A feeding control method based on deep learning of spatiotemporal characteristics of fish feeding behavior

Abstract

This invention relates to the field of intelligent aquaculture, and more particularly to a feeding control method based on deep learning of the spatiotemporal features of fish feeding behavior. The feeding control method based on deep learning of the spatiotemporal features of fish feeding behavior includes the following steps: S1: Continuously acquiring video frame sequences of the water surface in the feeding area; preprocessing the video frame sequences to extract spatiotemporal features, obtaining information on the fragmentation of the splash area and the turbulence intensity gradient; S2: Calculating instantaneous turbulence flow scheduling based on the fragmentation information and the turbulence intensity gradient information, wherein the instantaneous turbulence flow scheduling reflects the degree of interference caused by the current splash turbulence on the lateral line perception of the fish. This invention constructs a comprehensive index of dual vulnerability states—instantaneous turbulence flow scheduling and physiological perception—from the fragmentation of splashes and the turbulence intensity gradient in the water surface video, and combines this with fluctuation characteristics to achieve advanced early warning and phased adaptive feeding control for fish feeding collapse.

Description

Technical Field

[0001] This invention relates to the field of intelligent aquaculture, and in particular to a feeding control method based on deep learning of the spatiotemporal characteristics of fish feeding behavior. Background Technology

[0002] In intensive aquaculture, the feeding intensity and group behavior of fish are the core basis for judging the amount of feed and adjusting the feeding strategy. Existing technologies usually assess the feeding activity by collecting video images of the water surface and analyzing the visual characteristics of the splash area and the fish aggregation degree. However, in actual high-density feeding scenarios, the physiological stress and sensory function inhibition of individual fish are not entirely manifested as an obvious cessation of feeding or a sharp reduction in splash area. They are often reflected in the continuous interference of underwater turbulence caused by fierce competition for feeding on the lateral line senses of the fish, as well as the oxygen debt burden accumulated as a result. This kind of physiological and sensory dual stress has the characteristics of concealment and lag, and it is difficult to identify by directly observing the splash size or the apparent information of the fish outline. However, it will gradually accumulate during continuous feeding, eventually leading to the sudden collective cessation of feeding by the fish, forming a feeding collapse. Existing technologies for feeding control mostly rely on overt behavioral indicators such as splash area thresholds, changes in fish swimming speed, or feeding time windows for judgment. At the regulatory level, they depend on preset fixed rules or empirical intermittent strategies. Decisions are mainly based on directly observable optical features of the water surface, lacking the ability to deeply decode the coupled physiological perception information of the fish within the splash turbulence. However, the degree of interference of turbulence on the lateral line perception of the fish and the resulting physiological vulnerability of the group are implicit variables that are difficult to measure directly and cannot be characterized by conventional image features or single behavioral parameters. Due to the lack of technical means to establish a mapping relationship between the visual texture features of splash fragmentation and turbulent spatial distribution and the dual vulnerability of the fish's physiological perception, existing methods have difficulty in timely identifying the critical phase transition precursors of the group state in the early warning stage of feeding collapse. At the same time, the intervention process usually adopts a uniform feeding stop or fixed intermittent ratio, lacking the ability to refine and stage-by-stage control based on the dynamic evolution of the fish's vulnerable state, thus limiting the real-time and adaptive level of feeding control. Summary of the Invention

[0003] To overcome the difficulty in early detection of the dual stress of fish's concealed physiological perception, this invention provides a feeding control method based on deep learning of the spatiotemporal characteristics of fish feeding behavior.

[0004] The technical implementation scheme of the present invention is as follows: a feeding control method based on deep learning of the spatiotemporal features of fish feeding behavior, comprising the following steps: S1: Continuously acquire video frame sequences of the water surface in the feeding area, and extract spatiotemporal features after preprocessing the video frame sequences to obtain information on the fragmentation of the water splash area and the turbulence intensity gradient. S2: Calculate the instantaneous turbulence flow schedule based on the fragmentation information and the turbulence intensity gradient information. The instantaneous turbulence flow schedule reflects the degree of interference caused by the current splash turbulence on the lateral line perception of the fish school. S3: Combining the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area, calculate a comprehensive index characterizing the severity of the dual vulnerability state of physiological perception within the fish population. S4: Perform statistical behavior analysis on the sequence of the comprehensive index within the sliding time window, and extract the fluctuation characteristics of the comprehensive index near the phase transition critical point; S5: Determine the current phase transition stage based on the fluctuation characteristics and output the corresponding feeding control command. The feeding control command includes maintaining feeding, reducing feeding and inserting an interval, stopping feeding, and resuming feeding.

[0005] Preferably, the continuous acquisition of video frame sequences of the water surface in the feeding area, followed by preprocessing of the video frame sequences to extract spatiotemporal features, yields information on the fragmentation of the splash area and turbulence intensity gradient information. This includes: the fragmentation information being a splash fragmentation segmentation map, which classifies each pixel in the splash area into large, continuous splashes or small, fragmented splashes; the ratio of the area of ​​small splashes to the total splash area in the splash fragmentation segmentation map reflects the uniformity of the fish's feeding actions and the degree of disorder in individual competition; and the turbulence intensity gradient information being a turbulence intensity gradient direction distribution map of the splash area, which characterizes the rate and direction of change in turbulence intensity within a local area, reflecting the spatial pattern of turbulent energy diffusion from the center outwards.

[0006] Preferably, the step of calculating the instantaneous turbulence flow schedule based on the fragmentation information and the turbulence intensity gradient information, wherein the instantaneous turbulence flow schedule reflects the degree of interference caused by the current splash turbulence to the lateral line perception of the fish, includes: quantifying the intensity of interference caused by the underwater acoustic disturbance generated by the splash turbulence to the normal functioning of the fish's lateral line sensory perception by constructing the instantaneous turbulence flow schedule; when the fish are feeding in an orderly manner and the turbulence energy is concentrated in the central area of ​​the splash, the spatial non-uniformity of the turbulence is low, the energy concentration is high, and the proportion of fragmented splashes is low, the instantaneous turbulence flow schedule is in a low value range, indicating that the lateral line perception of the fish is slightly disturbed; when the fish are feeding in an disorderly manner and the turbulence energy is dispersed, the proportion of fragmented splashes increases, the instantaneous turbulence flow schedule increases sharply, indicating that the interference to the lateral line perception of the fish is more severe.

[0007] Preferably, the step of quantifying the interference intensity of underwater acoustic disturbances generated by splash turbulence on the normal functioning of the fish's lateral line sensory perception by constructing the instantaneous turbulence flow scheduling includes: the instantaneous turbulence flow scheduling is calculated according to the following formula: ;in, For the first The instantaneous turbulence scheduling at that moment; For the first The average value of the turbulent gradient amplitude within the water splash area at a given time reflects the spatial non-uniformity of the turbulence; For the first The standard deviation of the turbulence intensity within the splash area at a given time reflects the degree of concentration of turbulent energy; the larger the standard deviation, the more concentrated the energy. For the first The ratio of the area of ​​fine water splashes to the total water splash area at any given time; the lower the proportion of fine water splashes, the better. The closer it is to 0, the more its complement term... The closer to 1; The preset minimum positive number is used to prevent division by zero errors when the complement term is 0; when the fish feed in an orderly manner, the molecules... Smaller denominator Larger and complementary terms close to 1, making At low values; when turbulence is diffuse and feeding is disordered, the numerator increases while the denominator is low. Decrease the complement term, so that It increased dramatically.

