Method for automatic measurement and verification of multi-source wave parameters based on neural network

Abstract

This invention proposes an automated measurement and verification method for multi-source wave parameters based on neural networks. This method can quantitatively calculate and acquire wave elements, improving the accuracy of wave element measurement. The invention extracts video images frame by frame from nearshore monitoring videos, extracts the features of the buoy based on its motion state on the sea surface, and quantitatively calculates and acquires wave elements, mainly including effective wave period, effective wave height, and wave direction. Simultaneously, it acquires effective wave period, effective wave height, and wave direction information through the wave sensor built into the float, and calculates the wave direction using the float's Beidou module and anchor position. By comparing and verifying in-situ measurements from the float with non-contact measurements from a camera, the methods complement each other, improving the accuracy of wave element measurement.

Description

Technical Field

[0001] This invention relates to the field of ocean wave measurement technology, and specifically to an automated measurement and verification method for multi-source wave parameters based on neural networks. Background Technology

[0002] Ocean waves are a common physical phenomenon in the ocean, and monitoring and studying them has significant scientific and practical value. During their propagation towards the nearshore area, ocean waves are significantly influenced by seabed topography, shoreline boundaries, and environmental currents (nearshore currents and tidal currents), exhibiting more complex evolutionary patterns and faster spatiotemporal changes than in the deep sea and open continental shelf waters. Current research and understanding of ocean waves are still incomplete. Research on ocean waves mainly includes the study of wave elements and the relationships between these elements. Wave elements include effective wave period, mean wave height, effective wave height, wave direction, and wave steepness. Currently, nearshore wave element detection primarily relies on buoy observation, supplemented by manual observation, although radar observation has been actively promoted in recent years. Buoy observation is a point-based method; accurate wave measurement in the complex topography of harbors requires high-density deployment and incurs high operation and maintenance costs. Manual observation relies on experienced forecasters visually estimating wave information, which demands highly skilled personnel, and the prediction frequency and accuracy are difficult to guarantee. Ground-wave radar can achieve large-area, long-term automatic measurements, but radar measurement equipment is expensive, and microwave pulse and echo propagation are affected by the ionosphere, stratosphere, and air-sea interface, resulting in significant interference and spatiotemporal variations in data quality. Furthermore, measurement accuracy depends on signal inversion. Using shore-based monitoring video for wave element measurement offers advantages such as non-contact operation, low cost, and spatiotemporal continuity. Additionally, if a buoy or float is lost, it can be retrieved to some extent using BeiDou navigation. In-situ float measurement and non-contact camera measurement can be compared and verified, complementing each other and improving wave element measurement accuracy. Moreover, with advancements in artificial intelligence technology, camera measurement solutions show promising prospects, with low maintenance costs. Once the technology matures, the float module can be eliminated.

[0003] Currently, there are some studies both domestically and internationally on the detection of ocean wave elements using visual data. These methods mainly fall into two categories: photogrammetry-based methods and image / video feature-based methods, including statistical features, transform domain features, and texture features. Among these, stereo vision-based wave element analysis largely relies on video images and uses stereo vision systems for wave element detection. While it offers high wave height resolution accuracy, the models are complex, requiring parameter resetting for different marine environments, resulting in poor robustness and low computational efficiency, thus failing to adequately meet practical applications. Video-based wave element detection primarily detects wave direction and wave height level, but it cannot achieve high-precision wave height acquisition. Wave direction detection models are also highly complex and computationally inefficient. Image feature-based wave level threshold models suffer from incompleteness in manually designed features, leading to instability in wave height level detection and preventing refined wave height detection.

