A method of error prediction for an inertial navigation device

Abstract

The application discloses a kind of error prediction of inertial navigation equipment, belong to inertial positioning technical field, including data acquisition and preprocessing, error modeling and preliminary fusion, LSTM modeling based on physical constraint, LSTM optimization, error correction and feedback optimization, solve the technical problem of improving the real-time and accuracy of navigation system, the application fuses IMU drift formula and environmental temperature error modeling, enhances prediction accuracy and endows model with physical explanation, improves the adaptability of system to dynamic environment, uses real-time error as reward signal, dynamically optimizes LSTM weight and covariance matrix, improves the prediction ability under long time work, feedback optimization is carried out to the corrected navigation data, realizes model self-learning and performance iteration.

Description

Technical Field

[0001] This invention belongs to the field of inertial positioning technology, and particularly relates to a method for error prediction of inertial navigation devices. Background Technology

[0002] Inertial navigation systems (INS) combine IMU sensors and GNSS receivers to achieve precise navigation. However, IMU sensors are susceptible to drift errors, GNSS signals are unstable in obstructed environments, and changes in ambient temperature can also affect navigation accuracy. Current technologies for inertial navigation often employ Kalman filtering to fuse data or rely on data-driven deep learning models, but both have shortcomings:

[0003] 1. Traditional Kalman filters cannot adapt to complex dynamic environments in real time.

[0004] 2. Purely data-driven models lack physical constraints, are difficult to interpret, and are prone to failure under external disturbances. Summary of the Invention

[0005] The purpose of this invention is to provide a method for error prediction of inertial navigation devices, which solves the technical problem of improving the real-time performance and accuracy of navigation systems by directly embedding the IMU physical drift formula and environmental error into the LSTM model, combined with dynamic adaptive mechanisms and reinforcement learning optimization.

[0006] To achieve the above objectives, the present invention employs the following technical invention:

[0007] A method for error prediction of an inertial navigation device includes the following steps:

[0008] Step 1: The data acquisition module acquires IMU data P from the IMU sensor, GNSS receiver, and environmental sensor respectively. IMU (t), GNSS positioning data P GNSS (t) and ambient temperature data T env (t), and preprocessed; IMU data P IMU (t) includes acceleration a(t) and angular velocity ω(t);

[0009] Preprocessing includes processing IMU data P IMU (t) performs noise filtering on the ambient temperature data T env (t) is smoothed and the IMU data P is aligned using timestamps. IMU (t) and GNSS positioning data P GNSS (t), and interpolate to extend GNSS positioning data P GNSS (t) to IMU data P IMU The time frequency of (t); the time interval error E ΔtPerform mathematical modeling; combine the preprocessed data with the time interval error E Δt By combining these features, we obtain the feature vector X(t).

[0010] Step 2: The data fusion module processes the IMU data. IMU The drift error E of (t) IMU (t) and the temperature error E of the ambient temperature data T Mathematical modeling was performed separately, based on the drift error E. IMU (t), Temperature error E T and time interval error E Δt This generates an initial fusion error E. 融合 (t); the drift error E IMU (t) and temperature error E T Add it to the feature vector X(t);

[0011] Step 3: The LSTM computation module retrieves the feature vector X(t) and the fusion error E. 融合 (t), using an LSTM model combined with physical constraints to predict sensor errors, the error prediction result E is obtained. LSTM (t);

[0012] Step 4: Optimize the module to obtain real-time ambient temperature data T env (t), IMU signal-to-noise ratio (SNR) IMU And error prediction results E LSTM (t), establish a dynamic adaptive mechanism to adjust the noise covariance matrix according to environmental changes, use reinforcement learning optimization to establish a reward function, dynamically adjust LSTM weights and noise covariance matrix, obtain optimized LSTM model weights and dynamically adjusted noise covariance matrix, and improve the real-time performance of error prediction and correction.

