Low-feature corridor autonomous navigation method and system based on multi-sensor fusion

Abstract

The application provides a low-feature corridor autonomous navigation method and system based on multi-sensor fusion, and relates to the technical field of mobile robot path planning and intelligent control. The steps of the method are as follows: S100, acquiring synchronous collection multi-source data of a distributed sensor network, and extracting cross-medium boundary features; S200, constructing a cross-medium propagation coupling model based on a multi-physical field equation, and calculating physical field distribution parameters; S300, inputting historical data and boundary features into a physical information neural network, and predicting a physical field evolution trend; S400, executing a risk assessment algorithm, and generating a comprehensive safety map containing multi-dimensional risk indicators; and S500, generating a safe navigation trajectory considering multiple constraints based on a multi-objective optimization algorithm. The method has the beneficial effects that the problem of failure of a traditional single physical model when a medium mutates is overcome, the accuracy and physical consistency of the prediction of the physical field evolution trend are improved, and comprehensive environmental safety information is provided for path planning.

Description

Technical Field

[0001] This invention relates to the field of mobile robot path planning and intelligent control technology, specifically to a low-feature corridor autonomous navigation method and system based on multi-sensor fusion. Background Technology

[0002] In the underground corridor environments of modern industrial facilities, especially the underground utility tunnels of critical infrastructure such as nuclear power plants and chemical plants, autonomous navigation of robots faces extremely complex challenges. These corridors typically contain interfaces of multiple media, strong sources of electromagnetic interference, and complex acoustic environments.

[0003] Existing technologies mainly employ navigation methods based on a single sensor, such as vision-based SLAM, lidar-based positioning methods, or sonar-based underwater navigation systems. These methods can achieve good navigation results in specific environments, but they have significant shortcomings in complex multi-media corridor environments.

[0004] Existing technologies suffer from the following drawbacks: First, the propagation characteristics of physical fields vary greatly in different media, causing sensor systems based on a single physical model to fail when crossing media boundaries. Second, the boundaries of strong magnetic field regions are blurred and dynamically changing, interfering not only with magnetic navigation devices such as electronic compasses but also potentially affecting the normal operation of other electronic systems. Finally, the speed of sound in underwater environments is affected by factors such as temperature and salinity, and severe reverberation significantly reduces the accuracy of sonar positioning. These combined factors create technical challenges for robots in achieving accurate perception and safe navigation in complex corridor environments. Summary of the Invention

[0005] The main objective of this invention is to provide a low-feature corridor autonomous navigation method and system based on multi-sensor fusion, thereby solving the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a low-feature corridor autonomous navigation method based on multi-sensor fusion, comprising the following steps: S100: Acquire synchronous multi-source data from a distributed sensor network and extract cross-medium boundary features; S200. Construct a cross-medium propagation coupling model based on multiphysics equations and calculate the physical field distribution parameters; S300: Input historical data and boundary features into the physical information neural network to predict the evolution trend of the physical field; S400 executes a risk assessment algorithm to generate a comprehensive security map containing multi-dimensional risk indicators; S500 generates safe navigation trajectories that take into account multiple constraints based on a multi-objective optimization algorithm.

[0007] Furthermore, in step S100, the distributed sensor network includes: a distributed magnetic field sensor array, a multi-beam sonar system, and a conductivity sensor; Real-time data acquired by distributed sensor networks needs to undergo preprocessing operations, including: removing magnetic field measurement noise, separating direct waves and reverberation components in sonar echoes, and smoothing conductivity measurements. The process of extracting cross-medium boundary features is as follows: preprocessing and aligning multi-source data in time; identifying the spatial location of the air-water interface based on the relationship between the conductivity gradient change rate and a set threshold; marking the boundary range of strong magnetic field regions based on outlier detection of the spatial distribution of magnetic field strength; and extracting acoustic propagation characteristic parameters of the underwater environment based on the time delay and intensity characteristics of acoustic echo signals. Acoustic propagation characteristic parameters include: echo delay, echo intensity, spectral center frequency, and spectral bandwidth.

[0008] Further, the detailed process of step S200 is as follows: establish a propagation model of the electromagnetic field in different media based on Maxwell's equations; establish a propagation model of the sound field in the underwater environment based on the sound wave equations; couple the electromagnetic field model and the sound field model at the boundary of the medium to establish a set of boundary condition equations; solve the coupled equations to obtain numerical solutions of the electromagnetic field intensity distribution and sound pressure field distribution in different medium regions. The boundary condition equations include: electromagnetic field boundary conditions and acoustic field boundary conditions; Electromagnetic field boundary conditions include: the continuity condition of the tangential component of the electric field and the continuity condition of the tangential component of the magnetic field; The boundary conditions of the sound field include: the sound pressure continuity condition and the normal particle vibration velocity continuity condition; The distribution of physical fields includes: electromagnetic field intensity distribution and sound pressure field distribution.

[0009] Furthermore, before step S200 is executed, a preprocessing step is also included as follows: the corridor space is meshed using the finite element method, and an adaptive mesh refinement strategy is adopted near the medium boundary; The adaptive mesh refinement strategy determines whether the mesh needs to be refined based on the rate of change of the physical field gradient between adjacent cells. When the rate of change of the physical field gradient is greater than the refinement threshold, the cell is subdivided into smaller sub-cells.

[0010] Furthermore, the encoder module of the physical information neural network uses a Long Short-Term Memory (LSTM) network to extract temporal features of the magnetic field strength time series, acoustic feature vectors, and boundary position information. Historical information is selectively retained through gating mechanisms such as forget gate, input gate, and output gate. The output of the LSTM is mapped to the intermediate feature space through a fully connected layer. The physical constraint module receives the intermediate feature vector output by the encoder and combines it with the physical equation constraints in step S200. This can constrain the model fitting trend and improve the rationality and solution accuracy of the field quantity calculation results. The decoder module consists of a multi-layer fully connected neural network, which outputs predicted values ​​of magnetic field strength and sound pressure distribution at different spatial locations over a future period of time. The predicted values ​​output by the decoder are denormalized to obtain the magnetic field strength and sound pressure distribution. The optimization strategy for the physical information neural network is to use the Adam optimizer to minimize the composite loss function, which is expressed as follows: (1); in, The loss value is the composite loss function. , These are the data fitting loss and the physical equation residual loss, respectively. This is the balance coefficient; The data fitting loss is defined as the mean square error between the predicted and the true values; the physical equation residual loss is defined as the sum of squared residuals of the network output on the physical equation.

[0011] Furthermore, the detailed process of step S400 is as follows: Based on the physical field prediction results, the magnetic field hazard level and acoustic unreliability of each spatial grid point are calculated. The magnetic field hazard level and acoustic unreliability are mapped to the interval 0~1 using the mean normalization method based on the range. Then, a comprehensive risk value is generated by weighted fusion. The calculated comprehensive risk value is mapped to the corridor spatial coordinate system to generate a three-dimensional comprehensive safety map.

[0012] Furthermore, in step S400, the expression for the magnetic field hazard level is as follows: (2); in, The magnetic field hazard level is [not specified]. This is the predicted value of the magnetic field strength. As a safety threshold, It is a very small positive number; The expression for acoustic unreliability is as follows: (3); in, Acoustic unreliability The reverberant power spectral density; The expression for the overall risk value is as follows: (4); in, For the comprehensive risk value, For normalized , For normalized , , These are the magnetic field risk coefficient and the acoustic unreliability coefficient, respectively, and their sum is 1.

[0013] Furthermore, the method for calculating the reverberation power spectral density is as follows: First, perform a short-time Fourier transform on the reverberation components separated in step S100 to obtain the time spectrum; calculate the square of the time spectrum modulus to obtain the power spectrum; finally, integrate the power spectrum in the frequency domain to obtain the reverberation power spectral density, as expressed below: (5); in, For the sonar operating frequency, , These are the lower and upper limits of the sonar operating frequency, respectively. This is the power spectrum.

