CN115014394B - A multi-sensor graph optimization integrated navigation and fault diagnosis method - Google Patents

Abstract

The present invention discloses a multi-sensor graph optimization combined navigation and fault diagnosis method, which relates to the solution and robustness improvement of a GNSS/INS/airspeed meter/altimeter multi-sensor combined navigation system, and belongs to the technical field of calculation and extrapolation. The method obtains system sensor measurements, processes the measurements into INS factors and AVP factors, constructs a system navigation factor graph, and adds a sliding window to ensure the real-time performance of the algorithm. Next, the navigation system is optimized by factor graph optimization to realize the navigation function. On this basis, a fault diagnosis system is constructed by the residual of factor graph optimization, and the existing combined navigation data is classified and trained using a recurrent convolutional neural network to obtain a normal model and a fault model of the system, thereby identifying and diagnosing the faults of the navigation system.

Description

Multi-sensor graph optimized integrated navigation and fault diagnosis method

Technical Field

The invention relates to the technical field of resolving and fault diagnosis of a multi-sensor combined navigation system, in particular to a GNSS/INS/airspeed meter/altimeter multi-sensor diagram optimization combined navigation fault diagnosis algorithm.

Background

Navigation problems have been an important research topic, and have abundant application and development potential in various fields. The navigation system has the advantages of high precision, high fault tolerance, wide application prospect and remarkable technical advantages, such as aircraft cluster formation, aerial refueling, meeting butt joint and the like, become an important direction of current research hot spots and future development, and meanwhile, unmanned, robot formation, mobile position service, intelligent earth and other fast steps come to us. How to effectively improve the navigation precision and high fault tolerance, fully exert the application efficiency of high-performance navigation, and become one of the core problems of the current navigation research. The Global Navigation Satellite System (GNSS) has all-weather positioning, attitude measurement, speed measurement and other capabilities, and can be further fused with an inertial sensor to complement advantages and improve the overall navigation performance. However, the conventional GNSS/INS integrated navigation technology is far from meeting the increasing high-performance navigation requirements.

At present, the GNSS/INS integrated navigation research at home and abroad is mainly based on a Bayesian filtering method. High-performance navigation also often faces a complex unknown dynamic environment, and the traditional combined navigation design has poor compatibility, so that plug and play is difficult to achieve. Factor graph optimization is an emerging optimization theoretical branch, has the capability of processing large-scale constraint, is superior to a filtering method in precision, and has high adaptability in architecture. At present, the graph optimization theory has become a research hotspot of instant localization and map building (SLAM), but is still rarely applied to GNSS navigation and GNSS/INS integrated navigation. However, in the traditional graph optimization algorithm, because modeling optimization is performed on all historical information of the whole system, the corresponding matrix scale is continuously increased along with the time, the calculation amount is continuously increased, the real-time performance of the algorithm is difficult to ensure, a new method is required to be found, and the real-time performance of the algorithm is ensured under the condition that the performance of the algorithm is ensured.

Fault tolerance and diagnosis of faults in high performance relative navigation algorithms is another challenge to be addressed. Due to GNSS non-line-of-sight (NLOS) problems, multipath effects, susceptibility to occlusion, etc. in complex environments, navigation system models have significant nonlinearity, non-Gaussian, and uncertainty. Standard graph optimization frameworks, kalman filtering, etc. are based on gaussian error models, and navigation performance will be significantly degraded in complex environments. At present, researches for improving navigation fault tolerance performance, such as GNSS integrity monitoring and the like, mostly have strict requirements on priori knowledge of a system model, noise statistics characteristics and the like, but the priori information is difficult to acquire in a complex environment, and the requirements on good navigation positioning robustness are difficult to meet.

Regarding fault diagnosis of the integrated navigation system, the traditional method is mainly a fault diagnosis method based on a model, and the fault diagnosis is efficiently completed by establishing a model suitable for the characteristics of data of a fault diagnosis object. The system has the defects that for a system with a complex structure and multiple units, an accurate mathematical model is difficult to build, and the system is easily interfered by environmental factors in the actual operation process, so that other fault diagnosis means are needed to be supplemented, and an ideal fault diagnosis effect can be obtained. The data-based fault diagnosis technology does not need to establish an accurate model of a diagnosis object, can well process the difficulty that the fault diagnosis of a complex nonlinear system is difficult to model, and can well supplement a model-based method.

