Multi-Sensor Fusion for Underwater Vehicle Localization by Augmentation of RBF Neural Network and Error-State Kalman Filter

The Kalman filter variants extended Kalman filter (EKF) and error-state Kalman filter (ESKF) are widely used in underwater multi-sensor fusion applications for localization and navigation. Since these filters are designed by employing first-order Taylor series approximation in the error covariance matrix, they result in a decrease in estimation accuracy under high nonlinearity. In order to address this problem, we proposed a novel multi-sensor fusion algorithm for underwater vehicle localization that improves state estimation by augmentation of the radial basis function (RBF) neural network with ESKF. In the proposed algorithm, the RBF neural network is utilized to compensate the lack of ESKF performance by improving the innovation error term. The weights and centers of the RBF neural network are designed by minimizing the estimation mean square error (MSE) using the steepest descent optimization approach. To test the performance, the proposed RBF-augmented ESKF multi-sensor fusion was compared with the conventional ESKF under three different realistic scenarios using Monte Carlo simulations. We found that our proposed method provides better navigation and localization results despite high nonlinearity, modeling uncertainty, and external disturbances.


Introduction
The ocean floor has billions of dollars of natural resources in the form of precious elements and medicinal herbs. To take advantage of ocean resources, seabed mapping is the ultimate tool that depends on precise sensors and robust navigation fusion algorithms for autonomous underwater vehicles (AUVs) and remotely operated underwater vehicles (ROVs). Navigational accuracy is a key requirement for complex seabed mapping tasks [1]. However, primary sensors, three-axis gyros, and accelerometers have biases and drifts, which vary with time and are affected by noise. In contrast, most commonly used fusion algorithms based on extended Kalman filter (EKF) and its variant error-state Kalman (ESKF) suffer from divergence and degraded mean square error (MSE) performance in the nonlinear underwater condition because of linear approximation [2]. Thus, the higher the nonlinearity present in the system, the greater the error of the EKF state prediction, and it can also induce filter divergence.
Backpropagation multi-layer neural networks (BPNN) and radial basis neural networks (RBFNN) have excellent learning abilities and are well-known for their nonlinear system identification [3][4][5]. In comparison, the RBFNN has much greater accuracy of based on RFBNN, to improve the performance of central difference Kalman filter (CDKF) for attitude-of-the-satellite estimation. They proposed the range of forgetting at 0.2 to 2 as a multiplier to the Kalman gain, but limited description of the selection of the forgetting factor was provided. Recently, the RBFNN-aided Kalman Filter was proposed to improve the state estimation accuracy for spacecraft navigation [37]. Moreover, they did not use multi-sensor fusion of high-rate and low-rate sensors.

Contributions of the Paper
The proposed work fills the gap by proposing a novel multi-sensor fusion architecture based on the strengths of the RBF neural network and error-state Kalman filter for underwater navigation, which has not been proposed to date to the authors' knowledge. We named this algorithm RBF-ESKF. The augmentation of both algorithms improves the navigation of underwater vehicles in GPS-less environments. Our major contribution is the derivation of a multi-sensor fusion algorithm that improves the accuracy of underwater localization by taking advantage of a radial basis function (RBF) neural network that has the capability of nonlinear universal approximation via recursive learning [38]. Moreover, a simple structure of the RBF network can be trained online with less computation cost compared to the backpropagation neural network (BPNN) [39].
The structure of this paper is organized as follows. Section 2 discusses kinematic mathematical modeling with an ellipsoid Earth model. Section 3 discusses external and internal sensors for underwater navigation. In addition, the mathematical models of the sensors are described with their operating noise characteristics. The error dynamic model is presented in Section 4. In Section 5, a multi-sensor fusion algorithm is derived for accurate underwater integrated navigation. Section 6 shows the test results of the proposed multisensor fusion filter in three different conditions. In addition, acoustic communication lost and on-board sensor malfunctioning are tested and analyzed. Moreover, the performance of the RBF-ESKF is compared with ESKF under different scenarios. A comparative analysis is performed, showing that the proposed algorithm has promising results.

