A quick identification method for underwater launch bore mouth cavity

Abstract

The application provides a fast recognition method suitable for underwater launch bore cavity, comprising: acquiring a cavity sequence image of underwater launch; pre-processing the sequence image to obtain a column image with optical correction and normalized uniform size; inputting the pre-processed image into a recognition model, and providing physical constraints for the neural network based on prior knowledge to obtain a cavity volume, a closing rate and an asymmetry degree; calculating a total cavity evolution energy level by using the cavity volume, the closing rate and the asymmetry degree, setting an energy level threshold and an asymmetry degree threshold, and evaluating the cavity dynamics by the total cavity evolution energy level and the asymmetry degree. The application solves the technical problems of large recognition error and inability to quantitatively analyze the dynamic characteristics of the cavity in turbid water in the traditional method.

Description

Technical Field

[0001] This invention relates to the field of underwater weapon testing and intelligent monitoring technology, and in particular to a method for rapid identification of underwater launch muzzle cavitation that integrates prior knowledge of fluid dynamics and deep learning technology. Background Technology

[0002] When underwater weapons (such as underwater rifles and torpedoes) are launched, the high-pressure gas from the muzzle is expelled into the water, creating complex cavitation phenomena. The dynamic evolution of these cavitation bubbles directly affects launch stability, stealth, and weapon structural safety. Traditionally, cavitation observation has relied primarily on high-speed photography and manual post-processing. This method has significant drawbacks: processing times can be as long as several hours, failing to meet feedback requirements; in turbid waters, the image signal-to-noise ratio is low, and existing image processing algorithms often have an error exceeding 30% in identifying cavitation edges; and there is a lack of effective means to quantify the microsecond-level dynamic characteristics of cavitation bubbles (such as closure rate and asymmetry). These technical bottlenecks severely restrict the optimization of underwater weapon launch processes and fault diagnosis. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a rapid identification method for underwater launch muzzle cavitation.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A rapid identification method for underwater firing muzzle cavitation includes:

[0006] Acquire images of the cavitation sequence emitted underwater;

[0007] Preprocessing the sequence of images yields optically corrected and normalized column images of uniform size;

[0008] The preprocessed image is input into the recognition model, and physical constraints are provided to the neural network based on prior knowledge to obtain the cavitation volume V, closure rate R, and asymmetry A.

[0009] The total cavitation evolution energy level is calculated using the cavitation volume V, closure rate R, and asymmetry A. Energy level thresholds and asymmetry thresholds are set, and cavitation dynamics are evaluated using the total cavitation evolution energy level and asymmetry.

[0010] The significant advantages of this invention compared to existing technologies are:

[0011] (1) High efficiency: By setting up lightweight convolutional neural networks (CNN), the time for cavitation analysis is greatly shortened, and the efficiency is improved by 1000 times.

[0012] (2) Strong adaptability: With the help of matching optical correction technology, it can still operate stably in turbid waters.

[0013] (3) By combining fluid dynamics physics methods with deep learning networks, a new approach has been provided for the intelligent testing of underwater weapons. Attached Figure Description

[0014] Figure 1 This is a flowchart of the underwater launch muzzle cavitation rapid identification method of the present invention.

[0015] Figure 2 This is a schematic diagram of the lightweight convolutional neural network structure of the present invention, with the nodes for injecting prior knowledge marked in the diagram.

[0016] Figure 3 This is an example of the symmetrical cavitation morphology of an underwater projectile.

[0017] Figure 4 It is an example of an asymmetric cavitation morphology of an underwater projectile (with an angle of attack).

[0018] Figure 5 This is the flowchart for step S1.

[0019] Figure 6 This is the flowchart for step S2.

[0020] Figure 7 This is the flowchart for step S3.

[0021] Figure 8 This is the flowchart for step S4. Detailed Implementation

[0022] The following is in conjunction with the appendix Figure 1-8 The various steps of the method described in this invention will be explained in detail below. This embodiment takes the monitoring of the underwater rifle firing process as an example.

[0023] Step S1: Acquire underwater emission cavitation sequence images (e.g., using an optical acquisition module) Figure 5 (as shown);

[0024] S1.1 Establishing the optical acquisition module

[0025] A Phantom VEO 710L camera was selected, with the resolution set to 1280×800, the frame rate to 2000 fps, and the exposure time to 1 second. To ensure image sharpness, the lens aperture was set to f / 8.0.

