CN116147617B - Fault positioning and recovering method for underwater SINS/DVL/PS tightly-integrated navigation system - Google Patents

Abstract

The invention discloses a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system, which includes the following steps: S1. Construct a system virtual model with the same principle as the target experimental equipment, and generate a virtual training set; S2. Use the virtual training set to train the LSTM neural network to obtain the pre-trained model LSTM‑1; S3. Collect a small amount of SINS/DVL/PS tight integrated navigation system experimental data to obtain the experimental training set and migrate LSTM‑1 to experimental applications. Scenario, the final neural network model LSTM‑2 is obtained; S4, the LSTM‑2 neural network enters the classification mode, and the LSTM‑2 neural network model outputs the location of the fault at the current moment based on real-time fault statistics; S5. Based on the fault location results, it automatically Take appropriate failure recovery measures. The present invention adopts the above fault location and recovery method, which can quickly diagnose and locate common faults such as gradual faults under a small amount of experimental data, and automatically implement corresponding fault recovery strategies to improve the reliability of the integrated navigation system.

Description

A fault location and recovery system for underwater SINS/DVL/PS tight integrated navigation system compound method

Technical field

The invention relates to the technical field of navigation system fault diagnosis, and in particular to a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system.

Background technique

A navigation system with high precision and reliability is a necessary condition for autonomous underwater vehicles to successfully complete their missions. Common underwater navigation systems are Strapdown Inertial Navigation System (SINS)/Doppler Log (DVL)/Pressure Sensor (PS) integrated navigation system. Under the interference of the complex underwater environment, DVL and PS sensors may malfunction at any time due to their reliance on environmental information, causing errors in the positioning results of the navigation system to diverge, thereby affecting the work of autonomous underwater vehicles. In this case, users need a set of fault diagnosis technology for underwater integrated navigation systems that can detect the occurrence and disappearance of faults in DVL and PS navigation sensors in real time.

Mutation faults and gradient faults are common faults in navigation systems. Gradient faults are the most difficult to diagnose. To solve the problem of fault diagnosis of gradient faults in navigation sensors, some scholars have used the powerful nonlinear fitting ability of neural networks. and classification capabilities. The specific idea is to use neural networks to fit navigation data for fault diagnosis or to use neural networks to directly divide navigation data into fault-free categories. For example, in the patent application number CN202010380332.2, the name is "A method based on integrated neural networks" In the patent document "INS/GPS Integrated Navigation Fault Detection and Diagnosis Method", the integrated neural network is used as a classification decision function to detect faults in the navigation system; another example is in the patent application number CN201910953487.8, titled "A SVR-based In the patent document "Integrated Navigation System Fault Diagnosis Method", the fitting ability of SVR is used to construct a fault detection function. However, the above invention requires a large amount of navigation experimental data support to train a high-performance neural network model, and the above method does not have the internal capabilities of the sensor. fault locating function.

Contents of the invention

The purpose of the present invention is to provide a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system, which can quickly diagnose and locate common faults such as gradual faults under a small amount of experimental data, and based on The fault location results automatically implement corresponding fault recovery strategies to improve the reliability of the integrated navigation system.

In order to achieve the above objectives, the present invention provides a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system, which includes the following steps:

S1. Construct a virtual model of the SINS/DVL/PS tight integrated navigation system with the same principle as the target experimental equipment, and use the virtual model to generate a virtual training set;

S2. Build an LSTM neural network, and use the virtual training set to train the LSTM neural network to obtain the pre-trained model LSTM-1;

S3. Use the real ship experiment to collect a small amount of SINS/DVL/PS tight integrated navigation system experimental data, use the same method as step S1 to obtain the experimental training set based on the real ship experimental data, and use the transfer learning mechanism to pre-train the model: LSTM- 1. Migrate to the experimental application scenario and obtain the final LSTM neural network model: LSTM-2;

S4, the SINS/DVL/PS tightly integrated navigation system enters the working process, the LSTM-2 neural network enters the classification mode, and the LSTM-2 neural network model outputs the location of the fault at the current moment based on real-time fault statistics;

S5. Based on the fault location results output by the LSTM-2 neural network model, the SINS/DVL/PS tightly integrated navigation system automatically takes corresponding fault recovery measures and uses the Kalman filter to achieve effective fusion of SINS/DVL/PS information.