[0013] Step 5: Correct the inertial data using the prediction error, and continuously optimize the model through a feedback mechanism to output high-precision navigation data.

[0014] Preferably, when performing step 1, the feature vector X(t) is represented by the following formula:

[0015] X(t)=[a(t),ω(t),P GNSS (t),T env (t),E Δt ];

[0016] Time interval error E Δt It is calculated using the following formula:

[0017] E Δt = β × Δt;

[0018] Where Δt is the time interval, representing the IMU data P IMU (t) and GNSS positioning data P GNSS (t) is the time synchronization error at the same moment; β is the time error influence coefficient, representing the influence of time delay between sensors.

[0019] Preferably, when performing step 2, the specific steps are as follows:

[0020] Step 2-1: Process IMU data P IMU The drift error E of (t) IMU (t) modeling, as shown in the following formula:

[0021] E IMU (t)=E IMU (t−1)+α×Δt;

[0022] Among them, E IMU (t) represents the IMU error at the current time t, E IMU (t−1) represents the IMU error at the previous time point t-1, α represents the drift rate, and Δt represents the time interval;

[0023] Step 2-2: Temperature error E of ambient temperature data T The modeling is illustrated in the following formula:

[0024] E T = γ × (T env (t)− T0);

[0025] Among them, T env (t) represents the ambient temperature, T0 represents the calibration temperature of the ambient sensor, and γ represents the temperature sensitivity coefficient; the ambient sensor is a temperature sensor.

[0026] Steps 2-3: Calculate the drift error E IMU (t), Temperature error E T and time interval error E Δt The fusion is performed to obtain the fusion error E. 融合 (t):

[0027] E 融合 (t) = E IMU (t) + E T + E Δt ;

[0028] Steps 2-4: Calculate the drift error E IMU (t) and temperature error E T After being added to the eigenvector X(t), the eigenvector X(t) is represented as follows:

[0029] X(t)=[a(t),ω(t),P GNSS (t),T env (t),E IMU (t),E T E Δt ].

[0030] Preferably, when performing step 3, the specific steps are as follows:

[0031] Step 3-1: Feature input, specifically, the feature vector X(t) obtained in step 2-4 is retrieved as the feature vector input to the input layer of the LSTM model;

[0032] Step 3-2: Embed physical constraints in the LSTM model, specifically by incorporating the drift error formula E IMU (t)=E IMU (t−1)+α×Δt and temperature error formula E T = γ × (T env (t) − T0) is directly embedded in the input and gating mechanism of the LSTM model;

[0033] Introduce the drift rate α and time interval β into the forget gate;

[0034] LSTM model fitting prediction results E LSTM1 (t);

[0035] Step 3-3: Physical constraint regularization, specifically by introducing a regularization term L. phy Constraining the LSTM loss function:

[0036] L phy = λ phy × ||E L (t) − (E IMU (t) + E T + E Δt )|| 2 ;

[0037] Among them, E L (t) represents the error calculation for the intermediate process, used to constrain the degree to which the LSTM model conforms to physical laws; λ phy It is the regularization coefficient, which adjusts the strength of physical error constraints;

[0038] Physical constraint regularization is used to encourage prediction results E LSTM1 (t) satisfies physical constraints;

[0039] Steps 3-4: Combine the error loss and physical regularization loss to optimize the LSTM model weights and generate the final prediction result E. LSTM (t), the total loss function L of LSTM is:

[0040] L total = L data + λL phy ;

[0041] L data = ||E true (t) − E LSTM1 (t)|| 2 ;

[0042] Where λ is the weighting factor, L data It is based on data-driven error loss.