[0004] In summary, a survey of existing wave monitoring equipment both domestically and internationally revealed that automated observation equipment is expensive, while manual observation requires a high level of experience from the observer, is highly subjective, and is prone to data inconsistency. Summary of the Invention

[0005] In view of this, the present invention proposes an automated measurement and verification method for multi-source wave parameters based on neural networks, which can quantitatively calculate and obtain wave elements and improve the measurement accuracy of wave elements.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A method for automated measurement and verification of multi-source wave parameters based on neural networks includes the following steps: A float with a built-in BeiDou module transmits wave sensor data to a shore base station via BeiDou satellites. The wave sensor data includes wave height, wave period, and wave direction, used for subsequent wave calibration and verification. A neural network model is trained using the single-stage target detection algorithm YOLOv5. The neural network model is then used to locate and identify the float and buoy in an image dataset, obtaining an ordered set of positions for the float and buoy in the image dataset. The relationship between the longitudinal changes in the positions of the float and buoy in the dataset and the float wave height is confirmed through fitting calibration. The wave height is obtained using the relationship with the float wave height and the changes. The wave period is confirmed using the longitudinal change period of the positions of the float and buoy in the dataset. The wave direction is confirmed using the positional relationship between the positions of the float and buoy in the dataset and the anchor placement position. This achieves automated measurement of wave height, wave direction, and wave period, and the wave results from the float and buoy are cross-verified.

[0008] The specific implementation steps are as follows:

[0009] Step 1: Deploy anchors, buoys, and surface drift buoys on the shore, and install cameras. The anchor is deployed at position P, and the camera is at a vertical distance H from the sea surface. The anchors, buoys, and surface drift buoys are all within the camera's field of view, and the camera video observes the undulation of the buoys and surface drift buoys on the sea surface.

[0010] Step 2: The surface drift transmits wave sensor data, including wave height H, to the shore base station via the BeiDou satellite. 表漂 Wave cycle T 表漂 Wave towards D 表漂 This is used for subsequent wave calibration and verification;

[0011] Step 3: Acquire and store sea surface images with floats and buoys every 5 minutes using shore-based cameras to accumulate neural network data training set DataA;

[0012] Step 4: Label the positions of the float and the float in the data, train the model ModelA using the single-stage target detection algorithm YOLOv5, then use the model to label all the data, and manually proofread and modify it to obtain the labeled dataset DataB.

[0013] Step 5: Train the labeled dataset DataB using the single-stage object detection algorithm YOLOv5 to obtain the model ModelB until it approaches saturation;

[0014] Step 6: If the sea surface remains unchanged within a set time, then the shore-based camera will acquire one image every m frames, accumulating n frames to obtain the image dataset DataC.

[0015] Step 7: Use the neural network model ModelB to locate and identify the floats and buckets in the dataset DataC, and obtain the ordered set of positions of the floats and buckets in the image dataset.

[0016] Step 8: Compare the longitudinal change Δh between the central float and buoy positions with the float wave height H. 表漂 Fitting and calibration were performed to confirm the relationship between the change Δh and the surface drift height H. 表漂 relation;

[0017] Step 9, through Δh and H 表漂 The relationship between the wave height H and the change Δh yields the wave height H. 光电 ;

[0018] Step 10: Confirm the wave period T by analyzing the longitudinal changes in the positions of the float and buoy in the data set. 光电 ;

[0019] Step 11: Confirm the wave direction D by analyzing the positional relationship between the float and buoy positions in the data set and the anchor placement position P. 光电 ;

[0020] Step 12: Compare the wave height, wave direction, and wave period results of the calibration table drift and photoelectric data.

[0021] The surface drifting buoy has a built-in wave sensor module that transmits wave height, wave direction, and wave period results to the buoy's internal ARM program. At the same time, the ARM program communicates normally with the Beidou module and sends serial port messages to Beidou through a fixed signal transmission protocol. The central station's Beidou module receives the Beidou messages and decodes, verifies, and parses the messages through the transmission protocol to obtain the wave height, wave direction, and wave period results of the buoy.

[0022] Specifically, step 10 includes the following steps: calculating the vertical height difference R between each frame and the first frame using the vertical change Δh from the location set List; quantifying the difference between each frame and the first frame in the image set; obtaining the interval frame number dNum between the smaller values ​​in the set using the sum of squares error set Rs; obtaining the total number of frames in one wave cycle using m*dNum; and dividing by the camera's frame rate per second to obtain the wave cycle T. 光电 .

