An underwater inertial navigation method and system based on an expansion time sequence convolution enhanced fusion of a transformer and a Kalman filter

Abstract

The application discloses an underwater inertial navigation method and system based on inflation time series convolution enhanced fusion of a Transformer and Kalman filtering, an electronic device and a storage medium. The method constructs a two-stage online correction network, extracts local short-time IMU error features using inflation time series convolution, models long-time dependence by combining a Transformer encoder, and corrects the original angular velocity. The original acceleration and the corrected angular velocity are jointly input for the same network processing to obtain corrected acceleration. The corrected angular velocity and acceleration are used to perform inertial navigation mechanical arrangement, recursively obtain predicted attitude, velocity and position, and combine Kalman filtering to update the state when the DVL velocity measurement is available, and only rely on the corrected IMU to maintain navigation when the DVL fails. The application effectively solves the problem that the traditional method is difficult to adapt to dynamic errors, and reduces the navigation drift in a complex marine environment.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and autonomous navigation technology for underwater vehicles, specifically to an underwater inertial navigation method, system, electronic device, and storage medium based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter. Background Technology

[0002] In missions such as seabed mapping, pipeline inspection, marine environmental monitoring, target detection, and autonomous operation, underwater vehicles need to continuously acquire their attitude, velocity, and position in environments without or where satellite navigation signals are unavailable. Inertial Measurement Units (IMUs), with their advantages of high output frequency, strong autonomy, and independence from external signals, are the core sensors in underwater vehicle navigation systems.

[0003] However, the three-axis angular velocities and three-axis accelerations output by the IMU are typically affected by factors such as bias, scale factor error, non-orthogonality error, random noise, temperature drift, and vehicle vibration. These errors accumulate over time during the attitude, velocity, and displacement integration processes of inertial navigation, causing velocity drift, position divergence, and track deviation. Especially in underwater environments, external auxiliary information is affected by underwater acoustic propagation conditions, seabed topography, navigation altitude, and maneuvering status, making it impossible to continuously and stably provide high-precision positioning corrections.

[0004] Existing technologies mainly include the following categories.

[0005] The first category is the traditional offline calibration method. This type of method estimates the IMU's zero bias, scale factor, and installation error through static multi-position tests or turntable tests. Its drawback is that the offline calibration parameters are usually fixed and difficult to adapt to dynamic error changes, temperature changes, and carrier vibrations during the actual navigation of the underwater vehicle.

[0006] The second category is traditional integrated navigation filtering methods. A typical method is INS / DVL Kalman filter fusion, which uses the IMU for inertial navigation prediction and the DVL velocity measurement for filter updates. This type of method can effectively suppress inertial navigation drift when the DVL is working normally, but when the DVL malfunctions, loses lock, or the measurement quality deteriorates, the filter will revert to relying on the original IMU for recursion, and the error will still increase rapidly.

[0007] The third category comprises deep learning-based IMU calibration or inertial odometry methods. Some methods employ CNNs or dilated convolutional networks to compensate for errors in the IMU sequence, capturing local temporal features, but they are insufficient for modeling slowly varying zero bias and complex maneuver dependencies over long time periods. Other methods use Transformers to model the IMU sequence, handling long-term dependencies, but if high-frequency IMU sequences are directly input, they are easily affected by local high-frequency noise and lack close integration with inertial navigation physical recursion and integrated navigation filtering.

[0008] The fourth category consists of methods that directly predict velocity, displacement, or trajectory end-to-end. While these methods can bypass some integration errors, the physical interpretability of the output is weak, making it difficult to integrate with traditional inertial navigation mechanical orchestration and Kalman filtering systems. They also struggle to fully utilize the measurement advantages of commonly used underwater navigation sensors such as DVL. Summary of the Invention

[0009] The main objective of this application is to provide an underwater inertial navigation method, system, electronic device, and storage medium based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter. This application aims to solve the following technical problems: 1. The original IMU error accumulates rapidly during integration, leading to underwater inertial navigation position drift; 2. Traditional offline calibration struggles to handle dynamic time-varying IMU errors; 3. Simple dilated convolutional networks are insufficient for modeling long-term dependencies; 4. Simple Transformer networks are insufficient for characterizing local short-term errors and high-frequency noise; 5. Real acceleration labels are difficult to obtain directly in actual underwater experiments; 6. Directly differentiating DVL velocity into acceleration amplifies measurement noise.

