An energy efficiency optimization method for diving equipment based on environmental data prediction

Abstract

This application belongs to the technical field of energy efficiency optimization of diving equipment, and relates to an energy efficiency optimization method for diving equipment based on environmental data prediction. By collecting key environmental data of the diving equipment and performing standardized processing, a three-dimensional spatio-temporal data matrix is constructed based on the standardized data set, and the spatial correlation between sensor nodes is modeled through feature extraction, so as to accurately capture the spatio-temporal features in the data. The LSTM network is used for time series prediction to extract the non-linear time series features of environmental data, and the physical channel and the data channel are fused through a dual-channel structure, overcoming the errors caused by the traditional method's neglect of environmental dynamic changes. The fusion strategy enables the model to not only cope with complex spatio-temporal changes, but also realize the complementary advantages of theoretical knowledge and data-driven methods; by obtaining accurate predicted values of the environmental state, the subsequent multi-objective optimization algorithm can efficiently generate an optimal control sequence to achieve the optimization of the energy efficiency of the diving equipment.

Description

Technical Field

[0001] This application belongs to the technical field of energy efficiency optimization of diving equipment. More specifically, it relates to an energy efficiency optimization method for diving equipment based on environmental data prediction. Background Art

[0002] In the control and optimization process of modern diving equipment (such as underwater robots, submersibles, etc.), the complexity and uncertainty of the environment greatly affect the energy efficiency and operating performance of the equipment. Diving equipment usually needs to perform tasks in variable environments such as the ocean, and environmental factors such as water depth, salinity, and flow velocity will have a significant impact on the performance of the equipment. For example, changes in the intensity and direction of water flow, differences in water temperature and salinity, as well as the attitude and operating state of the equipment, will all cause fluctuations in the energy efficiency of the equipment. Therefore, how to accurately predict the impact of these dynamic environmental factors on the equipment and perform precise control and optimization based on these predicted values has become a key issue in improving the energy efficiency of diving equipment.

[0003] To effectively solve this problem, traditional methods usually rely on experience-based control strategies or single prediction models, which often cannot fully consider the complexity and spatio-temporal variability of environmental factors.

[0004] For example, physical models can theoretically provide guidance for the operation of the equipment, but their complexity and accuracy requirements make it difficult to provide accurate control instructions in real time when applied in variable ocean environments.

[0005] Secondly, when using machine learning models (such as deep neural networks) for environmental prediction and control optimization, problems such as data scarcity and difficulty in capturing spatio-temporal correlations are usually faced. Especially in the ocean environment, the sensor data obtained is often affected by noise and environmental changes, which makes the direct data-based models perform unsatisfactorily when dealing with these problems. Summary of the Invention

[0006] The present invention provides an energy efficiency optimization method for diving equipment based on environmental data prediction, aiming to solve the technical problem that traditional methods rely on experience-based control strategies or single prediction models and cannot fully consider the complexity and spatio-temporal variability of environmental factors.

[0007] An energy efficiency optimization method for diving equipment based on environmental data prediction includes the following steps:

[0008] Step 1: Collect data on depth, salinity, depth, three-dimensional flow velocity, equipment attitude, and equipment operating parameters, and perform preprocessing based on the collected data to obtain a standardized data set;

[0009] Step 2: Based on the standardized dataset, construct a three-dimensional spatio-temporal data matrix, and then perform feature extraction on the three-dimensional spatio-temporal data matrix to model the spatial correlation between sensor nodes, obtaining spatio-temporal features;

[0010] Step 3: Based on the spatio-temporal features, use the LSTM network for time series prediction, extract the non-linear time series features of the data, fuse the hydrodynamic equation with the data-based LSTM model, and optimize through a two-channel structure, where the two channels include a physical channel and a data channel. The physical channel provides theoretical values, and the data channel provides predicted values of actual data. By fusing the values of the two channels, obtain the environmental state prediction value;

[0011] Step 4: Use the environmental state prediction value as the input of the multi-objective optimization algorithm, and iteratively output the optimal control sequence based on the multi-objective optimization algorithm.

[0012] In the present invention, by collecting and standardizing the key environmental data of the diving equipment, then, based on the standardized dataset, a three-dimensional spatio-temporal data matrix is constructed, and the spatial correlation between sensor nodes is modeled through feature extraction, thus accurately capturing the spatio-temporal features in the data. Using the LSTM network for time series prediction, the non-linear time series features of the environmental data are successfully extracted, and through a two-channel structure, the physical channel (theoretical value based on the hydrodynamic equation) and the data channel (predicted value based on historical data) are fused, overcoming the errors caused by the traditional method's neglect of environmental dynamic changes, thereby achieving accurate prediction of the environmental state; the fusion strategy enables the model to not only cope with complex spatio-temporal changes but also realize the complementary advantages of theoretical knowledge and data-driven methods; by obtaining accurate environmental state prediction values, the subsequent multi-objective optimization algorithm can efficiently generate the optimal control sequence to optimize the energy efficiency of the diving equipment.

[0013] Preferably, the preprocessing includes performing distribution analysis on the data using the K-S test based on the collected data, then checking the physical rationality of the data in combination with the hydrodynamic equation, removing abnormal data to obtain a credible dataset, and then standardizing the data within each time window in the credible dataset to obtain the standardized data.

