Underwater intelligent light supplementing method and system for deep and far sea culture net cage

Abstract

The invention relates to the technical field of deep and far sea culture, in particular to an underwater intelligent light supplementing method and system for a deep and far sea culture net cage, and the method comprises the steps: obtaining video image data of fish school movement, and carrying out the preprocessing of the obtained video image data; taking the preprocessed video image data as input to construct a fish identification model, performing fish school behavior analysis by using an output result of the identification model, generating a dynamic light supplementing strategy through a fuzzy PID control algorithm based on a fish school behavior analysis result, converting the dynamic light supplementing strategy into electric signal driving, and performing fish school behavior analysis; the light supplement strategy is adaptively optimized according to the real-time illumination parameters, the weight of the fuzzy rule base is updated online through the recursive least square method, the light intensity error and the change rate of the light intensity error serve as input, the fuzzy rule base is constructed through the fuzzy PID control algorithm, the PID parameter adjustment amount is calculated in a defuzzification mode in combination with the gravity center method, and the light supplement strategy is optimized. The light supplementing strategy can be accurately adjusted according to the fish shoal behavior and the actual illumination condition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep-sea aquaculture, and in particular to an underwater intelligent light supplementing method and system for a deep-sea aquaculture net cage. BACKGROUND

[0002] As an important way to break the bottleneck of near-sea resources and realize the sustainable development of blue economy, the large-scale development of deep-sea net cage aquaculture has strategic significance for guaranteeing high-quality protein supply and food security. However, the high dynamic environment (such as strong ocean current, salinity change, and extreme weather) of open deep-sea waters and the high-density aquaculture demand pose severe challenges to the precision and intelligence of aquaculture technology. As a core means of regulating fish behavior and physiological metabolism, the technical level of net cage light supplementing directly affects the fish aggregation efficiency, growth rate, and optimization of aquaculture density.

[0003] However, traditional light supplementing systems mostly use fixed spectra (such as single blue light or white light) and constant light intensity, which cannot dynamically match the phototactic characteristics of different fish (such as grouper preference for blue light and salmon sensitivity to red light) and growth stage requirements, such as high-spectrum stimulation of metabolism during the juvenile stage and low light to suppress stress during the adult stage. Secondly, existing systems lack remote intelligent diagnosis capabilities for equipment operating status. In the high-corrosion and high-humidity environment of deep-sea, the response to lamp failure is lagging, manual maintenance costs are high, and there are safety hazards, and even safety accidents such as electrical leakage. Finally, existing technologies cannot realize real-time coordination of light parameters (spectrum, intensity, and photoperiod) and fish behavior rhythms, such as rapidly increasing the intensity of specific wavebands to stimulate appetite during the active feeding period of the fish population, and automatically reducing the light intensity to reduce energy consumption during the rest period, resulting in low light energy utilization and limited aquaculture efficiency. At present, there is a need for an underwater intelligent light supplementing method and system for a deep-sea aquaculture net cage. SUMMARY

[0004] To solve the problems of insufficient light parameter adaptability and lack of dynamic coordination and regulation in traditional light supplementing systems, the present application provides an underwater intelligent light supplementing method and system for a deep-sea aquaculture net cage, which builds a deep computer vision model to identify, analyze, and adjust the light supplementing strategy in real time, thereby improving the growth rate of farmed fish.

[0005] In the first aspect, the present application provides an underwater intelligent light supplementing method for a deep-sea aquaculture net cage, which adopts the following technical solution:

[0006] An underwater intelligent light supplementing method for a deep-sea aquaculture net cage, comprising:

[0007] Obtaining video image data of fish movement and pre-processing the obtained video image data;

[0008] The pre-processed video image data is taken as input to build a fish recognition model, including building a backbone network, a feature fusion network and a detection head based on a YOLOv8 deep learning model;

[0009] The output result of the recognition model is used for fish school behavior analysis, including performing gray difference operation on continuous video frames by using two-frame difference method, and generating a feeding intensity quantitative index by matching a fuzzy rule base through a Gaussian membership function;

[0010] A dynamic light supplement strategy is generated based on the fish school behavior analysis result by using a fuzzy PID control algorithm, including solving the optimal wave band according to the phototaxis characteristics of fish species by solving the PID parameter adjustment amount through the barycentric method;

[0011] The dynamic light supplement strategy is converted into an electrical signal to drive the LED lamp to adjust the illumination parameter;

[0012] The light supplement strategy is adaptively optimized according to the real-time illumination parameter, including updating the weight of the fuzzy rule base online by using the recursive least square method.

[0013] Further, the backbone network, the feature fusion network and the detection head based on the YOLOv8 deep learning model include using a CSPDarknet structure with RepVGG reparameterization as the backbone network, introducing cross-stage partial connection into the backbone network for multi-scale feature extraction, and inputting the extracted feature map into the feature fusion network, the feature fusion network uses a PANet++ network combined with a bidirectional feature pyramid structure to fuse multi-scale features through an adaptive weight distribution mechanism, and the multi-scale feature fusion formula is:

[0014]

[0015] wherein ω i represents the learnable weight of the i-th layer feature, F i represents the feature map of the i-th layer.

[0016] Further, the backbone network, the feature fusion network and the detection head based on the YOLOv8 deep learning model further include inputting the fused feature map into the detection head for target detection and classification, the detection head has a classification branch and a regression branch decoupled and independent, wherein the classification branch uses binary cross entropy loss to predict fish species labels, the regression branch uses CIoULoss to optimize the boundary box coordinates, and three-dimensional space positioning is performed based on the optimized boundary box coordinates, and the optimization formula is:

[0017]

[0018] wherein, p is the distance between the prediction box and the real box center point, c is the minimum bounding box diagonal length, v measures the consistency of the aspect ratio, a is the balance coefficient, loU is the intersection over union, indicating the overlap degree of the prediction box b and the real box b gt .