[0043] Preferably, step 4 includes the following steps:

[0044] Step 4-1: Obtain real-time ambient temperature data T env (t), IMU signal-to-noise ratio (SNR) IMU And the predicted result E LSTM (t);

[0045] Step 4-2: Establish a dynamic adaptive mechanism to adjust the noise covariance matrix according to environmental changes:

[0046] Q adaptive = Q0 + k1 × ∣T env (t) − T0∣;

[0047] R adaptive = R0 + k2 × SNR IMU ;

[0048] Among them, Q adaptive It is the process noise covariance matrix, which is affected by temperature changes; R adaptive R0 is the measurement noise covariance matrix, which is affected by the signal-to-noise ratio; k1 and k2 are both adjustment coefficients; Q0 is the initial value of the process noise covariance matrix; R0 is the initial value of the measurement noise covariance matrix.

[0049] Step 4-3: Establish the reward function Reward using reinforcement learning optimization methods:

[0050] Reward = −||E true (t) − E LSTM (t)|| 2 ;

[0051] The LSTM model and noise covariance matrix are optimized by dynamically adjusting the LSTM weights and Kalman filter covariance matrix through the reward function Reward.

[0052] Preferably, when performing step 5, the IMU data P is processed according to the following formula.IMU (t) is corrected to obtain the corrected navigation data P. corrected (t):

[0053] P corrected (t) = P IMU (t) − E LSTM (t);

[0054] The corrected data is used to update the parameters of the LSTM model and Kalman filter, and a feedback mechanism is used to continuously optimize the model, outputting navigation data P. corrected (t).

[0055] This invention provides a method for error prediction in inertial navigation devices. It addresses the technical challenge of improving the real-time performance and accuracy of navigation systems by directly embedding the IMU physical drift formula and environmental errors into the LSTM model, combined with dynamic adaptive mechanisms and reinforcement learning optimization. This invention integrates the IMU drift formula with environmental temperature error modeling, enhancing prediction accuracy and providing the model with physical interpretability. It adjusts the noise covariance matrix based on environmental temperature changes and the IMU signal-to-noise ratio, improving the system's adaptability to dynamic environments. Using real-time errors as a reward signal, it dynamically optimizes the LSTM weights and covariance matrix, enhancing prediction capabilities over long periods. Feedback optimization of the corrected navigation data enables model self-learning and performance iteration. Attached Figure Description

[0056] Figure 1 This is the main flowchart of the present invention;

[0057] Figure 2 This is a flowchart of step 1 of the present invention;

[0058] Figure 3 This is a flowchart of step 2 of the present invention;

[0059] Figure 4 This is a flowchart of step 3 of the present invention;

[0060] Figure 5 This is a flowchart of step 4 of the present invention;

[0061] Figure 6 This is a flowchart of step 5 of the present invention. Detailed Implementation

[0062] Depend on Figures 1-6 The method for error prediction of an inertial navigation device, as shown, includes the following steps:

[0063] Step 1: The data acquisition module acquires IMU data P from the IMU sensor, GNSS receiver, and environmental sensor respectively. IMU (t), GNSS positioning data PGNSS (t) and ambient temperature data T env (t), and preprocessed; IMU data P IMU (t) includes acceleration a(t) and angular velocity ω(t);

[0064] Preprocessing includes processing IMU data P IMU (t) performs noise filtering on the ambient temperature data T env (t) is smoothed and the IMU data P is aligned using timestamps. IMU (t) and GNSS positioning data P GNSS (t), and interpolate to extend GNSS positioning data P GNSS (t) to IMU data P IMU The time frequency of (t); the time interval error E Δt Perform mathematical modeling; combine the preprocessed data with the time interval error E Δt By combining these features, we obtain the feature vector X(t).

[0065] When performing step 1, the eigenvector X(t) is represented by the following formula:

[0066] X(t)=[a(t),ω(t),P GNSS (t),T env (t),E Δt ];

[0067] Time interval error E Δt It is calculated using the following formula:

[0068] E Δt = β × Δt;

[0069] Where Δt is the time interval, representing the IMU data P IMU (t) and GNSS positioning data P GNSS (t) is the time synchronization error at the same moment; β is the time error influence coefficient, representing the influence of time delay between sensors.

