Unmanned feeding ship dynamic feeding control method based on fish school feeding behavior signal analysis

Abstract

The invention relates to the technical field of aquaculture, in particular to an unmanned feeding ship dynamic feeding control method based on fish school feeding behavior signal analysis. Comprising the following steps: collecting and synchronously transmitting a fish school feeding signal; analyzing feeding behavior signals and calculating fish school hunger state parameters; generating a dynamic feeding decision; accurate feeding execution and feed bin state feedback are realized. According to the method, an acoustic sensor acoustic signal is taken as a reference, and a near-infrared optical signal is aligned through wavelet packet time-frequency interpolation; compensating environmental interference by using a frequency domain sub-coefficient associated with the reciprocal of the turbidity, a time domain sub-coefficient positively associated with the water flow and a space domain sub-coefficient positively associated with the reciprocal of the turbidity; and then the ingestion intensity gradient is extracted based on the comprehensive weight through attention weighted wavelet packet transformation, continuous fish school hunger parameters are obtained through quantification, and the problems that in the prior art, environment interference is not compensated, only discrete grades are output, and accurate feeding adjustment is difficult to support are solved.

Description

Technical Field

[0001] This invention relates to the field of aquaculture technology, and more specifically, to a dynamic feeding control method for unmanned feeding vessels based on the analysis of fish feeding behavior signals. Background Technology

[0002] With the development of large-scale and intensive aquaculture, unmanned feeding boats have become mainstream due to cost reduction and efficiency improvement. However, the traditional "fixed feeding amount + preset route" model does not take into account the hunger level of fish and the influence of the aquatic environment, which can easily lead to feed waste or insufficient feeding. Therefore, it is necessary to develop dynamic feeding technology based on the analysis of fish feeding behavior to accurately capture feeding characteristics, quantify hunger status, and achieve adaptive adjustment of feeding parameters.

[0003] In existing technologies, relevant patents have explored fish feeding identification and feeding control. For example, Chinese patent CN202411621346.3 discloses a method and system for hierarchical identification of fish feeding behavior based on vibration signals. This includes controlling multiple vibration sensors to collect feeding vibration signals, extracting features through synthesis preprocessing, filtering, framing, and wavelet transform, and inputting the signals into a deep learning model after dimensionality reduction to obtain hierarchical identification results. This achieves precise feeding and improves aquaculture efficiency at low cost. Another example is Chinese patent CN202310428730.0, which discloses a feeding amount decision method based on acoustic and water surface undulation information. This method divides feeding activity into four levels, collects acoustic signals, water surface undulation data, and feeding videos, extracts features after preprocessing, performs principal component analysis, and uses a classifier to determine activity levels and issue feeding instructions, providing support for precise feeding.

[0004] Despite the design advantages of the above technical solutions, they also have the following technical defects: First, the information dimensions and environmental adaptability of feeding behavior recognition are insufficient: CN202411621346.3 relies solely on a single vibration signal for recognition, without incorporating acoustic signals reflecting swallowing details, image signals presenting fish distribution, or considering the interference of turbidity and water flow on recognition, making the recognition basis singular; CN202310428730.0, although combining acoustic and water surface wave signals, does not resolve signal time deviation or compensate for environmental interference, and both CN202411621346.3 and CN202310428730.0 only output discrete levels, unable to provide continuous quantitative results, making it difficult to support precise feeding adjustments. Secondly, the feeding control lacks systematicity and closed-loop functionality: CN202411621346.3 and CN202310428730.0 both lack a dynamic correction mechanism for the feeding direction, making it impossible to match the location of fish aggregation, and they do not incorporate feedback on remaining feed in the hopper. Specifically, CN202411621346.3 does not clearly define the correlation logic between the identification results and the feeding amount and interval, while CN202310428730.0 does not form an "adjustment-execution-feedback-optimization" closed loop, making it difficult to adapt to complex aquaculture scenarios. Therefore, we propose a dynamic feeding control method for unmanned feeding vessels based on the analysis of fish feeding behavior signals. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic feeding control method for unmanned feeding vessels based on the analysis of fish feeding behavior signals, so as to solve the problems of insufficient information dimension and environmental adaptability in feeding behavior recognition and insufficient systematicity and closed-loop of feeding control mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention aims to provide a dynamic feeding control method for unmanned feeding vessels based on the analysis of fish feeding behavior signals, comprising the following steps: S1. Fish feeding signal acquisition and synchronous transmission: Vibration sensors, acoustic sensors, near-infrared sensors, turbidity sensors, and water flow sensors are deployed on the bottom of the unmanned feeding vessel to synchronously acquire micro-vibration signals of the water body, feeding and swallowing sound signals, fish activity signals, turbidity parameters, and water flow parameters; using the acoustic signal acquired by the acoustic sensor as a reference, wavelet packet time-frequency interpolation technology is used to align the optical signal acquired by the near-infrared sensor, and the synchronized signal and parameters are encapsulated into a standard data frame and transmitted to the feeding behavior signal analysis stage via the CAN bus; S2. Feeding Behavior Signal Analysis and Fish School Hunger State Parameter Calculation: Based on the received signals and parameters, the energy ratio of a specific frequency band of feeding and swallowing sounds is calculated as the frequency domain weight, the ratio of transient energy of feeding and swallowing sounds to total energy is calculated as the time domain weight, and the reciprocal of the spatial aggregation degree of the fish school is calculated as the spatial domain weight. Frequency domain sub-coefficients that are reciprocally correlated with turbidity parameters, time domain sub-coefficients that are positively correlated with water flow parameters, and spatial domain sub-coefficients that are positively correlated with the reciprocal of turbidity parameters are calculated respectively. After coupling each weight with the corresponding dynamic sub-coefficients, the weights are normalized to obtain the comprehensive weight. The attention-weighted wavelet packet transform algorithm is used to extract the feeding intensity gradient based on the comprehensive weight, and the continuous fish school hunger state parameters are obtained by quantization mapping. S3. Dynamic Feeding Decision Generation: Based on the average body weight of fish in the current breeding cycle, the baseline feeding amount is calculated. The feeding parameters are adjusted according to the continuous fish population hunger state parameters. The feeding amount increases as the continuous fish population hunger state parameters increase and decreases as the continuous fish population hunger state parameters decrease. The feeding interval decreases as the continuous fish population hunger state parameters increase and increases as the continuous fish population hunger state parameters decrease. Combining the real-time GPS positioning signal of the unmanned feeding vessel with the coordinates of the preset target feeding area, the offset between the vessel and the target feeding area is calculated, and the feeding direction is corrected to form a feeding control strategy that includes parameters such as feeding amount, feeding interval, and feeding direction. S4. Precise feeding execution and silo status feedback: Based on the feeding control strategy, the feeding speed is adjusted by the stepper motor feeding mechanism to control the feeding amount, the feeding range is controlled by the direction and opening and closing of the electric spray nozzle, and the silo level sensor collects the silo level data and feeds it back to S3.