[0023] In step 11, the wave direction D is obtained by the positional relationship between the anchor, the buoy, and the surface drifting buoy. 光电 This includes the following steps:

[0024] Calculate the angle Angle1 between the float P(x2, y2) and the anchor placement position P(x1, y1), and the angle Angle2 between the buoy P(x3, y3) and the float P(x2, y2):

[0025] Angle1=(90-math.degrees(math.atan2(y2-y1,x2-x1))+360)%360

[0026] Angle2=(90-math.degrees(math.atan2(y3-y2,x3-x2))+360)%360

[0027] Compare Angle1 and Angle2. If the difference between them is large, discard the result; otherwise, store it in the collection ListDir.

[0028] Correct each wave direction dir in the set ListDir to obtain the wave direction set ListDir2;

[0029] Take the average value of the wave directions dir in the set ListDir2 as the wave direction D. 光电 Output.

[0030] Beneficial effects:

[0031] 1. This invention extracts video images frame by frame from nearshore monitoring videos, extracts the float's features based on its movement on the sea surface, and quantitatively calculates wave elements, mainly including effective wave period, effective wave height, and wave direction. Simultaneously, it acquires effective wave period, effective wave height, and wave direction information through the float's built-in wave sensor, and calculates the wave direction using the float's BeiDou module and anchor position. By comparing and verifying in-situ measurements from the float with non-contact measurements from a camera, the accuracy of wave element measurement is improved.

[0032] 2. This invention serves as an auxiliary wave observation method to compensate for the lack of data during periods when buoys are not in place and the low resolution of data during manual observations. It improves the continuity and accuracy of wave observation data from marine stations, significantly reduces the manual observation component of wave observation projects, enhances the reliability and scientific rigor of marine observations, and solves key problems such as wave image recognition, analysis and processing, data quality control, data acquisition, and system integration.

[0033] 3. In this invention, the float has a built-in Beidou module that sends wave sensor data to the shore base station via Beidou satellite. The model is trained using the single-stage target detection algorithm YOLOv5. The neural network model is used to locate and identify the float and buoy in the image dataset, resulting in an ordered set of positions of the float and buoy in the image dataset, which can achieve stable detection.

[0034] 4. This invention uses the longitudinal change Δh of the position of the central float and buoy in the data and the wave height H of the float. 表漂 Fitting and calibration were performed to confirm the relationship between the change Δh and the surface drift height H. 表漂 Relationship; using the above relationship and the change Δh, the wave height H is obtained. 光电 The wave period T is determined by the longitudinal change period of the positions of the float and buoy in the data. 光电 The wave direction D was confirmed by using the positional relationship between the float and buoy positions in the data set and the anchor placement position P. 光电 This allows for the automated measurement of wave height, wave direction, and wave period, and cross-verification based on wave results from the float and buoy, thereby improving the accuracy of wave element measurement and enabling refined detection of wave height values. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the principle of the method of the present invention.

[0036] Figure 2 This is a schematic diagram showing the longitudinal position change of the pontoon as the waves change in this invention.

[0037] Figure 3 This is a schematic diagram of the movement of the float and the buoy with the waves in this invention.

[0038] Figure 4 This is a schematic diagram of the viewing angle of the photoelectric camera in this invention. Detailed Implementation

[0039] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0040] This invention proposes an automated measurement and verification method for multi-source wave parameters based on neural networks. The principle is as follows: Figure 1 As shown, the process includes the following steps: The float with a built-in Beidou module transmits wave sensor data to the shore base station via Beidou satellites. The wave sensor data includes wave height H.表漂 Wave cycle T 表漂 And the waves towards D 表漂 This is used for subsequent wave calibration and verification. A neural network model is trained using the single-stage target detection algorithm YOLOv5. This neural network model is then used to locate and identify floats and buoys in the image dataset, resulting in an ordered set of float and buoy positions in the image dataset. The vertical change Δh between the float and buoy positions in the dataset and the float wave height H are then compared. 表漂 Fitting and calibration were performed to confirm the relationship between the change Δh and the surface drift height H. 表漂 Relationship; using Δh and the drift height H 表漂 The wave height H is obtained from the relationship and the change Δh. 光电 The wave period T is determined by the longitudinal change period of the positions of the float and buoy in the data set. 光电； The wave direction D is determined by using the positional relationship between the float and buoy positions in the data set and the anchor placement position P. 光电 This enables automated measurement of wave height, wave direction, and wave period, and cross-validation based on wave results from the float and buoy.