[0019] Further, the three-dimensional space positioning based on the optimized bounding box coordinates comprises: acquiring an internal parameter matrix and a distortion coefficient of a binocular camera, for correcting image distortion and establishing a mapping relationship between pixel coordinates and a camera coordinate system; calculating an external parameter matrix of the two cameras, establishing a conversion relationship of a binocular vision coordinate system; performing feature matching on left and right camera images, acquiring homonymous point pixel coordinates; calculating fish three-dimensional coordinates based on the homonymous point pixel coordinates through a triangulation principle; and finally associating the three-dimensional position information with a fish species label.

[0020] Further, the fish school behavior analysis using the output result of the recognition model comprises: acquiring a fish species classification label; calculating a feeding intensity index through behavior recognition; constructing a multi-dimensional feature vector using the feeding intensity index, and defining a feeding intensity fuzzy set through a Gaussian membership function; and triggering a preliminary light supplement control instruction based on the feeding intensity fuzzy set and a preset rule library.

[0021]

[0022] wherein, N active is a feeding behavior pixel area, N total is a total area of a monitoring region, and T is a statistical frame number.

[0023] Further, the two-frame difference method is used to perform a gray scale difference operation on continuous video frames, which comprises: defining K states for each pixel point in a video image, then using a mixed model of K Gauss distributions to establish a background model for the pixel points, performing a gray scale difference between a current frame and a previous frame based on the background model, introducing an Otsu algorithm to adaptively obtain a threshold value, and then obtaining a foreground region in the current image.

[0024]

[0025] wherein, η i represents a probability density function of the i-th Gauss distribution, describing the probability of a pixel value I t belonging to the distribution, μ i,t represents a mean value of the i-th Gauss distribution at time t, represents a central tendency of a background pixel value, σ i,t represents a standard deviation of the i-th Gauss distribution at time t, and ω i represents a weight value of the i-th Gauss distribution.

[0026] Further, the dynamic light supplement strategy is generated based on the fish behavior analysis result through a fuzzy PID control algorithm, including constructing a fuzzy rule base with light intensity error and its change rate as input, fuzzifying the input parameters through a Gaussian membership function, combining a preset fuzzy rule table to solve the fuzzy calculation PID parameter adjustment amount by a gravity method, solving the optimal wave band according to the obtained fish species label, and generating the light supplement strategy in combination with the PID parameter adjustment result and the optimal wave band, wherein the optimal wave band solving formula is:

[0027] λ opt = arg max λ∈[400,700] (A(λ)η LED (λ)),

[0028] wherein λ represents the wavelength of the light supplement spectrum, the value range is 400-700nm, A(λ) represents the spectral responsivity of the fish species to the wavelength λ, η LED (λ) represents the luminous efficiency of the LED lamp at the wavelength λ.

[0029] Further, the fuzzy rule base weight is updated online by using a recursive least square method, including collecting underwater light intensity data through a distributed illuminometer network, performing Kalman filter processing on the collected light intensity data, constructing a feature vector of the light intensity data, defining a fuzzy rule base weight vector, updating the weight vector by using an iteration formula, and completing the fuzzy rule base weight update, wherein the iteration formula is represented as:

[0030]

[0031] wherein φ(k) represents the feature vector of the light intensity data at time k, θ(k) represents the weight vector of the fuzzy rule base at time k, λ1 represents a forgetting factor, P(k) represents a covariance matrix, measuring the uncertainty of parameter estimation, and y(k) represents the light intensity feedback value at time k.

[0032] In the second aspect, an underwater intelligent light supplement system for a deep-sea aquaculture net cage includes:

[0033] An image acquisition module is connected with a global binocular camera, pre-processes the video stream of the global binocular camera, and transmits the pre-processed video stream to a fish species identification module through an armored optical fiber transmission assembly;

[0034] The fish species identification module integrates a species identification model, outputs fish species labels, bounding boxes and three-dimensional coordinates, and transmits the outputs to a fish behavior identification module and a light strategy control module;

[0035] The fish behavior identification module performs gray difference operation on consecutive video frames by using a two-frame difference method, generates a feeding intensity quantitative index, and transmits the feeding intensity quantitative index to the light strategy control module;

[0036] The light strategy control module is used for receiving the fry label and the feeding intensity quantification index, generating a dynamic light supplement strategy through a fuzzy PID control algorithm, and transmitting the light supplement strategy to the driving module.

[0037] The driving module is internally provided with a constant current LED driving circuit, which realizes a linear mapping relationship between the PWM duty cycle and the current value, converts the command of the light supplement strategy control into an electric signal, and transmits the electric signal to the LED lamp assembly.

[0038] The LED lamp module adopts a multi-channel programmable LED array and integrates a wavelength selection type light emitting chip group, which is used for flexibly combining and outputting different wavebands of light spectrum.

[0039] Further, the constant current LED driving circuit comprises a sampling capacitor, an endless dimming control chip and a plurality of LEDs, one end of the sampling capacitor is connected to the positive pole of the circuit input end, the other end of the sampling capacitor is connected to the CMOD capacitor pin of the endless dimming control chip, which is used for collecting signals in the circuit, and the PWM signal output by the endless dimming control chip controls the transistor connected to the LED through a current limiting resistor, thereby realizing the driving of the LED.

