Gas-liquid interface recognition and ship sloshing liquid level correction method and system

Abstract

The application provides a gas-liquid interface recognition and ship sloshing liquid level correction method and system, which extracts image edges and gray texture features, uses a light convolutional neural network model to realize liquid-gas interface state recognition, and dynamically generates a collection strategy accordingly; then, the original liquid level signal is collected and IMU data is synchronously acquired, multi-modal components are extracted through signal decomposition, a disturbance recognition model is constructed in combination with the IMU, disturbance-related components are filtered out and a stable liquid level is reconstructed; further, the disturbance components are modeled as residual sequences, a time series prediction model is used to realize interference trend prediction and liquid level dynamic correction, and more stable and reliable liquid level data is output. It can effectively deal with the problem of unstable liquid level measurement under the LNG ship sloshing environment, fuse image recognition and IMU data, realize liquid level disturbance suppression and dynamic correction, has the advantages of self-adaptation, anti-interference and measurement stability, and is suitable for high-precision liquid level monitoring under complex working conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of liquefied natural gas storage and transportation, and more particularly to a gas-liquid interface identification and ship sloshing liquid level correction method and system. BACKGROUND

[0002] As a clean and efficient energy form, liquefied natural gas (LNG) has been continuously increasing in the global energy transportation field in recent years. Due to the need to maintain extremely low temperature (about -162℃) and stable pressure during transportation, the safety and accurate monitoring of LNG storage and transportation are particularly important. In the process of liquefied natural gas transportation, the detection of the liquid level in the ship's storage tank is one of the core links to ensure the safety of loading, unloading and navigation stability.

[0003] At present, LNG transport ships generally use differential pressure type, float type, capacitance type, radar type and other liquid level meters for liquid level monitoring. However, during the navigation of liquefied natural gas transport ships, especially in severe sea conditions (such as strong wind, heavy waves, etc.), the liquid surface in the storage tank will experience violent sloshing and irregular fluctuations. The violent fluctuation of the liquid surface not only causes the output signal of the liquid level meter to jump frequently, but also may cause the measured value to be severely distorted or even the measurement sensor to be temporarily disabled. This dynamic disturbance causes a large deviation in the tracking of the real liquid level by the traditional liquid level meter, affecting the credibility and response efficiency of the storage and transportation system to the liquid level data. SUMMARY

[0004] The present application provides a gas-liquid interface identification and ship sloshing liquid level correction method and system, which can improve the accuracy of gas-liquid interface identification, dynamically correct the liquid level by fusing multi-source data, has strong anti-interference ability and prediction function, and reduces the influence of ship sloshing on the liquid level measurement results.

[0005] In a first aspect, the present application provides a gas-liquid interface identification and ship sloshing liquid level correction method, which comprises: extracting edge information and gray texture features contained in real-time collected image data and classifying liquid-gas interface state by using a trained lightweight convolutional neural network model, and generating a collection strategy according to the liquid-gas interface state classification result; collecting real-time liquid level original signals and synchronously acquiring IMU inertial measurement unit data according to the collection strategy; processing the liquid level original signals by a preset decomposition method, generating multi-modal component and constructing a liquid surface disturbance prediction model according to the IMU inertial measurement unit data, then identifying the disturbance principal component modal component and filtering out the modal component related to the disturbance, and reconstructing a stable liquid level signal; modeling the filtered disturbance modal component as a disturbance residual sequence and constructing a residual time series prediction model, predicting the liquid level disturbance trend and realizing dynamic correction of the liquid level, and outputting the dynamically corrected liquid level data.

[0006] In an optional implementation of the first aspect, when the acquisition strategy is generated, the method comprises: acquiring an edge map of the image data by using an edge detection algorithm and extracting texture features of the image data by using a gray level co-occurrence matrix, the texture features at least including contrast, energy, homogeneity and entropy features; inputting the edge map and the texture features into a trained MobileNet or EfficientNet-Lite light-weight convolutional neural network recognition model and outputting a predicted liquid-gas interface height and a liquid-gas state classification vector; and dynamically switching a weight source signal according to the liquid-gas state classification vector result: wherein w imu is an IMU channel weight, w liq is a liquid level sensor channel weight, W is a weight mapping matrix, and b is a bias term; generating a corresponding acquisition strategy according to the weight source signal: s t =[f t ,w imu ,w liq ,F pre ], wherein f t is a sampling frequency, and F pre is a filtering strategy selected according to a disturbance level; and outputting a gas-liquid interface state label, the gas-liquid interface state label at least including liquid surface stability, existence of bubbles or foam, severe shaking and serious splashing disturbance.

[0007] In an optional implementation of the first aspect, when the filtering strategy is selected according to the disturbance level, the method comprises: inputting image data within a time window and performing time sequence encoding on the image data by using a time sequence encoder to extract time sequence disturbance features; decoding the time sequence disturbance features to generate a vertical disturbance index, a horizontal disturbance index and a local disturbance index; constructing a disturbance three-dimensional vector according to the vertical disturbance index, the horizontal disturbance index and the local disturbance index, introducing a disturbance level label, and mapping the disturbance three-dimensional vector to the disturbance level; and dynamically selecting a filtering strategy according to the disturbance level, the filtering strategy at least including mean filtering, Kalman filtering and wavelet denoising.

[0008] In an optional implementation of the first aspect, when real-time data is acquired according to the acquisition strategy, the method comprises: acquiring real-time liquid level raw pressure difference signal data, capacitance signal data and radar signal data according to the liquid level sampling frequency and the filtering strategy corresponding to the acquisition strategy; and acquiring real-time three-axis attitude angle and acceleration data of an IMU (inertial measurement unit) according to the IMU sampling frequency and the filtering strategy corresponding to the acquisition strategy.

[0009] In an optional implementation of the first aspect, when the stable liquid level signal is reconstructed, the method comprises: decomposing the liquid level raw signal into a plurality of modal components by using empirical mode decomposition or variational mode decomposition:

[0010] Where u k (t) represents the k-th modal component, K is the total number of modes obtained from the decomposition, and L(t) is the original liquid level signal, including the original differential pressure, capacitance, and radar signal data; for each modal component u k (t), calculate the cross-correlation strength between it and the acceleration signal: Where R k Let a be the average cross-correlation between the k-th mode and the z-axis acceleration. z (t) represents the vertical acceleration, and T is the integration time window. If |R k If the value is greater than the disturbance discrimination threshold, the corresponding modal component is considered to be the principal component of the disturbance; an acceleration-driven liquid surface disturbance prediction model is established: in For the predicted disturbance signal,

[0011] a i (t)∈{a x (t),a y (t),a z (t)} represents the triaxial acceleration of the IMU inertial measurement unit, β i For regression weights, τ i The lag time for each acceleration channel reflects the dynamic response; modal components identified as disturbance-related are removed and the stable liquid level signal is reconstructed. Where I non-disturb ={1,2,…,K}\I disturb I disturb The set of modal components that are identified as disturbances.