[0014] Furthermore, the detailed process of step S500 is as follows: Acquire the current depth value measured by the pressure sensor, the flow velocity vector measured by the flow sensor, and the current position and attitude information provided by the inertial measurement unit; integrate this real-time data with the comprehensive safety data generated in step S400. Figure 1 The input is fed into a multi-objective path optimization algorithm; the multi-objective path optimization algorithm simultaneously optimizes three objective functions, including: minimizing path length, minimizing comprehensive risk value, and minimizing the number of cross-media steps; the improved A* algorithm is used to search for a set of feasible paths that satisfy multiple constraints in a discretized spatial grid, and then a smooth navigation trajectory is generated by Bézier curve fitting; the navigation trajectory includes position sequence, velocity planning, and expected arrival time; The improved A* algorithm takes the starting position, target position, integrated safety map, and a set of constraints as input, and outputs the optimal path node sequence that satisfies multiple constraints. The improved A* algorithm includes the following steps: S5001. Initialize the open list and the closed list, add the starting node to the open list, and set the cost value of the starting node to 0; S5002. Select the node with the smallest comprehensive evaluation function value from the open list as the current node; the expression of the comprehensive evaluation function is as follows: (6); in, For any node, This is the comprehensive evaluation function value for this node. The actual cost from the starting point to this node. For a heuristic estimate of the cost from this node to the target point, This refers to the risk penalty item at this node; S5003. Determine if the current node is the target node. If it is, backtrack the parent node chain from the target node to generate a complete path and terminate the search. Otherwise, move the current node from the open list to the closed list. The parent node is the previous node in the path that leads to the current node. S5004. Expand all neighboring nodes of the current node. For each neighboring node, calculate the temporary value of the distance from the starting point to that neighboring node via the current node. Neighbor nodes are defined as nodes that are adjacent to the current node in the spatial grid, and the distance difference between each neighbor node and the current node in the x, y, and z directions does not exceed one grid spacing; the temporary cost value is the candidate value of the node's actual cost, used for path cost prediction and comparison; after filtering the optimal candidate value, it is assigned and updated to the node's actual cost value; The expression for the temporary cost value is as follows: (7); in, For the current node The temporary value, For the current node any of its neighboring nodes, Let $\frac{ ... (8); in, For the Euclidean norm, This is the depth penalty coefficient. For reference depth, The depth difference between the two nodes; S5005. Perform constraint checks on each neighboring node to determine whether the node satisfies the magnetic field strength constraint, acoustic positioning error constraint, and kinematic constraint. If it does not meet the constraints, skip the neighboring node. The magnetic field strength constraint is specifically: the magnetic field strength of neighboring nodes is not greater than the magnetic field strength constraint threshold. The acoustic positioning error constraint is specifically: the acoustic positioning error of neighboring nodes is not greater than the acoustic positioning error constraint threshold. The kinematic constraints are as follows: the Euclidean norm of the moving speed is not greater than the speed threshold; the Euclidean norm of the acceleration is not greater than the acceleration threshold; and the turning radius is not less than the minimum turning radius threshold. S5006. For neighboring nodes that satisfy the constraints, calculate the risk penalty term and heuristic cost; The expression for the risk penalty item is as follows: (9); in, For neighboring nodes Risk penalty items, , These are the risk weighting coefficient and the cross-media penalty coefficient, respectively. For neighboring nodes The overall risk value at the location, This is a cross-medium indicator function, which takes a value of 1 when the two nodes are located in different media, and 0 otherwise. The expression for the heuristic cost is as follows: (10); in, For neighboring nodes The heuristic cost, , Neighboring nodes With the target node The Euclidean distance and the weighted distance considering risk distribution between them. , These are the balance weights for Euclidean distance and weighted distance, respectively; The expression for the weighted distance between neighboring nodes and the target node, taking into account risk distribution, is as follows: (11); in, For the movement path; S5007. If a neighboring node is not in the open list, add it to the open list and record its parent node as the current node; if a neighboring node is already in the open list and the newly calculated temporary value is less than its original actual cost, update the actual cost of the neighboring node to the temporary value, and update the parent node at the same time. S5008. Return to step S5002 and continue searching until the target node is found or the open list is empty; In addition to step S500, step 501 is also included, which is: implementing a rolling time-domain optimization strategy, replanning the local path at fixed intervals based on the latest sensor data and prediction results, and realizing dynamic adjustment of the trajectory; The input to the dynamic temporal optimization strategy is the current robot position, current velocity, global path node sequence, latest sensor data, and the prediction result output from step S300. The output is the local trajectory within the future time window. The rolling temporal optimization strategy extracts the forward path segment starting from the current position on the global path as a reference trajectory within each planning cycle. Based on the latest environmental information and the predicted physical field distribution, a local optimization problem is constructed around the reference trajectory. The objective function of the dynamic temporal optimization is: (12); in, The objective function value, The length of the future time window. For reference trajectory, To control the input, , , These are the weighting coefficients for the trajectory control term, risk control term, and input control term, respectively. Step S500 also includes step S502, the details of which are as follows: Establish an emergency obstacle avoidance mechanism to quickly generate a temporary avoidance path when a sudden magnetic field anomaly or obstacle is detected; Emergency obstacle avoidance mechanisms include: anomaly detection and obstacle avoidance path generation; The anomaly detection section determines whether to trigger emergency obstacle avoidance by monitoring the rate of change of magnetic field strength and the sonar echo distance in real time. When the rate of change of magnetic field strength is greater than the emergency threshold of the transformation rate or the sonar echo distance is less than the emergency safe distance, it is determined to be a sudden anomaly and emergency obstacle avoidance is immediately initiated.

[0015] Furthermore, the system includes: The distributed sensor network module, including a magnetic field sensor array, a multi-beam sonar system, and a conductivity sensor, is used for multi-source data acquisition and cross-medium boundary feature extraction. The physical field modeling and prediction module includes a cross-medium propagation coupling model construction unit and a physical information neural network prediction unit. It is used to build an artificial potential field model, calculate the global potential energy distribution, and predict the dynamic trend of the potential field as the environment and obstacle positions change, providing field basis for robot trajectory planning and obstacle avoidance control. The risk assessment module is used to calculate the magnetic field hazard level and acoustic positioning reliability, and generate a comprehensive safety map; The path planning module is used to execute multi-objective optimization algorithms to generate safe navigation trajectories; The execution control module controls the robot's movement according to the navigation trajectory and implements rolling time-domain optimization and emergency obstacle avoidance.

[0016] The beneficial effects are as follows: (1) This invention uses a cross-medium propagation coupling model based on multi-physics field equations to accurately describe the propagation characteristics of electromagnetic fields and sound fields in different media and their interaction at the boundary. It overcomes the problem of failure of traditional single physical models when the medium changes abruptly, and realizes accurate modeling of the physical field distribution in complex multi-medium environments.

[0017] (2) By introducing a physical information neural network, the physical equation constraints are combined with data-driven learning, which improves the accuracy and physical consistency of the prediction of the evolution trend of the physical field.

[0018] (3) The generation mechanism of the comprehensive safety map quantifies and integrates multiple risk factors, providing comprehensive environmental safety information for route planning.

