Intelligent control underwater screed signal data calibration method

Abstract

This invention discloses a method for calibrating signal data of an intelligent underwater leveling machine, comprising: acquiring a raw signal dataset and constructing a multi-dimensional signal feature matrix; enhancing signal quality through a deep convolutional generative adversarial network (GAN) to output an initial clean signal set; inputting the initial clean signal set into a multimodal graph attention transform network (MPTN) to model the spatiotemporal dependence and physical coupling constraints between signals through a heterogeneous graph structure to obtain a fused calibration signal; establishing a physical information neural network model, embedding the acoustic wave propagation equation, water flow disturbance equation, and structural dynamics equation as physical constraints into the loss function for physical consistency correction; constructing an adaptive Kalman filter framework to estimate the noise covariance matrix online and dynamically adjust the filter gain to filter out residual noise and output the final calibration signal. This invention achieves high-precision calibration of underwater multi-source signals through the synergistic processing of generative adversarial enhancement, graph attention fusion, physical information constraints, and adaptive filtering.

Description

Technical Field

[0001] This invention relates to the field of signal processing and control technology for underwater leveling machines, and more specifically, to a method for calibrating signal data of an intelligent control underwater leveling machine. Background Technology

[0002] Subsea foundation laying is a crucial step in marine engineering construction, and its flatness directly affects the safety and stability of the superstructure. As the core equipment for achieving precise flatness of the seabed working surface, the accuracy of the sensor signals during operation of the underwater leveling machine is a fundamental prerequisite for achieving automated leveling control.

[0003] Underwater leveling machines are typically deployed in complex and variable marine environments, and their signal acquisition process faces multiple interference factors: First, the non-uniformity of the seawater medium, the scattering effect of suspended sediment, and multipath propagation phenomena lead to severe reverberation and distortion in sonar echo signals; second, underwater currents, eddies, and other hydrodynamic disturbances can couple and interfere with attitude sensors and current meters, introducing non-stationary noise; third, the vibration of the leveling machine's own hydraulic system and the impact excitation generated by the contact between the outriggers and the seabed result in vibration and pressure signals containing a large amount of structural noise. The superposition of these interference factors results in low signal-to-noise ratio and severe distortion in the raw signals acquired by the sensors, making them unsuitable for direct use in subsequent leveling control decisions.

[0004] Existing underwater signal processing methods mostly employ traditional filtering techniques, such as bandpass filtering, median filtering, or fixed-parameter Kalman filtering. While these methods can suppress noise to some extent, they still have the following shortcomings: First, traditional filters are difficult to adapt to the time-varying characteristics of the underwater environment, and the filtering parameters require manual experience for adjustment, resulting in poor universality. Second, a single filtering method cannot simultaneously handle the coupling distortion problem between multiple heterogeneous signals, such as the impact of attitude error on sonar topography measurements. Third, existing methods lack effective utilization of the physical laws of underwater signal propagation, making it difficult to achieve accurate correction of signal distortion at the physical mechanism level. Summary of the Invention

[0005] The purpose of this invention is to provide a method for calibrating signal data of an intelligent control underwater leveling machine, in order to solve the above-mentioned problems existing in the prior art.

[0006] The application is as follows: A method for calibrating signal data of an intelligent control underwater leveling machine includes: A heterogeneous sensor cluster deployed on multiple actuators and the main structure of the underwater leveling machine is used to collect raw signal datasets in real time. The raw signal datasets include attitude signals, pressure signals, vibration signals, sonar echo signals, and hydrodynamic environment signals. The raw signal datasets are preprocessed, and a multidimensional signal feature matrix is ​​constructed based on the preprocessed data. A signal quality enhancement model based on a deep convolutional generative adversarial network is constructed. The multidimensional signal feature matrix is ​​input into the signal quality enhancement model. The generator network learns the distribution characteristics of high signal-to-noise ratio signals to generate enhanced signal feature maps. The discriminator network performs adversarial training on the enhanced signal feature maps and real high-quality signals to output a calibrated initial clean signal set. The initial pure signal set is input into a multimodal graph attention transformation network. The multimodal graph attention transformation network models the spatiotemporal dependencies and physical coupling constraints between different types of signals through a heterogeneous graph structure, calculates the attention weight of each signal node, and performs nonlinear weighted fusion of the signals based on the attention weight to obtain a fused calibration signal. A physical information neural network model for underwater signal propagation is established. The physical information neural network model uses the acoustic wave propagation equation, the water flow disturbance equation, and the structural dynamics equation as physical constraints embedded in the network loss function to perform physical consistency correction on the fused calibration signal, thereby obtaining a physically consistent signal. An adaptive Kalman filter framework is constructed, with the physical consistency signal as the input observation. By estimating the system noise covariance matrix and the measurement noise covariance matrix online, the filter gain is dynamically adjusted to filter out residual Gaussian noise and random impulse noise in the signal, and the final calibration signal is output. The calibration signal is used for the automatic leveling control of the underwater leveling machine.

[0007] Further, the original signal dataset is preprocessed, and a multidimensional signal feature matrix is ​​constructed based on the preprocessed data, including: Perform timestamp alignment on each type of signal in the original signal dataset, and use cubic spline interpolation to synchronize signals with different sampling frequencies to a unified time reference. The aligned signal is initially denoised using an adaptive wavelet thresholding algorithm. The adaptive wavelet thresholding algorithm determines the optimal threshold for each decomposition layer based on the Stein unbiased risk estimation criterion and performs soft thresholding on the high-frequency detail coefficients. The denoised multi-channel signals are truncated by sliding window according to the time dimension. The signal data in each time window constitutes a signal sample. The signal samples are arranged into an initial matrix according to the sensor type. The initial matrix is ​​standardized to have a mean of 0 and a variance of 1. The position and type codes of each sensor are then concatenated along the feature dimension to form the multidimensional signal feature matrix. The threshold function of the adaptive wavelet threshold denoising algorithm is defined as follows:

[0008] Where, λj1 σ is the optimal threshold for the j1-th decomposition layer. j1 This is the estimated standard deviation of the noise in this layer, where N is the signal length and SNR is the signal strength. j1 The signal-to-noise ratio of this layer is given by α and β, which are adaptive adjustment parameters; the soft thresholding function is defined as follows: , Where, ω j1,k1 For the j1-th wavelet coefficient, γ is the coefficient after thresholding, and γ is the smoothing factor used to suppress signal abrupt changes at the threshold point.

[0009] Furthermore, a signal quality enhancement model based on a deep convolutional generative adversarial network is constructed, and the multidimensional signal feature matrix is ​​input into the signal quality enhancement model, including: A generator network is constructed, which adopts an encoder-decoder symmetric structure. The encoder extracts low-dimensional latent features of the signal through multiple downsampling convolutional layers, and the decoder reconstructs the latent features into an enhanced signal feature map through upsampling layers and transposed convolutional layers. Skip connections are introduced between corresponding layers of the encoder and decoder to preserve signal detail information. A discriminator network is constructed, which adopts a block-based discrimination structure. The input signal is divided into multiple overlapping image blocks, and the authenticity of each image block is judged separately before the overall discrimination result is output. The loss function of the adversarial training is composed of a weighted sum of adversarial loss, cycle consistency loss and feature matching loss. The adversarial loss is used to drive the generator network to produce realistic signals, the cycle consistency loss is used to ensure the reversibility of signal transformation, and the feature matching loss is used to stabilize the training process of adversarial training. After adversarial training is completed, the parameters of the generator network are frozen, the multidimensional signal feature matrix is ​​input into the generator network, the enhanced signal feature map is output, and the enhanced signal feature maps of multiple channels are fused by average pooling to obtain the initial clean signal set. The feature extraction process of the generator network is represented as follows: z=Ed(X)=σ(W4 σ(W3 σ(W2 σ(W1 X+b1)+b2)+b3)+b4), Where X is the input multidimensional signal feature matrix, and Ed is the encoder function. The convolution operation is represented by W1, W2, W3, and W4, which are the convolution kernel weight matrices, b1, b2, b3, and b4, which are the bias terms, σ(•) is the Leaky ReLU activation function, and z represents the extracted feature vector.