[0007] As a further improvement to this technical solution, in step S1, the process of collecting and synchronously transmitting the fish feeding signals includes the following steps: S11. Based on the deployment location requirements of the unmanned feeding vessel, the vibration sensor, sound sensor, near-infrared sensor, turbidity sensor and water flow sensor are arranged in a preset symmetrical layout along the central axis of the bottom of the vessel. S12. Using the sensor from S11, it synchronously collects water micro-vibration signals, feeding and swallowing sound signals, fish activity signals, turbidity parameters, and water flow parameters; S13. Based on the acoustic signal collected by the acoustic sensor, wavelet packet time-frequency interpolation technology is used to align the optical signal collected by the near-infrared sensor to eliminate the time deviation of multi-sensor data. S14. Based on the synchronized and aligned signals and parameters, the data is encapsulated into standard data frames using the CAN bus communication protocol and transmitted to S2.

[0008] As a further improvement to this technical solution, in step S2, the process of analyzing feeding behavior signals and calculating fish hunger state parameters includes the following steps: S21. Based on the collected signals, calculate the energy proportion of a specific frequency band of swallowing sounds as the frequency domain weight. Calculate the ratio of the transient energy of the swallowing sound to the total energy as the time-domain weight. Calculate the reciprocal of the spatial clustering degree of the fish school as the spatial weight. ; S22. Based on the collected parameters, calculate the corresponding frequency domain weights respectively. Time domain weights Spatial weights Frequency domain sub-coefficients Time-domain sub-coefficients Spatial sub-coefficients ; S23, The frequency domain weights With frequency domain sub-coefficients Time-domain weights and time-domain sub-coefficients Spatial weights With spatial sub-coefficients The components are merged separately to obtain the overall weight. ; S24. Employ the attention-weighted wavelet packet transform algorithm, based on the aforementioned comprehensive weights. Extracting the feeding intensity gradient ; S25, Based on the feeding intensity gradient The parameters of the continuous fish population's starvation state were calculated. .

[0009] As a further improvement to this technical solution, in step S21, the process of calculating weights based on the acquired signals includes the following steps: S21.1 Extract the feeding and swallowing sound signal transmitted in S1, and retain the characteristic frequency band of fish feeding by bandpass filtering. Calculate the frequency domain weight by the proportion of energy in this frequency band to the total energy of the entire frequency band. ; S21.2 Perform time-domain analysis on the swallowing sound signal to define the transient interval reflecting the eating action, and calculate the time-domain weight by the ratio of the energy in this interval to the energy of the entire frequency band. ; S21.3 Receive the near-infrared fish activity image transmitted by S1, extract the spatial distribution information of the fish school after denoising and target segmentation, and calculate the spatial weight based on the spatial clustering degree of the fish school. .

[0010] As a further improvement to this technical solution, in step S22, the process of calculating the dynamic sub-coefficients based on the collected parameters includes the following steps: S22.1 Turbidity parameters based on S1 Calculate the frequency domain sub-coefficients based on the reciprocal correlation. To correct the interference of turbidity on acoustic signal feature extraction; S22.2, Water flow parameters collected based on S1 Calculate time-domain sub-coefficients based on positive correlation. The influence of water flow velocity on the timing characteristics of fish feeding behavior; S22.3, Based on turbidity parameters The spatial sub-coefficients are calculated based on the positive correlation of their reciprocals. To compensate for the impact of turbidity on the accuracy of near-infrared image recognition.

[0011] As a further improvement to this technical solution, in step S23, the comprehensive weight is... The calculation process includes the following steps: S23.1, Frequency domain weights With frequency domain sub-coefficients Time domain weights With time-domain sub-coefficients Spatial weights With spatial sub-coefficients Weighted coupling is performed separately to obtain weighted components in three frequency-temporal-spatial dimensions; S23.2. Normalize the three weighted components to eliminate the dimensional differences between parameters of different dimensions, and obtain the comprehensive weight. .

[0012] As a further improvement to this technical solution, in step S24, the process of extracting the feeding intensity gradient includes the following steps: S24.1 Perform wavelet packet time-frequency decomposition on the swallowing sound signal to obtain multi-band sub-signals covering the characteristic frequency band of swallowing; S24.2, Based on comprehensive weights Assign attention coefficients to sub-signals in each frequency band Strengthen the signal weights of frequency bands related to feeding behavior; S24.3, Based on Attention Coefficient Weighted reconstruction of multi-band sub-signals is performed, and the energy change rate of the reconstructed signal is extracted to obtain the feeding intensity gradient reflecting feeding activity. .

[0013] As a further improvement to this technical solution, in step S25, the continuous fish school hunger state parameters are calculated. The process includes the following steps: S25.1, Call the feeding intensity gradient output from S24.3 The overall weights output from S23.2 Simultaneously extract varietal characteristic parameters from the sound signals of feeding and swallowing. The This is obtained by analyzing the characteristic frequency band distribution width and peak interval patterns of the signal, and is used to match the differences in feeding behavior among different fish species. S25.2, Gradient of feeding intensity Perform moving average processing based on variety characteristic parameters. By dynamically adjusting the sliding window parameters, a stable gradient value is obtained after eliminating instantaneous signal fluctuations. ; S25.3, Based on comprehensive weights , Stabilization gradient and variety characteristic parameters Construct association logic to obtain initial hunger parameters through the mapping relationship between signal features and hunger state. ; S25.4, to Standardization corrections are made when When the value exceeds the [0,1] interval, it is truncated according to the boundary value, and the final output is a continuous fish swarm hunger state parameter. .