[0019] In summary, this invention solves the technical problem of robots struggling to achieve accurate perception and safe navigation in complex corridor environments containing multiple media interfaces and strong electromagnetic interference sources, and achieves the technical effect of improving the reliability and safety of autonomous navigation. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the overall process of the low-feature corridor autonomous navigation method based on multi-sensor fusion of the present invention; Figure 2 It is a graph showing the variation trend of magnetic field strength and conductivity along the longitudinal position of the corridor; Figure 3 This is a comparison chart of the predicted magnetic field strength and the actual measured value; Figure 4 This is a graph showing the change in the overall risk value of the robot as it travels along the optimal navigation path. Detailed Implementation

[0021] Example 1 like Figures 1-4 As shown, the low-feature corridor autonomous navigation method based on multi-sensor fusion includes the following steps: S100. Acquire multi-source synchronous data from the distributed sensor network and extract cross-medium boundary features. The detailed process is as follows: Real-time data from the distributed magnetic field sensor array, multibeam sonar system, and conductivity sensor deployed in the corridor are acquired; the triaxial magnetic field strength data acquired by the magnetic field sensor at a frequency of 10 Hz, the echo signal data acquired by the sonar system at a frequency of 20 Hz, and the continuous measurement data from the conductivity sensor are synchronized in time; the spatial location of the air-water interface is initially identified based on the location where the conductivity gradient change rate exceeds a set threshold; the boundary range of the strong magnetic field area is marked based on the anomaly detection of the spatial distribution of magnetic field strength; and the acoustic propagation characteristic parameters of the underwater environment are extracted based on the time delay and intensity characteristics of the acoustic echo signal. The threshold expression for the rate of change of conductivity gradient is as follows: (1); in, The threshold for the rate of change of conductivity gradient. For electrical conductivity, This refers to the spatial distance; the threshold value is typically greater than or equal to 10, and it is determined based on the difference between the conductivity of air and the conductivity of water, which can effectively identify the location of the medium interface. Three methods were selected for detecting outliers in the spatial distribution of magnetic field strength. The rule is that any spatial location The magnetic field strength at a given location is compared with the mean and standard deviation of the magnetic field strength in the surrounding area. If the absolute value of the difference between the magnetic field strength at a given location and the mean magnetic field strength in the surrounding area is greater than three times the standard deviation of the magnetic field strength in the surrounding area, then the location is considered... The magnetic field strength value at that location is an anomaly; it will be marked. This location marks the boundary of a strong magnetic field region. Acoustic propagation characteristic parameters include: echo delay, echo intensity, spectral center frequency, and spectral bandwidth. The echo delay is obtained by detecting the time difference between the peak position of the echo signal and the transmission time. The echo intensity is obtained by calculating the root mean square value of the echo signal. The spectral center frequency and bandwidth are obtained by analyzing the spectral energy distribution after performing a Fast Fourier Transform (FFT) on the echo signal. It should be noted that a magnetic field sensor array refers to multiple triaxial fluxgate sensors distributed in a grid or linear pattern in a corridor space, with each sensor node capable of measuring the magnetic field strength vector at its location; a multibeam sonar system refers to an acoustic imaging device capable of simultaneously emitting multiple sound beams and receiving echoes, used to acquire three-dimensional acoustic images of the underwater environment; and a conductivity sensor determines whether the current location is in an air or water environment by measuring the conductivity value of the medium.

[0022] Since the raw signals collected contain a lot of noise, the real-time data collected by the distributed sensor network needs to be preprocessed, including: using a Kalman filter to remove magnetic field measurement noise, using an adaptive thresholding method to separate the direct wave and reverberation components in the sonar echo, and using a moving average algorithm to smooth the conductivity measurement values. The input to the Kalman filter is the triaxial magnetic field strength time series data acquired by the magnetic field sensor array. The output is the denoised magnetic field strength estimate. The Kalman filter includes a prediction step and an update step. The prediction step obtains the prior state estimate based on the system state transition equation, and the update step corrects the state estimate based on the observation equation and the Kalman gain matrix. The state transition equation is expressed as follows: (2); in, for The system state vector at time t, i.e., the time series data of the three-axis magnetic field strength. This is the state transition matrix; here, we take the identity matrix. This is process noise; The expression for the observation equation is as follows: (3); in, for Sensor observations at time [time] The observation matrix is ​​taken here as the identity matrix. for Zero-mean Gaussian observation noise at time t; The expression for the Kalman gain matrix is ​​as follows: (4); in, for The Kalman gain matrix at time 10:00. for The transpose of the matrix, for The prior estimate error covariance at time 1 / 2. The noise covariance matrix is ​​observed; this matrix is ​​a diagonal matrix. The expression for the state estimation correction is as follows: (5); in, After filtering The optimal state estimate at time t, i.e. the magnetic field strength after denoising; The adaptive thresholding method takes the time series of echo signals acquired by the sonar system as input and outputs the direct wave component and reverberation component. The adaptive thresholding method calculates the time-varying threshold based on the energy distribution characteristics of the echo signal. The expression for the time-varying threshold is as follows:

[0023] in, For time-varying thresholds, The local mean of the signal. For the local standard deviation of the signal, The adjustment coefficient is used to determine the portion of the signal amplitude that is greater than the time-varying threshold as the direct wave and the portion that is less than the time-varying threshold as the reverberation.

[0024] The adjustment coefficient ranges from 1.5 to 3, with a preferred value of 2; the local mean and local standard deviation are calculated at time... The calculation was performed within a sliding time window of 0.1 seconds before and after the sampling. Since the sonar sampling frequency is 20Hz, the number of sampling points within the window is 4. The moving average algorithm takes as input a conductivity measurement sequence collected by a conductivity sensor and outputs a smoothed conductivity value sequence. The calculation process of the moving average algorithm is as follows: first, determine the window length of the sliding window; then, calculate the average conductivity value within the window; then, move the window one unit to the right and calculate the average conductivity value within the window again. Repeat the above operations to obtain the smoothed conductivity value sequence. The window length is an integer between 5 and 15.

[0025] S200. Based on the multiphysics equations, a cross-medium propagation coupling model is constructed, and the physical field distribution parameters are calculated. The detailed process is as follows: Based on Maxwell's equations, a propagation model of the electromagnetic field in different media is established, considering the difference in relative permittivity between air and water. A propagation model of the sound field in an underwater environment is established based on the acoustic wave equations. The electromagnetic field model and the sound field model are coupled at the media boundary to establish a set of boundary condition equations. Solving the coupled equations yields numerical solutions for the electromagnetic field intensity distribution and sound pressure field distribution in different media regions. The relative permittivity of air is 1, and that of water is 80. The relationship between sound speed and the temperature, salinity, and depth of the underwater environment is expressed as follows: (6); in, For the speed of sound, For temperature, Salinity The depth is measured in real-time by temperature sensors deployed underwater, ranging from 0 to 40 degrees Celsius. Salinity ranges from 0 to 40 PSU; in freshwater environments within industrial corridors, salinity is 0, while in saline wastewater environments, salinity is determined through actual measurements. Depth ranges from 0 to 50 m, calculated by converting water pressure measured by pressure sensors. The conversion relationship is as follows: (7); in, To measure pressure, Atmospheric pressure. For the density of water, It is the acceleration due to gravity; Boundary conditions include: electromagnetic field boundary conditions and acoustic field boundary conditions; electromagnetic field boundary conditions include: continuity condition of the tangential component of the electric field and continuity condition of the tangential component of the magnetic field. Let the medium interface be denoted as The expression for the continuity condition of the tangential component of the electric field at the interface of the medium is as follows: (8); in, Let be the unit normal vector of the interface between the media. , These are the electric field intensity vectors in medium 1 and medium 2, respectively. This represents the vector cross product, used to extract the tangential component; The expression for the continuity condition of the tangential component of the magnetic field at the interface of the medium is as follows: (9); in, , These are the magnetic field strength vectors in medium 1 and medium 2, respectively; The boundary conditions of the sound field include: the sound pressure continuity condition and the normal particle vibration velocity continuity condition; The sound pressure continuity condition at the medium interface is specifically defined as follows: the sound pressure levels of medium 1 and medium 2 at the medium interface are equal. The expression for the continuous condition of normal particle vibration velocity at the interface of a medium is as follows: (10); in, , The sound pressures in medium 1 and medium 2 are respectively. , These are the densities of medium 1 and medium 2, respectively. It should be noted that the cross-medium propagation coupling model is a comprehensive physical model that simultaneously considers the propagation characteristics of electromagnetic waves and sound waves in different media. The cross-medium propagation coupling model can describe the reflection, refraction and transmission phenomena of physical fields at the boundaries of media. In this embodiment of the application, in order to improve the model calculation efficiency, a preprocessing step is included before step 2: the corridor space is meshed using the finite element method, and an adaptive mesh refinement strategy is adopted near the medium boundary to ensure that the boundary conditions are met; The specific method for solving the coupled equation system is as follows: The linear equations obtained by discretization using the finite element method are expressed as follows: (11); in, The overall stiffness matrix, Let be the vector of unknown quantities at the nodes, namely electric field, magnetic field, and sound pressure. This is the load vector; The system of linear equations is solved iteratively using the conjugate gradient method, and parallel computing techniques are employed to accelerate the solution process. The convergence criterion for the iteration is that the relative residuals satisfy the following: (12); in, For the first The residual vector of the next iteration For the first The result of the iteration calculation value, It is the vector norm, used to characterize the overall size of the vector; parallel computing techniques are used to accelerate the solution process of large-scale matrix equation systems; The adaptive mesh refinement strategy determines whether mesh refinement is needed based on the rate of change of the physical field gradient between adjacent cells. When the rate of change of the physical field gradient is greater than 0.1, the cells are subdivided into smaller sub-cells. The mesh generation uses tetrahedral elements, with an initial mesh size of 0.2m, and a minimum mesh size of 0.05m after refinement near the boundary. The expression for the rate of change of the physical field gradient is as follows: (13); in, The rate of change of the gradient of the unit physical field. For Hamiltonian operators, For the physical field quantities of the current computing unit, These are the physical field quantities of adjacent computational units.