[0010] Furthermore, the initial pure signal set is input into a multimodal graph attention transform network, which models the spatiotemporal dependencies and physical coupling constraints between different types of signals through a heterogeneous graph structure, including: Each sensor node is treated as a vertex of the heterogeneous graph. The signal correlation between different sensors at the same time is defined as a spatial edge, and the signal correlation between the same sensor at different times is defined as a temporal edge. Physical constraint edges are defined according to the mechanical structure connection relationship of the underwater leveling machine to construct the heterogeneous graph structure. Spatial attention coefficients are obtained by calculating the attention coefficients between nodes on the spatial edge through the graph attention layer. The spatial attention coefficients are calculated by a single-layer feedforward network based on the spliced ​​vector of node features and normalized by the Softmax function. Temporal dependency weights are obtained by calculating long-term dependency weights on temporal edges through a temporal convolutional attention network. The temporal convolutional attention network uses dilated convolution to capture multi-scale temporal features and introduces a channel attention mechanism to recalibrate feature channels. The spatial attention coefficient, temporal dependence weight, and preset coupling strength of physical constraint edges are fused to obtain a comprehensive propagation weight. Based on the comprehensive propagation weight, the features of neighboring nodes are weighted and aggregated to update the feature representation of each node. After multi-layer graph attention transformation, the fused calibration signal is output. The formula for calculating the spatial attention coefficient from node i to node j in the graph attention layer is as follows: , , Among them, h i Let h be the feature vector of node i. j Let W be the feature vector of node j. s To share the linear transformation matrix, a is the attention weight vector, a T Let || denote the transpose of the attention weight vector, || denote the vector concatenation operation, N(i) be the neighborhood set of node i, and μ be the balance coefficient. p is a priori coupling function based on spatial location. i p j All are sensor space coordinates, θ ij Let κ be the angle between the sensor points, κ be the attenuation coefficient, and exp(•) represent the exponential function. This represents the reciprocal of the Euclidean distance, and k represents the index variable; The update process of the node features is represented as follows: , in, W represents the updated feature vector of node i. v Let W be the value transformation matrix. r Let σs(•) be the residual connection matrix, and let σs(•) denote the nonlinear activation function. The output of the temporal convolutional attention network is represented as follows: , Where T1 represents the output feature matrix of the temporal convolutional attention network, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k M represents the dimension of the key vector. mask This is a causal mask matrix used to prevent future information leaks.

[0011] Furthermore, a physical information neural network model for underwater signal propagation is established. This model incorporates the acoustic wave propagation equation, water flow disturbance equation, and structural dynamics equation as physical constraints into the network loss function. This loss function is used to perform physical consistency correction on the fused calibration signal, resulting in a physically consistent signal, including... A signal correction module based on a physical information neural network is constructed. The fused calibration signal is used as the input feature of the signal correction module. The physical consistency signal after correction by physical laws is output through the forward propagation process of the physical information neural network. The underwater sound wave propagation equation is embedded as the first physical constraint into the network loss function. The sound wave propagation equation describes the physical relationship between the propagation speed and attenuation characteristics of the sound pressure signal in the underwater medium. By calculating the partial derivatives of the sound pressure signal output by the network with respect to time and space, the residual term of the wave equation is constructed, and the residual term of the wave equation is used as the first physical constraint loss. The flow perturbation equation is embedded as the second physical constraint in the network loss function. The flow perturbation equation describes the coupling relationship between the flow velocity field and the pressure field and the perturbation effect on the signal propagation path. By calculating the partial derivatives of the flow velocity field and pressure field output by the network with respect to time and space, the residual term of the Navier-Stokes equation is constructed. The residual term of the Navier-Stokes equation is used as the second physical constraint loss. The flow perturbation equation includes the Navier-Stokes equation. The structural dynamics equation of the underwater leveling machine's main structure is embedded as the third physical constraint in the network loss function. The structural dynamics equation describes the dynamic relationship between the elastic deformation and vibration modes of the structure under external excitation. By calculating the partial derivatives of the structural displacement output by the network with respect to time and space, the residual terms of the Euler-Bernoulli beam equation are constructed. The residual terms of the Euler-Bernoulli beam equation are used as the third physical constraint loss. The structural dynamics equation includes the Euler-Bernoulli beam equation. The total loss function of the physical information neural network is composed of the weighted sum of the first physical constraint loss, the second physical constraint loss, and the third physical constraint loss.

[0012] Furthermore, an adaptive Kalman filter framework is constructed, using the physical consistency signal as the input observation. By estimating the system noise covariance matrix and the measurement noise covariance matrix online, the filter gain is dynamically adjusted to filter out residual Gaussian noise and random impulse noise in the signal, outputting the final calibration signal, including: A discrete state-space model of the underwater leveling machine's leveling process is established, comprising a state transition equation and an observation equation. A system noise covariance matrix is ​​defined to describe the statistical characteristics of system noise in the state transition equation, and a measurement noise covariance matrix is ​​defined to describe the statistical characteristics of measurement noise in the observation equation. Based on the state estimate of the previous moment, the system state and observation value at the current moment are predicted. The deviation between the physical consistency signal and the predicted observation value is defined as the innovation sequence. The actual covariance matrix of the innovation sequence is calculated within a sliding window, and the theoretical covariance matrix of the innovation sequence is calculated according to Kalman filtering theory. The ratio of the traces of the two is defined as the adaptive adjustment factor. The system noise covariance matrix is ​​dynamically updated based on the value of the adaptive adjustment factor: when the adaptive adjustment factor is greater than 1, the system noise covariance matrix is ​​increased; when the adaptive adjustment factor is less than 1, the system noise covariance matrix is ​​decreased. Simultaneously, a baseline value for the measurement noise covariance matrix is ​​calculated based on the sensor's nominal accuracy and environmental parameters, and the root mean square error of the innovation sequence is calculated within the sliding window; the root mean square error is compared with the baseline root mean square error, and the measurement noise covariance matrix is ​​dynamically corrected accordingly to obtain a time-varying measurement noise covariance matrix; Based on the updated system noise covariance matrix and the time-varying measurement noise covariance matrix, the Kalman filter gain is calculated; the state prediction value is corrected using the Kalman filter gain, and the corrected state is output as a calibration signal to filter out residual noise in the physical consistency signal.

