A method and system for optimizing baiting decisions

Abstract

The application discloses a kind of baiting decision optimization method and system, specifically related to aquatic products cultivation automation baiting technology field, for solving the baiting strategy of existing neglecting fish school phototaxis and dynamic behavior resulting in bait waste and low feeding efficiency problem;By real-time monitoring the optical intensity distribution of fry cluster area and extracting the center position, combined with time series analysis to determine the periodic aggregation rule and phototaxis threshold;Dynamic adjustment baiting device to fish aggregation area, based on the divergence of motion vector field calculation generates initial baiting amount;According to the fractal characteristics of fish school bait trajectory and state duration corrects baiting amount, and periodically triggers baiting decision closed loop.The system realizes the collaborative optimization of bait delivery position, quantity and time by the fusion of optical sensing, kinematic analysis and adaptive algorithm, significantly improves the utilization rate of bait and fry growth uniformity.

Description

Technical Field

[0001] This invention relates to the field of automated feeding technology in aquaculture, and more specifically, to a feeding decision optimization method and system. Background Technology

[0002] In aquaculture, the feeding of fish fry and fingerlings during their breeding stage directly affects survival rates and the efficiency of selective breeding. Existing feeding methods are mostly based on fixed placement locations or timed, quantitative strategies, such as manually or mechanically scattering feed evenly into the rearing pond. These methods assume that the activity area of ​​the fry and the feed distribution space are the same, and they do not optimize for the biological characteristics of the fry, leading to significant deviations between actual feeding results and theoretical designs.

[0003] Traditional feeding strategies do not fully consider the group activity patterns caused by the phototaxis of fish fry, resulting in a misalignment between the spatial distribution of feed in the rearing pond and the natural aggregation areas of the fry. This spatial mismatch forces the fry to expend extra energy to feed, and some feed is wasted due to not being consumed in time, ultimately leading to decreased uniformity of fry growth and fluctuations in survival rate, thus restricting the efficiency of large-scale breeding of superior species. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a feeding decision optimization method and system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A feeding decision optimization method includes the following steps:

[0007] S1. Monitor the optical intensity distribution in the fish fry cluster area in the aquaculture pond in real time, and record the coordinates of the center position of the fish fry cluster area;

[0008] S2. Based on continuous time series data of optical intensity distribution, the periodic aggregation pattern and phototactic response threshold of fish fry cluster areas are determined through time series analysis.

[0009] S3. Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to the preset offset range of the center position coordinates corresponding to the phototactic response threshold.

[0010] S4. At the adjusted horizontal position, calculate the area ratio of the divergence region based on the divergence analysis of the fish movement vector field, generate the initial feeding amount based on the difference between the ratio and the preset activation threshold, and execute the feeding.

[0011] S5. After the initial feeding amount is released, extract the dynamic behavioral characteristics of the fish's feeding trajectory and determine whether the fish are in a state of random dispersion or group feeding. Adjust the initial feeding amount according to the state type and duration.

[0012] S6. Determine the next feeding time node based on the periodic aggregation pattern, and repeat S1 to S5 until the preset feeding cycle is completed.

[0013] In a preferred embodiment, real-time monitoring of the optical intensity distribution in the fish fry cluster area within the rearing pond, and recording the coordinates of the center position of the fish fry cluster area, includes:

[0014] An optical sensor array is installed on the top of the aquaculture pond to divide the water surface of the pond into multiple evenly distributed gridded zones.

[0015] The optical intensity data of each gridded partition is collected synchronously by an optical sensor array, and the average optical intensity of each gridded partition is calculated.

[0016] Based on the average optical intensity gradient change of adjacent gridded partitions, areas where the average optical intensity continuously exceeds a preset cluster threshold are identified as fish fry cluster areas.

[0017] Geometric mean calculation is performed on the geometric centers of all gridded partitions within the fish fry cluster area, and the calculation result is used as the coordinates of the center position of the fish fry cluster area.

[0018] In a preferred embodiment, the monitoring frequency of the optical sensor group is dynamically adjusted according to the area change of the fish fry cluster area. When the area increases, the monitoring frequency is increased proportionally, and when the area decreases, it is restored to the reference frequency.

[0019] In a preferred embodiment, based on continuous time-series data of optical intensity distribution, the periodic aggregation pattern and phototactic response threshold of the fish fry cluster area are determined through time-series analysis, including:

[0020] The continuous time series data of optical intensity distribution in the fish fry cluster area is smoothed by a sliding window, with a preset time window and step size set.

[0021] Perform a Fourier transform on the smoothed data and extract the period corresponding to the main frequency component whose energy proportion exceeds the preset energy proportion threshold in the amplitude spectrum as the candidate period;

[0022] The optical intensity data is periodically segmented based on the candidate period, and the distribution of the difference between the maximum and minimum light intensity in each time period is calculated.

[0023] The phototactic response threshold is determined based on the cumulative probability curve of the difference distribution as the quantile point corresponding to the light intensity difference reaching the preset quantile value.

[0024] Periods in the candidate periods where the corresponding light intensity difference quantiles continuously exceed the phototactic response threshold are marked as effective periodic clustering patterns.

[0025] In a preferred embodiment, the horizontal position of the feeding device is dynamically adjusted to a preset offset range within the center position coordinates corresponding to the phototactic response threshold, based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment. This includes:

[0026] Calculate the absolute value of the difference between the actual light intensity in the aquaculture pond at the current moment and the phototactic response threshold;

[0027] The direction of movement of the feeding device is determined by comparing the absolute value of the difference with the preset difference range.

[0028] The dynamic adjustment step size is generated by multiplying the moving direction and the absolute value of the difference. The feeding device is controlled to move along the moving direction with the dynamic adjustment step size until it enters the preset offset range. The offset distance between the current position of the feeding device and the center position coordinates is detected in real time to see if it meets the preset offset range constraint condition.

[0029] In a preferred embodiment, when the absolute value of the difference exceeds the upper limit of the preset difference range, the feeding device is controlled to move in the positive direction of the center position coordinate corresponding to the phototactic response threshold, and when the absolute value of the difference is lower than the lower limit of the preset difference range, it moves in the negative direction.

[0030] In a preferred embodiment, at the adjusted horizontal position, the area ratio of the divergence region is calculated based on the divergence analysis of the fish movement vector field. An initial feeding amount is generated based on the difference between this ratio and a preset activation threshold, and the feeding is then executed, including:

[0031] A continuous sequence of images of fish movement at an adjusted horizontal position is obtained using a high-speed camera device.

[0032] Optical flow is used to calculate the motion vector field of the fish school for adjacent frames;

[0033] Divergence calculation is performed on the motion vector field, and grid regions with divergence values ​​greater than zero are marked as divergence regions;

[0034] The proportion of the area of ​​the divergent region to the total area covered by the fish movement vector field is statistically analyzed.

[0035] When generating the initial feeding amount based on the difference between the ratio and the preset activation threshold, if the ratio is greater than or equal to the preset activation threshold, the initial feeding amount is the product of the baseline feeding amount and the difference ratio; if the ratio is less than the preset activation threshold, the initial feeding amount is the output value of a linearly decreasing function of the baseline feeding amount and the absolute value of the difference.

[0036] In a preferred embodiment, after the initial feeding amount is released, the dynamic behavioral characteristics of the fish's feeding trajectory are extracted and it is determined whether the fish are in a random, dispersed, or aggregated feeding state. The initial feeding amount is then adjusted based on the state type and duration, including:

[0037] The movement of fish schools towards the bait is continuously captured by underwater camera devices, and the movement trajectory is constructed by extracting the coordinate sequence of the fish school's center of gravity.

[0038] The fractal dimension of the movement trajectory is calculated. If the fractal dimension is lower than the preset fractal threshold, it is determined to be a state of aggregated feeding. If the fractal dimension is higher than the preset fractal threshold, it is determined to be a state of random dispersion.

[0039] The duration of the current state type is recorded. When the duration of the feeding state exceeds the preset duration threshold, the initial feeding amount is increased by the first correction coefficient.

[0040] When the duration of the random dispersion state exceeds the preset duration threshold, the initial feeding amount is reduced by the second correction coefficient.

[0041] If the duration does not exceed the preset duration threshold, the initial feeding amount will remain unchanged.

[0042] In a preferred embodiment, the next feeding time node is determined according to the periodic aggregation pattern, and S1 to S5 are repeated until the preset feeding cycle is completed, including:

[0043] Extract the principal period length from the effective periodic clustering pattern;

[0044] The next feeding interval is calculated based on the sliding window average of the main cycle length and historical feeding cycle data.