[0070] In applications, the data sampling frequencies of IMU sensors and GNSS receivers are typically different. IMU sensors have a higher sampling frequency (e.g., 100 Hz or higher), while GNSS receivers have a lower sampling frequency (e.g., 1 Hz). Because the time bases for data acquisition are not completely consistent, this can lead to time discrepancies.

[0071] The internal clocks of IMU sensors and GNSS receivers may drift or become out of sync, which can also lead to inconsistent timestamps.

[0072] Environmental factors (such as temperature) can also affect the internal clock, causing it to drift.

[0073] When interpolating GNSS data to IMU frequencies, additional time errors may be introduced due to the imperfections of the interpolation algorithm.

[0074] This invention corrects the above errors, reducing the time synchronization error Δt, thereby obtaining more accurate positioning data.

[0075] In this embodiment, the timestamp t of each IMU data is... IMU and the corresponding GNSS data timestamp t GNSS Compare them.

[0076] Because GNSS data has a low frequency, each data point of the IMU may correspond to multiple GNSS data points. In this case, this embodiment uses interpolated GNSS data to align to the IMU's sampling time.

[0077] The time interval Δt is the result of each comparison; it is the difference between the IMU sampling time and the timestamp of the corresponding GNSS data, i.e.:

[0078] Δt=t IMU -t GNSS ;

[0079] If the GNSS data does not perfectly match the IMU timestamp, interpolation can be performed on the GNSS data to obtain the closest timestamp t. GNSS .

[0080] In this embodiment, various sensor data are standardized and aligned through preprocessing to provide a unified data input format for subsequent modeling. Noise filtering and data smoothing improve data quality and reduce the interference of random errors on subsequent modeling. By interpolating and extending the GNSS data frequency to match the IMU frequency, the problem of time resolution mismatch is reduced. The modeling of time interval error directly addresses the time asynchrony problem between sensors, improving the accuracy of error prediction.

[0081] Step 2: The data fusion module processes the IMU data. IMU The drift error E of (t) IMU (t) and the temperature error E of the ambient temperature data T Mathematical modeling was performed separately, based on the drift error E. IMU (t), Temperature error E T and time interval error E Δt This generates an initial fusion error E. 融合 (t); the drift error E IMU (t) and temperature error E TAdd it to the feature vector X(t);

[0082] When performing step 2, the specific steps are as follows:

[0083] Step 2-1: Process IMU data P IMU The drift error E of (t) IMU (t) modeling, as shown in the following formula:

[0084] E IMU (t)=E IMU (t−1)+α×Δt;

[0085] Among them, E IMU (t) represents the IMU error at the current time t, E IMU (t−1) represents the IMU error at the previous time point t-1, α represents the drift rate, and Δt represents the time interval;

[0086] Step 2-2: Temperature error E of ambient temperature data T The modeling is illustrated in the following formula:

[0087] E T = γ × (T env (t) − T0);

[0088] Among them, T env (t) represents the ambient temperature, T0 represents the calibration temperature of the ambient sensor, and γ represents the temperature sensitivity coefficient; the ambient sensor is a temperature sensor.

[0089] Steps 2-3: Calculate the drift error E IMU (t), Temperature error E T and time interval error E Δt The fusion is performed to obtain the fusion error E. 融合 (t):

[0090] E 融合 (t) = E IMU (t) + E T + E Δt ;

[0091] Steps 2-4: Calculate the drift error E IMU (t) and temperature error E T After being added to the eigenvector X(t), the eigenvector X(t) is represented as follows:

[0092] X(t)=[a(t),ω(t),P GNSS (t),T env (t),E IMU (t),E T E Δt ].

