

[0221] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0222] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0223] It will be appreciated by those skilled in the art that the present invention may relate to a device for performing one or more of the operations described in the present application. The device may be specially designed and manufactured for the desired purpose, or may include a known device in a general-purpose computer that is selectively activated or reconfigured by a program stored therein. Such a computer program may be stored in a device (e.g., computer) readable medium or in any type of medium suitable for storing electronic instructions and coupled to a bus, respectively, the computer readable medium including but not limited to any type of disk (including floppy disk, hard disk, optical disk, CD-ROM, and magneto-optical disk), random access memory (RAM), read-only memory (ROM), electrically programmable ROM, electrically erasable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic card or optical card. The readable medium includes any mechanism for storing or transmitting information in a form readable by a device (e.g., computer). For example, the readable medium includes a random access memory (RAM), a read-only memory (ROM), a magnetic disk storage medium, an optical storage medium, a flash memory device, a signal (e.g., a carrier wave, an infrared signal, a digital signal) propagated in an electrical, optical, acoustic or other form, etc.

[0224] It can be understood by those skilled in the art that each block in these structure diagrams and / or block diagrams and / or flow charts and combinations of blocks in these structure diagrams and / or block diagrams and / or flow charts can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a specialized computer, or a processor of other programmable data processing methods to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing method create a method for implementing the block or multiple blocks specified in the structure diagram and / or block diagram and / or flow chart.

[0225] It is understood by those skilled in the art that the various operations, methods, steps, measures, and schemes in the processes discussed in the present invention may be alternated, changed, combined, or deleted. Furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in the present invention may also be alternated, changed, rearranged, decomposed, combined, or deleted. Furthermore, the steps, measures, and schemes in the prior art that are similar to those disclosed in the present invention may also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0226] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A graph deviation network optimization method for robust intelligent fusion of multi-source navigation information, characterized in that: The method comprises the following steps: Step 1: Analyze the multi-source navigation system and establish i The state variable x of the multi-source navigation system at i ; Step 2: Construct a multi-source navigation system fusion architecture based on the abstract auxiliary sensor factor node; Step 3: Propose the graph deviation network (GDN) autonomous decision-making algorithm in the multi-source navigation system fusion architecture; Step 4: Propose a robust graph optimization algorithm in the multi-source navigation information fusion architecture, design a robust M-estimator to adaptively adjust the fusion weights, and integrate the graph deviation network GDN autonomous decision-making algorithm to achieve robust, stable and efficient fusion of multi-source navigation information.

2. According to claim 1, a graph deviation network optimization method for robust intelligent fusion of multi-source navigation information is characterized in that: The step 1 is: Considering the local northeastern sky coordinate system, established at time t i The state variable x of the multi-source navigation system at i as follows: in represents the attitude angle vector, and are the position and velocity vectors, Represents the gyroscope bias parameter ε used to correct the inertial measurement unit IMU g and accelerometer bias parameters 3. The graph deviation network optimization method for robust intelligent fusion of multi-source navigation information according to claim 1 is characterized in that: The step 2 comprises the following steps: Step 2.1: Establishment and concrete access of abstract auxiliary sensor factor nodes: Abstract factor nodes have horizontal positions Vertical Position Horizontal speed Vertical speed Roll Angle Heading Angle Pitch Angle There are seven types of measurement information, and their measurement equations are described as follows: in is the measurement noise corresponding to different measurement information and obeys Gaussian distribution, H * Indicates measurement information and the multi-source navigation system state variable x k The measurement matrix of the relationship between them is: Get the abstract auxiliary sensor factor node f with unary factor properties * (x k )as follows: It only connects the state variable x at time k k , where d(·) represents the loss function; In order to realize the plug-and-play function of the multi-source fusion navigation system, the abstract auxiliary sensor factor node is assembled according to the actual sensor measurement information to complete the access of the actual sensor measurement information: First, when the actual sensor is connected, the multi-source navigation system first receives the first row of sensor data containing the measurement information attributes, so as to determine the measurement information contained in the connected sensor; and selects the corresponding abstract auxiliary sensor factor node according to the type of measurement information received; Step 2.2: Establishment of other factor nodes: First, define the prior factor with the properties of a unary factor node as follows: Where d(·) represents the loss function, x0 represents the state variable x of the multi-source navigation system i The initial value of f x INS and f α bias They represent the initial value x0 and the initial deviation estimate α0 respectively; Then the multi-source navigation system state variable x i It is further expressed by the following nonlinear discrete equation: where x k is the state variable of the multi-source navigation system at time k, is a given containing the gyroscope f b and accelerometer ω b Output of IMU measurement, h IMU (·) represents the nonlinear measurement function, is different from x k+1 The estimated value of the multi-source navigation system state; the correction parameter α of the inertial measurement unit IMU k It is described by the following nonlinear equation: in represents the predicted value of the correction parameter, and g(·) represents the nonlinear functional relationship of the deviation; Based on the multi-source navigation system state variable x i Nonlinear discrete equation and correction parameter α k Nonlinear equation, INS factor with binary factor properties is as follows: This formula represents the estimated value x k+1 With the predicted value The error function between Similar to the definition of the INS factor, the deviation factor, which also has the characteristics of a binary factor, is defined as follows: It is connected to the correction parameter α k and the predicted amount of the modified parameters Based on the INS factors and deviation factors, the INS factor graph is established; Step 2.3: Establishment of multi-source navigation system fusion architecture based on factor graph model: The fusion architecture of the multi-source navigation system based on the factor graph model is established as follows: The factors are described as follows: in Represents the error between different sensor measurements and states, δr i ∈δr k Represents the corresponding new information, which is defined as follows: where h i (·) represents the nonlinear functional relationship between different sensor measurements and states; The factor graph model can be further converted into a least squares problem as follows: Therefore, the multi-source information fusion navigation problem is converted into a nonlinear least squares problem.