[0026] Step S300: Input historical data and boundary features into the physical information neural network to predict the evolution trend of the physical field. The detailed process is as follows: The magnetic field strength time series data of the past 24 hours, the acoustic feature vector after reverberation separation processing, and the boundary position information extracted in step S100 are input into the Physical Information Neural Network (PINN). An uncertainty quantification module is introduced to output the predicted value and give the confidence interval, providing probability distribution information for subsequent risk assessment. Before inputting the three types of feature data into the network, Z-score standardization is required to eliminate the differences in dimensions and numerical ranges between different physical quantities. Z-score standardization standardizes each type of feature data, and the standardized data has a mean of 0 and a standard deviation of 1, so that different physical quantities are in the same numerical range. To improve prediction accuracy, the three types of feature data need to be processed using an attention mechanism after being input into the network. Specifically, the attention mechanism is used to adaptively weight the importance of different sensor data and dynamically adjust the contribution of each sensor data in the prediction based on the current environmental characteristics. The physical information neural network consists of three parts: an encoder module, a physical constraint module, and a decoder module. The encoder module uses a Long Short-Term Memory (LSTM) network to extract temporal features of magnetic field strength time series, acoustic feature vectors, and boundary position information. It selectively retains historical information through gating mechanisms such as forget gate, input gate, and output gate. The output of the LSTM is mapped to the intermediate feature space through a fully connected layer. The physical constraint module receives the intermediate feature vector output by the encoder and combines it with the physical equation constraints in step S200. This can constrain the model fitting trend and improve the rationality and solution accuracy of the field quantity calculation results. The decoder module consists of a multi-layer fully connected neural network, with each layer connected by the ReLU activation function. It outputs predicted values ​​of magnetic field strength and sound pressure distribution at different spatial locations within the next hour. The predicted values ​​output by the decoder are denormalized to obtain magnetic field strength and sound pressure distribution with actual physical meaning. The training of the physical information neural network adopts a physically constrained supervised learning mode, and the optimization strategy is to use the Adam optimizer to minimize the composite loss function. The expression of the composite loss function is as follows: (14); in, The loss value is the composite loss function. , These are the data fitting loss and the physical equation residual loss, respectively. This is the balance coefficient; the balance coefficient is used to adjust the relative importance of data-driven and physical constraints, and its value ranges from 0.1 to 10, with a preferred value of 1. The data fitting loss is defined as the mean squared error between the predicted and actual values, and its expression is as follows: (15); in, The number of training samples. , The first The true physical network value and network prediction value of each training sample; The residual loss of the physical equations is defined as the sum of squared residuals of the network output on the physical equations, and its expression is as follows: (16); in, To constrain the number of sampling points physically, For the first The physical equation operator for each physical constraint sampling point; the optimal number of physical constraint sampling points is 500; The expression for the physical equation operator is as follows: (17); in, For physical equation operators, here referring to the physical equation operator at any physical constraint sampling point. , These are the residual terms of Faraday's law of electromagnetic induction and the residual terms of the sound wave equation, respectively. The residual term of Faraday's law of electromagnetic induction is expressed as follows: (18); in, For electric field strength, The magnetic field strength; The residual term expression of the acoustic wave equation is as follows: (18); in, For the Laplace operator, Sound pressure level; The Hamiltonian operator in the residual loss of the physical equation is discretized and calculated on spatial grid points using the finite difference method, and the time derivative is calculated using the backward difference method. The training data preparation process for the neural network includes: collecting historical sensor data in the corridor environment and generating training samples by sliding segmentation according to the time window; dividing the samples into training set, validation set and test set; the parameters of the neural network are initialized using the Xavier initialization method; the training process uses mini-batch stochastic gradient descent, and training is stopped early when the validation set loss does not decrease for several consecutive rounds to prevent overfitting; It should be noted that physical information neural networks are a deep learning method that embeds physical equation constraints into the training process of neural networks. By adding physical equation residual terms to the loss function, the network output automatically satisfies physical laws, thereby improving the physical consistency and reliability of predictions. The aforementioned long short-term memory network selectively retains or forgets historical information through gating units, which can effectively capture long-term dependencies in time series.

[0027] The aforementioned fully connected layer maps the hidden state of the LSTM to a feature space suitable for the decoder to process; The aforementioned physical equation residual loss is calculated by automatic differentiation technique, which calculates the derivative of the network output with respect to the input spatial coordinates. The derivative is then substituted into the physical equation to obtain the residual value. During training, by minimizing the physical equation residual loss, the network is constrained to the physical feasible solution space, thus enabling it to provide prediction results that conform to physical laws even in sparse data regions. The uncertainty quantification module is implemented using the Monte Carlo Dropout method, which keeps the Dropout layer in the decoder active during the prediction phase. For the same input sample, multiple forward propagations are performed, and the mean of the prediction results is calculated as the final prediction value. The standard deviation is calculated as a measure of prediction uncertainty. Confidence intervals are constructed based on the normal distribution assumption. When the uncertainty is high, a more conservative approach is needed in the risk assessment.

[0028] S400 executes a risk assessment algorithm to generate a comprehensive security map containing multi-dimensional risk indicators. The detailed process is as follows: Based on the physical field prediction results of step S300, the magnetic field hazard level and acoustic unreliability of each spatial grid point are calculated. The magnetic field hazard level and acoustic unreliability are mapped to the interval 0~1 using the mean normalization method based on the range, eliminating the differences in the numerical range of different indicators. Then, a comprehensive risk value is generated by weighted fusion. The calculated comprehensive risk value is mapped to the corridor spatial coordinate system to generate a three-dimensional comprehensive safety map. The expression for the magnetic field hazard level is as follows: (19); in, The magnetic field hazard level is [not specified]. This is the predicted value of the magnetic field strength. As a safety threshold, It should be a very small positive number to avoid the occurrence of Too small a value is meaningless; the predicted magnetic field strength is obtained by calculating the magnitude of the predicted triaxial magnetic field strength components; for the magnetic field hazard level, a value greater than 0 indicates high risk, and a value less than 0 indicates low risk; the safety threshold is set to 50. better; The expression for acoustic unreliability is as follows: (20); in, Acoustic unreliability Let Reverberation power spectral density be denoted as follows: First, perform a short-time Fourier transform on the reverberation components separated in step S100 to obtain the time spectrum. Calculate the square of the time spectrum modulus to obtain the power spectrum. Finally, integrate the power spectrum in the frequency domain to obtain the reverberation power spectral density, as expressed below: (twenty one); in, For the sonar operating frequency, , These are the lower and upper limits of the sonar's operating frequency, typically taken as 20kHz and 80kHz, respectively. Power spectrum; The expression for the overall risk value is as follows: (twenty two); in, For the comprehensive risk value, For normalized , For normalized , , These are the magnetic field risk coefficient and the acoustic unreliability coefficient, respectively, and their sum is 1; under normal circumstances, the magnetic field risk coefficient is taken as 0.6; It should be noted that the comprehensive safety map is a spatial risk distribution map generated by taking into account risk factors such as magnetic field strength and acoustic positioning accuracy. Each spatial location is marked with a corresponding safety level to guide route planning.