[0010] To achieve the above objectives, the first aspect of this application provides an underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter, comprising: The raw angular velocity and raw acceleration data output by the inertial measurement unit (IMU) mounted on the underwater vehicle are acquired; a sliding time window of length L is constructed, and the raw IMU sequence within the sliding time window is input into a two-stage online calibration network for error compensation processing to obtain the corrected angular velocity and corrected acceleration. The error compensation processing of the two-stage online correction network includes: Dilated temporal convolution is performed on the original angular velocity and original acceleration within the sliding time window to extract local short-time IMU error features. After adding time position encoding, it is input into the Transformer encoder to model long-term dependencies. The gyroscope error correction term at the current moment is obtained through the output projection layer. The original angular velocity is compensated by the gyroscope error correction term to obtain the corrected angular velocity. The original acceleration and the corrected angular velocity within the sliding time window are used as joint inputs. Then, dilated temporal convolution, time position encoding, Transformer encoder modeling long-term dependencies, and output projection layer processing are performed sequentially to obtain the accelerometer error correction term at the current moment. The original acceleration is then compensated using the accelerometer error correction term to obtain the corrected acceleration. The inertial navigation mechanical arrangement is performed using the corrected angular velocity and corrected acceleration, and the predicted attitude, predicted velocity and predicted position are obtained recursively. The predicted attitude, predicted velocity, and predicted position are used as state vectors for covariance prediction. When the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement. When the DVL fails or is abnormal, the DVL update step is skipped, and navigation recursion is maintained solely by calibrating the IMU.

[0011] Optionally, the dilated temporal convolution employs a multi-layer structure, and the features of the l-th layer are calculated as follows: Where is the kernel size, is the dilation rate, is the batch normalization, and is the GELU activation function; The dilation rate of the dilated convolution is set to an increasing sequence.

[0012] Optionally, the Transformer encoder includes a self-attention calculation module and a multi-head attention calculation module; The features encoded by the Transformer are then aggregated through global averaging or by extracting the features corresponding to the end of the window to obtain the feature vector at the current time.

[0013] Optionally, the method further includes: During the training phase, the corrected angular velocity is recursively calculated using SO(3) attitude to obtain the estimated attitude, and the attitude consistency loss between the estimated attitude and the true attitude is calculated. Transform the DVL velocity into the navigation coordinate system, integrate the corrected acceleration between adjacent DVL sampling times after attitude rotation to obtain the estimated velocity increment, and calculate the velocity increment consistency loss between the estimated velocity increment and the DVL velocity increment. During the online inference phase, the two-stage online correction network only inputs the raw IMU sequence and does not input DVL data.

[0014] Optionally, the method further includes: The total training loss function is determined by weighted summation of attitude consistency loss, velocity increment consistency loss, and smoothing loss to suppress high-frequency jitter in the correction acceleration.

[0015] Optionally, when using a calibrated IMU for state prediction, the calibrated angular velocity and calibrated acceleration are substituted into the inertial navigation differential equation as control quantities; when DVL velocity measurements are available, velocity observation residuals are constructed, Kalman gain is calculated based on the velocity observation residuals, and the state vector and covariance matrix are updated based on the Kalman gain.

[0016] To achieve the above objectives, a second aspect of this application also provides an underwater inertial navigation system based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter, comprising: The IMU data acquisition module is used to acquire the raw angular velocity and raw acceleration data output by the IMU mounted on the underwater vehicle; The sliding window building module is used to build a sliding time window of length L; A two-level online calibration network module is used to perform error compensation processing on the original IMU sequence within the sliding time window to obtain the calibrated angular velocity and calibrated acceleration. This module includes an angular velocity calibration branch and an acceleration calibration branch. The angular velocity calibration branch extracts local short-time IMU error features through dilated temporal convolution and combines it with a Transformer encoder to model long-term dependencies, outputting a gyroscope error correction term to obtain the calibrated angular velocity. The acceleration calibration branch takes the original acceleration and calibrated angular velocity as joint inputs and performs the same network processing to output an accelerometer error correction term and obtain the calibrated acceleration. The inertial navigation mechanical orchestration module is used to recursively obtain the predicted attitude, predicted velocity, and predicted position using the corrected angular velocity and corrected acceleration. The Kalman filter fusion module is used to perform covariance prediction using the predicted attitude, predicted velocity, and predicted position as state vectors. When the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement. When the DVL fails or is abnormal, the DVL update step is skipped, and navigation recursion is maintained solely by the calibration IMU.