[0014] As a further improvement to this technical solution, in step S3, the process of generating dynamic feeding decisions includes the following steps: S31. Receive the average body weight data of fish in the current breeding cycle, and calculate the benchmark feeding amount to meet the basic nutritional needs of farmed fish in combination with the characteristics of the growth stage of farmed fish. S32. Call the continuous fish swarm hunger state parameters output by S2. Based on the baseline feeding amount, a continuous fish population starvation status parameter was established. The logic relating the feeding amount and feeding interval is used to implement the feeding amount based on the continuous fish population hunger state parameters. Increases with the rise in continuous fish population hunger parameters The feeding interval is adjusted in a reverse linkage with the reduction. S33. Receive the real-time GPS positioning signal and preset feeding area coordinate information of the unmanned feeding vessel, calculate the offset between the current position of the vessel and the target feeding area, and output the correction angle of the sprinkling direction. S34. Simultaneously receive the baseline feeding amount, adjusted feeding parameters, and scattering direction correction angle, integrate them into a standardized control command, and transmit it to S4.

[0015] As a further improvement to this technical solution, in step S4, the process of precise feeding execution and silo status feedback includes the following steps: S41. Receive the standardized control command output by S3 and parse it to obtain the target feeding amount, feeding interval and scattering direction correction angle; S42. Based on the analyzed target feeding amount, the feeding speed is adjusted through the stepper motor feeding mechanism to achieve precise control of the feeding amount and ensure consistency with the decision command. S43. Adjust the angle according to the direction of spraying, control the rotation and opening degree of the electric spraying head, adjust the feeding coverage range, and match the preset feeding area; S44. Collect the remaining material data in the silo in real time through the material level sensor, and feed the remaining information back to S3 to provide material status support for the dynamic adjustment of feeding parameters.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention addresses the limitations of existing technologies that rely solely on a single signal to identify feeding behavior by deploying vibration sensors, acoustic sensors, near-infrared sensors, turbidity sensors, and water flow sensors at the bottom of the unmanned feeding vessel. Simultaneously, using the acoustic signal collected by the acoustic sensor as a reference, wavelet packet time-frequency interpolation is employed to align the optical signal collected by the near-infrared sensor. Environmental interference is compensated by calculating frequency domain coefficients reciprocally correlated with turbidity parameters, time domain coefficients positively correlated with water flow parameters, and spatial domain coefficients positively correlated with the reciprocal of turbidity parameters. Furthermore, an attention-weighted wavelet packet transform algorithm is used to extract the feeding intensity gradient based on comprehensive weights, quantizing and mapping continuous fish hunger state parameters. This solves the problems of existing technologies that fail to compensate for environmental interference and only output discrete levels, making precise feeding adjustments difficult.

[0017] 2. This invention calculates the baseline feeding amount based on the average body weight of fish in the current breeding cycle, and clarifies the correlation logic between continuous fish hunger state parameters and feeding amount and interval. At the same time, it combines the real-time GPS positioning signal of the unmanned feeding vessel with the coordinates of the preset target feeding area to calculate the offset and correct the feeding direction to match the location of fish gathering. It also collects the remaining data of the feed bin through the feed level sensor and feeds it back to the dynamic feeding decision generation stage, forming a complete closed loop of "fish feeding signal collection - analysis - dynamic feeding decision - precise execution - remaining data feedback". This solves the problems of existing technologies that lack a dynamic correction mechanism for feeding direction, lack of feed bin remaining data feedback, unclear correlation logic, and lack of closed loop, making it difficult to adapt to complex breeding scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this embodiment provides a dynamic feeding control method for unmanned feeding vessels based on the analysis of fish feeding behavior signals, including: S1. Fish feeding signal acquisition and synchronous transmission: Vibration sensors, acoustic sensors, near-infrared sensors, turbidity sensors, and water flow sensors are deployed on the bottom of the unmanned feeding vessel to synchronously acquire micro-vibration signals of the water body, feeding and swallowing sound signals, fish activity signals, turbidity parameters, and water flow parameters; using the acoustic signal acquired by the acoustic sensor as a reference, wavelet packet time-frequency interpolation technology is used to align the optical signal acquired by the near-infrared sensor, and the synchronized signal and parameters are encapsulated into a standard data frame and transmitted to the feeding behavior signal analysis stage via the CAN bus; In this step, the process of collecting and synchronously transmitting the fish feeding signals in S1 includes the following steps: S11. Based on the deployment location requirements of the unmanned feeding vessel, the vibration sensor, sound sensor, near-infrared sensor, turbidity sensor and water flow sensor are arranged in a preset symmetrical layout along the central axis of the bottom of the vessel. Specifically, all sensors are fixed to the bottom of the boat with waterproof metal brackets. The connection between the brackets and the bottom of the boat is sealed with sealant (the sealant is a corrosion-resistant sealant specifically for aquaculture) to prevent water from seeping into the hull. At the same time, the height of the brackets can be finely adjusted (the adjustment range is 5-10cm, such as threaded adjustment brackets), ensuring that the sensor probes can be close to the water but not in contact with the bottom of the pool (or the silt at the bottom), avoiding probe wear or obstruction by silt.

[0021] Specifically, the vibration sensor, acoustic sensor, turbidity sensor, and water flow sensor all adopt a "symmetrical layout on both sides of the central axis"—with the central axis as the axis of symmetry, one of each type of sensor is deployed on the left and right sides of the axis (if the hull is wide, two can be deployed on each side to ensure comprehensive signal coverage), and the distance from the left-side sensor to the corresponding right-side sensor to the central axis is equal (e.g., the distance is 20-30cm). The near-infrared sensor is deployed separately directly below the central axis of the hull, with its detection lens pointing vertically into the water. A waterproof and light-transmitting cover (made of corrosion-resistant acrylic, suitable for freshwater and seawater aquaculture environments) is installed on the outside of the lens to avoid the hull structure (such as the keel) from blocking the near-infrared light and to ensure clear capture of fish activity images.