Disclosure of Invention

Aiming at the defects of the background technology, the invention provides a GNSS/INS/airspeed meter/altimeter graph optimization combined navigation fault diagnosis algorithm, and under the condition of ensuring the optimization precision of a traditional combined navigation system, airspeed meter and altimeter information are introduced, the running efficiency of the system is improved by increasing a sliding window, and the accuracy and fault tolerance performance of the algorithm are improved by more accurate noise modeling and increasing residual error-based switch constraint. On the basis, a new combined navigation fault diagnosis thought is provided by combining the factor graph optimization residual error and MRNN, so that the fault tolerance and diagnosis problems of the combined navigation system on faults are solved.

The invention adopts the following technical scheme for realizing the purposes of the invention:

the multi-sensor map optimized integrated navigation and fault diagnosis method is characterized by comprising the following steps of:

step1, initializing a navigation system, and acquiring initial measurement values of GNSS, INS, an airspeed meter and altimeter sensors to obtain initial values of system measurement;

Step 2, obtaining measured values of all sensors at the current moment, and constructing an observed value based on all the measured values to serve as the input quantity of the navigation system;

Step 3, constructing a factor graph based on an observed value by using a graph optimization method, namely carrying out residual calculation on the observed value to obtain measurement factors, namely INS factors and AVP factors, comparing the residual with a given threshold value, judging and adding a switch constraint, and finishing a factor graph at the current moment;

Step 4, updating a factor graph based on the measurement factors at the moment, and adding a sliding window for the factor graph, so that when a new measurement factor is input, the calculated amount is maintained within a certain range;

Step 5, carrying out nonlinear optimization on INS factors and AVP factors in the window in the step 4, sequentially calculating the state transition probability at the moment corresponding to each factor, the state transition posterior probability of the whole factor graph and the state when the posterior probability is maximum, converting the graph optimization problem into a nonlinear least square problem, and sequentially obtaining an optimal navigation solution;

and 6, constructing a fault diagnosis system based on the residual error optimized by the factor graph in the step 3, and identifying and diagnosing faults of the combined navigation result of the navigation system. Further, the system sensor measurement comprises GNSS measurement, INS measurement, airspeed meter measurement and barometric altimeter measurement, the measurement is processed into INS factors and AVP factors through measurement construction observables, the measurement residual is modeled through maximum posterior estimation, and the observation residual of the system is modeled through a Gaussian mixture model.

Further, in step 2, the sensor measurement values include INS measurement values, GNSS measurement values, airspeed meter measurement values, barometer measurement values, and the measurement values are used to construct an observation value, wherein the observation value includes information of a gesture a, a velocity v, and a position p, and is expressed as z= [ a, v, p ], and the observation value z at the current k moment is constructed as the observation value z under the optimal combinationAndExpressed as:

wherein, The pre-integrated resolved three-dimensional pose is measured for the INS,For the three-dimensional velocity measured by the INS,Three-dimensional position information measured by INS is selected by considering the accuracy advantages of different sensorsThree-dimensional gestures measured for GNSS; three-dimensional speed measured for airspeed meter; Comprising Three components, whereinFor the longitude information measured by the GNSS,For the latitude information measured by the GNSS,Is the height information measured by the height meter.

Further, step 3 constructs INS factors at k time by residual errors respectivelyAnd AVP factorThe method comprises the following steps:

wherein, i·i denote mahalanobis distance, AndRepresents the k-time pose, velocity and position information estimated from the k-1 time according to the dynamics modeling.

Further, the switch constraint is designed in such a way that if the residual error corresponding to a certain measurement factor at the current moment of the system is greater than a threshold value, an additional constraint factor is added to the measurement factor at the moment, so that the influence of the measurement factor on the whole system factor graph is limited and improved.

