Multi-factor fusion-based bullfrog breeding water quality early warning and intelligent feeding system

Abstract

The application discloses a bullfrog breeding water quality early warning and intelligent feeding system based on multi-factor fusion, relates to the technical field of intelligent aquaculture, and solves the technical problems that in the existing bullfrog breeding, water quality monitoring is difficult to exclude frog activity interference, feeding monitoring is difficult to handle individual overlap, and early warning and feeding lack causal alignment bidirectional linkage, comprising: removing bullfrog activity false readings through a transient disturbance filtering module, realizing identity re-identification through a frog group overlap segmentation module to accurately count feeding intensity, calculating lag steps and aligning sequences through a cross-modal causal time lag discovery module, fusing water quality and appetite information through bidirectional cross-attention and dynamic gating, generating a joint instruction of feeding speed regulation, protein reduction switching and oxygen enrichment guidance in a first fusion state, and generating an implicit water quality mutation early warning signal based on a feeding attenuation gradient in a second fusion state. The application realizes multi-factor deep fusion of water quality early warning and intelligent feeding in bullfrog breeding.

Description

Technical Field

[0001] This invention belongs to the field of intelligent aquaculture technology, specifically involving a water quality early warning and intelligent feeding system for bullfrog farming based on multi-factor fusion. Background Technology

[0002] In bullfrog farming, water quality management is a crucial factor in determining the success or failure of the farming. Under high-density farming conditions, uneaten feed and excrement accumulate rapidly, and increased ammonia nitrogen concentration and decreased dissolved oxygen are common forms of water quality deterioration. Therefore, water quality deterioration directly affects the feeding behavior and health of bullfrogs, and in severe cases, it can lead to large-scale mortality.

[0003] In water quality monitoring, existing online water quality monitoring systems typically collect physicochemical indicators such as ammonia nitrogen, dissolved oxygen, and pH at fixed frequencies. However, bullfrogs are highly mobile amphibians, and their swarming feeding and tail-wagging movements can create instantaneous water quality disturbances around the sensors, causing a false increase in ammonia nitrogen concentration and turbidity. These false readings caused by biological activity are difficult to distinguish from genuine water quality anomalies in a single sampling, easily leading to frequent false alarms and reducing the accuracy of water quality trend factors. Regarding feeding behavior monitoring, bullfrogs exhibit swarming feeding habits, with multiple individuals highly clustered near the feed distribution point, frequently overlapping each other. Existing vision-based feeding behavior analysis techniques typically rely on target detection and tracking algorithms to count the number of feeding individuals. When bullfrogs... When there is significant overlap in feeding activity, the algorithm is prone to losing the target identity or confusing different individuals, leading to a large deviation between the statistical results of foraging intensity and the actual feeding situation. In terms of the linkage between early warning and feeding, the existing water quality early warning system and the automatic feeding system usually operate independently. The water quality early warning system issues an alarm when the water quality index exceeds the standard, while the feeding system executes feeding according to preset times or preset intervals. There is a lack of dynamic coordination between the two based on multi-factor fusion. In addition, when the existing system faces a slight decline in water quality but the frogs still have a strong need to feed, it lacks a refined compensation feeding strategy. Once an abnormal water quality is detected, the system usually stops feeding directly or reduces the amount of feed significantly. This one-size-fits-all approach not only wastes feed, but also causes stress and cannibalism among the frogs due to the sudden cessation of feeding.

[0004] Therefore, a water quality early warning and intelligent feeding system for bullfrog farming based on multi-factor fusion has emerged. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a water quality early warning and intelligent feeding system for bullfrog farming based on multi-factor fusion, which is used to solve the technical problem of how to achieve a dynamic balance between protecting water quality and meeting feeding needs, and to achieve a smooth transition of feed ratio in feeding regulation.

[0006] To address the above problems, this invention provides a multi-factor fusion-based water quality early warning and intelligent feeding system for bullfrog farming, comprising: A water quality sensing array is used to acquire a water quality trend factor sequence at a first sampling frequency, the water quality trend factor sequence including the ammonia nitrogen change rate and the dissolved oxygen change rate; The feeding behavior analysis unit is used to acquire an appetite dynamic factor sequence at a second sampling frequency, the appetite dynamic factor sequence including the peak value of cluster feeding intensity and the feeding decay gradient; The multi-factor fusion decision engine is connected to the water quality sensing array and the feeding behavior analysis unit respectively, and runs a water quality-appetite bidirectional compensation model internally. The water quality-appetite bidirectional compensation model: In the first fusion state, in response to the water quality trend factor sequence pointing to a downward trend in dissolved oxygen and the peak value of the cluster foraging intensity > a first threshold, a joint instruction is generated, consisting of a feeding rate reduction instruction, a low-protein feed ratio switching instruction, and a directional feeding control signal for oxygen-rich areas; In the second fusion state, the foraging attenuation gradient is used as the trigger variable for water quality early warning. When the foraging attenuation gradient exceeds the non-satiation stress attenuation benchmark value within a preset time window after feeding is initiated, a latent water quality mutation early warning signal based on biological behavior is generated. The mixed feeding actuator is used to receive and execute the feeding rate reduction command, the low protein feed ratio switching command, and the oxygen-enriched area directional feeding control signal generated in the first fusion state.

[0007] Preferably, the water quality sensing array includes an ion-selective electrode array, a miniature underwater spectral probe, and a transient disturbance filtering module electrically connected to the ion-selective electrode array and the miniature underwater spectral probe, respectively. The first sampling frequency is a dynamically adjusted frequency based on event triggering; The transient disturbance filtering module is used to trigger a high-speed pulse sampling sequence when the ion-selective electrode array or the micro underwater spectral probe detects an ammonia nitrogen concentration or turbidity greater than a preset disturbance trigger threshold in a single sampling; when the high-speed pulse sampling sequence is less than a complete bullfrog feeding or tail-wagging action cycle, it is used to capture the complete pulse waveform of water quality trend factors in disturbance events. The transient disturbance filtering module is also used to extract features from the complete pulse waveform. If its peak half-width and waveform attenuation shape meet the morphological criteria for bullfrog activity disturbance, the data segment corresponding to the pulse waveform is removed from the water quality trend factor sequence input to the multi-factor fusion decision engine.

[0008] Preferably, the feeding behavior analysis unit includes an underwater binocular vision module, a skeletal point behavior analysis module, and a frog swarm overlapping individual segmentation module connected to the skeletal point behavior analysis module; The underwater binocular vision module is used to simultaneously acquire individual bullfrog image pairs in the feeding area; The skeletal point behavior analysis module is used to extract the skeletal point coordinates of individual bullfrogs based on the image, and to track the skeletal point trajectory of each bullfrog in consecutive frames. The frog swarm overlapping individual segmentation module is used to re-identify each bullfrog after the intersection when the skeletal point trajectories of at least two bullfrogs intersect in space and the number of frames of the overlap exceeds the overlap threshold, based on the snout morphological features and body surface pattern features of each bullfrog before the trajectory intersection. Among them, the peak of the cluster foraging intensity is determined by the number of independent bullfrog individuals whose snout vertical positive displacement exceeds the predation threshold after identity re-identification within a unit of time; the foraging attenuation gradient is the attenuation slope of the peak of the cluster foraging intensity within a preset time window.