[0022] S12. Using the sensor from S11, it synchronously collects water micro-vibration signals, feeding and swallowing sound signals, fish activity signals, turbidity parameters, and water flow parameters; Specifically, the shipboard central controller (such as the STM32H7 series industrial-grade controller) of the unmanned feeding vessel serves as the synchronization trigger source, outputting pulse synchronization trigger signals with a 1ms interval. These signals are transmitted to the trigger interface of each sensor via a dedicated signal cable. Upon receiving the trigger signal, all sensors immediately initiate a signal acquisition. The vibration sensor's acquisition frequency is set to 100-500Hz (outputting analog signals), the acoustic sensor's acquisition frequency band is set to 100-1000Hz, and the sampling rate is set to 44.1kHz. The near-infrared sensor's acquisition frame rate is set to 30 frames / second, the image resolution is set to 640×480 pixels, and the detection distance is set to 0.5-2m (outputting grayscale images). The turbidity sensor and the water flow sensor (acquiring parameters including water flow velocity and direction, with the water flow direction referenced to the bow direction as 0°) have the same acquisition frequency as the synchronization trigger signal (outputting digital signals). All signals and parameters acquired by the sensors are transmitted to the shipboard data preprocessing module via a dedicated data cable and temporarily stored in the temporary buffer area within the shipboard data preprocessing module, categorized according to the format of "sensor identifier + acquisition timestamp".

[0023] S13. Based on the acoustic signal collected by the acoustic sensor, wavelet packet time-frequency interpolation technology is used to align the optical signal collected by the near-infrared sensor to eliminate the time deviation of multi-sensor data. Specifically, acoustic and optical signals from the same time period are extracted from the buffer of the shipborne data preprocessing module. The acoustic signals are first filtered by a 50Hz notch filter and a 100-1000Hz bandpass filter, and the optical signals are converted into grayscale time-series signals. Using a db4 wavelet basis and a 5-level decomposition wavelet packet transform, time-frequency matrices are obtained by decomposing the two types of signals in the time-frequency domain. Based on the millisecond-level timestamp of the acoustic signals, the cross-correlation coefficient is calculated, and the region with a coefficient lower than 0.9 is identified as the time deviation region. The time-frequency matrix of the optical signals in the deviation region is supplemented with data by linear interpolation, and then the synchronous optical signal is reconstructed by inverse wavelet packet transform. The cross-correlation coefficient is calculated again. If it is ≥0.95, the alignment is qualified; otherwise, the decomposition and interpolation are re-executed.

[0024] S14. Based on the synchronized and aligned signals and parameters, the data is encapsulated into standard data frames using the CAN bus communication protocol and transmitted to S2.

[0025] Specifically, the synchronized signals and parameters are formatted in a unified way, including a 32-bit sensor identifier code (e.g., vibration sensor 0x0001, sound sensor 0x0002), a 64-bit millisecond-level timestamp, and a 32-bit parameter value with unit (e.g., turbidity NTU, water flow velocity m / s), and the data is converted to hexadecimal; the data frame is encapsulated according to the CAN2.0B protocol: 1 bit frame start, 29 bits arbitration field (11 bits standard ID 0x001, 18 bits extended ID including sensor identifier), 2 bits control field (data field length 8 bytes), 8 bytes data field (timestamp high bit → low bit → parameter value high bit → low bit), 15 bits CRC check field, and 2 bits frame end; it is transmitted to the CAN receiving module of the feeding behavior signal analysis stage (i.e., S2) through a twisted-pair CAN bus (rate 500kbps). The CAN receiving module verifies the data. If it is qualified, it is stored in the buffer of S2 for subsequent signal analysis. If it fails, it requests the sending end to retransmit within 10ms.

[0026] It should be added that the sensor in this embodiment adopts the conventional fault degradation strategy of industrial equipment: For a single non-core sensor (such as a water flow sensor) failure: replace it with the average historical data of that sensor stored locally on the device for the past hour; Failure of core sensors (such as sound sensors): Switch to "timed feeding mode" (feeding parameters use the average baseline value of this breeding cycle) and send an alarm to the breeding terminal at the same time; Meanwhile, this embodiment adopts the general redundancy strategy of industrial bus: after three failed retransmissions, the controller switches to local cache mode (temporarily stores data) and executes the pre-stored historical feeding strategy; it attempts to reconnect to the bus every 30 seconds and synchronizes the cached data after recovery.

[0027] S2. Analysis of Feeding Behavior Signals and Calculation of Fish School Hunger State Parameters: Based on the received signals and parameters, the energy ratio of a specific frequency band of feeding and swallowing sounds is calculated as the frequency domain weight, the ratio of transient energy of feeding and swallowing sounds to total energy is calculated as the time domain weight, and the reciprocal of the spatial aggregation degree of the fish school is calculated as the spatial domain weight. Frequency domain sub-coefficients that are reciprocally correlated with turbidity parameters, time domain sub-coefficients that are positively correlated with water flow parameters, and spatial domain sub-coefficients that are positively correlated with the reciprocal of turbidity parameters are calculated respectively. Each weight is coupled with its corresponding dynamic sub-coefficient and normalized to obtain a comprehensive weight. An attention-weighted wavelet packet transform algorithm is used to extract the feeding intensity gradient based on the comprehensive weight, and the continuous fish school hunger state parameters are obtained through quantization mapping. In S2, the process of analyzing feeding behavior signals and calculating fish school hunger state parameters includes the following steps: In this step, S21, based on the collected signal, the energy proportion of a specific frequency band of the swallowing sound is calculated as the frequency domain weight. Calculate the ratio of the transient energy of the swallowing sound to the total energy as the time-domain weight. Calculate the reciprocal of the spatial clustering degree of the fish school as the spatial weight. In step S21, the process of calculating weights based on the acquired signals includes the following steps: S21.1 Extract the feeding and swallowing sound signal transmitted in S1, and retain the characteristic frequency band of fish feeding by bandpass filtering. Calculate the frequency domain weight by the proportion of energy in this frequency band to the total energy of the entire frequency band. ; Specifically, the feeding and swallowing sound signal transmitted by S1 is extracted. Based on the feeding and swallowing sound characteristics of the target farmed fish, a preset bandpass filter frequency band is established (e.g., 200-800Hz for freshwater farmed fish). After bandpass filtering the signal, the frequency domain weights are calculated. : ; in, This represents the frequency domain weight, reflecting the energy proportion of the characteristic frequency bands of swallowing sounds. This represents the energy of the swallowing sound signal after bandpass filtering (preserving the characteristic frequency bands of fish feeding); It represents the full-frequency energy of the original ingestion and swallowing sound signal.