[0014] Preferably, the specific steps to obtain the credible dataset are as follows:

[0015] Benchmark distribution construction: Based on the continuous historical data of the equipment in the normal state in the past n hours, use the kernel density estimation method to calculate the probability density of each data point, smooth the data to obtain the probability distribution function of the data, and obtain the benchmark distribution based on the probability distribution function;

[0016] Empirical distribution function calculation: Calculate the empirical distribution function based on the real-time data of the sensor within a predetermined time window;

[0017] Calculation of K-S statistic: Calculate the K-S statistic based on the reference distribution and the empirical distribution function to measure the maximum difference between the reference distribution and the real-time data distribution, and obtain the K-S statistic;

[0018] Initial anomaly judgment: Compare the K-S statistic with a preset significance threshold. If the K-S statistic is greater than the significance threshold, the data is determined to be abnormal data, and physical constraints are executed. If the K-S statistic is less than the significance threshold, the data is retained;

[0019] Physical constraints: Calculate the real-time seawater density according to the physical properties of the fluid;

[0020] Calculation of the time variation and gradient of fluid flow: Calculate the change of fluid density over time and the gradient in space to obtain the density change rate and the fluid convection term;

[0021] Residual calculation and physical consistency verification: Add the calculated density change rate and the fluid convection term to obtain the residual difference. If the calculated residual difference is greater than the physical threshold, it indicates a physical anomaly, which is marked as abnormal data and repaired. If the residual difference is less than or equal to the physical threshold, it is determined to be a suspicious item, and the suspicious item is manually reviewed.

[0022] Preferably, the time window described in step 1 is adjusted by a dynamic adjustment method, including:

[0023] Three-dimensional vector synthesis: Convert the three-dimensional flow velocity data at each moment into a vector modulus value to obtain a vector modulus value sequence;

[0024] Standard deviation calculation: Calculate the window standard deviation based on the vector modulus value sequence;

[0025] Window dynamic adjustment: Perform window dynamic adjustment based on the calculated window standard deviation: ;

[0026] Where: represents the window length at the current moment; represents the minimum value of the window length; represents the maximum value of the window length; represents the lower limit of the flow velocity standard deviation; represents the upper limit of the flow velocity standard deviation; represents the standard deviation at the current moment; represents a limiting function that limits the value between and . If the calculated is less than then output . If the calculated is greater than Then the output ;

[0027] Based on the calculated window length , adjust the data processing window lengths of all sensors. When the change in window length results in insufficient data points for non-flow sensors, cubic spline interpolation is used for filling.

[0028] Preferably, in step 2, a three-dimensional spatio-temporal data matrix in three dimensions of time, space, and channel is constructed by using the time step, the number of spatial sensor nodes, and the data characteristics of each sensor. Then, dilated causal convolution and graph convolutional network are used to extract features from the three-dimensional spatio-temporal data matrix, model the spatial correlation between sensor nodes, and obtain spatio-temporal features.

[0029] Preferably, obtaining the spatio-temporal features includes the following steps:

[0030] Spatial topology modeling: Calculate the distance matrix based on the physical distances between sensors, construct an adjacency matrix based on the distance matrix to represent the connection strength between nodes, then introduce a flow direction correction factor, calculate the angle between the connection direction between nodes and the main flow direction, and adjust the weights of adjacent adjacency matrices according to the flow direction to obtain the adjusted adjacency matrix with weights; ]>

[0031] Multi-modal data fusion: Organize the standardized data into a three-dimensional tensor according to the dimensions of time, space, and channel. The data in each time window is retained in different channels to obtain the fused three-dimensional tensor;

[0032] Spatio-temporal convolutional encoding: Use dilated causal convolution in the time dimension to extract short-term and long-term time series features by gradually increasing the dilation rate of the convolution sum to obtain time series features. In the space dimension, calculate the adjacency matrix and the connection information of nodes through graph convolution to extract the patterns in the spatial topology structure to obtain spatial features;

[0033] Dynamic weight adjustment: Calculate the turbulence intensity of the environment, measure the degree of dynamic change of the environment by calculating the fluctuation of the flow velocity, and dynamically adjust the fusion ratio of the time series features and the spatial features according to the turbulence intensity, thereby obtaining the spatio-temporal feature fusion result.

[0034] Preferably, step 3 includes the following steps:

[0035] Physical prior knowledge embedding: Calculate the energy dissipation by using the Navier-Stokes equation by extracting the flow velocity tensor components from the spatio-temporal features;

[0036] Fluid resistance formula embedding: Using the flow velocity and the cross-sectional area facing the flow, combined with the resistance coefficient of the equipment shape and the seawater density, calculate the resistance of the equipment; use the Sigmoid function to smooth the influence of the resistance, and combine the smoothed resistance with the bias term of the LSTM network to adjust the initial state of the LSTM network;

[0037] Correction factor calculation: By calculating the ratio of the theoretical energy consumption to the actual power, obtain the physical correction factor for correcting the predicted value of the physical channel;

[0038] Physical channel output: At each moment, according to the current speed and position, solve the kinematic differential equation of the equipment by the fourth-order Runge-Kutta method to obtain the predicted value of the theoretical trajectory at the next moment;

[0039] Data channel output: The spatio-temporal features are fed into the LSTM network after embedding based on the fluid resistance formula. The encoder of the LSTM network processes the time series data step by step to generate the hidden state at each moment. The hidden state is fed into the decoder, and the decoder outputs the prediction result of the data channel through the fully connected layer;

[0040] Fusion weight adjustment: Dynamically adjust the fusion weight of the physical channel output and the data channel data based on the magnitude of the flow velocity variance; after correcting the output of the physical channel by introducing the physical correction factor, fuse the corrected physical channel output and the data channel output based on the fusion weight to obtain the fused prediction result; where the adjustment logic of the flow velocity variance is that the larger the flow velocity variance, the larger the fusion weight of the data channel output, and the smaller the flow velocity variance, the larger the fusion weight of the physical channel output.