[0040] In summary, the present application has the following beneficial technical effects:

[0041] 1. The present application is based on the introduction of the CSPDarknet structure of RepVGG reparameterization to construct the backbone network, combined with the feature fusion network of PANet++ and bidirectional feature pyramid structure, and the decoupled detection head (the classification branch adopts binary cross entropy loss, and the regression branch adopts CIoULoss), which can effectively extract multi-scale features, accurately detect and classify targets, and realize high-precision fish recognition and three-dimensional spatial positioning. Compared with the traditional recognition method, the detection ability of small target fish is improved.

[0042] 2. The present application adopts two-frame difference method combined with mixed Gaussian background model for gray difference operation, which can dynamically distinguish foreground fish school and static background, and effectively eliminate environmental noise interference. By calculating the feeding intensity index and constructing a multi-dimensional feature vector, the fuzzy set trigger preset rule library is defined by using the Gaussian membership function, which can accurately quantify the feeding intensity of fish school and generate reasonable preliminary light supplement control instructions.

[0043] 3. The present application takes light intensity error and its change rate as input, constructs a fuzzy rule library through a fuzzy PID control algorithm, combines the barycentric method to solve the fuzzy calculation PID parameter adjustment amount, and solves the optimal waveband according to the phototactic characteristics of fish species, which can dynamically and accurately adjust the light supplement strategy according to the behavior of fish school and the actual lighting condition.

[0044] 4、The present application adopts recursive least squares method to update fuzzy rule base weight online, combines with distributed illuminometer network to collect light intensity data and Kalman filter processing, can realize real-time sensing of underwater light intensity change, and dynamically adjusts the light supplement strategy. It can quickly adapt to the complex changes of light conditions in deep sea environment, such as light intensity fluctuation caused by factors such as water flow, weather, etc., to ensure that the light supplement system is always in the optimal working state.

[0045] 5、The constant current LED driving circuit of the present application realizes linear mapping of PWM duty cycle and current value through reasonable configuration of sampling capacitor, non-polar dimming control chip, current limiting resistor and transistor, and stabilizes the driving of LED lamps. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a whole process schematic diagram of a kind of underwater intelligent light supplement method for deep sea aquaculture net cage of embodiment 1 of the present application.

[0047] Figure 2 is the network structure diagram of fish identification model of embodiment 1 of the present application.

[0048] Figure 3 is the basic flow chart of two-frame difference method of embodiment 1 of the present application.

[0049] Figure 4 is a kind of underwater intelligent light supplement system schematic diagram for deep sea aquaculture net cage of embodiment 2 of the present application.

[0050] Figure 5 is the constant current LED driving circuit schematic diagram of embodiment 2 of the present application. DETAILED DESCRIPTION

[0051] The present application is further described in detail below in conjunction with the drawings.

[0052] Embodiment 1

[0053] Referring to Figure 1 , the underwater intelligent light supplement method for deep sea aquaculture net cage of the present embodiment comprises:

[0054] Obtain video image data of fish movement, and pre-process the obtained video image data;

[0055] The pre-processed video image data is used as input to construct a fish identification model, which includes constructing a backbone network, a feature fusion network and a detection head based on a YOLOv8 deep learning model;

[0056] The output result of the identification model is used for fish school behavior analysis, which includes performing gray difference operation on consecutive video frames by using two-frame difference method, and generating feeding intensity quantitative index by matching fuzzy rule base through Gaussian membership function;

[0057] Based on the fish behavior analysis results, a dynamic light supplement strategy is generated by a fuzzy PID control algorithm, including solving the optimal wave band according to the phototaxis characteristics of fish species by adjusting the gravity method to solve the fuzzy calculation PID parameter adjustment amount;

[0058] The dynamic light supplement strategy is converted into an electrical signal to drive the LED lamp to adjust the illumination parameters;

[0059] The light supplement strategy is adaptively optimized according to real-time illumination parameters, including updating the fuzzy rule base weight online using the recursive least squares method.

[0060] As shown in Figure 1 , specifically, an underwater intelligent light supplement method for a deep-sea aquaculture net cage includes the following steps:

[0061] S1, obtaining video image data of fish movement, and pre-processing the obtained video image data;

[0062] At the preset monitoring point of the deep-sea aquaculture net cage, a global binocular camera is fixedly installed for obtaining video image data of fish movement, and a median filter algorithm is used to process the collected video image. In the specific operation, for each pixel point in the video image, a filter window is set, the pixel values in the window are sorted according to the gray value size, and the middle value is taken to replace the gray value of the current pixel point. In this way, the impulse noise such as salt and pepper noise in the image can be effectively removed, and the edge and detail information of the image can be preserved to a certain extent, so that the noise enhancement of the video image is realized. Then, in view of the blue-green color deviation caused by the underwater environment, a white balance algorithm is used for color correction. First, the video image is converted from the common RGB color space to the Lab color space, in which the L channel represents the brightness, and the a channel and the b channel represent the red-green and yellow-blue components of the color, respectively. By analyzing the color components of the representative white or gray areas in the image, the corresponding correction coefficients are calculated, and then the a channel and the b channel of the image are adjusted according to these correction coefficients to balance the color and reduce the blue-green color deviation, so that the image color is closer to the real scene.

[0063] The video image after color correction is uniformly scaled to the pre-set model input size. Bilinear interpolation or other methods can be used for scaling operation, and the video image after size scaling is normalized. The pixel value range of the image is normalized to a specific range such as [0, 1] or [-1, 1] from [0, 255]. The specific method is that for the gray value of each pixel point, the formula is:

[0064]

[0065] Where x is the original pixel value, x minx is the minimum pixel value in the image max x is the maximum pixel value in the image, through normalization operation, the image data can have better numerical stability and consistency in subsequent model processing, improve the training and recognition effect of the model.