[0041] Based on the characteristics of nearshore ocean wave video images, the specific steps of the method of the present invention are as follows:

[0042] Step 1: Deploy anchors, buoys, and surface drifting buoys (hereinafter referred to as surface drifters) 1-2 km offshore. Install cameras. The anchor is deployed at position P, the camera is vertically distanced from the sea surface at distance H, and the camera is located at coordinate S. The anchor, buoy, and surface drifting buoy must all be within the camera's field of view, and the camera video must clearly show the undulating state of the buoy and surface drifting buoy on the sea surface; the longitudinal position of the buoy changes with the waves as follows: Figure 2 As shown, the float and the buoy move with the waves as follows: Figure 3 As shown, the view frustum of the photoelectric camera is illustrated at various angles. Figure 4 As shown; the float must have a built-in wave sensor and Beidou module. In this embodiment, the float has a built-in wave sensor module, Beidou module, battery, solar controller and solar charging panel, which are used for data acquisition, signal transmission, power storage, charging control and power supply to maintain the operation of the system module, respectively.

[0043] Step 2: The float's built-in wave sensor measures waves, and its BeiDou module transmits wave sensor data, including wave height H, to the shore base station via BeiDou satellites. 表漂 Wave cycle T 表漂 Wave towards D 表漂This is used for wave calibration and verification in subsequent photoelectric measurements. The surface drifting buoy's built-in wave sensor module can transmit wave height, direction, and period results to the buoy's internal ARM program. Simultaneously, the ARM program can communicate normally with the BeiDou module and send serial messages to BeiDou via a fixed signal transmission protocol. The central station's BeiDou module can receive BeiDou messages and decode, verify, and parse them according to the transmission protocol to obtain the buoy's wave height, direction, and period results.

[0044] Step 3: Acquire and store sea surface images with floats and buoys every 5 minutes using a shore-based high-magnification star-level camera to accumulate neural network data training set DataA (tens of thousands of images).

[0045] Step 4: Label the positions of the float and the float in the data in a small range (hundreds of images). Train the model ModelA using the single-stage object detection algorithm YOLOv5. Then use the model to label all the data and manually proofread and modify it to obtain the labeled dataset DataB.

[0046] The overall architecture of YOLOv5 remains the same for different network sizes (n, s, m, l, x). The difference lies in the different depths and widths used in each submodule. The model is downsampled by 32 times and uses three prediction feature layers, corresponding to the depth_multiple and width_multiple parameters in the yaml file. Since the labeled data in this case has a simple shape and obvious features, mask_yolov5s.yaml is sufficient.

[0047] Step 5: Train the labeled dataset DataB using the single-stage object detection algorithm YOLOv5 to obtain model ModelB. Observe whether the localization and recognition training set loss gradually decreases and then tends to saturate during the training process, and whether the accuracy reaches the expected standard, confirming that ModelB has achieved the expected localization and recognition accuracy. In this embodiment, the accuracy reaches over 99.5%.

[0048] In YOLOv5, the loss consists of two parts: positive and negative samples. The negative samples correspond to the background of the image. If there are far more negative samples than positive samples, the negative samples will overwhelm the loss of the positive samples, thereby reducing the efficiency of network convergence and detection accuracy. The loss formula is as follows.

[0049]

[0050]

[0051]

[0052] L box =Lbox +L cls +L obj (4)

[0053] In the formula:

[0054] S: S×S grids;

[0055] B: Each grid generates B candidate anchor boxes;

[0056] If there is a target (positive sample) in the box at position i or j, its value is 1; otherwise, it is 0.

[0057] If there is no target (negative sample) in the box at position i, j, its value is 1; otherwise, it is 0.

[0058] The box loss function at points i and j;

[0059] The cls loss function at points i and j;

[0060] The loss function of obj at points i and j.

[0061] Step 6: During the automated measurement of wave height, wave direction, and wave period, it is assumed that the sea state remains unchanged for a short period of time (10 minutes). An image is acquired every m frames using a shore-based camera, and a dataset of n frames, DataC, is accumulated.