[0008] Preferably, the comprehensive index characterizing the severity of the dual vulnerability state of physiological perception within the fish population is calculated by combining the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area. This includes: recording the continuous duration for which the splash area has remained above a preset high water level threshold since the end of the most recent event of a significant decrease in splash area, as the steady-state duration; calculating the theoretical safe recovery period based on the instantaneous turbulent flow scheduling, where the theoretical safe recovery period represents the shortest time that the fish population needs to suspend feeding to breathe and recover its senses under the current degree of turbulent mismatch, thereby reducing the continuous accumulation of the dual vulnerability state of physiological perception. This is obtained by performing an exponential mapping of the instantaneous turbulent flow scheduling with a natural constant as the base, with the aim of mapping the high value range of mismatch to a longer recovery time requirement in a non-linear manner. The theoretical safe recovery period shows an accelerating growth trend with the increase of the instantaneous turbulent flow scheduling, reflecting the physiological law that the recovery time required by the fish population is disproportionately extended when the degree of turbulent mismatch is higher.

[0009] Preferably, the step of calculating the theoretical safe recovery period based on the instantaneous turbulence flow scheduling includes: the calculation formula for the theoretical safe recovery period is: ;in, For the first The theoretical safe recovery period at any given time; The base recovery time represents the minimum recovery time required for the fish swarm to recover when the instantaneous turbulent flow scheduling approaches 0. For the natural constant An exponential function with base 0; The growth rate adjustment coefficient controls the growth rate of the theoretical safe recovery period as the instantaneous turbulent flow scheduling changes. For the first The instantaneous turbulence scheduling at that moment; The reference mismatch constant is used to normalize the instantaneous turbulent flow scheduling to a value similar to... Matching scale range; when Less than At that time, the theoretical safe recovery period increases at a slow rate. Exceed Subsequently, the theoretical safe recovery period increases rapidly due to the exponential effect, causing the theoretical recovery time required for the fish population under high turbulence and mismatch conditions to exceed the basic recovery time.

[0010] Preferably, the calculation of a comprehensive index characterizing the severity of the dual vulnerability state of physiological perception within the fish population, combining the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area, includes: the comprehensive index being calculated according to the following formula: ;in, For the first The comprehensive indicators mentioned at any given time; For the first The instantaneous turbulence scheduling at that moment; As of the date The duration of the steady state recorded at each moment; For the first The theoretical safe recovery period at any given time; The saturation rate adjustment coefficient controls the degree to which the ratio of the steady-state duration to the theoretical safe recovery period affects the saturation rate of the second factor. The first factor is a preset, extremely small positive number used to prevent division by zero errors when the theoretical safe recovery period is 0; The instantaneous turbulent flow schedule increases monotonically with a gradually decreasing rate of increase, and a compression transformation with a continuously decreasing rate of increase is applied to the unbounded instantaneous turbulent flow schedule; the second factor in the formula It is an exponential saturation function, which approaches 0 when the steady-state duration is much shorter than the theoretical safe recovery period, and approaches 1 when the steady-state duration is close to or exceeds the theoretical safe recovery period; the product of the first factor and the second factor makes the comprehensive index significantly increase only when both the instantaneous turbulence outflow is high and the steady-state duration exceeds the theoretical safe recovery period are met simultaneously. When one condition is met but the other condition is missing, the comprehensive index is suppressed to a low value range.

[0011] Preferably, the step of performing statistical behavior analysis on the sequence of the comprehensive index within the sliding time window to extract the fluctuation characteristics of the comprehensive index near the phase transition critical point includes: calculating three indicators—local variance, skewness, and autocorrelation decay time—of the sequence of the comprehensive index within a preset length sliding time window; the local variance is the variance of the comprehensive index values ​​within the sliding time window, used to measure the degree of fluctuation of the comprehensive index within the window, and the local variance increases significantly compared to the stable period when the internal state of the fish population tends to be unstable; the skewness is the variance of the comprehensive index values ​​within the sliding time window. The third standard moment is used to measure the asymmetry of the distribution of the comprehensive index. When the skewness is positive, it indicates that the comprehensive index is frequently testing higher levels. When the skewness is negative, it indicates that the comprehensive index is in the process of continuously falling back from a high level. The autocorrelation decay time is the time required for the autocorrelation function value of the comprehensive index to decay from the peak to a preset ratio. It is used to measure the strength of the time dependence of the comprehensive index sequence. When the comprehensive index approaches the phase transition critical point, the time inertia of the comprehensive index increases, and the autocorrelation decay time becomes abnormally prolonged, constituting an early warning signal of the phase transition critical point.

[0012] Preferably, the step of determining the current phase transition stage based on the fluctuation characteristics and outputting corresponding feeding control instructions includes maintaining feeding, reducing feeding and inserting intervals, stopping feeding, and resuming feeding, including: when the comprehensive index is in a low value range, the local variance is minimal, and the skewness fluctuates around 0; when the autocorrelation decay time is short, it is determined that the current feeding stage is normal, the fish are feeding in an orderly manner and their perception is smooth, and an instruction to maintain the current feeding rate is output; when the comprehensive index begins to rise continuously, the local variance increases compared to the normal feeding stage, and the autocorrelation decay time is prolonged; when the skewness turns positive, it is determined that the current stage is a vulnerable warning stage, the statistical behavior of the comprehensive index indicates that the fish are approaching the critical point of a physiologically and perceptually vulnerable state, and an instruction to reduce the feeding rate to a preset ratio of the baseline rate and forcibly insert an interval of a preset duration within a unit time is output to actively create a time window for the fish to breathe and recover their perception; when the comprehensive index When the skewness shifts from positive to negative after a significant jump to a high level and approaches saturation, it is determined that the current stage is a vulnerable interruption phase. The fish have been forcibly stopped feeding due to the combined effects of oxygen debt accumulation and sensory inhibition. An instruction to immediately stop feeding is output, and the duration of the interruption is recorded. It is determined that the current stage is a trial recovery phase, in which feeding is slowly restarted while the comprehensive index is continuously monitored. After feeding stops, when the comprehensive index falls from a high level to a preset decline range and the skewness turns from negative to positive again, it is determined that the current stage is a trial recovery phase. An instruction to slowly restart feeding at a preset ratio lower than the normal baseline rate is output, while the trend of the comprehensive index is continuously monitored. The triggering of the trial recovery phase is solely determined by the decline range and the skewness turning positive. When the comprehensive index returns to a low value and stabilizes, the normal feeding phase resumes. When the comprehensive index rises again during the trial feeding process, a stop feeding instruction is output again, extending the waiting time before attempting recovery again.

[0013] Preferably, the determination that the current stage of trial recovery has been entered, and the trial recovery is carried out by slowly starting feeding and continuously monitoring the comprehensive indicators, includes: recording the occurrence time of each vulnerable interruption stage, the peak value of the comprehensive indicators during this interruption, and the actual recovery time experienced from stopping feeding to successfully resuming feeding, forming an interruption event record sequence for the breeding batch; when a vulnerable interruption event occurs, retrieving the records of the previous interruption events; when the actual recovery time in the records of the previous interruption events shows a gradual increasing trend, determining that the fish population's responsiveness to the trial recovery stage tends to decrease, increasing the reduction of the rate of slowly starting feeding and the extension of the waiting time for resuming the trial recovery stage, wherein the reduction and extension are determined by a preset mapping function based on the trend of the recovery time; when the frequency of the vulnerable interruption event exceeds the preset frequency upper limit within a preset long-term observation window, triggering the instruction to reduce the feeding rate and insert an intermittent when entering the vulnerable warning stage again, and actively intervening before the comprehensive indicators have continuously risen to the original trigger level.