[0093] In this embodiment, physical characteristics and environmental factors are decomposed into independent models, which helps to analyze the sources of error more intuitively. The environmental temperature error model (such as temperature sensitivity) allows the system to adapt to different working environments, and the fused error E 融合 (t) provides a fast means of estimating error, providing a basic reference for subsequent complex models, adding physical error to the feature vector, and enhancing the prediction ability of subsequent LSTM models.

[0094] Step 3: The LSTM computation module retrieves the feature vector X(t) and the fusion error E. 融合 (t), using an LSTM model combined with physical constraints to predict sensor errors, the error prediction result E is obtained. LSTM (t);

[0095] When performing step 3, the specific steps are as follows:

[0096] Step 3-1: Feature input, specifically, the feature vector X(t) obtained in step 2-4 is retrieved as the feature vector input to the input layer of the LSTM model;

[0097] Step 3-2: Embed physical constraints in the LSTM model, specifically by incorporating the drift error formula E IMU (t)=E IMU (t−1)+α×Δt and temperature error formula E T = γ × (T env (t) − T0) is directly embedded in the input and gating mechanism of the LSTM model;

[0098] Introduce the drift rate α and time interval β into the forget gate;

[0099] LSTM model fitting prediction results E LSTM1 (t);

[0100] Step 3-3: Physical constraint regularization, specifically by introducing a regularization term L. phy Constraining the LSTM loss function:

[0101] L phy = λ phy × ||E L (t) − (E IMU (t) + E T + E Δt )|| 2 ;

[0102] Among them, E L (t) represents the error calculation for the intermediate process, used to constrain the degree to which the LSTM model conforms to physical laws; λ phyIt is the regularization coefficient, which adjusts the strength of physical error constraints;

[0103] Physical constraint regularization is used to encourage prediction results E LSTM1 (t) satisfies physical constraints;

[0104] Steps 3-4: Combine the error loss and physical regularization loss to optimize the LSTM model weights and generate the final prediction result E. LSTM (t), the total loss function L of LSTM is:

[0105] L total = L data + λL phy ;

[0106] L data = ||E true (t) − E LSTM1 (t)|| 2 ;

[0107] Where λ is the weighting factor, L data It is based on data-driven error loss. In this embodiment, L data E represents the model prediction error. LSTM1 (t) and actual error E true The goal is to minimize the squared error between (t) and the LSTM model parameters to optimize the prediction result E. LSTM1 (t) is closer to the true error E true (t);

[0108] E true (t) is the true error, representing the deviation between the IMU sensor error and the true ground value in the actual measurement. When training the LSTM model, E true (t) Pre-input precise location information P true (t) and IMU sensor measurement value P IMU The difference between (t); precise location information P true (t) Obtained through a high-precision reference system, such as a high-precision GNSS system;

[0109] In this embodiment, when optimizing the LSTM model, the time synchronization error is embedded as a feature into the model, and the model automatically optimizes its impact on the error through learning; the physical constraint regularization term L... phy Including E Δt This encourages the model's fitted predictions to conform to the physical laws of time synchronization, thereby reducing the impact of time synchronization errors on the final results.

[0110] In this embodiment, data-driven and physical modeling are combined to make up for the lack of physical interpretability in traditional deep learning models. By regularizing physical constraints, the risk of overfitting during training of the LSTM model is reduced, and the accuracy of error prediction is improved. The LSTM model can learn dynamic patterns in time series data and can adapt to various navigation scenarios. The total loss function combines physical error and data error to effectively balance the model's predictive ability and physical consistency.

[0111] Step 4: Optimize the module to obtain real-time ambient temperature data T env (t), IMU signal-to-noise ratio (SNR) IMU And error prediction results E LSTM (t), establish a dynamic adaptive mechanism to adjust the noise covariance matrix according to environmental changes, use reinforcement learning optimization to establish a reward function, dynamically adjust LSTM weights and noise covariance matrix, obtain optimized LSTM model weights and dynamically adjusted noise covariance matrix, and improve the real-time performance of error prediction and correction.