And for the INS factors outside the window, adopting a ShuerBu algorithm to convert INS factor information before the window into priori information according to operation frequency, namely calculating the priori probability, and reducing the node number of the factor graph, thereby ensuring that the sliding window does not greatly influence the accuracy of the algorithm while reducing the number of the factors, namely the calculated amount. The dynamic sliding window strategy decides to use a ShuerBu algorithm to marginalize factor nodes before a period of time in the factor graph according to the operation frequency, converts the information of the factors into prior information, reduces the node number of the factor graph, and therefore reduces the operation amount of optimization of each factor graph.

Further, the Gaussian mixture model is modeled, the system residual is regarded as the coupling of a plurality of Gaussian noises, the duty ratio weight pi _n of each Gaussian distribution and the parameter mu _n,Σ_n of the independent Gaussian distribution are determined through the parameters of the sensor, and the Gaussian mixture model is built.

Further, step5 is specifically that step 5.1, calculating a state transition probability expression at k moment corresponding to the INS factor and the AVP factor, expressed asWherein, Step 5.2, calculating the state transition posterior probability P (X _k|Z_k) of the whole factor graph, which is expressed as: Wherein K _ins,K_avp is the number of INS and AVP factors, P ₀ is the prior probability calculated according to the factors outside the previous sliding window after the sliding window is added, step 5.3, the state when the posterior probability is maximum is calculated Represented as Wherein, AndX _k, which respectively represent the maximum and minimum of the expression f, is finally obtained

Further, in step 6, based on the fault diagnosis system, a training set is constructed by using the residual error optimized by the factor graph, a memory convolutional neural network MRNN is constructed, parameters of MRNN are obtained through training, then a new factor graph optimized residual error is input MRNN, and the working state of the system at the moment is judged.

Further, the memory convolutional neural network MRNN comprises MRNN blocks which are connected in sequence, wherein the output of the last MRNN block is linearly rectified and then is input into the next MRNN block, and each MRNN block simultaneously extracts the space provided by the convolutional layer and the time information provided by the memory layer, so that the working state of the integrated navigation is diagnosed.

Further, the residuals of the factor graph optimization are respectively expressed asAnd selecting a certain number of residual errors at intervals, wherein the number of the selected residual errors is equal to the classification number adopted in MRNN, respectively calculating the score of each group of residual errors corresponding to each classification type, and using the classification type with the highest final score as the working state of the system.

Further, the working state is mainly divided into normal operation, GNSS positioning abrupt change, GNSS positioning gradual change, airspeed meter speed measurement abrupt change, airspeed meter speed measurement gradual change, altimeter positioning abrupt change, altimeter positioning gradual change and INS bias gradual change.

The memory convolutional neural network MRNN is constructed based on a training set, specifically, for any k time, residuals r _k+1,r_k+2,…,r_k+c-1,r_k+c and MRNN of c continuous time are selected, and the fault diagnosis is expressed as f ^MRNN:

f^MRNN([r_k+1,r_k+2,…,r_k+c-1,r_k+c])＝[s_k+1,s_k+2,…,s_k+c-1,s_k+c]

Wherein s _k+1,s_k+2,…,s_k+c-1,s_k+c represents the working state corresponding to r _k+1,r_k+2,…,r_k+c-1,r_k+c, and comprises normal operation, GNSS positioning abrupt change, GNSS positioning gradual change, airspeed meter speed measurement abrupt change, airspeed meter speed measurement gradual change, altimeter positioning abrupt change, altimeter positioning gradual change and INS bias gradual change.

The technical scheme has the advantages that the GNSS/INS graph optimization fault diagnosis method is improved on the basis of a traditional graph optimization algorithm, the robustness of a navigation system to non-Gaussian errors in measurement is enhanced by means of a Gaussian mixture model, the factor number of the system is reduced by means of a sliding window, the operation amount of the system is further reduced, the operation efficiency of the navigation system is enhanced, real-time requirements of real-time navigation are met, and finally the switching constraint is further increased by means of residual errors, so that fault tolerance performance of the system to fault errors is further enhanced. As a fault diagnosis system, the intelligent method of MRNN training is adopted, and the residual information of each sensor optimized by the graph can be effectively utilized to match and acquire the working state of the integrated navigation system. And through combination of factor graph optimization and MRNN, the problem that INS faults cannot be diagnosed in the traditional fault diagnosis based on a filtering algorithm can be solved, and meanwhile, fault tolerance and fault diagnosis of navigation are achieved.