[0014] Beneficial Effects: This invention extracts information on the fineness of water splashes and the gradient of turbulence intensity from surface videos and establishes a quantitative mapping relationship between these information and the degree of lateral line sensory interference in fish schools, achieving cross-modal inference from visible surface texture to invisible underwater sensory states. By constructing an instantaneous turbulence outflow scheduling index, the interference intensity of water splashes and turbulence on the lateral line sensory perception of fish schools is transformed into a calculable indirect representation, enabling a quantitative characterization of the degree of sensory disturbance within the fish school, thereby improving the early identification capability of hidden physiological stress. By multiplicatively coupling instantaneous turbulence outflow scheduling with the steady-state duration of high water splash area, a comprehensive index characterizing the dual vulnerability state of physiological perception is constructed, achieving unified modeling of the two dimensions of sensory disturbance and oxygen debt accumulation. The index only significantly increases when both conditions are met simultaneously, accurately pinpointing the true vulnerability state within the fish school and reducing the impact of a single factor. This approach addresses misjudgments and omissions; by conducting statistical behavioral analysis on the fluctuation characteristics of comprehensive indicators within a sliding time window, it extracts local variance, skewness, and autocorrelation decay time as precursor signals of critical phase transitions, enabling a leap from delayed response to proactive early warning, and proactively intervening and regulating before fish feeding collapse occurs; by dividing the phase transition into four stages—normal feeding, vulnerability warning, vulnerability interruption, and tentative recovery—and outputting corresponding graded control commands such as maintaining, reducing feeding and inserting intermittent feeding, stopping feeding, and slow recovery, the regulation strategy achieves stage-specific targeting and precision; by recording each vulnerability interruption event and feeding it back to the early warning threshold adjustment process, an adaptive correction closed loop based on aquaculture process data is formed, enabling autonomous adaptation to the evolution of fish vulnerability under specific aquaculture species, densities, and environmental conditions during long-term operation, continuously optimizing the timing and intensity of regulation. Attached Figure Description

[0015] Figure 1 This is a flowchart of the feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to the present invention; Figure 2 This is a flowchart of the phase transition stage determination and control command output method based on fluctuation characteristics of the present invention. Detailed Implementation

[0016] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example: A feeding control method based on deep learning of spatiotemporal features of fish feeding behavior, such as... Figure 1 and Figure 2 As shown, it includes the following steps: S1: Continuously acquire video frame sequences of the water surface in the feeding area, and extract spatiotemporal features after preprocessing the video frame sequences to obtain information on the fragmentation of the water splash area and the turbulence intensity gradient. S2: Calculate the instantaneous turbulence flow schedule based on the fragmentation information and the turbulence intensity gradient information. The instantaneous turbulence flow schedule reflects the degree of interference caused by the current splash turbulence on the lateral line perception of the fish school. S3: Combining the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area, calculate a comprehensive index characterizing the severity of the dual vulnerability state of physiological perception within the fish population. S4: Perform statistical behavior analysis on the sequence of the comprehensive index within the sliding time window, and extract the fluctuation characteristics of the comprehensive index near the phase transition critical point; S5: Determine the current phase transition stage based on the fluctuation characteristics and output the corresponding feeding control command. The feeding control command includes maintaining feeding, reducing feeding and inserting an interval, stopping feeding, and resuming feeding.

[0018] The fragmentation information is a water splash fragmentation segmentation map, which classifies each pixel in the water splash area into large continuous water splashes or small splashes. The ratio of the area of ​​small splashes to the total area of ​​water splashes in the water splash fragmentation segmentation map reflects the uniformity of the fish's feeding behavior and the degree of disorder in individual competition. The turbulence intensity gradient information is a turbulence intensity gradient direction distribution map of the water splash area. The turbulence intensity gradient direction distribution map characterizes the rate and direction of change of turbulence intensity in a local area, reflecting the spatial pattern of turbulence energy spreading from the center outward.

[0019] It should be explained that the process of obtaining the fragmentation information includes uniform preprocessing of the video frame sequence and pixel-level classification based on a deep learning segmentation model. The preprocessing includes denoising and filtering consecutive video frames to suppress water surface reflection noise, correcting for illumination changes through grayscale normalization, and extracting the water splash region as the region of interest based on edge detection. The water splash fragmentation segmentation map is obtained through a pre-constructed and trained semantic segmentation network. This semantic segmentation network employs an encoder and decoder structure, taking consecutive frames or inter-frame difference results as input, and outputting a classification result with the same size as the input image through convolutional feature extraction and multi-scale feature fusion. The semantic segmentation network is constructed by selecting a convolutional neural network structure with pixel-level classification capabilities and initializing the network parameters. The training process involves supervised learning of the network based on manually labeled training samples. Forward propagation is used to calculate the predicted segmentation result, which is then compared with the labeled result to obtain the loss value. The training samples for the semantic segmentation network are obtained by manually labeling water splash areas in historical video frames. Regions with continuous morphology, smooth boundaries, and strong spatial connectivity are labeled as large, continuous water splashes, while regions exhibiting discrete splashes, broken boundaries, and sparse spatial distribution are labeled as small, fragmented splashes. The network parameters are optimized by minimizing the segmentation loss function. The iterative update of the network parameters; the segmentation loss function consists of pixel-level classification error terms, used to measure the difference between the predicted classification result and the labeled classification result. The iterative update of the network parameters is achieved by calculating the gradient of each layer parameter according to the loss function through the backpropagation algorithm and gradually adjusting the parameter values ​​to reduce the value of the loss function; the ratio of the area of ​​fine splashes to the total area of ​​splashes is calculated by counting the number of pixels belonging to the fine splash category in the water splash fragmentation image and calculating the ratio with the total number of pixels in the water splash area. The calculation process involves traversing and counting the category labels of each pixel in the segmentation image and accumulating the pixel counts belonging to the fine splash category. The value is compared with the total pixel count within the splash area, and the ratio of the pixel count to the splash area is calculated to obtain the ratio of the area of ​​the fine splashes to the total splash area; the number of pixels is obtained by traversing the classification label matrix within the region of interest; the turbulence intensity gradient information is obtained by obtaining the motion velocity field within the splash area based on the grayscale changes between adjacent video frames, and the amplitude of the motion velocity field is used as an approximate representation of the turbulence intensity. The gradient of the turbulence intensity is calculated in space. The gradient calculation is performed by performing differential calculations on the turbulence intensity distribution map in the horizontal and vertical directions to obtain the gradient vector field, and then calculating the gradient amplitude and direction at each pixel position;The turbulence intensity gradient direction distribution map is constructed by discretizing the gradient direction according to a preset angle interval, statistically analyzing the distribution of gradient vectors within each direction interval, and forming a distribution map characterizing the turbulence change direction features within the splash region. The preset angle intervals are divided according to directional resolution requirements. The preset angle intervals are determined by performing a full-range statistical analysis of the gradient direction to obtain a direction distribution histogram. Based on the resolution requirements of directional changes in the direction distribution histogram, the minimum distinguishable directional difference is set as the dividing criterion, and the directional angle range is equally divided into multiple intervals as the directional distribution statistical results. These directional distribution statistical results reflect the dominant diffusion direction and uniformity of turbulent energy in space.

[0020] By constructing the instantaneous turbulence flow schedule, the intensity of underwater acoustic disturbances caused by splash turbulence on the normal functioning of fish's lateral line sensory perception of water flow is quantified. When the fish are feeding in an orderly manner and the turbulent energy is concentrated in the central area of ​​the splash, the spatial non-uniformity of the turbulence is low, the energy concentration is high, and the proportion of fine splashes is low. The instantaneous turbulence flow schedule is in the low value range, indicating that the fish's lateral line perception is slightly disturbed. When the fish are feeding in an disorderly manner and the turbulent energy is dispersed, the proportion of fine splashes increases, and the instantaneous turbulence flow schedule increases sharply, indicating that the disturbance to the fish's lateral line perception is more severe.