[0041] Preferably, the objective function of the multi-objective optimization algorithm is as follows: ;

[0042] In the formula: Both represent weight parameters, and their sum is 1; Represents the motor output power at the k-th time step; Represents the predicted equipment position at the k-th time step; Represents the task planning target position at the k-th time step; Represents the equipment heading angle at the k-th time step; Represents the second derivative of the heading angle; Represents the time interval of the control period.

[0043] Preferably, the objective function of the multi-objective optimization algorithm includes power constraint, equipment speed constraint, equipment acceleration constraint, and soft constraint; the soft constraint is to introduce a slack variable and a deviation tolerance threshold to constrain the path deviation.

[0044] Preferably, the multi-objective optimization algorithm adopts MPC rolling optimization, converts the objective function and constraint conditions into a standard non-linear programming problem, and outputs it as a problem description file, which is used for MPC rolling optimization;

[0045] When performing MPC rolling optimization, first dynamically adjust the optimization time domain length according to the predicted result of the real-time turbulence intensity. If the turbulence intensity is less than the threshold, increase the time domain; if the turbulence intensity is greater than the threshold, shorten the time domain. The time domain length determines the number of future time steps to be considered for each optimization;

[0046] During the rolling optimization process, the optimal control instruction is obtained by solving through the interior point method. According to the output control instruction, the first set of control quantities is executed, and the remaining control sequences are reserved for the next cycle's warm start.

[0047] The beneficial effects of the present invention include:

[0048] The present invention collects key environmental data of the diving equipment and performs standardization processing. Then, based on the standardized data set, a three-dimensional spatio-temporal data matrix is constructed, and the spatial correlation between sensor nodes is modeled through feature extraction, so as to accurately capture the spatio-temporal features in the data. The LSTM network is used for time series prediction, and the non-linear time series features of the environmental data are successfully extracted. Through the dual-channel structure, the physical channel (the theoretical value based on the hydrodynamic equation) and the data channel (the predicted value based on historical data) are fused, overcoming the errors caused by the traditional method's neglect of environmental dynamic changes, thus realizing the accurate prediction of the environmental state. The fusion strategy enables the model to not only cope with complex spatio-temporal changes but also achieve the complementary advantages of theoretical knowledge and data-driven methods. By obtaining accurate predicted values of the environmental state, the subsequent multi-objective optimization algorithm can efficiently generate the optimal control sequence to optimize the energy efficiency of the diving equipment. Description of the Drawings

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0050] Figure 1 It is the overall step block diagram provided by the embodiment of the present invention.

[0051] Figure 2 It is the preprocessing step block diagram provided by the embodiment of the present invention.

[0052] Figure 3 It is the step 3 step schematic block diagram provided by the embodiment of the present invention. Detailed implementation manners

[0053] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0054] See Figure 1 As shown, an energy efficiency optimization method for a diving device based on environmental data prediction includes the following steps:

[0055] Step 1: Collect data on depth, salinity, depth, three-dimensional flow velocity, device attitude, and device operation parameters, and perform preprocessing on the collected data to obtain a standardized data set;

[0056] As a possible implementation manner of this embodiment, the preprocessing includes performing distribution analysis on the data using the K-S test based on the collected data, and then checking the physical rationality of the data in combination with the hydrodynamic equation, removing abnormal data to obtain a credible data set, and then standardizing the data in each time window in the credible data set to obtain standardized data.

[0057] See Figure 2 As shown, as a possible implementation manner of this embodiment, the specific steps for obtaining the credible data set are as follows:

[0058] Benchmark distribution construction: Based on the continuous historical data of the device in the normal state in the past n hours, use the kernel density estimation method to calculate the probability density of each data point, smooth the data to obtain the probability distribution function of the data, and obtain the benchmark distribution based on the probability distribution function; ;

[0059] In the formula: n represents the total number of data points; h represents the bandwidth; K represents the kernel function, such as the Gaussian kernel function; represents the probability density function, the probability density at the given point x; represents the value of the i-th data point; x represents the value of the point where the probability density is desired to be estimated;

[0060] Among them, h is selected using Silverman's rule to ensure the best smoothing effect, for example: ;

[0061] In the formula: is the interquartile range, measuring the dispersion degree of the data; represents the total number of data points;

[0062] Empirical distribution function calculation: Calculate the empirical distribution function based on the real-time data of the sensor within a predetermined time window (such as the flow velocity data in the past minute); ;

[0063] In the formula: represents the indicator function, when is true, the value is 1, otherwise it is 0; represents the empirical distribution function, which is the proportion of data points less than or equal to x; n represents the total number of data points; represents the value of the i-th data point; x represents the point at which the empirical distribution function is calculated;

[0064] Describe the cumulative distribution of real-time data through the empirical distribution function, indicating the proportion of data points that appear before a certain value x;

[0065] Kolmogorov-Smirnov (K-S) statistic calculation: Calculate the K-S statistic based on the reference distribution and the empirical distribution function to measure the maximum difference between the reference distribution and the real-time data distribution, and obtain the K-S statistic ; ;

