Underwater vehicle sideslip angle intelligent identification method and system based on flow field perception

Abstract

This invention discloses an intelligent method and system for recognizing the sideslip angle of underwater vehicles based on flow field perception, belonging to the field of underwater detection. Addressing the difficulty in accurately obtaining sideslip angles in complex marine environments, this invention involves arranging N pressure sensors at intervals along the surface of the vehicle to collect time-series signals of surface pressure fluctuations. A Daubechies 4 wavelet algorithm is used to decompose each signal into 7 layers, extracting detailed energy at each layer and approximate energy at the 7th layer. Each pressure sensor obtains 8-dimensional energy features, and the N pressure sensors are concatenated into an 8×N-dimensional feature vector. An XGBoost model is used as a classifier, and a nonlinear mapping between the high-dimensional feature vector and the sideslip angle is established through supervised learning. The model automatically optimizes hyperparameters through Bayesian optimization and cross-validation, ultimately outputting the sideslip angle. This invention achieves high-precision sideslip angle recognition based solely on surface pressure without relying on acoustic or inertial equipment, improving accuracy by 19.33% compared to time-domain features, and has promising application prospects.

Description

Technical Field

[0001] This invention belongs to the field of underwater detection, specifically relating to a method and system for intelligent identification of the sideslip angle of an underwater vehicle based on flow field perception. Background Technology

[0002] In real marine environments, ocean currents exhibit significant spatiotemporal nonuniformity and are often accompanied by complex flow phenomena such as internal waves, vortices, and turbulent fluctuations. This results in the underwater vehicle experiencing highly random and rapidly changing directions and intensities of the incoming current. The sideslip angle of the underwater vehicle relative to the incoming current, as a key parameter characterizing the relationship between the vehicle's attitude and the external flow field, directly affects the vehicle's track maintenance, attitude stability, and control law design. The sideslip angle primarily represents the lateral deflection of the underwater vehicle relative to the velocity direction of the incoming current, i.e., the angle between the vehicle's axis and the direction of the current in the horizontal plane (usually...). x – y The degree of deviation from the direction of the incoming flow velocity within a plane, and the angle formed between its axial projection and the projection of the incoming flow velocity direction, such as... Figure 1 As shown; at the same time, abnormal changes in the sideslip angle may also reflect the arrival and evolution of external flow disturbances (such as internal wave shear flow, local strong current, etc.), so the perception and early warning of flow phenomena in complex marine environments are also of great significance.

[0003] Existing methods for obtaining sideslip angles mostly rely on indirect calculations such as inertial devices, trajectory / velocity calculations, external acoustic positioning, or flow velocity and direction measurement devices. However, several shortcomings still exist in practical applications: First, acoustic and external positioning methods are susceptible to multipath effects, environmental noise and blockage, which leads to the accumulation of attitude / velocity calculation errors and makes it difficult to stably invert the relative inflow direction; Secondly, ocean current disturbances often have unsteady and multi-scale characteristics. Traditional estimation methods based on average flow or low-frequency information are difficult to characterize the multi-frequency features contained in signals such as wall pressure, which limits the accuracy and robustness of sideslip angle identification. Third, the statistical characteristics of flow structure and pressure pulsation differ significantly under different incoming flow velocities. Existing methods are prone to model mismatch when velocity changes, making it difficult to meet the comprehensive requirements of engineering scenarios for real-time performance, accuracy, and cross-condition adaptability.

[0004] Based on this, the present invention proposes an intelligent sideslip angle sensing method that relies on the pressure information of the ship's wall and can characterize multi-frequency and multi-scale features, in order to overcome the problems of difficulty in obtaining sideslip angle and low recognition rate in the complex marine environment of the existing technology. Summary of the Invention

[0005] The technical problem to be solved: To overcome the shortcomings of existing technologies, this invention provides an intelligent method and system for recognizing the sideslip angle of underwater vehicles based on flow field perception. The sideslip angle characterizes the lateral deflection state of a vehicle during navigation caused by heading adjustments, external lateral disturbances, or control inputs, and is one of the important characteristic parameters describing the attitude changes of an underwater vehicle. This invention collects the wall pressure pulsation response caused by changes in sideslip angle by arranging a pressure sensor array on the vehicle wall, uses multi-scale wavelet features to characterize the differences in flow structure at different frequency bands, and combines this with a machine learning model to achieve automatic classification and recognition of the sideslip angle. This improves the accuracy and robustness of sideslip angle recognition, overcoming the problems of existing underwater vehicle attitude / sideslip angle perception methods being susceptible to noise interference, insufficient feature extraction, and low recognition rates in complex flow field environments.

[0006] The technical solution of this invention is: an intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception, comprising the following steps: Step 1: Acquire the wall pressure fluctuation time-series signal synchronously collected by the pressure sensor array arranged on the wall of the underwater vehicle; the wall pressure fluctuation time-series signal includes the pressure pulsation information induced by the change of the boundary layer separation point position caused by the change of the sideslip angle of the underwater vehicle relative to the direction of the incoming flow and the evolution of the multi-scale vortex structure. Step 2: Perform multi-level discrete wavelet decomposition on the wall pressure fluctuation time-series signal collected by any pressure sensor in the pressure sensor array, extract the wavelet coefficient energy of each decomposition layer, and construct the multi-scale energy feature of the pressure sensor; combine the multi-scale energy features of all pressure sensors to construct a high-dimensional feature vector that reflects the multi-band structural energy distribution of the flow field. Step 3: Input the high-dimensional feature vector into the pre-trained classification model. The classification model learns and establishes a nonlinear mapping relationship between the high-dimensional feature vector and the sideslip angle through supervised learning, using high-dimensional feature vector samples labeled with sideslip angles as the training sample set. The classification model judges the currently input high-dimensional feature vector according to the mapping relationship and outputs the corresponding sideslip angle as the recognition result.