[0028] S21.2 Perform time-domain analysis on the swallowing sound signal to define the transient interval reflecting the eating action, and calculate the time-domain weight by the ratio of the energy in this interval to the energy of the entire frequency band. ; Specifically, time-domain analysis was performed on the swallowing sound signal. A signal amplitude threshold was set (the threshold is 1.5 times the average amplitude of the signal across the entire frequency band). Continuous signal segments with amplitudes exceeding this threshold were defined as the "transient interval of the feeding action," and time-domain weights were calculated. : ; in, This represents the time-domain weight, reflecting the energy proportion during the transient interval of the feeding action; This represents the energy of the swallowing sound signal during the transient interval of the feeding action; It represents the full-frequency energy of the original ingestion and swallowing sound signal.

[0029] S21.3 Receive the near-infrared fish activity image transmitted by S1, extract the spatial distribution information of the fish school after denoising and target segmentation, and calculate the spatial weight based on the spatial clustering degree of the fish school. .

[0030] Specifically, the near-infrared fish activity images transmitted by S1 are received, and after Gaussian filtering for noise reduction, the fish targets are extracted using the Otsu adaptive threshold segmentation algorithm to obtain the spatial distribution information of the fish school. The spatial clustering degree of the fish school is calculated using the ratio of the area of ​​the minimum enclosing rectangle of the fish school to the area of ​​the effective region of the image. : ; in, This represents the spatial weight, reflecting the degree of spatial dispersion of the fish school (the inverse of the aggregation degree). Represents the area of ​​the smallest rectangle that contains all the fish targets; This represents the area of ​​the image region within the detection range of the near-infrared sensor.

[0031] In this step, S22, based on the acquired parameters, calculate the corresponding frequency domain weights. Time domain weights Spatial weights Frequency domain sub-coefficients Time-domain sub-coefficients Spatial sub-coefficients In step S22, the process of calculating the dynamic sub-coefficients based on the collected parameters includes the following steps: S22.1 Turbidity parameters based on S1 Calculate the frequency domain sub-coefficients based on the reciprocal correlation. To correct the interference of turbidity on acoustic signal feature extraction; Specifically, based on the turbidity parameters collected by S1 Calculate the frequency domain sub-coefficients according to the "reciprocal correlation" relationship. Correcting the interference of turbidity on acoustic signal feature extraction: ; in, Represents the frequency domain sub-coefficients used to compensate for turbidity. Interference; This represents the temporary normalization coefficient (which will be used later for the unified normalization of the overall weights). This represents the turbidity parameter of the water sampled by S1, in NTU.

[0032] S22.2, Water flow parameters collected based on S1 Calculate time-domain sub-coefficients based on positive correlation. The influence of water flow velocity on the timing characteristics of fish feeding behavior; Specifically, based on the water flow parameters collected by S1 Calculate time-domain sub-coefficients based on "positive correlation" relationship The influence of water flow velocity on the timing characteristics of feeding actions: ; in, Represents time-domain sub-coefficients used for matching water flow pairs The impact; This represents the temporary normalization coefficient (which will be used later for the unified normalization of the overall weights). This represents the water flow velocity parameters collected by S1.

[0033] S22.3, Based on turbidity parameters The spatial sub-coefficients are calculated based on the positive correlation of their reciprocals. To compensate for the impact of turbidity on the accuracy of near-infrared image recognition.

[0034] Specifically, based on the turbidity parameters collected by S1 Spatial sub-coefficients are calculated based on the positive correlation between the reciprocal of turbidity. The impact of turbidity compensation on near-infrared image recognition accuracy: ; in, Represents the spatial sub-coefficients used to compensate for turbidity pairs. Interference; This represents the temporary normalization coefficient (which will be used later for the unified normalization of the overall weights). This represents the turbidity parameter of the water sampled by S1, in NTU.

[0035] In this step, S23, the frequency domain weights are... With frequency domain sub-coefficients Time-domain weights and time-domain sub-coefficients Spatial weights With spatial sub-coefficients The components are merged separately to obtain the overall weight. In S23, the comprehensive weight The calculation process includes the following steps: S23.1, Frequency domain weights With frequency domain sub-coefficients Time domain weights With time-domain sub-coefficients Spatial weights With spatial sub-coefficients Weighted coupling is performed separately to obtain weighted components in three frequency-temporal-spatial dimensions; Specifically, each weight is multiplied by its corresponding dynamic sub-coefficient to obtain weighted components in three frequency-time-space dimensions: ; in, Represents the frequency domain weighted components, for and The coupling result; ; in, Representing the time-domain weighted components, as and The coupling result; ; in, Represents the spatial weighted components, for and The coupling result.

[0036] S23.2. Normalize the three weighted components to eliminate the dimensional differences between parameters of different dimensions, and obtain the comprehensive weight. .

[0037] Specifically, the three weighted components are normalized to eliminate the dimensional differences between different dimensions, resulting in a comprehensive weight: ; in, This represents the comprehensive weight, which integrates the feeding feature weights across the frequency, time, and space dimensions, and its value ranges from [0,1]. express The maximum value in; when (When there is no effective feeding signal) 0.

[0038] In this step, S24, the attention-weighted wavelet packet transform algorithm is used, based on the comprehensive weights. Extracting the feeding intensity gradient In step S24, the process of extracting the feeding intensity gradient includes the following steps: S24.1 Perform wavelet packet time-frequency decomposition on the swallowing sound signal to obtain multi-band sub-signals covering the characteristic frequency band of swallowing; Specifically, using the db4 wavelet basis function and setting the decomposition level to 5, wavelet packet time-frequency decomposition was performed on the swallowing sound signal to obtain a multi-band sub-signal covering the characteristic frequency band of swallowing (denoted as...). , (This represents the number of sub-signals in the decomposed frequency band).

[0039] S24.2, Based on comprehensive weights Assign attention coefficients to sub-signals in each frequency band Strengthen the signal weights of frequency bands related to feeding behavior; Specifically, based on comprehensive weighting For each frequency band sub-signal Assign attention coefficient Strengthen the signal weighting of frequency bands related to feeding behavior: ; in, Indicates the first Attention coefficients of each frequency band sub-signal; Indicates the overall weight; Indicates the first The initial weights of each frequency band sub-signal are determined by the matching degree between the frequency band sub-signal and the feeding characteristic frequency band; the higher the matching degree, the higher the weights. The larger (the larger).