Mathematical Modeling
This section discusses the mathematical modeling of the underwater vehicle taken from [40][41][42][43][44]. The first subsection explains the notations used for modeling. The second subsection describes the frame of references used for the mathematical formulation of IMU. The third subsection formulates the kinematics equation of motion for six degrees of freedom. The last subsection provides brief information on the sensors used for underwater navigation.

Frame of References
The importance of frame of reference transformation for underwater navigation arises from the fact that sensors are mounted on the vehicle's body. The origin of the body is defined as the center of body frame (b) External position fixes are in a rotating, Earthcentered, Earth-fixed (ECEF) frame (e). Moreover, Newton's laws are applicable to the Earth-centered inertial (ECI) frame (i). However, the north-east-down (NED) frame or navigation frame (n) is the tangent plane to the Earth's surface at the location of the underwater vehicle and its x-axis points toward the true north of the Earth. Figure 1 shows a frame of references used to develop navigation equations for underwater vehicles, with slight modification from [40].

Mathematical Notation
The mathematical notations used in this work are standard notations used to model underwater vehicle position, velocity, and attitude [40]. The position and velocity are three dimensional (3D) vectors in Euclidean space. Rotations are represented by quaternion. Subscripts and superscripts are used to represent the relationship between the frames. For example, ω e ie shows an angular velocity of frame (e) with respect to (i) represented in (e) frame. The following Table 1 shows a list of symbols used in the development of underwater vehicle navigation equations.
Linear velocity in the NED frame (north, east, and down)

Navigation Equations
The underwater vehicles are most commonly equipped with a strapdown inertial navigation system [41]. In this configuration, measurements are obtained directly in the body frame and sensors experience full rotation during maneuvers. Moreover, this configuration requires an initial condition for position, velocity, and attitude. The kinematics equations for the underwater vehicles used in this work are described in [42][43][44]. The rate of change in positionṗ e and velocity v e of vehicle in (e) frame is related by the following differential equation: The rate of change in velocityv e of the underwater vehicle in (e) frame is dependent on accelerometer output f b , angular velocity ω e ie , and gravity vector g e in ECEF and can be expressed by the following differential equation: where Ω e ie is a skew symmetric matrix of angular velocity ω e ie and R e b is the rotation matrix that transforms a specific force vector from the (b) frame to the (e) frame.
The rate of change of rotation matrixṘ e b represented in frame (e) is dependent on angular velocity Ω b ib of the body with respect to frame (i), and angular velocity ω b ie of Earth with respect to frame (i) is expressed by the following equation: The relationship between the change in latitude of the vehicleφ lat and velocity is represented by the following differential equation: The change in longitude of the vehicleλ long in the form of the east velocity is represented by following mathematical relationship: The change in height of the vehicleḣ d is expressed in the form of the down velocity aṡ The latitude φ lat , longitude λ long , and depth h d are given as The rate of change in velocity of the vehicle in the (n) framev n eb is expressed by the following differential equation: where Ω n en v n is the centripetal acceleration related to the motion of the (n) frame with respect to the (e) frame and 2Ω n ie v n is the Coriolis acceleration.
The local gravity vector g n eb = 0 0 g T depending on the latitude, longitude, radius of curvature of the meridian, and radius of curvature of the prime vertical is given by where g 0 = 9.780318 × 1 + 5.3024 × 10 −3 sin 2 φ − 5.9 × 10 −6 sin 2 2φ [44]. The rate of change in the rotation matrixṘ n b represented in frame (n) is dependent on the angular velocity of the body with respect to frame (i) and on the angular velocity of frame (n) with respect to frame (i): The attitude of the vehicle is represented by quaternion. It has only one constraint compared to the direction cosine matrix (DCM), and this method is also singularity free in comparison to the Euler angle representation, which has a singularity problem [45]. The attitude of the vehicle represented in quaternionq n b is given aṡ where q n b has two parts: η is the scalar part; i is the vector part; and i = 1,2,3. The ⊗ sign represents a quaternion product.
This section developed the motion equations of an underwater vehicle in starpdown configuration. The measurements from accelerometers were obtained in (b) frame as a specific force vector and were not the true acceleration of the vehicle. To obtain the true acceleration or rate of change in velocity of the vehicle in e frame, as shown by Equation (2), the (b) frame specific forces are transformed into the e frame by a rotational matrix and by compensating for the gravitation and rotation effects of the Earth. Equation (1) shows that the rate of change in the position and acceleration has an integral relationship. Since the position from an acoustic fix is obtain in the geoid ECEF frame, for compatibility, Equations (4)-(9) convert the position of the vehicle in latitude, longitude, and depth. Equation (13) shows a quaternion representation of the roll, pitch, and yaw angles of the vehicle. Interested readers can find more details for mathematical modeling of an underwater vehicle in starpdown configuration in [40][41][42][43][44].