[0007] A further technical solution of the present invention is that the training process of the classification model includes the following steps: Step 3.1: Under various sideslip angle conditions, the pressure sensor array collects the wall pressure fluctuation time series signal at different positions of the underwater vehicle, and preprocesses and extracts features from the wall pressure fluctuation time series signal under each sideslip angle condition to obtain a high-dimensional feature vector; Step 3.2: Label the true sideslip angle for each high-dimensional feature vector to form a training sample set; Step 3.3: Construct an ensemble learning classification model based on a gradient boosting decision tree framework. The input of the classification model is a high-dimensional feature vector, and the output is the probability distribution of each sideslip angle. Step 3.4: Supervised training of the classification model is performed using the training sample set. The goal is to minimize the loss function between the model's predicted output and the true label. The model parameters are iteratively optimized so that the model learns and converges to the mapping relationship between the high-dimensional feature vector and the sideslip angle.

[0008] A further technical solution of the present invention is: in step 3.1: The various sideslip angle conditions include at least 7 discrete angle categories: -20°, -10°, 0°, 10°, 20°, 30°, and 40°; The high-dimensional feature vector extraction method is as follows: Daubechies 4 wavelet is used to perform 7-level discrete wavelet decomposition on each wall pressure fluctuation time series signal, extracting the detail energy corresponding to the detail coefficients of the 1st to 7th levels, and the approximate energy corresponding to the approximation coefficients of the 7th level, so that each pressure sensor obtains 8-dimensional energy features; the 8-dimensional energy features of N pressure sensors are spliced ​​together to form an 8×N high-dimensional feature vector.

[0009] A further technical solution of the present invention is as follows: In step 3.3, the classification model is an XGBoost model, the core architecture of which is an additive ensemble of multiple CART regression trees; in each iteration, a CART regression tree is generated for each sideslip angle condition, each tree fits the first derivative residual between the current predicted value and the true label, and uses the second derivative information to split nodes and update leaf weights; a greedy algorithm is used to traverse all feature dimensions to select the split position with the largest gain; splitting stops when the gain is less than or equal to a preset threshold or when the maximum depth is reached; each leaf node calculates the corrected weight based on the sum of the gradients falling into that node and the sum of the second derivatives; after each iteration, the prediction results of all newly constructed trees are multiplied by the learning rate and then accumulated into the current XGBoost model.

[0010] A further technical solution of the present invention is as follows: In step 3.4, the hyperparameters of the classification model are automatically optimized by Bayesian optimization combined with five-fold cross-validation; the hyperparameters to be optimized include learning rate, maximum tree depth, number of iterations, L1 regularization coefficient, L2 regularization coefficient, leaf node number penalty coefficient, subsampling rate, and column sampling rate; the optimal combination of hyperparameters is searched through multiple iterations with the average log loss of cross-validation as the objective function.

[0011] A further technical solution of the present invention is: the optimized hyperparameters of the XGBoost model are: Number of iteration rounds: 740 rounds, with one CART regression tree generated for each of the 7 sideslip angles in each round, for a total of 5180 decision trees; Maximum depth of each tree: 3; Learning rate: 0.0318; L1 regularization coefficient: 9.24 × 10 -8 L2 regularization coefficient: 9.08 × 10 -6 ; Subsampling rate: 0.6277, column sampling rate: 0.8735; Leaf node number penalty coefficient: 0.0145, minimum child node weight: 1.

[0012] A further technical solution of the present invention is: in step 1, six pressure sensors with a sampling frequency of 2500Hz and a single sampling duration of 15 seconds are arranged at intervals from the bow to the stern of the underwater vehicle to form a pressure sensor array, and the sensitive surface of each pressure sensor is flush with the wall surface to eliminate the step effect. The raw wall pressure fluctuation time-series signals collected by all pressure sensors were preprocessed by removing the mean and using a 50Hz power frequency notch filter. The sample enhancement was performed using an overlapping sliding window technique with a window length of 5 seconds and a sliding step of 1 second.

[0013] A further technical solution of the present invention is as follows: In step 2, the multi-layer discrete wavelet decomposition adopts Daubechies 4 wavelet, and the number of decomposition layers is 7; 8-dimensional energy features are extracted from each pressure sensor signal, and the 8-dimensional energy features of 6 pressure sensors are spliced ​​together to form a 48-dimensional feature vector.