[0026] Using a wavelength of 532 A green laser is used in conjunction with a cylindrical mirror to form a sheet light source. The wavelength was chosen based on the absorption coefficient of water for green light (…). ) and scattering coefficient ( ≈0.003 The light intensity is relatively low, yet it can effectively penetrate turbid water. Set it to 1000 lux to ensure a signal-to-noise ratio greater than 30 dB.

[0027] The camera, laser, and transmitter are connected to the synchronization controller (model: BNC Model 575) via BNC cables. The transmission signal is set to rise-edge triggering, and the camera delays upon receiving the trigger signal. Data acquisition begins after 100 μs to capture the initial formation stage of cavitation. Acquisition duration =500 ms, ensuring the complete cavitation lifecycle is recorded.

[0028] S1.2 Acquire sequence images

[0029] A single launch event will generate a sequence of images. }, t=1,2,...,N, where N is the total number of effective image frames acquired by the high-speed camera, which is determined by the acquisition time and the camera frame rate, and its calculation formula is N= × Frame rate = 500 × 2000 = 1,000,000 frames. The t-th frame is... It is a three-dimensional matrix with dimensions H×W×3 (height×width×color channels).

[0030] Step S2: Preprocess the image sequence, including optical correction based on adaptive exposure control techniques (such as...). Figure 6 (As shown); S2 aims to reduce image degradation caused by complex underwater optical environments and improve the quality of subsequent input images.

[0031] S2.1 Optical Correction Based on Physical Methods

[0032] The underwater image degradation method can be simplified as follows:

[0033] ,

[0034] in: y is the horizontal coordinate (width direction), and y is the vertical coordinate (height direction). The intensity of the observed image; The true radiation intensity of the target; This is a transmittance diagram, representing the degree of light attenuation as it travels from the target to the camera. β(λ) is the spectral attenuation coefficient of the water body, which is related to turbidity, and λ is the wavelength of light. In this invention, λ = 532. ; It is a distance map from the target to the camera, obtained through binocular vision or prior calibration; Background ambient light intensity; It is additive noise.

[0035] This invention employs the following iterative algorithm for solving:

[0036] (1) Initial estimation: Assuming an initial image = .

[0037] in, These are raw observation images, that is, images directly captured by a high-speed camera without any processing; This is the target's true image estimated in the 0th iteration.

[0038] (2) Transmittance update: Obtained by referring to a table based on the turbidity of the water body. Combined with the calibrated distance map Calculate the estimated value of the transmittance map at the k-th iteration. .

[0039] (3) Background light estimation: Take the average value of the darkest 0.1% pixel area of ​​the image to estimate the background light. .

[0040] (4) Real Image Update: Based on the result of the k-th iteration, estimate the image after the (k+1)-th iteration. True radiation intensity at coordinates :

[0041] ,

[0042] (5) Noise suppression: The image obtained in the (k+1)th iteration Apply guided filtering to smooth the noise η. .

[0043] (6) When the mean square error (MSE) of two consecutive iterations is less than the threshold Stop iteration when:

[0044] ,

[0045] in, The height of the image. The width of the image. Let be the true radiation intensity of the target at the k-th iteration, and let be the final result obtained after the iteration converges. As the corrected true image intensity , for subsequent processing.

[0046] S2.2 Image Normalization and Resizing

[0047] The corrected image Pixel values ​​are normalized to the [0,1] interval:

[0048] ,

[0049] in, The normalized image intensity represents the coordinates. The normalized and dimensionless intensity value of the pixel; The intensity of the corrected true image, Representing an image The maximum value of all pixels in the range. Representing an image The minimum value of all pixels in the array; Representing an image The dynamic range.

[0050] Subsequently, a bicubic interpolation algorithm was used to uniformly adjust the image size to 512×512 to meet the network input requirements. The interpolation weight function is:

[0051] ,

[0052] in, The distance from the point to be interpolated to a neighboring source pixel is measured in units of 1, which is the center-to-center distance between adjacent source pixels. Used to calculate distance The contribution weight of the source pixels to the final interpolation result; parameters The shape of the weighting function is used to balance image smoothness and edge sharpness, and is usually set to -0.5, which can achieve smooth scaling while maintaining edge sharpness.