[0013] Furthermore, the deployment scheme of the heterogeneous sensor cluster includes: arranging six-axis inertial measurement units at the four corners of the main frame of the leveling machine to collect attitude signals; integrating pressure-temperature composite sensors inside the cylinder of each hydraulic outrigger to collect pressure and temperature signals; arranging piezoelectric accelerometers at the connection between the outrigger end and the base plate to collect vibration signals; installing multi-beam sonar arrays at equal intervals along the longitudinal direction at the bottom of the leveling machine to collect sonar echo signals; and arranging electromagnetic flowmeters on both sides of the leveling machine to collect hydrodynamic environment signals; all sensors have built-in time synchronization modules and perform global clock calibration through underwater acoustic synchronization signals. The signal acquisition model of the multibeam sonar array is as follows: , Among them, S sonar For the received sonar echo signal, θ and These are the azimuth and elevation angles, respectively; M is the number of sonar elements; A m Let R(t) be the directional gain of the m-th array element, and R(t) be the target reflected signal. m Let n be the propagation delay, h(t) be the underwater acoustic channel impulse response, and n be the propagation delay. clutter For reverberation and ambient noise, t represents the time scale; The attitude calculation of the six-axis inertial measurement unit adopts a quaternion complementary filtering fusion algorithm, and the discrete form of the attitude calculation is represented as follows: , , Where, q c+1 Let q represent the attitude quaternion at time c+1. c Let c be the attitude quaternion at time c. B represents quaternion multiplication. s ω is the sampling period. c b is the angular velocity measurement value of the gyroscope. ω,c Let b be the zero-bias estimate of the gyroscope at time c. ω,c+1 Let q be the zero bias estimate of the gyroscope at time c+1. e K is the quaternion for attitude error measured by the accelerometer and magnetometer. p and K s For the proportional and integral gain of the complementary filter, This is the integral term for angular velocity error.

[0014] Compared with the prior art, the present invention achieves the following beneficial effects: This invention acquires raw signal datasets and constructs a multi-dimensional signal feature matrix; it then enhances signal quality using a deep convolutional generative adversarial network (GAN) to output an initial clean signal set; this initial clean signal set is input into a multimodal graph attention transform network (MPT), which models the spatiotemporal dependencies and physical coupling constraints between signals through a heterogeneous graph structure to obtain a fused calibration signal; a physical information neural network model is established, embedding the acoustic propagation equation, water flow disturbance equation, and structural dynamics equation as physical constraints into the loss function for physical consistency correction; finally, an adaptive Kalman filter framework is constructed to estimate the noise covariance matrix online and dynamically adjust the filter gain to remove residual noise and output the final calibration signal. This invention achieves high-precision calibration of underwater multi-source signals through the coordinated processing of generative adversarial enhancement, graph attention fusion, physical information constraints, and adaptive filtering. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for calibrating signal data of an intelligent underwater leveling machine according to an embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings.

[0017] Example 1

[0018] This invention provides a method for calibrating signal data of an intelligent control underwater leveling machine, such as... Figure 1 As shown, the method includes: S1. A heterogeneous sensor cluster deployed on multiple actuators and the main structure of the underwater leveling machine is used to collect raw signal datasets in real time. The raw signal datasets include attitude signals, pressure signals, vibration signals, sonar echo signals and hydrodynamic environment signals. The raw signal datasets are preprocessed, and a multidimensional signal feature matrix is ​​constructed based on the preprocessed data. S2. Construct a signal quality enhancement model based on a deep convolutional generative adversarial network. Input the multidimensional signal feature matrix into the signal quality enhancement model. Learn the distribution characteristics of high signal-to-noise ratio signals through the generator network to generate an enhanced signal feature map. Perform adversarial training on the enhanced signal feature map and real high-quality signals through the discriminator network to output a calibrated initial clean signal set. S3. Input the initial pure signal set into the multimodal graph attention transformation network. The multimodal graph attention transformation network models the spatiotemporal dependencies and physical coupling constraints between different types of signals through a heterogeneous graph structure, calculates the attention weight of each signal node, and performs nonlinear weighted fusion of the signals based on the attention weight to obtain the fused calibration signal. S4. Establish a physical information neural network model for underwater signal propagation. The physical information neural network model uses the acoustic wave propagation equation, the water flow disturbance equation, and the structural dynamics equation as physical constraints embedded in the network loss function to perform physical consistency correction on the fused calibration signal, thereby obtaining a physically consistent signal. S5. Construct an adaptive Kalman filter framework, take the physical consistency signal as the observation input, estimate the system noise covariance matrix and the measurement noise covariance matrix online, dynamically adjust the filter gain, filter out residual Gaussian noise and random impulse noise in the signal, and output the final calibration signal. The calibration signal is used for the automatic leveling control of the underwater leveling machine.

[0019] Specifically, this embodiment acquires the original signal dataset and constructs a multi-dimensional signal feature matrix; it enhances signal quality using a deep convolutional generative adversarial network (GAN) to output an initial clean signal set; it inputs the initial clean signal set into a multimodal graph attention transform network (MPT), and obtains a fused calibration signal by modeling the spatiotemporal dependencies and physical coupling constraints between signals through a heterogeneous graph structure; it establishes a physical information neural network model, embedding the acoustic propagation equation, water flow disturbance equation, and structural dynamics equation as physical constraints into the loss function for physical consistency correction; and it constructs an adaptive Kalman filter framework to estimate the noise covariance matrix online and dynamically adjust the filter gain to filter out residual noise and output the final calibration signal. This embodiment achieves high-precision calibration of underwater multi-source signals through the coordinated processing of generative adversarial enhancement, graph attention fusion, physical information constraints, and adaptive filtering.

[0020] In the above embodiments, specifically, the original signal dataset is preprocessed, and a multidimensional signal feature matrix is ​​constructed based on the preprocessed data, including: Perform timestamp alignment on each type of signal in the original signal dataset, and use cubic spline interpolation to synchronize signals with different sampling frequencies to a unified time reference. The aligned signal is initially denoised using an adaptive wavelet thresholding algorithm. The adaptive wavelet thresholding algorithm determines the optimal threshold for each decomposition layer based on the Stein unbiased risk estimation criterion and performs soft thresholding on the high-frequency detail coefficients. The denoised multi-channel signals are truncated by sliding window according to the time dimension. The signal data in each time window constitutes a signal sample. The signal samples are arranged into an initial matrix according to the sensor type. The initial matrix is ​​standardized to have a mean of 0 and a variance of 1. The position and type codes of each sensor are then concatenated along the feature dimension to form the multidimensional signal feature matrix. The threshold function of the adaptive wavelet threshold denoising algorithm is defined as follows:

[0021] Where, λ j1 σ is the optimal threshold for the j1-th decomposition layer. j1 This is the estimated standard deviation of the noise in this layer, where N is the signal length and SNR is the signal strength. j1 The signal-to-noise ratio of this layer is given by α and β, which are adaptive adjustment parameters; the soft thresholding function is defined as follows: , Where, ω j1,k1 For the j1-th wavelet coefficient, γ is the coefficient after thresholding, and γ is the smoothing factor used to suppress signal abrupt changes at the threshold point.