[0012] In one alternative embodiment of the first aspect, when predicting liquid level disturbance trends and implementing dynamic correction of the liquid level, the method includes: extracting disturbance residuals from the original liquid level signal and the stable liquid level signal.

[0013] r(t) = L raw (t)-L stable r(t) is the disturbance residual signal, representing the influence of the sloshing disturbance on the liquid level. Based on the time dependence of the disturbance residual signal, a residual time series prediction model is constructed to predict the disturbance trend at the next moment. in For predicting the residuals at the next time step, n is the length of the input sequence of the residual time series prediction model, and M is the residual time series prediction model, including at least one of LSTM, GRU, 1D-CNN, Transformer, and ARIMA models; after predicting the perturbation trend, the liquid level data is corrected based on the predicted values: Where L corrected (t+1) represents the corrected liquid level data, taking into account both the current stable trend and future disturbance trends; Lstable (t+1) is a stable liquid level trend.

[0014] In an optional solution of the first aspect, when the liquid level data is corrected, the method further comprises: extracting spatio-temporal dynamic embedding features according to real-time collected image data; introducing a deformation convolution network enhanced model to enhance the perception ability of the disturbance, and performing dynamic offset identification of the edge region: ΔL img (t) = V deformCNN (I t-n ), wherein ΔL img (t) is an image rollback disturbance compensation term, I t-n is an image frame at time t-n; the image rollback disturbance compensation term is combined with the stable liquid level trend and the disturbance residual prediction value to dynamically correct the liquid level data:

[0015]

[0016] In an optional solution of the first aspect, when the liquid level data is dynamically corrected, the method further comprises: constructing a liquid level sensing signal confidence factor, an IMU confidence factor and an image confidence factor according to the disturbance degree, signal-to-noise ratio and prediction residual mean of the liquid level sensor channel, IMU channel and image acquisition channel within a preset time range; and dynamically weighting and fusing the corrected liquid level data according to the constructed confidence factors:

[0017] wherein ∈1, ∈2, ∈3 are the liquid level sensing signal confidence factor, the IMU confidence factor and the image confidence factor respectively, L fused (t+1) is the dynamically weighted and fused corrected liquid level data; and the dynamically weighted and fused corrected liquid level data is output.

[0018] In an optional solution of the first aspect, when the dynamically weighted and fused corrected liquid level data is corrected, the method comprises: constructing a liquid level variation rate monitoring index within a sliding window according to the stable liquid level trend:

[0019] wherein n is the size of the sliding window; if the variation rate monitoring index exceeds a preset threshold or any value of the confidence factor is lower than a preset tolerance value, it is determined that the current state is abnormal, and an emergency liquid level correction process is entered, the emergency liquid level correction process comprising: reducing the sampling frequency and prolonging the stable period of the sliding window, temporarily shielding the channel with the lowest confidence and increasing the trust weight of other channels, or introducing a sliding average trend of the historical liquid level stable sequence for short-term compensation.

[0020] In a second aspect, the application provides a gas-liquid interface recognition and ship sloshing liquid level correction system using the above liquid level correction method, comprising: a sensor assembly, including a liquid level sensor, an image acquisition unit, and an IMU inertial measurement unit, the liquid level sensor is used to collect liquid level original differential pressure signal data, capacitance signal data and radar signal data of a target container, the image acquisition unit is used to collect image data in the target container, and the IMU inertial measurement unit is used to collect three-axis attitude angle and acceleration data of the target container; a control unit connected with the sensor assembly, the control unit comprises: an image processing module, used to extract edge information and gray texture features contained in the real-time collected image data, and classify the liquid-gas interface state by using a trained lightweight convolutional neural network model, and generate a collection strategy according to the liquid-gas interface state classification result; a data collection module, used to collect image data, and also used to collect real-time liquid level original signals and synchronously acquire IMU inertial measurement unit data according to the collection strategy; a stable reconstruction module, used to process the liquid level original signals by a preset decomposition method, generate multi-modal component, and construct a liquid surface disturbance prediction model according to the IMU inertial measurement unit data, then identify a disturbance principal component modal component and filter out the modal component related to the disturbance, and reconstruct a stable liquid level signal; and a liquid level correction module, used to model the filtered disturbance modal component as a disturbance residual sequence, construct a residual time series prediction model, predict a liquid level disturbance trend, and realize dynamic correction of the liquid level, and output the dynamic corrected liquid level data.

[0021] It should be understood that the general description above and the following detailed description are only exemplary and do not limit the application. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate one or more embodiments of the present application and, together with the description, explain the principles of the application and enable a person skilled in the relevant art to make and use the application.

[0023] Figure 1 is an exemplary module connection schematic diagram of a gas-liquid interface recognition and ship sloshing liquid level correction system according to some embodiments of the application.

[0024] Figure 2 is an exemplary flowchart of a gas-liquid interface recognition and ship sloshing liquid level correction method according to some embodiments of the application.

[0025] Figure 3 is an exemplary flowchart of a collection strategy generation method according to some embodiments of the application.

[0026] Figure 4is a schematic diagram of an example light-weight convolutional neural network recognition model output prediction of a liquid-gas interface height and a liquid-gas state classification vector according to some embodiments of the present application.

[0027] Figure 5 is a flowchart of an example filter strategy selection method according to some embodiments of the present application.

[0028] Figure 6 is a flowchart of an example real-time data collection method according to some embodiments of the present application.

[0029] Figure 7 is a flowchart of an example stable liquid level signal reconstruction method according to some embodiments of the present application.

[0030] Figure 8 is a flowchart of an example liquid level dynamic correction method according to some embodiments of the present application.

[0031] Figure 9 is a flowchart of an example liquid level data correction method according to some embodiments of the present application.

[0032] Figure 10 is a flowchart of an example dynamic liquid level data correction method according to some embodiments of the present application.

[0033] Figure 11 is a flowchart of an example dynamic weighted fusion liquid level data correction method according to some embodiments of the present application.

[0034] Figure 12 is a connection diagram of an example electronic device according to some embodiments of the present application DETAILED DESCRIPTION

[0035] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example implementations to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more implementations. In the following description, numerous specific details are provided to give a thorough understanding of example implementations. Implementations can be used in any number of ways.

[0036] At present, liquefied natural gas (LNG) transportation is usually carried out by dedicated liquefied natural gas transport ships, which generally use differential pressure type, float type, capacitance type, radar type and other liquid level meters for liquid level monitoring.