[0045] If the current feeding time deviates from the peak time of light intensity corresponding to the periodic aggregation pattern by more than the preset time tolerance, the next feeding time interval will be adjusted according to the deviation ratio.

[0046] Update the next feeding time node to the sum of the current time node and the corrected feeding time interval, and trigger S1 to S5 to execute until the cumulative number of feedings reaches the preset total number of feeding cycles.

[0047] On the other hand, the present invention provides a feeding decision optimization system, comprising the following modules:

[0048] Cluster positioning module: Real-time monitoring of the optical intensity distribution in the fish fry cluster area within the aquaculture pond, and recording the coordinates of the center position of the fish fry cluster area;

[0049] Periodic threshold module: Based on continuous time series data of optical intensity distribution, it determines the periodic aggregation pattern and phototactic response threshold of fish fry cluster areas through time series analysis;

[0050] Dynamic positioning module: Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to a preset offset range of the center position coordinates corresponding to the phototactic response threshold.

[0051] Divergence Feeding Module: At the adjusted horizontal position, the divergence area ratio is calculated based on the divergence analysis of the fish movement vector field. The initial feeding amount is generated based on the difference between the ratio and the preset activation threshold, and the feeding is executed.

[0052] Behavior correction module: After the initial feeding amount is released, the dynamic behavioral characteristics of the fish's feeding trajectory are extracted and the fish are determined to be in a state of random dispersion or group feeding. The initial feeding amount is corrected according to the state type and duration.

[0053] Periodic scheduling module: Determines the next feeding time node based on the periodic aggregation pattern, and repeatedly executes the cluster positioning module, periodic threshold module, dynamic positioning module, dispersion feeding module and behavior correction module until the preset feeding cycle is completed.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. This invention constructs a closed-loop feedback mechanism for precise feeding decisions by integrating the phototactic response patterns of fish schools with dynamic analysis of their movement trajectories. Based on the spatiotemporal distribution characteristics of optical intensity, fish aggregation areas are calibrated in real time. Combined with periodic patterns and vector field divergence analysis, the feeding position and amount are dynamically adjusted, effectively solving the problem of misalignment between bait distribution and fish activity in traditional methods. By quantifying the complexity of fish movement trajectories using fractal dimensions and coupling the duration of states to correct feeding strategies, the matching accuracy between bait placement and fish feeding needs is significantly improved, reducing resource waste caused by environmental fluctuations or changes in group behavior.

[0056] 2. By deeply integrating optical monitoring, kinematic analysis, and periodic scheduling, an adaptive feeding control logic is formed. Time-based optimization is driven by periodic aggregation patterns, and parameters are dynamically adjusted based on real-time behavioral feedback, overcoming the environmental adaptability limitations of static feeding models. This multi-dimensional data coupling and closed-loop iterative mechanism reduces human intervention while ensuring simultaneous improvement in fry feeding efficiency and growth uniformity, providing reliable technical support for large-scale aquaculture. Attached Figure Description

[0057] Figure 1 This is a flowchart of a feeding decision optimization method according to the present invention;

[0058] Figure 2 This is a schematic diagram of the structure of a feeding decision optimization system according to the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1: Figure 1 This invention provides a feeding decision optimization method, which includes the following steps:

[0061] S1. Monitor the optical intensity distribution in the fish fry cluster area in the aquaculture pond in real time, and record the coordinates of the center position of the fish fry cluster area;

[0062] S2. Based on continuous time series data of optical intensity distribution, the periodic aggregation pattern and phototactic response threshold of fish fry cluster areas are determined through time series analysis.

[0063] S3. Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to the preset offset range of the center position coordinates corresponding to the phototactic response threshold.

[0064] S4. At the adjusted horizontal position, calculate the area ratio of the divergence region based on the divergence analysis of the fish movement vector field, generate the initial feeding amount based on the difference between the ratio and the preset activation threshold, and execute the feeding.

[0065] S5. After the initial feeding amount is released, extract the dynamic behavioral characteristics of the fish's feeding trajectory and determine whether the fish are in a state of random dispersion or group feeding. Adjust the initial feeding amount according to the state type and duration.

[0066] S6. Determine the next feeding time node based on the periodic aggregation pattern, and repeat S1 to S5 until the preset feeding cycle is completed.

[0067] S1. Real-time monitoring of the optical intensity distribution in the fish fry cluster area within the aquaculture pond, and recording the coordinates of the center position of the fish fry cluster area. Specific implementation details are as follows:

[0068] An optical sensor array, consisting of 16 optical sensors, is installed on the top of the aquaculture pond. The pond surface is divided into square grids with sides of 1 meter at 2m x 2m intervals. Each optical sensor monitors the optical intensity of four adjacent grids. Data from the sensors is synchronously transmitted to the embedded controller via an RS-485 bus. The optical sensor array collects optical intensity data for each grid at a reference frequency of once per second. The average optical intensity of each grid is calculated as the arithmetic mean of the values ​​measured by the optical sensors at the four vertices of the grid. For example, if the measured values ​​at a grid vertex are 1200 Lux, 1250 Lux, 1180 Lux, and 1230 Lux, then the average optical intensity of that grid is (1200 + 1250 + 1180 + 1230) / 4 = 1215 Lux. In practice, the light intensity value can be adjusted according to the sensor accuracy.

[0069] When identifying fish fry clustering areas based on the average optical intensity gradient changes of adjacent gridded zones, the absolute value of the average optical intensity difference between each grid and its two adjacent grids to the east and south is first calculated. If the difference in both directions exceeds 50 Lux and the average optical intensity of the current grid is greater than the preset clustering threshold of 800 Lux for three consecutive samplings, then the grid is marked as a candidate clustering grid. Adjacent candidate clustering grids are merged into a continuous region, excluding isolated regions with an area less than 0.5 square meters. This threshold is determined based on observation data of fish fry feeding behavior. When the area is less than 0.5 square meters, the feeding efficiency of the fish population is lower than the effective value, and the fish fry clustering region is finally determined. For example, when the intensities of three adjacent grids are detected to be 810 Lux, 830 Lux, and 820 Lux, and the differences with the eastern grid are 60 Lux and 70 Lux, respectively, they are merged into one fish fry clustering region. In actual implementation, the gradient threshold and clustering threshold can be fine-tuned according to the size of the aquaculture pond.

[0070] When calculating the geometric mean of the geometric centers of all gridded zones within the fish fry cluster area, first determine the geometric center coordinates of each grid. For example, the geometric center coordinates of a square grid with a side length of 1 meter are the coordinates of the lower left corner of the grid offset by (0.5 meters, 0.5 meters). Take the arithmetic mean of the horizontal and vertical coordinates of the geometric centers of all grids within the area to obtain the center position coordinates of the fish fry cluster area. For example, if an area contains four grids with geometric center coordinates of (1.5, 2.5), (2.5, 2.5), (1.5, 3.5), and (2.5, 3.5), then the center position coordinates are ((1.5 + 2.5 + 1.5 + 2.5) / 4, (2.5 + 2.5 + 3.5 + 3.5) / 4) = (2.0, 3.0). In actual implementation, the grid size can be adjusted according to the shape of the aquaculture pond.

[0071] When dynamically adjusting the monitoring frequency of the optical sensor array based on changes in the area of ​​the fish fry cluster, a baseline monitoring frequency of once per second is set. As the area of ​​the fish fry cluster increases, the monitoring frequency increases by 1.1 times for every 0.2 square meters increase, but not exceeding 5 times per second to prevent sensor overheating. Conversely, as the area decreases, the monitoring frequency decreases by 0.9 times for every 0.2 square meters decrease, until the baseline frequency is restored. For example, if the current area is 2 square meters, and the next detected area is 2.4 square meters, the monitoring frequency is adjusted to 1.1 × 1 = 1.1 times per second; if the area continues to increase to 2.8 square meters, the frequency is adjusted to 1.1 × 1.1 ≈ 1.21 times per second. The adjusted monitoring frequency acts as the data acquisition command for the optical sensor array, which is then sent to each sensor node via the embedded controller.

[0072] In the anomaly handling mechanism, when an optical sensor fails to return valid data five times consecutively, the gridded area it covers is automatically marked as a failure region, and interpolation compensation calculations for that region are initiated by adjacent sensors. The interpolation calculation method is to take the average optical intensity values ​​of the eight adjacent grids around the failure region as the estimated value for that region. For example, if the intensity of the grids surrounding a failure region is 790 Lux, 810 Lux, 800 Lux, and 805 Lux, then the estimated value is (790 + 810 + 800 + 805) / 4 = 801.25 Lux. In actual implementation, the number of grids for interpolation compensation can be adjusted according to the sensor layout.