[0022] It should be noted that during seabed foundation laying operations in a certain sea area, an underwater leveling machine descended to a depth of 50 meters. A heterogeneous sensor cluster deployed on the leveling machine collected signals at different frequencies: the six-axis inertial measurement unit output attitude signals at 200Hz, the pressure sensor output hydraulic signals at 100Hz, the multibeam sonar output topographic echo signals at 30Hz, and the current meter output hydrodynamic signals at 50Hz. The above steps are explained below: Preprocessing implementation process: Step 1: Timestamp Alignment. Due to the different sampling frequencies of each sensor (200Hz, 100Hz, 30Hz, 50Hz), direct fusion will result in time misalignment. The system first obtains the timestamps of each signal and uses the highest sampling frequency of 200Hz as a unified time reference. The pressure signal, sonar signal, and flow velocity signal are resampled using the cubic spline interpolation method. Specifically, a cubic polynomial curve is inserted between two adjacent original sampling points to make the interpolated signal completely aligned with the attitude signal on the time axis. The interpolation error is controlled within 0.5%. Step 2: Adaptive wavelet threshold denoising. Taking the sonar echo signal as an example, the signal is affected by the scattering of suspended sediment at a water depth of 50 meters, resulting in severe reverberation noise. The system performs 5-layer wavelet decomposition on the sonar signal. The first to third layers are high-frequency detail layers (mainly containing noise), and the fourth to fifth layers are approximation layers (mainly containing effective signals). Using the Stein unbiased risk estimation criterion, the system automatically calculates the optimal threshold for each layer: Layer 1 (highest frequency): Signal-to-noise ratio (SNR) = 3dB, noise standard deviation (σ) = 0.35, and the calculated threshold (λ) = 0.52; Layer 2: Signal-to-noise ratio (SNR) = 5dB, σ = 0.28, threshold λ = 0.41; Layer 3: Signal-to-noise ratio (SNR) = 8dB, σ = 0.19, threshold λ = 0.28; The adaptive adjustment parameters α and β are dynamically adjusted based on the real-time signal-to-noise ratio: when the signal-to-noise ratio of the first layer is low, α automatically increases to 1.8 to enhance the denoising strength; when the signal-to-noise ratio of the third layer increases, α decreases to 1.2 to retain more signal details. When using soft thresholding, the smoothing factor γ is set to 0.1, which makes the coefficients at the threshold point transition smoothly rather than abruptly, avoiding the Gibbs oscillation phenomenon caused by traditional hard thresholding. For example, if the absolute value of a wavelet coefficient is 0.55, the threshold of this layer is 0.52, and the coefficient after processing is about 0.03 instead of being directly set to zero, thus preserving the continuity of weak and effective signals. Step 3: Sliding window capture and matrix construction. The denoised multi-channel signals are captured by sliding window with a window length of 1 second and a sliding step of 0.5 seconds. Within each window: 200 sampling points for attitude signal, 100 sampling points for pressure signal, 30 sampling points for sonar signal, and 50 sampling points for flow velocity signal. The samples are arranged according to sensor type to form an initial matrix with a dimension of 380×1. Step 4: Standardization and Encoding. The initial matrix is ​​Z-score standardized so that the mean of each feature dimension is 0 and the variance is 1. At the same time, a three-dimensional position code (such as the X, Y, and Z coordinates of the sensor on the leveling machine) and a one-dimensional type code (0=attitude, 1=pressure, 2=sonar, 3=flow velocity) are spliced ​​for each sensor node, and finally a multi-dimensional signal feature matrix with a dimension of 380×8 is formed.

[0023] The above steps synchronize signals of different frequencies to a unified time base using cubic spline interpolation. Compared with linear interpolation, cubic spline interpolation has a second-order continuous derivative and a smooth, oscillating interpolation curve, which can more accurately restore the true value of the signal at non-sampling times, providing a time-accurately aligned signal input for subsequent multimodal fusion. By introducing an adaptive threshold determination mechanism based on Stein's unbiased risk estimation, this criterion can automatically calculate the optimal threshold based on the statistical characteristics of the signal itself without needing to know the noise distribution. At the same time, the α and β parameters are dynamically adjusted with the signal-to-noise ratio, achieving an adaptive balance between denoising intensity and signal fidelity: the low signal-to-noise ratio layer enhances denoising, while the high signal-to-noise ratio layer preserves details, significantly improving the algorithm's adaptability to signals under different water depths and bottom conditions.

[0024] In the above embodiments, specifically, a signal quality enhancement model based on a deep convolutional generative adversarial network is constructed, and the multidimensional signal feature matrix is ​​input into the signal quality enhancement model, including: A generator network is constructed, which adopts an encoder-decoder symmetric structure. The encoder extracts low-dimensional latent features of the signal through multiple downsampling convolutional layers, and the decoder reconstructs the latent features into an enhanced signal feature map through upsampling layers and transposed convolutional layers. Skip connections are introduced between corresponding layers of the encoder and decoder to preserve signal detail information. A discriminator network is constructed, which adopts a block-based discrimination structure. The input signal is divided into multiple overlapping image blocks, and the authenticity of each image block is judged separately before the overall discrimination result is output. The loss function of the adversarial training is composed of a weighted sum of adversarial loss, cycle consistency loss and feature matching loss. The adversarial loss is used to drive the generator network to produce realistic signals, the cycle consistency loss is used to ensure the reversibility of signal transformation, and the feature matching loss is used to stabilize the training process of adversarial training. After adversarial training is completed, the parameters of the generator network are frozen, the multidimensional signal feature matrix is ​​input into the generator network, the enhanced signal feature map is output, and the enhanced signal feature maps of multiple channels are fused by average pooling to obtain the initial clean signal set. The feature extraction process of the generator network is represented as follows: z=Ed(X)=σ(W4 σ(W3 σ(W2 σ(W1 X+b1)+b2)+b3)+b4), Where X is the input multidimensional signal feature matrix, and Ed is the encoder function. The convolution operation is represented by W1, W2, W3, and W4, which are the convolution kernel weight matrices, b1, b2, b3, and b4, which are the bias terms, σ(•) is the Leaky ReLU activation function, and z represents the extracted feature vector.