[0017] To achieve the above objectives, a third aspect of this application also provides an electronic device, including a processor and a memory, the memory storing a computer program, and the processor being configured to execute the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as provided above.

[0018] To achieve the above objectives, a fourth aspect of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as provided above.

[0019] To achieve the above objectives, the fifth aspect of this application also provides an autonomous navigation system for underwater vehicles, comprising: The underwater vehicle body; IMU and DVL sensors are installed on the body of the underwater vehicle; And as mentioned above, an underwater inertial navigation system based on dilated temporal convolution enhanced Transformer and Kalman filter fusion, wherein the two-level online correction network module is deployed in the edge computing terminal or cloud server of the underwater vehicle body, and is used to output correction IMU information in real time to drive the inertial navigation mechanical orchestration module and Kalman filter fusion module to perform state estimation.

[0020] This application discloses an underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filtering. The method includes acquiring raw angular velocity and raw acceleration data output by the inertial measurement unit (IMU) mounted on an underwater vehicle; constructing a sliding time window of length L; inputting the raw IMU sequence within the sliding time window into a two-stage online calibration network for error compensation processing to obtain corrected angular velocity and corrected acceleration; wherein the error compensation processing of the two-stage online calibration network includes: performing dilated temporal convolution on the raw angular velocity and raw acceleration within the sliding time window to extract local short-time IMU error features; adding time-position encoding and inputting it into a Transformer encoder to model long-term dependencies; obtaining the gyroscope error correction term at the current moment through the output projection layer; and using the gyroscope error correction term to compensate the raw angular velocity to obtain the corrected angular velocity. The original acceleration and the corrected angular velocity within the sliding time window are used as joint inputs. These are then subjected to dilated temporal convolution, time-position encoding, Transformer encoder modeling of long-term dependencies, and output projection layer processing to obtain the accelerometer error correction term for the current moment. This accelerometer error correction term is used to compensate for the original acceleration to obtain the corrected acceleration. The corrected angular velocity and corrected acceleration are used to perform inertial navigation mechanical orchestration, recursively obtaining the predicted attitude, predicted velocity, and predicted position. The predicted attitude, predicted velocity, and predicted position are used as state vectors for covariance prediction. When the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement. When the DVL fails or malfunctions, the DVL update step is skipped, and navigation recursion is maintained solely by the calibrated IMU.

[0021] This application effectively overcomes the technical shortcomings of traditional offline calibration, such as difficulty in adapting to dynamic time-varying errors, insufficient modeling of long-term dependencies by simple CNNs, and susceptibility of simple Transformers to high-frequency noise, by extracting local short-term IMU error features through dilated temporal convolution and combining it with Transformer modeling of long-term dependencies. It employs a cascaded structure of a two-stage online calibration network that first corrects angular velocity and then assists in acceleration correction, reducing the impact of attitude errors on velocity and position recursion. During the training phase, the angular velocity branch is supervised by real attitude data, and the acceleration branch is indirectly supervised by DVL velocity increments, allowing network training to be completed without real acceleration labels. In the online inference phase, only the original IMU sequence is input to output the calibration results. Fusion estimation is achieved by combining inertial navigation mechanical orchestration and Kalman filtering. Filtering updates are performed when DVL is available, and stable navigation can still be maintained by relying on the calibrated IMU even when DVL is abnormal or fails. Compared with existing technologies, this application can effectively suppress the accumulation of inertial navigation errors, improve the accuracy of underwater track estimation and system robustness, and reduce navigation drift in complex marine environments. Attached Figure Description

[0022] Figure 1 A schematic diagram of a two-stage IMU online correction network for a dilated temporal convolutional enhancement Transformer provided in this application embodiment; Figure 2 A diagram illustrating the architecture of an underwater inertial navigation and Kalman filter fusion system based on a dual-level IMU online calibration network, provided in this application embodiment; Figure 3 A comparison chart of errors of different underwater inertial navigation methods provided in the embodiments of this application; Figure 4 Visual comparison chart of underwater vehicle trajectory estimation provided in the embodiments of this application; Figure 5 A flowchart illustrating the method provided in this application embodiment.