[0009] Preferably, in the multi-factor fusion decision engine, the water quality-appetite bidirectional compensation model includes a cascaded cross-modal causal time-delay discovery module, a temporal feature encoder, and a dual-head prediction decoder: The cross-modal causal time-delay detection module, before concatenating the water quality trend factor sequence and the appetite dynamic factor sequence, takes the ammonia nitrogen change rate sequence within the first preset time period before the start of the current feeding, and records it as... And the food decay gradient sequence within the second preset time period after the start of feeding, denoted as ,in, and These are the lengths of the ammonia nitrogen change rate sequence and the feeding decay gradient sequence, respectively. The feeding attenuation gradient sequence Fixed, ammonia nitrogen change rate sequence Slide successively in the negative direction of the time axis with a step size of 1, for the offset... ,in, Take the preset maximum offset. Tail subsequence and Equal-length subsequences ,in, The overlap length between the two subsequences; The offset is calculated using the Pearson correlation coefficient formula. Cross-correlation strength .

[0010] Preferably, the cross-correlation strength With offset In the cross-correlation curve formed by the changes, Offset corresponding to the maximum value The real-time causal lag step number is determined within the current feeding cycle; based on the real-time causal lag step number... Subtract the timestamp of each data point in the water quality trend factor sequence. A time step was used to align the translated water quality trend factor sequence with the appetite dynamic factor sequence in causal time.

[0011] Preferably, the dual-head predictive decoder includes a bidirectional cross-attention layer, a dynamic gating unit, a water quality prediction head, and an appetite decline prediction head. The bidirectional cross-attention layer includes a first cross-attention sub-layer and a second cross-attention sub-layer. The first cross-attention sub-layer uses the water quality trend factor sequence after causal temporal alignment as the query matrix and the appetite dynamic factor sequence as the key matrix and value matrix to output the appetite context sequence of water quality perception. The second cross-attention sub-layer uses the appetite dynamic factor sequence as the query matrix and the water quality trend factor sequence after causal temporal alignment as the key matrix and value matrix to output the water quality context sequence of appetite perception.

[0012] Preferably, the dynamic gating unit is connected to the transient disturbance filtering module, and obtains two statistical values ​​for the current feeding cycle from the transient disturbance filtering module. and Wherein, is the number of pulse waveform data segments identified as bullfrog activity interference and removed during the feeding cycle. The total number of times the high-speed pulse sampling sequence is triggered within the feeding cycle; the dynamic gating unit calculates the gating coefficient. , among which, when hour, ;when hour, ; Multiply the appetite context sequence perceived by water quality Multiply the water quality context sequence of appetite perception by The two are added together and then concatenated with the aligned water quality trend factor sequence, and then input into the temporal feature encoder; The fused temporal feature vector output by the temporal feature encoder is input to the water quality prediction head and the appetite decline prediction head, respectively.

[0013] Preferably, the combined control command for generating the feeding rate reduction command, the low-protein feed ratio switching command, and the oxygen-enriched area directional feeding control signal includes: The joint control command is used to execute the feeding trajectory, specifically: to perform the first feeding at the current gathering location of the frog swarm identified by the feeding behavior analysis unit, and then to feed at preset time intervals. The mobile chassis will be moved a preset distance toward the oxygen-rich area. The angle of the variable speed throwing disc is simultaneously adjusted to point to the water surface area within a preset radius centered on the current position of the mobile chassis, until the mobile chassis reaches the oxygen-rich area. Each movement is followed by a feeding, with the amount of food given increasing sequentially along the direction of movement. ; The oxygen-rich area is identified by a water quality sensing array and is defined as a local water area where the dissolved oxygen saturation is greater than the second threshold.

[0014] Preferably, the mixing and feeding actuator includes a dual hopper, a twin-screw feeder, a mixing chamber, and a variable-speed throwing disc; The dual silos store conventional protein feed and low-protein feed respectively; The low-protein feed formulation switching command specifies a protein content decreasing curve. ,in, To switch the real-time feed protein content at the start time. The preset deceleration rate, This is the time counted from the start of the switchover; At each time step, according to Calculate the target protein content at the current time step, and based on the target protein content, back-calculate the rotational speed ratio between the conventional protein feed conveyor and the low-protein feed conveyor in the double-helix feeder. Drive the two feeders according to the stated rotational speed ratio, so that the protein content of the feed output from the mixing chamber decreases along the stated curve from... The protein content is continuously reduced until it reaches the preset target low protein content value.

[0015] The beneficial effects of this invention are: This invention captures the complete pulse waveform caused by bullfrogs pouncing or wagging their tails through a transient disturbance filtering module. It distinguishes between biological disturbances and real water quality anomalies by using peak half-width and waveform attenuation morphology and eliminates false readings. Combined with a frog swarm overlapping individual segmentation module, it performs identity re-identification based on snout morphology and body surface patterns when individuals overlap, thereby improving the accuracy of water quality trend factors and appetite dynamic factors from the data source. This invention uses a cross-modal causal time delay discovery module to calculate the optimal lag step between the ammonia nitrogen change rate and the feeding decline gradient in real time using sliding cross-correlation. After causal time alignment of the two sequences, they are then fused by a bidirectional cross-attention layer and a dynamic gating unit. The gating coefficient is derived from the interference statistics of the transient disturbance filtering module, so that the fusion weight is adaptively adjusted according to the degree of interference of the water quality data, thus achieving robust fusion of water quality and appetite information. This invention guides the frog population to migrate to an oxygen-rich area by smoothly switching the low-protein feed ratio along a decreasing curve in the first fusion state and using an incremental feeding trajectory. Combined with the implicit water quality change warning based on the feeding decline gradient in the second fusion state, it not only maintains the precise satisfaction of feeding needs when the water quality declines, but also achieves early warning before the physicochemical indicators are obviously exceeded. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1 As shown, this invention is a multi-factor fusion-based water quality early warning and intelligent feeding system for bullfrog farming, comprising: A water quality sensing array is used to acquire a water quality trend factor sequence at a first sampling frequency, the water quality trend factor sequence including the ammonia nitrogen change rate and the dissolved oxygen change rate; The feeding behavior analysis unit is used to acquire the appetite dynamic factor sequence at a second sampling frequency. The appetite dynamic factor sequence includes the peak value of cluster feeding intensity and the feeding decay gradient. The multi-factor fusion decision engine is connected to the water quality sensing array and the feeding behavior analysis unit respectively, and runs a water quality-appetite bidirectional compensation model internally. The water quality-appetite bidirectional compensation model: In the first fusion state, in response to the water quality trend factor sequence pointing to a downward trend in dissolved oxygen and the peak value of the cluster foraging intensity > a first threshold, a joint instruction is generated, consisting of a feeding rate reduction instruction, a low-protein feed ratio switching instruction, and a directional feeding control signal for oxygen-rich areas; In the second fusion state, the foraging attenuation gradient is used as the trigger variable for water quality early warning. When the foraging attenuation gradient exceeds the non-satiation stress attenuation benchmark value within a preset time window after feeding is initiated, a latent water quality mutation early warning signal based on biological behavior is generated. The mixed feeding actuator is used to receive and execute the feeding rate reduction command, the low protein feed ratio switching command, and the oxygen-enriched area directional feeding control signal generated in the first fusion state.