[0073] The preset clustering threshold of 800 Lux is based on the following: In fish fry rearing experiments, when the light intensity is below 800 Lux, the phototactic activity of fish is significantly weakened, and the feeding efficiency decreases by more than 30%. This threshold was determined by comparing experimental data on the amount of feed consumed per unit time by fish fry under different light intensities, with experimental conditions of water temperature 25℃±1℃ and dissolved oxygen 6mg / L±0.5mg / L. The gradient change threshold of 50 Lux is based on the following: When the intensity difference between adjacent grids exceeds 50 Lux, the distribution of fish swarms exhibits significant non-uniformity, and the accuracy of clustering area identification can reach over 95% at this point.

[0074] The proportional coefficient for adjusting the monitoring frequency was determined through data acquisition delay experiments simulating different area change rates. When the area change rate exceeds 0.1 square meters per second, increasing the monitoring frequency ensures that the location coordinate update delay is less than 0.5 seconds. The area threshold of 0.5 square meters for excluding isolated areas in the geometric mean calculation is based on the following: In actual observation, the minimum effective feeding area of ​​the natural gathering area of ​​fish fry is 0.5 square meters. Areas below this area are mostly caused by sensor noise or individual fish fry wandering.

[0075] S2. Based on continuous time-series data of optical intensity distribution, the periodic aggregation pattern and phototactic response threshold of fish fry clustering areas are determined through time-series analysis. The specific implementation is as follows:

[0076] When performing sliding window smoothing on continuous time-series data of optical intensity distribution in fish fry aggregation areas, the sliding window length is set to five minutes, with a step size of one minute. In practice, the optical intensity data within each five-minute interval is arranged chronologically, and the sliding window is used at one-minute intervals. The median value of all data points within the window is taken as the smoothed value for the current time point. For example, if the optical intensity data from timestamp T0 to T5 is [800 Lux, 810 Lux, 790 Lux, 820 Lux, 805 Lux], then the smoothed value at time point T2 is the median of the data in that window, 805 Lux. The sliding window length is set based on the average response time of the fish fry's phototactic activity. Experimental observations show that the aggregation behavior of fish groups typically reaches a stable state within three to eight minutes after changes in light intensity; therefore, a five-minute window is chosen to cover the typical activity cycle. The one-minute step size is set based on the minimum temporal resolution for updating the fish group's location, ensuring a balance between data timeliness and processing efficiency. In practice, if the area of ​​the aquaculture pond exceeds 200 square meters, the window length can be adjusted to ten minutes to adapt to a wider range of light changes.

[0077] When performing a Fourier transform on the smoothed data, a Fast Fourier Transform (FFT) algorithm is used to convert the time-domain optical intensity data into a frequency-domain amplitude spectrum. Specifically, the smoothed optical intensity data is input into the algorithm at equally spaced sampling points, with each sampling point corresponding to a one-minute time interval. The number of sampling points is five data points within a five-minute sliding window. After calculating the amplitude spectrum using the FFT, the dominant frequency components whose energy percentage exceeds a preset energy percentage threshold are extracted. The preset energy percentage threshold is set to 15% of the total energy, meaning only frequencies in the amplitude spectrum whose individual frequency components account for more than 15% of the total energy of all frequency components are retained. For example, if the amplitude energy of a certain frequency component is 18% of the total energy, then the period corresponding to that component is retained; if it is 12%, it is excluded. The period corresponding to the dominant frequency component is calculated by the reciprocal of the frequency; for example, a frequency of 0.002 Hz corresponds to a period of 500 seconds. The preset energy percentage threshold was determined based on the significance experiment of the periodic activity of fish. When the energy percentage is less than 15%, the periodic activity cannot be effectively distinguished from random noise. This threshold was verified by comparing experimental data on the feeding efficiency of fish under different light fluctuation frequencies. The experimental conditions were water temperature 25℃±1℃ and dissolved oxygen 6mg / L±0.5mg / L.

[0078] When periodically segmenting optical intensity data based on candidate periods, the time series data is divided into multiple time segments that are integer multiples of the candidate period length. For example, with a candidate period of 500 seconds, the optical intensity data with a total duration of 6000 seconds is divided into twelve time segments, each containing 500 seconds of continuous data. To calculate the distribution of the difference between the maximum and minimum illuminance within each time segment, all time segments are iterated through, and the difference between the maximum and minimum illuminance within each time segment is calculated, forming a difference dataset. For example, in a time segment with a period of 500 seconds, if the maximum illuminance is 850 Lux and the minimum is 780 Lux, then the difference is 70 Lux.

[0079] The cumulative probability curve of the difference distribution is generated by calculating the cumulative probability after arranging the differences in ascending order. The phototropic response threshold is set as the quantile value corresponding to a cumulative probability of 75%. For example, if 75% of the data points in the difference dataset have a difference less than or equal to 65 Lux, then the phototropic response threshold is 65 Lux.

[0080] The cumulative probability curve is generated as follows: After sorting the difference dataset in ascending order, calculate the cumulative probability value corresponding to each difference, using the following formula:

[0081]

[0082] Where P(u) represents the cumulative probability that the difference is less than or equal to u; pm(u) represents the rank of the difference u in the sorted dataset (counting from 1); N represents the total number of difference datasets; N+1 is a correction factor used to avoid computational overflow when the cumulative probability reaches 100%; 100% means converting the probability value to a percentage.

[0083] When a candidate period is marked as a valid periodic clustering pattern if the light intensity difference quantile consistently exceeds the phototactic response threshold, all candidate periods are iterated through, and it is determined whether the light intensity difference in all time periods within each period exceeds the threshold. If the proportion of time periods exceeding the threshold in a certain period is greater than 80%, it is considered a valid periodic clustering pattern. For example, if a candidate period contains ten time periods, and the difference in eight of them exceeds 65 Lux, it is marked as a valid period. The determination of the threshold persistence is based on the stability of the fish's phototactic activity. When the compliance rate of time periods is less than 80%, it is considered an occasional fluctuation rather than a regular clustering. The marking conditions for valid periods are verified through historical feeding data. When the compliance rate is higher than 80%, feeding according to the periodic pattern can increase the survival rate of fish fry by more than 12%. This data is based on statistics from a three-month breeding experiment. In the experiment, the control group used fixed-time feeding, while the experimental group used a periodic marking strategy.

[0084] In the anomaly handling mechanism, if the data missing rate exceeds 30% within a certain time window after smoothing, interpolation compensation for adjacent windows is initiated. The interpolation method is to take the average of the corresponding time points of the previous and next windows. For example, the smoothed value of the missing window T3 is calculated from the smoothed values ​​of T2 and T4 as (805Lux + 815Lux) / 2 = 810Lux. The triggering condition for interpolation compensation is set according to data integrity requirements. When the missing rate is below 30%, it can be ignored through local smoothing; when it exceeds 30%, compensation is required to avoid periodic analysis bias. If the no-dominant-frequency component after Fourier transform meets the energy proportion threshold, the sliding window length is automatically increased to ten minutes and reanalyzed to avoid periodic missed detections due to short-term data noise. The adjusted window length must be limited to a maximum of twenty minutes in the specification to prevent the loss of periodic details due to an excessively long window.

[0085] The preset energy percentage threshold of 15% is based on the following: In the phototactic experiment of fish fry, when the energy percentage of periodically changing light intensity exceeds 15%, the correlation between fish aggregation behavior and the light cycle significantly increases, and the fluctuation range of feeding efficiency narrows to within 10%. The phototactic response threshold quantile of 75% is based on the following: When the light intensity difference corresponding to the 75th percentile of the difference distribution reaches the threshold, the correlation between fish feeding activity and light intensity changes reaches its peak. After exceeding this threshold, the rate of increase in feeding efficiency tends to level off. This threshold was determined by comparing experimental data on the amount of food consumed per unit time by fish fry at different quantiles. The experimental results show that the feeding efficiency at the quantile of 75% is 8% higher than that at the 70% quantile, but not significantly different from that at the 80% quantile. The sliding window length of five minutes is based on the following: The shortest duration of phototactic activity in fish fry is three to eight minutes. The window length covers the typical activity cycle to avoid data truncation. This conclusion is based on the analysis of fish movement trajectories recorded by a high-speed camera.