[0014] An intelligent sideslip angle recognition system for underwater vehicles based on flow field perception, used to execute the intelligent sideslip angle recognition method for the underwater vehicle, includes: A pressure sensor array is installed on the wall of an underwater vehicle to synchronously acquire multi-channel wall pressure fluctuation time-series signals. The feature extraction module is used to perform multi-level wavelet decomposition on each channel of wall pressure fluctuation time series signal and extract the energy of wavelet coefficients at each level to construct a high-dimensional feature vector. The classification and recognition module is a pre-trained classification model. The classification model learns and establishes a non-linear mapping relationship between the high-dimensional feature vectors labeled with sideslip angles through supervised learning, using high-dimensional feature vectors labeled with sideslip angles as the training sample set. The classification and recognition module receives high-dimensional feature vectors extracted in real time and outputs the corresponding sideslip angles according to the mapping relationship. The model training module is used to perform supervised training of the classification model in the offline stage using a labeled training sample set, and to optimize the model parameters to converge to the mapping relationship.

[0015] A further technical solution of the present invention is as follows: the feature extraction module uses Daubechies 4 wavelet to perform a 7-level decomposition on the time series signal of pressure fluctuations on each wall, extracting 7 detail energies and 1 approximate energy; the pressure sensor array includes 6 pressure sensors of model HM91-H2-1-V2-F0-W1-T, arranged at intervals along the wall of the hull from bow to stern; the classification and recognition module is an XGBoost model, with the following hyperparameters: 740 iterations, tree depth 3, learning rate 0.0318, and L1 regularization coefficient 9.24 × 10⁻⁶. -8 The L2 regularization coefficient is 9.08 × 10⁻⁶. -6 .

[0016] Beneficial effects The beneficial effects of this invention are as follows: This invention proposes an intelligent recognition method for the sideslip angle of underwater vehicles based on flow field perception, and verifies its recognition accuracy and robustness under multiple flow velocity conditions. Compared with the prior art, this invention has the following beneficial effects: 1. This invention utilizes a wall pressure sensor array deployed along the axial direction of the vessel to directly acquire the spatiotemporal distribution of wall pressure induced by changes in sideslip angle. This measurement method does not rely on hydrodynamic linearization models and integral calculations, thus achieving accurate differentiation of adjacent 10° sideslip angle states with an accuracy rate exceeding 90%.

[0017] 2. The present invention uses the energy of detail coefficients at each layer to form a multi-dimensional feature vector. Compared with conventional time-domain statistical features (such as root mean square value and variance), the multi-scale wavelet energy feature proposed in this invention can improve the accuracy of sideslip angle recognition by 19.33%, and enhance the feature's ability to characterize differences in flow structure.

[0018] 3. This invention uses the XGBoost model as the recognizer and introduces Bayesian optimization and five-fold cross-validation to automatically optimize hyperparameters. The optimal model obtained through multiple rounds of iterative search improves the final recognition accuracy by approximately 3.89% compared to the default parameter model. This eliminates the need for repeated adjustments based on human experience, significantly reducing model construction costs and subjective bias. Attached Figure Description

[0019] Figure 1 A schematic diagram showing the sideslip angle of an underwater vehicle relative to the incoming current. Figure 2 This is a schematic diagram of the experimental setup constructed in an embodiment of the present invention; Figure 3 This is a simplified schematic diagram illustrating the concept of additive training of the XGBoost model in an embodiment of the present invention; Figure 4 This refers to the average recognition accuracy of the XGBoost model on the test set at different decomposition scales in this embodiment of the invention. Figure 5 This is the test set recognition result of the XGBoost model in this embodiment of the invention at a 7-level decomposition scale.

[0020] Explanation of reference numerals in the attached drawings: 1. Towing pool, 2. Electric rotary table mechanism, 3. Support rod, 4. Vehicle model, 5. Pressure sensor array. Detailed Implementation

[0021] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the invention, and should not be construed as limiting the invention.

[0022] In recent years, inspired by the lateral line system of fish, researchers have begun to explore using pressure sensor arrays deployed on the walls of vehicles to perceive flow field information. For example, related studies have used artificial lateral lines combined with multilayer feedforward neural networks to predict incoming flow velocity and distance from the wall, or used wavelet transform and decision trees to identify the shape of target vehicles. However, existing methods still mainly rely on single statistics or low-dimensional features for feature extraction, resulting in relatively simple model structures. They are primarily geared towards continuous value regression or classification tasks with highly variable shapes. For classification problems that require accurately distinguishing sideslip angles of adjacent 10° deflections, the recognition accuracy and robustness remain significantly insufficient.

[0023] To address the above problems, this invention proposes an intelligent method for recognizing the sideslip angle of underwater vehicles based on flow field perception, comprising the following steps: Step 1: Acquire the wall pressure fluctuation time-series signal synchronously collected by the pressure sensor array arranged on the wall of the underwater vehicle; the wall pressure fluctuation time-series signal includes the pressure pulsation information induced by the change of the boundary layer separation point position caused by the change of the sideslip angle of the underwater vehicle relative to the direction of the incoming flow and the evolution of the multi-scale vortex structure. Step 2: Perform multi-level discrete wavelet decomposition on the wall pressure fluctuation time-series signal collected by any pressure sensor in the pressure sensor array, extract the wavelet coefficient energy of each decomposition layer, and construct the multi-scale energy feature of the pressure sensor; combine the multi-scale energy features of all pressure sensors to construct a high-dimensional feature vector that reflects the multi-band structural energy distribution of the flow field. Step 3: Input the high-dimensional feature vector into the pre-trained classification model. The classification model learns and establishes a nonlinear mapping relationship between the high-dimensional feature vector and the sideslip angle through supervised learning, using high-dimensional feature vector samples labeled with sideslip angles as the training sample set. The classification model judges the currently input high-dimensional feature vector according to the mapping relationship and outputs the corresponding sideslip angle as the recognition result.