Sensors on the Vehicle
The following subsections give a brief overview of the sensors used in the AUV with a mathematical formulation of errors. For detail about the mathematical model of the sensors used for navigation, readers can refer to [40,46,47].

Inertial Measurement Unit
The inertial measurement unit (IMU) provides a three-axis accelerometer and threeaxis gyro outputs [48]. These measurements are affected by variable factors such as temperature, manufacturing process, scale factor, noise, and drift.
The accelerometer measurement output vector f b acc in (b) frame is modeled as where acc is white noise and b acc is acceleration bias, which is modeled as a 1st-order Markov process. The accelerometer bias b acc is expressed by the random walk and the random constant shown asḃ where τ acc is the correlation time given by the manufacturer. Its value depends on the accelerometer sensors used in the IMU. The value ρ acc depends on standard deviation, and it is known as a process driving noise. The accelerometer in IMU does not directly measure the true kinematic acceleration of vehicles due to the presence of Earth's gravitation. For that reason, the measurement from the accelerometer is known as the relative acceleration or specific force f b ib and it is related to the kinematics acceleration of the vehicle as follows: wherep i is acceleration represented in an inertial frame and g b is gravitation sensed by the accelerometer in the body frame. The actual output of the gyro ω b g is influenced by noise and bias that is given by where g is the white noise and b g is the gyro bias, which is modeled as a 1st-order Markov process. The gyro bias b g is represented by a random walk and a random constant given as follows:ḃ where τ g is the gyro correlation time obtained from manufacturer documentation. Its value depends on the quality of the accelerometer sensors used in the IMU. The value ρ g depends on thte standard deviation, and it is called process driving noise. The IMU used for simulation is tactical grade Emcore SDI-1500. It uses high-precision micro-electromechanical systems (MEMS) quartz sensor technology with 1 • /h gyro bias and 1 mg accelerometer bias stability. The IMU offers the best cost to performance ratio compared to other technologies. It consists of three orthogonal accelerometer sensors that provide the measurement of specific forces. Three gyros provide the angular rates of the body with respect to the inertial frame of reference.

Underwater Acoustic Positing System
The underwater acoustic positing system measures the distance and direction of the vehicle from the reference positions. For this work, HiPaP 502 is used for simulation. This system provides a typical range detection accuracy of 0.2 m, with an operating range of 1 to 5000 m. It has an acoustic operating area of 200 • /200 • with the capability of narrow beamforming of 10 • , which improves the signal-to-noise ratio. It can be interfaced by GPS to provide Earth-related coordinates. However, acoustic position estimate is effected by GPS accuracy, system installation, ship attitude, sound velocity profile, ray bending, and measurement noise.
Assuming the system is precisely calibrated, installation and ship attitude have negligible effects. The mathematical model of the system actual outputp is given aŝ where p h is true output position, whereas b h is the time-varying bias modeled as a 1st-order Markov process and depends on the sound velocity profile and ray bending effect. h represents measurement white noise.

Doppler Velocity Log
The Doppler velocity log (DVL) measures the change in acoustic frequency for determining the speed of the vehicle with reference to the seabed. In deep water, when the seabed is not available, the DVL measures the speed with respect to water. The DVL sends a known frequency signal to the seabed and receives the signal that bounces back to the vehicle. The speed of the underwater vehicle is a dependent doppler effect [49].
The work used Nortek DVL-500, which has a range of 0.3 to 200 m and long-term accuracy of ±0.1% ±0.1 cm/s. The DVL can be operated in a 4-beam Janus configuration complete performance testing of DVL; readers may refer to [50].
Considering that the attitude and installation error are negligible, the actual output of DVLv dvl can be modeled for the true velocity measurement vector, noise, and bias given as [51].v where v dvl is the true output and random velocity error, b dvl is modeled as the 1st-order Markov process, and dvl is white noise.