[0053] Step S3: Input the preprocessed image into the fast identification method to obtain the cavitation volume V, closure rate R, and asymmetry A (e.g., ...). Figure 7 (as shown);

[0054] S3.1 Network forward propagation

[0055] Preprocessed image Input a lightweight CNN.

[0056] The forward propagation and feature learning process of this network is described as follows:

[0057] The network takes an underwater cavitation image of size 512×512×3 as input from the Input layer. This input is first processed by a convolutional layer with a kernel size of 7×7 and a stride of 2. This operation extracts low-level edge and texture features of the image and halves the spatial dimension (512→256), which greatly reduces the subsequent computation and outputs 32 feature channels.

[0058] Subsequently, the feature images enter the core feature extraction section, which consists of three multi-scale modules. Each module employs two paths, using convolutional kernels of different sizes: 3×3 and 5×5, to simultaneously capture both the fine local contours and global morphological features of the cavitation (3×3 for capturing fine local contours, and 5×5 for capturing global morphological features). During propagation, the network halves in spatial dimension, while the number of feature channels doubles accordingly. Specifically:

[0059] MSBlock1 (at 256×256 resolution): The network analyzes the local details and regional morphology of cavitation at a higher resolution, accurately locating and outlining the contours of the cavitation. The number of channels increases to 64, generating 64 different feature descriptions.

[0060] MSBlock2 (at 128×128 resolution): The image size is halved, allowing for the observation of larger structures by sacrificing some local details. The network begins to understand the regional composition of cavitation (such as head and tail regions) and the relationships between them. The number of channels is increased to 128, indicating a richer description of these structures.

[0061] MSBlock3 (at 64×64 resolution): At the lowest resolution, the network can no longer see details, but it gains the widest field of view, enabling it to grasp the global shape of cavitation. With 256 channels, the network generates 256 abstract concepts to describe this global information. These three modules constitute the main body of the network parameters.

[0062] Following feature fusion, the network uses a global average pooling layer to aggregate the entire feature map space information of each channel into a single scalar value, thereby compressing a 64×64×256 feature tensor into a 1×1×256 vector. This operation achieves significant dimensionality reduction, minimizing subsequent computational overhead.

[0063] Finally, the network constructs a physical parameter regression head through three independent fully connected layers in parallel, mapping the 256-dimensional feature vector to three key cavitation physical quantities: cavitation volume, closure rate, and asymmetry. Each fully connected layer contains only about 257 parameters, achieving decoupled regression of complex physical quantities.

[0064] In summary, this network achieves accurate identification of underwater cavitation dynamics under the lightweight constraint of strictly controlling the total number of parameters to approximately 1.06 megabytes.

[0065] S3.2 Injecting Prior Physical Knowledge

[0066] Figure 3The diagram illustrates the cavitation patterns of a projectile underwater under different incoming flow velocities (V=100 m / s, V=200 m / s, V=400 m / s). For a fully developed muzzle cavitation, its profile conforms to the evolution of natural supercavitation and can be described using a classical empirical formula derived from Logvinovich's principle of independence.

[0067] ,

[0068] in, The axial coordinate system is defined with the cavitation initiation point as the origin and the direction pointing towards the inside of the projectile as the positive direction. The radius at that point, Let L be the maximum radius of the cavitation bubble, and L be the total length of the cavitation bubble. The cavitation number is calculated based on the flow field parameters at the nozzle. This formula constrains the network's morphological recognition of steady-state, symmetrical cavitation bubbles, affecting its predicted aspect ratio L / The parameters must conform to the reasonable range defined by this physical law.

[0069] The knowledge base also integrates empirical relationships of the macroscopic geometric characteristics of cavitation, whose aspect ratios satisfy:

[0070] ,

[0071] in, The drag coefficient associated with the projectile head described in this invention is an empirical constant. This represents the maximum diameter of the cavitation bubble. This relationship is used to calculate the cavitation volume V, length L, and maximum diameter obtained from network regression. Perform cross-validation and physical consistency checks.