[0025] It should be noted that the underwater leveling machine operates at a water depth of 50 meters. After preprocessing, the multidimensional signal feature matrix obtained has a dimension of 380×8. This signal matrix includes: attitude signals (tilt angle changes, angular velocity), pressure signals (hydraulic outrigger pressure fluctuations), sonar signals (seabed topographic echoes), vibration signals (structural vibration response), and flow velocity signals (hydrodynamic environment). Due to the complexity of the underwater environment, the following problems exist in this signal matrix: the sonar signal is severely contaminated by reverberation noise, and some waveforms are submerged in noise; the pressure signal contains periodic pulsating noise from the hydraulic pump; and the vibration signal is mixed with broadband random noise from water flow impact. Traditional filtering methods are difficult to process the above multiple noises simultaneously and cannot recover the effective signals submerged by noise. Generative Adversarial Network Training and Inference Implementation Process: Step 1: Generator Network Construction and Training. The generator network adopts an encoder-decoder symmetrical U-shaped structure, with the following specific configuration: Encoder part (downsampling path): The first convolutional layer has an input dimension of 380×8, uses 32 3×3 convolutional kernels, has a stride of 2, and outputs a feature map size of 190×32. This layer extracts local temporal features of the signal, such as the peak position of the sonar echo and the rising edge of the pressure pulse. The second convolutional layer uses 64 3×3 convolutional kernels with a stride of 2, and outputs a feature map size of 95×64. This layer extracts intermediate semantic features, such as the correlation pattern between attitude angle and vibration amplitude. The third convolutional layer uses 128 3×3 convolutional kernels with a stride of 2, and outputs a feature map size of 48×128. This layer extracts high-order features, such as the collaborative variation patterns between signals from multiple sensors. The fourth convolutional layer uses 256 3×3 convolutional kernels with a stride of 2, and outputs a feature map of size 24×256, which is further compressed into a latent feature vector z (dimension 24×256). Implementation instructions for the decoder section (upsampling path): Layer 1 upsampling: z is upsampled to 48×256, and then spliced ​​with the output of the encoder layer 3 (48×128) through a skip connection. After transpose convolution, the output is 96×128. Layer 2 upsampling: Upsampled to 95×128, and after being spliced ​​with the encoder's Layer 2 output (95×64), the output is 190×64; Layer 3 upsampling: Upsampled to 190×64, and after being concatenated with the encoder layer 1 output (190×32), the output is 380×32; Layer 4 upsampling: Output the final enhanced signal feature map 380×8, which is consistent with the input dimension; The role of skip connections: Taking the weak terrain echo in the sonar signal as an example, this feature is well preserved in the first layer encoder, but details may be lost after multiple downsampling. Skip connections directly pass the features of the first layer to the third layer decoder, enabling the decoder to fuse high-resolution detail information with high-level semantic information and recover the weak echo signal that has been submerged by noise. Step 2: Discriminator network construction and training. The discriminator network adopts the PatchGAN (block-based discriminator) structure: The input enhanced signal feature map (380×8) is divided into 16 overlapping 96×4 image blocks with a 50% overlap between adjacent blocks. Each image block is independently processed by a 5-layer convolutional network to determine whether it is true or false, and outputs a scalar discrimination result (0 represents false and 1 represents true). Finally, the average of the 16 discrimination results is used as the overall discrimination output. The advantage of block discrimination is that traditional global discriminators tend to ignore local details, while PatchGAN forces the generator not only to generate a realistic signal as a whole, but also to maintain authenticity in each local block, thereby effectively improving the quality of signal detail recovery. For example, after the reverberation noise in a certain local time window of a sonar signal is eliminated, its waveform characteristics should be consistent with the real seabed echo, and the local block should be judged as real by the discriminator. Through the above steps, in the original sonar signal, the reverberation noise amplitude is about 80% of the effective echo amplitude, and the target terrain features are difficult to identify. After enhancement by generative adversarial network, the reverberation noise is suppressed to less than 15% of the effective echo amplitude, and the seabed terrain undulation features are clearly distinguishable. In the original pressure signal, the hydraulic pump pulsation noise amplitude is 25% of the effective outrigger pressure signal, which increases the difficulty of pressure control. After enhancement, the pulsation noise is suppressed to less than 5%, and the dynamic response characteristics of the outrigger pressure are completely preserved.

[0026] In the above embodiments, specifically, the initial pure signal set is input to a multimodal graph attention transform network. The multimodal graph attention transform network models the spatiotemporal dependencies and physical coupling constraints between different types of signals through a heterogeneous graph structure, including: Each sensor node is treated as a vertex of the heterogeneous graph. The signal correlation between different sensors at the same time is defined as a spatial edge, and the signal correlation between the same sensor at different times is defined as a temporal edge. Physical constraint edges are defined according to the mechanical structure connection relationship of the underwater leveling machine to construct the heterogeneous graph structure. Spatial attention coefficients are obtained by calculating the attention coefficients between nodes on the spatial edge through the graph attention layer. The spatial attention coefficients are calculated by a single-layer feedforward network based on the spliced ​​vector of node features and normalized by the Softmax function. Temporal dependency weights are obtained by calculating long-term dependency weights on temporal edges through a temporal convolutional attention network. The temporal convolutional attention network uses dilated convolution to capture multi-scale temporal features and introduces a channel attention mechanism to recalibrate feature channels. The spatial attention coefficient, temporal dependence weight, and preset coupling strength of physical constraint edges are fused to obtain a comprehensive propagation weight. Based on the comprehensive propagation weight, the features of neighboring nodes are weighted and aggregated to update the feature representation of each node. After multi-layer graph attention transformation, the fused calibration signal is output. The formula for calculating the spatial attention coefficient from node i to node j in the graph attention layer is as follows: , , Among them, h i Let h be the feature vector of node i. jLet W be the feature vector of node j. s To share the linear transformation matrix, a is the attention weight vector, a T Let || denote the transpose of the attention weight vector, || denote the vector concatenation operation, N(i) be the neighborhood set of node i, and μ be the balance coefficient. p is a priori coupling function based on spatial location. i p j All are sensor space coordinates, θ ij Let κ be the angle between the sensor points, κ be the attenuation coefficient, and exp(•) represent the exponential function. This represents the reciprocal of the Euclidean distance, and k represents the index variable; The update process of the node features is represented as follows: , in, W represents the updated feature vector of node i. v Let W be the value transformation matrix. r Let σs(•) be the residual connection matrix, and let σs(•) denote the nonlinear activation function. The output of the temporal convolutional attention network is represented as follows: , Where T1 represents the output feature matrix of the temporal convolutional attention network, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k M represents the dimension of the key vector. mask This is a causal mask matrix used to prevent future information leaks.

[0027] In the above embodiments, specifically, a physical information neural network model for underwater signal propagation is established. This model uses the acoustic wave propagation equation, the water flow disturbance equation, and the structural dynamics equation as physical constraints embedded in the network loss function to perform physical consistency correction on the fused calibration signal, resulting in a physically consistent signal, including... A signal correction module based on a physical information neural network is constructed. The fused calibration signal is used as the input feature of the signal correction module. The physical consistency signal after correction by physical laws is output through the forward propagation process of the physical information neural network. The underwater sound wave propagation equation is embedded as the first physical constraint into the network loss function. The sound wave propagation equation describes the physical relationship between the propagation speed and attenuation characteristics of the sound pressure signal in the underwater medium. By calculating the partial derivatives of the sound pressure signal output by the network with respect to time and space, the residual term of the wave equation is constructed, and the residual term of the wave equation is used as the first physical constraint loss. The flow perturbation equation is embedded as the second physical constraint in the network loss function. The flow perturbation equation describes the coupling relationship between the flow velocity field and the pressure field and the perturbation effect on the signal propagation path. By calculating the partial derivatives of the flow velocity field and pressure field output by the network with respect to time and space, the residual term of the Navier-Stokes equation is constructed. The residual term of the Navier-Stokes equation is used as the second physical constraint loss. The flow perturbation equation includes the Navier-Stokes equation. The structural dynamics equation of the underwater leveling machine's main structure is embedded as the third physical constraint in the network loss function. The structural dynamics equation describes the dynamic relationship between the elastic deformation and vibration modes of the structure under external excitation. By calculating the partial derivatives of the structural displacement output by the network with respect to time and space, the residual terms of the Euler-Bernoulli beam equation are constructed. The residual terms of the Euler-Bernoulli beam equation are used as the third physical constraint loss. The structural dynamics equation includes the Euler-Bernoulli beam equation. The total loss function of the physical information neural network is composed of the weighted sum of the first physical constraint loss, the second physical constraint loss, and the third physical constraint loss.