Further, the SINS/DVL/PS tight integrated navigation system virtual model setting method in step S1:

Obtain the state equation and measurement equation of the tightly integrated navigation system. The state equation: X(k)=φ(k,k-1)X(k-1)+W(k-1) is derived from the error equation of the inertial navigation system. , where X(k) is the state quantity of the navigation system at time k, Where [φ _e ,φ _n ,φ _u ] are the attitude angle errors of the navigation system in the northeast direction, [δv _e ,δv _n ,δv _u ] are the velocity errors of the navigation system in the northeast direction, [δL,δλ,δh ] is the position error of the longitude, latitude and height of the navigation system, [ε _x ,ε _y ,ε _z ] is the three-axis zero bias of the gyroscope,/> is the three-axis zero bias of the accelerometer, δK is the scale factor error of the Doppler velocimeter, φ(k,k-1) represents the state transition matrix of the system from time k-1 to time k, which is determined by the inertial navigation system The error equation is obtained, W is the system noise; the measurement equation: Z(k)=H(k)X(k)+V(k), the quantity measurement in the measurement equation is measured by the actual ground speed of the four DVL beams , the depth measured by the PS sensor and the corresponding value calculated by the inertial navigation system are constructed:/> where Z is the quantity measurement of the navigation system,/> Represents the projection component of the velocity calculated by the inertial navigation in the i-th beam direction of the Doppler sensor, HSINS represents the altitude measurement result of the inertial navigation system,/> represents the speed measurement result of the i-th beam of the Doppler sensor, HPS represents the height measurement result of the depth gauge, H(k) is the measurement matrix at time k, and V is the measurement noise;

The information fusion method applied in the SINS/DVL/PS tight integrated navigation system is the Kalman filter algorithm, which includes time update and parameter update, as shown in the formula:

P(k,k-1)＝φ(k,k-1)P(k-1)φ(k,k-1) ^T +Q(k-1)

K(k)＝P(k,k-1)H(k) ^T (H(k)·P(k,k-1)H(k) ^T +R(k)) ^-1

P(k)＝(I-K(k)H(k))P(k,k-1)

in, is the status update value from time k-1 to time k,/> is the state estimate value at time k-1, P(k,k-1) is the updated value of the covariance matrix from time k-1 to time k, P(k) is the covariance matrix at time k, Q(k) is The system noise matrix at time k, R(k) is the system measurement noise matrix at time k, and K(k) is the gain matrix at time k.

Further, the method of generating a virtual training set in step S1 is:

According to the state equation, measurement equation and Kalman filter algorithm of the experimental equipment SINS/DVL/PS tight integrated navigation system, a set of virtual models of the SINS/DVL/PS tight integrated navigation system with the same principle are constructed. Added glitches to the various beams of the model's Doppler velocimeter and depth gauge: H _PS ^* = H _PS +f, where * represents the fault state that has been added, and f represents the fault that has been artificially added at different locations. The fault location includes: a certain beam of DVL, multiple beams of DVL and depth gauge, and the time when the fault occurs is recorded. time and location;

Using Kalman filter iteration, the residual amount at time k can be obtained: The normalized residual amount is r(k)'=A(k) ^-1/2 r(k), where A(k) is the variance matrix of the residual at time k, and the normalized residual amount is set to the reaction For the navigation statistics of the fault location, the normalized residual sequence generated by the Kalman filter with a fixed time length is stored as a sample, and the fault location at the last moment is stored as the label of the sample to generate a virtual training set.