[0023] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0024] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0025] Current underwater inertial navigation technologies primarily rely on traditional offline calibration and INS / DVL Kalman filter fusion. Offline calibration struggles to adapt to dynamic, time-varying errors underwater; traditional integrated navigation systems revert to the original IMU recursion when DVL anomalies or lock-out occur, leading to rapid error growth; among deep learning-based correction methods, CNNs are insufficient for long-term dependency modeling, and Transformers are susceptible to local high-frequency noise and lack close integration with inertial navigation physical recursion. Therefore, existing technologies struggle to simultaneously address local short-term feature extraction, long-term dynamic dependency modeling, and navigation robustness under DVL failure conditions.

[0026] refer to Figures 1-5 The first embodiment of this application provides a two-stage IMU online calibration method based on dilated temporal convolution enhanced Transformer to solve the technical problems mentioned in the background art, such as rapid accumulation of original IMU error, difficulty of traditional offline calibration to adapt to dynamic changes, and modeling defects of simple CNN or Transformer networks. This method can be executed by a processor, which can be set in a terminal or server. The execution process of the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter can be as follows: Step S101: Obtain the raw angular velocity and raw acceleration data output by the inertial measurement unit (IMU) mounted on the underwater vehicle; construct a sliding time window of length L, and input the raw IMU sequence within the sliding time window into a two-stage online calibration network for error compensation processing to obtain the corrected angular velocity and corrected acceleration.

[0027] In one embodiment of this application, let the first The original angular velocity at each sampling time is The original acceleration is .

[0028]

[0029]

[0030] A single-frame IMU vector is represented as: .

[0031] Build length is A sliding window.

[0032]

[0033] In one-dimensional convolution calculations, it is represented in channel-priority form.

[0034] .

[0035] Step S102: Dilated temporal convolution is performed on the original angular velocity and original acceleration within the sliding time window to extract local short-time IMU error features. After adding time position encoding, the input is fed into the Transformer encoder to model long-term dependencies. The gyroscope error correction term at the current moment is obtained through the output projection layer. The gyroscope error correction term is used to compensate the original angular velocity to obtain the corrected angular velocity.

[0036] Furthermore, the known error model of the IMU is:

[0037] in It is the original measurement value, It is a real physical value. It is the scaling factor error, Non-orthogonal error It is zero bias. It is noise, and can be transformed into:

[0038]

[0039] in For the correction matrix, Since it is an error correction term, it can be transformed into:

[0040] 1. Angular velocity correction Angular velocity correction branch is denoted as .

[0041]

[0042] The corrected angular velocity is:

[0043] This branch first extracts local error features through multi-layer dilated temporal convolution. The features of the l-th layer are calculated as follows:

[0044] in, The kernel size is [size]. For expansion rate, For batch normalization, This is the GELU activation function.

[0045] The single-point form of dilated convolution is:

[0046] Preferably, the expansion ratio is set as follows:

[0047] To preserve the temporal order, temporal position encoding is added after the dilated convolution output.

[0048]

[0049] in, This refers to the feature sequence after dilated convolution and temporal location encoding. Encodes the time location.

[0050] The self-attention of the Transformer encoder is calculated as follows:

[0051] Where Q, K, and V refer to the query, key, and value matrices obtained from the input sequence after linear transformation, respectively, and d is the dimension of the key vector, used to scale the dot product.

[0052] Multi-head attention is calculated as follows:

[0053] MHA refers to multi-head attention, which involves concatenating the results from multiple attention heads and then projecting them onto the output. It involves concatenating the outputs of all heads along the feature dimensions. This refers to the i-th attention head, where each head independently performs the above Attention calculation, focusing on different feature subspaces. This refers to the output projection matrix, which maps the stitched high-dimensional features back to the original dimensions, enabling the fusion of multi-head information.