[0072] The physical constraints for typical cavitation morphologies are defined in the prior knowledge base. For sheet-like cavitation, the aspect ratio L / D is typically greater than 5; for cloud-like cavitation, the radius of curvature is... It satisfies an empirical relationship with volume V. These constraints are injected into the network during training via the loss function:

[0073] ,

[0074] in, This is the total loss function; For the predicted cavitation volume True label of cavitation volume The mean square error between them For the predicted cavitation closure rate True label of cavitation closure rate The mean square error between them Predicted cavitation asymmetry True label of cavitation asymmetry The mean square error between them (true labels refer to data obtained through more authoritative and accurate measurement methods independent of this method); , The physical constraint loss is responsible for performing physical law checks and penalizing violations; ϕ represents derived parameters such as aspect ratio and curvature. These are the derived parameters for prediction (secondary parameters with explicit physical meanings derived from the original parameters V, R, and A output by the neural network, such as aspect ratio L / D). ϕ represents the prior physical parameter (the empirical range of values ​​that the derived parameter ϕ should have); τ is the tolerance threshold. This is the weighting coefficient (set to 0.1).

[0075] S3.3 Calculate key parameters

[0076] Cavity volume V: The volume value directly regressed by the network is learned based on a large amount of calibration data. Its physical basis is the 3D shape reconstruction of the segmented cavitation regions. For monocular images, assuming the cross-section of the cavitation bubble perpendicular to the image plane is an ellipse, the volume can be approximated as:

[0077] ,

[0078] in, The actual area represented by the i-th pixel (determined by pixel scale and camera calibration); This represents the depth information estimated at this pixel based on the curvature of the cavitation contour; k is the shape correction factor, which for axisymmetric cavitation is k≈π / 4.

[0079] Closure rate R: Calculated by the rate of change of cavitation volume in consecutive frames.

[0080] ,

[0081] in, The frame interval is 0.5ms.

[0082] Asymmetry A: Calculated by measuring the offset of the geometric center of the cavitation binary mask from the principal axis.

[0083] ,

[0084] in, It is the center of gravity of the cavitation mask. This represents the total number of pixels within the bubble region. For the first The coordinates of a bubble pixel; It is the center of the principal axis of the cavitation fitting ellipse (the standard ellipse that best represents the overall position, orientation, and size of the cavitation shape). =(h,j), where h represents the x-coordinate of the center point of the ellipse and j represents the y-coordinate of the center point of the ellipse. It is the maximum diameter of the cavitation bubble. =0 indicates perfect symmetry, such as Figure 4 The cavitation morphology diagram and pressure cloud diagram shown can intuitively reflect whether the underwater cavitation of the projectile is symmetrical. It can be clearly seen that when the angle of attack is not zero... It is definitely not zero.

[0085] Step S4: The intensity of cavitation evolution can be determined by the total cavitation evolution, and the stability of cavitation morphology can be determined by the asymmetry threshold (e.g., ...). Figure 8 (As shown). S4 performs comprehensive analysis on the identified parameters to achieve intelligent diagnosis of cavitation state.

[0086] S4.1 Calculate the total cavitation evolution energy levels

[0087] To comprehensively evaluate the dynamic behavior of cavitation, the total cavitation evolution energy level is defined. :

[0088] ,

[0089] in, The value approximately represents the kinetic energy released during cavitation collapse and is used to measure the intensity of cavitation evolution. Parameters α and β are determined through a specially designed energy calibration experiment, which simultaneously acquires high-speed camera sequences and dynamic pressure data of the cavitation collapse region. Based on historical experimental data, in this embodiment, α = 0.1 and β = 0.5.

[0090] S4.2 Set anomaly detection threshold

[0091] Energy level threshold Based on historical normal launch data, calculate The statistical distribution of the mean. The threshold is set as the mean plus three standard deviations: = +3 (Normal distribution, the probability of data within this range is 99.73%).

[0092] in, This represents the average energy level of the cavitation evolution. Energy levels for cavitation evolution The standard deviation in this embodiment Therefore, it is set =150.

[0093] Asymmetry threshold Based on experience, when A > 0.3, the risk of asymmetric collapse of cavitation bubbles increases significantly; therefore, we set... =0.3. When > Or A> This means that unstable phenomena such as asymmetric collapse exist.