Further, in step S2, the LSTM neural network consists of an input layer, a long short-term memory layer, a fully connected layer, a softmax layer and an output layer; the long short-term memory layer contains 128 LSTM units, and the forgetting gate and input of the LSTM unit Gates and output gates control the flow of information within the network. The specific calculation formula is as follows:

i _t =σ(U _i h _t-1 +W _i x _t +b _i )

f _t =σ(U _f h _t-1 +W _f x _t +b _f )

o _t =σ(U _o h _t-1 +W _o x _t +b _o )

c _t ＝f _t ⊙c _t-1 +i _t ⊙tanh(U _c h _t-1 +W _c x _t +b _c )

h _t ＝o _t ⊙tanh(c _t )

The subscript t represents time t, i, f and o represent the input gate, forget gate and output gate respectively, c and h are the state of the neuron and the output of the neuron respectively, U, W and b are obtained by the network after training. Parameters, tanh is the activation function, σ is the sigmoid function;

The specific way to generate the pre-training model LSTM-1 is to use the normalized residual sequence in the virtual training set generated in step S1 as the input of the LSTM neural network: X _input = [r(k-l+1)',r (k-l+2)',...,r(k)'], the fault location at the current k moment is used as the label of the network: Y _output , train the LSTM neural network, and obtain the pre-trained model: LSTM-1, store the pre-trained model Training model standby.

Further, the specific process of migrating the pre-trained model to the experimental application scenario based on the transfer learning mechanism in step S3 is as follows:

Retain the pre-trained model: the long short-term memory layer in the LSTM-1 neural network, reset the fully connected layer of LSTM-1, set a smaller learning rate for the long short-term memory layer, and set a larger learning rate for the fully connected layer efficiency, use a small amount of experimental data to migrate the modified neural network to the experimental scenario, that is, the fixed-length normalized residual sequence generated by the experimental data is used as the input of the modified neural network, and the fault location at the last moment is used as the input of the neural network The output, train the modified neural network, and obtain the final LSTM-2 neural network model.

Further, in step S4, in the SINS/DVL/PS tight integrated navigation system work project, the working process of the classification mode of the LSTM-2 neural network is: calculating the normalized residual statistics r generated by the Kalman filter (k)', and store these statistics at a fixed time step l to generate time series data corresponding to moment k: X _input = [r(k-l+1)',r(k-l+2)',...,r(k)'], at this time, the LSTM-2 neural network model trained in step S3 enters the classification mode, inputs time series data to the model, and the model outputs fault location results.

Further, the fault recovery measures in step S5 are: when the fault location result in step S4 is no fault, then continue to follow the default navigation model information fusion in step S1; when the fault location result is that a certain beam of the Doppler speedometer occurs When a fault occurs, according to the inherent relationship between the four beam measurements of the Doppler velocity meter: Use the remaining three beam measurements to replace the faulty beam, taking beam 1 as an example: /> The measurement equation in this case is as follows:/> When multiple beams fail, the Doppler speedometer is isolated. The measurement equation in this case is as follows: Z=[H _SINS -H _PS ]=HX+V. In the integrated navigation system, only the depth gauge measurements are integrated. information; when the depth gauge fails, the depth gauge is isolated. The measurement equation in this case is as follows: In the integrated navigation system, only the measurement information from the Doppler speedometer is integrated.

The advantages and positive effects of the fault location and recovery method for underwater SINS/DVL/PS tight integrated navigation systems described in the present invention are:

1. The present invention is an intelligent fault location method based on transfer learning and long-short-term memory neural network. Compared with the traditional neural network algorithm, the introduced long-short-term memory network can learn time series data of longer time intervals and shorter time intervals at the same time. characteristics, suitable for processing navigation data with time correlation.

2. The present invention uses the normalized residual sequence as the input of the intelligent fault location module, and the fault location as the output of the module. At the same time, it introduces a transfer learning mechanism, so that the method can be used indefinitely when the sample size is not enough to train the model alone. The obtained virtual data assists in training the model, allowing the model to successfully complete the task, solving the problem of difficulty in obtaining experimental data, and has certain engineering application value.