Depth Sensor
Depth and underwater pressure have a direct relationship [52]. As the vehicle goes deep into water, the pressure reading increases linearly. The depth sensor modeled in this work is from Paroscientific, Inc. part number 8CDP700-I, which provides an accuracy of 0.01% and high stability under tough conditions. The depth sensor actual outputĥ d is modeled by adding true depth h d with noise: where d is measurement noises modeled as white noise.

Magnetometer
The magnetometer or compass measures the magnitude and direction of Earth's magnetic field [53]. The magnetometer used in the work is jewel instrument ECS-AC-RS232 e-compass, which has a 3-axis magnetometer and a 2-axis tilt sensor. The tilt sensor is used for the initialization of roll and pitch [40]. It offers an accuracy of ±0.5 • root mean square RMS, a repeatability of ±0.3 • , and a response time of 36 milliseconds. The pitch and roll are ±42 • whereas the dip angle range ±80 • . The major error sources of the magnetometer include the declination angle, which is the difference between true north and sensor north; the hard and soft magnetic distortion due to motors and ferromagnetic materials [54]; and sensor imperfection, misalignment, and noise. However, with proper calibration, most of the errors in measurement can be removed.
The actual output q m of the magnetometer is a combination of noise and true output q m written asq where m is the sensor noise modeled as white noise.

Error Dynamic Model
The position p, velocity v, attitude q, gyro bias b g , accelerometer bias b acc , and acoustic fix bias b h in full state vector x form are given as where s v is the sound velocity model underwater used to calculation underwater acoustic transmission.
The estimated state vectorx is written aŝ The design used in this worked estimate an error-state vector, which offers the main advantages of flexible sampling rate, robustness, and low computation burden [55,56]. The error-state vector δẋ is the difference between the true state and estimated stateẋ of the model and is given as The error-state vector can be represented as where the position, velocity, and attitude error equations are provided as under. For complete derivation of this error model, readers can refer to [42,57].
Assuming that the underwater vehicle travel at low speed and that the depth of operation is much less than the Earth's radius, the rate of change in errors in longitude δλ long ,latitude δφ lat , and depth δḣ d are given by The velocity error δv n eb is given by the following differential equation: If the vehicle operates at low speed underwater, angler velocity ω n ie and δω n ie can be neglected. Gravity error is also neglected because of the small operating area and accurate estimation [44,58].
Thus, the velocity error δv n eb differential equation can be re-written as Assuming that the angular velocity of rotation of Earth with respect to the inertial frame ω ie is accurately known, using the attitude error model in the form of the quaternion, it is given asδ The error of gyro bias δḃ g is given as The error of accelerometer bias δb acc is given as The error of hydro-acoustic system bias δb acc is given as Since the ESKF employed in this work used an error-state or indirect-form-state vector z, it is obtained by subtracting the outputs of the inertial measurement unit (INS) and aiding sensors measurement [59].
The position error δz p between the INS position p I NS and acoustic position system measurement p h is written as The velocity error δz v between the INS velocity v I NS and DVL measurement v d is given by the following equation: The attitude error δz v between the INS attitude q I NS and magnetometer measurement q m is not a vector quantity; it cannot be subtracted as position and velocity. Quaternion multiplication ⊗ is used to find the error term as follows:

RBF-ESKF Mulit-Sensor Fusion
The proposed modifications improve ESKF performance and make use of the advantages of the RBF neural network. The RBF neural network can approximate any nonlinear function, and they are also known as universal function approximators [60]. The RBF center, its width, and the linear weights for each output neuron are altered at every iteration of a learning algorithm. When each RBF center is as close to the input vector as possible and the network output error is within the target limit, the training phase is completed. Therefore, it is possible to express the approximation of any functional dependency between variables as a linear combination of a best possible number of RBF neurons with appropriate weight and center. The top level block diagram of our proposed fusion algorithm is depicted in Figure 2. As shown by Figure 2, the algorithm takes the error of the aiding sensors and INS as input. The RBF-ESKF fusion algorithm after processing gives an output to INS for error correction and reset. The RBF neural network used for processing has a basic three layer structure; input, hidden, and output layer. Furthermore, compared to BPNN, the RBF variants have less computational load and fast online learning. The first layer is the input layer, which provides an interface between the data and neural network. The data from the input layer to the second hidden layer is transferred in such a manner that the output value of every hidden neuron is inversely related to the Euclidean distance from that neuron's input vector to the RBF neuron's center. The third layer is the output layer, which takes into account the cumulative weights and biases of all RBF neuron outputs.
Several variants of RBF exist in literature that depend on the application [61]. However, in this work, we used the Gaussian-type RBF function. The weights of neurons in the hidden layer for the Gaussian RBF function indicate the center of the symmetrical Gaussian distribution curve. The novelty of the RBF-ESKF algorithm is that it includes system information within the weight and center learning update rules.

RBF-ESKF Mathematical Formulation
The underwater vehicle navigation system has nonlinear dynamics and measurement characteristics, which can be represented in state space form by Equation (37) and Equation (38) where the state of the system is represented byẋ(t). The u(t) is a known system input and α(t) is Gaussian white noise. The measurement output z(t) in state space can be written as where f and h are nonlinear functions. The measurement is corrupted by Gaussian white noise β(t).
Dropping time and noise in the above equations for simplification, the linear form of error state can represented by Equation (39) and Equation (40): [40] δẋ = F(x, u, t)δx (39) where F(x, u, t) = ∂ f ∂x x=x However because the deterministic component F(x, u, t), is always incomplete, which means the model does not incorporate, for instance, AUV underwater movements due to waves, the stochastic component α(t) takes these effects into account.
The linear residual measurement output δz can represented by where H(x, t) = ∂h ∂x x=x ,x is defined as the trajectory obtained by using vehicle kinematics equations.
The Kalman filter [62], in its basic form, is based on linear system and measurement models, which in reality might not be the case. In our underwater vehicle, navigation equations are used to construct the mathematical model and are not linear with respect to the state variables. This is done by linearization of any predicted trajectory, leading to an error-state model as discussed in Section 4. This linear approximation was proven to be incomplete and requires special consideration in underwater navigation, which leads to derivation of the RBF-ESKF algorithm.
For mathematical formulation of RBF-ESKF, we start with ESKF implementation, which utilizes definitions of measurement error and state dynamics error, as described by the Equations (39) and (40) [63]. The ESKF algorithm has three major components. The first component is initialization, in which state, covariance, process, and measurement covariance are initialized [64].
Q 0 = Initialization process noise covariance. R 0 = Initialization of measurement noise covariance. where superscript minus − denotes the a priori state that occurs before innovation is updated. Superscript plus + denotes the posteriori state after innovation calculation.
The second component is time update, in which the error state and error-state covariance are updated, give by where δx − k+1 is the predicted error state and Φ k is the state transition matrix in discrete form.
where P − k+1 is the predicted error covariance and Q k is the process noise covariance v k . The third component is measurement update, in which the residual of measurement is updated. The residual of measurement δz k is given by the difference between the actual measurement z k and the prediction of measurement h(x k ) The kalman K k gain is given as The posteriori error estimate δx + k is given as The expression δz k − H k δx − k is referred to as innovation. It is the difference between the error of observation and its expected error,represented by s k as The posteriori error-state covariance P + k is given as The complete corrected navigation statex + k can be written as the sum of the error estimate from Equation (6) and prior full state estimatex − k aŝ The major assumption for obtaining proper results from ESKF is that the time interval should be short for error calculation and nonlinearity should not be dominant in the calculation of the innovation term. To compensate the effect of nonlinearity in the innovation Equation (46), we propose to modify it by incorporating an RBF neural network. The modified innovation terms k is given as where the term W k y k is the output of the RBF neural network. The term y k is the output of the hidden layer of the RBF neural network and W k is the weight matrix that provides the link between output and hidden layers of RBF neural network. These weights can be designed by minimizing the the mean square error (MSE) cost function J, defined by To improve the estimation of multi-sensor fusion under nonlinear conditions, the Gaussian RBF function utilize an a priori error-state estimate, which is given as where c ik is the neuron center and σ i is the width of the Gaussian RBF function. The weights of the neurons in matrix form w k are represented as where M donates the size of the state vector and N c represents the number of neuron centers. The size of W k is M × Nc. The size of W k y k is M × 1, and the size of y k is N c × 1.