[0086] In the difference distribution calculation, time periods with differences less than 10 Lux are excluded because experiments show that such small changes do not induce significant fish aggregation behavior. The exclusion rule is verified by comparing fish density changes across different difference intervals; when the difference is less than 10 Lux, the fish density change rate is less than 5%, and is considered an invalid fluctuation. In actual implementation, if the turbidity of the aquaculture pond water is higher than 50 NTU (turbidity unit), the phototactic response threshold quantile is automatically lowered to 70% to compensate for light intensity measurement errors. The turbidity compensation mechanism is set based on the transmittance curve of the optical sensor; when the turbidity exceeds 50 NTU, the sensor measurement error increases to ±10 Lux, requiring a reduction in the threshold to offset the error.

[0087] The effective periodic aggregation pattern marking conditions (80% compliance rate over a given time period) were determined by comparing historical feeding data with fish growth rates. When the compliance rate exceeded 80%, periodic feeding could increase fry survival rates by more than 12%. This data was based on a comparative experiment over a six-month rearing period. The experimental group used a dynamic periodic marking strategy, while the control group used fixed-time feeding. The results showed that the experimental group fry had an average body length increase of 15% and feed waste was reduced by 22%. Time periods with differences less than 10 Lux were excluded in the difference distribution calculation, as experiments showed that such small changes did not induce significant fish aggregation behavior. In actual implementation, if the turbidity of the rearing pond water exceeded a preset threshold, the phototactic response threshold quantile was automatically lowered to 70% to compensate for light intensity measurement errors.

[0088] S3. Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to within a preset offset range of the center position coordinates corresponding to the phototactic response threshold. Specifically, this is implemented as follows:

[0089] To calculate the absolute value of the difference between the actual light intensity and the phototactic response threshold in the aquaculture pond at the current moment, data is collected in real time by four light intensity sensors evenly distributed along the edge of the pond. The arithmetic mean of the four sensor measurements is taken as the actual light intensity. For example, if the four sensors measure 820 Lux, 815 Lux, 830 Lux, and 810 Lux at time stamp T0, the actual light intensity is (820 + 815 + 830 + 810) / 4 = 818.75 Lux. The phototactic response threshold is determined to be 65 Lux in step S1, and the absolute value of the difference is calculated as |818.75 - 65| = 753.75 Lux. The arrangement density of the light intensity sensors is set according to the area of ​​the aquaculture pond, with at least one sensor installed for every 50 square meters. The sensor model is TSL2561, with a measurement range of 0-2000 Lux and an accuracy of ±5%. In actual implementation, if the aquaculture pond has an irregular shape, additional sensors should be installed at the corners to avoid blind spots.

[0090] When determining the movement direction of the feeding device based on the comparison between the absolute value of the light difference and a preset light difference range, the upper limit of the preset light difference range is 800 Lux, and the lower limit is 600 Lux. This range was determined through experiments on the phototactic sensitivity of fish: when the absolute value of the light difference exceeds 800 Lux, the fish disperse to the edge area due to excessive light, and the feeding efficiency decreases by more than 25%; when the absolute value of the light difference is below 600 Lux, insufficient light causes the fish to stop feeding. For example, if the absolute value of the light difference is 850 Lux (above 800 Lux), the feeding device is controlled to move in the positive direction of the center position coordinate corresponding to the phototactic response threshold; if it is 550 Lux (below 600 Lux), it moves in the negative direction. The positive and negative definitions of the movement direction are based on the coordinate system of the aquaculture pond, with the positive direction being the direction of increasing light intensity and the negative direction being the opposite. The direction determination logic is implemented through an embedded controller (model STM32F407). After receiving sensor data, the controller performs a comparison calculation and outputs a direction command to the drive motor of the feeding device.

[0091] When generating a dynamically adjusted step size based on the product of the direction of movement and the absolute value of the difference, the base step size is set to 0.1 m / Lux. The formula for calculating the dynamically adjusted step size is: Step size = Direction coefficient × Absolute value of difference × Base step size. The direction coefficient is defined as +1 for the positive direction and -1 for the negative direction. For example, when the absolute value of the difference is 850 Lux and the direction is positive, the step size is +1 × 850 × 0.1 = 85 meters; if it is 550 Lux and the direction is negative, the step size is -1 × 550 × 0.1 = -55 meters. The base step size is determined through a calibration experiment on the phototactic movement speed of fish. In a 10 m × 10 m aquaculture pond, the feeding device can cover the entire pond area within 100 seconds with a step size of 0.1 m / Lux. In practice, the step length calculation result is sent to the stepper motor (model 42BYGHW208) of the feeding device via RS-485 communication protocol. The drive motor controls the ball screw transmission system to move the corresponding distance according to the step length value.

[0092] The feeding device is controlled to move along the direction of movement by dynamically adjusting the step size until it enters the preset offset range, which is set to ±3 meters from the center position coordinates. When the offset distance between the current position of the feeding device and the center position coordinates is detected in real time, the current position coordinates are obtained through the GPS positioning module (model UBLOX NEO-M8N, positioning accuracy ±0.1 meters) at the bottom of the feeding device, and the Euclidean distance between it and the center position coordinates is calculated.

[0093] For example, if the center position coordinates are (10.0, 20.0) and the current position is (12.0, 22.0), then the offset distance is:

[0094]

[0095] If the offset distance exceeds 3 meters, repeat the adjustment steps until the requirements are met. The preset offset range is determined based on bait diffusion experiments: when the baiting point offset exceeds 3 meters, the bait concentration in the edge area drops to the effective feeding threshold (0.5 g / m³). 3 The threshold below was determined by experiments on the amount of feed consumed by fish fry per unit time.

[0096] In the anomaly handling mechanism, if the light intensity sensor returns an abnormal value three times consecutively (e.g., exceeding the range of 2000 Lux or falling below 0 Lux), the backup sensor group is activated to collect data. The backup sensor group is deployed 5 meters apart from the main sensor group, and the data replacement logic is to take the arithmetic mean of the measurements from the backup group. For example, if the main sensor group returns an abnormal value of 2500 Lux consecutively at position P1, the measurement value from the backup group P1' is used to replace it. If the feeding device is obstructed during movement, causing a positional deviation of more than 5 meters, an emergency braking is triggered and an alarm signal is sent to the monitoring terminal. At the same time, the current position coordinates are recorded for manual intervention. For example, when the feeding device stops due to an obstacle at the bottom of the pool, the controller detects that the position coordinates have not changed three times consecutively, determines that it is obstructed, and initiates the braking procedure. The braking response time is less than 0.5 seconds.

[0097] The upper and lower limits of the preset difference range were determined by comparing the feeding efficiency of fish under different light intensity differences. The experimental conditions were a water temperature of 25℃±1℃ and dissolved oxygen of 6mg / L±0.5mg / L. Experimental data showed that when the difference was between 600Lux and 800Lux, the feeding efficiency of the fish was stable at over 85%; beyond this range, the efficiency decreased at a rate exceeding 5% / 10Lux. The step size of 0.1m / Lux was set based on the fact that in a 10m×10m rearing pond, the feeding device could complete the movement of the entire pond within 100 seconds at a maximum difference of 1000Lux, matching the average migration speed of the fish in response to phototaxis (0.1m / s). The offset range of ±3m was set based on a feed diffusion experiment: when the feeding point was offset by 3m, the concentration gradient of the feed within a 5m radius was 0.2g / m². 3 / meter, which can cover areas where fish gather.

[0098] The mechanism of dynamically adjusting the directional coefficient of the step size and binding it to the direction of movement was verified through the phototaxis of fish. During forward movement, the area of ​​increasing light intensity aligned with the direction of fish aggregation. For example, in a forward movement test, after the feeding device moved to an area of ​​higher light intensity, the fish density increased from 10 fish / m² within 5 minutes. 3 Increased to 12 tails / m 3The GPS positioning module for position offset detection updates data at a frequency of 1Hz, ensuring a real-time position error of less than 0.2 meters. In actual implementation, if the aquaculture pond is irregularly shaped, the preset offset range can be adjusted to a polygonal area. The vertex coordinates are generated manually or automatically through mapping. The mapping method involves driving the feeding device along the pond wall and recording the coordinates of the trajectory points.