[0037] Differential pressure level gauges, based on the principle of static pressure, calculate the liquid level by measuring the pressure difference between the bottom and top of the LNG storage tank. They are simple in structure, low in cost, and easy to maintain. However, they are significantly affected by changes in LNG density (temperature and composition), changes in tank pressure, condensation / vaporization of the pressure tapping pipe, and surface fluctuations and dynamic pressure changes caused by ship swaying. Specifically, LNG density fluctuates with temperature and composition, and density errors directly lead to deviations in liquid level calculations. Condensation / vaporization of the pressure tapping pipe can easily cause measurement lag, drift, or even blockage and failure. Ship swaying causes instability in the liquid surface, resulting in abnormal instantaneous dynamic pressure differences and misjudgment of the liquid level.

[0038] Servo / float level gauges use a motor to control the movement of a small ball or float up and down, measuring its equilibrium point when it floats on the liquid surface. They use the principle of buoyancy to sense the height of the float. However, under the long-term shaking and alternating hot and cold conditions at sea, they are prone to problems such as jamming and zero drift. The float needs to move with the liquid surface, and it will frequently impact the support when it shakes, making it unsuitable for environments with violent shaking.

[0039] Capacitive level gauges can change the capacitance between probes by using liquid as a medium, and calculate the liquid level based on the capacitance change. They can also indirectly measure the change in dielectric constant, helping to determine the gas-liquid interface and changes in the medium (such as density). However, they are also affected by liquid surface fluctuations caused by ship swaying, medium adhering to the wall, electrode icing / contamination, and small changes in the dielectric constant of LNG. Calibration is complex and requires frequent adjustments in dynamic environments. Specifically, ship swaying causes liquid to adhere to the probe, leading to misjudgment due to electrode adhering; electrode icing / contamination affects the measurement capacitance, causing long-term drift; and small changes in the dielectric constant of LNG require high signal discrimination capability and have a low signal-to-noise ratio.

[0040] Radar level gauges determine the level by emitting high-frequency electromagnetic waves (usually K-band or W-band) to the liquid surface and receiving the reflected wave signals, and calculating the reflection time. They are highly accurate and unaffected by density. However, the upper layer of LNG storage tanks is filled with low-temperature vapor, and the change in refractive index affects the propagation of radar waves. Furthermore, the refractive interface between the high-temperature layer and the low-temperature layer is prone to generating false echoes. Especially under violent shaking, the repeated reflection of liquid droplets in the gas phase space can cause measurement errors. In low-temperature environments, the antenna is prone to frost and ice formation, leading to signal attenuation or misjudgment.

[0041] Therefore, for reference Figure 1 As shown, Figure 1 The diagram illustrates the module connections of a gas-liquid interface identification and ship sloshing level correction system according to some embodiments of this application. To address measurement problems in liquefied natural gas (LNG) shipping, this application designs a gas-liquid interface identification and ship sloshing level correction system 1, including a sensor assembly 10 and a control unit 11.

[0042] Specifically, the sensor assembly 10 includes a liquid level sensor 101 for collecting liquid level raw differential pressure signal data, capacitive signal data and radar signal data of a target container, an image acquisition unit 102 for collecting image data in the target container, and an IMU inertial measurement unit 103 for collecting three-axis attitude angle and acceleration data of the target container.

[0043] The liquid level sensor 101 includes a differential pressure liquid level meter for collecting liquid level raw differential pressure signal data of the target container, a capacitive liquid level meter for collecting liquid level raw capacitive signal data of the target container, and a radar liquid level meter for collecting liquid level raw radar signal data of the target container. The image acquisition unit 102 includes an industrial camera capable of acquiring high-definition image data in the target container and / or an infrared thermal imager capable of acquiring infrared image data in the target container. The IMU inertial measurement unit 103 includes at least a three-axis accelerometer for measuring linear acceleration and reflecting movement trends in each direction, a three-axis gyroscope for measuring angular velocity and acquiring rotation and attitude change conditions, and a three-axis magnetometer for providing a reference for orientation and assisting in calibrating gyroscope offset and drift. Optionally, the IMU inertial measurement unit 103 can also be provided with a temperature sensor to compensate for temperature drift and improve measurement accuracy.

[0044] Specifically, the control unit 11 is connected to the sensor assembly 10, and the control unit 11 includes an image processing module 111 for extracting edge information and gray texture features contained in real-time collected image data and classifying liquid-gas interface states using a trained lightweight convolutional neural network model, and generating a collection strategy according to the liquid-gas interface state classification result; a data collection module 112 for collecting image data and also for collecting real-time liquid level raw signal and synchronously acquiring IMU inertial measurement unit 103 data according to the collection strategy; a stable reconstruction module 113 for processing the liquid level raw signal by a preset decomposition method, generating multi-modal component and constructing a liquid surface disturbance prediction model according to the IMU inertial measurement unit 103 data, then identifying a disturbance principal component modal component and filtering out modal components related to the disturbance, and reconstructing a stable liquid level signal; and a liquid level correction module 114 for modeling the filtered disturbance modal component as a disturbance residual sequence and constructing a residual time series prediction model, predicting liquid level disturbance trends and realizing dynamic correction of the liquid level, and outputting dynamic corrected liquid level data.

[0045] In some embodiments of the present application, referring to Figure 2 , the image acquisition unit 102 includes an industrial camera capable of acquiring high-definition image data in the target container and / or an infrared thermal imager capable of acquiring infrared image data in the target container. The IMU inertial measurement unit 103 includes at least a three-axis accelerometer for measuring linear acceleration and reflecting movement trends in each direction, a three-axis gyroscope for measuring angular velocity and acquiring rotation and attitude change conditions, and a three-axis magnetometer for providing a reference for orientation and assisting in calibrating gyroscope offset and drift. Optionally, the IMU inertial measurement unit 103 can also be provided with a temperature sensor to compensate for temperature drift and improve measurement accuracy. Figure 2The diagram illustrates a flow chart of a gas-liquid interface identification and ship sloshing level correction method according to some embodiments of this application. This application also designs a gas-liquid interface identification and ship sloshing level correction method, comprising the following steps:

[0046] S1: Extract edge information and grayscale texture features from real-time acquired image data and use a trained lightweight convolutional neural network model to classify the liquid-gas interface state. Generate an acquisition strategy based on the liquid-gas interface state classification results.

[0047] Specifically, refer to Figure 3 and Figure 4 As shown, Figure 3 The following is a flowchart illustrating a method for generating acquisition strategies according to some embodiments of this application. Figure 4 This diagram illustrates the predicted liquid-gas interface height and liquid-gas state classification vector output by a lightweight convolutional neural network recognition model according to some embodiments of this application. When generating the acquisition strategy in S1, the method includes:

[0048] S11: An edge detection algorithm is used to obtain the edge map of the image data and the gray-level co-occurrence matrix is ​​used to extract the texture features of the image data. The texture features include at least contrast, energy, homogeneity and entropy features.