[0066] In the formula: is the cumulative distribution function of the reference distribution; represents taking the supremum over the range of x; By calculating the value of the K-S statistic to quantify the difference between the real-time data and the reference distribution;

[0067] Preliminary anomaly judgment: Compare the K-S statistic with a preset significance threshold. If the K-S statistic is greater than the significance threshold, it is determined as abnormal data, and physical constraints are executed. If the K-S statistic is less than the significance threshold, the data is retained; The significance threshold is calculated based on the following formula: ;

[0068] In the formula: represents a constant; represents the significance threshold; n represents the total number of data points; represents the significance level, which is used to control the strictness of anomaly judgment;

[0069] Physical constraint: Calculate the real-time seawater density according to the physical properties of the fluid; ;

[0070] In the formula: is the standard seawater density; represents the density change caused by the changes in temperature T, salinity S, and pressure p; represents the real-time seawater density;

[0071] Temporal Variation and Gradient Calculation of Fluid Flow: Calculate the temporal variation of fluid density and its gradient in space to obtain the density change rate and the fluid convection term. The formula for calculating the density change rate is as follows: ;

[0072] In the formula: represents the time step; represents the density at the current moment; represents the density change rate; represents the seawater density at time ;

[0073] The formula for calculating the convection term is as follows: ;

[0074] In the formula: and are the densities of adjacent nodes; and are the flow velocities of adjacent nodes; and represent the cross-sectional areas of adjacent nodes; represents the distance between adjacent nodes; represents the convection term of the fluid, which is the spatial variation of the combination of fluid density and velocity;

[0075] Residual Calculation and Physical Consistency Verification: Add the calculated density change rate and the fluid convection term to obtain the residual ; ;

[0076] In the formula: represents the density change rate; represents the convection term;

[0077] If the calculated residual is greater than the physical threshold, it indicates a physical anomaly, which is marked as abnormal data and repaired; if the residual is less than or equal to the physical threshold, it is determined as a suspicious item, and the suspicious item is manually reviewed.

[0078] The above judgment logic for abnormal data is only an exemplary technical solution of this application. It is also possible to simultaneously judge statistical anomalies and physical anomalies, and any data that meets either anomaly will be determined as abnormal data; therefore, the above judgment logic should not be regarded as a limitation of the present invention.

[0079] In this embodiment, the kernel density estimation method can effectively smooth data fluctuations, enabling the model to extract long-term trends and patterns from the data, and being less susceptible to noise. Moreover, the K-S statistic provides a quantitative measure that can clearly reveal the difference between real-time data and the reference distribution. When the K-S statistic exceeds the significance threshold, the abnormal data can be clearly identified. It is objective and operable, avoiding subjectivity in traditional empirical judgments. Secondly, through physical consistency verification and real-time data repair, the high quality and reliability of the data are ensured, enhancing the safety of equipment operation.

[0080] As a possible implementation of this embodiment, the time window described in step 1 is adjusted by a dynamic adjustment method, including:

[0081] Three-dimensional vector synthesis: Convert the three-dimensional flow velocity data at each moment into a vector modulus value to obtain a vector modulus value sequence;

[0082] Standard deviation calculation: Calculate the window standard deviation based on the vector modulus value sequence;

[0083] Window dynamic adjustment: Perform window dynamic adjustment based on the calculated window standard deviation: ;

[0084] Where: represents the window length at the current moment; represents the minimum value of the window length; represents the maximum value of the window length; represents the lower limit of the flow velocity standard deviation; represents the upper limit of the flow velocity standard deviation; represents the standard deviation at the current moment; represents the limiting function, which limits the value between and . If the calculated is less than , then output . If the calculated is greater than and ;

[0085] Based on the calculated window length adjust the data processing window lengths of all sensors. When the change in the window length results in insufficient data points for non-flow velocity sensors, cubic spline interpolation is used for filling.

[0086] Step 2: Construct a three-dimensional spatio-temporal data matrix based on the standardized data set, and then perform feature extraction on the three-dimensional spatio-temporal data matrix to model the spatial correlation between sensor nodes and obtain spatio-temporal features;

[0087] As a possible implementation of this embodiment, in step 2, a three-dimensional spatio-temporal data matrix in three dimensions of time, space, and channel is constructed using the time step, the number of spatial sensor nodes, and the characteristics of each sensor data. Then, dilated causal convolution and a graph convolutional network are used to extract features from the three-dimensional spatio-temporal data matrix, model the spatial correlation between sensor nodes, and obtain spatio-temporal features.

[0088] As a possible implementation of this embodiment, obtaining the spatio-temporal features includes the following steps:

[0089] Before performing spatial topology modeling, it is necessary to cut the continuous data stream into overlapping time windows to balance temporal continuity and computational efficiency. Among them, the time window inherits the adaptive window length in the anomaly detection stage, and the sliding step retains an overlap rate of 20% to ensure temporal coherence; and mirror symmetry padding is used to process the window boundary data;

[0090] Spatial topology modeling: Calculate the distance matrix based on the physical distance between sensors, construct an adjacency matrix based on the distance matrix to represent the connection strength between nodes, then introduce a flow direction correction factor, calculate the angle between the connection direction between nodes and the main flow direction, and adjust the weights of adjacent adjacency matrices according to the flow direction to obtain the adjusted adjacency matrix; ;

[0091] In the formula: represents the angle between the node connection direction and the main flow direction; represents the connection strength between the original sensor nodes i and j; represents the connection strength between nodes i and j after flow direction correction;