3. The present invention also adds a fault recovery module in the next step of the fault location module, and takes corresponding fault recovery measures based on the fault location results, thereby improving the fault tolerance performance of the system.

The technical solution of the present invention will be further described in detail below through the accompanying drawings and examples.

Description of the drawings

Figure 1 is a flow chart of an embodiment of a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system according to the present invention;

Figure 2 is a structural diagram of a navigation system with a fault location and fault tolerance module according to an embodiment of a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system according to the present invention;

Figure 3 is a speed error curve diagram of an embodiment of a fault location and recovery method for an underwater SINS/DVL/PS tight integrated navigation system according to the present invention.

Detailed ways

The technical solution of the present invention will be further described below through the drawings and examples.

Example

Figure 1 is a flow chart of an embodiment of a fault location and recovery method for an underwater SINS/DVL/PS tight combination navigation system according to the present invention. Figure 2 is a flow chart for an underwater SINS/DVL/PS tight combination navigation system according to the present invention. Structural diagram of the navigation system with fault location and fault tolerance module according to the embodiment of the fault location and recovery method of the navigation system. as the picture shows,

A fault location and recovery method for underwater SINS/DVL/PS tightly integrated navigation system, including the following steps:

S1. Construct a virtual model of the SINS/DVL/PS tight integrated navigation system with the same principle as the target experimental equipment, artificially add faults to the virtual model, and collect a sample set as a virtual training set.

SINS/DVL/PS tight integrated navigation system virtual model setting method:

Obtain the state equation and measurement equation of the tightly integrated navigation system. The state equation: X(k)=φ(k,k-1)X(k-1)+W(k-1) is derived from the error equation of the inertial navigation system. , where X(k) is the state quantity of the navigation system at time k, Where [φ _e ,φ _n ,φ _u ] are the attitude angle errors of the navigation system in the northeast direction, [δv _e ,δv _n ,δv _u ] are the velocity errors of the navigation system in the northeast direction, [δL,δλ,δh ] is the position error of the longitude, latitude and height of the navigation system, [ε _x ,ε _y ,ε _z ] is the three-axis zero bias of the gyroscope,/> is the three-axis zero bias of the accelerometer, δK is the scale factor error of the Doppler velocimeter, φ(k,k-1) represents the state transition matrix of the system from time k-1 to time k, which is determined by the inertial navigation system The error equation is obtained, and W is the system noise. Measurement equation: Z(k)=H(k) And the corresponding values calculated by the inertial navigation system are constructed:/> where Z is the quantity measurement of the navigation system,/> Represents the projection component of the speed calculated by the inertial navigation in the i-th beam direction of the Doppler sensor, H _SINS represents the altitude measurement result of the inertial navigation system, /> represents the speed measurement result of the i-th beam of the Doppler sensor, H _PS represents the height measurement result of the depth gauge, H(k) is the measurement matrix at time k, and V is the measurement noise.

The information fusion method applied in the SINS/DVL/PS tight integrated navigation system is the Kalman filter algorithm, which includes time update and parameter update, as shown in the formula:

P(k,k-1)＝φ(k,k-1)P(k-1)φ(k,k-1) ^T +Q(k-1)

K(k)＝P(k,k-1)H(k) ^T (H(k)·P(k,k-1)H(k) ^T +R(k)) ^-1

P(k)＝(I-K(k)H(k))P(k,k-1)

in, is the status update value from time k-1 to time k,/> is the state estimate value at time k-1, P(k,k-1) is the updated value of the covariance matrix from time k-1 to time k, P(k) is the covariance matrix at time k, Q(k) is The system noise matrix at time k, R(k) is the system measurement noise matrix at time k, and K(k) is the gain matrix at time k.