[0049] Among them, the Sobel or Canny edge detection algorithm is used to obtain the edge map of the image data:

[0050] E(x,y,t) = Sobel(I(x,y,t)), or E(x,y,t) = Canny(I(x,y,t)), where x and y are pixel coordinates, t is the image data acquisition time, and E(x,y,t) is the edge map. The Sobel algorithm uses the gradient of gray-level differences in the image to detect edges, while the Canny algorithm is a multi-stage edge detection algorithm that includes noise reduction, gradient calculation, non-maximum suppression, and double thresholding.

[0051] Wherein, the gray level co-occurrence matrix is a statistical tool for describing the gray level relationship between pixels, reflecting the spatial structure distribution of image texture, and the texture feature indexes including contrast, energy, homogeneity and entropy can be calculated by calculating the gray level co-occurrence matrix of the original image data. The contrast feature can reflect the degree of gray level change, the energy feature can reflect the repetition degree of image texture, the homogeneity feature can measure whether the gray level distribution of the image is concentrated (the closer to the diagonal line, the higher), and the entropy feature can reflect the complexity or uncertainty of the image (the higher the value, the more complex the texture). The specific process of extracting the texture features of image data by using the gray level co-occurrence matrix is as follows: gray-scale the image data, set the gray level number (such as 256 levels), and the relative position. Traverse each pair of adjacent pixels of the gray-scale image data, count the gray level, and fill in the matrix P(i,j). Normalize the matrix to make all element values represent the co-occurrence probability; (i,j)∈[0,G-1], G is the gray level number, and thus:

[0052] The extraction process of the contrast feature is as follows: The greater the difference value, the greater the value, and the higher the contrast when the texture changes sharply (such as edges).

[0053] The extraction process of the energy feature is as follows: The greater the energy value, the more uniform the image, the more regular the texture, and the higher the energy of the flat area.

[0054] The extraction process of the homogeneity feature is as follows: When the gray level values of adjacent pixels are close (i≈j), the score is larger, reflecting the smoothness of the image.

[0055] The extraction process of the entropy feature is as follows: Wherein, ψ is a small number (such as 1e-10) to avoid log(0); the more complex the image, the more random the gray level distribution, and the greater the entropy.

[0056] S12: input the edge map and the texture features into the trained MobileNet or EfficientNet-Lite lightweight convolutional neural network recognition model and output the predicted liquid-gas interface height and liquid-gas state classification vector.

[0057] Wherein, the MobileNet or EfficientNet-Lite lightweight convolutional neural network recognition model is selected by the technical personnel from the existing architecture of the lightweight convolutional neural network recognition model and is adaptively trained according to the actual demand, and then the trained lightweight convolutional neural network recognition model is deployed at the required position, and then the edge map and the texture features are received to output the vertical position of the liquid surface in the image and the liquid-gas state classification of the current image, such as stable liquid surface, bubbles / foam, severe shaking or serious liquid droplet splashing interference.

[0058] S13: Dynamically switch the weight source signal based on the liquid-gas state classification vector results:

[0059] Where w imu For IMU channel weights, w liq Here, W represents the weights for the 101 channels of the level sensor, W is the weight mapping matrix, and b is the bias term; a corresponding acquisition strategy is generated based on the weight source signals.

[0060] s t =[f t ,w imu ,w liq ,F pre ], where f t F is the sampling frequency. pre The filtering strategy is selected based on the disturbance level; at the same time, the gas-liquid interface status label is output, which includes at least the following: stable liquid surface, presence of bubbles or foam, violent shaking, and severe splashing interference.

[0061] In some examples of this application, see reference Figure 5 As shown, Figure 5 A flowchart illustrating a filtering strategy selection method according to some embodiments of this application is shown. In S13, when selecting a filtering strategy based on the disturbance level, the method includes:

[0062] S131: Input image data within the time window and use a timing encoder to perform timing encoding on the image data to extract timing perturbation features.

[0063] S132: Decode the temporal disturbance characteristics to generate vertical disturbance indices, horizontal disturbance indices, and local disturbance indices.

[0064] The vertical disturbance index is as follows: Where L top (y,t) represents the height of the liquid level edge in the y-th row of the image at time t. σ is the stable mean of the corresponding position in historical images. v This is the normalized scaling factor.

[0065] The horizontal disturbance index is: Where θ t The slope angle of the liquid surface in the current frame. For reference average angle, σ h This is the scaling factor.

[0066] The local disturbance index is: Among them, R i Let σ represent the i-th local perturbation region in the image (e.g., a bubble spot). Entropy is the entropy feature used to reflect the perturbation complexity. o This is the normalization factor.

[0067] S133: Based on the vertical perturbation index, horizontal perturbation index and local perturbation index, construct the perturbation three-dimensional vector: U(t)=[Z(t),h(t),O(t)], and introduce the perturbation level label C(t)={0,,1,2,3,4}, and then map the perturbation three-dimensional vector to the perturbation level through the classifier D: C(t)=D(U(t)).

[0068] The disturbance levels are as follows:

[0069] Disturbance level Description Features 0 Calm (flat surface) All perturbation terms < threshold 1 Light disturbance (micro waves or slight tilt) Z(t) and H(t) near threshold 2 Moderate disturbance (obvious tilt or bubble interference) H(t) and O(t) elevated 3 Strong disturbance (violent surges, large-scale liquid overflow) Z(t), H(t) and O(t) exceed threshold

[0070] S134: Dynamically select a filtering strategy based on the disturbance level, wherein the filtering strategy includes at least one of mean filtering, Kalman filtering, and wavelet denoising.

[0071] By combining vertical disturbance index, horizontal disturbance index, and local disturbance index with a disturbance level classifier and dynamic filtering strategy selection, adaptive control of liquid level measurement error is achieved.

[0072] For example, the data collection strategy is shown in the table below:

[0073]

[0074] By adopting the above technical solution, multi-dimensional texture features extracted from image edge detection and gray-level co-occurrence matrix are integrated. A lightweight convolutional neural network model is used to accurately identify the height and state of the liquid-gas interface, and to classify disturbances such as liquid surface stability, bubbles, foam, and splashing. Furthermore, based on the state classification results, the trust weights of the IMU and the liquid level sensing channel are adaptively adjusted through softmax mapping, and a dynamic acquisition strategy including sampling frequency and filtering strategy is generated. In addition, a time encoder is introduced to extract image disturbance evolution features, construct a three-dimensional disturbance vector and map the disturbance level to realize the level adjustment of the filtering strategy.

[0075] S2: Acquire real-time raw liquid level signals according to the acquisition strategy and simultaneously acquire data from the IMU inertial measurement unit 103.