Derivation of Weight Update of RBF-ESKF
For the weight matrix W k , we only consider weight associated with the mth-specific output. Thus, the weight vector of the mth output of RBF can be denoted by w mk . Therefore, the weight update rule of the specific mth output is given as where η w is the learning rate, and its value is selected by experimentation. Too small a value of η w can cause unsuitability, and too large a value can make the response sluggish. A gradient descent method, namely the steepest decent, is used to minimize the cost function J relative to the RBF neural network weight. By using Equation (50), the gradient of cost function can be written as The result of the derivative is given as Thus, putting Equation (54) in Equation (52), we get the complete weight update equation:

Derivation of Center Update of RBF-ESKF
The center update rule for the jth element of the mth neuron center is given as where η c is the learning rate of center update. It is determined experimentally. By applying the steepest descent method, we minimized cost function J with respect to weight. The result of derivative is shown in the following equation: Plugging Equation (57) into Equation (56), the new center update equation becomes It is evident from the weight update Equation (55) and center update Equation (58) that our algorithm uses system information to train the RBF neural network to overcome the drawbacks of ESKF.
The complementary form block diagram representation of the RBF-ESKF fusion algorithm is shown in Figure 3. The vehicle kinematics combined with IMU high-rate sensors of IMU provide the estimated output. This estimated output is then subtracted from low-rate sensors and fed into RBF-ESKF. RBF and ESKF work together to find the best error estimate, which is then added into the underwater vehicle kinematics in feed-forward fashion to obtain the total state.  Algorithm 1 shows iterative steps of the proposed method. The terms k is RBF modified innovation term used by RBF-ESKF to improve accuracy of filter in highly nonlinear conditions.

RBF Gaussian function update:
4: Learning non-linearity of error state vector Innovation update:

5: Non-linearity influence is minimized by using output of RBF neural network in innovation term
Measurement update: 6: Estimate error state by using innovation term and Kalman gain Error state covariance update. 8: Full state is corrected by error adding error estimate.

Full State correction
x =x + δx + k RBF Neural Network Weight and Center update: 9: RBF weight update Time propagation: 11: Time propagation of error state and covariance 12: Next iteration (posterior becomes prior)

Complexity of RBF-ESKF
The proposed algorithm uses RBFNN to enhance the performance of underwater vehicle localization with a slight increased in time and space complexity compared to ESKF due to matrix multiplications. Compared to BPNN and deep learning neural networks, RBFNN has less complexity because of its simple three-layer structure because the time and memory space complexity of the neural networks is directly related to the structure and number of layers. However, faster matrix multiplication algorithms such as the Strassen algorithm [65] can be used to decrease execution time. Table 2 shows the structure of the RBFNN used in this work.

RBFNN Numbers
Input layer neurons 20 Hidden layer neurons 50 output layer neurons 20 Learning rate of weights 0.001 Learning rate of centers 0.001

Results and Discussion
In order to compare performances, the proposed algorithm and ESKF were simulated in three different realistic scenarios. As the ESKF structure was modified by RBF in our fusion algorithm, low-level functions were written for simulation. The noise specifications of the sensors used in this work are comparable to their datasheets. The main purpose of the simulation was to compare the maximum error (max) and root means square error (RMSE) of position, velocity, and attitude. Practical failure mode tests cases were developed and simulated with DVL and acoustic positioning loss of measurements for a short duration. The simulation results were compared with conventional ESKF for performance evaluation. Furthermore, for simulation, the assumption was made that underwater vehicles can move in any direction and with any roll, pitch, and yaw angle. A reference trajectory of the vehicle was generated by angular velocities and acceleration.
To consider the effects of random variations in the accuracy of the fusion algorithms, a Monte Carlo simulation was used. The test consisted of 100 runs. Two filters processed the same data during the test to ensure a fair contrast. For all three cases, the same RBF neural network structure was used, as listed in Table 2. For the simulation, the RBF weights, centers, and sigma were initialized randomly.