[0062] Step 7: Using the neural network model ModelB, locate and identify the floats and buoys in the dataset DataC, obtaining a list of ordered positions of the floats and buoys in the image dataset. Approximately, the ordered position changes of the floats and buoys in the image dataset can be seen as the single-point movement trend of ocean waves on the sea surface, indirectly reflecting wave height, wave direction, and wave period.

[0063] Step 8: Compare the longitudinal change Δh of the position of the float and buoy in the data set with the float wave height H. 表漂 Fitting and calibration were performed to confirm the relationship between the change Δh and the surface drift height H. 表漂 relation.

[0064] H 表漂 =F(Δh) (5)

[0065] Step 9: Obtain the wave height H using formula (5) in process eight and the change Δh. 光电 .

[0066] Step 10: Confirm the wave period T by analyzing the vertical changes in the positions of the float and buoy in the dataset List. 光电The forward movement of waves as perceived by the eye is the result of the combined action of ocean currents and waves. Waves cause vertical undulation at a single point, while ocean currents cause the waves to move forward. The vertical changes in the positions of the floats and buoys in the dataset are transformed into periodic changes in the waves. By denoising, smoothing, peak value extraction, and calculating the average interval Δt between peak values ​​on the list, the wave period T can be converted into the wave period T. 光电 .

[0067] The vertical position difference quantization includes the following steps: Calculate the vertical height difference R between each frame and the first frame using the vertical change Δh from the position set List; quantize the difference between each frame and the first frame in the image set; obtain the interval frame number dNum between the smaller values ​​in the sum of squares error set Rs; obtain the total number of frames in one wave cycle using m*dNum; divide this by the camera's frame rate per second to obtain the wave cycle T. 光电 .

[0068] Step 11: Theoretically, the positions of the buoy and the float will move in the direction of wave energy propagation. Furthermore, due to the anchor chain pulling the buoy, and the buoy connecting to the float, the anchor, buoy, and float will be aligned in a straight line, in the same direction as the waves. Field verification shows that the positional relationship between the float and buoy and the anchor placement position P conforms to the theoretical assumptions.

[0069] The wave direction D was determined by analyzing the positional relationship between the float and buoy positions in the dataset and the anchor placement position P. 光电 This includes the following steps:

[0070] Calculate the angle Angle1 between the float P(x2, y2) and the anchor placement position P(x1, y1), and the angle Angle2 between the buoy P(x3, y3) and the float P(x2, y2):

[0071] Angle1=(90-math.degrees(math.atan2(y2-y1,x2-x1))+360)%360

[0072] Angle2=(90-math.degrees(math.atan2(y3-y2,x3-x2))+360)%360

[0073] Theoretically, the positions of the buoy and the float will move in the direction of wave energy propagation. Furthermore, due to the anchor chain pulling the buoy, and the buoy connecting to the float, the anchor, buoy, and float will be aligned in a straight line, in the same direction as the waves. Field verification shows that the positional relationship between the float and buoy and the anchor placement position P conforms to the theoretical assumptions.

[0074] Compare angles Angle1 and Angle2. If the difference between them is large, discard the result. Otherwise, consider (Angle1 + Angle2) / 2 as the wave direction dir for a single image and store it in the set ListDir. Due to the tilted angle of the camera, the image is offset. Correct each wave direction dir in the set ListDir to obtain the wave direction set ListDir2.

[0075] The average value of the wave directions dir in the set ListDir2 can be used as the wave direction D. 光电 Output.

[0076] Step 12: Compare the wave height, wave direction, and wave period results of the calibration table and photoelectric sensor to achieve automatic measurement and calibration (correctness or incorrectness) of wave height, wave direction, and wave period.