[0092] Multi-modal data fusion: Organize the standardized data into a three-dimensional tensor according to the dimensions of time, space, and channel. The data in each time window is retained in different channels to obtain the fused three-dimensional tensor. For example , where T represents the time step; S represents the number of spatial nodes; C represents the number of channels;

[0093] Spatio-temporal convolutional encoding: Includes time convolution and graph convolution;

[0094] Time convolution: Use dilated causal convolution to capture multi-scale temporal patterns: ;

[0095] In the formula: K represents the convolution kernel size; represents the result of time convolution, which is the output value at time t; represents the weight of the convolution kernel; represents the value at time where d represents the dilation factor; K represents the size of the convolution kernel; represents the time step;

[0096] Graph Convolution: Based on the Chebyshev polynomial approximation graph convolution: ;

[0097] In the formula: represents the scaled Laplacian matrix; represents the Chebyshev polynomial basis; K = 3 is the polynomial order; represents the output result of the graph convolution; represents the Chebyshev polynomial basis weights; represents the input data;

[0098] Dynamic Weight Adjustment: Calculate the turbulence intensity of the environment : ;

[0099] In the formula: T represents the time step; represents the flow velocity value at time t; represents the average value of the flow velocity;

[0100] Measure the dynamic change degree of the environment by calculating the fluctuation of the flow velocity, dynamically adjust the fusion ratio of the temporal feature and the spatial feature according to the turbulence intensity, and then obtain the spatio-temporal feature fusion result; ;

[0101] In the formula: k represents the adjustment slope; represents the feature turning threshold; represents the dynamically adjusted weight factor; e represents the base of the natural logarithm; ;

[0102] In the formula: represents the fused spatio-temporal feature; represents the temporal feature obtained by temporal convolution; represents the spatial feature obtained by graph convolution.

[0103] In this embodiment, the Dilated Causal Convolution can capture short-term and long-term temporal features without increasing the computational complexity by gradually increasing the convolution dilation rate, can effectively process the information with a long time span in the temporal data, and at the same time avoids the information loss problem caused by too large a convolution kernel size; by calculating the adjacency matrix through graph convolution, the spatial topological relationship between sensor nodes can be fully considered, so as to extract spatial features. Therefore, by combining the Dilated Causal Convolution and graph convolution, the model can obtain effective features in both the time and space dimensions, improving the processing ability of multi-dimensional data.

[0104] Step 3: Use the LSTM network for time series prediction based on spatio-temporal features, extract the non-linear time series features of the data, fuse the hydrodynamic equations with the data-based LSTM model, and optimize through a dual-channel structure, where the dual-channel includes a physical channel and a data channel. The physical channel provides theoretical values, and the data channel provides predicted values of actual data. By fusing the values of the two channels, the predicted value of the environmental state is obtained;

[0105] See Figure 3 As shown, as a possible implementation of this embodiment, step 3 includes the following steps:

[0106] Embedding of physical prior knowledge: Calculate the energy dissipation using the Navier-Stokes equation by extracting the velocity tensor components from the spatio-temporal features; ;

[0107] In the formula: represents the theoretical energy dissipation rate; represents the seawater dynamic viscosity, which is obtained by looking up the temperature-salinity table; and both represent the velocity tensor components, where i and j are indices; represents the seawater density; represents the acceleration due to gravity; represents the current water depth; represents the vertical component of the flow velocity; represents the flow velocity component with respect to the spatial coordinate partial derivative; represents the flow velocity component with respect to the spatial coordinate partial derivative;

[0108] In this embodiment, by introducing the energy dissipation calculation, effects such as friction and turbulence in fluid flow can be accurately simulated, which is very important for fluid mechanics related problems; the embedding of physical priors helps to accurately model the system dynamics, thereby improving the prediction accuracy.

[0109] Embedding of fluid resistance formula: Use the flow velocity and the cross-sectional area facing the flow, and calculate the resistance of the device in combination with the resistance coefficient of the device shape and the seawater density; ;

[0110] In the formula: represents the relative flow velocity magnitude; represents the resistance coefficient related to the device shape; A represents the cross-sectional area facing the flow, which is dynamically calculated according to the attitude angle of the device; represents the fluid resistance; represents the seawater density;

[0111] Use the Sigmoid function to smooth the influence of the resistance, combine the smoothed resistance with the bias term of the LSTM network, and adjust the initial state of the LSTM network; ;

[0112] In the formula: represents the initial bias term of the LSTM network; represents the learnable parameter; represents using the Sigmoid activation function to map the resistance to a range for network training.

[0113] In this embodiment, combining the smoothed resistance with the bias term of the LSTM network helps to introduce physical constraints during network initialization, reduce the uncertainty during the training process, thereby accelerating the convergence of the network and improving the model training efficiency and prediction performance.

[0114] Correction factor calculation: Calculate the physical correction factor by calculating the ratio of the theoretical energy consumption to the actual power, and use it to correct the predicted value of the physical channel; ;

[0115] In the formula: represents the actual propulsion power; represents the space integration of the energy consumption rate to obtain the overall energy consumption, where , , respectively represent the small increments in the x, y, and z directions in space;

[0116] In this embodiment, the ratio between the theoretical energy consumption and the actual power reflects the energy efficiency deviation of the device in a specific environment; by calculating the correction factor and applying it to the output of the physical channel, the error caused by environmental factors (such as changes in flow velocity, temperature, etc.) can be reduced, thereby making the energy consumption prediction more accurate.