[0066] S2, the pre-processed video image data is used as input to construct a fish recognition model, including constructing a backbone network, a feature fusion network and a detection head based on a YOLOv8 deep learning model;

[0067] As shown in Figure 2 YOLOv8 deep learning model is improved to realize industrial real-time recognition and positioning, which is suitable for fast detection demand in actual application scenarios. After image data is pre-processed, including size standardization and normalization operation, it is sent to the improved CSPDarknet backbone network for multi-scale feature extraction.

[0068] Subsequently, the feature map enters the fusion network (Neck), which adopts PANet++ combined with BiFPN structure, and the features of different scales are fused through the adaptive weight distribution mechanism. In the PANet++ network, first, the path enhancement from bottom to top is performed, that is, the information is transmitted from the feature map with low resolution and high semantic information to the feature map with high resolution and low semantic information. Through upsampling and feature fusion operation, the high-resolution feature map can also obtain more semantic information. Then, the path fusion from top to bottom is performed, which transmits the detailed information in the high-resolution feature map to the low-resolution feature map, and further fuses the features of different scales.

[0069] The fused features are transmitted to the detection head (Head), in which the classification and regression tasks are separated into independent branches: the classification branch predicts the fish species label based on the binary cross entropy loss, and the regression branch optimizes the boundary box coordinates using the CIoU loss function. At the same time, combined with the stereo calibration data of the binocular camera, the three-dimensional position information of the fish school is calculated through the spatial coordinate positioning algorithm.

[0070] Compared with YOLOv5, the C3 module of YOLOv8 is replaced by C2f module, which improves the feature extraction capability by increasing the gradient flow branch and jump connection. The C2f module integrates more residual structures, reduces the calculation redundancy, and retains more rich multi-scale features, which is especially suitable for fish small target detection in complex underwater scenes; compared with the coupled head (CoupledHead) of YOLOv5, YOLOv8 uses decoupled head (DecoupledHead) to process classification and positioning through independent branches, which can more accurately capture the shape and species characteristics of fish.

[0071] Finally, the model outputs the detection results containing fish species, bounding boxes, and depth coordinates, and filters redundant predictions through non-maximum suppression (NMS) to ensure the accuracy and robustness of the detection results. The improved model continues the single-stage detection framework of the YOLO series, and the core structure includes three parts: backbone, neck, and head. Specifically, it includes:

[0072] The improved Backbone adopts the improved CSPDarknet structure, introduces cross-stage partial connection, reduces computational redundancy through gradient shunting, and the core module integrates the reparameterization idea of RepVGG. During training, the multi-branch structure is used to improve the feature extraction capability. The specific formula is:

[0073] F conv =Conv(Concat(X,Conv(X))),

[0074] where X is the input feature map, Concat represents channel concatenation, and Conv is the convolution operation. During inference, it is combined into a single path to improve speed, using the PANet++ structure, combining BiFPN and ASFF, balancing the importance of different scale features through learnable weights. The specific formula is:

[0075] and ∑ω i =1,

[0076] where ω is the weight coefficient of the i-th layer feature. Through backpropagation optimization, the network can automatically adjust the fusion weight of each layer feature according to different input images and task requirements, thereby enhancing the detection ability of different scale targets (especially small target fish).

[0077] The classification task and the regression task are separated into two independent branches to avoid performance degradation caused by feature coupling. The classification branch uses binary cross-entropy loss, and the regression branch uses CIoULoss (Complete Intersection over Union Loss) to optimize the position of the bounding box. The specific formula is:

[0078]

[0079] where ρ is the distance between the predicted box and the real box center point, c is the diagonal length of the minimum bounding box, v measures the consistency of the aspect ratio, α is the balance coefficient, loU is the intersection over union, and represents the predicted box b and the real box b gtThe degree of overlap of the bounding boxes, the fused feature map is input into a regression sub-network, and the coordinate parameters (such as the center point coordinates, width and height) of the predicted bounding box are output. During training, the CIoU Loss between the predicted bounding box and the real bounding box is calculated, and the parameters of the regression branch are updated through the back propagation algorithm, so that the predicted bounding box continuously approaches the real bounding box, and the accuracy of target positioning is improved.

[0080] Finally, based on the optimized bounding box coordinates, three-dimensional space positioning is performed. First, the intrinsic matrix and distortion coefficient of the binocular camera are obtained. Specifically, a chessboard calibration plate with a known size is used to take multiple pictures at different positions and angles in the field of view of the binocular camera. Then, the corner points of the chessboard are detected using image processing algorithms. According to the pixel coordinates and actual physical coordinates of the corner points, the intrinsic matrix (including focal length, principal point coordinates, etc.) and distortion coefficient (used to correct image distortion) are calculated through optimization algorithms such as least squares. At the same time, the extrinsic matrix (rotation matrix and translation vector) of the two cameras is calculated to establish the conversion relationship between the binocular vision coordinate system.

[0081] Feature matching is performed on the left and right camera images. A feature point-based matching algorithm is used. First, feature points are extracted from the left and right images, and feature descriptors are calculated. Then, feature point matching is performed by measuring the distance between descriptors (such as Euclidean distance) to find the same point pixel coordinates in the left and right images. To improve the accuracy and efficiency of matching, the RANSAC algorithm and other methods are used to remove false matching points.

[0082] Finally, based on the same point pixel coordinates, the three-dimensional coordinates of the fish school are calculated using the triangulation principle. According to the geometric model of the binocular camera and the obtained intrinsic matrix and extrinsic matrix, the coordinates of the same point in three-dimensional space are calculated through trigonometric relationships and spatial geometric transformation formulas. Finally, the calculated three-dimensional coordinate information of the fish school is associated with the fish species label obtained through the classification branch, thereby realizing the species identification and three-dimensional space positioning of the fish school.