[0094] The detailed implementation process described above demonstrates that this invention provides a complete and feasible solution, from hardware configuration, image preprocessing, intelligent recognition to comprehensive evaluation and judgment, effectively solving the problem of accurate monitoring of underwater launch muzzle cavitation.

Claims

1. A rapid identification method for underwater launch muzzle cavitation, characterized in that, include: Acquire images of the cavitation sequence emitted underwater; Preprocessing the sequence of images yields optically corrected and normalized column images of uniform size; The preprocessed image is input into the recognition model, and physical constraints are provided to the neural network based on prior knowledge to obtain the cavitation volume V, closure rate R, and asymmetry A. The total cavitation evolution energy level is calculated using the cavitation volume V, closure rate R, and asymmetry A. Energy level thresholds and asymmetry thresholds are set, and cavitation dynamics are evaluated using the total cavitation evolution energy level and asymmetry. The identification model employs a convolutional neural network, including: The Input layer is used to perform convolution processing on the preprocessed underwater cavitation image; Three multi-scale modules, each with two paths and two different sizes of convolutional kernels, are used to capture the local fine contours and global morphological features of cavitation. The global average pooling layer aggregates the entire feature map space information of each channel into a single scalar value, resulting in a compressed feature vector. The physical parameter regression head maps the feature vectors to three cavitation physical quantities: cavitation volume, closure rate, and asymmetry. Total cavitation evolution energy level The calculation method is as follows: ， Energy level threshold Set as: = +3 ， when > Or A> That is, there exists asymmetric collapse; among which The asymmetry threshold is α, and β are calibration parameters. Energy levels for cavitation evolution The mean, Energy levels for cavitation evolution The standard deviation.

2. The rapid identification method for underwater firing muzzle cavitation according to claim 1, characterized in that, Physical constraints are injected during network training via the loss function: ， in, This is the total loss function; For the predicted cavitation volume True label of cavitation volume The mean square error between them For the predicted cavitation closure rate True label of cavitation closure rate The mean square error between them Predicted cavitation asymmetry True label of cavitation asymmetry The mean square error between them These are the weighting coefficients. As a physical constraint loss, it is responsible for performing physical law checks and punishing behaviors that violate the laws.

3. The rapid identification method for underwater firing muzzle cavitation according to claim 1, characterized in that, The cavitation volume V, closure rate R, and asymmetry A are calculated using the following formula: ， ， ， in, Let be the actual area represented by the i-th pixel. represents the depth information estimated at the i-th pixel based on the cavitation contour curvature; k is the shape correction coefficient. The frame interval time. It is the center of gravity of the cavitation mask. This represents the total number of pixels within the bubble region. For the first The coordinates of a bubble pixel; It is the center of the principal axis of the cavitation-fitted ellipse. =(h,j), where h represents the x-coordinate of the center point of the ellipse and j represents the y-coordinate of the center point of the ellipse. It is the maximum diameter of the cavitation bubble.

4. The rapid identification method for underwater firing muzzle cavitation according to claim 1, characterized in that, The following iterative algorithm is used to solve the optical correction problem: (1) Initial estimation: Assuming an initial image = ; in, The original observation image, This is the target's true image estimated in the 0th iteration; (2) Transmittance update: The spectral attenuation coefficient of the water body is obtained by looking up the table based on the turbidity of the water body. Combined with the distance map from the calibrated target to the camera Calculate the estimated value of the transmittance map at the k-th iteration. ; (3) Background light estimation: Take the average value of the darkest 0.1% pixel area of ​​the image to estimate the background ambient light intensity. ; (4) Real Image Update: Based on the result of the k-th iteration, estimate the image after the (k+1)-th iteration. True radiation intensity at coordinates : ， Among them The intensity of the observed image; (5) Noise suppression: The image obtained in the (k+1)th iteration Smoothing is achieved by applying guided filtering to suppress additive noise η. ; (6) Stop iteration when the mean square error (MSE) of two consecutive iterations is less than the threshold: ， in, The height of the image. The width of the image. Let be the true radiation intensity of the target at the k-th iteration, and let be the final result obtained after the iteration converges. As the corrected true image intensity .

Images