Test Case 1: Normal Working Condition
In the first case, the vehicle was considered to be working in a normal operating mode without any on-board and off-board sensor failure. To simulate a real situation, the noise and drift characteristics of the sensors used for simulation were almost the same as listed in the manufacturer's documentation stated under Section 3. Both ESKF and RBF-ESKF were tested on the same operating conditions. The performances of ESKF and RBF-ESKF are compared side by side in Table 3.
From the results of the north position prediction displayed in Table 3, it can be observed that ESKF has an almost three times higher maximum error than that of the RBF-ESKF. Furthermore, for the east position, the ESKF maximum error was twice as high as the RBF-ESKF. In the case of maximum error in the prediction of depth, the performance of the two filters are approximately identical. The RBF-ESKF RMSE for position estimation was even better, almost twice as high. Significant improvement was seen in the north, where it were three times higher than the one achieved by the ESKF. Overall, the RMSE sum for the RBF-ESKF position estimate was approximately two times better than that of the ESKF.
The estimation of velocity showed considerable improvement compared to the ESKF. The overall northern velocity error was almost double than that of the ESKF compared to the RBF-ESKF. Small but noticeable improvements were seen in the maximum error for east and depth velocities. The RMSE of the RBF-ESKF was roughly two times better than the ESKF for the north, east, and down velocities. Overall, with our proposed fusion algorithm, the sum of all RMSE states was almost three times better than ESKF. Overall, in terms of attitude, relative to ESKF, the sum of all state RMSEs improved significantly by about double as much. Figure 4 shows the 2D and 3D trajectories. It can be seen that the trajectory is not linear. The position, velocity, and attitude errors are estimated on this trajectory. The star symbol in the 2D trajectory shows an acoustic fix from the external source. The performances of these ESKF and ESKF-RBF were evaluated by how close the estimated path was to the actual value.    (e,f) Euler angle errors for roll and pitch. It is evident from (c-f) that ESKF has an oscillatory response that contributes to compromised accuracy compared to RBF-ESKF.

Test Case 2: Acoustic Fix Not Available
In this case, the robustness of the multi-sensor fusion algorithm was tested by simulating the loss of the underwater acoustic fix for a short period. The unavailability of the position information from acoustic fix mostly influenced the position estimate of both filters, predominantly ESKF. In this case, the position estimate was only available from the integration of the DVL velocity, which compensated the acoustic fix loss effect and reduced the drift error. The detailed comparison is shown in Table 4. It can be noted from Table 4 that the maximum error of RBF-ESKF was almost five times better for the north and east directions as compared to ESKF. The down position estimate was also two times better than that of ESKF. The overall RMSE in all three directions is considerably improved by almost one and a half times that of ESKF. Overall, compared to normal working conditions, the position accuracy had a detrimental effect, but RBF-ESKF has proven to be more robust.
In contrast to standard operating conditions, the velocity estimation was marginally influenced by acoustic fix unavailability. The maximum ESKF error was worse than that of RBF-ESKF for velocity in all directions. For RBF-ESKF, the RMSE of velocity was significantly better by almost two times that for ESKF. Overall, with our proposed method of fusion, the sum of all RMSE velocity states was almost two times better than ESKF.
For RBF-ESKF, the RMSE of the roll, pitch, and yaw angles were substantially better by about one and a half times that of ESKF. The maximum roll, pitch, and yaw angle errors were two times better for RBF-ESKF. Overall, the sum of all state for RBF-ESKF was better than that for ESKF. Figure 6 below illustrates a comparison of the position and velocity errors when acoustic fix was not available for a short period from 1500 to 1700 s. The performances of ESKF and ESKF-RBF were evaluated on similar trajectory. Moreover, the RBF-EKF is consistent in the estimation and is close to the actual value. (d-f) RBF-EKF has a minimal effect of acoustic fix loss on velocity estimation, and the response is much smoother than that of ESKF.