[0024] The above technical solution will be further analyzed below with reference to the accompanying drawings and examples: In one embodiment, a method for intelligent recognition of the sideslip angle of an underwater vehicle based on flow field perception is provided, the specific scheme of which is as follows: Step 1: Setting up the experimental platform; Reference Figure 2 As shown, a flow field detection experimental platform was built, including a towed water tank 1, a high-precision electric rotary table mechanism 2, and an artificial sideline system.

[0025] The towing pool 1 is used to provide controllable and stable inflow conditions, and different operating speeds (such as 0.3 m / s, 0.4 m / s, 0.5 m / s) and operating distances (10 m) are set to simulate the operation of the aircraft.

[0026] The electric rotary table mechanism 2 includes a motion controller, an electric rotary table, a support rod 3, and a vehicle model 4. The vehicle model 4 is fixed to the bottom end of the support rod 3, and the top end of the support rod 3 is connected to the turntable of the electric rotary table. When the electric rotary table rotates around its axis under the command of the motion controller, the vehicle model 4 rotates synchronously, and the angle between its axis and the direction of the incoming flow velocity changes, thereby effectively adjusting the sideslip angle of the vehicle model 4 relative to the incoming flow.

[0027] The artificial flank system consists of a pressure sensor array 5 arranged on the wall of the model aircraft 4. In this embodiment, six HM91-H2-1-V2-F0-W1-T type pressure sensors are arranged at intervals from bow to stern on the same side wall of the model aircraft 4. The sensitive surface of each pressure sensor is strictly flush with the outer surface of the model aircraft 4. If necessary, thin-layer encapsulation and surface polishing processes are used to reduce the step effect, thereby reducing additional flow disturbances caused by the installation protrusions.

[0028] Before the experiment, each pressure sensor was calibrated in static water and zero-point to obtain the calibration relationship between the pressure sensor output voltage / charge and the pressure value. Before the experiment, a undisturbed baseline signal was acquired for zero drift correction and consistency verification to improve data reliability.

[0029] Step 2: Fix the vehicle model and establish the coordinate system; The vehicle model 4 is fixed below the electric rotary table mechanism 2 by the support rod 3, and the coordinate system of the vehicle model 4 is established (heading is ). x Axis and transverse are y Axis and vertical are z (Axis). Initially, the sideslip angle β of the vehicle model is set to 0°, meaning the heading axis is aligned with the direction of the incoming flow. The sideslip angle β is the angle between the longitudinal axis of the vehicle and the velocity vector of the incoming flow in the horizontal plane; leftward deviation is negative, and rightward deviation is positive.

[0030] Step 3: Setting flow field parameters and acquiring time-series signals of wall pressure fluctuations; The operating speed of the towing pool 1 is set to 0.3 m / s, 0.4 m / s and 0.5 m / s at the control terminal, and the operating distance is 10m.

[0031] After the speed and attitude stabilized, the time-series signals of the pressure fluctuations on the surface of the vehicle were collected under different sideslip angle conditions (-20°, -10°, 0°, 10°, 20°, 30°, 40°). in, s = 1,…, S , S The total number of pressure sensor channels, indicated by the subscript. s This is the pressure sensor channel number, in this embodiment... S = 6; t For time. Record the sampling frequency. f s = 2500 Hz, single sampling duration T 0 = 15 s, and simultaneously save the experimental label information (speed, sideslip angle, experiment number, etc.).

[0032] For each slip angle condition, the timing signal acquisition of pressure fluctuations on the wall in the stable section was repeated to form a raw pressure database covering multiple velocities and slip angles. Each condition was acquired 5 times to enhance statistical reliability, and obvious abnormal sections (such as sudden changes in mechanical vibration) were removed.

[0033] Step 4: Preprocessing of wall pressure fluctuation timing signals; To increase the sample size, an overlapping sliding window technique was used to enhance the original wall pressure fluctuation time-series signal. The window length was set to... The sliding step size is T h =1s. Timing signal of wall pressure fluctuation for each path. Slicing to obtain subsample sequences ,in, q This indicates the sample number within the time window after the sliding window slice. This represents the index of the discrete sampling point within the slice sequence.

[0034] Each window signal is preprocessed: first, the DC component is removed (mean is removed), and then a 50 Hz power frequency notch filter is applied to suppress noise interference and improve feature robustness.

[0035] Step 5: Multi-scale wavelet decomposition and energy feature extraction; Each preprocessed window signal was decomposed into 7 levels of discrete wavelet using the Daubechies 4 (db4) wavelet. The db4 wavelet has good orthogonality and compact support, and its vanishing moment order is 4. It can effectively capture the transient change characteristics in turbulent pressure fluctuations and has a good characterization ability for the local singularity of the signal.

[0036] As the sideslip angle of the airship model changes, the location of the fluid boundary layer separation point and the vortex structure (such as horseshoe vortices and wake vortices) near the wall evolve. This change in flow field structure is directly reflected in the energy distribution of wall pressure fluctuations. Specifically: (1) Low-frequency band characteristics (large-scale decomposition layer) reflect the separation characteristics of the motion of large-scale vortex structure and the overall boundary layer; (2) High-frequency band characteristics (small-scale decomposition layer) correspond to high-frequency pressure fluctuations generated by small-scale turbulent pulsation and eddy shedding.