[0021] It should be explained that the intensity of the underwater acoustic disturbance caused by the splashing turbulence on the normal functioning of the fish's lateral line sensory perception of water flow is considered in three aspects: the spatial non-uniformity of turbulence reflected by the amplitude of the turbulence gradient in the splashing area, the degree of turbulence energy concentration reflected by the standard deviation of turbulence intensity, and the degree of disorder in feeding behavior reflected by the ratio of the area of ​​the fragmented splashes to the total area of ​​the splashes. The construction process of the instantaneous turbulence flow scheduling is based on the fragmentation information and the turbulence intensity gradient information. By uniformly calculating and fusing multiple intermediate quantities reflecting the dynamic characteristics of splashes, a quantitative index is formed to characterize the degree of disturbance to the lateral line sensory perception of the fish. Specifically, within the water splash area, the gradient amplitude distribution is calculated based on the turbulence intensity gradient information. The overall characteristic of the turbulence gradient amplitude is obtained by statistically averaging the gradient amplitudes in space. This overall characteristic is used to characterize the spatial non-uniformity of the flow structure within the water splash area. The calculation process involves accumulating the gradient amplitudes corresponding to each pixel in the gradient vector field and taking the average. The turbulence intensity distribution is discretized, and the statistical standard deviation is calculated. The turbulence intensity is characterized by the amplitude of the motion velocity field. The standard deviation is calculated by taking the mean of the turbulence intensity values ​​of all pixels within the water splash area and then applying this mean to the turbulence intensity of each pixel. The root mean square of the squared deviations between the intensity and the mean is calculated to obtain a statistical measure reflecting the degree of turbulent energy concentration. A larger standard deviation indicates a more significant energy accumulation in a local area. The ratio of the area of ​​the fine splashes to the total splash area is used as a characterizing parameter describing the spatial dispersion of feeding behavior. This ratio parameter has been obtained statistically from the pixel classification results of the splash fragmentation segmentation map. By combining the area of ​​the fine splashes, the total splash area, and the ratio parameter into a model, the instantaneous turbulent flow scheduling can simultaneously reflect the coupling effects of spatial structure, energy distribution, and behavioral patterns. During the calculation process, the intermediate values ​​are... The data are uniformly time-aligned to make the instantaneous turbulent flow scheduling comparable across different time frames. When the gradient amplitude within the splash region is generally low and the standard deviation is high, and the ratio of the area of ​​fine splashes to the total splash area is low, it indicates that the splash flow structure tends to be concentrated and the feeding behavior remains orderly, corresponding to a low level of instantaneous turbulent flow scheduling. Conversely, when the gradient amplitude generally increases and the standard deviation decreases, and the ratio of the area of ​​fine splashes to the total splash area increases, it indicates that the turbulent structure tends to be diffuse and the degree of disorder in feeding behavior increases, leading to a rapid increase in instantaneous turbulent flow scheduling, which is used to characterize the state where the lateral line perception of the fish school is significantly disturbed.

[0022] The instantaneous turbulence flow scheduling is calculated according to the following formula: ;in, For the first The instantaneous turbulence scheduling at that moment; For the first The average value of the turbulent gradient amplitude within the water splash area at a given time reflects the spatial non-uniformity of the turbulence; For the first The standard deviation of the turbulence intensity within the splash area at a given time reflects the degree of concentration of turbulent energy; the larger the standard deviation, the more concentrated the energy. For the first The ratio of the area of ​​fine water splashes to the total water splash area at any given time; the lower the proportion of fine water splashes, the better. The closer it is to 0, the more its complement term... The closer to 1; The preset minimum positive number is used to prevent division by zero errors when the complement term is 0; when the fish feed in an orderly manner, the molecules... Smaller denominator Larger and complementary terms close to 1, making At low values; when turbulence is diffuse and feeding is disordered, the numerator increases while the denominator is low. Decrease the complement term, so that It increased dramatically.

[0023] It should be explained that the calculation of the instantaneous turbulent flow scheduling adopts a fractional structure to normalize and couple different physical quantities; wherein, the numerator term The spatial average of the gradient amplitude directly derived from the water splash region is used to characterize the degree of spatial non-uniformity of the turbulent structure. A larger value indicates more drastic changes in local flow, which will create a stronger disturbance to the lateral line perception of fish schools; the denominator term is the standard deviation of the turbulence intensity. With complement Multiplication constitutes, where Used for weighted modulation of the energy concentration of the molecule, when the energy is concentrated Increasing the value of the denominator amplifies its overall value, thus inhibiting the effect on the numerator; the complement term... By inversely mapping the proportion of fine water splashes, the complement term approaches a larger value when the feeding behavior is more ordered, thereby further increasing the denominator value to suppress the instantaneous turbulent flow scheduling. Conversely, when the proportion of fine water splashes increases, the complement term decreases, weakening the constraint effect of the denominator on the numerator term; the preset minimum positive number The method for determining this ratio is as follows: during the initialization phase, historical data distribution of the ratio of the area of ​​fine water splashes to the total area of ​​water splashes is statistically analyzed to obtain the minimum non-zero change in the ratio parameter. Based on this, a value not exceeding the minimum non-zero change is selected as the ratio. The calculation process of the formula is as follows: at each moment... Regarding the , and Synchronous reading is performed, and the instantaneous turbulent flow scheduling is solved by first calculating the complement term, then performing denominator multiplication, and finally executing numerator and denominator operations. To ensure the stability of the calculation results across different time series, the... and Before input calculation, a uniform scale normalization process is performed. This normalization method involves constructing a reference interval based on historical stable phase data and linearly mapping the current value. Small value and When the complement term is large, the overall denominator is large, making the... Keep in the low range; when the Increase and When the complement term decreases synchronously, the decrease in the denominator leads to a rapid amplification of the fractional result, thereby achieving a sensitive response to the degree of mismatch in turbulence.

[0024] The duration of the steady-state period is recorded as the continuous length of time during which the water surface area remains above a preset high water level threshold since the end of the most recent event of a significant decrease in water surface area. The theoretical safe recovery period is calculated based on the instantaneous turbulent flow scheduling. This theoretical safe recovery period represents the shortest time that the fish population needs to suspend feeding to breathe and recover their senses under the current degree of turbulent mismatch, thereby reducing the continuous accumulation of physiological and sensory vulnerability. It is obtained by performing an exponential mapping of the instantaneous turbulent flow scheduling with a natural constant as the base, aiming to non-linearly map the high-value range of mismatch to a longer recovery time requirement. The theoretical safe recovery period shows an accelerating growth trend with the increase of the instantaneous turbulent flow scheduling, reflecting the physiological law that the recovery time required by the fish population increases disproportionately with the higher the degree of turbulent mismatch.