According to the above equation, the designed ship simulation trajectory, the sensor data generated by the simulation and the Kalman filter principle are used to build a virtual model of the SINS/DVL/PS tight integrated navigation system, and the real-time process quantities and solution results of the integrated navigation system are calculated and generated.

Add faults to each beam and depth gauge of the virtual model's Doppler velocimeter at different times: H _PS ^* = H _PS +f, where * represents the fault state that has been added, and f represents the fault that has been artificially added at different locations. The fault location includes: a certain beam of DVL, multiple beams of DVL and depth gauge, and the time when the fault occurs is recorded. time and location.

Using Kalman filter iteration, the residual amount at time k can be obtained: The normalized residual amount is r(k)'=A(k) ^-1/2 r(k), where A(k) is the variance matrix of the residual at time k, and the normalized residual amount is set to the reaction Navigation statistics for fault locations. The normalized residual sequence generated by the Kalman filter with a fixed time length is stored as a sample, and the fault location at the last moment is stored as the label of the sample to generate a virtual training set. The sample set includes the input amount X _input of the neural network and the label Y _output . The input amount of the neural network is designed to be a time series data of length l composed of the normalized residual amount: X _input = [r(k-l+1 )',r(k-l+2)',...,r(k)'], the label is designed as the fault location at the current moment (k moment): Y _output .

S2. Build an LSTM neural network, and use the virtual training set to train the LSTM neural network to obtain the pre-trained model LSTM-1.

The LSTM neural network consists of an input layer, a long short-term memory layer, a fully connected layer, a softmax layer and an output layer; the long-short-term memory layer contains 128 LSTM units. The forgetting gate, input gate and output gate of the LSTM unit control the inside of the network. The specific calculation formula for the flow of information is as follows:

i _t =σ(U _i h _t-1 +W _i x _t +b _i )

f _t =σ(U _f h _t-1 +W _f x _t +b _f )

o _t =σ(U _o h _t-1 +W _o x _t +b _o )

c _t ＝f _t ⊙c _t-1 +i _t ⊙tanh(U _c h _t-1 +W _c x _t +b _c )

h _t ＝o _t ⊙tanh(c _t )

The subscript t represents time t, i, f and o represent the input gate, forgetting gate and output gate respectively, _c and h are the state of the neuron and the output of the neuron respectively, U, W and b are obtained by the network after training. Parameters, tanh is the activation function, σ is the sigmoid function;

The specific way to generate the pre-training model LSTM-1 is to use the normalized residual sequence in the virtual training set generated in step S1 as the input of the LSTM neural network: X _input = [r(k-l+1)',r (k-l+2)',...,r(k)'], the fault location at the current time (k time) is used as the label of the network: Y _output , the Adam optimization algorithm is used to train the LSTM neural network, and the pre-trained model is obtained :LSTM-1, stores the pre-trained model for later use.

S3. Use the real ship experiment to collect a small amount of SINS/DVL/PS tight integrated navigation system experimental data, use the same method as step S1 to inject faults into the experimental data, and collect the sample set as the experimental training set. The input volume of the same network is designed as _: Based on the transfer learning mechanism, the pre-trained model: LSTM-1 was migrated to the experimental application scenario, and the final LSTM neural network model was obtained: LSTM-2.

Retain the pre-trained model: the long short-term memory layer in the LSTM-1 neural network, reset the fully connected layer of LSTM-1, set a smaller learning rate (0.0001) for the long short-term memory layer, and set a smaller learning rate for the fully connected layer. A large learning rate (0.001), using a small amount of experimental data to migrate the modified neural network to the experimental scenario, that is, the fixed-length normalized residual sequence generated by the experimental data is used as the input of the modified neural network. At the last moment The fault location is used as the output of the neural network, the modified neural network is trained to obtain the final LSTM-2 neural network model, and the model is saved for later use.

S4, the SINS/DVL/PS tight integrated navigation system enters the working process, the LSTM-2 neural network enters the classification mode, and the LSTM-2 neural network model outputs the location of the fault at the current moment based on real-time fault statistics.