Test Case 3: DVL Unavailable
In this case, the robustness of both filters was tested when DVL was not available for a short duration. The velocity estimate had a larger influence than the position estimate because, when DVL was not available, it was calculated by taking the derivative of the position measurement, which suffers from noise amplification from the differentiation process. Moreover, acoustic fix measurement bias was negatively influenced by DVL measurements that contribute to increasing the position error. Table 5 compares the performance robustness multi-sensor fusion algorithm with DVL failure.
It can be observed from the above data that position estimate in this case was better than case 2 but slightly less accurate than normal working conditions. However, in the north direction, ESKF maximum error was almost two times worse and almost one and a half times worse for east and depth. Furthermore, the RBF-ESKF RMSE for the position was notably better than that for ESKF by around one and a half times in all directions. Thus, the position of estimation for the overall mission was improved in all directions by the RBF-ESKF algorithm.
The velocity estimation without DVL was less accurate compared to case 1 and case 2. However, the overall results of our method are much better than those of ESKF, which were almost doubly improved with respect to the maximal error and RMSE. In addition, the sum of all estimated velocity states from RBF-ESKF in all directions was approximately two times better than that of ESKF. Hence, the RBF-ESKF velocity estimation was more robust than that of ESKF. The RBF-ESKF attitude estimation of the roll, pitch, and yaw angles had a lower maximum estimation error. Moreover, the RMSE for RBF-ESKF showed considerable improvement. Overall, the sum of all estimated attitude states was one and half times better than that of ESKF.  Figure 7 below shows comparison velocity errors and velocity when the DVL measurement was not available for a short duration from 1900 to 2000 s. The performances of ESKF and ESKF-RBF were evaluated on a similar trajectory as test case 1.
The time complexity of ESKF and ESKF-RBF was tested by running the algorithms on an Intel I7 CPU with 4 GB RAM without any GPU. The testing software used was Matlab 2020b on Windows 7 platform. The timing comparison of both methods is given in Table 6. On a high-speed microcontroller or field-programmable gate array FPGA, the execution time difference will be further reduced. Furthermore, the speed of surveying the type of underwater vehicle is in the range of 2 km/h to 10 km/h; this difference has no significant effect on navigation and localization.
The work tried to fill the gap by proposing a noval ESKF and RBF-augmented fusion solution, which was assessed in three different underwater cases. The ESKF performance strongly relies on the knowledge of the system models and noise properties, which was degraded by nonlinearity. The errors in the RBF-ESKF are smaller than the errors in the ESKF because of the recurcive learning of RBF. Moreover, the fusion algorithm based on RBF-ESKF with help from the aiding sensors was able to correct the drift problems in the INS with better accuracy.
Nevertheless, the proposed algorithm is not limited to underwater; it can also be used in other applications [66,67] such as improving aircraft navigation and tracking by using aerial sensors. Moreover, autonomous ground vehicles are another area where this method can be employed. Furthermore, by improving the Kalman filer response, our methodology can also enhance accuracy satellite attitude estimation [68].

Conclusions
This paper discusses the performance of the proposed algorithm RBF-ESKF for underwater vehicle localization. The primary aim of this work is to take advantage of the RBF neural network to improve the estimation performance of the conventional ESKF for the position, velocity, and attitude of an underwater vehicle. It contrasts outcomes with ESKF in three separate simulations, showing that RBF-ESKF performs better in estimating position, velocity, and attitude. Why the standard ESKF has an inferior performance is because it is designed by employing first-order Taylor series approximation in the error covariance matrix estimation, which results in a decrease in estimation accuracy under high nonlinearity. Thus, important information about the dynamic of underwater is lost because of this realization. However, RBF-ESKF efficiently handles nonlinearity due to its inherent capability for nonlinear function approximation and learning ability.
The research also compared robustness in cases when there is no available position information from the acoustic fix. Here, relative to ESKF, RBF-ESKF demonstrated better accuracy. When acoustic fix becomes available, RBF-ESKF converges quickly. In addition, when DVL fails due to short durations, RBF-ESKF also demonstrates less estimation error.
This work is part of an ongoing work on underwater vehicle localization. The novel proposed algorithm is intended for use as a state estimation for underwater seabed mapping application.
In the future, we would like to test and analyze our algorithm with a different set of sensors with different data rates in underwater environments. For instance, an enhanced version of the multi-sensor fusion algorithm could be designed by using stereo vision and underwater wireless sensor nodes to improve localization.