[0083] S3, using the output results of the recognition model to analyze the behavior of the fish school, including using two-frame difference method to perform gray difference operation on consecutive video frames, and generating feeding intensity quantitative index through Gaussian membership function matching fuzzy rule base;

[0084] As shown in Figure 3 For each pixel point in the video image, define K states (generally K takes a value of 3-5, which can be adjusted according to actual conditions), and for each pixel point, use a mixture model of K Gaussian distributions to establish a background model. According to the formula:

[0085]

[0086] and ωi，t > 0 (i e [1, K]),

[0087] wherein η i represents a probability density function of the i-th Gaussian distribution, describing the probability of a pixel value I t belonging to this distribution, μ i,t represents the mean value of the i-th Gaussian distribution at time t, indicating the central tendency of the background pixel value, σ i,t represents the standard deviation of the i-th Gaussian distribution at time t, ω i represents the weight of the i-th Gaussian distribution,

[0088] The two-frame difference method is selected for target behavior recognition monitoring. The gray difference of adjacent two frames of images in a video sequence is calculated. By setting a gray difference threshold, the pixel points in the image are classified into foreground and background, and the foreground region in the current image is obtained.

[0089] After the background model is established, the gray difference between the current frame and the previous frame is calculated. For each pixel point in the video image, the difference between the current frame pixel value I t and the background model prediction value (i.e. ) is calculated to obtain the gray difference image D(x, y):

[0090]

[0091] wherein (x, y) is the pixel point coordinate. When a moving target appears in the video image, the gray value of the pixel point where the target is located will change greatly between two frames of images. When the absolute value of the gray difference is greater than the threshold set by the algorithm, it is determined that there is a moving object, and the pixel point is determined as foreground, otherwise it is determined as background.

[0092] The specific operation process of the two-frame difference method is expressed as:

[0093] f1(x, y) = [d t (x, y) - d t-1 (x, y)],

[0094] f1(x, y) > R1 FG,

[0095] f1(x, y) ≤ R1 BG,

[0096] wherein d t (x, y) and d t-1(x, y) and (x, y) are the gray values of the pixel point in the current image (t frame) and the previous frame image respectively; f1(x, y) is the two-frame difference of the pixel point, R1 is the threshold value set by Otsu algorithm, Otsu algorithm is usually introduced to adaptively obtain the best threshold value of each two-frame difference, and FG and BG are the detection results, representing the foreground and background respectively.

[0097] After obtaining the foreground region, the feeding intensity index (0-5 levels) is calculated by behavior recognition. First, the total area Ntotal of the monitoring region is determined total , and the pixel area Nactive belonging to the feeding behavior in the foreground region of each frame image is calculated according to the camera field of view range and image resolution. active (t) according to the formula:

[0098]

[0099] Where Nactive is the pixel area of feeding behavior, Ntotal is the total area of the monitoring region, and the calculated feeding intensity index F is combined with the fish species classification label S and other related environmental parameters (water temperature T w , dissolved oxygen concentration DO, environmental light intensity L amb ) to construct a multi-dimensional feature vector X = [S, F, T w , DO, L amb ] T , and the fuzzy logic is combined to define the feeding intensity fuzzy set (such as “low feeding, high feeding”) and its Gaussian membership function to trigger the preset rule base. The specific function can be represented as:

[0100]

[0101] Where b is the width parameter of the membership function. Replace the feeding intensity index F in the multi-dimensional feature vector X with the corresponding Gaussian membership function to calculate the membership degree of each feeding intensity fuzzy set. According to the membership degree, trigger the rules in the preset rule base, for example, if the membership degree of the “high feeding” fuzzy set is greater than a certain threshold value (such as 0.7), trigger the corresponding light supplement control rule to generate the preliminary light supplement control instruction (such as increasing the light intensity, adjusting the spectrum, etc.).

[0102] S4, based on the fish behavior analysis results, generate dynamic light supplement strategy through fuzzy PID control algorithm, including solving the PID parameter adjustment amount through the barycentric method, and solving the optimal wave band according to the phototactic characteristics of fish species;

[0103] Receive fish behavior information, retrieve the spectrum and light intensity requirement of the fish at the real-time growth state from the database, adjust the corresponding light supplement strategy, monitor the fish feeding behavior and growth state in real time, adjust the LED lamp start and stop time according to the fish feeding behavior rhythm, receive the feeding behavior recognition information, and adjust the spectrum and light intensity, including:

[0104] Based on the light intensity error e(t) = L target -L actual And the error change rate Δe(t), the fuzzy rule base is constructed based on the light intensity error (e) and the error change rate (Δe), both of which are divided into fuzzy sets (such as "negative big NB", "negative small NS", "zero ZO", "positive small PS", "positive big PB"), and the rule table adopts the form of "if the light intensity error (e) is A, the error change rate (Δe) is B, then the proportional coefficient K p is C, the integral coefficient K i is D, and K p is E", for example:

[0105] If the light intensity error is large and increases rapidly (e = PB, Δe = PB), increase the proportional coefficient (K p ↑), reduce the integral coefficient (K i ↓); if the error is close to the target and the fluctuation is small (e = ZO, Δe = NS), fine-tune K p , K i , and keep K d smooth to maintain steady-state accuracy, map the light intensity error and error change rate to fuzzy quantities through Gaussian membership functions, and calculate the triggering weight of each rule; activate the matching rule according to the current e and Δe, and calculate the PID parameter adjustment amount (ΔK p , ΔK i , ΔK d ) through the gravity method.