[0037] Therefore, in this embodiment, after performing a 7-level discrete wavelet decomposition on the signal from each pressure sensor, the energy values ​​of the detail coefficients from levels 1 to 7 (a total of 7 detail energies) and the energy value of the approximation coefficients from level 7 (1 approximation energy) are extracted respectively. For each window signal... Perform 7-level discrete wavelet decomposition to obtain the th j Layer detail factor Passing the exam J Layer approximation coefficient ;in, J This indicates the preset maximum wavelet decomposition level. J = 7; j This indicates the sequence number of the discrete wavelet decomposition layer.

[0038] The energy of each layer of detail is defined as follows:

[0039] in, Indicates the first q The time window sample at the ... s The first pressure sensor channel, the first j Detail energy under layer wavelet decomposition Indicates the first q The time window sample at the ... s The first pressure sensor channel, the first j The first layer wavelet decomposition obtained n Individual wavelet coefficients, This represents the index of discrete sampling points for the detail coefficients at the current scale. Indicates the sequence number of the discrete wavelet decomposition layer. This indicates the preset maximum wavelet decomposition level. Indicates the pressure sensor channel number. This represents the sample number of the time window after the sliding window slice.

[0040] The approximate energy is defined as:

[0041] in, Indicates the first q Samples within a time window in the decomposition layer J The approximate energy below, superscript ( A ) represents the approximation coefficient.

[0042] Step 6: Constructing high-dimensional feature vectors; The 8-dimensional energy features of the six pressure sensors are sequentially concatenated according to the order of the pressure sensors to form a 48-dimensional high-dimensional feature vector. Specifically, for the... q The feature vector of a time window sample can be represented as:

[0043] in, Indicates the first q Feature vectors of samples within a time window This indicates that the first pressure sensor is in the [number]th [year]. q The energy of the first-level detail coefficients extracted from samples within a time window. This indicates that the first pressure sensor is in the [number]th [year]. q Approximate energy extracted from samples within a time window.

[0044] At the same time, the corresponding sideslip angle β ( 20° 10°, 0°, 10°, 20°, 30°, 40°) were used as labels for supervised learning.

[0045] Step 7: Divide the sample database and dataset; Under different incoming flow velocities (0.3 m / s, 0.4 m / s, 0.5 m / s) and 7 different sideslip angles ( 20° Data was repeatedly collected at 10°, 0°, 10°, 20°, 30°, and 40°, with each condition repeated 5 times. After sliding window enhancement, a large number of samples were obtained. Each sample contains a 48-dimensional feature vector and a corresponding sideslip angle label.

[0046] A hierarchical random sampling strategy is adopted, dividing the samples into training and test sets in a 4:1 ratio to ensure that the sample distribution is consistent across the three speeds and each sideslip angle, thus avoiding model bias and generalization degradation caused by imbalanced sideslip angle samples.

[0047] Step 8: XGBoost model construction and training; This embodiment uses the XGBoost model as the classifier. The model input is a 48-dimensional high-dimensional feature vector, and the output is a probability distribution of 7 sideslip angles.

[0048] (1) Model structure and hyperparameter settings; The base learner is a CART regression tree. The key hyperparameters determined after Bayesian optimization are as follows: Number of iteration rounds (number of base learners) K: The search ranges from 200 to 800, and after Bayesian optimization, it is determined to be 740 rounds. Each iteration generates a CART regression tree for each of the 7 sideslip angles, so the total number of decision trees is 740 × 7 = 5180.

[0049] The maximum depth of each tree is limited to 3-10 to maintain the weak learner characteristics of a single tree. After Bayesian optimization, the maximum depth of each tree is determined to be 3.

[0050] The learning rate η is adjusted within the logarithmic scale of 0.01 to 0.3 to control the contribution step size of each tree. After Bayesian optimization, it is determined to be 0.0318.

[0051] Regularization parameters: Both the L1 regularization coefficient α and the L2 regularization coefficient λ are within the logarithmic scale of 1×10⁻⁶. -8 Search within ~10.0; the leaf node number penalty coefficient γ, minimum child node weight, subsampling rate, and column sampling rate are all automatically determined through Bayesian optimization. After Bayesian optimization, the L1 regularization coefficient α is 9.24 × 10⁻⁶. -8 The L2 regularization coefficient λ is 9.08 × 10⁻⁶. -6 .

[0052] Leaf node number penalty coefficient γ: 0.0145 (search range 0~10); Minimum child node weight: 1 (search range 1~15); Subsampling rate: 0.6277 (search range 0.5~1.0); Column sampling rate: 0.8735 (search range 0.5 to 1.0).

[0053] (2) Additive training strategy; The XGBoost model employs an additive training strategy, the specific process of which is as follows: Initialization: The model's initial prediction value is 0.

[0054] Iterative Construction: In each iteration, a CART regression tree is constructed for each of the seven sideslip angle scenarios. Each tree fits the residual between the current model prediction and the true label, using the first derivative (gradient) information to fit the residual, and using the second derivative (Hessian matrix) information to perform node splitting and weight updates.