[0028] It should be noted that the first physical constraint formula of the wave equation is: , Where p(x) i ,t i c(x) represents the sound pressure signal output by the network. i ) is the sound speed distribution function, γ(x) i () represents the absorption attenuation coefficient. ▽ 2 For the Laplace operator; The second physical constraint formula for the Navier-Stokes equations is: , Where u is the velocity field, P is the pressure field, ρ is the fluid density, and ν is the kinematic viscosity coefficient. The second physical constraint formula for the Euler-Bernoulli beam equation is: , Where, ω(x) i ,t i ρ is the lateral displacement of the beam, EI is the bending stiffness, and ρ is the bending stiffness. s Where A is the structural density and A is the cross-sectional area. fext is the external excitation load.

[0029] In the above embodiments, specifically, an adaptive Kalman filter framework is constructed, using the physical consistency signal as the input observation. By estimating the system noise covariance matrix and the measurement noise covariance matrix online, the filter gain is dynamically adjusted to filter out residual Gaussian noise and random impulse noise in the signal, outputting the final calibration signal, including: A discrete state-space model of the underwater leveling machine's leveling process is established, comprising a state transition equation and an observation equation. A system noise covariance matrix is ​​defined to describe the statistical characteristics of system noise in the state transition equation, and a measurement noise covariance matrix is ​​defined to describe the statistical characteristics of measurement noise in the observation equation. Based on the state estimate of the previous moment, the system state and observation value at the current moment are predicted. The deviation between the physical consistency signal and the predicted observation value is defined as the innovation sequence. The actual covariance matrix of the innovation sequence is calculated within a sliding window, and the theoretical covariance matrix of the innovation sequence is calculated according to Kalman filtering theory. The ratio of the traces of the two is defined as the adaptive adjustment factor. The system noise covariance matrix is ​​dynamically updated based on the value of the adaptive adjustment factor: when the adaptive adjustment factor is greater than 1, the system noise covariance matrix is ​​increased; when the adaptive adjustment factor is less than 1, the system noise covariance matrix is ​​decreased. Simultaneously, a baseline value for the measurement noise covariance matrix is ​​calculated based on the sensor's nominal accuracy and environmental parameters, and the root mean square error of the innovation sequence is calculated within the sliding window; the root mean square error is compared with the baseline root mean square error, and the measurement noise covariance matrix is ​​dynamically corrected accordingly to obtain a time-varying measurement noise covariance matrix; Based on the updated system noise covariance matrix and the time-varying measurement noise covariance matrix, the Kalman filter gain is calculated; the state prediction value is corrected using the Kalman filter gain, and the corrected state is output as a calibration signal to filter out residual noise in the physical consistency signal.

[0030] In the above embodiments, specifically, the deployment scheme of the heterogeneous sensor cluster includes: arranging six-axis inertial measurement units at the four corners of the main frame of the leveling machine to collect attitude signals; integrating pressure-temperature composite sensors inside the cylinder of each hydraulic outrigger to collect pressure and temperature signals; arranging piezoelectric accelerometers at the connection between the outrigger end and the base plate to collect vibration signals; installing multi-beam sonar arrays at equal intervals along the longitudinal direction at the bottom of the leveling machine to collect sonar echo signals; and arranging electromagnetic flowmeters on both sides of the leveling machine to collect hydrodynamic environment signals; all sensors have built-in time synchronization modules and perform global clock calibration through underwater acoustic synchronization signals. The signal acquisition model of the multibeam sonar array is as follows: , Among them, S sonar For the received sonar echo signal, θ and These are the azimuth and elevation angles, respectively; M is the number of sonar elements; A m Let R(t) be the directional gain of the m-th array element, and R(t) be the target reflected signal. m Let n be the propagation delay, h(t) be the underwater acoustic channel impulse response, and n be the propagation delay. clutter For reverberation and ambient noise, t represents the time scale; The attitude calculation of the six-axis inertial measurement unit adopts a quaternion complementary filtering fusion algorithm, and the discrete form of the attitude calculation is represented as follows: , , Where, q c+1 Let q represent the attitude quaternion at time c+1. c Let c be the attitude quaternion at time c. B represents quaternion multiplication. s ω is the sampling period. c b is the angular velocity measurement value of the gyroscope. ω,c Let b be the zero-bias estimate of the gyroscope at time c. ω,c+1 Let q be the zero bias estimate of the gyroscope at time c+1. e K is the quaternion for attitude error measured by the accelerometer and magnetometer. p and K s For the proportional and integral gain of the complementary filter, This is the integral term for angular velocity error.

[0031] It should be noted that the adaptive adjustment parameters α, β, and smoothing factor γ in the adaptive wavelet thresholding denoising algorithm need to be dynamically adjusted according to the real-time statistical characteristics of the signal. The specific constraints are as follows: The parameter α is inversely proportional to the expected value of the signal-to-noise ratio of the current decomposition layer, and is directly proportional to the relative magnitude of the noise standard deviation of this layer in the entire wavelet decomposition layer. When the signal-to-noise ratio is low, α automatically increases to enhance the denoising strength, and when the signal-to-noise ratio is high, α automatically decreases to retain more signal details. The value of parameter β is determined by the ratio of the signal-to-noise ratio of the highest decomposition layer to the signal-to-noise ratio of the lowest decomposition layer. This ratio is compressed by a logarithmic function and used as the benchmark value of β, ensuring reasonable differences in thresholds between different decomposition layers. The value of the smoothing factor γ is closely related to the maximum amplitude of the wavelet coefficients and the size of the current layer threshold. γ is set to the reciprocal of the product of the two multiplied by a fixed multiple, so that the threshold function can smoothly transition near the threshold point. Furthermore, when the absolute value of the wavelet coefficients is less than the sum of the threshold and the transition band width coefficient, the soft threshold processing function automatically switches to the hard threshold correction mode, and the value range of the transition band width coefficient is limited to between 0.05 and 0.15. This switching mechanism ensures that when the signal amplitude is close to the threshold boundary, it can both preserve the continuity of the signal and avoid Gibbs oscillation, and suppress the introduction of spurious fluctuation components.

[0032] The comprehensive propagation weight is calculated by a multi-dimensional adaptive fusion function using the spatial attention coefficient, the time-dependent weight, and the physical constraint edge coupling strength. The fusion function adopts a square sum normalization strategy, that is, the comprehensive propagation weight is equal to the product of the spatial attention coefficient, the time-dependent weight, and the physical constraint edge coupling strength, divided by the square root of the sum of the squares of the three in the neighborhood. This normalization method ensures the numerical stability of the comprehensive weight. The coupling strength of the physical constraint edge is pre-calculated based on the stiffness matrix and damping matrix of the underwater leveling machine's mechanical structure. Specifically, the coupling strength is determined by the ratio of the product of the coupling stiffness element between the two sensor nodes and their respective stiffness elements in the stiffness matrix, while the exponential decay factor of the corresponding element in the damping matrix is ​​introduced for correction. When the calculated coupling strength is lower than the pre-set weak coupling threshold, the physical constraint edge is identified as a redundant connection and dynamically pruned from the graph structure to reduce the computational complexity of the graph attention network and prevent interference information introduced by weak connections.