[0117] Physical channel output: At each moment, according to the current speed and position, use the fourth-order Runge-Kutta method to solve the kinematic differential equation of the device to obtain the predicted value of the theoretical trajectory at the next moment;

[0118] Among them, the kinematic differential equation of the device is as follows: ;

[0119] In the formula: x represents the position of the device; v represents the speed of the device; , represents the propulsion force, represents the propulsion efficiency coefficient, calibrated based on experiments; represents the real-time propulsion power of the device; m represents the mass of the device; represents the angular velocity; represents the derivative with respect to time;

[0120] The fourth-order Runge-Kutta method is used to numerically solve the kinematic differential equation of the device to obtain the predicted value of the theoretical trajectory ;

[0121] Data channel output: The spatio-temporal features are used as inputs and fed into the LSTM network embedded based on the fluid resistance formula. The encoder of the LSTM network gradually processes the time series data to generate the hidden state at each moment. The hidden state is fed into the decoder, and the decoder outputs the prediction result of the data channel through the fully connected layer;

[0122] Exemplarily, the LSTM network includes an input gate, a forget gate, and an output gate, and each gate includes a bias term. The fluid resistance formula can be embedded into these bias terms, that is, based on the calculation formula of is embedded into the bias term of each gate;

[0123] For example: ;

[0124] In the formula: represents the original bias term of the input gate; represents the bias term of the input gate after embedding the fluid resistance;

[0125] Referring to the above formula, the fluid resistance can be embedded into other gates;

[0126] Based on this, it is ensured that the resistance information in the physical model can affect the initialization of the LSTM network, so that the model can combine the physical background at the beginning of learning and adapt to the changes of fluid dynamics faster.

[0127] Fusion weight adjustment: Based on the magnitude of the flow velocity variance, dynamically adjust the fusion weight of the physical channel output and the data channel data; after correcting the output of the physical channel by introducing the physical correction factor, fuse the corrected physical channel output and the data channel output based on the fusion weight to obtain the fused prediction result; the adjustment logic of the flow velocity variance is that the larger the flow velocity variance, the larger the fusion weight of the data channel output, and the smaller the flow velocity variance, the larger the fusion weight of the physical channel output; the exemplary calculation formula is as follows: ;

[0128] In the formula: represents the scaling factor; represents the flow velocity variance, which characterizes the turbulence intensity; the fusion weight at time t; represents the base of the natural logarithm;

[0129] In this embodiment, the flow velocity variance, as a measure of environmental changes, can reflect the stability of the fluid environment; by dynamically adjusting the fusion weights of the physical channel and the data channel based on the flow velocity variance, more reliance can be placed on the prediction results of the data channel when the environment fluctuates greatly, while more reliance can be placed on the prediction results of the physical channel when the environment is stable, which can improve the accuracy and robustness of the prediction.

[0130] Output the final prediction result based on the calculated dynamic weight , and the calculation formula is as follows: ;

[0131] In the formula: represents the prediction result of the data channel at time t; represents the prediction result of the physical channel at time t.

[0132] Step 4: Use the environmental state prediction value as the input of the multi-objective optimization algorithm, and iteratively output the optimal control sequence based on the multi-objective optimization algorithm;

[0133] As a possible implementation of this embodiment, the objective function of the multi-objective optimization algorithm is as follows: ;

[0134] In the formula: , , all represent weight parameters, and their sum is 1; represents the motor output power at the k-th time step; represents the predicted device position at the k-th time step; represents the task planning target position at the k-th time step; represents the device heading angle at the k-th time step; represents the second derivative of the heading angle at the k-th time step; the time interval of the control period; ;

[0135] In the formula: represents the seawater density; represents the drag coefficient; A represents the cross-sectional area of the device; represents the speed of the device; represents the flow velocity at position k in the predicted flow velocity field; represents the motor power at the k-th time step;

[0136] The objective function of the multi-objective optimization algorithm includes power constraint, device speed constraint, device acceleration constraint and soft constraint; the soft constraint is to introduce a relaxation variable and a deviation tolerance threshold to constrain the path deviation;

[0137] Exemplarily, the power constraint shall not exceed the maximum power of the device; if there is a minimum power constraint, the minimum power can also be added to the power constraint; secondly, for the device speed constraint, it shall not exceed the maximum speed and minimum speed of the device; the acceleration constraint of the device is that it shall not exceed the maximum acceleration of the device;

[0138] Among them, the expression form of the soft constraint is as follows: ;

[0139] In the formula: represents the allowable path deviation threshold; represents the slack variable of the path deviation, which is included in the penalty term in the objective function; represents the predicted device position at the k-th time step; represents the task planning target position at the k-th time step;

[0140] The multi-objective optimization algorithm adopts MPC rolling optimization, converts the objective function and constraint conditions into a standard nonlinear programming problem, and outputs it as a problem description file, and uses the problem description file for MPC rolling optimization;

[0141] When performing MPC rolling optimization, first dynamically adjust the optimization time domain length according to the real-time prediction result of the turbulence intensity. If the turbulence intensity is less than the threshold, increase the time domain; if the turbulence intensity is greater than the threshold, shorten the time domain. The time domain length determines the number of future time steps to be considered for each optimization. Exemplary expressions are as follows: ;

[0142] In the formula: represents the variance of the predicted flow velocity field; the optimization time domain length; t represents the current time;

[0143] During the rolling optimization process, the optimal control instruction is obtained by solving with the interior point method. According to the output control instruction, the first set of control quantities is executed, and the remaining control sequence is reserved for the next cycle's warm start. Exemplary expressions of the interior point method are as follows: ; where, u represents the control variable (propulsion power, rudder angle, etc.); x represents the state variable (position, speed, heading angle, etc.); p represents the environmental parameter (predicted flow velocity, density, etc.); represents the objective function; represents the inequality constraint, such as power constraint, speed constraint, etc.; represents the equality constraint, such as path tracking constraint; for the constraint conditions, they are only exemplary, and the constraint conditions can be dynamically adjusted and optimized according to the actual situation, and different constraint conditions can be added according to different requirements.