In the SINS/DVL/PS tight integrated navigation system work project, the normalized residual statistics r(k)' generated by the Kalman filter are calculated, and these statistics are stored with a fixed time step l to generate k moments Corresponding time series _data : -2 The neural network model enters the classification mode as a fault location determiner, and inputs the time series data into the fault location determiner. The determiner will generate the location of the current navigation sensor fault (or no fault).

S5. Based on the fault location results output by the LSTM-2 neural network model, the SINS/DVL/PS tightly integrated navigation system automatically takes corresponding fault recovery measures and uses the Kalman filter to achieve effective fusion of SINS/DVL/PS information.

When the fault location result in step S4 is no fault, then continue to follow the default navigation model information fusion in step S1; when the fault location result is that a beam of the Doppler speedometer is faulty, according to the four beam quantities of the Doppler speedometer The inherent relationship of measurement: Use the remaining three beam measurements to replace the faulty beam, taking beam 1 as an example: /> The measurement equation for this case is as follows: When multiple beams fail, the Doppler speedometer is isolated. The measurement equation in this case is as follows: Z=[H _SINS -H _PS ]=HX+V. In the integrated navigation system, only the depth gauge measurements are integrated. information; when the depth gauge fails, the depth gauge is isolated. The measurement equation in this case is as follows:/> In the integrated navigation system, only the measurement information from the Doppler speedometer is integrated.

This embodiment illustrates the effectiveness of the algorithm by combining an example of navigation data collected through experiments. The experimental parameters are set as follows: assume that the zero bias of each axis gyroscope is 0.2°/h, the angle random walk is 0.15°/h, the accelerometer zero bias is 500 μg, and the speed random walk is The random noise of the Doppler velocimeter is 0.01m/s, the scaling factor is 0.1%, and the random noise of the pressure gauge is 2m (due to the limitations of objective experimental conditions, the data of the Doppler velocimeter and pressure gauge are artificial reference data. obtained by adding noise). The experimental training set samples are sample data with a sailing time of 30s, and the test set samples are navigation data with a sailing time of 500s. From 101 to 129s, 201 to 229s, 301 to 329s and 401 to 429s, gradient faults with a growth rate of 0.5m/s ² are applied to the four beams of the Doppler velocimeter respectively. As can be seen from Figure 3, in the small sample The fault location and recovery module under the data reduces the impact of the Doppler speedometer fault on the navigation system solution results.

Therefore, the present invention adopts the above-mentioned fault location and recovery method for the underwater SINS/DVL/PS tight integrated navigation system, which can quickly diagnose and locate common faults such as gradual faults under the condition of a small amount of experimental data, and locate the fault according to the fault location. As a result, corresponding fault recovery strategies are automatically implemented to improve the reliability of the integrated navigation system.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: The technical solution of the present invention may be modified or equivalently substituted, but these modifications or equivalent substitutions cannot cause the modified technical solution to depart from the spirit and scope of the technical solution of the present invention.

Claims (7)

1. The fault locating and recovering method for the underwater SINS/DVL/PS tightly-integrated navigation system is characterized by comprising the following steps:

s1, constructing a set of SINS/DVL/PS tightly-combined navigation system virtual model with the same principle as that of target experimental equipment, and generating a virtual training set by using the virtual model;

s2, building an LSTM neural network, and training the LSTM neural network by utilizing a virtual training set to obtain a pre-training model LSTM-1;

s3, acquiring a small amount of SINS/DVL/PS tightly-integrated navigation system experimental data by using a real ship experiment, obtaining an experimental training set based on the real ship experimental data by using the same method as the step S1, and pre-training a model based on a migration learning mechanism: and migrating the LSTM-1 to an experimental application scene to obtain a final LSTM neural network model: LSTM-2;

s4, enabling the SINS/DVL/PS tightly-combined navigation system to enter a working process, enabling the LSTM-2 neural network to enter a classification mode, and enabling the LSTM-2 neural network model to output the fault occurrence position at the current moment according to real-time fault statistics;

s5, according to a fault positioning result output by the LSTM-2 neural network model, the SINS/DVL/PS tightly combined navigation system automatically takes corresponding fault recovery measures, and effective fusion of SINS/DVL/PS information is realized by using a Kalman filter.