[0099] Boundary condition coverage includes: when the absolute value of the difference exceeds 2000 Lux, the feeding device position is forcibly reset to the center of the aquaculture pond; when the calculated step length exceeds the maximum stroke of the ball screw (20 meters), the movement command is executed in segments. For example, if the step length is 85 meters and the screw stroke is 20 meters, the movement is divided into four steps (20 meters each time, and 5 meters for the fifth step).

[0100] S4. At the adjusted horizontal position, calculate the area ratio of the divergence region based on the divergence analysis of the fish movement vector field. Generate the initial feeding amount based on the difference between the ratio and the preset activation threshold, and execute the feeding. The specific implementation is as follows:

[0101] When acquiring a continuous image sequence of fish movement at the adjusted horizontal position using a high-speed camera, an industrial camera (model DAHENG MER-131-75U3M) with a frame rate of 30 frames per second was used for image acquisition. The camera was mounted 2 meters directly above the feeding device and fixed with an adjustable bracket, covering a circular area with a radius of 5 meters centered on the adjusted horizontal position. The continuous image sequence was acquired for 10 seconds, totaling 300 frames, with an image resolution of 1280×1024 pixels and stored in RAW12 format to preserve lighting details. For example, after the feeding device moved to coordinates (15.0, 25.0), the camera continuously captured images of the fish movement in that area, with each frame timestamped 33 milliseconds apart to ensure temporal continuity. In practice, if the turbidity of the aquaculture pond water is higher than 50 NTU (turbidity unit), it will automatically switch to near-infrared light source (wavelength 850nm, power 10W) ​​to assist imaging. The light source is installed on both sides of the camera at a tilt angle of 30° to avoid disturbing the fish with direct light.

[0102] When calculating the fish movement vector field using optical flow on adjacent frames, the Lucas-Kanade optical flow algorithm is employed. The sampling window size is set to 16×16 pixels, the maximum number of iterations is 20, the displacement tolerance is 0.01 pixels, and the maximum displacement detection range is ±32 pixels. Specifically, two consecutive frames are input into the algorithm to calculate the fish movement vector within each 16×16 pixel region. The vector direction represents the direction of fish movement, and the magnitude represents the speed. For example, if a motion vector of (3.2, -1.5) pixels / frame is detected in a certain region, it indicates that the fish are moving to the upper right, with a horizontal displacement of 3.2 pixels and a vertical displacement of -1.5 pixels. The motion vector field covers the entire image area and is then converted into a physical space vector field with a conversion coefficient of 0.1 meters / pixel (based on camera calibration parameters; the calibration method involves taking images on a calibration board of known dimensions and calculating the ratio of pixels to the actual length).

[0103] When calculating the divergence of a moving vector field, the field is divided into a 1m × 1m grid region, with each grid containing 10 × 10 vector points. The divergence value for each grid is calculated as the sum of the partial derivatives of all vector points within that grid in the X and Y directions, using the following formula:

[0104]

[0105] in, V represents the horizontal velocity component of the vector point. x The partial derivative with respect to the horizontal coordinate x, i.e., the rate of change of velocity in the horizontal direction; V represents the velocity component in the vertical direction of the vector point. y The partial derivative with respect to the vertical coordinate y, i.e., the rate of change of velocity in the vertical direction; V x V represents the velocity component of the vector point in the horizontal direction (x-axis). y x represents the velocity component of the vector point in the vertical direction (y-axis); x represents the spatial coordinate in the horizontal direction; y represents the spatial coordinate in the vertical direction. The symbol for partial differential is used to represent the differentiation operation with respect to a single variable.

[0106] If the divergence value is greater than zero, the grid is marked as a divergent region, indicating that the fish are spreading out; if it is less than or equal to zero, it is marked as a convergent region. For example, if the sum of the partial derivatives of vector points within a grid is +0.5, it is determined to be a divergent region; if it is -0.3, it is determined to be a convergent region. Divergence calculation is implemented using an embedded processor (NVIDIA Jetson TX2), with a processing latency of less than 50 milliseconds, ensuring real-time performance. In actual implementation, if the missing vector point rate at the grid edge exceeds 20%, interpolation compensation is initiated, using the arithmetic mean of the divergence values ​​of four adjacent grids.

[0107] When calculating the proportion of the divergent region area to the total area covered by the fish movement vector field, the ratio of the sum of the areas of all grids marked as divergent regions to the total grid area is used. For example, if the fish movement vector field covers 100 1m × 1m grids, and 30 of them are divergent regions, the proportion is 30%. In actual implementation, if the water flow velocity exceeds 0.2 m / s, the edge 5% of the grid area is automatically removed to avoid water flow interference. The removal rule is to remove all grids within 2 meters of the image edge. The preset activation threshold is set to 25%, determined based on experiments on fish feeding efficiency: when the proportion of divergent regions is higher than 25%, the fish activity is positively correlated with feeding demand. This threshold is verified by comparing experimental data on the amount of feed consumed per unit time by fish fry under different proportions. The experimental conditions were: water temperature 25℃±1℃, dissolved oxygen 6mg / L±0.5mg / L, and light intensity 800Lux±50Lux.

[0108] When generating the initial feeding amount based on the difference between the ratio and the preset activation threshold, if the ratio is greater than or equal to 25%, the initial feeding amount is the product of the baseline feeding amount (e.g., 10 kg) and the difference ratio ((ratio - 25%) / 25%). For example, if the ratio is 30%, the difference ratio is (30 - 25) / 25 = 0.2, and the initial feeding amount is 10 × (1 + 0.2) = 12 kg. If the ratio is less than 25%, the initial feeding amount is the output value of a linearly decreasing function of the baseline feeding amount and the absolute value of the difference (25% - ratio), with the formula: initial feeding amount = baseline feeding amount × (1 - 0.8 × absolute value of difference / 25%). For example, if the ratio is 20% and the absolute value of the difference is 5%, the initial feeding amount is 10 × (1 - 0.8 × 5 / 25) = 10 × 0.84 = 8.4 kg. The linear decrease coefficient of 0.8 was determined experimentally. When the activity of the fish population decreases, overfeeding will lead to a 12% increase in the waste rate. This coefficient was determined based on comparative experimental data within a three-month breeding cycle. The experimental group adopted a dynamic feeding strategy, while the control group adopted a fixed feeding amount. The results showed that the feed waste rate of the experimental group was reduced by 18%.

[0109] In the anomaly handling mechanism, if the high-speed camera fails to acquire five consecutive frames of images (e.g., due to data transmission interruption), a backup camera (of the same model, installed 1 meter apart) is activated to continue acquisition, and the image alignment parameters are recalibrated. The calibration method involves extracting feature points in the overlapping areas of the adjacent camera's field of view for matching. If the vector point missing rate of a certain grid in the divergence calculation exceeds 40%, interpolation compensation is initiated. The interpolation method is to take the arithmetic mean of the divergence values ​​of the four adjacent grids. For example, if the divergence values ​​of the four grids surrounding a missing grid are +0.4, +0.3, +0.5, and +0.2, the interpolation result is (0.4 + 0.3 + 0.5 + 0.2) / 4 = 0.35. The interpolated grid is then used for area statistics.

[0110] The preset activation threshold of 25% is based on the following: In fish feeding experiments, when the proportion of the divergent area reaches 25%, the feeding efficiency of the fish reaches a peak of 80%. Efficiency decreases by 5% / 5% when the threshold is exceeded or fallen below this value. The optical flow parameters (16×16 pixels window, 20 iterations) were determined by comparing the motion vector errors under different settings. Experiments showed that the displacement error under this parameter combination was less than 0.5 pixels, achieving an optimal balance between computation time and accuracy. The linear decrease coefficient of 0.8 is based on the following: When activity decreases by 5%, reducing the feeding amount by 4% (10kg → 9.6kg) can reduce the waste rate to below 8%. This data is based on a comparative experiment over a three-month breeding cycle. The feeding amount in the experimental group was dynamically adjusted, while the control group received a fixed feeding amount. The results showed that the uniformity of fish fry growth in the experimental group increased by 15%.

[0111] Boundary condition coverage includes: if the proportion exceeds 50%, the initial feeding amount is forcibly limited to twice the baseline value to prevent overfeeding; if the proportion is below 10%, a manual inspection command is triggered. For example, when the proportion is 55%, the maximum feeding amount is 20kg; when the proportion is 8%, the system sends an alarm to the monitoring terminal and suspends automatic feeding. In actual implementation, if the pond area exceeds 500 square meters, the grid size is scaled proportionally to 2m × 2m to reduce the computational load. The divergence calculation error rate after scaling is less than 3%, and the error compensation method is to multiply the divergence value by the square of the scaling factor (2). 2 =4) To maintain consistency in physical meaning.