[0029] S500 generates a safe navigation trajectory considering multiple constraints based on a multi-objective optimization algorithm. The detailed process is as follows: Acquire the current depth value measured by the pressure sensor, the flow velocity vector measured by the flow sensor, and the current position and attitude information provided by the inertial measurement unit; integrate this real-time data with the comprehensive safety data generated in step S400. Figure 1 The input is fed into a multi-objective path optimization algorithm. This algorithm simultaneously optimizes three objective functions: path length minimization, overall risk minimization, and cross-medium number minimization. An improved A* algorithm is used to search for a set of feasible paths satisfying multiple constraints in a discretized spatial grid, and then a smooth navigation trajectory is generated through Bézier curve fitting. The navigation trajectory includes a position sequence, velocity planning, and expected arrival time. The Bézier curve fitting method uses a cubic Bézier curve to smooth the discrete path node sequence output by the A* algorithm. Specifically, for ease of representation, three adjacent path nodes are denoted as follows: , , Based on these 3 path nodes, 4 control points are constructed, and the control points are denoted as follows: , , , A cubic Bézier curve segment is constructed based on control points, and linear interpolation is used to ensure the continuity of the curve. For each Bézier curve segment, 20 uniformly distributed parameter points are sampled to generate smooth trajectory points. Velocity planning is calculated based on the smoothed trajectory points, and a trapezoidal velocity curve is used to ensure that the acceleration satisfies kinematic constraints, with the initial and final velocities being 0. It should be noted that here... , Path nodes and , Control points and subsequent points, such as target nodes and neighbor nodes, represent the path nodes or control points themselves and their corresponding coordinates. In addition, the coordinates here include the horizontal and vertical coordinates. The relationship between control points and path nodes is as follows: (twenty three); The expression for a cubic Bézier curve segment is as follows: (twenty four); in, For normalized position parameters, For parameters The corresponding coordinates of the fitted path; The improved A* algorithm takes the starting position, target position, integrated safety map, and a set of constraints as input, and outputs the optimal path node sequence that satisfies multiple constraints. The improved A* algorithm includes the following steps: S5001. Initialize the open list and the closed list, add the starting node to the open list, and set the cost value of the starting node to 0; S5002. Select the node with the smallest comprehensive evaluation function value from the open list as the current node; the expression of the comprehensive evaluation function is as follows: (twenty four); in, For any node, This is the comprehensive evaluation function value for this node. The actual cost from the starting point to this node. For a heuristic estimate of the cost from this node to the target point, This refers to the risk penalty item at this node; S5003. Determine if the current node is the target node. If it is, backtrack the parent node chain from the target node to generate a complete path and terminate the search. Otherwise, move the current node from the open list to the closed list. The parent node is the previous node in the path that leads to the current node. S5004. Expand all neighboring nodes of the current node. For each neighboring node, calculate the temporary value of the distance from the starting point to that neighboring node via the current node. Neighbor nodes are defined as the 26 nodes adjacent to the current node in the spatial grid, with the distance difference between each neighbor node and the current node in the x, y, and z directions not exceeding one grid spacing; the temporary cost value is the candidate value of the node's actual cost, used for path cost prediction and comparison; after filtering the optimal candidate value, it is assigned and updated to the node's actual cost value; The expression for the temporary cost value is as follows: (25); in, For the current node The temporary value, For the current node any of its neighboring nodes, Let $\frac{ ... (26); in, For the Euclidean norm, This is the depth penalty coefficient. For reference depth, This represents the depth difference between the two nodes; the depth here refers to the hierarchical span of selecting neighboring nodes outwards. For example, a depth of 1 means selecting the first node immediately before or after the current node, while a depth of 2 means selecting one more level of nodes outwards. The reference depth is the maximum depth of the corridor, which is determined according to the actual structure of the corridor. For typical industrial corridors, it is generally taken as 20m. S5005. Perform constraint checks on each neighboring node to determine whether the node satisfies the magnetic field strength constraint, acoustic positioning error constraint, and kinematic constraint. If it does not meet the constraints, skip the neighboring node. The magnetic field strength constraint is specifically defined as follows: the magnetic field strength of neighboring nodes must not exceed a magnetic field strength constraint threshold, which is typically set to 100. The magnetic field strength constraint threshold is the upper limit of the magnetic field strength that the equipment can withstand. Exceeding the magnetic field strength constraint threshold may cause electronic system failure or permanent damage. The acoustic positioning error constraint is as follows: the acoustic positioning error of neighboring nodes is not greater than the acoustic positioning error constraint threshold, which is generally taken as 0.5m. The acoustic positioning error constraint threshold is the maximum allowable positioning error required to ensure navigation safety. When the positioning error exceeds the acoustic positioning error constraint threshold, the robot may deviate from the safe area. The kinematic constraints are as follows: the Euclidean norm of the moving velocity is not greater than a velocity threshold, which is typically taken as 1.5 m / s²; the Euclidean norm of the acceleration is not greater than an acceleration threshold, which is typically taken as 0.5 m / s². 2 The turning radius shall not be less than the minimum turning radius threshold, which is generally taken as 0.8m; these constraints are determined based on the robot's physical characteristics and dynamic capabilities. S5006. For neighboring nodes that satisfy the constraints, calculate the risk penalty term and heuristic cost; The expression for the risk penalty item is as follows: (27); in, For neighboring nodes Risk penalty items, , These are the risk weighting coefficient and the cross-media penalty coefficient, respectively. For neighboring nodes The overall risk value at the location, This is a cross-medium indicator function, which takes a value of 1 when the two nodes are located in different media, and 0 otherwise. In the preferred scheme, a risk weighting coefficient of 10 is more effective. The risk weighting coefficient converts the dimensionless comprehensive risk value into a numerical order of magnitude equivalent to the distance cost, which significantly increases the cost in high-risk areas. A cross-medium penalty coefficient of 5m is preferred. The cross-medium penalty coefficient is the equivalent distance cost added when crossing the medium boundary, reflecting the additional costs such as sensor switching and parameter recalibration caused by medium conversion. The preferred value for the depth penalty coefficient is 0.3, which is used to balance the cost difference between horizontal and vertical movement. The risk penalty term directly integrates risk information from the comprehensive safety map into the cost calculation of the A* algorithm, causing the path search process to automatically tend towards low-risk areas; the cross-medium indicator function increases the cost of crossing medium boundaries and reduces the number of medium transitions in the path, thereby improving the stability and reliability of navigation. The expression for the heuristic cost is as follows: (28); in, For neighboring nodes The heuristic cost, , Neighboring nodes With the target node The Euclidean distance and the weighted distance considering risk distribution between them. , These are the balance weights for Euclidean distance and weighted distance, respectively; The expression for the weighted distance between neighboring nodes and the target node, taking into account risk distribution, is as follows: (29); in, The path is defined as the movement path. In actual calculations, the weighted distance is approximated by discretely sampling and summing the straight path. To eliminate the influence of the difference in numerical magnitude between the weighted distance and the Euclidean distance, a normalization method is used to map the weighted distance to the same numerical range as the Euclidean distance. The balancing weights of the Euclidean distance and the weighted distance are generally set to 0.7 and 0.3, respectively, indicating that the heuristic estimation places more emphasis on geometric distance while moderately considering risk distribution, thus ensuring the search efficiency of the algorithm. The number of sampling points for discrete sampling is determined based on the distance from neighboring nodes to the target node and the sampling interval. The sampling interval is set to 0.5m, which is consistent with the resolution of the spatial grid division, ensuring the accuracy of the risk integral calculation. Heuristic cost estimation combines traditional Euclidean distance and risk distribution information, ensuring a reasonable estimate of the target distance and guiding the search process to avoid high-risk areas, thus improving the efficiency and safety of path search. Heuristic cost estimation satisfies the acceptability condition, ensuring that the algorithm can find the optimal solution.