Drawings

FIG. 1 is a flow chart of an integrated navigation fault diagnosis system according to one embodiment of the present invention;

FIG. 2 is a schematic diagram of a factor graph optimization model according to one embodiment of the present invention;

FIG. 3 is a diagram of MRNN blocks in accordance with one embodiment of the present invention;

Fig. 4 is a schematic diagram of MRNN according to one embodiment of the present invention.

Detailed Description

The technical scheme of the invention is described in detail below with reference to the accompanying drawings.

The invention discloses a GNSS/INS/airspeed meter/altimeter map optimization combined navigation fault diagnosis algorithm, which is used for obtaining system sensor measurement, processing the measurement into INS factors and AVP factors, constructing a system navigation factor map, and further optimizing the factor map of a navigation system to realize a navigation function. Then, a fault diagnosis system is constructed through residual error optimized by the factor graph, and the existing combined navigation data is classified, and a memory convolutional neural network is used for training to obtain a model, so that the fault of the navigation system is identified and diagnosed.

The robust graph optimized integrated navigation algorithm of the invention is described below by taking into account an integrated navigation system consisting of a GNSS, an INS, an airspeed meter and an altimeter carried by a conventional aircraft, and a flow chart of the whole method is shown in FIG. 1, and specifically comprises the following 6 steps.

Step 1, initializing the whole navigation system, initializing priori information at a given initial moment to obtain initial position, speed and attitude information [ A ₀,v₀,p₀ ], and obtaining initial readings of GNSS, INS, airspeed meter and altimeter sensors.

And 2, collecting the observation values of the GNSS, the INS, the airspeed meter and the altimeter at the current moment as system input to obtain the gesture of the navigation system expressed by the current carrier coordinate system and the speed, longitude, latitude and altitude position information of the navigation system in the north-east-earth coordinate system.

The state quantity and observed quantity of the system are then clarified. The observed quantity is collected to update the estimation of the state quantity, and the optimal estimation is obtained through the optimization processing of the estimation, and is taken as the true value of the system state at the moment.

The state quantity x _k of the system selects navigation parameters, which are specifically as follows:

wherein A, v, p, The pose, velocity, position, accelerometer bias and gyroscope bias of the navigation object, respectively.

Without loss of generality, the discussion will be made on time k. Constructing an observed quantity of a system at time kAnd

Wherein, The pre-integrated resolved three-dimensional pose is measured for the INS,For the three-dimensional velocity measured by the INS,Three-dimensional position information measured for INS. Next, taking the accuracy advantages of different sensors into consideration, selectingThree-dimensional gestures measured for GNSS; For measuring three-dimensional velocity of airspeed meter ComprisingThree components, whereinFor the longitude information measured by the GNSS,For the latitude information measured by the GNSS,Is the height information measured by the height meter.

And 3, constructing a factor graph of the navigation system, namely modeling the factor graph of the observed quantity of the system by using a graph optimization method, wherein the factor graph is shown in figure 2. And carrying out residual calculation on the observed value z, constructing a measurement factor at the moment k, comparing the residual with a set threshold value, judging and adding a switch constraint, and finishing a factor graph at the current moment.

The residual error is the difference value between the estimated value and the measurement reading at each moment, and INS factors at k moment are respectively constructed by utilizing the residual errorAnd AVP factor

Wherein, the terms "and" represent the Mahalanobis distance,AndRepresents the k-time pose, velocity and position information estimated from the k-1 time according to the dynamics modeling. Modeling the residual, taking into account that the actual situation is often a non-gaussian error, using a gaussian mixture model.

Modeling a Gaussian mixture model, namely taking the system residual as the coupling of a plurality of Gaussian noises, determining the duty ratio weight pi _n of each Gaussian distribution and the parameter mu _n,Σ_n of the independent Gaussian distribution through the parameters of the sensor, and establishing the Gaussian mixture model.