[0076] Specifically, refer to Figure 6 As shown, Figure 6 A flowchart illustrating a method for acquiring real-time data according to some embodiments of this application is shown. In S2, when acquiring real-time data according to an acquisition strategy, the method includes:

[0077] S21: Based on the corresponding liquid level sampling frequency and filtering strategy of the acquisition strategy, acquire real-time raw liquid level differential pressure signal data, capacitance signal data and radar signal data.

[0078] S22: Based on the IMU sampling frequency and filtering strategy corresponding to the acquisition strategy, acquire the real-time three-axis attitude angle and acceleration data of the IMU inertial measurement unit 103.

[0079] S23: Acquire real-time image data from image acquisition unit 102 according to the visual sampling frequency and filtering strategy corresponding to the acquisition strategy.

[0080] By adopting the above technical solution, based on differentiated sampling and filtering configuration, the acquisition accuracy and resource consumption can be flexibly adjusted according to the needs of the scenario. While ensuring the dynamic response capability of key disturbances, invalid high-frequency interference is suppressed. This multi-channel, configurable data acquisition mechanism provides a high-quality, time-consistent data foundation for subsequent liquid level fusion estimation, disturbance identification and liquid level correction steps, thereby improving the overall accuracy and stability of liquid level measurement.

[0081] S3: The original liquid level signal is processed by a preset decomposition method to generate multimodal components and a liquid surface disturbance prediction model is constructed based on the data of the IMU inertial measurement unit 103. Then, the main disturbance modal components are identified and the modal components related to the disturbance are filtered out to reconstruct a stable liquid level signal.

[0082] Specifically, refer to Figure 7 As shown, Figure 7 A flowchart illustrating a method for reconstructing a stable liquid level signal according to some embodiments of this application is shown. When reconstructing the stable liquid level signal in S3, the method includes:

[0083] S31: The original liquid level signal is decomposed into several modal components using empirical mode decomposition or variational mode decomposition.

[0084] Where u k L(t) represents the k-th modal component, K represents the total number of modes obtained from the decomposition, and L(t) represents the original liquid level signal, including the original differential pressure, capacitance, and radar signal data.

[0085] The original liquid level signals in this application include the original differential pressure signal acquired in real time by the differential pressure level gauge, the capacitance signal acquired in real time by the capacitive level gauge, and the radar signal acquired in real time by the radar level gauge.

[0086] In empirical mode decomposition, the nonlinear, non-stationary signal L(t) is decomposed into several modal components with intrinsic frequency characteristics: Empirical Mode Decomposition (EMD) is an adaptive data-driven method suitable for processing noisy raw liquid level signals.

[0087] In variational mode decomposition, the signal is considered to consist of several band-limited modes, and L(t) is decomposed into several modal components through variational optimization: Each modal component uk (t) is a band-limited signal characterized by different center frequencies, and the optimal decoupling between modes is achieved by minimizing the total modal bandwidth.

[0088] By selectively retaining and reconstructing the modal components, the interference information is suppressed and the true liquid level change trend is extracted, enhancing the system's adaptability to sloshing disturbances.

[0089] S32: For each modal component u k (t), calculate the cross-correlation strength between it and the acceleration signal:

[0090] where R k is the average cross-correlation of the kth mode with the z-axis acceleration, a z (t) is the vertical direction acceleration, T is the integral time window, and if |R k | is greater than the interference discrimination threshold, it is considered that the corresponding modal component is a disturbance principal component.

[0091] where a disturbance discrimination threshold δ is set, and if a certain modal component satisfies |R k |>δ, it means that the mode is highly correlated with the vertical acceleration and belongs to the component dominated by sloshing disturbance.

[0092] S33: Establish an acceleration-driven liquid surface disturbance prediction model:

[0093] where is the predicted disturbance signal,

[0094] a i (t)∈{a x (t),a y (t),a z (t)} is the three-axis acceleration of the IMU inertial measurement unit 103, β i is the regression weight, and τ i is the lag time of each acceleration channel, reflecting the dynamic response.

[0095] where, when the ship moves on the sea surface, the change in attitude will cause fluctuations in the liquid level signal measured by the liquid level sensor 101. These fluctuations are not real liquid level changes, but disturbance components caused by ship acceleration. By learning the mapping relationship between the acceleration signal and the liquid level disturbance, a regression model can be established to predict the disturbance value. Subtracting the predicted disturbance from the original liquid level signal can obtain a relatively stable liquid surface trend signal.

[0096] S34: Remove the modal components judged to be related to disturbance and reconstruct the stable liquid level signal:

[0097] Among them I non-disturb ={1,2,…,K}\I disturb I disturb The set of modal components that are identified as disturbances.

[0098] The reconstructed stable liquid level signal is the liquid level estimate after removing swaying disturbances, which is more stable than the original signal.

[0099] By adopting the above technical solution, the influence of factors such as swaying, tilting, and vibration on liquid level measurement can be significantly reduced in a strongly disturbed environment, improving the accuracy and stability of liquid level estimation, and possessing good adaptive capabilities. It is particularly suitable for liquid level monitoring applications in complex working conditions such as ships.

[0100] S4: The filtered disturbance mode components are modeled as disturbance residual sequences and residual time series prediction models are constructed to predict the liquid level disturbance trend and realize the dynamic correction of the liquid level, and output the dynamically corrected liquid level data.

[0101] Specifically, refer to Figure 8 As shown, Figure 8 A flowchart illustrating a method for dynamically correcting liquid levels according to some embodiments of this application is shown. In step S4, when predicting liquid level disturbance trends and implementing dynamic liquid level correction, the method includes:

[0102] S41: Extract the disturbance residual from the original liquid level signal and the stable liquid level signal:

[0103] r(t) = L raw (t)-L stable r(t) is the disturbance residual signal, which represents the influence of sloshing disturbance on liquid level. r(t) ≈ disturbance component + measurement error.

[0104] Among them, the extraction of disturbance residuals is mainly used to quantify the disturbance effect in the liquid level signal caused by non-real liquid level change factors such as sloshing and vibration.

[0105] S42: Based on the time dependence of the perturbation residual signal, construct a residual time series prediction model to predict the perturbation trend at the next time step: in For the prediction of the residual at the next time step, n is the length of the input sequence of the residual time series prediction model, and M is the residual time series prediction model, including at least one of LSTM, GRU, 1D-CNN, Transformer, and ARIMA models.