[0144] Furthermore, in order to ensure real-time performance, the computational efficiency is improved by sparse matrix of the Jacobian matrix during the solving process: ;

[0145] In the formula: represents the sparsified Jacobian matrix; represents the partial derivative of the objective function q with respect to the control variable u; represents the partial derivative of the constraint function h with respect to the state variable x;

[0146] Among them, the proportion of non - zero elements being less than 15% significantly reduces the computational amount, compressing the single - step solution time to within 200 ms to ensure the real - time response of the system.

[0147] Sparsification of the Jacobian matrix is a commonly used technique in optimization algorithms. Therefore, details of how to sparsify and how to calculate are not elaborated in this application.

[0148] In this embodiment, the interior - point method transforms the optimization problem into a series of linearized problems, solves for the optimal control instruction by iteratively calculating the control input updated each time, and in order to accelerate the calculation, non - zero elements are retained by sparsifying the Jacobian matrix to reduce the computational amount.

[0149] The above are only the preferred embodiments of this application and are not used to limit this application. Any modifications, equivalent replacements, and improvements made within the spirit and principle of this application shall be included within the protection scope of this application.

Claims

1. An energy efficiency optimization method for diving equipment based on environmental data prediction, characterized in that, It includes the following steps: Step 1: Collect data on depth, salinity, depth, three-dimensional flow velocity, device attitude, and device operating parameters, and perform preprocessing on the collected data to obtain a standardized data set; Step 2: Based on the standardized data set, construct a three-dimensional spatio-temporal data matrix, and then perform feature extraction on the three-dimensional spatio-temporal data matrix to model the spatial correlation between sensor nodes and obtain spatio-temporal features; Step 3: Use an LSTM network for time series prediction based on the spatio-temporal features, extract the non-linear time series features of the data, fuse the hydrodynamic equations with the data-based LSTM model, and optimize through a dual-channel structure, where the dual-channel includes a physical channel and a data channel. The physical channel provides the theoretical value, and the data channel provides the predicted value of the actual data. By fusing the values of the two channels, the predicted value of the environmental state is obtained; Physical prior knowledge embedding: Calculate the energy dissipation by using the Navier-Stokes equation by extracting the velocity tensor components from the spatio-temporal features; Fluid resistance formula embedding: Use the flow velocity and the cross-sectional area facing the flow, combine the resistance coefficient of the device shape and the seawater density to calculate the resistance of the device; Use the Sigmoid function to smooth the influence of the resistance, and combine the smoothed resistance with the bias term of the LSTM network to adjust the initial state of the LSTM network; Correction factor calculation: Calculate the physical correction factor by calculating the ratio of the theoretical energy consumption to the actual power, which is used for the correction of the predicted value of the physical channel; Physical channel output: At each moment, according to the current speed and position, solve the kinematic differential equation of the device by the fourth-order Runge-Kutta method to obtain the predicted value of the theoretical trajectory at the next moment; Data channel output: The spatio-temporal features are used as input and fed into the LSTM network after the fluid resistance formula is embedded. The encoder of the LSTM network gradually processes the time series data to generate the hidden state at each moment. The hidden state is fed into the decoder, and the decoder outputs the prediction result of the data channel through the fully connected layer; Fusion weight adjustment: Dynamically adjust the fusion weight of the physical channel output and the data channel data based on the magnitude of the flow velocity variance; After correcting the output of the physical channel by introducing the physical correction factor, fuse the corrected physical channel output and the data channel output based on the fusion weight to obtain the fused prediction result; The adjustment logic of the flow velocity variance is that the larger the flow velocity variance, the larger the fusion weight of the data channel output, and the smaller the flow velocity variance, the larger the fusion weight of the physical channel output; Step 4: Use the predicted value of the environmental state as the input of the multi-objective optimization algorithm, and iteratively output the optimal control sequence based on the multi-objective optimization algorithm.

2. The energy efficiency optimization method of a diving device based on environmental data prediction according to claim 1, wherein, The preprocessing includes performing distribution analysis on the data using the K-S test based on the collected data, and then checking the physical rationality of the data in combination with the hydrodynamic equations, removing abnormal data to obtain a credible data set, and then standardizing the data within each time window in the credible data set to obtain the standardized data.