[0019] In one embodiment of the present invention, the water quality sensing array includes an ion-selective electrode array, a miniature underwater spectral probe, and a transient disturbance filtering module electrically connected to the ion-selective electrode array and the miniature underwater spectral probe, respectively. The first sampling frequency is a dynamically adjusted frequency based on event triggering; The transient disturbance filtering module is used to trigger a high-speed pulse sampling sequence when the ion-selective electrode array or the micro underwater spectral probe detects an ammonia nitrogen concentration or turbidity greater than a preset disturbance trigger threshold in a single sampling; when the high-speed pulse sampling sequence is less than a complete bullfrog feeding or tail-wagging action cycle, it is used to capture the complete pulse waveform of water quality trend factors in disturbance events. The transient disturbance filtering module is also used to extract features from the complete pulse waveform. If its peak half-width and waveform attenuation shape meet the morphological criteria for bullfrog activity disturbance, the data segment corresponding to the pulse waveform is removed from the water quality trend factor sequence input to the multi-factor fusion decision engine.

[0020] Specifically, in the bullfrog breeding pond, at least one deployment point is selected in each of the feeding area, resting shallow area, and concealed vegetation area. At each deployment point, a set of ion-selective electrode arrays and miniature underwater spectrometers are fixedly installed at a preset depth below the water surface. The preset depth is the average depth at which the bullfrog's torso is submerged in water during normal activity. The average depth is determined by taking a side view image of the bullfrog resting normally in the pond, extracting the torso submersion depth, and taking the average value. The ion-selective sensitive membrane of the ion-selective electrode array faces the open direction of the water body, and the optical window of the miniature underwater spectrometer faces the unobstructed water area. The distance between the two is preset so that the water volume detected by the two overlaps but does not completely coincide, to ensure that the same disturbance event can be captured by at least one sensor. The sensor group at each deployment point is electrically connected to the transient disturbance filtering module in the onshore data processing unit through a waterproof cable. The sensors continuously collect water quality data at a reference sampling frequency and calculate the ammonia nitrogen change rate and dissolved oxygen change rate in real time, forming the basic data stream of the water quality trend factor sequence. The transient disturbance filtering module monitors the results of each single sampling in real time. When the ammonia nitrogen concentration detected by the ion-selective electrode array exceeds a preset disturbance trigger threshold, or the turbidity detected by the miniature underwater spectral probe exceeds a preset disturbance trigger threshold, the transient disturbance filtering module determines that a suspected disturbance event has occurred. The preset disturbance trigger threshold is determined by recording the ammonia nitrogen concentration and turbidity readings of the sensor at a calm water surface during several consecutive periods without feeding or human interference in the bullfrog breeding pond, and taking the maximum value of the readings during each period as the baseline fluctuation limit; multiplying the baseline fluctuation limit by a preset multiple results in the disturbance trigger threshold. Value; After a suspected disturbance event is detected, the transient disturbance filtering module immediately triggers a high-speed pulse sampling sequence. The sampling frequency of the high-speed pulse sampling sequence is higher than the reference sampling frequency, and its duration is preset to be less than a complete bullfrog feeding or tail-wagging action cycle. The bullfrog feeding action cycle is obtained by recording the complete process of the bullfrog from stillness to feeding and then back to stillness multiple times through underwater camera recording, and taking the average of the recording durations as the feeding action cycle; the tail-wagging action cycle is obtained in the same way; the high-speed pulse sampling sequence samples intensively within a duration less than this action cycle to ensure that the complete pulse waveform of the disturbance event is captured. The transient disturbance filtering module extracts features from the complete pulse waveform captured by the high-speed pulse sampling sequence. The extracted features include peak half-width (HWHM) and waveform decay morphology. HWHM refers to the time span during which the pulse amplitude rises from half its peak value to the peak value and then falls back to half its peak value. Waveform decay morphology refers to the curve shape of the amplitude change over time as the pulse falls from the peak value back to the baseline. The extracted HWHM and waveform decay morphology are matched with preset bullfrog activity disturbance morphological criteria. The bullfrog activity disturbance morphological criteria are obtained by artificially simulating bullfrog feeding and tail-wagging movements in the breeding pond, creating typical water quality disturbances around the sensor, and collecting multiple sets of pulse waveform samples from known sources using the high-speed pulse sampling sequence. HWHM and waveform decay morphology features are extracted from these samples, and after statistical clustering, the HWHM range and decay morphology feature parameters corresponding to bullfrog activity disturbances are determined and stored as morphological criteria in the transient disturbance filtering module. Simultaneously, sewage discharge operations are collected. Water quality pulse waveform samples generated by non-bullfly activities such as the start-up and shutdown of aeration equipment are extracted using the same features as a distinguishing reference. If the peak half-width of the current pulse waveform falls within the peak half-width range corresponding to bullfrog activity interference, and the matching degree between the waveform attenuation pattern and the attenuation pattern feature parameter of bullfrog activity interference exceeds a preset matching threshold (wherein, the preset matching threshold is determined by using the bullfrog activity interference pulse waveform samples from the above-mentioned known sources as the positive sample set, and the water quality pulse waveform samples generated by sewage discharge operations and the start-up and shutdown of aeration equipment as the negative sample set; calculating the matching degree between each sample in the positive sample set and the morphological criterion center value of bullfrog activity interference, and the matching degree between each sample in the negative sample set and the morphological criterion center value of bullfrog activity interference; selecting a matching degree value such that the proportion of samples with a matching degree higher than this value in the positive sample set is the largest, and the proportion of samples with a matching degree higher than this value in the negative sample set is the smallest, and this matching degree value is used as the preset matching threshold), then the pulse waveform is determined to meet the morphological criteria of bullfrog activity interference. For pulse waveforms that are determined to meet the morphological criteria for bullfrog activity disturbance, the transient perturbation filtering module removes the data segment corresponding to the pulse waveform from the water quality trend factor sequence that is about to be input into the multi-factor fusion decision engine. The water quality trend factor sequence after removal retains the data generated by the actual water quality changes and filters out the transient pseudo-readings introduced by bullfrog activity. The water quality trend factor sequence processed by the transient perturbation filtering module is output to the multi-factor fusion decision engine in a time-series format corresponding to the reference sampling frequency.