[0030] S5007. If a neighboring node is not in the open list, add it to the open list and record its parent node as the current node; if a neighboring node is already in the open list and the newly calculated temporary value is less than its original actual cost, update the actual cost of the neighboring node to the temporary value, and update the parent node at the same time. 5008. Return to step 5002 and continue searching until the target node is found or the open list is empty.

[0031] In this embodiment of the application, in order to adapt to dynamic environmental changes, step 501 is also included in step S500, which is: implementing a rolling time-domain optimization strategy, replanning the local path every 5 seconds based on the latest sensor data and prediction results, so as to realize dynamic adjustment of the trajectory; The input to the dynamic temporal optimization strategy is the current robot position, current velocity, global path node sequence, latest sensor data, and the prediction result output from step S300. The output is the local trajectory within the future time window. The rolling temporal optimization strategy extracts a forward path segment from the current position on the global path as a reference trajectory within each planning cycle. The length of the forward path segment corresponds to the maximum distance the robot can reach within the future time window. Based on the latest environmental information and the predicted physical field distribution, a local optimization problem is constructed around the reference trajectory. The objective function of the dynamic temporal optimization is: (30); in, The objective function value, The length of the future time window. For reference trajectory, To control the input, , , These are the weight coefficients for the trajectory control term, risk control term, and input control term, respectively. To eliminate the dimensional differences between different terms in the objective function, a normalization algorithm is needed to make the ranges of the trajectory control term and input control term consistent with the range of the risk control term. After solving the local optimization problem and obtaining the locally optimal trajectory, the control instructions for the first planning cycle are executed only, and the above process is repeated in the next cycle to achieve rolling optimization. Here, the control input is the acceleration vector. The time integral in the objective function is also converted into a finite sum form for solution using a discretization method. Since the control input is an acceleration vector, the relationship between the control input and the trajectory point sequence can be constructed based on the kinematic equations that include the relationship between distance, time, velocity, and acceleration. Then, the discrete optimization problem is solved using a sequential quadratic programming algorithm, iteratively updating the control input until the objective function converges. In the preferred scheme, the trajectory control term is normalized based on the reference distance, and the control input is normalized based on the acceleration reference value, which can directly use the acceleration threshold mentioned earlier. The future time window length is set to 10 seconds. At this time, the time window length can provide sufficient look-ahead distance to deal with dynamic obstacles and maintain the real-time performance of the calculation. The weight coefficients of the trajectory control term, risk control term, and input control term are set to 1.0, 0.5, and 0.2, respectively. This set of weights prioritizes ensuring trajectory tracking accuracy, then considers risk avoidance, and finally takes into account control energy consumption.

[0032] Step S500 also includes step S502 to address the issue of detecting sudden magnetic field anomalies or obstacles. The detailed process of step S502 is as follows: Establish an emergency obstacle avoidance mechanism to quickly generate a temporary avoidance path when a sudden magnetic field anomaly or obstacle is detected; Emergency obstacle avoidance mechanisms include: anomaly detection and obstacle avoidance path generation; The anomaly detection section determines whether to trigger emergency obstacle avoidance by real-time monitoring of the magnetic field strength change rate and sonar echo distance. When the magnetic field strength change rate is greater than the emergency threshold of 100μT / s or the sonar echo distance is less than the emergency safety distance of 2m, it is determined to be a sudden anomaly and emergency obstacle avoidance is immediately activated. The obstacle avoidance path generation section uses an artificial potential field method to quickly plan a temporary path, setting the anomaly source as a repulsive potential field and the target point as a gravitational potential field; the avoidance direction is determined by calculating the negative gradient of the synthetic potential field, and the robot moves along this direction at an emergency speed; once the robot leaves the anomaly's influence range, the rolling time-domain optimization in step 501 is re-executed to return to the global path; the emergency speed is generally taken as 0.5 times the speed threshold; the anomaly's influence range is the area centered on the anomaly source with a radius of 1.5 times the emergency safety distance; The expression for the repulsive potential field is as follows: (31); in, Current location The repulsive potential energy, This is the repulsion gain coefficient, typically taken as 50. This represents the distance between the current location and the anomaly source. For emergency safety distance; The expression for the gravitational potential field is as follows: (32); in, The gravitational potential energy at the current position; This is the gravitational gain coefficient, typically taken as 5; The expression for the negative gradient of the composite potential is as follows: (33); in, The negative gradient of the synthesized potential field; This implementation method, by employing a cross-medium propagation coupling model based on multi-physics equations, accurately describes the propagation characteristics of electromagnetic and acoustic fields in different media and their interactions at boundaries. This overcomes the problem of traditional single-physics models failing when media changes abruptly, achieving accurate modeling of the physical field distribution in complex multi-medium environments. Simultaneously, by introducing a physical information neural network, it combines physical equation constraints with data-driven learning, improving the accuracy and physical consistency of physical field evolution trend prediction. Furthermore, the integrated safety map generation mechanism quantifies and integrates multiple risk factors, providing comprehensive environmental safety information for path planning. Therefore, this implementation method solves the technical problem of robots struggling to achieve accurate perception and safe navigation in complex corridor environments containing multiple media interfaces and strong electromagnetic interference sources, improving the reliability and safety of autonomous navigation.

[0033] This section provides further details regarding an inspection task of an underground cooling water circulation corridor at a nuclear power plant. The corridor is approximately 150 meters long and includes an air section (0-50 meters), an air-water transition section (50-60 meters), and an underwater section (60-150 meters, with a water depth of approximately 8 meters). A high-power circulation pump (located at 70 meters) and a transformer (located at 90 meters) are deployed within the corridor, generating localized strong magnetic field interference. The inspection robot needs to navigate from the corridor entrance (origin) to the inspection point at the end of the corridor. (150, 0, -8), in meters.

[0034] The distributed sensor network was deployed along the corridor, and some of the raw data collected at 14:00:00 on October 29, 2024 are shown in Table 1. Table 1. Raw sensor data (at some points in time)

[0035] After preprocessing in step S100, the extracted trans-medium boundary features are shown in Table 2. Table 2. Results of cross-medium boundary feature extraction

[0036] Among them, the conductivity gradient change rate at position (52,0,-0.5) is 12.8 S / m / m, which exceeds the set threshold θc=10 S / m / m, and is successfully identified as an air-water interface; the magnetic field strength at positions (70,0,-5) and (90,0,-7) deviates from the mean of the surrounding area by more than 3 times the standard deviation, and is marked as the boundary of the strong magnetic field region.

[0037] A cross-medium propagation coupling model based on Maxwell's equations and the acoustic wave equations was used to calculate the physical field distribution parameters of the corridor space. Historical data from the past 24 hours was input into a physical information neural network to predict the physical field distribution for the next hour (15:00:00). The neural network outputs the triaxial magnetic field strength components and sound pressure at each spatial sampling point. Prediction results for some spatial sampling points are shown in Table 3. Table 3. Prediction results of the physical field (partial spatial points for the next hour).