[0054] The features encoded by Transformer are:

[0055] in, It is the feature sequence output by the encoder, which retains the context information of each time step within the window.

[0056] The feature vector at the current time is obtained by taking the feature corresponding to the end of the window or by globally averaging and pooling the window features.

[0057]

[0058] in, Represents the final extracted global feature vector at the current moment (fixed dimension), representing the network's comprehensive estimate of the current IMU error. This refers to the read operation, with options for "retrieving features from the last moment of the window" or "performing global average pooling on features from all moments within the window".

[0059] Error correction terms are output through a linear projection layer.

[0060] .

[0061] in, , The weight matrix and bias of the linear projection layer are trainable parameters. This is the error correction term for the network prediction at the current time.

[0062] Step S103: The original acceleration and corrected angular velocity within the sliding time window are used as joint inputs, and dilated temporal convolution, time position encoding, Transformer encoder modeling long-term dependencies, and output projection layer processing are performed sequentially to obtain the accelerometer error correction term at the current moment. The accelerometer error correction term is used to compensate the original acceleration to obtain the corrected acceleration.

[0063] In one embodiment of this application, the acceleration correction branch is denoted as... Its input consists of the original acceleration sequence and the corrected angular velocity sequence.

[0064]

[0065] The corrected acceleration is:

[0066] This branch also uses dilated temporal convolution to enhance the Transformer structure.

[0067]

[0068] After adding time and location encoding, we get:

[0069] The features encoded by Transformer are:

[0070] Read the features at the current moment or perform global average aggregation.

[0071]

[0072] The acceleration error correction term is output through the linear projection layer.

[0073]

[0074] Ultimately, this embodiment captures local IMU errors through dilated temporal convolution, models dynamic dependencies within a longer time window using a Transformer, and enhances the stability of acceleration correction through a cascaded structure prioritizing angular velocity correction. This core module can serve as the online correction front-end for the IMU in a complete underwater inertial navigation system.

[0075] Step S104: Perform inertial navigation mechanical arrangement using the corrected angular velocity and corrected acceleration to recursively obtain the predicted attitude, predicted velocity and predicted position.

[0076] Furthermore, during the online phase, the corrected angular velocity output by the network is used for attitude recursion, and the corrected acceleration is used for velocity and position recursion.

[0077]

[0078] The acceleration of the navigation system is:

[0079] The velocity recursion is as follows:

[0080] The position is recursively determined as follows: .

[0081] Step S105: Covariance prediction is performed using the predicted attitude, predicted velocity, and predicted position as state vectors; when the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement.

[0082] Furthermore, the system state vector is defined as:

[0083] in, It is the attitude angle, It is an acceleration bias. It is the gyroscope bias.

[0084] State prediction is performed using a calibrated IMU.

[0085]

[0086] Covariance prediction is:

[0087] in, Let be the state covariance matrix, representing the uncertainty of the current state estimate. This is the state transition matrix, used to propagate the uncertainty from the previous time step to the current time step. The process noise covariance matrix represents the increment of uncertainty caused by the noise of the IMU (accelerometer and gyroscope) and random walk.

[0088] When DVL velocity measurements are available, construct velocity observations.

[0089]

[0090] The observation residuals are:

[0091] in, The residual is equal to the measured speed in DVL minus the estimated speed predicted based on the current state.

[0092] The Kalman gain is:

[0093] in, The observation matrix (the Jacobian matrix that maps the 15-dimensional state to the 3-dimensional velocity measurement space). The observation noise covariance matrix represents the magnitude of noise in the DVL velocity measurement. The Kalman gain (a 15×3 matrix) determines the proportion by which the residuals are corrected (the greater the gain, the more reliable the DVL measurement).

[0094] Status updated to:

[0095] Covariance updated to:

[0096] Skip DVL updates when DVL fails or malfunctions.

[0097]

[0098]

[0099] Ultimately, in this embodiment, the neural network front-end is responsible for correcting the IMU data, and the Kalman filter back-end is responsible for fusing the DVL velocity measurements. The DVL is used for velocity increment supervision during the training phase and for filter updates during the online phase, but it is not used as input to the neural network, thus avoiding the problem of reusing real-time DVL information in the network and filter.

[0100] Step S106: When the DVL fails or malfunctions, skip the DVL update step and rely solely on the calibration IMU to maintain navigation recursion.