[0055] The regularization objective function for model optimization is:

[0056] in, i Indicates the training sample index. and They represent the first i The true labels of each training sample and the probability distribution predicted by the model. For multi-class classification loss function (softmax cross-entropy):

[0057] in, y c Represents the th element in the true label vector. c The components of the sideslip angle, The vector representing the probability distribution predicted by the model. The model predicts that the sample belongs to the sideslip angle. c The posterior probability, C = 7 is the total number of sideslip angles. c It is the index of the target sideslip angle, with a value ranging from 1 to... C .

[0058] For the first k The complexity regularization term for trees is specifically defined as:

[0059] in, Indicates the first k A decision tree, T k For the first k The number of leaf nodes in a tree. The penalty coefficient for the number of leaf nodes. It is the L2 regularization coefficient. For the first l The output weights of the leaf nodes l The index of the leaf node.

[0060] Node splitting: A greedy algorithm is used to traverse all 48 feature dimensions and calculate the gain for each possible split point. The gain calculation formula is:

[0061] in, G The sum of gradients This represents the sum of the first derivatives (gradients) of all samples in the left child node. This represents the sum of the first derivatives (gradients) of all samples in the right child node.H It is the sum of the second derivatives. This represents the sum of the second derivatives (Hessian matrix) of all samples in the left child node. This represents the sum of the second derivatives (Hessian matrix) of all samples in the right child node. Splitting stops if Gain ≤ 0.

[0062] Leaf node weight calculation: The weight of each leaf node is calculated based on the sum of the gradient and second derivative of the samples falling into that node.

[0063] in, It is the first l The output weights (or predicted values) of each leaf node. Is it falling into the first l The sum of the first-order gradients (gradients) of all training samples in each leaf node. Is it falling into the first l The sum of the second derivatives of all training samples at each leaf node. It is the L2 regularization coefficient.

[0064] Model update: Multiply the cumulative weights of the leaf nodes of the 7 newly constructed CART regression trees in this iteration by the learning rate η=0.0318, and then add them to the cumulative scores of the corresponding categories in the current model to complete one round of update.

[0065] Repeat the above process until 740 iterations are completed.

[0066] Reference Figure 3 The diagram shown is a simplified illustration of the additive training concept of the XGBoost model. Tree_1 to Tree_k represent the iteration sequence, with each iteration showing only a single tree example structure. In the actual multi-class implementation, each iteration generates a CART regression tree for each sideslip angle (7 trees in total), and all classes are trained in parallel. Result_k represents the prediction correction for a certain class in that iteration, and the final Weighted Sum is the sum of predictions from all trees for all classes. Detailed parameter identification (including tree splitting and leaf weight calculation) relies on the gain formula and regularization objective function in step 12, while hyperparameter identification is completed through Bayesian optimization in the following steps.

[0067] (3) Hyperparameter optimization; To improve generalization ability and recognition accuracy, a Bayesian optimization algorithm is introduced to automatically optimize the key hyperparameters of XGBoost, and a five-fold cross-validation strategy is used to evaluate the model performance for each set of parameters. The hyperparameters to be optimized and their search range include: (1) Learning rate: 0.01 to 0.3 (log-scale sampling); (2) Maximum tree depth: 3 to 10 (integer); (3) Subsampling rate: 0.5–1.0; (4) Column sampling rate: 0.5~1.0; (5) Leaf node number penalty coefficient γ: 0.0~10.0; (6) L1 regularization coefficient α: 1×10 -8 ~10.0 (log-scale sampling); (7) L2 regularization coefficient λ: 1×10 -6 ~10.0 (log-scale sampling); (8) Minimum child node weight: 1 to 15 (integer); (9) Number of weak learners (number of iterations): 200 to 800 (integer).

[0068] The Bayesian optimization process uses the average log loss of five-fold cross-validation as the objective function:

[0069] in, Represents a given hyperparameter vector The objective function value at that time, in this invention, refers to the average log loss under five-fold cross-validation. For hyperparameter vectors, For the first r The average log loss on the validation set. The fold index represents the cross-validation number, traversing from 1 to 5. The optimal hyperparameter combination is searched using a probabilistic surrogate model to efficiently sample in the continuous or discrete parameter space, with a finite number of objective function evaluations (300 in this example), ultimately determining the aforementioned optimal parameter configuration.

[0070] Step 9: Final model training and inference; Using the optimal hyperparameter combination determined through Bayesian optimization and five-fold cross-validation in step 8, the XGBoost model is retrained on the complete training set to obtain the final sideslip angle recognition model. For the time-series signal samples of underwater vehicle wall pressure fluctuations to be identified, the same preprocessing and multi-scale wavelet energy feature extraction process as in the training phase is first followed to obtain a 48-dimensional feature vector. The feature vector is input into the trained XGBoost model. The model performs forward inference by traversing all 5180 decision trees generated in 740 iterations: each tree reaches a leaf node along the branch path based on the feature value, and outputs the corresponding leaf node weight; the weights of the leaf nodes of all trees are accumulated by category to obtain 7 raw scores; finally, the probability distribution is mapped through the Softmax function.

[0071] in, This indicates that the model predicts the sample to belong to the sideslip angle. c The posterior probability, The index representing the target sideslip angle, with a value ranging from 1 to 7. This indicates that the XGBoost model is for the first... c The original accumulated score output by each side slip angle. It is an index variable that represents the traversal of all sideslip angles. It is the natural constant (approximately 2.718); The model is indexed as m The original accumulated score is output by the sideslip angle.