[0106] Wherein, by learning the time evolution law of disturbance residual r(t), the disturbance trend at the next moment is predicted, which is used for future liquid level disturbance estimation and advance correction, and enhances the adaptability to sloshing or fluctuating environment, and improves the stability and accuracy of multi-sensor liquid level measurement; LSTM long short-term memory network model, which is suitable for capturing long-time dependent nonlinear disturbance characteristics; GRU gate recurrent unit, which has fewer parameters and faster training, and is suitable for edge deployment; 1D-CNN one-dimensional convolution network, which is suitable for extracting local time sequence features in disturbance signal; Transformer multi-head attention mechanism, which can consider long and short term dependencies at the same time, and is suitable for complex disturbance; ARIMA model is suitable for linear disturbance trend modeling, and has strong parameter interpretability.

[0107] By fitting the time evolution law of disturbance, the forward-looking prediction of disturbance trend is realized, and dynamic compensation support is provided for liquid level correction.

[0108] S43: After predicting the disturbance trend, correct the liquid level data based on the predicted value:

[0109] Wherein, L corrected (t+1) is the corrected liquid level data, which comprehensively considers the current stable trend and future disturbance trend; L stable (t+1) is the stable liquid level trend.

[0110] Wherein, the predictable disturbance trend is added back to the stable trend, so as to more accurately simulate the real liquid level reading at the next moment. Therefore, the correction process comprehensively considers the stable liquid level change trend and future disturbance trend, and can realize more accurate liquid level data estimation in disturbance environment such as liquid surface sloshing.

[0111] By using the above technical scheme, the disturbance residual can be extracted from the liquid level original signal and the stable liquid level signal, and a residual time sequence prediction model can be constructed based on the time dependence, which can effectively realize the prediction of the disturbance trend at the next moment, and can also make the final liquid level estimation retain the long-term stable trend while having the forward-looking correction ability to short-time disturbance, so that the data can still be smooth and continuous in strong disturbance or measurement delay scene, and the accuracy of liquid level monitoring is improved.

[0112] In some examples of the present application, in complex environment (such as foam, water splash, spot shielding), deviation is often caused by visual error, in order to improve the accuracy of liquid level estimation, image space-time information and deformation convolution perception ability are introduced to compensate the image disturbance and jointly correct the final liquid level with stable liquid level and disturbance prediction. Thus, referring to Figure 9 , it is shown that Figure 9 shows a flowchart of the method for correcting liquid level data according to some embodiments of the present application. When correcting the liquid level data, the specific method includes:

[0113] S431: Extract the spatio-temporal dynamic embedding features according to the real-time collected image data.

[0114] Among them, the spatio-temporal dynamic embedding features of the image data are extracted by using a three-dimensional convolution network (3DCNN) or an image time series Transformer model, that is, the liquid level region features of each frame of image in the image data are extracted as the input of the deformation convolution network in S432.

[0115] S432: Introduce a deformation convolution network enhanced model to enhance the perception ability of the disturbance, and perform dynamic offset recognition of the edge region: ΔL img (t) = V deformCNN (I t-n ), wherein ΔL img (t) is an image rollback disturbance compensation term for correcting the liquid level error caused by visual disturbance; I t-n is the image frame at time t-n.

[0116] Among them, the DeformableConvNet is introduced to enhance the recognition and positioning ability of the model to irregular disturbances (such as bubbles, foam contours, spot occlusion, etc.), so as to infer the disturbance error and compensate; wherein the backbone network can be selected from ResNet-18 / ResNet-50+DeformableConvolution (DCNv2) or YOLOv5backbone+deformablefeaturemodule, and the training method adopts a dataset with real liquid level labels and disturbance image samples (foam, occlusion) and uses a regression loss (such as SmoothL1Loss) to train the network output disturbance error.

[0117] S433: Combine the image rollback disturbance compensation term with the stable liquid level trend and the disturbance residual prediction value to dynamically correct the liquid level data:

[0118] By using the above technical solution, the irregular disturbance area is accurately perceived by using the deformation convolution, the edge recognition and error positioning ability is improved, the adaptability to complex disturbance scenes is enhanced, the dynamic deformation evolution features are extracted, the liquid surface three-dimensional disturbance recognition is realized, and the liquid level error under visual disturbance is reduced through the residual rollback mechanism guided by the image.

[0119] In some examples of the present application, in a complex dynamic environment (such as ship sway, image occlusion, liquid surface fluctuation), information distortion may exist in a single channel. By analyzing the disturbance degree, signal-to-noise ratio, prediction residual amplitude and other parameters of each channel, the confidence factor is calculated, multi-source data fusion is realized, and the accuracy of liquid level estimation is improved. Thus, as shown in Figure 10 , the liquid level estimation accuracy is improved. Figure 10The flowchart of the dynamic correction method of liquid level data of some embodiments of the application is shown. When the liquid level data is dynamically corrected, the specific method includes:

[0120] S434: According to the disturbance degree, signal-to-noise ratio and predicted residual mean of the liquid level sensor 101 channel, IMU channel and image acquisition channel in the preset time range, the liquid level sensor signal confidence factor, IMU confidence factor and image confidence factor are constructed.

[0121] Wherein, the preset time range is set by a technician as a sliding time window [t-Δt, t] for analyzing the stability of each channel, and the time window size can be set to 1-3 seconds; the calculation process of the disturbance degree of the liquid level sensor 101 channel is:

[0122] D L =Var(L(t-Δt:t)), wherein Var is a variance operator for measuring the dispersion degree of data, and the calculation process of the signal-to-noise ratio is: Power signal is the effective power of the liquid level signal, which is usually the mean value of the energy of the liquid level signal, and Power noise is the noise power generated in the measurement process of the sensor; the calculation process of the disturbance degree of the IMU inertial measurement unit 103 channel is: Wherein a x (t) is the lateral acceleration, and a z (t) is the longitudinal acceleration; the image edge change amount (such as the area fluctuation of the liquid surface edge) of the image acquisition channel can be used as the disturbance degree index, and the signal-to-noise ratio SNR can be indirectly estimated by the image definition (such as the Laplacian variation coefficient). Then the weight is constructed by normalization processing: Wherein each α i (i=1, 2, 3) represents the channel reliability score:

[0123] Wherein is the predicted residual mean of the liquid level sensor 101 channel, and δ is a positive number to prevent the denominator from being zero, is the predicted residual mean of the image acquisition channel.

[0124] S435: According to the constructed confidence factors, the liquid level data is dynamically weighted and fused to correct:

[0125] Wherein ∈1, ∈2, ∈3 are respectively the liquid level sensor signal confidence factor, IMU confidence factor and image confidence factor, and L fused (t+1) is the dynamically weighted and fused corrected liquid level data.

[0126] S436: Output the dynamically weighted and fused corrected liquid level data.

[0127] By adopting the technical solution, a multi-channel confidence evaluation mechanism based on disturbance amplitude, signal-to-noise ratio and prediction residual is introduced, dynamic weight fusion of the liquid level sensor 101, IMU and image information is realized, compared with the traditional single source liquid level estimation method, the stability and accuracy of liquid level measurement can be improved in a complex disturbance environment, and the interference of abnormal fluctuations on the final liquid level result is reduced.