[0101] Furthermore, DVL updates are skipped when DVL fails or malfunctions. Ultimately, in this embodiment, the neural network front-end is responsible for correcting the IMU data, and the Kalman filter back-end is responsible for fusing the DVL velocity measurements. DVL is used for velocity increment supervision during the training phase and for filter updates during the online phase, but it is not used as neural network input, thus avoiding the problem of reusing real-time DVL information in the network and filter.

[0102] The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter also includes: During the training phase, the corrected angular velocity is used to recursively calculate the attitude to obtain the estimated attitude, and the attitude consistency loss between the estimated attitude and the true attitude is calculated. Transform the DVL velocity into the navigation coordinate system, integrate the corrected acceleration between adjacent DVL sampling times after attitude rotation to obtain the estimated velocity increment, and calculate the velocity increment consistency loss between the estimated velocity increment and the DVL velocity increment. During the online inference phase, the two-stage online correction network only inputs the raw IMU sequence and does not input DVL data.

[0103] The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter also includes: The total training loss function is determined by weighted summation of attitude consistency loss, velocity increment consistency loss, and smoothing loss to suppress high-frequency jitter in the correction acceleration.

[0104] Specifically, during the training phase, the actual posture is used to constrain the angular velocity correction results, and the DVL velocity increment is used to indirectly constrain the acceleration correction results. During the online inference phase, the neural network only inputs the IMU and outputs the corrected angular velocity and corrected acceleration; the DVL, when available, is used as a Kalman filter velocity measurement and is not used as network input.

[0105] 1. Real pose constraints The attitude is derived recursively from the corrected angular velocity.

[0106]

[0107] in, It is the attitude matrix at the current moment. It is the attitude matrix of the previous moment. This is the antisymmetric matrix of angular velocity.

[0108]

[0109] The attitude consistency loss is:

[0110] 2. DVL (Dynamic Value Chain) velocity increment constraint This invention does not directly differentiate the DVL velocity into an acceleration label, but instead uses the DVL velocity increment to constrain the acceleration correction branch.

[0111] Convert DVL velocity to navigation coordinates.

[0112]

[0113] Transform the corrected acceleration into the navigation coordinate system.

[0114]

[0115] At adjacent DVL times Between these points, the estimated velocity increment is obtained by integrating the corrected acceleration.

[0116]

[0117] DVL speed increment is:

[0118] The consistency loss for speed increments is:

[0119] To suppress high-frequency jitter in the correction acceleration, a smoothing loss is set.

[0120]

[0121] The total training loss is:

[0122] To address the aforementioned technical problems, the second embodiment of this application provides an underwater inertial navigation system based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter, to solve the same technical problems as the method embodiment. The system 1000 may include the following modules: IMU data acquisition module 1001, sliding window construction module 1002, two-level online correction network module 1003, inertial navigation mechanical arrangement module 1004, and Kalman filter fusion module 1005.

[0123] The IMU data acquisition module 1001 is used to acquire the raw angular velocity and raw acceleration data output by the IMU mounted on the underwater vehicle.

[0124] The sliding window building module 1002 is used to build a sliding time window of length L.

[0125] The dual-level online calibration network module 1003 is used to perform error compensation processing on the original IMU sequence within the sliding time window to obtain the calibrated angular velocity and calibrated acceleration. The module includes an angular velocity calibration branch and an acceleration calibration branch. The angular velocity calibration branch extracts local short-time IMU error features through dilated temporal convolution and combines it with a Transformer encoder to model long-term dependencies and outputs a gyroscope error correction term to obtain the calibrated angular velocity. The acceleration calibration branch takes the original acceleration and the calibrated angular velocity as joint inputs and performs the same network processing to output an accelerometer error correction term and obtain the calibrated acceleration.

[0126] The inertial navigation mechanical orchestration module 1004 is used to recursively obtain the predicted attitude, predicted velocity, and predicted position using the corrected angular velocity and corrected acceleration.

[0127] The Kalman filter fusion module 1005 is used to predict covariance using predicted attitude, predicted velocity, and predicted position as state vectors. When the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement. When the DVL fails or is abnormal, the DVL update step is skipped, and navigation recursion is maintained solely by the calibration IMU.