[0106] Adjust the PID parameters K' p , K' i , K' d through fuzzy reasoning, calculate the adjustment factor through the gravity method, and take K' p as an example. The specific formula is:

[0107]

[0108] Where μ i is the rule triggering weight, K p,i is the preset value of the rule base, and n is the total of the weighted adjustment amount of each rule. The final output PID control amount is:

[0109]

[0110] where e(t) represents the light intensity error, K p is the proportional coefficient, K i is the integral coefficient, is the integral cumulative value of historical error, reflecting the total error from the initial time to the current time, K d is the integral coefficient.

[0111] The light intensity error and error change rate are mapped to fuzzy quantities by Gaussian membership functions, and the rule triggering weights are calculated. Then, for different fish species (such as large yellow croaker and grouper), the feeding activity, growth rate and other indicators under different wavelengths (400-700nm) are measured by experiments, and the spectral response function A(λ) is constructed. According to the feeding behavior of fish species and the light emitting efficiency η LED of LED, the optimal waveband is solved:

[0112] λ opt = arg max λ∈[400,700] (A(λ)η LED (λ)),

[0113] where λ represents the wavelength of the light supplement spectrum, with a value range of 400-700nm, A(λ) represents the spectral response of fish species to wavelength λ, η LED (λ) represents the light emitting efficiency of LED lamps at wavelength λ, combined with the PID output u t and the ambient light compensation, the target light intensity is calculated:

[0114] L output =u(t)+α(L amb -L min ),

[0115] where α is the coupling coefficient, L amb represents the ambient light intensity, and L min represents the compensation threshold. According to the fish species label S, the corresponding A(λ) model is dynamically switched to realize the waveband optimization specific to fish species.

[0116] S5, convert the dynamic light supplement strategy into an electrical signal to drive LED lamps to adjust the lighting parameters;

[0117] map the PID controller output to the LED driving current,

[0118]

[0119] where u(k) is the output value of the PID controller at the kth sampling time, I min is the lower limit of the safe LED driving current, I max is the upper limit of the safe LED driving current, and u maxThe maximum design output value of the PID controller is taken as the target spectrum, and then multi-channel spectrum control is performed. According to the calculated optimal wavelength λ opt The preset weight distribution function is used to allocate the driving current proportion of the three LED channels of 450 nm blue light, 520 nm green light and 630 nm red light. The current of each channel is the overall driving current multiplied by the corresponding weight, so that the target spectrum synthesis is realized.

[0120] Finally, at the hardware level, a multi-channel constant current driving circuit is built by using a dimming control chip, and the linear mapping between the PWM duty cycle and the current is realized by using the characteristics of the chip, such as 10% duty cycle corresponding to 200mA current. At the same time, an independent driving channel is configured for each wavelength of the LED array, and the duty cycle command sent by the MCU is received through the SPI interface. Each channel is isolated from each other to ensure accurate control of the output of each LED channel, and finally the dynamic and accurate adjustment of the illumination parameters of the LED lamp is realized.

[0121] S6, adaptively optimizing the light supplement strategy according to real-time illumination parameters, including using recursive least squares method to update the weight of fuzzy rule base online.

[0122] Firstly, the real-time light intensity data of each point in the water is collected through the distributed illuminometer network, and the environmental parameters such as water temperature and depth are also collected. In order to eliminate the noise interference in the data, Kalman filtering is used to process the collected light intensity data to obtain stable and reliable light intensity values. Combining time, space and environmental factors, a light intensity data feature vector containing current light intensity, light intensity change rate, adjacent node light intensity difference and other information is constructed. Secondly, the weight vector of the fuzzy rule base is defined, and the covariance matrix and the forgetting factor are initialized. The forgetting factor is used to balance the influence degree of historical data and new data in weight update. Then, the iterative formula is used:

[0123]

[0124] Wherein, φ(k) represents the feature vector of the light intensity data at time k, θ(k) represents the weight vector of the fuzzy rule base at time k, λ1 represents the forgetting factor, P(k) represents the covariance matrix, which measures the uncertainty of parameter estimation, and y(k) represents the light intensity feedback value at time k. Finally, according to the updated weight vector, the activation strength of each rule in the fuzzy rule base is adjusted, so that the triggering of the rule is more in line with the actual light intensity. Through the evaluation indexes such as rule contribution degree and stability, the effectiveness of the rule is judged, and the invalid rules are removed. When a new light intensity change mode is detected, the rule base is dynamically expanded, so as to realize the adaptive optimization of the light supplement strategy according to the real-time illumination parameters.

[0125] Example 2

[0126] The embodiment differs from embodiment 1 in that the embodiment provides an underwater intelligent light supplementing system for a deep-sea aquaculture net cage, comprising:

[0127] As shown in Figure 4 The underwater intelligent light supplementing system for a deep-sea aquaculture net cage provided by the embodiment comprises multiple functional modules working in cooperation. The image acquisition module is connected with a global binocular camera. The video stream captured by the camera is first preprocessed, such as denoising and color correction, and then transmitted to the fish species recognition module through armored optical fiber. The fish species recognition module integrates a specific recognition model, which can recognize the fish species from the video stream and output the corresponding species label, bounding box and three-dimensional coordinate information. These information is transmitted to the fish behavior recognition module and the light strategy control module. The fish behavior recognition module uses two-frame difference method to perform gray difference operation on consecutive video frames to identify the moving fish school area, and then generates the feeding intensity quantitative index and transmits it to the light strategy control module. After receiving the fish species label and the feeding intensity quantitative index, the light strategy control module generates a dynamic light supplementing strategy by means of fuzzy PID control algorithm, combined with factors such as light intensity error, error change rate and fish species phototaxis characteristics, and transmits the strategy to the driving module.