2. The fault locating and recovering method for the underwater SINS/DVL/PS compact navigation system according to claim 1, wherein the SINS/DVL/PS compact navigation system virtual model setting method in step S1 is as follows:

acquiring a state equation and a measurement equation of the tightly combined navigation system, wherein the state equation is as follows: x (k) =phi (k, k-1) X (k-1) +w (k-1) derives from the error equation of the inertial navigation system, where X (k) is the state quantity of the navigation system at time k,wherein [ phi ] _e ,φ _n ,φ _u ]For the attitude angle error of the navigation system in the northeast direction, [ δv ] _e ,δv _n ,δv _u ]For speed error of navigation system in northeast direction, [ delta L, delta lambda, delta h]For position error of longitude and latitude altitude of navigation system, [ epsilon ] _x ,ε _y ,ε _z ]Is the three-axis zero offset of the gyroscope, +.>For the triaxial zero offset of the accelerometer, δK is the scale factor error of the Doppler velocimeter, phi (K, K-1) represents the state transition matrix of the system from K-1 time to K time, the state transition matrix is obtained by the error equation of the inertial navigation system, and W is the system noise; measurement equation: z (k) =h (k) X (k) +v (k), the measurement in the measurement equation is constructed from the real measurement of the earth speed of the four beams of DVL, the depth measured by the PS sensor, and the corresponding values solved by the inertial navigation system: />Wherein Z is the measurement of the navigation system, +.>Representing the projected component of the velocity of the inertial navigation solution in the direction of the ith beam of the Doppler sensor, H _SINS Representing altitude measurements of an inertial navigation system, < >>Representing the speed measurement result of the ith wave beam of the Doppler sensor, H _PS Representing the height measurement result of the depth gauge, wherein H (k) is a measurement matrix at k moment, and V is measurement noise;

the information fusion method applied by the SINS/DVL/PS tightly-integrated navigation system is a Kalman filtering algorithm, and comprises time updating and parameter updating, and is shown as the formula:

P(k,k-1)＝φ(k,k-1)P(k-1)φ(k,k-1) ^T +Q(k-1)

K(k)＝P(k,k-1)H(k) ^T (H(k)·P(k,k-1)H(k) ^T +R(k)) ^-1

P(k)＝(I-K(k)H(k))P(k,k-1)

wherein,update value for the state from moment k-1 to moment k,/>For the state estimation value at time K-1, P (K, K-1) is the updated value of the covariance matrix from time K-1 to time K, P (K) is the covariance matrix at time K, Q (K) is the system noise matrix at time K, R (K) is the system measurement noise matrix at time K, and K (K) is the gain matrix at time K.

3. The fault locating and recovering method for underwater SINS/DVL/PS tightly-integrated navigation system according to claim 2, wherein the method for generating the virtual training set in step S1 is as follows:

according to a state equation, a measurement equation and a Kalman filtering algorithm of an SINS/DVL/PS tightly combined navigation system of experimental equipment, constructing a set of SINS/DVL/PS tightly combined navigation system virtual model with the same principle, and adding faults to each beam and depth meter of a Doppler velocimeter of the virtual model at different moments:H _PS ^* ＝H _PS +f, where x represents the added fault state, f represents the fault artificially added to a different location, the fault location comprising: a certain beam of the DVL, a plurality of beams of the DVL and a depth gauge, and recording the occurrence time and the occurrence position of faults;

the residual quantity at the moment k can be obtained by iteration of Kalman filtering:the normalized residual amount is r (k)' =a (k) ^-1/2 r (k), where A (k) is the variance matrix of the time k residualsSetting the normalized residual quantity as navigation statistics reflecting the fault position, storing a normalized residual sequence generated by a Kalman filter for a fixed time length as one sample, storing the fault position at the last moment as a label of the sample, and generating a virtual training set.