The Gaussian mixture model is specifically as follows:

Wherein p (x) is a probability function of the Gaussian mixture model, N is the total number of components, N (x|mu _n,Σ_n) is the N-component in the mixture model, each component conforms to the Gaussian distribution, mu _n,Σ_n is the mean vector and covariance matrix of the Gaussian distribution respectively, and in this example, mu _n∈R⁹,Σ_n∈R^9×9;π_n is satisfied to represent the weight of each component, and the weight is satisfied to The gaussian mixture model represents the model distribution obeyed by the residual random variable, which affects the calculation of the covariance matrix needed in the mahalanobis distance, and the variables of the factor graph appear in the form of posterior probability in the final optimization, so the distribution can change the expression of the probability. In this example, 3 gaussian distribution combinations are used to simulate the distribution of the fit residuals, i.e. n=3.

Adding switching constraints to the measurement factors tends to vary greatly from one type of disturbance to another, when a simple expression constructed using Mahalanobis (Mahalanobis) distance results in a system failure, the switching constraints need to be introduced, as shown in fig. 3. The switchable constraint is applied to the residuals at the current k moment by giving a threshold epsilon _ins,ε_avp for each measurement residualWhen the residual error is larger than the threshold value, introducing a new switching factor based on the original factor:

Wherein, without loss of generality, the superscript Q ε { ins, avp } describes a particular measurement factor; Represents a weight function, satisfies Expression of the corresponding measurement factor at time k under switch constraintCan be written as:

And 4, in order to reduce the operation complexity and improve the algorithm real-time performance, adding a sliding window to the factor graph. When new measurements are entered, to ensure that the calculation does not increase over time, the factors that are too large in time interval need to be discarded, thereby maintaining the calculation within a certain range. In this example, the INS factor is limited. The number of INS factors participating in optimization is limited to be less than 1000, an edge method is adopted for factors which are earlier than the 1000 factors in time, a ShuerBu algorithm is used for calculating the prior probability P ₀ of a corresponding sliding window, and therefore the accuracy of the algorithm is not greatly influenced by the sliding window while the number of the factors, namely the calculated amount, is reduced.

And 5, performing nonlinear optimization on the factor graph with the sliding window added. Firstly, calculating a state transition probability expression at k moment corresponding to an INS factor and an AVP factor:

wherein, And respectively corresponding probability normalization parameters. Next, the state transition posterior probability P (X _k|Z_k) of the entire factor graph is calculated:

Wherein K _ins,K_avp is the number of INS and AVP factors, respectively, and P ₀ represents the prior probability calculated from factors outside the previous sliding window after the sliding window is added. In this example, consider that INS is much more frequent than other sensors, with 2000. Gtoreq.K _ins≥K_avp. Next, the state when the posterior probability is maximum is found The method can obtain the following steps:

wherein, AndX _k that maximizes and minimizes the expression f, respectively. Finally, the method can obtain:

and then the graph optimization problem is converted into a nonlinear least square problem, and the optimal navigation solution can be obtained by solving the nonlinear least square problem through a Gaussian-Newton method. And (3) waiting for reading the measurement information at the moment k+1, continuing to return to the step (2), and circulating the steps (2) to (5) to finish the normal navigation operation of the integrated navigation system.

And 6, diagnosing whether the working state of the navigation system is normal or not and whether faults occur or not on the basis of normal operation of the integrated navigation system. The system uses MRNN to judge the working state of the system. The working state mainly comprises normal operation, GNSS positioning abrupt change, GNSS positioning gradual change, airspeed meter speed measurement abrupt change, airspeed meter speed measurement gradual change, altimeter positioning abrupt change, altimeter positioning gradual change and INS bias gradual change.

Firstly, constructing a test set through residual errors, and selecting the residual errors r _k+1,r_k+2,…,r_k+c-1,r_k+c of continuous c times as a training set and a test set of MRNN for any k times.