[0025] It should be explained that the steady-state duration is determined based on the time-series change in the splash area. This is achieved by calculating the area of ​​the splash region frame-by-frame in consecutive video frames and constructing a time series. The occurrence of a significant decrease in splash area is then identified by combining this with the trend analysis. Specifically, the splash area is calculated by statistically analyzing the pixels within the splash region to obtain area values. Abrupt points are identified by calculating the area difference or rate of change between consecutive frames. When the rate of change in area exceeds a threshold set based on historical data fluctuations, it is marked as a significant decrease event. The threshold is determined based on the statistical analysis of the splash area change rate during historical normal feeding phases. The upper quantile boundary, which can distinguish between natural fluctuations and anomalous mutations, is selected as the criterion. This upper quantile boundary is determined by performing quantile statistics on the historical rate of change sequence and selecting quantiles that allow the false detection rate to be below a preset tolerance level. The preset high water level threshold is determined based on the historical statistical distribution of the splash area during the normal feeding phase. Area data within the stable feeding interval is extracted and quantile intervals are calculated. The upper quantile boundary reflecting the high-intensity feeding state is used as the preset high water level threshold. This upper quantile boundary is determined by statistically analyzing the area distribution during the stable phase and selecting quantiles that can stably distinguish the high-intensity feeding state. Starting from the end time of the most recent significant decrease event, the threshold is... The time period during which the splash area remains above the preset high water level threshold in the subsequent time series is accumulated until the current moment. The accumulated time length is the steady-state duration. The end of the significant decrease event is defined as the moment when the splash area recovers and remains stably above the preset high water level threshold within a continuous preset time window. The construction of the theoretical safe recovery period is based on the instantaneous turbulent flow schedule. By introducing an exponential mapping relationship, the instantaneous turbulent flow schedule is converted into a recovery time scale. The exponential mapping is implemented by using the instantaneous turbulent flow schedule as the input variable of an exponential function and constructing a nonlinear growth model with the natural constant as the base, so that the theoretical safe recovery period is low. The mismatch interval changes gradually but expands rapidly in the high mismatch interval. The relevant parameters in the exponential mapping are obtained by statistically analyzing the actual time distribution required for fish to recover to normal feeding status under different instantaneous turbulent mismatch levels based on feeding interruption event data recorded during historical aquaculture. By fitting the relationship curve between the actual recovery time and instantaneous turbulent mismatch, the parameter values ​​used to control the growth rate and scale mapping in the exponential function are determined. The theoretical safe recovery period reflects the changing pattern of the physiological recovery needs of fish under different degrees of turbulent mismatch. When the instantaneous turbulent mismatch is low, the recovery period is close to the baseline level. When the instantaneous turbulent mismatch increases, the recovery period shows an accelerated growth trend.

[0026] The formula for calculating the theoretical safe recovery period is as follows: ;in, For the first The theoretical safe recovery period at any given time; The base recovery time represents the minimum recovery time required for the fish swarm to recover when the instantaneous turbulent flow scheduling approaches 0. For the natural constant An exponential function with base 0; The growth rate adjustment coefficient controls the growth rate of the theoretical safe recovery period as the instantaneous turbulent flow scheduling changes. For the first The instantaneous turbulence scheduling at that moment; The reference mismatch constant is used to normalize the instantaneous turbulent flow scheduling to a value similar to... Matching scale range; when Less than At that time, the theoretical safe recovery period increases at a slow rate. Exceed Subsequently, the theoretical safe recovery period increases rapidly due to the exponential effect, causing the theoretical recovery time required for the fish population under high turbulence and mismatch conditions to exceed the basic recovery time.

[0027] It should be explained that the calculation of the theoretical safe recovery period adopts an exponential mapping form, converting the instantaneous turbulent flow scheduling into a physiologically meaningful time-scale parameter, thereby achieving a nonlinear characterization of the fish population's recovery needs; the basic recovery time... This is used to characterize the shortest time required for the fish swarm to perform basic ventilation and sensory recovery when the instantaneous turbulent flow scheduling approaches 0. Within the time period when the instantaneous turbulent flow scheduling is below a preset low disturbance threshold, time samples are extracted showing the fish swarm recovering to a stable feeding state after a corresponding feeding interruption event. The mean of these time samples is then calculated as... The value of ; scheduling the instantaneous turbulence. With the reference mismatch constant Ratio processing is performed to create dimensionless input variables, eliminating the influence of differences in different breeding environments or observation scales on the calculation results, thus ensuring a uniform scale for the input of the exponential function; a growth rate adjustment coefficient is introduced. Linear scaling is applied to the dimensionless variable to control the growth slope of the exponential function output, wherein... The value of is determined by constructing sample pairs with the instantaneous turbulent flow scheduling as the independent variable and the corresponding actual recovery time as the dependent variable. The least squares method is used to fit the parameters of the exponential model to obtain the value that minimizes the fitting error. The value of the exponential function; The calculation process is as follows: Perform an exponentiation operation to solve for the base recovery time, and use the result as the amplification factor. Multiply to get the first... The theoretical safe recovery period corresponding to the time. ; For each instantaneous moment, the instantaneous turbulence flow schedule is synchronously input and an exponential mapping operation is performed to obtain a recovery period sequence that dynamically changes over time; when the When in the low level range A smaller value makes the exponential function output close to 1, making the theoretical safe recovery period close to the basic recovery time. When the stated Gradually increase and exceed the above After the corresponding scale, the exponential term grows rapidly, causing the theoretical safe recovery period to show an accelerated amplification trend, which is used to characterize the nonlinear enhancement feature of the fish population recovery demand under high turbulence mismatch conditions.

[0028] The comprehensive index is calculated according to the following formula: ;in, For the first The comprehensive indicators mentioned at any given time; For the first The instantaneous turbulence scheduling at that moment; As of the date The duration of the steady state recorded at each moment; For the first The theoretical safe recovery period at any given time; The saturation rate adjustment coefficient controls the degree to which the ratio of the steady-state duration to the theoretical safe recovery period affects the saturation rate of the second factor. The first factor is a preset, extremely small positive number used to prevent division by zero errors when the theoretical safe recovery period is 0; The instantaneous turbulent flow schedule increases monotonically with a gradually decreasing rate of increase, and a compression transformation with a continuously decreasing rate of increase is applied to the unbounded instantaneous turbulent flow schedule; the second factor in the formula It is an exponential saturation function, which approaches 0 when the steady-state duration is much shorter than the theoretical safe recovery period, and approaches 1 when the steady-state duration is close to or exceeds the theoretical safe recovery period; the product of the first factor and the second factor makes the comprehensive index significantly increase only when both the instantaneous turbulence outflow is high and the steady-state duration exceeds the theoretical safe recovery period are met simultaneously. When one condition is met but the other condition is missing, the comprehensive index is suppressed to a low value range.

[0029] It should be explained that the comprehensive index is constructed using the product of a logarithmic compression factor and a time saturation factor, aiming to describe the dual-dimensional coupling effect of disturbance intensity and duration; wherein, the first factor originates from the nonlinear transformation of the instantaneous turbulence runoff scheduling, through the... Taking the reciprocal and then performing a natural logarithm operation followed by a negative value, a compression transformation with a continuously decaying increasing rate is applied to the unbounded instantaneous turbulent mismatch scheduling. This makes the numerical growth more gradual as the mismatch increases, thus avoiding excessive amplification of changes in the high mismatch interval, while maintaining a monotonically increasing relationship to reflect the severity of perceived interference caused by turbulent mismatch. The second factor is constructed based on the relative relationship between the steady-state duration and the theoretical safe recovery period. By inputting the ratio of the two into an exponential decay function and then taking the complement form, a continuous transition from insufficient time to sufficient time is achieved. When the steady-state duration is short, the exponential term is close to 1, making the overall factor close to 0. As the steady-state duration gradually approaches or exceeds the theoretical safe recovery period, the exponential term decays rapidly, making the overall factor approach 1, reflecting the degree to which recovery demand is continuously accumulated. The saturation rate adjustment coefficient... The method for determining this is based on the correspondence between different durations and recovery effects during historical feeding processes, through the ratio... The response curve between the exponential function and the observed degree of deterioration in feeding behavior was fitted, and parameter values ​​that matched the inflection point of the exponential function with the actual physiological turning point were selected as the optimal values. During the calculation process, the instantaneous turbulent flow scheduling, the steady-state duration, and the theoretical safe recovery period are synchronously input at each moment. The logarithmic compression factor and the time saturation factor are calculated first, and then the product operation is performed to obtain the comprehensive index. When the instantaneous turbulent flow scheduling is low, even if the steady-state duration is long, the first factor is kept at a small value to suppress the result. When the steady-state duration is insufficient, even if the instantaneous turbulent flow scheduling is high, the second factor limits the growth of the result. The comprehensive index only rises rapidly when the instantaneous turbulent flow scheduling and the insufficient steady-state duration reach a high level at the same time.