[0128] As shown in Figure 5 The driving module is built-in constant current LED driving circuit. The circuit acquires circuit signals through a sampling capacitor. The stepless dimming control chip outputs PWM signals according to the signals. The current limiting resistor controls the transistor connected to the LED, realizes the linear mapping of PWM duty cycle and current value, and thus converts the light supplementing strategy instruction into an electrical signal, which is transmitted to the LED lamp assembly. Specifically, the sampling capacitor acquires the current signal in the circuit in real time and transmits it to the CMOD pin of the stepless dimming control chip, forming a closed-loop feedback mechanism. The chip outputs a PWM signal with a corresponding duty cycle according to the preset PID control amount and the real-time current deviation. The PWM signal is input to the transistor base through the current limiting resistor. The amplified current drives the LED to turn on. Each wavelength of LED channel (such as 450nm blue light and 520nm green light) is independently configured with a transistor and a current limiting resistor, and the current proportion of each channel is independently controlled by the PWM signal output from different pins of the chip.

[0129] The channel programmable LED array integrates wavelength selection type light emitting chip group. According to the electrical signal of the driving module, the target spectrum is dynamically combined and output, for example, by adjusting the current proportion of the blue, green and red light channels, any wavelength combination within the range of 400-700nm can be synthesized, which meets the needs of different fish species in the feeding active period (enhancing blue and green light to stimulate appetite), resting period (switching to low-intensity red light to maintain ambient light) and other scenes. The lamp housing is made of corrosion-resistant aluminum alloy material, and the temperature sensor inside the housing is linked to the temperature control module. When the temperature exceeds the safety threshold, the protection mechanism such as power reduction and forced shutdown is automatically triggered to ensure long-term stable operation.

[0130] Embodiment 3

[0131] The embodiment differs from embodiments 1 and 2 in that the embodiment provides a specific integrated manner of an underwater intelligent light supplementing system for a deep-sea aquaculture net cage:

[0132] The fish culture light source intelligent regulation and control system control cabinet is installed with the light strategy control module, the driving module and the temperature control module.

[0133] The temperature control module is configured with an industrial-grade temperature transmitter module, which monitors the temperature change of the equipment in real time, dynamically adjusts the operation state of the heat dissipation device in combination with an active heat dissipation control strategy, and triggers power reduction operation, forced shutdown and remote alarm in sequence when the temperature exceeds a safety threshold, effectively avoiding equipment damage caused by overheating, and balancing system efficiency and stability through optimized heat dissipation energy consumption.

[0134] The outer surface of the control cabinet is sprayed with a plastic process to increase the corrosion resistance of the shell, the cabinet is installed in a floor-mounted manner, the cabinet door is outwardly and right-opens, the heat dissipation fan is built-in, and the temperature control switch can set the heat dissipation temperature and start and stop the heat dissipation fan.

[0135] Optionally, the front panel of the control cabinet has 18 waterproof quick connector female sockets for connecting the controllable light source. The design is free of wiring, plug-and-play, supports hot plug, the socket adopts IP68 protection design, and can ensure the use condition in a rain environment outdoors.

[0136] The controllable light source is linked with the output end of the driving module, is installed with a temperature sensor, is connected with the temperature control module, and monitors the temperature of the lamp; the controllable light source is integrally processed with an aluminum alloy shell and adopts an oxidation method to increase the corrosion resistance, and the lamp panel is protected by tempered glass to prevent damage caused by collision; the controllable light source adopts a fanless passive cooling method, reduces the later maintenance cost, and further improves the equipment operation stability.

[0137] The above are preferred embodiments of the present application, and do not limit the protection scope of the present application, so that: equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. An underwater intelligent lighting method for deep-sea aquaculture cages, characterized in that: include: Acquire video image data of fish school movement and pre-process the acquired video image data; The preprocessed video image data is used as input to build a fish recognition model, including a backbone network, feature fusion network, and detection head based on the YOLOv8 deep learning model. The output of the recognition model is used to analyze fish behavior, including grayscale difference calculation of consecutive video frames using the two-frame difference method and generating a quantitative index of feeding intensity by matching the fuzzy rule base with Gaussian membership functions. Based on the results of fish school behavior analysis, a dynamic fill light strategy is generated through a fuzzy PID control algorithm, including defuzzification calculation of PID parameter adjustments using the center of gravity method and determination of the optimal wavelength based on the phototaxis characteristics of the fish species. Convert the dynamic fill light strategy into an electrical signal to drive the LED lamp to adjust the lighting parameters; The fill-light strategy is adaptively optimized according to real-time lighting parameters, including online updating of fuzzy rule base weights using recursive least squares method.

2. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 1, characterized in that: The backbone network, feature fusion network and detection head are constructed based on the YOLOv8 deep learning model, including the use of the CSPDarknet structure with RepVGG re-parameterization as the backbone network. The backbone network introduces cross-stage partial connections to perform multi-scale feature extraction, and the extracted feature maps are input into the feature fusion network. The feature fusion network uses the PANet++ network combined with a bidirectional feature pyramid structure to fuse multi-scale features through an adaptive weight distribution mechanism. The multi-scale feature fusion formula is: Among them, ω i represents the learnable weight of the i-th layer feature, F i Represents the feature map of the i-th layer.

3. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 2, characterized in that: The method constructs a backbone network, a feature fusion network, and a detection head based on the YOLOv8 deep learning model, and also includes inputting the fused feature map into the detection head for target detection and classification. The detection head has a built-in classification branch and a regression branch that are decoupled and independent. The classification branch uses binary cross entropy loss to predict fish species labels, and the regression branch uses CIoU Loss to optimize bounding box coordinates. Three-dimensional spatial positioning is performed based on the optimized bounding box coordinates. The optimization formula is: Among them, ρ is the distance between the center point of the predicted box and the real box, c is the minimum diagonal length of the bounding box, v measures the consistency of the aspect ratio, α is the balance coefficient, and loU is the intersection-union ratio, which indicates the difference between the predicted box b and the real box b gt degree of overlap.

4. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 3, characterized in that: The three-dimensional spatial positioning based on the optimized bounding box coordinates includes obtaining the binocular camera intrinsic parameter matrix and distortion coefficient for correcting image distortion and establishing a mapping relationship between pixel coordinates and the camera coordinate system, calculating the two-camera extrinsic parameter matrices, establishing a binocular vision coordinate system conversion relationship, performing feature matching on the left and right camera images, obtaining the pixel coordinates of the same-name points, calculating the three-dimensional coordinates of the fish school based on the pixel coordinates of the same-name points through the principle of triangulation, and finally associating the three-dimensional position information with the fish species label.

5. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 1, characterized in that: The fish behavior analysis using the output results of the recognition model includes obtaining fish species classification labels, calculating the feeding intensity index through behavior recognition, constructing a multidimensional feature vector using the feeding intensity index, and defining a feeding intensity fuzzy set through a Gaussian membership function. Based on the feeding intensity fuzzy set, a preset rule library is triggered to generate preliminary fill light control instructions. The expression for calculating the feeding intensity index is: Among them, N active is the pixel area of ​​feeding behavior, N total is the total area of ​​the monitoring area, and T is the number of statistical frames.

6. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 1, characterized in that: The two-frame difference method is used to perform grayscale difference calculation on continuous video frames, including defining K states for each pixel in the video image, then using a mixture model of K Gaussian distributions to establish a background model for the pixel, performing grayscale difference between the current frame and the previous frame based on the background model, introducing the Otsu algorithm to adaptively obtain the threshold, and then obtaining the foreground area in the current image. The background model formula is: Among them, η i Represents the probability density function of the i-th Gaussian distribution, describing the pixel value I t The probability of belonging to this distribution, μ i,t represents the mean of the i-th Gaussian distribution at time t, represents the central trend of the background pixel value, σ i,t represents the standard deviation of the i-th Gaussian distribution at time t, ω i represents the weight of the i-th Gaussian distribution.

7. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 1, characterized in that: The method generates a dynamic fill light strategy based on the results of fish school behavior analysis through a fuzzy PID control algorithm, including building a fuzzy rule base with light intensity error and its change rate as input, fuzzifying the input parameters through a Gaussian membership function, defuzzifying and calculating the PID parameter adjustment amount using a center of gravity method in combination with a preset fuzzy rule table, solving the optimal band according to the obtained fish species label, and generating a fill light strategy by combining the PID parameter adjustment result and the optimal band. The optimal band solution formula is: l opt =argmax λ∈[400,700] (A(l)h LED (l)), Where λ represents the wavelength of the supplementary light spectrum, ranging from 400 to 700 nm, A(λ) represents the spectral response of the fish species to the wavelength λ, and η LED (λ) represents the luminous efficiency of the LED lamp at wavelength λ.

8. The underwater intelligent light-filling method for deep-sea aquaculture cages according to claim 1, characterized in that: The online updating of the fuzzy rule base weights using the recursive least squares method includes collecting underwater light intensity data through a distributed illuminometer network, performing Kalman filtering on the collected light intensity data, constructing a feature vector of the light intensity data, and defining a fuzzy rule base weight vector. The weight vector is updated using an iterative formula to complete the fuzzy rule base weight update. The iterative formula is expressed as: Among them, φ(k) represents the eigenvector of the light intensity data at time k, θ(k) represents the weight vector of the fuzzy rule base at time k, λ1 represents the forgetting factor, P(k) represents the covariance matrix, which measures the uncertainty of parameter estimation, and y(k) represents the light intensity feedback value at time k.

9. An underwater intelligent lighting system for deep-sea aquaculture cages, implemented according to the method of claim 1, characterized in that: include: An image acquisition module is connected to the global binocular camera, preprocesses the video stream of the global binocular camera, and transmits the preprocessed video stream to the fish recognition module through the armored optical fiber transmission component; The fish species recognition module integrates a class recognition model to output fish species labels, bounding boxes, and three-dimensional coordinates, and transmits the output to the fish behavior recognition module and the lighting strategy control module; The fish behavior recognition module uses a two-frame difference method to perform grayscale difference calculations on continuous video frames to generate a quantitative index of feeding intensity and transmit it to the lighting strategy control module; The lighting strategy control module is used to receive the fish species label and the feeding intensity quantitative index, generate a dynamic fill light strategy through the fuzzy PID control algorithm, and transmit the fill light strategy to the driving module; The driving module has a built-in constant current LED driving circuit to realize the linear mapping relationship between PWM duty cycle and current value, convert the instruction of fill light strategy control into an electrical signal, and transmit the electrical signal to the LED lamp assembly; The LED lamp module adopts a multi-channel programmable LED array and integrates a wavelength-selective light-emitting chip group for flexibly combining and outputting spectra of different bands.

10. The underwater intelligent light supplement system for deep sea aquaculture cages according to claim 9, characterized in that: The constant current LED drive circuit includes a sampling capacitor, a stepless dimming control chip and multiple LEDs. One end of the sampling capacitor is connected to the positive electrode of the circuit input end, and the other end of the sampling capacitor is connected to the CMOD capacitor pin of the stepless dimming control chip for collecting signals in the circuit. The PWM signal output by the stepless dimming control chip controls the transistor connected to the LED through a current limiting resistor, thereby driving the LED.

Images