MRNN the fault diagnosis can be expressed as f ^MRNN:

f^MRNN([r_k+1,r_k+2,…,r_k+c-1,r_k+c])＝[s_k+1,s_k+2,…,s_k+c-1,s_k+c]

Wherein s _k+1,s_k+2,…,s_k+c-1,s_k+c represents the working state corresponding to r _k+1,r_k+2,…,r_k+c-1,r_k+c, and comprises normal operation, GNSS positioning abrupt change, GNSS positioning gradual change, airspeed meter speed measurement abrupt change, airspeed meter speed measurement gradual change, altimeter positioning abrupt change, altimeter positioning gradual change and INS bias gradual change.

Initializing MRNN parameters, training MRNN network by taking a training set as input, updating MRNN network parameters through back propagation when the score is unstable, namely the state classification is unstable, and determining MRNN network parameters when the score tends to be stable, namely the state classification is stable, so as to obtain a trained MRNN network.

The test set is input into a trained MRNN network, and the working state of the system is output so as to judge whether the system is in a normal working state or other fault states. Wherein the working state comprises:

Normal operation, abrupt change in GNSS positioning, gradual change in GNSS positioning, abrupt change in airspeed meter speed measurement, abrupt change in altimeter positioning, gradual change in altimeter positioning, and gradual change in INS bias.

MRNN is similar to other neural networks in basic structure and can be seen as being divided into individual modules, MRNN being shown in figure 3. Each module comprises a convolution layer of the traditional RNN and an additionally added memory layer M layer, the convolution layer reduces the information size, but the semanteme is enhanced greatly, and the memory layer does not change the information size, and the semanteme is enhanced slowly. The overall neural network architecture implementation is shown in fig. 4. By the structure of fig. 4, the neural network can be divided into different layers, and each layer simultaneously extracts the space provided by the convolution layer and the time information provided by the memory layer. The information among multiple layers enters subsequent operation at the same time, which is beneficial to improving the operation speed and reducing the time required by convergence. Through training the training set, the obtained neural network can be used for judging the working state of the system in actual operation, namely fault diagnosis.

Claims (8)

1. The multi-sensor map optimized integrated navigation and fault diagnosis method is characterized by comprising the following steps of:

step1, initializing a navigation system, and acquiring initial measurement values of GNSS, INS, an airspeed meter and altimeter sensors to obtain initial values of system measurement;

step 2, obtaining measured values of each sensor at the current k moment, and constructing an observed value z based on each measured value as the input quantity of the navigation system;

Step 3, constructing a factor graph based on an observed value by using a graph optimization method, namely carrying out residual calculation on the observed value z to obtain measurement factors, namely INS factors and AVP factors, comparing the residual with a given threshold value, judging and adding a switch constraint, and finishing a factor graph at the current moment;

step 4, updating a factor graph based on the measurement factors at the moment k, and adding a sliding window for the factor graph, so that when new measurement factors enter, the calculated amount is maintained within a certain range;

Step 5, carrying out nonlinear optimization on INS factors and AVP factors in the window in the step 4, sequentially calculating the state transition probability at k moment, the state transition posterior probability of the whole factor graph and the state when the posterior probability is maximum, converting the graph optimization problem into a nonlinear least square problem, and sequentially obtaining an optimal navigation solution;

Step 6, constructing a fault diagnosis system based on the residual error optimized by the factor graph in step 3, identifying and diagnosing faults of the combined navigation result of the navigation system, specifically, constructing a training set based on the fault diagnosis system by using the residual error optimized by the factor graph, constructing a memory convolutional neural network MRNN, obtaining parameters of MRNN through training, and then inputting a new factor graph optimized residual error into MRNN to judge the working state of the system at the moment;

The memory convolutional neural network MRNN comprises MRNN blocks which are connected in sequence, wherein the output of the last MRNN block is linearly rectified and then is input into the next MRNN block, and each MRNN block simultaneously extracts the space provided by a convolutional layer and the time information provided by a memory layer, so that the working state of the integrated navigation is diagnosed.