[0033] The improved Sage-Husa adaptive filtering algorithm introduces a convergence monitoring mechanism, which automatically triggers a reset operation when a divergence trend is detected. The divergence trend is detected by comprehensively judging two conditions: The first condition is that the actual covariance trace of the new sequence exceeds the set multiple of the theoretical covariance trace. This multiple is set between 2 and 5 times based on engineering experience. When it exceeds this threshold, it indicates that the deviation between the observed value and the predicted value exceeds the normal range. The second condition is that the change magnitude of the state estimation vector relative to the normalized ratio of the state vector at the previous time exceeds a preset jump threshold. This condition is used to capture abrupt changes in the state variables. When either of the above two conditions is met, the filter is determined to be in a divergent state. At this time, the following reset operation is performed: the system noise covariance matrix is ​​reset to its initial value, the measurement noise covariance matrix is ​​reset to the nominal value of the sensor, and the adaptive forgetting factor is also reset to its initial set value. Within a certain number of consecutive filtering steps after the reset operation is completed, the system is forced to adopt an open-loop covariance update strategy, that is, the state prediction covariance is determined only by the covariance of the previous moment and the initial system noise covariance, and no adaptive adjustment term is introduced. After the transition period ends, the complete adaptive covariance update mechanism is reactivated. This reset mechanism ensures the robustness and convergence stability of the filter in harsh underwater environments.

[0034] It should be understood that the above embodiments are one or more embodiments of the present invention, and there are many other embodiments and variations based on the present invention; any variations and modifications made by those skilled in the art through the present invention without making pioneering innovations are all within the protection scope of the present invention.

Claims

1. A method for calibrating signal data of an intelligent control underwater leveling machine, characterized in that, include: A heterogeneous sensor cluster deployed on multiple actuators and the main structure of the underwater leveling machine is used to collect raw signal datasets in real time. The raw signal datasets include attitude signals, pressure signals, vibration signals, sonar echo signals, and hydrodynamic environment signals. The raw signal datasets are preprocessed, and a multidimensional signal feature matrix is ​​constructed based on the preprocessed data. A signal quality enhancement model based on a deep convolutional generative adversarial network is constructed. The multidimensional signal feature matrix is ​​input into the signal quality enhancement model. The generator network learns the distribution characteristics of high signal-to-noise ratio signals to generate enhanced signal feature maps. The discriminator network performs adversarial training on the enhanced signal feature maps and real high-quality signals to output a calibrated initial clean signal set. The initial pure signal set is input into a multimodal graph attention transformation network. The multimodal graph attention transformation network models the spatiotemporal dependencies and physical coupling constraints between different types of signals through a heterogeneous graph structure, calculates the attention weight of each signal node, and performs nonlinear weighted fusion of the signals based on the attention weight to obtain a fused calibration signal. A physical information neural network model for underwater signal propagation is established. The physical information neural network model uses the acoustic wave propagation equation, the water flow disturbance equation, and the structural dynamics equation as physical constraints embedded in the network loss function to perform physical consistency correction on the fused calibration signal, thereby obtaining a physically consistent signal. An adaptive Kalman filter framework is constructed, with the physical consistency signal as the input observation. By estimating the system noise covariance matrix and the measurement noise covariance matrix online, the filter gain is dynamically adjusted to filter out residual Gaussian noise and random impulse noise in the signal, and the final calibration signal is output. The calibration signal is used for the automatic leveling control of the underwater leveling machine.

2. The method for calibrating signal data of an intelligent control underwater leveling machine according to claim 1, characterized in that, The original signal dataset is preprocessed, and a multidimensional signal feature matrix is ​​constructed based on the preprocessed data, including: Perform timestamp alignment on each type of signal in the original signal dataset, and use cubic spline interpolation to synchronize signals with different sampling frequencies to a unified time reference. The aligned signal is initially denoised using an adaptive wavelet thresholding algorithm. The adaptive wavelet thresholding algorithm determines the optimal threshold for each decomposition layer based on the Stein unbiased risk estimation criterion and performs soft thresholding on the high-frequency detail coefficients. The denoised multi-channel signals are truncated by sliding window according to the time dimension. The signal data in each time window constitutes a signal sample. The signal samples are arranged into an initial matrix according to the sensor type. The initial matrix is ​​standardized to have a mean of 0 and a variance of 1. The position and type codes of each sensor are then concatenated along the feature dimension to form the multidimensional signal feature matrix. The threshold function of the adaptive wavelet threshold denoising algorithm is defined as follows: Where, λ j1 σ is the optimal threshold for the j1-th decomposition layer. j1 This is the estimated standard deviation of the noise in this layer, where N is the signal length and SNR is the signal strength. j1 The signal-to-noise ratio of this layer is given by α and β, which are adaptive adjustment parameters; the soft thresholding function is defined as follows: , Where, ω j1,k1 For the j1-th wavelet coefficient, γ is the coefficient after thresholding, and γ is the smoothing factor used to suppress signal abrupt changes at the threshold point.

3. The method for calibrating signal data of an intelligent control underwater leveling machine according to claim 1, characterized in that, Constructing a signal quality enhancement model based on a deep convolutional generative adversarial network, and inputting the multidimensional signal feature matrix into the signal quality enhancement model, including: A generator network is constructed, which adopts an encoder-decoder symmetric structure. The encoder extracts low-dimensional latent features of the signal through multiple downsampling convolutional layers, and the decoder reconstructs the latent features into an enhanced signal feature map through upsampling layers and transposed convolutional layers. Skip connections are introduced between corresponding layers of the encoder and decoder to preserve signal detail information. A discriminator network is constructed, which adopts a block-based discrimination structure. The input signal is divided into multiple overlapping image blocks, and the authenticity of each image block is judged separately before the overall discrimination result is output. The loss function of the adversarial training is composed of a weighted sum of adversarial loss, cycle consistency loss and feature matching loss. The adversarial loss is used to drive the generator network to produce realistic signals, the cycle consistency loss is used to ensure the reversibility of signal transformation, and the feature matching loss is used to stabilize the training process of adversarial training. After adversarial training is completed, the parameters of the generator network are frozen, the multidimensional signal feature matrix is ​​input into the generator network, the enhanced signal feature map is output, and the enhanced signal feature maps of multiple channels are fused by average pooling to obtain the initial clean signal set. The feature extraction process of the generator network is represented as follows: z=Ed(X)=σ(W4 σ(W3 σ(W2 σ(W1 X+b1)+b2)+b3)+b4), Where X is the input multidimensional signal feature matrix, and Ed is the encoder function. The convolution operation is represented by W1, W2, W3, and W4, which are the convolution kernel weight matrices, b1, b2, b3, and b4, which are the bias terms, σ(•) is the Leaky ReLU activation function, and z represents the extracted feature vector.