[0021] In one embodiment of the present invention, the feeding behavior analysis unit includes an underwater binocular vision module, a skeletal point behavior analysis module, and a frog swarm overlapping individual segmentation module connected to the skeletal point behavior analysis module. The underwater binocular vision module is used to simultaneously acquire individual bullfrog image pairs in the feeding area; The skeletal point behavior analysis module is used to extract the skeletal point coordinates of individual bullfrogs based on the image, and to track the skeletal point trajectory of each bullfrog in consecutive frames. The frog swarm overlapping individual segmentation module is used to re-identify each bullfrog after the intersection when the skeletal point trajectories of at least two bullfrogs intersect in space and the number of frames of the overlap exceeds the overlap threshold, based on the snout morphological features and body surface pattern features of each bullfrog before the trajectory intersection. Among them, the peak of the cluster foraging intensity is determined by the number of independent bullfrog individuals whose snout vertical positive displacement exceeds the predation threshold after identity re-identification within a unit of time; the foraging attenuation gradient is the attenuation slope of the peak of the cluster foraging intensity within a preset time window.

[0022] Specifically, at least one underwater binocular vision module is deployed below the water surface in the feeding area of ​​the aquaculture pond. Each underwater binocular vision module consists of two cameras calibrated for underwater distortion. The two cameras are installed parallel to each other within a waterproof housing at a preset baseline distance. The preset baseline distance is determined by placing a known-sized calibration object in the feeding area, adjusting the distance between the two cameras, capturing images of the calibration object, calculating the 3D reconstruction error of the calibration object at different distances, and selecting the distance that minimizes the 3D reconstruction error as the preset baseline distance. The waterproof housing is fixedly installed on the side of the feeding area near the pond wall. The optical axes of the two cameras are perpendicular to the pond wall and parallel to the water surface, covering the main water surface area of ​​the feeding area. The field of view coverage is determined by... Buoys are deployed at the water surface boundary of the feeding area. It is checked whether the buoys are all within the edge of the frame in the images captured synchronously by the two cameras. If any buoys are not in the frame, the installation height or pitch angle of the waterproof housing is adjusted until all buoys are in the frame. After the feeding is started, the binocular vision module continuously and synchronously acquires individual bullfrog image pairs in the feeding area at the second sampling frequency. The second sampling frequency is obtained by capturing continuous images of the bullfrog completing one complete feeding action, and calculating the average number of frames required from fixation to swallowing. The lowest frequency in which at least a preset number of frames are acquired within the time interval corresponding to the average number of frames is taken as the second sampling frequency. Each image pair includes a left eye image and a right eye image. The timestamps of the two images are strictly aligned by a hardware synchronization trigger signal. The skeletal point behavior analysis module receives image pairs from the underwater binocular vision module. For each frame of the left eye image, it extracts two-dimensional coordinates using a pre-trained bullfrog skeletal point detection network. This skeletal point detection network employs a high-resolution network as its backbone feature extraction network, while maintaining high-resolution subnetworks in parallel to preserve spatial location information. Its output is connected to a fully connected layer and outputs data representing the snout, forelimb joints, and hindlimb joints. Heatmaps of key points are used to obtain the coordinates of two-dimensional skeletal points by taking the peak values ​​of the heatmaps. The training method of the bullfrog skeletal point detection network is as follows: underwater images of bullfrogs under different lighting conditions and different frog population densities are collected, and the two-dimensional coordinates of three skeletal points, namely the snout, forelimb joints, and hindlimb joints, are manually labeled. The labeled images are used as training samples, and the mean squared error is used as the loss function for supervised training. For each pair of left and right eye images, the skeletal point behavior analysis module performs stereo matching on the same bullfrog in the left and right eye images. Based on the principle of binocular parallax, the three-dimensional spatial coordinates of each skeletal point are calculated. In continuous frames, the skeletal point behavior analysis module tracks the three-dimensional coordinates of the skeletal points of each individual bullfrog frame by frame through Kalman filtering to form the three-dimensional motion trajectory of the skeletal points of each bullfrog. The frog swarm overlapping individual segmentation module is connected to the skeletal point behavior analysis module. It receives the 3D motion trajectory of each bullfrog's skeletal point. In each frame, the frog swarm overlapping individual segmentation module calculates the minimum Euclidean distance in 3D space between the skeletal point trajectories of any two bullfrogs. When the minimum Euclidean distance between the skeletal point trajectories of at least two bullfrogs is less than a preset crossover determination distance in multiple consecutive frames, and the number of frames of overlap exceeds an overlap threshold, it is determined that the two bullfrogs have crossed trajectories and entered an overlapping state. The preset crossover determination distance is determined by collecting multiple side-view video clips of two bullfrogs moving independently and then making physical contact during a bullfrog swarm's feeding process. In each video clip, the Euclidean distance between the skeletal points of the two bullfrogs is labeled frame by frame at the moment their torsos make contact. The average of all labeled distances is taken as the preset crossover determination distance. The overlap threshold is determined by recording the number of frames in the same batch of labeled video clips from the point where the Euclidean distance between the skeletal points of the two bullfrogs is less than the crossover determination distance until the Euclidean distance becomes greater than the crossover determination distance again. The average of all recorded frame numbers is taken as the preset crossover determination distance. The mean value is used as the overlap threshold. When the overlap state is determined, the frog group overlapping individual segmentation module traces back to the frames before the trajectory intersection, and extracts the snout region and trunk surface region from each bullfrog image extracted by the skeletal point behavior analysis module. For the snout region, the contour morphological features of the snout are extracted, including the ratio of snout width to length and the snout edge curvature. For the trunk surface region, the surface pattern features are extracted, including the main color distribution and spot density of the pattern. The snout morphological features and surface pattern features are combined into the identity feature vector of each bullfrog. The snout morphology features and body surface pattern features are combined to form the identity feature vector of each bullfrog. After the overlap ends and the skeletal point trajectories are separated again, the frog group overlapping individual segmentation module re-extracts the snout morphology features and body surface pattern features of each bullfrog after the crossover ends, calculates the cosine similarity between the snout morphology features and the identity feature vectors of each bullfrog before the crossover, and associates each bullfrog after the crossover ends with the bullfrog with the highest cosine similarity before the crossover to complete the identity re-identification. After identity re-identification, the skeletal point trajectory of each bullfrog maintains the same identity before and after the overlap. Based on the completed identity re-identification, the skeletal point behavior analysis module monitors the vertical displacement of the snout skeletal points of each individual bullfrog frame by frame. When the vertical positive displacement of a bullfrog's snout exceeds the predation threshold, it is determined that the bullfrog has completed a predation action. The predation threshold is determined by collecting the vertical displacement of the snout of each bullfrog during predation actions in multiple segments of normal feeding, as well as the vertical displacement of the snout in multiple segments of non-feeding states (such as natural swimming and surfacing for air). The probability density distribution of the vertical displacement in feeding and non-feeding states is calculated separately. A displacement value is selected such that the probability of the displacement being greater than the value in feeding state is equal to the probability of the displacement being less than the value in feeding state. The difference is maximized, and the displacement value is used as the predation threshold. The peak of the cluster foraging intensity is determined by counting the number of independent bullfrog individuals whose snout vertical displacement exceeds the predation threshold after identity re-identification within a preset time window after feeding begins, with a statistical period of one unit time. The maximum value of the statistical count within each unit time within the entire preset time window is taken as the peak of the cluster foraging intensity. The foraging attenuation gradient is determined by calculating the linear fitting slope of the cluster foraging intensity value changing with time within the attenuation observation window, with the occurrence time of the peak of the cluster foraging intensity as the starting point and the preset time window after that time as the attenuation observation window. The absolute value of this slope is the foraging attenuation gradient.