[0077] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automated measurement and verification method for multi-source wave parameters based on neural networks, characterized in that, The process includes the following steps: The float with a built-in Beidou module sends wave sensor data to the shore base station via Beidou satellite. The wave sensor data includes wave height, wave period, and wave direction, which are used for subsequent wave calibration and verification. A neural network model is trained using the single-stage target detection algorithm YOLOv5. The neural network model is then used to locate and identify the float and buoy in the image dataset, resulting in an ordered set of positions of the float and buoy in the image dataset. By fitting and calibrating the longitudinal changes in the positions of the float and buoy in the data set with the float wave height, the relationship between the changes and the float wave height is confirmed; the wave height is obtained by using the relationship with the float wave height and the changes; the wave period is confirmed by using the longitudinal change period of the float and buoy positions in the data set; and the wave direction is confirmed by using the positional relationship between the float and buoy positions in the data set and the anchor placement position. This achieves automated measurement of wave height, wave direction, and wave period, and cross-validates the wave results of the float and buoy. The specific implementation steps are as follows: Step 1: Deploy anchors, buoys, and surface drift buoys on the shore, and install cameras. The anchor is deployed at position P, and the camera is at a vertical distance H from the sea surface. The anchors, buoys, and surface drift buoys are all within the camera's field of view, and the camera video observes the undulation of the buoys and surface drift buoys on the sea surface. Step 2: The surface drift transmits wave sensor data, including wave height H, to the shore base station via the BeiDou satellite. 表漂 Wave cycle T 表漂 Wave towards D 表漂 This is used for subsequent wave calibration and verification; Step 3: Acquire and store sea surface images with floats and buoys every 5 minutes using shore-based cameras to accumulate neural network data training set DataA; Step 4: Label the positions of the float and the float in the data, train the model ModelA using the single-stage target detection algorithm YOLOv5, then use the model to label all the data, and manually proofread and modify it to obtain the labeled dataset DataB. Step 5: Train the labeled dataset DataB using the single-stage object detection algorithm YOLOv5 to obtain the model ModelB until it approaches saturation; Step 6: If the sea surface remains unchanged within a set time, then the shore-based camera will acquire one image every m frames, accumulating n frames to obtain the image dataset DataC. Step 7: Use the neural network model ModelB to locate and identify the floats and buckets in the dataset DataC, and obtain the ordered set of positions of the floats and buckets in the image dataset. Step 8: Compare the longitudinal change Δh between the central float and buoy positions with the float wave height H. 表漂 Fitting and calibration were performed to confirm the relationship between the change Δh and the surface drift height H. 表漂 relation; Step 9, through Δh and H 表漂 The relationship between the wave height H and the change Δh yields the wave height H. 光电 ; Step 10: Confirm the wave period T by analyzing the longitudinal changes in the positions of the float and buoy in the data set. 光电 ; Step 11: Confirm the wave direction D by analyzing the positional relationship between the float and buoy positions in the data set and the anchor placement position P. 光电 ; Step 12: Compare the wave height, wave direction, and wave period results of the verification table drift and photoelectric data; Step 10 specifically includes the following steps: quantifying the vertical changes in the position set List. h calculates the vertical height difference R between each frame and the first frame, quantizes the difference between each frame and the first frame in the image set, and obtains the interval frame number dNum between the smaller values ​​in the set through the sum of squares error set Rs. The total number of frames in one wave cycle is obtained by m*dNum, and divided by the camera's frame rate per second, the wave cycle T is obtained. 光电 ; In step 11, the wave direction D is obtained by the positional relationship between the anchor, the buoy, and the surface drifting buoy. 光电 This includes the following steps: Calculate the angle Angle1 between the float P(x2, y2) and the anchor placement position P(x1, y1), and the angle Angle2 between the float P(x3, y3) and the float P(x2, y2): Angle1=(90-math.degrees(math.atan2(y2-y1,x2-x1))+360)%360 Angle2=(90-math.degrees(math.atan2(y3-y2,x3-x2))+360)%360 Compare Angle1 and Angle2. If the difference between them is large, discard the result; otherwise, store it in the collection ListDir. Correct each wave direction dir in the set ListDir to obtain the wave direction set ListDir2; Take the average value of the wave directions dir in the set ListDir2 as the wave direction D. 光电 Output.

2. The method according to claim 1, characterized in that, The surface drifting buoy's built-in wave sensor module transmits wave height, wave direction, and wave period results to the buoy's internal ARM program. At the same time, the ARM program communicates normally with the Beidou module and sends serial port messages to Beidou through a fixed signal transmission protocol. The central station's Beidou module receives the Beidou messages and decodes, verifies, and parses them through the transmission protocol to obtain the buoy's wave height, wave direction, and wave period results.

Images