[0030] For a sliding time window of a preset length, three indicators are calculated for the sequence of the comprehensive index: local variance, skewness, and autocorrelation decay time. The local variance is the variance of the comprehensive index values ​​within the sliding time window, used to measure the degree of fluctuation of the comprehensive index within the window. When the internal state of the fish population tends to become unstable, the local variance increases significantly compared to the stable period. The skewness is the third standard moment of the comprehensive index values ​​within the sliding time window, used to measure the asymmetry of the comprehensive index distribution. A positive skewness indicates that the comprehensive index frequently attempts to reach higher levels, while a negative skewness indicates that the comprehensive index is in the process of continuously declining from a high level. The autocorrelation decay time is the time required for the autocorrelation function value of the comprehensive index to decay from its peak to a preset proportion, used to measure the strength of the time dependence of the comprehensive index sequence. When the comprehensive index approaches the phase transition critical point, the time inertia of the comprehensive index increases, and the autocorrelation decay time becomes abnormally prolonged, constituting an early warning signal for the phase transition critical point.

[0031] It should be explained that the preset window length of the sliding time window is determined by comparing and analyzing the statistical stability and pre-phase transition response sensitivity of the historical comprehensive index sequence at different window scales, selecting a window length that balances stationarity and early warning lead time. The comprehensive index sequence is truncated moment-by-moment according to the preset window length, forming continuously updated subsequences. Local variance, skewness, and autocorrelation decay time are calculated synchronously within each window. The local variance is calculated based on the sum of the squared deviations between the comprehensive index and its mean at each moment within the window, and then averaged. This measures the increase in fluctuation amplitude within a short time scale. When the fish population transitions from a stable feeding state to a disorderly competition state, the fluctuation amplitude of the comprehensive index increases, causing the local variance to rise. The skewness is calculated based on the third-order central moment after standardizing the distribution of the comprehensive index within the window. This describes the skewness direction at the tail of the distribution. When in the critical leading edge stage, the comprehensive index exhibits intermittent transitions to higher values, causing the skewness to shift in the positive direction. During the recovery or decline phase, high values ​​decline predominate, and the skewness turns negative. The autocorrelation decay time is obtained by calculating the autocorrelation function of the comprehensive index sequence within the window. Using the autocorrelation value at the 0-lag position as a benchmark, the minimum lag step corresponding to the decay to a preset proportion threshold is found. The actual time length corresponding to the lag step is the autocorrelation decay time, which is used to characterize the sequence's ability to remember historical states. The preset proportion threshold is determined by statistically analyzing the decay characteristics of the autocorrelation function in the historical stable phase and verifying it with phase transition samples, selecting quantile values ​​that can distinguish between normal decay and abnormal delayed decay. The preset proportion threshold is also determined by statistically analyzing the historical autocorrelation decay curves and selecting decay proportions that can significantly distinguish between normal and abnormal decay modes. During the calculation process, the window length is fixed and the time axis is uniformly sampled. As the phase transition critical point is gradually approached, the recovery ability of the comprehensive index weakens, the duration of the disturbance is prolonged, the correlation between adjacent moments is enhanced, resulting in a slower decay rate of the autocorrelation function, which manifests as a significant extension of the autocorrelation decay time.

[0032] When the comprehensive index is in a low range, the local variance is minimal, and the skewness fluctuates around 0; when the autocorrelation decay time is short, it is determined that the fish are currently in a normal feeding phase, feeding in an orderly manner and with good sensory perception, and an instruction to maintain the current feeding rate is output; when the comprehensive index begins to rise continuously, the local variance increases compared to the normal feeding phase, and the autocorrelation decay time prolongs; when the skewness turns positive, it is determined that the fish are currently in a vulnerable warning phase, and the statistical behavior of the comprehensive index indicates that the fish are approaching the critical point of a state of dual physiological and sensory vulnerability, and an instruction to reduce the feeding rate to a preset proportion of the baseline rate and forcibly insert a preset duration of intermittent time within a unit of time is output to actively create a time window for the fish to breathe and recover their senses; when the comprehensive index jumps sharply to a high level and tends to saturate, the skewness changes from positive to negative. When the value is reached, it is determined that the current stage is a fragile interruption stage, where the fish have been forcibly stopped feeding due to the combined effects of oxygen debt accumulation and sensory inhibition. An instruction to immediately stop feeding is output, and the duration of the interruption is recorded. It is determined that the current stage is a trial recovery stage, in which feeding is slowly restarted and the comprehensive index is continuously monitored. After feeding stops, when the comprehensive index falls from a high level to a preset decline range and the skewness turns from negative to positive again, it is determined that the current stage is a trial recovery stage. An instruction to slowly restart feeding at a preset ratio lower than the normal baseline rate is output, while the trend of the comprehensive index is continuously monitored. When the comprehensive index returns to a low value and stabilizes, the normal feeding stage is resumed. When the comprehensive index rises again during the trial feeding process, a stop feeding instruction is output again, and the waiting time before attempting recovery is extended.

[0033] It should be explained that, using the statistical distribution constructed from historical local variance data accumulated during the normal feeding phase as a reference benchmark, the current local variance is dynamically compared with the historical normal fluctuation range to determine the minimum relative state and the triggering conditions for increase; the method for determining the low value range of the comprehensive index is to extract the lower quantile interval based on the historical stable feeding phase data distribution as the judgment boundary. The lower quantile interval is determined by statistically analyzing the distribution of the comprehensive index in the stable phase and selecting the low interval that can stably represent the normal feeding state. When the comprehensive index is in the interval for a long time, the local variance remains below the preset low fluctuation threshold, the skewness fluctuates slightly around 0, and the autocorrelation decays, the condition is considered met. When the time does not exceed the baseline time length, the fish are determined to be in a normal feeding phase. The determination of the vulnerable warning phase is based on the detection of the continuous monotonically increasing trend of the comprehensive index, combined with the conditions that the increase in the local variance relative to the baseline value exceeds a preset proportion, the autocorrelation decay time is prolonged in the relatively stable phase, and the skewness turns from near 0 to a positive value. The preset proportion of the normal baseline rate is determined by statistically analyzing the matching relationship between feeding rate and feeding efficiency in historical stable feeding phases, selecting a rate scaling factor that can maintain feeding activity without causing competitive disorder. The preset proportion of the baseline rate is determined by the effect of different rate reduction amplitudes on the suppression of the comprehensive index during the vulnerable warning phase. Comparative analysis was conducted to select a proportional parameter that could effectively delay the instability trend and reduce drastic behavioral fluctuations. The method for determining the feeding rate reduction ratio was based on historical early warning stage intervention effect statistics, selecting a proportional value that could inhibit the continued rise of the comprehensive index without significantly reducing feeding efficiency. The preset interval duration was determined through statistical modeling and optimization analysis of the relationship between recovery effect and feeding efficiency per unit time under different interval durations. The determination of the vulnerable interruption stage was based on the comprehensive index entering the historical high percentile range and tending to saturate in a short period of time, and the characteristic of the skewness changing from a positive value to a negative value, representing the state transition from upward surge to decline. After outputting the stop feeding command in the stage, the feeding rate was reduced by 10%. By recording the duration of the interruption and continuously monitoring the trend of the comprehensive indicator, when the comprehensive indicator falls back to the preset decline range obtained from the statistics of historical interruption recovery processes and the skewness turns positive again, it is determined that the trial recovery phase has begun. The preset decline range is determined by statistically analyzing the decline trajectory of the comprehensive indicator during historical interruption recovery processes and selecting an range that can stably represent the initial state of recovery. The preset ratio of the feeding start rate in the trial recovery phase is obtained by analyzing historical successful recovery samples. During the recovery process, dynamic switching control between normal feeding and re-interruption is achieved by comparing in real time whether the comprehensive indicator re-enters the low-value stable range or shows an upward trend again.