4. The graph deviation network optimization method for robust intelligent fusion of multi-source navigation information according to claim 3 is characterized in that: The step 3 comprises the following steps: Step 3.1: Data preprocessing when abstract auxiliary sensor factors are concretized: In step 2.1, we get the concrete abstract auxiliary sensor factor node f * (x k ) and define the new information δr k =f * (x k ); Then the new information is normalized to eliminate the adverse effects caused by the singular sample data: Where d is the normalized new information, μ is the mean of the new information of the abstract auxiliary sensor factor node, and σ is the standard deviation of the new information of the abstract auxiliary sensor factor node; Step 3.2: Calculation of the correlation degree of abstract auxiliary sensor factor nodes with nonlinear coupling characteristics and state transfer of graph deviation network GDN: The graph deviation network GDN represents the relationship between the abstract auxiliary sensor factor nodes with an embedding vector V i i∈{1,…,7}, the degree of association v between the abstract auxiliary sensor factor nodes is further calculated by the embedding vector ij as follows, Then, the state transfer of the abstract auxiliary sensor factor node is calculated, and the hidden layer state calculation at time t is defined as follows: z t =sigmod(W z D t +U z h t-1 ) r t =sigmod(W r D t +U r h t-1 ) Among them, D t is the input of the graph deviation network GDN at time t, which is composed of the normalized new information d in step 3.1, and the update gate z t represents the update gate, r t Represents the reset gate; W (·) ,U (·) represents the corresponding vector matrix, Represents the activation state h of the candidate activation map bias network GDN t Activated by Candidate and the previous activation state h t-1 Linear interpolation representation of ; The prediction layer consists of two fully connected layers stacked on top of each other, and is calculated as follows: Where W0 and W1 are parameter matrices, b0 and b1 are vectors, and thus the state of the graph deviation network GDN is transferred; Step 3.3: Training and validation of the graph bias network for calculating the nonlinear coupling between multi-sensor factors: Training data D Tr By train The new information generated by the multi-source sensors in time is composed of The kth time point Indicated by the new information δr k Obtained through data preprocessing in step 3.1; The processed and combined training input data D Tr-in Denoted as: Where w represents the sliding window size of fixed length; Training output data D Tr-out is the sensor data of the corresponding training data set at the next moment, recorded as: The graph deviation network GDN makes a one-step prediction of the new information, and the predicted output data of the GDN network is compared with the training output data D Tr-out The mean square error (MSE) is used to evaluate the training effect of the graph deviation network (GDN); The performance of the graph deviation network GDN needs to be verified before it is connected to the multi-source navigation system fusion architecture: take the time t0 to t va is the verification time period. In the verification phase, the difference between the pre-verification output value of the graph deviation network GDN at time t and the sensor's new information is calculated as: Among them, d t is the new information of all abstract auxiliary sensor factor nodes at time t, is the new information predicted by the graph deviation network GDN; further, for each dimension e of the above difference vector i (t) are standardized as follows: in, From t0 to t k The normalized median and quartile of the new information in the i-th row at time; Take the maximum value of the deviation score as the threshold λ for fault detection of the corresponding dimension i : λ i =max{s i (t0),…,s i (t),…,s i (t va )},t∈{t0,…,t va } Step 3.4: Access the prediction results of the map deviation network in the multi-source navigation system fusion architecture: In actual application, the prediction input set D Te-in Input graph deviation network GDN to get the predicted output deviation score s i (t), and further compare it with the threshold λ obtained in step 3.3 i contrast: When the deviation score crosses the threshold, α i =0, then it is determined that the dimension has a fault, and the system decides not to access the measurement information of this dimension. Otherwise, it is determined that the measurement value of this dimension is normal. The prediction result completes the intervention judgment when the multi-sensor factor is concretized.

5. The graph deviation network optimization method for robust intelligent fusion of multi-source navigation information according to claim 4 is characterized in that: The step 4 comprises the following steps: Step 4.1: M-estimation design for robust multi-sensor factor nodes: First, it is universal to consider that non-Gaussian noise is modeled as Gaussian mixture noise, which has the following form: where τ is the error model weight satisfying 0≤τ<<1, μ and Respectively represent the mean and variance of Gaussian distribution; For the modeling problem of non-Gaussian noise, M estimation is used as the maximum a posteriori probability estimation, which is different from the least squares estimation and is defined as follows: minρ(m k ), in, represents the Mahalanobis distance of the measurement error, h(·) represents the nonlinear functional relationship between the measurement and state of different sensors, Σ G is the noise covariance matrix, ρ(·) is the robustness loss function, which is a positive definite convex function with a unique minimum value; based on the iterative weight redistribution algorithm, the above formula can be transformed into: The adaptive weight ω(m k ) is defined as follows: Where c is the positive parameter to be designed; at this time, the sensor factor in step 2.3 can be further robust to ω(m k )f i (x k ); Step 4.2: Robust graph bias network optimization method in multi-source navigation system fusion architecture: For the least squares problem mentioned in step 2, after obtaining the fusion architecture of the concrete multi-source navigation system based on the autonomous decision weight of the graph deviation network in step 3 and the M estimation in step 4.1, we get: Therefore, the least squares problem mentioned above is solved incrementally in real time to obtain the multi-source information fusion navigation result.

Images