[0040] S24.3, Based on Attention Coefficient Weighted reconstruction of multi-band sub-signals is performed, and the rate of change of energy in the reconstructed signal is extracted to obtain the feeding intensity gradient, which reflects feeding activity. .

[0041] Specifically, based on attention coefficient The reconstructed signal is obtained by weighting and reconstructing the multi-band sub-signals. ; Calculate the reconstructed signal The rate of energy change, as a gradient of feeding intensity. : ; in, This represents the feeding intensity gradient, reflecting the changing trend of feeding activity. This represents the energy of the reconstructed signal at the current sampling moment; This represents the energy of the reconstructed signal at the previous sampling time.

[0042] In this step, S25, based on the feeding intensity gradient The parameters of the continuous fish population's starvation state were calculated. In step S25, the continuous fish population starvation state parameters are calculated. The process includes the following steps: S25.1, Call the feeding intensity gradient output from S24.3 The overall weights output from S23.2 Variety characteristic parameters were extracted simultaneously from the sound signals of feeding and swallowing. The This is obtained by analyzing the characteristic frequency band distribution width and peak interval patterns of the signal, and is used to match the differences in feeding behavior among different fish species; Specifically, the feeding intensity gradient output by S24.3 is called. S23.2 Output of the comprehensive weight Simultaneously analyze the "characteristic frequency band distribution width" and "peak interval pattern" of the feeding and swallowing sound signals, and match them with a preset fish species feature database (e.g., grass carp corresponds to 1.0, crucian carp corresponds to 1.0). =0.9), thus obtaining the varietal characteristic parameters. (The value range is [0.8, 1.2]).

[0043] S25.2, Gradient of feeding intensity Perform moving average processing based on variety characteristic parameters. By dynamically adjusting the sliding window parameters, a stable gradient value is obtained after eliminating instantaneous signal fluctuations. ; Specifically, regarding the feeding intensity gradient Perform a moving average process to eliminate instantaneous signal fluctuations and obtain a stable gradient value. The sliding window length is determined by the variety characteristic parameters. Dynamic adjustment: ; in, Represents the floor function; ; in, Indicates the number of sampling points for the moving average; Indicates the first [number]th ... The feeding intensity gradient at each sampling time.

[0044] S25.3, Based on comprehensive weights , Stabilization gradient and variety characteristic parameters Construct association logic to obtain initial hunger parameters through the mapping relationship between signal features and hunger state. ; Specifically, based on comprehensive weighting , Stabilization gradient and variety characteristic parameters Construct the association logic and calculate the initial starvation parameters. : .

[0045] S25.4, to Standardization corrections are made when When the value exceeds the [0,1] interval, it is truncated according to the boundary value, and the final output is a continuous fish swarm hunger state parameter. .

[0046] Specifically, regarding the initial starvation parameters After standardization correction to ensure it falls within the [0,1] interval, the final continuous fish population hunger state parameters are obtained: ; in, This parameter represents the continuous hunger state of a fish population, and its value ranges from [0,1] (the larger the value, the higher the degree of hunger in the fish population). This means that the lower limit of the parameter is limited to 0; This indicates that the upper limit of the parameter is limited to 1.

[0047] S3. Dynamic Feeding Decision Generation: Based on the average body weight of fish in the current breeding cycle, the baseline feeding amount is calculated. The feeding parameters are adjusted according to the continuous fish population hunger state parameters. The feeding amount increases as the continuous fish population hunger state parameters increase and decreases as the continuous fish population hunger state parameters decrease. The feeding interval decreases as the continuous fish population hunger state parameters increase and increases as the continuous fish population hunger state parameters decrease. Combining the real-time GPS positioning signal of the unmanned feeding vessel with the coordinates of the preset target feeding area, the offset between the vessel and the target feeding area is calculated, and the feeding direction is corrected to form a feeding control strategy that includes parameters such as feeding amount, feeding interval, and feeding direction. In this step, the dynamic feeding decision generation process in S3 includes the following steps: S31. Receive the average body weight data of fish in the current breeding cycle, and calculate the benchmark feeding amount to meet the basic nutritional needs of farmed fish in combination with the characteristics of the growth stage of farmed fish. Specifically, S31 calculates the baseline feeding amount to meet the basic nutritional needs based on the average body weight and growth stage characteristics of the fish during the current breeding cycle, using the following formula: ; in, Indicates the baseline feeding amount (unit: g); This indicates the average body weight of the fish during the current breeding cycle (unit: g / fish). This indicates the number of fish in the aquaculture area (for methods of counting fish, please refer to "Methods for Counting the Number of Fish in a Fishpond" - Baidu Wenku). This indicates the feeding rate for the corresponding fish growth stage (preset based on the growth stage).

[0048] S32. Call the continuous fish swarm hunger state parameters output by S2. Based on the baseline feeding amount, a continuous fish population starvation status parameter was established. The logic relating the feeding amount and feeding interval is used to implement the feeding amount based on the continuous fish population hunger state parameters. Increases with the rise in continuous fish population hunger parameters The feeding interval is adjusted in a reverse linkage with the reduction. Specifically, S32 calls the continuous fish swarm hunger state parameters output by S2. Based on the baseline feeding amount, a logical association is established between the feeding amount and the feeding interval, as follows: First, calculate the adjusted feeding amount: ; in, This indicates the adjusted actual feeding amount (unit: g). This represents the baseline feeding amount calculated by S31; This represents the continuous fish population hunger state parameter output by S2. The value range is [0,1]), which enables the feeding amount to be adjusted accordingly. Rise and rise, follow It decreases as it decreases.

[0049] Then, the adjusted feeding interval is calculated. Preset baseline feeding interval (Based on the pre-set parameters for the aquaculture species, such as 30 seconds per feeding), the adjusted feeding interval is as follows: ; in, Indicates the adjusted actual feeding interval (unit: s); Indicates the baseline feeding interval; This parameter represents the continuous hunger state of a fish population, allowing for adjustments to the feeding interval. Increase and shorten, follow It decreases and then prolongs.

[0050] S33. Receive the real-time GPS positioning signal and preset feeding area coordinate information of the unmanned feeding vessel, calculate the offset between the current position of the vessel and the target feeding area, and output the correction angle of the sprinkling direction. Specifically, S33 calculates the offset and outputs the corrected angle for the feeding direction based on the real-time GPS positioning of the unmanned feeding vessel and the coordinates of the preset feeding area, as follows: Coordinate acquisition: The real-time GPS coordinates of the unmanned feeding boat are (Coordinate values ​​in a Cartesian coordinate system); The center coordinates of the preset target feeding area are (In the same coordinate system as the ship's hull).