[0023] In one embodiment of the present invention, the multi-factor fusion decision engine includes a water quality-appetite bidirectional compensation model comprising a cascaded cross-modal causal time-delay discovery module, a temporal feature encoder, and a dual-head prediction decoder. The cross-modal causal time-delay detection module, before concatenating the water quality trend factor sequence and the appetite dynamic factor sequence, takes the ammonia nitrogen change rate sequence within the first preset time period before the start of the current feeding, and records it as... And the food decay gradient sequence within the second preset time period after the start of feeding, denoted as ,in, and These are the lengths of the ammonia nitrogen change rate sequence and the feeding decay gradient sequence, respectively. The feeding attenuation gradient sequence Fixed, ammonia nitrogen change rate sequence Slide successively in the negative direction of the time axis with a step size of 1, for the offset... ,in, Take the preset maximum offset. Tail subsequence and Equal-length subsequences ,in, The overlap length between the two subsequences; The offset is calculated using the Pearson correlation coefficient formula. Cross-correlation strength .

[0024] Specifically, after each feeding is initiated, the cross-modal causal time-delay detection module obtains the ammonia nitrogen change rate sequence within the first preset time period before the current feeding starts from the water quality sensing array, denoted as... ,in, This represents the ammonia nitrogen change rate value corresponding to the start time of the first preset duration. This represents the rate of change in ammonia nitrogen at the start of feeding. The number of sampling points for the ammonia nitrogen change rate within a first preset time period, wherein the first preset time period is determined by recording the average time taken for the ammonia nitrogen concentration to rise from the start to the peak value during the breeding cycle, and taking a preset multiple of this average time period as the first preset time period; simultaneously, the feeding attenuation gradient sequence within a second preset time period after the start of the current feeding is obtained from the feeding behavior analysis unit, denoted as ,in, This represents the food decay gradient value during the first statistical period after feeding begins. This represents the food decay gradient value at the end of the second preset duration. The number of sampling points for the feeding attenuation gradient within the second preset time period, wherein the second preset time period is the average time taken for bullfrogs from the start of feeding to the complete cessation of feeding during the breeding cycle, and this average time period is taken as the second preset time period; preset maximum offset. By recording the time interval from the initial rise in ammonia nitrogen concentration to the point where bullfrog feeding behavior begins to decline during the breeding cycle, the criteria for determining a decline in bullfrog feeding behavior are as follows: using the peak feeding intensity after the start of feeding as a benchmark, when the feeding intensity value drops from the peak and the feeding intensity values ​​at three consecutive sampling points are all lower than the peak value multiplied by a preset decline ratio, feeding behavior is determined to have declined. The time corresponding to the first sampling point among the three sampling points is taken as the time of decline in feeding behavior. The preset decline ratio is obtained by collecting data on the decline amplitude of feeding intensity from the peak to the point of satiation during multiple feeding cycles, and taking the minimum value of the ratio of the feeding intensity value to the peak value during the fluctuation decline phase as the preset decline ratio. The maximum value of the time interval from the initial rise in ammonia nitrogen concentration to the point where feeding behavior declines is taken from multiple recorded intervals, converted into the number of sampling points, rounded up, and used as the preset maximum offset. ; The feeding attenuation gradient sequence Fixed, ammonia nitrogen change rate sequence Slide successively along the negative direction of the time axis with a step size of 1, for the offset ,Pick Tail subsequence At the same time, take Equal-length subsequences ,in, for and The overlap length; the offset is calculated using the following Pearson correlation coefficient formula. Cross-correlation strength ,in, For subsequence The Middle The value of each element, For subsequence The Middle The value of each element, For subsequence The mean, For subsequence The mean, Let represent the overlap length between the two subsequences, where The value range of is [-1, 1]. The closer to 1, the higher the offset. The stronger the positive correlation between changes in ammonia nitrogen and nutrient decline, the closer to -1 indicates a stronger negative correlation, and the closer to 0 indicates no linear correlation. Complete all at each offset After calculation, the cross-correlation curve showing the cross-correlation strength as a function of the offset is obtained. In this cross-correlation curve, take Offset corresponding to the maximum value As the real-time causal lag step number within the current feeding cycle; based on the real-time causal lag step number Subtract the timestamp of each data point in the water quality trend factor sequence. The operation at each time step aligns the shifted water quality trend factor sequence with the appetite dynamic factor sequence in causal time order; in the aligned water quality trend factor sequence, the first time step... The timestamp corresponding to the data point and the first data point in the appetite dynamic factor sequence The timestamps of the data points correspond causally, and the water quality trend factor sequence and appetite dynamic factor sequence after causal time-series alignment are used as inputs for the subsequent time-series feature encoder.

[0025] In one embodiment of the present invention, the cross-correlation strength With offset In the cross-correlation curve formed by the changes, Offset corresponding to the maximum value The real-time causal lag step number is determined within the current feeding cycle; based on the real-time causal lag step number... Subtract the timestamp of each data point in the water quality trend factor sequence. A time step was used to align the translated water quality trend factor sequence with the appetite dynamic factor sequence in causal time.

[0026] In one embodiment of the present invention, the dual-head predictive decoder includes a bidirectional cross-attention layer, a dynamic gating unit, a water quality prediction head, and an appetite decline prediction head. The bidirectional cross-attention layer includes a first cross-attention sub-layer and a second cross-attention sub-layer. The first cross-attention sub-layer uses the water quality trend factor sequence after causal temporal alignment as the query matrix and the appetite dynamic factor sequence as the key matrix and value matrix to output the appetite context sequence of water quality perception. The second cross-attention sub-layer uses the appetite dynamic factor sequence as the query matrix and the water quality trend factor sequence after causal temporal alignment as the key matrix and value matrix to output the water quality context sequence of appetite perception.