[0128] In some examples of the present application, when dynamically weighting and fusing the modified liquid level data, in order to improve the stability of sudden abnormal disturbances (such as sensor failure, ship body collision, surge disturbance, etc.) or abnormal perception channels (such as image blur, signal loss), a liquid level abnormal variation detection is introduced to monitor and statistically identify abnormal signals in real time. Thus, referring to FIG. 8, a flowchart of a dynamic weighting and fusion method for modifying liquid level data is shown. The method comprises: Figure 11 Figure 11 A flowchart of a dynamic weighting and fusion method for modifying liquid level data is shown. The method comprises:

[0129] S4351: Set the sliding window size n, for example n = 5-20, then collect the stable liquid level trend L stable (t-i) at time points t-n-1 to t-1, and then construct a liquid level variation rate monitoring index within the sliding window according to the stable liquid level trend:

[0130] Thus, the average variation speed of the liquid level in the sliding time period is reflected by the liquid level variation rate monitoring index.

[0131] S4352: If the variation rate monitoring index exceeds the preset threshold or any value in the confidence factor is lower than the preset tolerance value, it is determined that the current state is abnormal, and an emergency liquid level correction process is entered, which comprises:

[0132] Reducing the sampling frequency and prolonging the stable period of the sliding window, temporarily shielding the channel with the lowest confidence and increasing the trust weight of other channels, or introducing the sliding average trend of the historical liquid level stable sequence for short-term compensation.

[0133] The preset threshold can be set based on historical data statistics, such as an empirical value or based on standard deviation.

[0134] The process of reducing the sampling frequency and prolonging the stable period of the sliding window comprises at least one of temporarily reducing the liquid level update frequency, for example from 1 Hz to 0.2 Hz, increasing the sliding window length n, improving the data smoothness (such as from n = 10 to n = 30), and reducing the interference of short-term fluctuations on the variation rate calculation.

[0135] ​The process of temporarily shielding the channel with the lowest confidence and increasing the trust weight of other channels includes: identifying the channel with the lowest current confidence (for example, the image confidence factor ∈3 is the lowest), temporarily shielding the channel (setting the corresponding weight to 0), and then proportionally readjusting and normalizing the remaining two channels.

[0136] The process of introducing the moving average trend of the stable sequence of historical liquid levels for short-term compensation includes: extracting the mean trend of the liquid level in the previous stable period, and then replacing the untrusted data with the moving average during the abnormal state for short-term compensation correction.

[0137] By adopting the above technical solutions, the stability of the system to sudden disturbances and extreme abnormalities can be improved, and liquid level errors caused by confidence decline can be avoided, and the continuity of liquid level monitoring can be ensured through redundant channel cooperation and historical trend compensation.

[0138] In some embodiments, with reference to Figure 12 as shown, Figure 12 A connection diagram of an electronic device for implementing the embodiments of the present application is shown. The electronic device 3 includes a memory 301 and a processor 302, and the memory 301 stores a computer program that can run on the processor 302. The processor 302 implements the method in the above embodiments when executing the computer program. The number of memories 301 and processors 302 can be one or more.

[0139] The electronic device 3 further includes:

[0140] A communication interface 303 for communicating with external devices and performing data interaction transmission.

[0141] If the memory 301, the processor 302 and the communication interface 303 are independently implemented, the memory 301, the processor 302 and the communication interface 303 can be connected to each other through a bus and complete communication between them.

[0142] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 12 only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0143] Optionally, in a specific implementation, if the memory 301, the processor 302 and the communication interface 303 are integrated on a chip, the memory 301, the processor 302 and the communication interface 303 can complete the communication among each other through an internal interface.

[0144] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed by the processor 302 to implement the method provided in the embodiment of the present application.

[0145] The embodiment of the present application further provides a chip, which comprises the processor 302, and the processor 302 is used for calling and running the instruction stored in the memory 301, so that the communication device installed with the chip executes the method provided in the embodiment of the present application.

[0146] The embodiment of the present application further provides a chip, which comprises an input interface, an output interface, a processor 302 and a memory 301, the input interface, the output interface, the processor 302 and the memory 301 are connected through an internal connection path, and the processor 302 is used for executing the code in the memory 301, and when the code is executed, the processor 302 is used for executing the method provided in the embodiment of the present application.

[0147] It should be understood that the processor 302 described above can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. It should be noted that the processor 302 can be a processor supporting an advanced RISC machine (ARM) architecture.

[0148] Further, the memory 301 can include a read-only memory and a random access memory, and can further include a non-volatile random access memory. The memory 301 can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. The non-volatile memory can include a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can include a random access memory (RAM) used as an external cache. By way of example, and not limitation, many forms of RAM are available. For example, a static random access memory (SRAM), a dynamic random access memory (DRAM), a synchronous dynamic random access memory (SDRAM), a double data rate SDRAM (DDR SDRAM), an enhanced SDRAM (ESDRAM), a Synchlink DRAM (SLDRAM), and a direct Rambus RAM (DRRAM) are available.

[0149] In the above-described embodiments, all or part of the embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the present disclosure are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable apparatus. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium.

[0150] The above description is merely a specific implementation of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present disclosure, and all such changes or replacements should be covered within the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be subject to the scope of protection of the claims.

Claims

1. A method for gas-liquid interface identification and ship sloshing level correction, characterized in that, The method includes: The edge information and grayscale texture features contained in the real-time acquired image data are extracted and the liquid-gas interface state is classified using a trained lightweight convolutional neural network model. The acquisition strategy is generated based on the liquid-gas interface state classification results. Real-time raw liquid level signals are acquired according to the acquisition strategy, and IMU inertial measurement unit data is acquired simultaneously. The original liquid level signal is processed by a preset decomposition method to generate multimodal components. A liquid surface disturbance prediction model is constructed based on the IMU inertial measurement unit data. Then, the main disturbance modal components are identified and the disturbance-related modal components are filtered out to reconstruct a stable liquid level signal. The filtered disturbance modal components are modeled as disturbance residual sequences and residual time series prediction models are constructed to predict the liquid level disturbance trend and realize the dynamic correction of the liquid level, and output the dynamically corrected liquid level data. The method for reconstructing a stable liquid level signal includes: The original liquid level signal is decomposed into several modal components using empirical mode decomposition or variational mode decomposition: ,in Let K be the k-th modal component, and K be the total number of modes obtained from the decomposition. The raw liquid level signal includes raw differential pressure, capacitance, and radar signal data; For each modal component Calculate the cross-correlation strength between it and the acceleration signal: ,in The average cross-correlation between the k-th mode and the z-axis acceleration. Let T be the vertical acceleration, and T be the integration time window. If the value is greater than the interference discrimination threshold, the corresponding modal component is considered to be the principal perturbation component. Establish an acceleration-driven liquid surface disturbance prediction model: ,in For the predicted disturbance signal, The triaxial acceleration of the IMU inertial measurement unit. For regression weights, The lag time for each acceleration channel reflects the dynamic response; The modal components identified as disturbance-related will be removed and the stable liquid level signal will be reconstructed. ,in , The set of modal components that are identified as disturbances.