[0112] Hardware Dependencies and Data Verification: The frame rate and resolution of the industrial camera were calibrated using the movement speed of a school of fish. Experiments showed that the maximum swimming speed of the fish was 0.5 m / s, and a sampling rate of 30 frames / second ensured a displacement capture accuracy of over 90%. The embedded processor was selected based on real-time requirements; the Jetson TX2's computing power could meet the optical flow calculations required to process 30 frames of 1280×1024 images per second.

[0113] S5. After the initial feeding amount is released, the dynamic behavioral characteristics of the fish's feeding trajectory are extracted, and it is determined whether the fish are in a state of random dispersion or group feeding. The initial feeding amount is adjusted according to the state type and duration. The specific implementation is as follows:

[0114] When continuously capturing images of fish movement towards the bait after feeding using an underwater camera device, a HIKVISIONDS-2CD3T86FWDV2-I5 underwater camera is used, mounted 1 meter below the feeding device. It captures images at a rate of 25 frames per second, covering a spherical area with a radius of 3 meters centered on the feeding point. The camera is encased in a waterproof, sealed housing, withstanding pressure up to 10 meters, and is suitable for both freshwater and seawater environments. The image resolution is 1920×1080 pixels, stored in H.265 format to compress data size, and transmitted using the RTSP (Real-Time Streaming Protocol) with a latency of less than 200 milliseconds. For example, after feeding, the fish disperse towards the bait under the influence of the initial feeding amount. The camera continuously captures video of the fish movement, with each frame timestamped and aligned with the feeding time, achieving a time synchronization accuracy of ±50 milliseconds. In actual implementation, if the turbidity of the water exceeds 40 NTU (turbidity unit), the supplementary light (520nm green light, 15W power) will be automatically activated. Green light has better penetration than white light and causes less disturbance to fish. The supplementary light is installed at a downward tilt of 45° to avoid direct light reflection interfering with imaging.

[0115] When extracting the centroid coordinate sequence of the fish school to construct its motion trajectory, each frame of the image is first binarized with a threshold of grayscale value 128, marking the fish fry pixels as foreground. Connectivity analysis is then performed using the `findContours` function from the OpenCV library to identify the minimum bounding rectangle of the fish fry cluster region, and the geometric center of each connected component is calculated as the centroid coordinates of the fish school. For example, if three fish school regions are detected in a frame with centroid coordinates of (2.5, 3.0), (5.1, 4.2), and (7.8, 1.5), the coordinate sequence is constructed in chronological order as [(2.5, 3.0), (5.1, 4.2), (7.8, 1.5), ...]. In practice, if no fish school region is detected in a single frame, the current centroid coordinates are estimated using linear interpolation, which takes the arithmetic mean of the coordinates of the preceding and following frames. For example, if the coordinates are missing in frame t, the estimated coordinates are [(coordinates of frame t-1) + (coordinates of frame t+1)] / 2.

[0116] When calculating the fractal dimension of the motion trajectory, the box counting method is used to divide the area covered by the motion trajectory into square grids with different side lengths. The side length is halved from 1 meter to 0.125 meters. The relationship between the minimum number of grids required to cover the trajectory and the side length is statistically analyzed.

[0117] The formula for calculating fractal dimension is:

[0118]

[0119] Where N(ε) represents the minimum number of grids required to cover the motion trajectory when the grid side length is ε; ε represents the side length of the grid, and ε>0; 1 / ε represents the reciprocal of the grid side length; ε→0 represents the limiting condition where the grid side length approaches zero.

[0120] For example, when the grid side length decreases from 1 meter to 0.5 meters, the required number of grids increases from 20 to 45, resulting in a fractal dimension of approximately 1.73. A preset fractal threshold of 1.5 was set, determined based on fish behavior experiments: when the fractal dimension is below 1.5, fish exhibit linear aggregation (feeding state), while above 1.5, they exhibit random diffusion (dispersion state). The fractal threshold was calibrated through a three-month aquaculture experiment, comparing the feeding efficiency of fish at different fractal dimensions. The feeding efficiency reached a peak of 85% when the dimension was 1.5.

[0121] When calculating the duration of the current state type, the timer starts counting from the moment the state is determined and continues until the state changes or the timer stops. The preset duration threshold is set to 5 minutes, determined based on experiments on fish feeding response delays: if the fish remain gathered for more than 5 minutes, it indicates stable feeding demand, and the amount of feed should be increased; if they remain dispersed for more than 5 minutes, the amount of feed should be reduced to avoid waste. For example, if the fish enter a gathered state 3 minutes after feeding and remain so for 6 minutes, the duration is 6 minutes, exceeding the threshold and triggering an upward adjustment; if it only lasts 4 minutes, no adjustment is triggered. The timer uses the system clock with a time error of less than 1 second and automatically resets when the state changes.

[0122] When the duration of concentrated feeding exceeds 5 minutes, the initial feed amount is increased by a first correction factor, set at 1.2, meaning the corrected feed amount equals the initial feed amount multiplied by 1.2. For example, if the initial feed amount is 10 kg, the corrected amount is 12 kg. When the duration of random dispersion exceeds 5 minutes, the initial feed amount is decreased by a second correction factor, set at 0.7, meaning the corrected feed amount equals the initial feed amount multiplied by 0.7. For example, if the initial feed amount is 10 kg, the corrected amount is 7 kg. The factor settings are based on historical feeding efficiency data: increasing by 1.2 times increases feeding coverage by 15%, while decreasing by 0.7 times reduces waste by 18%. The correction factors were validated through a six-month comparative experiment. The experimental group used a dynamic correction strategy, while the control group had a fixed feed amount. The results showed that the experimental group had a 22% higher feed utilization rate and an 18% higher average weight gain rate for fry.

[0123] If the duration does not exceed 5 minutes, the initial feeding amount remains unchanged. For example, if the clustered state lasts for 4 minutes or the dispersed state lasts for 3 minutes, the feeding amount will not be adjusted. In actual implementation, after each feeding cycle, the system records the duration of the state and the correction result, which is used to optimize subsequent coefficient settings. The optimization method is the sliding window averaging method, and the window size is the data from the most recent 10 feeding cycles.

[0124] In the anomaly handling mechanism, if the underwater camera fails to return valid images for 10 consecutive frames, a backup camera (of the same model, installed at a distance of 0.5 meters) is activated, and the overlapping area of ​​the field of view is recalibrated. The calibration method involves extracting SIFT feature points in the overlapping area for matching and calculating the coordinate transformation matrix. If the grid coverage is abnormal in the fractal dimension calculation (e.g., the trajectory exceeds the boundary of the aquaculture pond), the grid range is automatically expanded to the pond wall, and invalid areas are marked. After excluding invalid areas, the calculation is recalculated. For example, when the trajectory extends to the pond wall, the grid beyond the wall is not included in the statistics; only the fractal dimension of the valid area is calculated. If the fractal dimension calculation result exceeds a reasonable range (e.g., less than 1.0 or greater than 2.5), it is determined to be a sensor malfunction, triggering a manual inspection command. At the same time, the feeding amount correction logic is frozen, maintaining the current feeding amount until the malfunction is resolved.

[0125] The basis for setting the fractal threshold to 1.5 is as follows: In flowing water aquaculture, for every 0.1 m / s increase in water flow velocity, the fractal threshold decreases by 0.1 to compensate for the impact of water flow on the trajectory. For example, when the water flow velocity is 0.3 m / s, the fractal threshold is adjusted to 1.5 - 0.3 = 1.2. The dynamic adjustment rule for the 5-minute duration threshold is: when the water temperature is below 20℃, the threshold is extended to 8 minutes; when it is above 30℃, it is shortened to 3 minutes to adapt to changes in the fish's metabolic rate. Experimental data shows that the feeding waste rate can be further reduced by 5% under the temperature adaptive strategy.

[0126] S6. Determine the next feeding time node based on the periodic aggregation pattern, and repeat S1 to S5 until the preset feeding cycle is completed. The specific implementation is as follows:

[0127] When extracting the principal period length from effective periodic clustering patterns, all marked effective periodic clustering patterns are iterated through, and the period with the highest frequency is selected as the principal period length. For example, if an effective periodic clustering pattern contains three period values ​​of 500 seconds, 520 seconds, and 480 seconds, and the 500-second period occurs 60% of the time, then the principal period length is determined to be 500 seconds. If multiple period frequencies are similar (the difference between the highest and second-highest frequencies is less than 10%), then the arithmetic mean of these periods is taken as the principal period length. For example, if the periods 500 seconds and 520 seconds occur 45% and 40% of the time, respectively, then the principal period length is (500 + 520) / 2 = 510 seconds. The selection of the principal period length is based on the stability of the fish's phototactic activity. High-frequency periods indicate a stronger regularity in the fish's aggregation behavior. This rule has been verified through historical feeding data analysis. When the principal period is selected correctly, the fluctuation range of the fish's feeding efficiency is reduced to within 5%.