[0027] Specifically, the bidirectional cross-attention layer comprises a first cross-attention sub-layer and a second cross-attention sub-layer. The two sub-layers have the same structure but the allocation directions of the query matrix and the key-value matrix are opposite. The input to the first cross-attention sub-layer is the query matrix obtained by linearly transforming the water quality trend factor sequence after causal time-series alignment. The appetite dynamic factor sequence after linear transformation is used as the key matrix. Sum matrix The output of the first cross-attention sublayer is calculated using the following formula: ,in, For querying the matrix, The key matrix, For value matrices, This is the transpose of the key matrix. Let be the column dimension of the key matrix. Scaling factor The function is a row-normalized function; the output of this formula is the appetite context sequence perceived by water quality. The features of each time step in this sequence are fused with appetite information from multiple time steps that are causally related to that time step. The input of the second cross-attention sublayer is: the appetite dynamic factor sequence after linear transformation is used as the query matrix. The water quality trend factor sequence after causal time-series alignment is used as the key matrix after linear transformation. Sum matrix The second cross-attention sublayer is calculated using the same formula: The formula outputs a sequence of water quality contexts related to appetite perception. The dynamic gating unit is connected to the transient disturbance filtering module. During each feeding cycle, the dynamic gating unit obtains two statistical values ​​from the transient disturbance filtering module: and ,in, This refers to the number of pulse waveform data segments identified as bullfrog activity interference and removed during the feeding cycle. The total number of times the high-speed pulse sampling sequence is triggered within the feeding cycle; when hour, ;when hour, ;in, This represents the proportion of excluded interference pulse segments to the total number of high-speed sampling triggers. A higher proportion indicates more frequent interference from bullfrog activity with the water quality data, and thus lower reliability of the water quality data. The value range is from 0 to 1, where, The closer the value is to 1, the cleaner and more reliable the water quality data is; Multiply the appetite context sequence perceived by water quality Multiply the water quality context sequence of appetite perception by The sum of the two is concatenated with the aligned water quality trend factor sequence and input into the temporal feature encoder. The calculation formula is as follows: ,in, For gated fusion context sequence, The gating coefficient is used to concatenate the gating fusion context sequence and the aligned water quality trend factor sequence along the feature dimension to obtain the concatenated sequence, which is then input into the temporal feature encoder. The temporal feature encoder employs a recurrent neural network composed of a long short-term memory network. This recurrent neural network comprises four components: an input gate, a forget gate, an output gate, and a memory unit. At each time step, the temporal feature encoder receives the feature vector from the concatenated sequence at the current time step, which is input together with the hidden state from the previous time step, and updated according to the following formula: ; ; ; ; ; ;in, The input feature vector at the current time step. This is the hidden state from the previous time step. This refers to the state of the memory unit at the previous time step. , and These are the outputs of the input gate, forget gate, and output gate, respectively. Candidate memory cell state, This is the updated memory cell state. The hidden state at the current time step. , , and This is the weight matrix. , , and For bias terms, It is the Sigmoid activation function. The hyperbolic tangent activation function is used; the hidden state output at the last time step of the temporal feature encoder is taken as the fused temporal feature vector. The water quality prediction head receives a fused temporal feature vector as input, and maps the fused temporal feature vector to future values ​​through a fully connected layer. Dissolved oxygen prediction sequence at each time step ,in, The preset prediction step number is determined by recording the average time interval from the start of feeding to the start of the next feeding within the breeding cycle. This time interval is then converted into a time step number and rounded down to the nearest integer. The appetite loss prediction head receives the same fused temporal feature vector as input, and maps the fused temporal feature vector to future values ​​through another fully connected layer. Predicted sequence of grazing decay gradient values ​​at each time step ; In the water quality-appetite bidirectional compensation model, the cross-modal causal time lag discovery module is a parameter-free computation module that requires no training. The temporal feature encoder, bidirectional cross-attention layer, water quality prediction head, and appetite decay prediction head are trainable components, trained jointly. The training samples are constructed as follows: data from multiple feeding cycles within the aquaculture period are collected. Each feeding cycle includes a water quality trend factor sequence and an appetite dynamic factor sequence as input samples, and actual dissolved oxygen value sequences and actual feed decay gradient value sequences as labels. The input samples and labels are organized into sample-label pairs according to the feeding cycle, and the training and validation sets are divided chronologically. The loss function for joint training is: ,in, This is a sequence of dissolved oxygen prediction values ​​output by the water quality prediction head. This is the actual dissolved oxygen value sequence. This is the sequence of predicted gradient values ​​for appetite decline output by the appetite decline prediction head. This is the actual sequence of food decay gradient values. Let be the mean square error function. These are preset weighting coefficients, ranging from 0 to 1, used to balance the weights of the water quality prediction task and the appetite loss prediction task during training. By training models with the same structure separately on the training set for both water quality prediction and appetite decline prediction tasks, and recording the validation set loss values ​​at convergence for each task, the ratio of the reciprocals of the two loss values ​​is normalized and used as the result. The value of the loss function is calculated; during training, training samples are input into the model sequentially, the loss function value is calculated, and the trainable parameters are updated through gradient backpropagation. This process is repeated iteratively until the loss function value converges.

[0028] In one embodiment of the present invention, the dynamic gating unit is connected to the transient disturbance filtering module, and obtains two statistical values ​​for the current feeding cycle from the transient disturbance filtering module. and ,in, This refers to the number of pulse waveform data segments identified as bullfrog activity interference and removed during the feeding cycle. The total number of times the high-speed pulse sampling sequence is triggered within the feeding cycle; the dynamic gating unit calculates the gating coefficient. , among which, when hour, ;when hour, ; Multiply the appetite context sequence perceived by water quality Multiply the water quality context sequence of appetite perception by The two are added together and then concatenated with the aligned water quality trend factor sequence, and then input into the temporal feature encoder; The fused temporal feature vector output by the temporal feature encoder is input to the water quality prediction head and the appetite decline prediction head, respectively.

[0029] In one embodiment of the present invention, the combined control command for generating the feeding rate reduction command, the low-protein feed ratio switching command, and the oxygen-enriched area directional feeding control signal includes: The joint control command is used to execute the feeding trajectory, specifically: to perform the first feeding at the current gathering location of the frog swarm identified by the feeding behavior analysis unit, and then to feed at preset time intervals. The mobile chassis will be moved a preset distance toward the oxygen-rich area. The angle of the variable speed throwing disc is simultaneously adjusted to point to the water surface area within a preset radius centered on the current position of the mobile chassis, until the mobile chassis reaches the oxygen-rich area. Each movement is followed by a feeding, with the amount of food given increasing sequentially along the direction of movement. ; The oxygen-rich area is identified by a water quality sensing array and is defined as a local water area where the dissolved oxygen saturation is greater than the second threshold.