[0051] Offset angle calculation: Calculate the relative offset angle between the ship's hull and the center of the target feeding area: ; in, This indicates the offset angle of the ship's hull relative to the center of the target feeding area (unit: °, value range [-180°, 180°]). It is a four-quadrant arctangent function used to determine the quadrant of the offset angle.

[0052] Correction angle for the direction of spraying: Obtain the current heading angle of the unmanned feeding vessel. (Data collected by shipboard attitude sensors, with true north as the reference point), the corrected angle for the spraying direction is: ; in, Indicates the angle (unit: °) that needs to be adjusted in the direction of the spray; if If the value exceeds [-180°, 180°], adjust it to within that range by adding or subtracting 360°.

[0053] S34. Simultaneously receive the baseline feeding amount, adjusted feeding parameters, and scattering direction correction angle, integrate them into a standardized control command, and transmit it to S4.

[0054] Specifically, S34 synchronously receives the baseline feeding amount from S31, the adjusted feeding parameters from S32, and the correction angle for the scattering direction from S33. They are then integrated into standardized control commands, and the specific operation is as follows: Command field definition: Control commands contain three core parameter fields—"Actual feeding amount" "Actual feeding interval" "Direction of spraying correction angle" ”; Encapsulation format: The data is encapsulated into a data frame according to the control bus protocol of the unmanned feeding vessel (such as the CAN2.0B protocol consistent with S1). The frame contains the numerical values ​​of each parameter (with 1 decimal place) and the unit identifier in sequence. Transmission: The encapsulated control commands are transmitted to S4 (precision feeding execution stage) via the bus to drive the feeding execution mechanism.

[0055] S4. Precise feeding execution and silo status feedback: Based on the feeding control strategy, the feeding speed is adjusted by the stepper motor feeding mechanism to control the feeding amount, the feeding range is controlled by the direction and opening and closing of the electric spray nozzle, and the silo level sensor collects the silo level data and feeds it back to S3.

[0056] In this step, the process of precise feeding execution and silo status feedback in S4 includes the following steps: S41. Receive the standardized control command output by S3 and parse it to obtain the target feeding amount, feeding interval and scattering direction correction angle; Specifically, S41 receives the standardized control commands output by S3, parses the commands to obtain execution parameters, and the specific operation is as follows: Receive control command data frames (following the CAN2.0B protocol) transmitted by S3 via the CAN bus, and parse them according to the preset command field mapping relationship: Data frame bytes 1-4: Parsed as target feed amount (Unit: g); Data frame bytes 5-6: Parsed as feeding interval (Unit: s); Data frame bytes 7-8: Parsed as the spray direction correction angle (Unit: °); The parsed parameters are stored in the temporary buffer of the feeding execution module (the feeding execution module is a hardware control unit integrated on the unmanned feeding vessel, which communicates with the S3's stepper motor feeding mechanism, electric spray nozzles, material level sensors, and other components, and is responsible for parsing control commands, outputting execution drive signals, and collecting status data), as the basis for subsequent execution actions.

[0057] S42. Based on the analyzed target feeding amount, the feeding speed is adjusted through the stepper motor feeding mechanism to achieve precise control of the feeding amount and ensure consistency with the decision command. Specifically, the control unit of the unmanned feeding vessel, based on the analyzed target feeding amount, adjusts the speed of the stepper motor feeding mechanism to achieve precise feeding. The specific operation and formula are as follows: The "feeding coefficient per unit speed" of the stepper motor feeding mechanism is... (Unit: g / r, determined by inherent structural parameters such as the feed trough pitch and hopper outlet size of the feeding mechanism), feed rate per unit time. (Unit: g / min) and motor speed The correlation (unit: r / min) is as follows: ; To be at the feeding interval Complete the target feeding amount within (unit: seconds) (Unit: g), the required feed rate per unit time is: ; By combining the two equations, we can obtain the target speed of the stepper motor: ; in, The target speed of the stepper motor (unit: r / min); The target feeding amount obtained from the analysis; The feeding interval is obtained from the analysis; This is the feeding coefficient per unit rotational speed of the feeding mechanism.

[0058] The control unit outputs a PWM control signal to the stepper motor, driving the feeding mechanism to run at the target speed until the target feeding amount for a single cycle is completed and then the feeding stops.

[0059] It is understandable that the feeding coefficient per unit speed of the stepper motor feeding mechanism is... , The calibration is based on the actual parameters of the feed trough structure and feed characteristics. The calibration formula is: ; in, Indicates the feed trough pitch (the industry standard feed trough pitch is 2-3 cm / r); Indicates the cross-sectional area of ​​the discharge port (determined by the discharge port diameter); This indicates the bulk density of the feed (0.3-0.5 g / cm³ for extruded feed and 0.6-0.8 g / cm³ for pelleted feed; both are publicly available industry parameters). This indicates the feeding efficiency (0.92-0.95 when the inner wall of the trough is made of a smooth material, which is the typical efficiency range of the actual feeding mechanism).

[0060] S43. Adjust the angle according to the direction of spraying, control the rotation and opening degree of the electric spraying head, adjust the feeding coverage range, and match the preset feeding area; Specifically, the control unit of the unmanned feeding vessel adjusts the angle according to the direction of the feeding, controls the direction and opening / closing degree of the electric feeding nozzles, and matches the preset feeding area. The specific operation is as follows: Nozzle steering control: The control unit acquires the current heading angle from the ship's attitude sensors. (Unit: °, with true north as the reference), combined with the angle correction obtained from the analysis of the spraying direction. Calculate the nozzle target turning angle: ; in, The target steering angle (unit: °) for the electric spray nozzle; the control unit sends an angle drive signal to the nozzle steering motor to adjust the nozzle steering to... .

[0061] Nozzle opening and closing degree control: The inherent maximum coverage radius of the electric spray nozzle is (Unit: m, determined by nozzle structure parameters), the control unit calculates the feed area radius based on the preset feed area radius. (Unit: m), Calculate the nozzle opening / closing degree: ; in, The nozzle opening / closing degree (unit: %); the control unit outputs an electrical signal to the nozzle opening / closing control mechanism to adjust the opening / closing degree to... This ensures that the feeding coverage area matches the preset feeding area.