[0112] Step 4 includes the following steps:

[0113] Step 4-1: Obtain real-time ambient temperature data T env (t), IMU signal-to-noise ratio (SNR) IMU And the predicted result E LSTM (t);

[0114] Step 4-2: Establish a dynamic adaptive mechanism to adjust the noise covariance matrix according to environmental changes:

[0115] Q adaptive = Q0 + k1 × ∣T env (t) − T0∣;

[0116] R adaptive = R0 + k2 × SNR IMU ;

[0117] Among them, Q adaptive It is the process noise covariance matrix, which is affected by temperature changes; R adaptive R0 is the measurement noise covariance matrix, which is affected by the signal-to-noise ratio; k1 and k2 are both adjustment coefficients; Q0 is the initial value of the process noise covariance matrix; R0 is the initial value of the measurement noise covariance matrix.

[0118] Step 4-3: Establish the reward function Reward using reinforcement learning optimization methods:

[0119] Reward = −||E true (t) − E LSTM (t)|| 2 ;

[0120] The LSTM model and noise covariance matrix are optimized by dynamically adjusting the LSTM weights and Kalman filter covariance matrix through the reward function Reward.

[0121] IMU signal-to-noise ratio (SNR) IMU This is a typical signal-to-noise ratio value for the IMU sensor, provided by the manufacturer.

[0122] In this embodiment, by dynamically adjusting the noise covariance matrix, the system can continuously optimize performance during real-time operation. Reinforcement learning dynamically adjusts the model parameters and filter covariance matrix, enabling the model to adapt to complex changes in the real-time environment. Through a multi-layer optimization mechanism (LSTM optimization + noise adjustment), error accumulation is further reduced. The system can automatically adjust noise parameters according to the real-time ambient temperature and IMU signal-to-noise ratio, thereby improving the robustness of the system.

[0123] Step 5: Correct the inertial data using the prediction error, and continuously optimize the model through a feedback mechanism to output high-precision navigation data.

[0124] When performing step 5, the IMU data P is processed according to the following formula. IMU (t) is corrected to obtain the corrected navigation data P. corrected (t):

[0125] P corrected (t) = P IMU (t) − E LSTM (t);

[0126] The corrected data is used to update the parameters of the LSTM model and Kalman filter, and a feedback mechanism is used to continuously optimize the model, outputting navigation data P. corrected (t).

[0127] The following is an application scenario in this embodiment:

[0128] A drone is flying through complex valley terrain. Due to terrain obstruction, GNSS signals are intermittently lost, and the system needs to rely on inertial navigation equipment for high-precision positioning.

[0129] Input data:

[0130] IMU data: Acceleration a(t) = [0.1, 0.2, −9.8] (m / s²), angular velocity ω(t) = [0.01, −0.02, 0.03].

[0131] GNSS positioning data:

[0132] P GNSS(t)=[39.123456,−76.543210,100] (latitude and longitude and altitude, in degrees and meters).

[0133] Ambient temperature:

[0134] T env (t) = 35℃.

[0135] Preprocessing:

[0136] Time interval error: E Δt =0.005×Δt,Δt=0.1s.

[0137] Fusion error calculation:

[0138] E IMU (t) = 0.01 + 0.005 × 0.1 = 0.0105.

[0139] E T =0.002×(35−25)=0.02.

[0140] E 融合 (t)=0.0105+0.02+0.0005=0.031.

[0141] Error correction:

[0142] Correcting inertial data:

[0143] P corrected (t)=P IMU (t)−E LSTM (t), E LSTM (t)=0.03.

[0144] Output: P corrected (t)=[39.123455,−76.543209,99.97].