[0072] Take the sideslip angle corresponding to the highest probability As the final identification result:

[0073] Step 10: Model performance evaluation; Predictions are made on an independent test set, and a confusion matrix is ​​constructed and multiple classification evaluation metrics are calculated. The feature vectors of all samples in the test set are sequentially input into the optimal XGBoost model trained in step 9 to obtain the predicted sideslip angle for each sample. Its actual sideslip angle label Together they form the predicted-true label pair. Let the total number of sideslip angles be C = 7, and define the confusion matrix. Its elements This represents the number of samples whose true class is class u and which are predicted by the model to be class v.

[0074] in, The total number of samples in the test set. This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. The confusion matrix is... u line reflection u The distribution of real samples into different categories, the first v The column reflects the prediction as the first v The samples for each class represent the distribution of the real classes. The main diagonal elements of the confusion matrix. This represents the number of samples that were correctly identified in each category, while the off-diagonal elements reflect the degree of confusion between different sideslip angle states.

[0075] Step 11: Calculate the classification indicators; Based on the confusion matrix, this invention evaluates model performance by calculating the following metrics: (1) Overall accuracy: The proportion of samples correctly classified out of all categories to the total number of samples, reflecting the overall recognition accuracy of the model.

[0076]

[0077] (2) Precision for each category: For the first category u The class is the proportion of samples that actually belong to that class, reflecting the accuracy of the model's prediction for that class.

[0078]

[0079] (3) Recall rate for each category: For the first category u The class represents the proportion of samples that are correctly identified from those that actually belong to that class, reflecting the model's recall capability for that class.

[0080]

[0081] (4) F1 score for each category: the harmonic mean of precision and recall, which comprehensively reflects the classification performance of the model for each category.

[0082]

[0083] Step 12: Presentation of experimental results; Figure 4 and Figure 5 The comparison results of model recognition performance and the statistical results of average recognition accuracy under different wavelet decomposition levels are presented respectively. Table 1 shows the recognition effect of the model on the sideslip angle of the aircraft body at a 7-level decomposition scale (precision, recall, F1 score). Table 2 compares the sideslip angle recognition accuracy of the proposed model method, the method based on conventional temporal feature input, and the experimental condition without Bayesian optimization.

[0084] Step 13: Conclusion; from Figure 4 and Figure 5 As can be seen, the model achieves a sideslip angle recognition accuracy of over 90% at different decomposition scales, demonstrating good recognition performance. Among all decomposition scales, the 7-level wavelet decomposition achieves the optimal balance between recognition accuracy and computational efficiency, and can be considered the best feature scale configuration. Table 2 shows that, compared with conventional time-series features, the multi-scale wavelet energy features used in this invention improve the sideslip angle recognition accuracy by 19.33%; compared with the default parameter model, the Bayesian optimization method used improves the accuracy by approximately 3.89%.

[0085] Table 1. Model prediction performance at 7-level decomposition scale

[0086] Table 2 Comparison of Model Experimental Performance Results

[0087] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for intelligent recognition of sideslip angle of underwater vehicles based on flow field perception, characterized in that, Includes the following steps: Step 1: Acquire the wall pressure fluctuation time-series signal synchronously collected by the pressure sensor array arranged on the wall of the underwater vehicle; the wall pressure fluctuation time-series signal includes the pressure pulsation information induced by the change of the boundary layer separation point position caused by the change of the sideslip angle of the underwater vehicle relative to the direction of the incoming flow and the evolution of the multi-scale vortex structure. Step 2: Perform multi-level discrete wavelet decomposition on the wall pressure fluctuation time-series signal collected by any pressure sensor in the pressure sensor array, extract the wavelet coefficient energy of each decomposition layer, and construct the multi-scale energy characteristics of the pressure sensor. By combining the multi-scale energy characteristics of all pressure sensors, a high-dimensional feature vector reflecting the multi-band structural energy distribution of the flow field is constructed. Step 3: Input the high-dimensional feature vector into the pre-trained classification model. The classification model learns and establishes a nonlinear mapping relationship between the high-dimensional feature vector and the sideslip angle through supervised learning, using high-dimensional feature vector samples labeled with sideslip angles as the training sample set. The classification model judges the currently input high-dimensional feature vector according to the mapping relationship and outputs the corresponding sideslip angle as the recognition result.

2. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 1, characterized in that: The training process of the classification model includes the following steps: Step 3.1: Under various sideslip angle conditions, the pressure sensor array collects the wall pressure fluctuation time series signal at different positions of the underwater vehicle, and preprocesses and extracts features from the wall pressure fluctuation time series signal under each sideslip angle condition to obtain a high-dimensional feature vector; Step 3.2: Label the true sideslip angle for each high-dimensional feature vector to form a training sample set; Step 3.3: Construct an ensemble learning classification model based on a gradient boosting decision tree framework. The input of the classification model is a high-dimensional feature vector, and the output is the probability distribution of each sideslip angle. Step 3.4: Supervised training of the classification model is performed using the training sample set. The goal is to minimize the loss function between the model's predicted output and the true label. The model parameters are iteratively optimized so that the model learns and converges to the mapping relationship between the high-dimensional feature vector and the sideslip angle.

3. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 2, characterized in that: In step 3.1: The various sideslip angle conditions include at least 7 discrete angle categories: -20°, -10°, 0°, 10°, 20°, 30°, and 40°; The method for extracting the high-dimensional feature vector is as follows: Daubechies 4 wavelet is used to perform 7-level discrete wavelet decomposition on each wall pressure fluctuation time series signal, extracting the detail energy corresponding to the detail coefficients of the 1st to 7th levels, and the approximate energy corresponding to the approximation coefficients of the 7th level, so that each pressure sensor obtains 8-dimensional energy features; the 8-dimensional energy features of N pressure sensors are spliced ​​together to form an 8×N high-dimensional feature vector.

4. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 3, characterized in that: In step 3.3, the classification model is the XGBoost model, whose core architecture is the additive ensemble of multiple CART regression trees; in each iteration, a CART regression tree is generated for each sideslip angle condition, each tree fits the first derivative residual between the current predicted value and the true label, and uses the second derivative information to split nodes and update leaf weights; a greedy algorithm is used to traverse all feature dimensions to select the split position with the largest gain. Splitting stops when the gain is less than or equal to a preset threshold or when the maximum depth is reached; each leaf node calculates a corrected weight based on the sum of the gradients falling into that node and the sum of the second derivatives; After each iteration, the prediction results of all newly constructed trees are multiplied by the learning rate and then accumulated into the current XGBoost model.

5. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 4, characterized in that: In step 3.4, the hyperparameters of the classification model are automatically optimized by Bayesian optimization combined with five-fold cross-validation. The hyperparameters to be optimized include learning rate, maximum tree depth, number of iterations, L1 regularization coefficient, L2 regularization coefficient, leaf node number penalty coefficient, subsampling rate, and column sampling rate. The optimal combination of hyperparameters is searched through multiple iterations with the average log loss of cross-validation as the objective function.

6. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 5, characterized in that: The optimized hyperparameters of the XGBoost model are: Number of iteration rounds: 740 rounds, with one CART regression tree generated for each of the 7 sideslip angles in each round, for a total of 5180 decision trees; Maximum depth of each tree: 3; Learning rate: 0.0318; L1 regularization coefficient: 9.24 × 10 -8 L2 regularization coefficient: 9.08 × 10 -6 ; Subsampling rate: 0.6277, column sampling rate: 0.8735; Leaf node number penalty coefficient: 0.0145, minimum child node weight:

1.

7. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 1, characterized in that: In step 1, six pressure sensors with a sampling frequency of 2500Hz and a single sampling duration of 15 seconds are arranged at intervals from the bow to the stern on the wall of the underwater vehicle to form a pressure sensor array. The sensitive surface of each pressure sensor is flush with the wall to eliminate the step effect. The raw wall pressure fluctuation time-series signals collected by all pressure sensors were preprocessed by removing the mean and using a 50Hz power frequency notch filter. The sample enhancement was performed using an overlapping sliding window technique with a window length of 5 seconds and a sliding step of 1 second.

8. The intelligent recognition method for the sideslip angle of an underwater vehicle based on flow field perception according to claim 1, characterized in that: In step 2, the multi-layer discrete wavelet decomposition uses Daubechies 4 wavelet with 7 decomposition layers; 8-dimensional energy features are extracted from each pressure sensor signal, and the 8-dimensional energy features of 6 pressure sensors are spliced ​​together to form a 48-dimensional feature vector.

9. An intelligent recognition system for the sideslip angle of an underwater vehicle based on flow field perception, used to execute the intelligent recognition method for the sideslip angle of an underwater vehicle as described in any one of claims 1-8, characterized in that, include: A pressure sensor array is installed on the wall of an underwater vehicle to synchronously acquire multi-channel wall pressure fluctuation time-series signals. The feature extraction module is used to perform multi-level wavelet decomposition on each channel of wall pressure fluctuation time series signal and extract the energy of wavelet coefficients at each level to construct a high-dimensional feature vector. The classification and recognition module is a pre-trained classification model. The classification model learns and establishes a non-linear mapping relationship between the high-dimensional feature vectors labeled with sideslip angles through supervised learning, using high-dimensional feature vectors labeled with sideslip angles as the training sample set. The classification and recognition module receives high-dimensional feature vectors extracted in real time and outputs the corresponding sideslip angles according to the mapping relationship. The model training module is used to perform supervised training of the classification model in the offline stage using a labeled training sample set, and to optimize the model parameters to converge to the mapping relationship.

10. The intelligent recognition system for sideslip angle of an underwater vehicle based on flow field perception according to claim 9, characterized in that: The feature extraction module uses Daubechies 4 wavelets to perform a 7-level decomposition on the temporal signal of pressure fluctuations on each wall, extracting 7 detail energies and 1 approximate energy. The pressure sensor array includes 6 pressure sensors of model HM91-H2-1-V2-F0-W1-T, arranged at intervals along the wall of the hull from bow to stern. The classification and recognition module is an XGBoost model with the following hyperparameters: 740 iterations, tree depth 3, learning rate 0.0318, and L1 regularization coefficient 9.24 × 10⁻⁶. -8 The L2 regularization coefficient is 9.08 × 10⁻⁶. -6 .

Images