[0128] To address the aforementioned technical problems, a second embodiment of this application also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as provided above.

[0129] To address the aforementioned technical problems, the third embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as provided above.

[0130] To address the aforementioned technical problems, the fourth embodiment of this application also provides an autonomous navigation system for an underwater vehicle, comprising: an underwater vehicle body; an IMU and DVL sensor disposed on the underwater vehicle body; and an underwater inertial navigation system based on the fusion of dilated temporal convolutional enhanced Transformer and Kalman filter as described above, wherein a two-level online correction network module is deployed in the edge computing terminal or cloud server of the underwater vehicle body, for real-time output of correction IMU information to drive the inertial navigation mechanical orchestration module and the Kalman filter fusion module to perform state estimation.

[0131] It is understood that the above-described device embodiments correspond to the method embodiments. Therefore, the device embodiments have all the beneficial effects of the method embodiments, and will not be repeated here.

[0132] refer to Figure 3 and Figure 4 The fifth embodiment of this application provides an experimental verification and analysis scheme for verifying the effectiveness of the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter in the autonomous navigation mission of underwater vehicles.

[0133] In one embodiment of this application, experimental data can be collected by an underwater vehicle in a pool, lake, or sea trial environment. The underwater vehicle is equipped with an IMU, a DVL (Dynamic Vehicle Range), and a reference navigation system. The IMU outputs three-axis angular velocity and three-axis acceleration, the DVL outputs three-axis velocity, and the reference system provides the actual attitude and reference trajectory.

[0134] Further, data preprocessing includes: time synchronization of IMU, DVL, and reference attitude; DVL outlier removal; DVL velocity coordinate system transformation; DVL velocity increment calculation; IMU sliding window construction; and partitioning of training, validation, and test sets. A single training sample can be represented as: where is the IMU time window, is the true attitude, and is the DVL velocity in the navigation frame.

[0135] For example, the deep learning algorithms compared in the evaluation criteria include Transformer correction, CNN correction, etc. The evaluation metrics are: absolute pose error, velocity error RMSE, and ATE (absolute position error) and RTE (30s relative position error).

[0136] Furthermore, combined Figure 3 and Figure 4 Objective data presentation: Figure 3 The results show that the original IMU method has the highest error, indicating that the uncorrected IMU exhibits significant drift during inertial recursion. CNN and dilated convolution methods can reduce local errors, but they are insufficient for modeling long-term dependencies. The Transformer method can improve long-term trends, but its suppression of local short-term errors is not as stable as that of dilated temporal convolution. The original IMU + Kalman filter method can reduce velocity errors when DVL is available, but it is still limited by the prediction quality of the original IMU. The method in this application achieves lower errors across various metrics through the synergistic effect of dilated temporal convolution, Transformer, and Kalman filtering. Figure 4The results show that the original IMU integral track deviates significantly from the true track over time; the deep learning correction method alone can improve the track shape, but local offsets still exist; the traditional Kalman filter method can correct some errors in the DVL usable range, but drift still exists when the DVL fails or the update interval is long. The track of the method in this application is closest to the true track, with a smaller endpoint error, indicating that it can effectively suppress the cumulative error of underwater inertial navigation.

[0137] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. An underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter, characterized in that, include: Acquire raw angular velocity and raw acceleration data output from the inertial measurement unit (IMU) mounted on the underwater vehicle; A sliding time window of length L is constructed, and the original IMU sequence within the sliding time window is input into a two-stage online calibration network for error compensation processing to obtain the calibrated angular velocity and calibrated acceleration. The error compensation processing of the two-stage online correction network includes: Dilated temporal convolution is performed on the original angular velocity and original acceleration within the sliding time window to extract local short-time IMU error features. After adding time position encoding, it is input into the Transformer encoder to model long-term dependencies. The gyroscope error correction term at the current moment is obtained through the output projection layer. The original angular velocity is compensated by the gyroscope error correction term to obtain the corrected angular velocity. The original acceleration and the corrected angular velocity within the sliding time window are used as joint inputs. Then, dilated temporal convolution, time position encoding, Transformer encoder modeling long-term dependencies, and output projection layer processing are performed sequentially to obtain the accelerometer error correction term at the current moment. The original acceleration is then compensated using the accelerometer error correction term to obtain the corrected acceleration. The inertial navigation mechanical arrangement is performed using the corrected angular velocity and corrected acceleration, and the predicted attitude, predicted velocity and predicted position are obtained recursively. The predicted attitude, predicted velocity, and predicted position are used as state vectors for covariance prediction. When the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement. When the DVL fails or is abnormal, the DVL update step is skipped, and navigation recursion is maintained solely by calibrating the IMU.