[0145] This invention provides a method for error prediction in inertial navigation devices. It addresses the technical challenge of improving the real-time performance and accuracy of navigation systems by directly embedding the IMU physical drift formula and environmental errors into the LSTM model, combined with dynamic adaptive mechanisms and reinforcement learning optimization. This invention integrates the IMU drift formula with environmental temperature error modeling, enhancing prediction accuracy and providing the model with physical interpretability. It adjusts the noise covariance matrix based on environmental temperature changes and the IMU signal-to-noise ratio, improving the system's adaptability to dynamic environments. Using real-time errors as a reward signal, it dynamically optimizes the LSTM weights and covariance matrix, enhancing prediction capabilities over long periods. Feedback optimization of the corrected navigation data enables model self-learning and performance iteration.

Claims

1. A method for error prediction in an inertial navigation device, characterized in that: Includes the following steps: Step 1: The data acquisition module acquires IMU data P from the IMU sensor, GNSS receiver, and environmental sensor respectively. IMU (t), GNSS positioning data P GNSS (t) and ambient temperature data T env (t), and preprocessed; IMU data P IMU (t) includes acceleration a(t) and angular velocity ω(t); Preprocessing includes processing IMU data P IMU (t) performs noise filtering on the ambient temperature data T env (t) is smoothed and the IMU data P is aligned using timestamps. IMU (t) and GNSS positioning data P GNSS (t), and interpolate to extend GNSS positioning data P GNSS (t) to IMU data P IMU The time frequency of (t); the time interval error E Δt Perform mathematical modeling; The preprocessed data and time interval error E Δt By combining these features, we obtain the feature vector X(t). Step 2: The data fusion module processes the IMU data. IMU The drift error E of (t) IMU (t) and the temperature error E of the ambient temperature data T Mathematical modeling is performed separately, based on the drift error E IMU (t), Temperature error E T and time interval error E Δt This generates an initial fusion error E. 融合 (t); the drift error E IMU (t) and temperature error E T Add it to the feature vector X(t); Step 3: The LSTM computation module retrieves the feature vector X(t) and the fusion error E. 融合 (t), using an LSTM model combined with physical constraints to predict sensor errors, the error prediction result E is obtained. LSTM (t); Step 4: Optimize the module to obtain real-time ambient temperature data T env (t), IMU signal-to-noise ratio (SNR) IMU And error prediction results E LSTM (t), establish a dynamic adaptive mechanism to adjust the noise covariance matrix according to environmental changes, use reinforcement learning optimization to establish a reward function, dynamically adjust LSTM weights and noise covariance matrix, obtain optimized LSTM model weights and dynamically adjusted noise covariance matrix, and improve the real-time performance of error prediction and correction. Step 5: Correct the inertial data using the prediction error, and continuously optimize the model through a feedback mechanism to output high-precision navigation data.

2. The method for error prediction of an inertial navigation device as described in claim 1, characterized in that: When performing step 1, the eigenvector X(t) is represented by the following formula: X(t)=[a(t),ω(t),P GNSS (t),T env (t),E Δt ]; Time interval error E Δt It is calculated using the following formula: E Δt = β × Δt; Where Δt is the time interval, representing the IMU data P IMU (t) and GNSS positioning data P GNSS (t) is the time synchronization error at the same moment; β is the time error influence coefficient, representing the influence of time delay between sensors.