[0038] The physical information neural network was trained using 720 historical samples. After 158 iterations, the composite loss function value of the validation set converged to 0.0032, with the data fitting loss at 0.0021 and the physical equation residual loss at 0.0011. Table 3 shows that the prediction uncertainty is high in the strong magnetic field regions (70 m and 90 m), at 8.7 μT and 12.3 μT respectively, reflecting the complex dynamic characteristics of the physical field at these locations.

[0039] Based on the above prediction results, a risk assessment algorithm was executed; for each spatial grid point, the magnetic field hazard level and acoustic positioning reliability were calculated; after normalization and weighted fusion, a comprehensive risk value was generated; the risk assessment results for some spatial points are shown in Table 4. Table 4 Comprehensive Risk Assessment Results

[0040] As can be seen from Table 4, the comprehensive risk values ​​of spatial points located near the circulating pump (70 meters) and transformer (90 meters) reach 0.754 and 0.872 respectively, and are marked as high-risk areas that need to be avoided in route planning.

[0041] The A* algorithm was improved by integrating the safety map, current position (origin), target position (150,0,-8), and kinematic constraints as inputs. The algorithm underwent 2847 node expansions, with a search time of 1.23 seconds, and generated the optimal path node sequence as shown in Table 5. Table 5 Optimal Navigation Path Node Sequence

[0042] As shown in Table 5, when the planned path passes through the strong magnetic field area (around 70 meters and 90 meters), it actively avoids the high-risk center area by increasing the y-direction offset (up to 4.0 meters), keeping the local risk value below 0.612, significantly lower than the risk value of 0.872 when the straight path passes through this area. The total path length is 154.5 meters, only 4 meters (2.7%) longer than the straight distance (150.5 meters), but the overall risk score decreases from 72.5 for the straight path to 52.3, a risk reduction of 27.9%.

[0043] After smoothing with Bézier curves, the minimum radius of curvature of the generated continuous trajectory is 1.2 meters, which meets the minimum turning radius constraint. The speed planning adopts a trapezoidal curve, with the initial and final speeds being 0 and the maximum speed in the middle section reaching 1.5 m / s. The estimated execution time of the entire task is 113.8 seconds. In actual execution, the rolling time-domain optimization strategy replanned the local trajectory every 5 seconds. When the robot reached the 76-meter mark, it detected a magnetic field strength change rate of 125 μT / s, exceeding the emergency obstacle avoidance threshold. It immediately activated the artificial potential field method to generate a temporary avoidance path, shifting towards the positive y-direction to the position (76, 5.2, -6), successfully avoiding the sudden magnetic field anomaly. The robot then smoothly returned to the original planned path to continue its task. Ultimately, the robot safely reached the target position in 118.5 seconds, a deviation of 4.7 seconds (4.1%) from the expected time, with a navigation accuracy error of 0.32 meters.

[0044] Example 2 To further illustrate with reference to Example 1, the low-feature corridor autonomous navigation system based on multi-sensor fusion includes: The distributed sensor network module, including a magnetic field sensor array, a multi-beam sonar system, and a conductivity sensor, is used for multi-source data acquisition and cross-medium boundary feature extraction. The physical field modeling and prediction module includes a cross-medium propagation coupling model construction unit and a physical information neural network prediction unit. It is used to build an artificial potential field model, calculate the global potential energy distribution, and predict the dynamic trend of the potential field as the environment and obstacle positions change, providing field basis for robot trajectory planning and obstacle avoidance control. The risk assessment module is used to calculate the magnetic field hazard level and acoustic positioning reliability, and generate a comprehensive safety map; The path planning module is used to execute multi-objective optimization algorithms to generate safe navigation trajectories; The execution control module controls the robot's movement according to the navigation trajectory and implements rolling time-domain optimization and emergency obstacle avoidance.

[0045] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be defined as the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method and system for autonomous navigation in low-feature corridors based on multi-sensor fusion, characterized in that, The method includes the following steps: S100: Acquire synchronous multi-source data from a distributed sensor network and extract cross-medium boundary features; S200. Construct a cross-medium propagation coupling model based on multiphysics equations and calculate the physical field distribution parameters; S300: Input historical data and boundary features into the physical information neural network to predict the evolution trend of the physical field; S400 executes a risk assessment algorithm to generate a comprehensive security map containing multi-dimensional risk indicators; S500 generates safe navigation trajectories that take into account multiple constraints based on a multi-objective optimization algorithm.

2. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 1, characterized in that, In step S100, the distributed sensor network includes: a distributed magnetic field sensor array, a multi-beam sonar system, and a conductivity sensor; Real-time data acquired by distributed sensor networks needs to undergo preprocessing operations, including: removing magnetic field measurement noise, separating direct waves and reverberation components in sonar echoes, and smoothing conductivity measurements. The process of extracting cross-medium boundary features is as follows: preprocessing and aligning multi-source data in time; identifying the spatial location of the air-water interface based on the relationship between the conductivity gradient change rate and a set threshold; marking the boundary range of strong magnetic field regions based on outlier detection of the spatial distribution of magnetic field strength; and extracting acoustic propagation characteristic parameters of the underwater environment based on the time delay and intensity characteristics of acoustic echo signals. Acoustic propagation characteristic parameters include: echo delay, echo intensity, spectral center frequency, and spectral bandwidth.

3. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 2, characterized in that, The detailed process of step S200 is as follows: Based on Maxwell's equations, establish a propagation model of the electromagnetic field in different media; based on the sound wave equations, establish a propagation model of the sound field in the underwater environment; couple the electromagnetic field model and the sound field model at the boundary of the medium to establish a set of boundary condition equations; Solving the coupled equations yields numerical solutions for the electromagnetic field intensity distribution and sound pressure field distribution in different medium regions; The boundary condition equations include: electromagnetic field boundary conditions and acoustic field boundary conditions; Electromagnetic field boundary conditions include: the continuity condition of the tangential component of the electric field and the continuity condition of the tangential component of the magnetic field; The boundary conditions of the sound field include: the sound pressure continuity condition and the normal particle vibration velocity continuity condition; The distribution of physical fields includes: electromagnetic field intensity distribution and sound pressure field distribution.

4. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 3, characterized in that, Before step S200 is executed, the following preprocessing steps are also included: the corridor space is meshed using the finite element method, and an adaptive mesh refinement strategy is adopted near the medium boundary; The adaptive mesh refinement strategy determines whether the mesh needs to be refined based on the rate of change of the physical field gradient between adjacent cells. When the rate of change of the physical field gradient is greater than the refinement threshold, the cell is subdivided into smaller sub-cells.

5. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 4, characterized in that, The encoder module of the physical information neural network uses a Long Short-Term Memory (LSTM) network to extract temporal features of magnetic field strength time series, acoustic feature vectors, and boundary position information. Historical information is selectively retained through gating mechanisms such as forget gate, input gate, and output gate. The output of the LSTM is mapped to the intermediate feature space through a fully connected layer. The physical constraint module receives the intermediate feature vector output by the encoder and combines it with the physical equation constraints in step S200. This can constrain the model fitting trend and improve the rationality and solution accuracy of the field quantity calculation results. The decoder module consists of a multi-layer fully connected neural network, which outputs predicted values ​​of magnetic field strength and sound pressure distribution at different spatial locations over a future period of time. The predicted values ​​output by the decoder are denormalized to obtain the magnetic field strength and sound pressure distribution. The optimization strategy for the physical information neural network is to use the Adam optimizer to minimize the composite loss function, which is expressed as follows: （1）； in, The loss value is the result of the composite loss function. , These are the data fitting loss and the physical equation residual loss, respectively. This is the balance coefficient; The data fitting loss is defined as the mean square error between the predicted and the true values; the physical equation residual loss is defined as the sum of squared residuals of the network output on the physical equation.

6. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 5, characterized in that, The detailed process of step S400 is as follows: Based on the physical field prediction results, the magnetic field hazard level and acoustic unreliability of each spatial grid point are calculated. The magnetic field hazard level and acoustic unreliability are mapped to the interval 0~1 using the mean normalization method based on the range, and then a comprehensive risk value is generated by weighted fusion. The calculated comprehensive risk value is mapped onto the corridor spatial coordinate system to generate a three-dimensional comprehensive safety map.

7. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 6, characterized in that, In step S400, the expression for the magnetic field hazard level is as follows: （2）； in, The magnetic field hazard level is [not specified]. This is the predicted value of the magnetic field strength. As a safety threshold, It is a very small positive number; The expression for acoustic unreliability is as follows: （3）； in, Acoustic unreliability The reverberant power spectral density; The expression for the overall risk value is as follows: （4）； in, For the comprehensive risk value, For normalized , For normalized , , These are the magnetic field risk coefficient and the acoustic unreliability coefficient, respectively, and their sum is 1.

8. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 7, characterized in that, The method for calculating the reverberation power spectral density is as follows: First, perform a short-time Fourier transform on the reverberation components separated in step S100 to obtain the time spectrum. Calculate the square of the time spectrum modulus to obtain the power spectrum. Finally, integrate the power spectrum in the frequency domain to obtain the reverberation power spectral density, as expressed below: （5）； in, For the sonar operating frequency, , These are the lower and upper limits of the sonar operating frequency, respectively. This is the power spectrum.

9. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to claim 7, characterized in that, The detailed process of step S500 is as follows: The current depth value measured by the pressure sensor, the flow velocity vector measured by the water flow sensor, and the current position and attitude information provided by the inertial measurement unit are obtained; these real-time data, together with the integrated safety map generated in step S400, are input into the multi-objective path optimization algorithm. The multi-objective path optimization algorithm simultaneously optimizes three objective functions: path length minimization, comprehensive risk minimization, and cross-medium number minimization. It uses an improved A* algorithm to search for a set of feasible paths that satisfy multiple constraints in a discretized spatial grid, and then generates a smooth navigation trajectory through Bézier curve fitting. The navigation trajectory includes position sequence, velocity planning, and expected arrival time. The improved A* algorithm takes the starting position, target position, integrated safety map, and a set of constraints as input, and outputs the optimal path node sequence that satisfies multiple constraints. The improved A* algorithm includes the following steps: S5001. Initialize the open list and the closed list, add the starting node to the open list, and set the cost value of the starting node to 0; S5002. Select the node with the smallest comprehensive evaluation function value from the open list as the current node; the expression of the comprehensive evaluation function is as follows: （6）； in, For any node, This is the comprehensive evaluation function value for that node. The actual cost from the starting point to this node. For a heuristic estimate of the cost from this node to the target point, This refers to the risk penalty item at this node; S5003. Determine if the current node is the target node. If it is, backtrack the parent node chain from the target node to generate a complete path and terminate the search. Otherwise, move the current node from the open list to the closed list. The parent node is the previous node in the path that leads to the current node. S5004. Expand all neighboring nodes of the current node. For each neighboring node, calculate the temporary value of the distance from the starting point to that neighboring node via the current node. Neighbor nodes are defined as nodes that are adjacent to the current node in the spatial grid, and the distance difference between each neighbor node and the current node in the x, y, and z directions does not exceed one grid spacing; the temporary cost value is the candidate value of the node's actual cost, used for path cost prediction and comparison; after filtering the optimal candidate value, it is assigned and updated to the node's actual cost value; The expression for the temporary cost value is as follows: （7）； in, For the current node The temporary value, For the current node any of its neighboring nodes, Let $\frac{ ... （8）； in, For the Euclidean norm, This is the depth penalty coefficient. For reference depth, The depth difference between the two nodes; S5005. Perform constraint checks on each neighboring node to determine whether the node satisfies the magnetic field strength constraint, acoustic positioning error constraint, and kinematic constraint. If it does not meet the constraints, skip the neighboring node. The magnetic field strength constraint is specifically: the magnetic field strength of neighboring nodes is not greater than the magnetic field strength constraint threshold. The acoustic positioning error constraint is specifically: the acoustic positioning error of neighboring nodes is not greater than the acoustic positioning error constraint threshold. The kinematic constraints are as follows: the Euclidean norm of the moving speed is not greater than the speed threshold; the Euclidean norm of the acceleration is not greater than the acceleration threshold; and the turning radius is not less than the minimum turning radius threshold. S5006. For neighboring nodes that satisfy the constraints, calculate the risk penalty term and heuristic cost; The expression for the risk penalty item is as follows: （9）； in, Neighboring nodes Risk penalty items, , These are the risk weighting coefficient and the cross-media penalty coefficient, respectively. Neighboring nodes The overall risk value at the location, This is a cross-medium indicator function, which takes a value of 1 when the two nodes are located in different media, and 0 otherwise. The expression for the heuristic cost is as follows: （10）； in, Neighboring nodes The heuristic cost, , Neighboring nodes With the target node The Euclidean distance and the weighted distance considering risk distribution between them. , These are the balance weights for Euclidean distance and weighted distance, respectively; The expression for the weighted distance between neighboring nodes and the target node, taking into account risk distribution, is as follows: （11）； in, For the movement path; S5007. If a neighboring node is not in the open list, add it to the open list and record its parent node as the current node; if a neighboring node is already in the open list and the newly calculated temporary value is less than its original actual cost, update the actual cost of the neighboring node to the temporary value, and update the parent node at the same time. S5008. Return to step S5002 and continue searching until the target node is found or the open list is empty; In addition to step S500, step 501 is also included, which is: implementing a rolling time-domain optimization strategy, replanning the local path at fixed intervals based on the latest sensor data and prediction results, and realizing dynamic adjustment of the trajectory; The input to the dynamic temporal optimization strategy is the current robot position, current velocity, global path node sequence, latest sensor data, and the prediction result output from step S300. The output is the local trajectory within the future time window. The rolling temporal optimization strategy extracts the forward path segment starting from the current position on the global path as a reference trajectory within each planning cycle. Based on the latest environmental information and the predicted physical field distribution, a local optimization problem is constructed around the reference trajectory. The objective function of the dynamic temporal optimization is: （12）； in, The objective function value, The length of the future time window. For reference trajectory, To control the input, , , These are the weighting coefficients for the trajectory control term, risk control term, and input control term, respectively. Step S500 also includes step S502, the details of which are as follows: Establish an emergency obstacle avoidance mechanism to quickly generate a temporary avoidance path when a sudden magnetic field anomaly or obstacle is detected; Emergency obstacle avoidance mechanisms include: anomaly detection and obstacle avoidance path generation; The anomaly detection section determines whether to trigger emergency obstacle avoidance by monitoring the rate of change of magnetic field strength and the sonar echo distance in real time. When the rate of change of magnetic field strength is greater than the emergency threshold of the transformation rate or the sonar echo distance is less than the emergency safe distance, it is determined to be a sudden anomaly and emergency obstacle avoidance is immediately initiated.

10. The low-feature corridor autonomous navigation method and system based on multi-sensor fusion according to any one of claims 1 to 9, characterized in that, The system includes: The distributed sensor network module, including a magnetic field sensor array, a multi-beam sonar system, and a conductivity sensor, is used for multi-source data acquisition and cross-medium boundary feature extraction. The physical field modeling and prediction module includes a cross-medium propagation coupling model construction unit and a physical information neural network prediction unit. It is used to build an artificial potential field model, calculate the global potential energy distribution, and predict the dynamic trend of the potential field as the environment and obstacle positions change, providing field basis for robot trajectory planning and obstacle avoidance control. The risk assessment module is used to calculate the magnetic field hazard level and acoustic positioning reliability, and generate a comprehensive safety map; The path planning module is used to execute multi-objective optimization algorithms to generate safe navigation trajectories; The execution control module controls the robot's movement according to the navigation trajectory and implements rolling time-domain optimization and emergency obstacle avoidance.

Images