[0030] Specifically, in the first fusion state, the multi-factor fusion decision engine sends an oxygen-rich area identification request to the water quality sensing array. The water quality sensing array obtains the current dissolved oxygen saturation data of each deployment point in the aquaculture pond and defines the local water area with dissolved oxygen saturation greater than a second threshold as an oxygen-rich area. The second threshold is determined by recording the lowest dissolved oxygen saturation value of the feeding area during each feeding period when the bullfrogs are feeding normally and there is no oxygen deficiency surfacing phenomenon during the aquaculture cycle. The average of the multiple recorded lowest values ​​is taken as the second threshold. The boundary of the oxygen-rich area is determined by using the dissolved oxygen saturation data of each deployment point as sampling points and generating a dissolved oxygen saturation distribution map of the aquaculture pond surface through spatial interpolation. The contour lines where the dissolved oxygen saturation is equal to the second threshold are extracted from the distribution map. The area enclosed by these contour lines is the oxygen-rich area. The projection point of the geometric center of the oxygen-rich area onto the trajectory of the aquaculture pond shore is taken as the target endpoint position of the moving chassis in the joint control command. Among them, the multi-factor fusion decision engine sends a frog group location identification request to the feeding behavior analysis unit. The feeding behavior analysis unit extracts the three-dimensional coordinates of the skeletal points of all bullfrog individuals in the picture through the skeletal point behavior analysis module based on the image pair of the current frame of the underwater binocular vision module, calculates the geometric center of the three-dimensional coordinates of the skeletal points of all bullfrog individuals, projects the geometric center onto the horizontal plane, and obtains the two-dimensional coordinates of the current gathering position of the frog group, which serves as the placement position for the first feeding in the joint control command. Among them, the preset time interval in the joint control command The average time taken by bullfrogs from sensing the food falling into the water to swimming to the food's location and swallowing it was recorded as the average time. Preset distance The furthest diffusion radius of feed particles on the water surface during a single feeding is measured and taken as the [spread radius]. Feeding increment By recording the increase in feeding amount per unit time during the period from the start of feeding to the peak feeding time of the frog troop under normal feeding conditions, the average of these increases was taken as the mean. The preset radius is obtained by measuring the radius of the circular coverage area formed by the feed pellets being thrown onto the water surface by the throwing disc at the rated speed. After receiving the joint control command, the mixed feeding actuator executes the feeding trajectory in the following sequence: First, the mobile chassis of the mixed feeding actuator moves to the pond bank position corresponding to the current gathering position of the frog swarm identified in step two, adjusts the angle of the feeding plate to point to this gathering position, and performs the first feeding, with the feeding amount being the preset basic feeding amount. The preset basic feeding amount is determined by recording the total amount of food required for the frog swarm to reach a state of satiety during normal feeding, and taking a preset proportion of this total feeding amount as the baseline. The second step is to wait for the preset time interval. Then, move the mobile chassis a preset distance toward the geometric center of the oxygen-rich area. Simultaneously adjust the angle of the feeding disc to point it at the water surface area within a preset radius centered on the current position of the moving chassis, and perform the second feeding, with a feeding amount of [missing information]. Third, repeat the operation of step two, increasing the feeding amount after each movement by the amount of the previous feeding. The feeding process continues until the mobile chassis reaches the pool bank position corresponding to the geometric center of the oxygen-rich area, at which point the final feeding is performed, ending the current feeding trajectory.

[0031] In one embodiment of the present invention, the mixing and feeding actuator includes a dual hopper, a twin-screw feeder, a mixing chamber, and a variable-speed throwing disc; The dual silos store conventional protein feed and low-protein feed respectively; The low-protein feed formulation switching command specifies a protein content decreasing curve. ,in, To switch the real-time feed protein content at the start time. The preset deceleration rate, This is the time counted from the start of the switchover; At each time step, according to Calculate the target protein content at the current time step, and based on the target protein content, back-calculate the rotational speed ratio between the conventional protein feed conveyor and the low-protein feed conveyor in the double-helix feeder. Drive the two feeders according to the stated rotational speed ratio, so that the protein content of the feed output from the mixing chamber decreases along the stated curve from... The protein content is continuously reduced until it reaches the preset target low protein content value.

[0032] Specifically, the mixed feeding mechanism includes a dual-feed silo, a double-screw feeder, a mixing chamber, and a variable-speed throwing disc. The first silo stores conventional protein feed with a protein content that meets the standard protein content required for normal bullfrog growth during the breeding cycle. The second silo stores low-protein feed with a preset target low-protein content. This target low-protein content is determined by recording the lowest feed protein content that does not significantly inhibit bullfrog weight gain during the breeding cycle when water quality shows a downward trend but has not yet exceeded the standard, and taking the average of multiple recorded minimum values ​​as the target low-protein content. The outlet of the first silo is connected to the first screw feeder, and the outlet of the second silo is connected to the second screw feeder. The outlets of both screw feeders are connected to the inlet of the mixing chamber. The mixing chamber is equipped with stirring blades to mix the two feeds entering the mixing chamber evenly. The outlet of the mixing chamber is connected to the variable-speed throwing disc, which rotates to throw the mixed feed particles onto the surface of the breeding pond. Specifically, when the multi-factor fusion decision engine generates a low-protein feed ratio switching instruction in the first fusion state, the instruction specifies a protein content decreasing curve. ,in, The first time after the switch starts The feed protein content that the mixing chamber should output at each time step. : ,in, for , The current rotational speed of the first screw conveyor. This refers to the current rotational speed of the second screw conveyor. This refers to the protein content of a conventional protein feed. The protein content of low-protein feed; among which, The preset deceleration rate is expressed as a percentage of protein content per time step. By recording the average number of transition days required for bullfrogs to fully adapt to a low-protein diet from normal feeding, and converting this number of days into time steps, the following was followed... Obtain, among which, The target is a low protein content value. This represents the total number of time steps. At each time step, perform the following steps: First, calculate the target protein content for the current time step based on the protein content decline curve. ,like If the target low protein content value is reached, then... The value is set to the target low protein content value; then, based on the target protein content... The reverse calculation of the rotational speed ratio between the first and second screw conveyors in a twin-screw conveyor system is derived from the mixing rule: Let the rotational speed of the first screw conveyor be... The rotational speed of the second screw conveyor is The protein content of the feed output from the mixing chamber is Then there is ,and ,in, The preset total conveying rate is calculated by recording the total mass of feed output from the mixing chamber per unit time during normal feeding, and then converting this mass into the total rotational speed of the two conveyors. The result is obtained through inverse solution. and Finally, according to the calculation... Drive the first screw conveyor, press The second screw conveyor is driven, causing the two conveyors to deliver corresponding masses of conventional protein feed and low-protein feed to the mixing chamber within the current time step. After mixing, the feed is ejected by the variable-speed throwing disc, causing the protein content of the feed output from the mixing chamber to decrease along a curve. Continuously reduce to the target low protein content value.

[0033] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A multi-factor fusion-based water quality early warning and intelligent feeding system for bullfrog farming, characterized in that, include: A water quality sensing array is used to acquire a water quality trend factor sequence at a first sampling frequency, the water quality trend factor sequence including the ammonia nitrogen change rate and the dissolved oxygen change rate; The feeding behavior analysis unit is used to acquire an appetite dynamic factor sequence at a second sampling frequency, the appetite dynamic factor sequence including the peak value of cluster feeding intensity and the feeding decay gradient; The multi-factor fusion decision engine is connected to the water quality sensing array and the feeding behavior analysis unit respectively, and runs a water quality-appetite bidirectional compensation model internally. The water quality-appetite bidirectional compensation model: In the first fusion state, in response to the water quality trend factor sequence pointing to a downward trend in dissolved oxygen and the peak value of the cluster foraging intensity > a first threshold, a joint instruction is generated, consisting of a feeding rate reduction instruction, a low-protein feed ratio switching instruction, and a directional feeding control signal for oxygen-rich areas; In the second fusion state, the foraging attenuation gradient is used as the trigger variable for water quality early warning. When the foraging attenuation gradient exceeds the non-satiation stress attenuation benchmark value within a preset time window after feeding is initiated, a latent water quality mutation early warning signal based on biological behavior is generated. The mixed feeding actuator is used to receive and execute the feeding rate reduction command, the low protein feed ratio switching command, and the oxygen-enriched area directional feeding control signal generated in the first fusion state.

2. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion as described in claim 1, characterized in that, The water quality sensing array includes an ion-selective electrode array, a miniature underwater spectral probe, and a transient disturbance filtering module that is electrically connected to the ion-selective electrode array and the miniature underwater spectral probe, respectively. The first sampling frequency is a dynamically adjusted frequency based on event triggering; The transient disturbance filtering module is used to trigger a high-speed pulse sampling sequence when the ion-selective electrode array or the micro underwater spectral probe detects an ammonia nitrogen concentration or turbidity greater than a preset disturbance trigger threshold in a single sampling; when the high-speed pulse sampling sequence is less than a complete bullfrog feeding or tail-wagging action cycle, it is used to capture the complete pulse waveform of water quality trend factors in disturbance events. The transient disturbance filtering module is also used to extract features from the complete pulse waveform. If its peak half-width and waveform attenuation shape meet the morphological criteria for bullfrog activity disturbance, the data segment corresponding to the pulse waveform is removed from the water quality trend factor sequence input to the multi-factor fusion decision engine.

3. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion as described in claim 1, characterized in that, The feeding behavior analysis unit includes an underwater binocular vision module, a skeletal point behavior analysis module, and a frog swarm overlapping individual segmentation module connected to the skeletal point behavior analysis module. The underwater binocular vision module is used to simultaneously acquire individual bullfrog image pairs in the feeding area; The skeletal point behavior analysis module is used to extract the skeletal point coordinates of individual bullfrogs based on the image, and to track the skeletal point trajectory of each bullfrog in consecutive frames. The frog swarm overlapping individual segmentation module is used to re-identify each bullfrog after the intersection when the skeletal point trajectories of at least two bullfrogs intersect in space and the number of frames of the overlap exceeds the overlap threshold, based on the snout morphological features and body surface pattern features of each bullfrog before the trajectory intersection. Among them, the peak of the cluster foraging intensity is determined by the number of independent bullfrog individuals whose snout vertical positive displacement exceeds the predation threshold after identity re-identification within a unit of time; the foraging attenuation gradient is the attenuation slope of the peak of the cluster foraging intensity within a preset time window.

4. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion as described in claim 1, characterized in that, In the multi-factor fusion decision engine, the water quality-appetite bidirectional compensation model includes a cascaded cross-modal causal time-delay detection module, a temporal feature encoder, and a dual-head prediction decoder: The cross-modal causal time-delay detection module, before concatenating the water quality trend factor sequence and the appetite dynamic factor sequence, takes the ammonia nitrogen change rate sequence within the first preset time period before the start of the current feeding, and records it as... And the food decay gradient sequence within the second preset time period after the start of feeding, denoted as ,in, and These are the lengths of the ammonia nitrogen change rate sequence and the feeding decay gradient sequence, respectively. The feeding attenuation gradient sequence Fixed, ammonia nitrogen change rate sequence Slide successively in the negative direction of the time axis with a step size of 1, for the offset... ,in, Take the preset maximum offset. Tail subsequence and Equal-length subsequences ,in, The overlap length between the two subsequences; The offset is calculated using the Pearson correlation coefficient formula. Cross-correlation strength .

5. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion according to claim 4, characterized in that, The cross-correlation strength With offset In the cross-correlation curve formed by the changes, Offset corresponding to the maximum value The real-time causal lag step number is determined within the current feeding cycle; based on the real-time causal lag step number... Subtract the timestamp of each data point in the water quality trend factor sequence. A time step was used to align the translated water quality trend factor sequence with the appetite dynamic factor sequence in causal time.

6. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion according to claim 4, characterized in that, The dual-head predictive decoder includes a bidirectional cross-attention layer, a dynamic gating unit, a water quality prediction head, and an appetite decline prediction head. The bidirectional cross-attention layer includes a first cross-attention sub-layer and a second cross-attention sub-layer. The first cross-attention sub-layer uses the water quality trend factor sequence after causal temporal alignment as the query matrix and the appetite dynamic factor sequence as the key matrix and value matrix to output the appetite context sequence of water quality perception. The second cross-attention sublayer uses the appetite dynamic factor sequence as the query matrix and the water quality trend factor sequence after causal temporal alignment as the key matrix and value matrix to output the water quality context sequence of appetite perception.

7. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion according to claim 6, characterized in that, The dynamic gating unit is connected to the transient disturbance filtering module and obtains two statistical values ​​for the current feeding cycle from the transient disturbance filtering module. and ,in, This refers to the number of pulse waveform data segments identified as bullfrog activity interference and removed during the feeding cycle. The total number of times the high-speed pulse sampling sequence is triggered within the feeding cycle; the dynamic gating unit calculates the gating coefficient. , among which, when hour, ;when hour, ; Multiply the appetite context sequence perceived by water quality Multiply the water quality context sequence of appetite perception by The two are added together and then concatenated with the aligned water quality trend factor sequence, and then input into the temporal feature encoder; The fused temporal feature vector output by the temporal feature encoder is input to the water quality prediction head and the appetite decline prediction head, respectively.

8. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion according to claim 1, characterized in that, The combined control command for generating the feeding rate reduction command, the low-protein feed ratio switching command, and the oxygen-enriched area directional feeding control signal includes: The joint control command is used to execute the feeding trajectory, specifically: to perform the first feeding at the current gathering location of the frog swarm identified by the feeding behavior analysis unit, and then to feed at preset time intervals. The mobile chassis will be moved a preset distance toward the oxygen-rich area. The angle of the variable speed throwing disc is simultaneously adjusted to point to the water surface area within a preset radius centered on the current position of the mobile chassis, until the mobile chassis reaches the oxygen-rich area. Each movement is followed by a feeding, with the amount of food given increasing sequentially along the direction of movement. ; The oxygen-rich area is identified by a water quality sensing array and is defined as a local water area where the dissolved oxygen saturation is greater than the second threshold.

9. The bullfrog farming water quality early warning and intelligent feeding system based on multi-factor fusion according to claim 1, characterized in that, The mixed feeding actuator includes a dual hopper, a twin-screw conveyor, a mixing chamber, and a variable-speed throwing disc; The dual silos store conventional protein feed and low-protein feed respectively; The low-protein feed formulation switching command specifies a protein content decreasing curve. ,in, To switch the real-time feed protein content at the start time. The preset deceleration rate, This is the time counted from the start of the switchover; At each time step, according to Calculate the target protein content at the current time step, and based on the target protein content, back-calculate the rotational speed ratio between the conventional protein feed conveyor and the low-protein feed conveyor in the double-helix feeder. Drive the two feeders according to the stated rotational speed ratio, so that the protein content of the feed output from the mixing chamber decreases along the stated curve from... The protein content is continuously reduced until it reaches the preset target low protein content value.

Images