4. The fault locating and recovering method for underwater SINS/DVL/PS integrated navigation system according to claim 3, wherein the LSTM neural network in step S2 is composed of an input layer, a long-short-term memory layer, a full-connection layer, a softmax layer and an output layer; the long-term memory layer comprises 128 LSTM units, and forgetting gates, input gates and output gates of the LSTM units control the flow of information in the network, and a specific calculation formula is as follows:

i _t ＝σ(U _i h _t-1 +W _i x _t +b _i )

f _t ＝σ(U _f h _t-1 +W _f x _t +b _f )

o _t ＝σ(U _o h _t-1 +W _o x _t +b _o )

c _t ＝f _t ⊙c _t-1 +i _t ⊙tanh(U _c h _t-1 +W _c x _t +b _c )

h _t ＝o _t ⊙tanh(c _t )

wherein the subscript t represents the time t, i, f and o represent an input gate, a forgetting gate and an output gate respectively, c, h are the states of neurons and the outputs of the neurons respectively, U, W and b are parameters obtained by training a network, tanh is an activation function, and sigma is a sigmoid function;

the specific mode of generating the pre-training model LSTM-1 is that the normalized residual sequence in the virtual training set generated in the step S1 is used as the input quantity of the LSTM neural network: x is X _input ＝[r(k-l+1)',r(k-l+2)',...,r(k)']The fault location at the current time k is used as a label of the network: y is Y _output And training the LSTM neural network to obtain a pre-training model LSTM-1, and storing the pre-training model for later use.

5. The fault locating and recovering method for the underwater SINS/DVL/PS tightly-integrated navigation system according to claim 4, wherein the specific process of migrating the pre-training model to the experimental application scene based on the migration learning mechanism in the step S3 is as follows:

the method comprises the steps of reserving a pre-training model, namely, a long-period memory layer in an LSTM-1 neural network, resetting a full-connection layer of the LSTM-1, setting a small learning rate for the long-period memory layer, setting a large learning rate for the full-connection layer, and transferring the modified neural network to an experimental scene by using a small amount of experimental data, namely, taking a normalized residual sequence with a fixed length generated by the experimental data as input of the modified neural network, taking the fault position at the last moment as output of the neural network, and training the modified neural network to obtain a final LSTM-2 neural network model.

6. The fault locating and recovering method for underwater SINS/DVL/PS integrated navigation system according to claim 5, wherein in step S4, in the SINS/DVL/PS integrated navigation system working engineering, the classification mode working process of the LSTM-2 neural network is as follows: calculating normalized residual statistics r (k)' generated by a Kalman filter, storing the statistics with a fixed time step l, and generating time sequence data X corresponding to k moment _input ＝[r(k-l+1)',r(k-l+2)',...,r(k)']At this time, the trained LSTM-2 neural network model in the step S3 enters a classification mode, time series data is input into the model, and the model outputs a fault positioning result.

7. The fault locating and recovering method for underwater SINS/DVL/PS tightly integrated navigation system according to claim 6, wherein the fault recovering measure in step S5 is: when the fault positioning result in the step S4 is no fault, continuing to fuse the navigation model information according to the default in the step S1; when the fault positioning result is that a certain beam of the Doppler velocimeter fails, according to the inherent relation of measurement of four beams of the Doppler velocimeter:replacing the fault beam quantity by using the rest three beam quantities; when a plurality of beams fail, the Doppler velocimeter is isolated, and the measurement equation of the situation is as follows: z= [ H ] _SINS -H _PS ]=hx+v, only the measurement information of the depth gauge is fused in the integrated navigation system; when the depth gauge fails, the depth gauge is isolated, and the measurement equation of the situation is as follows: />Only the measurement information of the Doppler velocimeter is fused in the integrated navigation system.