[0128] When calculating the next feeding interval based on the sliding window average of the main cycle length and historical feeding cycle data, the sliding window size is set to the most recent 5 feeding cycles. The data within the window is arranged in chronological order, and the arithmetic mean is taken as the baseline interval. For example, if the historical cycle data is [500 seconds, 510 seconds, 490 seconds, 505 seconds, 495 seconds], then the sliding window average is (500+510+490+505+495) / 5 = 500 seconds. If the difference between the main cycle length and the sliding window average exceeds 10%, the baseline interval is adjusted proportionally according to the difference. The adjustment formula is: Next feeding interval = Sliding window average × (1 + (Main cycle length - Sliding window average) / Sliding window average). For example, if the main cycle length is 550 seconds, the sliding window average is 500 seconds, and the difference is 10%, then the interval = 500 × (1 + 0.1) = 550 seconds. The sliding window size was set based on the short-term stability analysis of periodic fluctuations. Experiments showed that 5 historical data points could cover 80% of the periodic fluctuation range. When the window was expanded to 10 points, the accuracy improved by less than 5% but the computational load increased by 50%.

[0129] If the current feeding time deviates from the peak light intensity time corresponding to the periodic aggregation pattern by more than the preset time tolerance, the next feeding time interval will be adjusted proportionally to the deviation. The preset time tolerance is set at 15% of the main cycle length; for example, a main cycle of 500 seconds corresponds to a tolerance of 75 seconds. The deviation ratio is calculated as follows: Deviation Ratio = Actual Deviation / Time Tolerance. For example, if the current feeding time lags behind the peak light intensity by 100 seconds, the deviation ratio = 100 / 75 ≈ 1.33. If it exceeds the tolerance, the time interval will be corrected as follows: Original Time Interval × (1 - 0.5 × Deviation Ratio), with a coefficient of 0.5 determined experimentally. For example, if the original time interval is 500 seconds, the corrected time interval would be 500 × (1 - 0.5 × 1.33) = 500 × 0.335 ≈ 167 seconds. The correction coefficient of 0.5 is determined experimentally based on the linear relationship between fish feeding efficiency and time deviation. When the deviation exceeds the tolerance, a 0.5% time interval adjustment is required for every 1% deviation. The coefficient was verified through a three-month comparative experiment. The experimental group adopted a dynamic correction strategy, while the control group had a fixed time interval. The results showed that the feed utilization rate of the experimental group increased by 25%, and the standard deviation of the fry growth rate decreased by 12%.

[0130] When updating the next feeding time to the sum of the current time and the corrected feeding time interval, the current time is obtained from the system clock, which uses a GPS synchronized clock module (model UBLOX NEO-M8T) with a time synchronization error of less than 1 millisecond. For example, if the current time is 14:00:00 and the corrected time interval is 167 seconds, then the next feeding time is 14:00:00 + 167 seconds = 14:02:47. After the time is updated, steps S1 to S5 are automatically triggered, including optical intensity monitoring, position adjustment, and feeding amount correction. During repeated execution, the cumulative feeding count increases until the preset total feeding cycle is reached. For example, if the preset total is 10 times, the process terminates after 10 repetitions. The cumulative feeding count is counted using an embedded counter, with an overflow value set to 10,000 to ensure long-term trouble-free operation.

[0131] In the anomaly handling mechanism, if there are fewer than 5 historical feeding cycle data points, the sliding window average calculation is skipped, and the main cycle length is directly used as the next feeding interval. For example, if only 3 historical data points exist, the interval is directly taken as the main cycle length of 500 seconds. If the peak light intensity time cannot be obtained due to sensor failure, the time deviation correction logic is disabled, and the uncorrected interval is used directly. For example, in the event of sensor failure, the next time node = current time + 500 seconds. If the aquaculture requirements are still not met after the cumulative feeding count reaches the preset total (e.g., the fry density does not reach the target value), a manual intervention command is triggered, pausing automatic feeding and sending an alarm to the monitoring terminal. The manual intervention command is transmitted to the central control system via the Modbus protocol, with a response delay of less than 1 second.

[0132] The preset time tolerance of 15% is based on the following: Fish feeding efficiency experiments show that when the feeding time deviates from the peak light intensity by more than 15% of the cycle length, feeding efficiency decreases by more than 20%. This threshold was determined by comparing the amount of feed residue under different deviation ratios; when the deviation is 15%, the residue increases from 10% to 30%. The sliding window size of 5 times is based on the following: Statistical analysis of periodic fluctuations shows that 5 historical data points can cover 80% of the periodic fluctuation range. Expanding the window to 10 times improves accuracy by less than 5% but increases computational load by 50%. The deviation correction coefficient of 0.5 was calibrated through a six-month aquaculture trial. The experimental group used a dynamic correction strategy, while the control group used a fixed time interval. The results showed that the feeding hit rate (time deviation <15%) in the experimental group increased from 68% to 92%.

[0133] Hardware Dependencies and Data Validation: The GPS synchronization clock module has a time synchronization accuracy of ±1 millisecond, ensuring the accuracy of time point calculations. Historical periodic data is stored in an embedded database (SQLite), with read / write latency of less than 10 milliseconds, supporting 1000 concurrent accesses per second. For experimental data validation, all parameter settings were verified through a six-month aquaculture trial. Data such as feeding time, peak light deviation, feeding efficiency, and fry density were recorded during the trial to ensure the reproducibility of the technical effects. For example, in 100 feeding cycles, the dynamic time correction strategy reduced the standard deviation of the time deviation from 35 seconds to 12 seconds, and the peak feed utilization rate reached 85%.

[0134] Boundary condition coverage includes: if the main cycle length exceeds 24 hours, it is determined to be an invalid cycle and a manual check is triggered; if the corrected time interval is less than 10 seconds, it is forcibly set to 10 seconds to avoid high-frequency feeding. For example, if the main cycle length is incorrectly calibrated to 86400 seconds (24 hours), the system sends an alarm and freezes the automatic process; if the corrected interval is 5 seconds, it is executed as 10 seconds. In actual implementation, if the aquaculture pond is in a multi-zone linkage mode, the feeding time node is calculated independently for each zone, and the time deviation between zones is controlled within 10 seconds, and the clock is synchronized via the NTP protocol.

[0135] This method constructs a closed-loop feedback mechanism for dynamic feeding decisions by coupling the spatiotemporal distribution characteristics of optical intensity with the vector field of fish movement in a multidimensional manner. Unlike traditional techniques that isolate light intensity or fish density, this approach combines the periodicity of fish fry's phototactic response with the complex fractal characteristics of their real-time movement trajectories. It quantifies the fish's diffusion state through divergence analysis and dynamically calibrates the feeding activity threshold based on the fractal dimension, thereby establishing a nonlinear mapping relationship between feeding amount and fish behavior patterns. In particular, the dynamic correction mechanism based on periodic clustering patterns and real-time illumination deviations enables adaptive optimization of feeding timing.

[0136] Example 2: Figure 2 A schematic diagram of a feeding decision optimization system according to the present invention is provided. The feeding decision optimization system includes the following modules:

[0137] Cluster positioning module: Real-time monitoring of the optical intensity distribution in the fish fry cluster area within the aquaculture pond, and recording the coordinates of the center position of the fish fry cluster area;

[0138] Periodic threshold module: Based on continuous time series data of optical intensity distribution, it determines the periodic aggregation pattern and phototactic response threshold of fish fry cluster areas through time series analysis;

[0139] Dynamic positioning module: Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to a preset offset range of the center position coordinates corresponding to the phototactic response threshold.

[0140] Divergence Feeding Module: At the adjusted horizontal position, the divergence area ratio is calculated based on the divergence analysis of the fish movement vector field. The initial feeding amount is generated based on the difference between the ratio and the preset activation threshold, and the feeding is executed.