3. The method for error prediction of an inertial navigation device as described in claim 2, characterized in that: When performing step 2, the specific steps are as follows: Step 2-1: Process IMU data P IMU The drift error E of (t) IMU (t) modeling, as shown in the following formula: E IMU (t)=E IMU (t−1)+α×Δt; Among them, E IMU (t) represents the IMU error at the current time t, E IMU (t−1) represents the IMU error at the previous time point t-1, α represents the drift rate, and Δt represents the time interval; Step 2-2: Temperature error E of ambient temperature data T The modeling is illustrated in the following formula: E T = γ × (T env (t)− T0); Among them, T env (t) represents the ambient temperature, T0 represents the calibration temperature of the ambient sensor, and γ represents the temperature sensitivity coefficient; the ambient sensor is a temperature sensor. Steps 2-3: Calculate the drift error E IMU (t), Temperature error E T and time interval error E Δt The fusion is performed to obtain the fusion error E. 融合 (t): AND 融合 (t) = E IMU (t) + E T + E Δt ; Steps 2-4: Calculate the drift error E IMU (t) and temperature error E T After being added to the eigenvector X(t), the eigenvector X(t) is represented as follows: X(t)=[a(t),ω(t),P GNSS (t),T env (t),E IMU (t),E T ,E Δt ]。 4. The method for error prediction of an inertial navigation device as described in claim 3, characterized in that: When performing step 3, the specific steps are as follows: Step 3-1: Feature input, specifically, the feature vector X(t) obtained in step 2-4 is retrieved as the feature vector input to the input layer of the LSTM model; Step 3-2: Embed physical constraints in the LSTM model, specifically by incorporating the drift error formula E IMU (t)=E IMU (t−1)+α×Δt and temperature error formula E T = γ × (T env (t) − T0) is directly embedded in the input and gating mechanism of the LSTM model; Introduce the drift rate α and time interval β into the forget gate; LSTM model fitting prediction results E LSTM1 (t); Step 3-3: Physical constraint regularization, specifically by introducing a regularization term L. phy Constraining the LSTM loss function: L phy = λ phy × ||E L (t) − (E IMU (t) + E T + E Δt )|| 2 4 Among them, E L (t) represents the error calculation for the intermediate process, used to constrain the degree to which the LSTM model conforms to physical laws; λ phy It is the regularization coefficient, which adjusts the strength of physical error constraints; Physical constraint regularization is used to encourage prediction results E LSTM1 (t) satisfies physical constraints; Steps 3-4: Combine the error loss and physical regularization loss to optimize the LSTM model weights and generate the final prediction result E. LSTM (t), the total loss function L of LSTM is: THE total = L data + λL phy ; L data = ||E true (t) − E LSTM1 (t)|| 2 ; Where λ is the weighting factor, L data E is based on data error loss. LSTM1 (t) represents the initial prediction result obtained by the LSTM model through forward propagation after embedding physical constraints, E LSTM (t) represents the loss L due to combined data error. data With physical constraint loss L phy The final prediction result obtained after optimization.

5. The method for error prediction of an inertial navigation device as described in claim 4, characterized in that: Step 4 includes the following steps: Step 4-1: Obtain real-time ambient temperature data T env (t), IMU signal-to-noise ratio (SNR) IMU And the predicted result E LSTM (t); Step 4-2: Establish a dynamic adaptive mechanism to adjust the noise covariance matrix according to environmental changes: Q adaptive = Q0+ k1 × ∣T env (t) − T0∣; R adaptive = R0 + k2 × SNR IMU ; Among them, Q adaptive It is the process noise covariance matrix, which is affected by temperature changes; R adaptive R0 is the measurement noise covariance matrix, which is affected by the signal-to-noise ratio; k1 and k2 are both adjustment coefficients; Q0 is the initial value of the process noise covariance matrix; R0 is the initial value of the measurement noise covariance matrix. Step 4-3: Establish the reward function Reward using reinforcement learning optimization methods: Reward = −||E true (t) − E LSTM (t)|| 2 ; Among them, E true (t) is the true error; The LSTM model and noise covariance matrix are optimized by dynamically adjusting the LSTM weights and Kalman filter covariance matrix through the reward function Reward.

6. The method for error prediction of an inertial navigation device as described in claim 4, characterized in that: When performing step 5, the IMU data P is processed according to the following formula. IMU (t) is corrected to obtain the corrected navigation data P. corrected (t): P corrected (t) = P IMU (t) − E LSTM (t); The corrected data is used to update the parameters of the LSTM model and Kalman filter, and a feedback mechanism is used to continuously optimize the model, outputting navigation data P. corrected (t).

Images