3. The energy efficiency optimization method of a diving device based on environmental data prediction according to claim 2, characterized in that The specific steps to obtain the credible data set are as follows: Benchmark distribution construction: Based on the continuous historical data of the device in the normal state in the past n hours, the kernel density estimation method is used to calculate the probability density of each data point, smooth the data, obtain the probability distribution function of the data, and obtain the benchmark distribution based on the probability distribution function; Empirical distribution function calculation: Calculate the empirical distribution function based on the real-time sensor data within a predetermined time window; K-S statistic calculation: Calculate the K-S statistic based on the benchmark distribution and the empirical distribution function, measure the maximum difference between the benchmark distribution and the real-time data distribution, and obtain the K-S statistic; Preliminary anomaly judgment: Compare the K-S statistic with a preset significance threshold. If the K-S statistic is greater than the significance threshold, it is determined as abnormal data, and physical constraints are executed. If the K-S statistic is less than the significance threshold, the data is retained; Physical constraint: Calculate the real-time seawater density according to the physical properties of the fluid; Calculation of the time variation and gradient of fluid flow: Calculate the change of fluid density over time and the gradient in space to obtain the density change rate and the fluid convection term; Residual calculation and physical consistency verification: Add the calculated density change rate and fluid convection term to obtain the residual. If the calculated residual is greater than the physical threshold, it indicates a physical anomaly, which is marked as abnormal data and repaired; if the residual is less than or equal to the physical threshold, it is determined as a suspicious item, and the suspicious item is manually reviewed.

4. The energy efficiency optimization method of a diving device based on environmental data prediction according to claim 2, characterized in that The time window described in step 1 is adjusted by a dynamic adjustment method, including: Three-dimensional vector synthesis: Convert the three-dimensional flow velocity data at each moment into a vector modulus value to obtain a vector modulus value sequence; Standard deviation calculation: Calculate the window standard deviation based on the vector modulus value sequence; Window dynamic adjustment: Perform window dynamic adjustment based on the calculated window standard deviation: ; Wherein: represents the window length at the current moment; represents the minimum value of the window length; represents the maximum value of the window length; represents the lower limit of the standard deviation of the flow velocity; represents the upper limit of the standard deviation of the flow velocity; represents the standard deviation at the current moment; represents a limiting function that limits the value between and If the calculated is less than then output If the calculated is greater than then output ; Based on the calculated window length , adjust the data processing window lengths of all sensors. When the change in the window length results in insufficient data points for non-flow rate sensors, cubic spline interpolation is used for filling.

5. The energy efficiency optimization method of a diving device based on environmental data prediction according to claim 1, characterized in that In step 2, a three-dimensional spatio-temporal data matrix in three dimensions of time, space, and channel is constructed using the time step, the number of spatial sensor nodes, and the characteristics of each sensor data. Then, dilated causal convolution and graph convolutional network are used to extract features from the three-dimensional spatio-temporal data matrix, model the spatial correlation between sensor nodes, and obtain spatio-temporal features.

6. The energy efficiency optimization method of a diving device based on environmental data prediction according to claim 5, characterized in that, Obtaining the spatio-temporal features includes the following steps: Spatial topology modeling: Calculate the distance matrix based on the physical distance between sensors, construct an adjacency matrix based on the distance matrix to represent the connection strength between nodes, then introduce a flow direction correction factor, calculate the angle between the connection direction between nodes and the main flow direction, and adjust the weights of adjacent adjacency matrices according to the flow direction to obtain the adjusted adjacency matrix; Multi-modal data fusion: Organize the standardized data into a three-dimensional tensor according to the dimensions of time, space, and channel, and the data in each time window is retained in different channels to obtain the fused three-dimensional tensor; Spatio-temporal convolutional encoding: Use dilated causal convolution in the time dimension to extract short-term and long-term temporal features by gradually increasing the dilation rate of the convolution sum to obtain temporal features. In the space dimension, calculate the adjacency matrix and the connection information of nodes through graph convolution to extract the patterns in the spatial topology structure to obtain spatial features; Dynamic weight adjustment: Calculate the turbulence intensity of the computing environment, measure the degree of dynamic change of the environment by calculating the fluctuation of the flow velocity, and dynamically adjust the fusion ratio of temporal features and spatial features according to the turbulence intensity, so as to obtain the spatio-temporal feature fusion result.

7. A method for optimizing the energy efficiency of a diving device based on environmental data prediction according to claim 1, characterized in that, The objective function of the multi-objective optimization algorithm is as follows: ; where: both represent weight parameters, and their sum is 1; represents the motor output power at the k-th time step; represents the predicted device position at the k-th time step; represents the task planning target position at the k-th time step; represents the device heading angle at the k-th time step; represents the second derivative of the heading angle; represents the time interval of the control period.

8. A method for optimizing the energy efficiency of a diving device based on environmental data prediction according to claim 1, characterized in that, The objective function of the multi-objective optimization algorithm includes power constraint, device speed constraint, device acceleration constraint and soft constraint; the soft constraint is to introduce a relaxation variable and a deviation tolerance threshold to constrain the path deviation.

9. The energy efficiency optimization method of a diving device based on environmental data prediction according to claim 1, characterized in that The multi-objective optimization algorithm adopts MPC rolling optimization, converts the objective function and constraint conditions into a standard nonlinear programming problem, and outputs it as a problem description file, and uses the problem description file for MPC rolling optimization; When performing MPC rolling optimization, first dynamically adjust the optimization time domain length according to the real-time turbulence intensity prediction result. If the turbulence intensity is less than the threshold, increase the time domain, and if the turbulence intensity is greater than the threshold, shorten the time domain; the time domain length determines the number of future time steps to be considered for each optimization; During the rolling optimization process, the optimal control instruction is obtained by solving with the interior point method. According to the output control instruction, execute the first set of control quantities, and retain the remaining control sequence for the next cycle warm start.

Images