4. The method for calibrating signal data of an intelligent control underwater leveling machine according to claim 1, characterized in that, The initial pure signal set is input into a multimodal graph attention transform network, which models the spatiotemporal dependencies and physical coupling constraints between different types of signals through a heterogeneous graph structure, including: Each sensor node is treated as a vertex of the heterogeneous graph. The signal correlation between different sensors at the same time is defined as a spatial edge, and the signal correlation between the same sensor at different times is defined as a temporal edge. Physical constraint edges are defined according to the mechanical structure connection relationship of the underwater leveling machine to construct the heterogeneous graph structure. Spatial attention coefficients are obtained by calculating the attention coefficients between nodes on the spatial edge through the graph attention layer. The spatial attention coefficients are calculated by a single-layer feedforward network based on the spliced ​​vector of node features and normalized by the Softmax function. Temporal dependency weights are obtained by calculating long-term dependency weights on temporal edges through a temporal convolutional attention network. The temporal convolutional attention network uses dilated convolution to capture multi-scale temporal features and introduces a channel attention mechanism to recalibrate feature channels. The spatial attention coefficient, temporal dependence weight, and preset coupling strength of physical constraint edges are fused to obtain a comprehensive propagation weight. Based on the comprehensive propagation weight, the features of neighboring nodes are weighted and aggregated to update the feature representation of each node. After multi-layer graph attention transformation, the fused calibration signal is output. The formula for calculating the spatial attention coefficient from node i to node j in the graph attention layer is as follows: ， ， Among them, h i Let h be the feature vector of node i. j Let W be the feature vector of node j. s To share the linear transformation matrix, a is the attention weight vector, a T Let || denote the transpose of the attention weight vector, || denote the vector concatenation operation, N(i) be the neighborhood set of node i, and μ be the balance coefficient. p is a priori coupling function based on spatial location. i p j All are sensor spatial coordinates, θ ij Let κ be the angle between the sensor's pointing points, κ be the attenuation coefficient, and exp(•) represent the exponential function. This represents the reciprocal of the Euclidean distance, and k represents the index variable; The update process of the node features is represented as follows: ， in, W represents the updated feature vector of node i. v Let W be the value transformation matrix. r Let σs(•) be the residual connection matrix, and let σs(•) denote the nonlinear activation function. The output of the temporal convolutional attention network is represented as follows: ， Where T1 represents the output feature matrix of the temporal convolutional attention network, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k M represents the dimension of the key vector. mask This is a causal mask matrix used to prevent future information leaks.

5. The method for calibrating signal data of an intelligent control underwater leveling machine according to claim 1, characterized in that, A physical information neural network model for underwater signal propagation is established. This model incorporates the acoustic wave propagation equation, water flow disturbance equation, and structural dynamics equation as physical constraints into the network loss function. This model then performs physical consistency correction on the fused calibration signal to obtain a physically consistent signal, including... A signal correction module based on a physical information neural network is constructed. The fused calibration signal is used as the input feature of the signal correction module. The physical consistency signal after correction by physical laws is output through the forward propagation process of the physical information neural network. The underwater sound wave propagation equation is embedded as the first physical constraint into the network loss function. The sound wave propagation equation describes the physical relationship between the propagation speed and attenuation characteristics of the sound pressure signal in the underwater medium. By calculating the partial derivatives of the sound pressure signal output by the network with respect to time and space, the residual term of the wave equation is constructed, and the residual term of the wave equation is used as the first physical constraint loss. The flow perturbation equation is embedded as the second physical constraint in the network loss function. The flow perturbation equation describes the coupling relationship between the flow velocity field and the pressure field and the perturbation effect on the signal propagation path. By calculating the partial derivatives of the flow velocity field and pressure field output by the network with respect to time and space, the residual term of the Navier-Stokes equation is constructed. The residual term of the Navier-Stokes equation is used as the second physical constraint loss. The flow perturbation equation includes the Navier-Stokes equation. The structural dynamics equation of the underwater leveling machine's main structure is embedded as the third physical constraint in the network loss function. The structural dynamics equation describes the dynamic relationship between the elastic deformation and vibration modes of the structure under external excitation. By calculating the partial derivatives of the structural displacement output by the network with respect to time and space, the residual terms of the Euler-Bernoulli beam equation are constructed. The residual terms of the Euler-Bernoulli beam equation are used as the third physical constraint loss. The structural dynamics equation includes the Euler-Bernoulli beam equation. The total loss function of the physical information neural network is composed of the weighted sum of the first physical constraint loss, the second physical constraint loss, and the third physical constraint loss.

6. The method for calibrating signal data of an intelligent control underwater leveling machine according to claim 1, characterized in that, An adaptive Kalman filter framework is constructed, using the physical consistency signal as the input observation. By estimating the system noise covariance matrix and the measurement noise covariance matrix online, the filter gain is dynamically adjusted to filter out residual Gaussian noise and random impulse noise in the signal, outputting the final calibration signal, including: A discrete state-space model of the underwater leveling machine's leveling process is established, comprising a state transition equation and an observation equation. A system noise covariance matrix is ​​defined to describe the statistical characteristics of system noise in the state transition equation, and a measurement noise covariance matrix is ​​defined to describe the statistical characteristics of measurement noise in the observation equation. Based on the state estimate of the previous moment, the system state and observation value at the current moment are predicted. The deviation between the physical consistency signal and the predicted observation value is defined as the innovation sequence. The actual covariance matrix of the innovation sequence is calculated within a sliding window, and the theoretical covariance matrix of the innovation sequence is calculated according to Kalman filtering theory. The ratio of the traces of the two is defined as the adaptive adjustment factor. The system noise covariance matrix is ​​dynamically updated based on the value of the adaptive adjustment factor: when the adaptive adjustment factor is greater than 1, the system noise covariance matrix is ​​increased; when the adaptive adjustment factor is less than 1, the system noise covariance matrix is ​​decreased. Simultaneously, a baseline value for the measurement noise covariance matrix is ​​calculated based on the sensor's nominal accuracy and environmental parameters, and the root mean square error of the innovation sequence is calculated within the sliding window; the root mean square error is compared with the baseline root mean square error, and the measurement noise covariance matrix is ​​dynamically corrected accordingly to obtain a time-varying measurement noise covariance matrix; Based on the updated system noise covariance matrix and the time-varying measurement noise covariance matrix, the Kalman filter gain is calculated; the state prediction value is corrected using the Kalman filter gain, and the corrected state is output as a calibration signal to filter out residual noise in the physical consistency signal.

7. The method for calibrating signal data of an intelligent control underwater leveling machine according to claim 1, characterized in that, The deployment scheme of the heterogeneous sensor cluster includes: arranging six-axis inertial measurement units at the four corners of the main frame of the leveling machine to collect attitude signals; integrating pressure-temperature composite sensors inside the cylinder of each hydraulic outrigger to collect pressure and temperature signals; arranging piezoelectric accelerometers at the connection between the outrigger end and the base plate to collect vibration signals; installing multi-beam sonar arrays at equal intervals along the longitudinal direction at the bottom of the leveling machine to collect sonar echo signals; and arranging electromagnetic flowmeters on both sides of the leveling machine to collect hydrodynamic environment signals. All sensors have built-in time synchronization modules and perform global clock calibration through underwater acoustic synchronization signals. The signal acquisition model of the multibeam sonar array is as follows: ， Among them, S sonar For the received sonar echo signal, θ and These are the azimuth and elevation angles, respectively; M is the number of sonar elements; A m Let R(t) be the directional gain of the m-th array element, and R(t) be the target reflected signal. m Let h(t) be the propagation delay, h(t) be the impulse response of the underwater acoustic channel, and n be the propagation delay. clutter For reverberation and ambient noise, t represents the time scale; The attitude calculation of the six-axis inertial measurement unit adopts a quaternion complementary filtering fusion algorithm, and the discrete form of the attitude calculation is represented as follows: ， ， Where, q c+1 Let q represent the attitude quaternion at time c+1. c Let c be the attitude quaternion at time c. B represents quaternion multiplication. s ω is the sampling period. c b is the angular velocity measurement value of the gyroscope. ω,c Let b be the zero-bias estimate of the gyroscope at time c. ω,c+1 Let q be the zero bias estimate of the gyroscope at time c+1. e K is the quaternion for attitude error measured by the accelerometer and magnetometer. p and K s For the proportional and integral gain of the complementary filter, This is the integral term for angular velocity error.

Images