2. The liquid level correction method according to claim 1, characterized in that, When generating the acquisition strategy, the method includes: An edge detection algorithm is used to obtain the edge map of the image data, and the gray-level co-occurrence matrix is ​​used to extract the texture features of the image data. The texture features include at least contrast, energy, homogeneity and entropy features. The edge map and texture features are input into the trained MobileNet or EfficientNet-Lite lightweight convolutional neural network recognition model, and the predicted liquid-gas interface height and liquid-gas state classification vector are output. Dynamically switch the weight source signal based on the liquid-gas state classification vector results: ,in For IMU channel weights, For the channel weights of the liquid level sensor, Here, b is the weight mapping matrix, and b is the bias term; the corresponding acquisition strategy is generated based on the weight source signal. ,in, Sampling frequency, The filtering strategy is selected based on the disturbance level; at the same time, the gas-liquid interface status label is output, which includes at least the following: stable liquid surface, presence of bubbles or foam, violent shaking, and severe splashing interference.

3. The liquid level correction method according to claim 2, characterized in that, When selecting a filtering strategy based on the disturbance level, the method includes: Input image data within the time window and use a time encoder to perform time-series encoding on the image data to extract time-series perturbation features; Decode the temporal disturbance characteristics to generate vertical disturbance indices, horizontal disturbance indices, and local disturbance indices; Based on the vertical disturbance index, horizontal disturbance index, and local disturbance index, a three-dimensional disturbance vector is constructed and a disturbance level label is introduced to map the three-dimensional disturbance vector to the disturbance level. The filtering strategy is dynamically selected based on the disturbance level. The filtering strategy includes at least one of mean filtering, Kalman filtering, and wavelet denoising.

4. The liquid level correction method according to claim 2 or 3, characterized in that, When collecting real-time data according to the collection strategy, the method includes: Based on the corresponding liquid level sampling frequency and filtering strategy of the acquisition strategy, real-time raw liquid level differential pressure signal data, capacitance signal data and radar signal data are acquired; Based on the corresponding IMU sampling frequency and filtering strategy, the real-time three-axis attitude angle and acceleration data of the IMU inertial measurement unit are acquired.

5. The liquid level correction method according to claim 1, characterized in that, When predicting liquid level disturbance trends and implementing dynamic liquid level correction, the method includes: Extract the disturbance residual from the original liquid level signal and the stable liquid level signal: ,in The disturbance residual signal represents the effect of sloshing disturbance on the liquid level; Based on the time dependence of the perturbation residual signal, a residual time series prediction model is constructed to predict the perturbation trend at the next time step: ,in To predict the residuals at the next time step, The input sequence length is the residual time series prediction model, and M is the residual time series prediction model, including at least one of LSTM, GRU, 1D-CNN, Transformer, and ARIMA models; After predicting the disturbance trend, the liquid level data is corrected based on the predicted values: ,in The corrected liquid level data takes into account both the current stable trend and future disturbance trends. To stabilize the liquid level trend.

6. The liquid level correction method according to claim 5, characterized in that, When correcting the liquid level data, the method further includes: Based on real-time acquired image data, extract spatiotemporal dynamic embedding features; A deformable convolutional network is introduced to enhance the model's ability to perceive perturbations and perform dynamic offset recognition of edge regions: ,in, For image backtracking disturbance compensation, The image frame at time tn; The image backtracking perturbation compensation term is combined with the stable liquid level trend and the predicted perturbation residual value to dynamically correct the liquid level data: 。 7. The liquid level correction method according to claim 6, characterized in that, When dynamically correcting liquid level data, the method further includes: Based on the disturbance level, signal-to-noise ratio, and mean prediction residual of the liquid level sensor channel, IMU channel, and image acquisition channel within a preset time range, a confidence factor for the liquid level sensing signal, an IMU confidence factor, and an image confidence factor are constructed. The liquid level data is dynamically weighted and corrected based on the constructed confidence factors: ,in These are the confidence factors for the liquid level sensing signal, the IMU confidence factor, and the image confidence factor, respectively. The liquid level data is dynamically weighted and corrected. Output the dynamically weighted fusion corrected liquid level data.

8. The liquid level correction method according to claim 7, characterized in that, When dynamically weighted and correcting liquid level data, the method includes: Based on the stable liquid level trend, a liquid level variation rate monitoring index is constructed within the sliding window: , where n is the size of the sliding window; If the variation rate monitoring index exceeds the preset threshold or any value of the confidence factor is lower than the preset tolerance value, the current state is determined to be abnormal, and the emergency liquid level correction process is initiated. The emergency liquid level correction process includes: reducing the sampling frequency and extending the sliding window stabilization period, temporarily shielding the channel with the lowest confidence and increasing the confidence weight of other channels, or introducing the moving average trend of historical liquid level stabilization sequences for short-term compensation.

9. A gas-liquid interface identification and ship sloshing level correction system using the liquid level correction method according to any one of claims 1-8, characterized in that, include: The sensor assembly includes a liquid level sensor, an image acquisition unit, and an IMU inertial measurement unit. The liquid level sensor is used to acquire raw differential pressure signal data, capacitance signal data, and radar signal data of the target container. The image acquisition unit is used to acquire image data inside the target container. The IMU inertial measurement unit is used to acquire three-axis attitude angle and acceleration data of the target container. A control unit, connected to the sensor assembly, the control unit comprising: The image processing module is used to extract edge information and grayscale texture features contained in real-time acquired image data and use a trained lightweight convolutional neural network model to classify the liquid-gas interface state, and generate an acquisition strategy based on the liquid-gas interface state classification results. The data acquisition module is used to acquire image data, and also to acquire real-time raw liquid level signals according to the acquisition strategy and simultaneously acquire IMU inertial measurement unit data. The stable reconstruction module is used to process the original liquid level signal using a preset decomposition method, generate multi-modal components, construct a liquid surface disturbance prediction model based on the IMU inertial measurement unit data, identify the main disturbance modal components, filter out the disturbance-related modal components, and reconstruct a stable liquid level signal. The liquid level correction module is used to model the filtered disturbance modal components as disturbance residual sequences and construct residual time series prediction models to predict liquid level disturbance trends and realize dynamic correction of liquid level, and output dynamically corrected liquid level data.

Images