2. The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in claim 1, characterized in that, The dilated temporal convolution employs a multi-layer structure, the first... l Layer features are calculated as follows: in, The kernel size is [size]. For expansion rate, For batch normalization, Use the GELU activation function; The dilation rate of the dilated convolution is set to an increasing sequence.

3. The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in claim 2, characterized in that, The Transformer encoder includes a self-attention calculation module and a multi-head attention calculation module; The features encoded by the Transformer are then aggregated through global averaging or by extracting the features corresponding to the end of the window to obtain the feature vector at the current time.

4. The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in claim 1, characterized in that, The method further includes: During the training phase, the corrected angular velocity is used to recursively calculate the attitude to obtain the estimated attitude, and the attitude consistency loss between the estimated attitude and the true attitude is calculated. Transform the DVL velocity into the navigation coordinate system, integrate the corrected acceleration between adjacent DVL sampling times after attitude rotation to obtain the estimated velocity increment, and calculate the velocity increment consistency loss between the estimated velocity increment and the DVL velocity increment. During the online inference phase, the two-stage online correction network only inputs the raw IMU sequence and does not input DVL data.

5. The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in claim 4, characterized in that, The method further includes: The total training loss function is determined by weighted summation of attitude consistency loss, velocity increment consistency loss, and smoothing loss to suppress high-frequency jitter in the correction acceleration.

6. The underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in claim 1, characterized in that, When using a calibrated IMU for state prediction, the calibrated angular velocity and calibrated acceleration are substituted into the inertial navigation differential equation as control variables. When DVL velocity measurements are available, velocity observation residuals are constructed, Kalman gain is calculated based on the velocity observation residuals, and the state vector and covariance matrix are updated based on the Kalman gain.

7. An underwater inertial navigation system based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter, characterized in that, include: The IMU data acquisition module is used to acquire the raw angular velocity and raw acceleration data output by the IMU mounted on the underwater vehicle; The sliding window building module is used to build a sliding time window of length L; A two-level online calibration network module is used to perform error compensation processing on the original IMU sequence within the sliding time window to obtain the calibrated angular velocity and calibrated acceleration. This module includes an angular velocity calibration branch and an acceleration calibration branch. The angular velocity calibration branch extracts local short-time IMU error features through dilated temporal convolution and combines it with a Transformer encoder to model long-term dependencies, outputting a gyroscope error correction term to obtain the calibrated angular velocity. The acceleration calibration branch takes the original acceleration and calibrated angular velocity as joint inputs and performs the same network processing to output an accelerometer error correction term and obtain the calibrated acceleration. The inertial navigation mechanical orchestration module is used to recursively obtain the predicted attitude, predicted velocity, and predicted position using the corrected angular velocity and corrected acceleration. The Kalman filter fusion module is used to perform covariance prediction using the predicted attitude, predicted velocity, and predicted position as state vectors. When the Doppler velocity measurement (DVL) is available, the state vector is updated using the DVL velocity measurement. When the DVL fails or is abnormal, the DVL update step is skipped, and navigation recursion is maintained solely by the calibration IMU.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor being configured to execute the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the underwater inertial navigation method based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in any one of claims 1 to 6.

10. An autonomous navigation system for an underwater vehicle, characterized in that, include: The underwater vehicle body; IMU and DVL sensors are installed on the body of the underwater vehicle; And the underwater inertial navigation system based on the fusion of dilated temporal convolution enhanced Transformer and Kalman filter as described in claim 7, wherein the dual-level online correction network module is deployed in the edge computing terminal or cloud server of the underwater vehicle body, and is used to output correction IMU information in real time to drive the inertial navigation mechanical orchestration module and the Kalman filter fusion module to perform state estimation.

Images