[0141] Behavior correction module: After the initial feeding amount is released, the dynamic behavioral characteristics of the fish's feeding trajectory are extracted and the fish are determined to be in a state of random dispersion or group feeding. The initial feeding amount is corrected according to the state type and duration.

[0142] Periodic scheduling module: Determines the next feeding time node based on the periodic aggregation pattern, and repeatedly executes the cluster positioning module, periodic threshold module, dynamic positioning module, dispersion feeding module and behavior correction module until the preset feeding cycle is completed.

[0143] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0144] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0145] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0146] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0147] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0148] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0149] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feeding decision optimization method, characterized in that, Includes the following steps: S1. Monitor the optical intensity distribution in the fish fry cluster area in the aquaculture pond in real time, and record the coordinates of the center position of the fish fry cluster area; S2. Based on continuous time-series data of optical intensity distribution, determine the periodic aggregation pattern and phototactic response threshold of fish fry clustering areas through time-series analysis, including: The continuous time series data of optical intensity distribution in the fish fry cluster area is smoothed by a sliding window, with a preset time window and step size set. Perform a Fourier transform on the smoothed data and extract the period corresponding to the main frequency component whose energy proportion exceeds the preset energy proportion threshold in the amplitude spectrum as the candidate period; The optical intensity data is periodically segmented based on the candidate period, and the distribution of the difference between the maximum and minimum light intensity in each time period is calculated. The phototactic response threshold is determined based on the cumulative probability curve of the difference distribution as the quantile point corresponding to the light intensity difference reaching the preset quantile value. Periods in the candidate periods where the corresponding light intensity difference quantiles continuously exceed the phototactic response threshold are marked as effective periodic clustering patterns; S3. Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to the preset offset range of the center position coordinates corresponding to the phototactic response threshold. S4. At the adjusted horizontal position, calculate the area ratio of the divergence region based on the divergence analysis of the fish movement vector field, generate the initial feeding amount based on the difference between the ratio and the preset activation threshold, and execute the feeding. S5. After the initial feeding amount is released, extract the dynamic behavioral characteristics of the fish's feeding trajectory and determine whether the fish are in a state of random dispersion or group feeding. Adjust the initial feeding amount according to the state type and duration. S6. Determine the next feeding time node based on the periodic aggregation pattern, and repeat S1 to S5 until the preset feeding cycle is completed.

2. The feeding decision optimization method according to claim 1, characterized in that, Real-time monitoring of the optical intensity distribution in the fish fry cluster area within the aquaculture pond, and recording the coordinates of the center location of the fish fry cluster area, including: An optical sensor array is installed on the top of the aquaculture pond to divide the water surface of the pond into multiple evenly distributed gridded zones. The optical intensity data of each gridded partition is collected synchronously by an optical sensor array, and the average optical intensity of each gridded partition is calculated. Based on the average optical intensity gradient change of adjacent gridded partitions, areas where the average optical intensity continuously exceeds a preset cluster threshold are identified as fish fry cluster areas. Geometric mean calculation is performed on the geometric centers of all gridded partitions within the fish fry cluster area, and the calculation result is used as the coordinates of the center position of the fish fry cluster area.

3. The feeding decision optimization method according to claim 2, characterized in that, The monitoring frequency of the optical sensor array is dynamically adjusted according to the area changes of the fish fry cluster area. When the area increases, the monitoring frequency is increased proportionally, and when the area decreases, it returns to the reference frequency.

4. The feeding decision optimization method according to claim 1, characterized in that, Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, the horizontal position of the feeding device is dynamically adjusted to within a preset offset range of the center position coordinates corresponding to the phototactic response threshold, including: Calculate the absolute value of the difference between the actual light intensity in the aquaculture pond at the current moment and the phototactic response threshold; The direction of movement of the feeding device is determined by comparing the absolute value of the difference with the preset difference range. The dynamic adjustment step size is generated by multiplying the moving direction and the absolute value of the difference. The feeding device is controlled to move along the moving direction with the dynamic adjustment step size until it enters the preset offset range. The offset distance between the current position of the feeding device and the center position coordinates is detected in real time to see if it meets the preset offset range constraint condition.

5. The feeding decision optimization method according to claim 4, characterized in that, When the absolute value of the difference exceeds the upper limit of the preset difference range, the feeding device is controlled to move in the positive direction of the center position coordinate corresponding to the phototactic response threshold; when the absolute value of the difference is lower than the lower limit of the preset difference range, it moves in the negative direction.

6. The feeding decision optimization method according to claim 1, characterized in that, At the adjusted horizontal position, the area ratio of the divergence region is calculated based on the divergence analysis of the fish movement vector field. The initial feeding amount is generated based on the difference between this ratio and a preset activation threshold, and the feeding is then executed, including: A continuous sequence of images of fish movement at an adjusted horizontal position is obtained using a high-speed camera device. Optical flow is used to calculate the motion vector field of the fish school for adjacent frames; Divergence calculation is performed on the motion vector field, and grid regions with divergence values ​​greater than zero are marked as divergence regions; The proportion of the area of ​​the divergent region to the total area covered by the fish movement vector field is statistically analyzed. When generating the initial feeding amount based on the difference between the ratio and the preset activation threshold, if the ratio is greater than or equal to the preset activation threshold, the initial feeding amount is the product of the baseline feeding amount and the difference ratio; if the ratio is less than the preset activation threshold, the initial feeding amount is the output value of a linearly decreasing function of the baseline feeding amount and the absolute value of the difference.

7. The feeding decision optimization method according to claim 1, characterized in that, After the initial feeding amount is released, the dynamic behavioral characteristics of the fish's feeding trajectory are extracted to determine whether the fish are in a random, dispersed, or aggregated feeding state. The initial feeding amount is then adjusted based on the state type and duration, including: The movement of fish schools towards the bait is continuously captured by underwater camera devices, and the movement trajectory is constructed by extracting the coordinate sequence of the fish school's center of gravity. The fractal dimension of the movement trajectory is calculated. If the fractal dimension is lower than the preset fractal threshold, it is determined to be a state of aggregated feeding. If the fractal dimension is higher than the preset fractal threshold, it is determined to be a state of random dispersion. The duration of the current state type is recorded. When the duration of the feeding state exceeds the preset duration threshold, the initial feeding amount is increased by the first correction coefficient. When the duration of the random dispersion state exceeds the preset duration threshold, the initial feeding amount is reduced by the second correction coefficient. If the duration does not exceed the preset duration threshold, the initial feeding amount will remain unchanged.

8. The feeding decision optimization method according to claim 1, characterized in that, The next feeding time is determined based on the periodic aggregation pattern. S1 to S5 are repeated until the preset feeding cycle is completed, including: Extract the principal period length from the effective periodic clustering pattern; The next feeding interval is calculated based on the sliding window average of the main cycle length and historical feeding cycle data. If the current feeding time deviates from the peak time of light intensity corresponding to the periodic aggregation pattern by more than the preset time tolerance, the next feeding time interval will be adjusted according to the deviation ratio. Update the next feeding time node to the sum of the current time node and the corrected feeding time interval, and trigger S1 to S5 to execute until the cumulative number of feedings reaches the preset total number of feeding cycles.

9. A feeding decision optimization system, used to implement the feeding decision optimization method according to any one of claims 1-8, characterized in that, Includes the following modules: Cluster positioning module: Real-time monitoring of the optical intensity distribution in the fish fry cluster area within the aquaculture pond, and recording the coordinates of the center position of the fish fry cluster area; Periodic threshold module: Based on continuous time series data of optical intensity distribution, it determines the periodic aggregation pattern and phototactic response threshold of fish fry cluster areas through time series analysis; Dynamic positioning module: Based on the difference between the phototactic response threshold and the actual light intensity in the aquaculture pond at the current moment, dynamically adjust the horizontal position of the feeding device to a preset offset range of the center position coordinates corresponding to the phototactic response threshold. Divergence Feeding Module: At the adjusted horizontal position, the divergence area ratio is calculated based on the divergence analysis of the fish movement vector field. The initial feeding amount is generated based on the difference between the ratio and the preset activation threshold, and the feeding is executed. Behavior correction module: After the initial feeding amount is released, the dynamic behavioral characteristics of the fish's feeding trajectory are extracted and the fish are determined to be in a state of random dispersion or group feeding. The initial feeding amount is corrected according to the state type and duration. Periodic scheduling module: Determines the next feeding time node based on the periodic aggregation pattern, and repeatedly executes the cluster positioning module, periodic threshold module, dynamic positioning module, dispersion feeding module and behavior correction module until the preset feeding cycle is completed.

Images