[0034] Record the occurrence time of each vulnerable interruption phase, the peak value of the comprehensive index during this interruption, and the actual recovery time from stopping feeding to successfully resuming feeding, forming an interruption event record sequence for each aquaculture batch; when a vulnerable interruption event occurs, retrieve the records of all previous interruption events; when the actual recovery time in the records of previous interruption events shows a gradual increasing trend, determine that the fish population's responsiveness to the trial recovery phase is decreasing, and increase the reduction in the rate of slow feeding in the trial recovery phase and the extension of the waiting time for resuming; when the frequency of the vulnerable interruption event exceeds the preset frequency limit within a preset long-term observation window, trigger the instruction to reduce the feeding rate and insert an intermittent when entering the vulnerable warning phase again, and actively intervene before the comprehensive index continuously rises to the original trigger level.

[0035] It should be explained that the interruption event recording sequence is based on breeding batches. Each time a vulnerable interruption phase occurs, the occurrence time, the peak value of the comprehensive index, and the actual recovery time from the cessation of feeding to the resumption of stable feeding are recorded, and the sequence is constructed in chronological order. Trend analysis is performed on the recovery time of each event; when the trend shows a gradual increase, it is determined that the fish population's recovery ability is declining. The slow-start feeding rate ratio is a scaling factor of the initial feeding rate during the recovery phase relative to the normal baseline rate, and its value is determined based on the interval in historical successful recovery samples that can reduce re-instability while maintaining feeding activity. The extension of the waiting time for re-attempting recovery is the increment of the interval between two adjacent recovery attempts, mapped according to the strength of the recovery time growth trend to provide more sufficient recovery time. The rate ratio is simultaneously reduced and the waiting time is extended. The frequency of interruption events is statistically analyzed by setting a preset long-term observation window, which covers the entire feeding fluctuation cycle. The time interval of the period, with the window length determined based on the historical event timescale; the preset frequency upper limit is the upper quantile threshold of the frequency of interruption events under normal aquaculture conditions, used to distinguish between normal fluctuations and abnormally frequent states; the upper quantile threshold is determined by statistically analyzing the historical frequency distribution and selecting quantiles that can distinguish between normal fluctuations and abnormally frequent states; when the frequency exceeds the upper limit, it is determined to enter a high-risk zone, and intervention is carried out in advance in subsequent control; the original trigger level is the standard level for triggering intervention when the comprehensive index continuously rises to the preset trigger benchmark level without frequency constraint adjustment. In the case of high-frequency interruption, the trigger condition is moved forward, so that the intervention is executed in advance when the comprehensive index has not reached the level; the preset trigger benchmark level is determined by statistically analyzing the key turning points of the comprehensive index from stable to unstable processes in the historical vulnerability warning stage, extracting the distribution range of the corresponding values, and selecting the quantile boundary that can stably distinguish between normal fluctuations and unstable trends.

[0036] The above embodiments are provided for those skilled in the art to implement or use the present invention. Those skilled in the art can make various modifications or changes to the above embodiments without departing from the inventive concept of the present invention. Therefore, the protection scope of the present invention is not limited to the above embodiments, but should be the maximum scope that conforms to the innovative features mentioned in the claims.

Claims

1. A feeding control method based on deep learning of spatiotemporal features of fish feeding behavior, characterized in that, Includes the following steps: S1: Continuously acquire video frame sequences of the water surface in the feeding area, and extract spatiotemporal features after preprocessing the video frame sequences to obtain information on the fragmentation of the water splash area and the turbulence intensity gradient. S2: Calculate the instantaneous turbulence flow schedule based on the fragmentation information and the turbulence intensity gradient information. The instantaneous turbulence flow schedule reflects the degree of interference caused by the current splash turbulence on the lateral line perception of the fish school. S3: Combining the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area, calculate a comprehensive index characterizing the severity of the dual vulnerability state of physiological perception within the fish population. S4: Perform statistical behavior analysis on the sequence of the comprehensive index within the sliding time window, and extract the fluctuation characteristics of the comprehensive index near the phase transition critical point; S5: Determine the current phase transition stage based on the fluctuation characteristics and output the corresponding feeding control command. The feeding control command includes maintaining feeding, reducing feeding and inserting an interval, stopping feeding, and resuming feeding.

2. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 1, characterized in that, The continuous acquisition of video frame sequences of the water surface in the feeding area, followed by preprocessing of the video frame sequences to extract spatiotemporal features, yields information on the fragmentation of the splash area and turbulence intensity gradient information. This includes: the fragmentation information being a splash fragmentation segmentation map, which classifies each pixel in the splash area into large, continuous splashes or small, fragmented splashes; the ratio of the area of ​​small splashes to the total splash area in the splash fragmentation segmentation map reflects the uniformity of the fish's feeding actions and the degree of disorder in individual competition; and the turbulence intensity gradient information being a turbulence intensity gradient direction distribution map of the splash area, which characterizes the rate and direction of change in turbulence intensity within a local area, reflecting the spatial pattern of turbulent energy diffusion from the center outwards.

3. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 1, characterized in that, The instantaneous turbulence flow schedule is calculated based on the fragmentation information and the turbulence intensity gradient information. The instantaneous turbulence flow schedule reflects the degree of interference caused by the current splash turbulence on the lateral line perception of the fish. This includes: quantifying the intensity of interference caused by the underwater acoustic disturbance generated by the splash turbulence on the normal functioning of the fish's lateral line sensory perception by constructing the instantaneous turbulence flow schedule; when the fish are feeding in an orderly manner and the turbulence energy is concentrated in the central area of ​​the splash, the spatial non-uniformity of the turbulence is low, the energy concentration is high, and the proportion of fragmented splash is low, the instantaneous turbulence flow schedule is in a low value range, indicating that the lateral line perception of the fish is slightly disturbed; when the fish are feeding in an disorderly manner and the turbulence energy is dispersed, the proportion of fragmented splash increases, the instantaneous turbulence flow schedule increases sharply, indicating that the interference on the lateral line perception of the fish is more severe.

4. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 3, characterized in that, The method of quantifying the interference intensity of underwater acoustic disturbances generated by splash turbulence on the normal functioning of fish's lateral line sensory perception of water flow by constructing the instantaneous turbulence flow scheduling includes: the instantaneous turbulence flow scheduling is calculated according to the following formula: ;in, For the first The instantaneous turbulence scheduling at that moment; For the first The average value of the turbulent gradient amplitude within the water splash area at a given time reflects the spatial non-uniformity of the turbulence; For the first The standard deviation of the turbulence intensity within the splash area at a given time reflects the degree of concentration of turbulent energy; the larger the standard deviation, the more concentrated the energy. For the first The ratio of the area of ​​fine water splashes to the total water splash area at any given time; the lower the proportion of fine water splashes, the better. The closer it is to 0, the more its complement term... The closer to 1; The preset minimum positive number is used to prevent division by zero errors when the complement term is 0; when the fish feed in an orderly manner, the molecules... Smaller denominator Larger and complementary terms close to 1, making At low values; when turbulence is diffuse and feeding is disordered, the numerator increases while the denominator is low. Decrease the complement term, so that It increased dramatically.

5. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 1, characterized in that, The method combines the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area to calculate a comprehensive index characterizing the severity of the dual vulnerability state of physiological perception within the fish population. This includes: recording the continuous duration for which the splash area has remained above a preset high water level threshold since the end of the most recent event of a significant decrease in splash area, as the steady-state duration; calculating the theoretical safe recovery period based on the instantaneous turbulent flow scheduling. The theoretical safe recovery period represents the shortest time that the fish population needs to suspend feeding to breathe and recover its senses under the current degree of turbulent mismatch, thereby reducing the continuous accumulation of the dual vulnerability state of physiological perception. This is obtained by performing an exponential mapping of the instantaneous turbulent flow scheduling with a natural constant as the base, with the aim of nonlinearly mapping the high value range of mismatch to a longer recovery time requirement. The theoretical safe recovery period shows an accelerating growth trend with the increase of the instantaneous turbulent flow scheduling, reflecting the physiological law that the recovery time required by the fish population is disproportionately extended when the degree of turbulent mismatch is higher.

6. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 5, characterized in that, The calculation of the theoretical safe recovery period based on the instantaneous turbulence flow scheduling includes: the formula for calculating the theoretical safe recovery period is: ;in, For the first The theoretical safe recovery period at any given time; The base recovery time represents the minimum recovery time required for the fish swarm to recover when the instantaneous turbulent flow scheduling approaches 0. For the natural constant An exponential function with base 0; The growth rate adjustment coefficient controls the growth rate of the theoretical safe recovery period as the instantaneous turbulent flow scheduling changes. For the first The instantaneous turbulence scheduling at that moment; The reference mismatch constant is used to normalize the instantaneous turbulent flow scheduling to a value similar to... Matching scale range; when Less than At that time, the theoretical safe recovery period increases at a slow rate. Exceed Subsequently, the theoretical safe recovery period increases rapidly due to the exponential effect, causing the theoretical recovery time required for the fish population under high turbulence and mismatch conditions to exceed the basic recovery time.

7. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 1, characterized in that, The comprehensive index, which combines the instantaneous turbulent flow scheduling with the current steady-state duration of maintaining a high splash area, is used to calculate the severity of the dual vulnerability state of physiological perception within the fish population. The comprehensive index is calculated according to the following formula: ;in, For the first The comprehensive indicators mentioned at any given time; For the first The instantaneous turbulence scheduling at that moment; As of the date The duration of the steady state recorded at each moment; For the first The theoretical safe recovery period at any given time; The saturation rate adjustment coefficient controls the degree to which the ratio of the steady-state duration to the theoretical safe recovery period affects the saturation rate of the second factor. The first factor is a preset, extremely small positive number used to prevent division by zero errors when the theoretical safe recovery period is 0; The instantaneous turbulent flow schedule increases monotonically with a gradually decreasing rate of increase, and a compression transformation with a continuously decreasing rate of increase is applied to the unbounded instantaneous turbulent flow schedule; the second factor in the formula It is an exponential saturation function, which approaches 0 when the steady-state duration is much shorter than the theoretical safe recovery period, and approaches 1 when the steady-state duration is close to or exceeds the theoretical safe recovery period; the product of the first factor and the second factor makes the comprehensive index significantly increase only when both the instantaneous turbulence outflow is high and the steady-state duration exceeds the theoretical safe recovery period are met simultaneously. When one condition is met but the other condition is missing, the comprehensive index is suppressed to a low value range.

8. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 1, characterized in that, The statistical behavior analysis of the comprehensive index sequence within the sliding time window, extracting the fluctuation characteristics of the comprehensive index near the phase transition critical point, includes: calculating three indicators—local variance, skewness, and autocorrelation decay time—of the comprehensive index sequence within a preset length sliding time window; the local variance is the variance of the comprehensive index values ​​within the sliding time window, used to measure the intensity of fluctuation of the comprehensive index within the window; when the internal state of the fish population tends to be unstable, the local variance increases significantly compared to the stable period; the skewness is the variance of the three comprehensive index values ​​within the sliding time window. The standard moment of the first order is used to measure the asymmetry of the distribution of the comprehensive index. When the skewness is positive, it indicates that the comprehensive index is frequently testing higher levels. When the skewness is negative, it indicates that the comprehensive index is in the process of continuously falling back from a high level. The autocorrelation decay time is the time required for the autocorrelation function value of the comprehensive index to decay from the peak to a preset proportion. It is used to measure the strength of the time dependence of the comprehensive index sequence. When the comprehensive index approaches the phase transition critical point, the time inertia of the comprehensive index increases, and the autocorrelation decay time becomes abnormally prolonged, constituting an early warning signal of the phase transition critical point.

9. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 1, characterized in that, The system determines the current phase transition stage based on the fluctuation characteristics and outputs corresponding feeding control commands. These commands include maintaining feeding, reducing feeding and inserting intervals, stopping feeding, and resuming feeding. Specifically: when the comprehensive index is in a low range, the local variance is minimal, and the skewness fluctuates around 0; and the autocorrelation decay time is short, it is determined that the fish are currently in a normal feeding phase, with orderly feeding and clear perception, and an instruction to maintain the current feeding rate is output; when the comprehensive index begins to rise continuously, the local variance increases compared to the normal feeding phase, and the autocorrelation decay time prolongs; and when the skewness turns positive, it is determined that the fish are currently in a vulnerable warning phase, where the statistical behavior of the comprehensive index indicates that the fish are approaching a critical point of physiological and sensory vulnerability, and an instruction to reduce the feeding rate to a preset ratio of the baseline rate and forcibly insert intervals of a preset duration within a unit of time is output to proactively create ventilation and sensory recovery opportunities for the fish. The time window is defined as follows: When the comprehensive index jumps significantly to a high level and approaches saturation, and the skewness changes from positive to negative, it is determined that the current stage is a fragile interruption stage. The fish have been forcibly stopped feeding due to the combined effects of oxygen debt accumulation and sensory inhibition. An instruction to immediately stop feeding is output, and the duration of the interruption is recorded. It is determined that the current stage is a trial recovery stage, and a trial recovery is carried out by slowly starting feeding and continuously monitoring the comprehensive index. After stopping feeding, when the comprehensive index falls from a high level to a preset decline range and the skewness changes from negative to positive again, it is determined that the current stage is a trial recovery stage. An instruction to slowly start feeding at a preset ratio lower than the normal baseline rate is output, while continuously monitoring the trend of the comprehensive index. When the comprehensive index returns to a low value and stabilizes, the normal feeding stage is resumed. When the comprehensive index rises again during the trial feeding process, a stop feeding instruction is output again, and the waiting time before attempting recovery is extended.

10. The feeding control method based on deep learning of spatiotemporal features of fish feeding behavior according to claim 9, characterized in that, The determination of whether the current stage has entered a trial recovery phase involves a slow initiation of feeding and continuous monitoring of the comprehensive indicators. This includes: recording the occurrence time of each vulnerable interruption phase, the peak value of the comprehensive indicators during the interruption, and the actual recovery time from the cessation of feeding to the successful resumption of feeding, forming an interruption event record sequence for the aquaculture batch; retrieving the records of previous interruption events when a vulnerable interruption event occurs; determining that the fish population's responsiveness to the trial recovery phase is decreasing when the actual recovery time in the records of previous interruption events shows a gradual increasing trend, and increasing the reduction in the rate of slow initiation of feeding and the extension of the waiting time for resuming the trial recovery phase; the gradual increasing trend is determined by trend fitting of the recovery time sequence of multiple consecutive interruption events, with a continuous increase in the fitting result as the criterion; when the frequency of the vulnerable interruption event exceeds the preset frequency upper limit within a preset long-term observation window, the instruction to reduce the feeding rate and insert an intermittent is triggered in advance when the next vulnerable warning phase begins, and proactive intervention is initiated before the comprehensive indicators continuously rise to the original trigger level.

Images