2. The method for optimizing integrated navigation and fault diagnosis according to claim 1, wherein the sensor measurement values in step 2 include INS measurement values, GNSS measurement values, airspeed meter measurement values, and barometric altimeter measurement values, and the measurement values are used to construct an observation value, which includes information of attitude a, velocity v, and position p, expressed as z= [ a, v, p ], and the observation value z at the current k moment is constructed as the optimal combinationAndExpressed as:

wherein, The pre-integrated resolved three-dimensional pose is measured for the INS,For the three-dimensional velocity measured by the INS,Three-dimensional position information measured by INS is selected by considering the accuracy advantages of different sensorsThree-dimensional gestures measured for GNSS; three-dimensional speed measured for airspeed meter; Comprising Three components, whereinFor the longitude information measured by the GNSS,For the latitude information measured by the GNSS,Is the height information measured by the height meter.

3. The multi-sensor map optimized integrated navigation and fault diagnosis method according to claim 2, wherein step 3 constructs INS factors at k time instants respectively using residualsAnd AVP factorThe method comprises the following steps:

wherein, i·i denote mahalanobis distance, AndRepresents the k-time pose, velocity and position information estimated from the k-1 time according to the dynamics modeling.

4. The method for optimizing integrated navigation and fault diagnosis of multiple sensor maps according to claim 3, wherein in step 3, the switch constraint is that if a residual error corresponding to a measurement factor at a current time of the system is greater than a threshold, an additional constraint factor is added to the measurement factor at the time, so as to limit and improve the influence of the measurement factor on the whole system factor map.

5. The multi-sensor graph optimization integrated navigation and fault diagnosis method according to claim 4 is characterized in that in step 4, a sliding window is added for a factor graph, the sliding window is moved in real time, the window size is N, namely the number of INS factors participating in optimization is limited to be N, for INS factors outside the window, a ShuerBunge algorithm is adopted, INS factor information in front of the window is converted into priori information according to operation frequency, namely the prior probability P ₀ is calculated, the node number of the factor graph is reduced, and therefore the sliding window is ensured not to influence the accuracy of the algorithm greatly while the number of factors, namely the calculated amount, is reduced.

6. The method for optimizing integrated navigation and fault diagnosis of multiple sensor maps according to claim 5, wherein step 5 is specifically:

Step 5.1, calculating a state transition probability expression at k moment corresponding to the INS factor and the AVP factor, which is expressed as

Wherein, Respectively corresponding probability normalization parameters;

Step 5.2, calculating the state transition posterior probability P (X _k|Z_k) of the whole factor graph, expressed as:

Wherein K _ins,K_avp is the number of INS and AVP factors respectively, and P ₀ represents the prior probability calculated according to factors outside the previous sliding window after the sliding window is added;

step 5.3, solving the state when the posterior probability is maximum Represented as

Wherein, AndX _k, which respectively represent the maximum and minimum of the expression f, is finally obtained

7. The multi-sensor map optimized integrated navigation and fault diagnosis method according to claim 1, wherein residuals of factor map optimization are respectively expressed asAnd selecting a certain number of residual errors at intervals, wherein the number of the selected residual errors is equal to the classification number adopted in MRNN, respectively calculating the score of each group of residual errors corresponding to each classification type, and using the classification type with the highest final score as the working state of the system.

8. The method for optimizing integrated navigation and fault diagnosis of multiple sensor maps according to claim 7, wherein the construction of the memory convolutional neural network MRNN based on the training set is specifically that, for any k time points, residuals r _k+1,r_k+2,…,r_k+c-1,r_k+c and MRNN of c consecutive time points are selected, and the fault diagnosis is expressed as f ^MRNN:

f^MRNN([r_k+1,r_k+2,…,r_k+c-1,r_k+c])＝[s_k+1,s_k+2,…,s_k+c-1,s_k+c]

Wherein s _k+1,s_k+2,…,s_k+c-1,s_k+c represents the working state corresponding to r _k+1,r_k+2,…,r_k+c-1,r_k+c, and comprises normal operation, GNSS positioning abrupt change, GNSS positioning gradual change, airspeed meter speed measurement abrupt change, airspeed meter speed measurement gradual change, altimeter positioning abrupt change, altimeter positioning gradual change and INS bias gradual change.