[0062] S44. Collect the remaining material data in the silo in real time through the material level sensor, and feed the remaining information back to S3 to provide material status support for the dynamic adjustment of feeding parameters.

[0063] Specifically, the control unit of the unmanned feeding vessel collects the remaining amount of feed in the hopper through a level sensor and feeds it back to S3. The specific operation is as follows: Inventory of remaining material in silos: Infrared level sensors are installed along the height direction on the inner wall of the silo. The control unit reads the sensor data in real time to obtain the real-time height of the feed in the silo. (Unit: m); Total height of the silo is (Unit: m, determined by silo structure), the silo clearance percentage is calculated as follows: ; in, This represents the percentage of remaining feed in the silo (unit: %). Furthermore, the above formula only applies to regular cylindrical silos (volume is proportional to height, and this is the most commonly used silo type in aquaculture equipment). Balance information feedback: The control unit will determine the proportion of remaining material in the hopper. Encapsulated as a CAN data frame (frame ID set to 0x002, identifying the hopper status signal), it is transmitted to the dynamic feeding decision-making stage of S3 via the CAN bus; when When the feed rate is below a preset threshold (e.g., 10%), S3 can dynamically adjust the feeding parameters based on this feedback (e.g., appropriately reduce the amount of feed per feeding) to avoid empty silos.

[0064] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0065] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic feeding control method for an unmanned feeding ship based on fish feeding behavior signal analysis, characterized in that, It comprises the following steps: S1, fish feeding signal collection and synchronous transmission: deploying vibration sensors, acoustic sensors, near-infrared sensors, turbidity sensors and flow sensors at the bottom of the unmanned feeding ship, synchronously collecting water micro-vibration signals, feeding swallowing sound signals, fish activity signals, turbidity parameters and flow parameters; taking the acoustic signals collected by the acoustic sensor as the reference, using wavelet packet time-frequency interpolation technology to align the optical signals collected by the near-infrared sensor, packaging the synchronized signals and parameters into standard data frames, and transmitting them to the feeding behavior signal analysis link through the CAN bus; S2, feeding behavior signal analysis and fish population hunger state parameter calculation: based on the received signals and parameters, calculating the specific frequency band energy proportion of the feeding swallowing sound as the frequency domain weight, the ratio of the feeding swallowing sound transient energy to the total energy as the time domain weight, and the inverse of the fish population spatial aggregation degree as the space domain weight; respectively calculating the frequency domain sub-coefficient inversely related to the turbidity parameter, the time domain sub-coefficient positively related to the flow parameter, and the space domain sub-coefficient positively related to the inverse of the turbidity parameter; coupling the weights with the corresponding dynamic sub-coefficients and normalizing to obtain the comprehensive weight, using the attention weighted wavelet packet transform algorithm to extract the feeding intensity gradient based on the comprehensive weight, and quantizing and mapping to obtain the continuous fish population hunger state parameter; S3, dynamic feeding decision generation: calculating the reference feeding amount based on the average body weight of the fish in the current breeding cycle, adjusting the feeding parameters according to the continuous fish population hunger state parameter, wherein the feeding amount increases as the continuous fish population hunger state parameter increases, and decreases as the continuous fish population hunger state parameter decreases, the feeding interval shortens as the continuous fish population hunger state parameter increases, and lengthens as the continuous fish population hunger state parameter decreases; combining the real-time GPS positioning signal of the unmanned feeding ship with the preset target feeding area coordinates, calculating the offset between the ship body and the target feeding area, correcting the feeding throwing direction, and forming a feeding control strategy including the feeding amount, feeding interval and throwing direction parameters; S4, precise feeding execution and bin state feedback: based on the feeding control strategy, adjusting the feeding speed through the stepping motor feeding mechanism to control the feeding amount, controlling the feeding range through the steering and opening degree of the electric throwing nozzle, and collecting the bin remaining data through the level sensor, and feeding the bin remaining data to S3.

2. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 1, characterized in that, In the S1, the process of fish feeding signal collection and synchronous transmission comprises the following steps: S11, based on the deployment location requirements of the unmanned feeding ship, arranging the vibration sensors, acoustic sensors, near-infrared sensors, turbidity sensors and flow sensors along the central axis of the ship bottom in a preset symmetrical layout; S12, using the sensors of S11, synchronously collecting water micro-vibration signals, feeding swallowing sound signals, fish activity signals, turbidity parameters and flow parameters; S13, based on the acoustic signals collected by the acoustic sensor, using wavelet packet time-frequency interpolation technology to align the optical signals collected by the near-infrared sensor, and eliminating the time deviation of multi-sensor data; S14, based on the synchronized and aligned signals and parameters, using the CAN bus communication protocol to package the data into standard data frames and transmit them to S2.

3. The dynamic feeding control method of the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 2, characterized in that, In the S2, the process of feeding behavior signal analysis and fish population hunger state parameter calculation includes the following steps: S21、based on the collected signal, calculate the specific frequency band energy proportion of the feeding swallowing sound as the frequency domain weight , calculate the ratio of the transient energy of the feeding swallowing sound to the total energy as the time domain weight , calculate the reciprocal of the spatial aggregation degree of the fish school as the space domain weight ; S22, based on the collected parameters, respectively calculate corresponding frequency domain weights , time domain weights , spatial domain weights frequency domain sub-coefficients time domain sub-coefficients spatial domain sub-coefficients ; S23, the frequency domain weight with the frequency domain sub-coefficients , the time domain weight and the time domain sub-coefficients , the spatial domain weight with the spatial domain sub-coefficients respectively, to obtain a comprehensive weight ; S24, using attention weighted wavelet packet transform algorithm, based on the comprehensive weight extracting the feeding intensity gradient ; S25、based on the feeding intensity gradient , calculate the continuous fish population hunger state parameter .

4. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 3, characterized in that, In the S21, the process of calculating the weight based on the collected signals includes the following steps: S21.1, extract the feeding swallowing sound signal transmitted in S1, reserve the frequency band characteristic of fish feeding through band-pass filtering, calculate the frequency domain weight through the proportion of energy in the frequency band to the energy in the whole frequency band ; S21.2, time domain analysis is performed on the feeding swallowing sound signal, a transient interval reflecting the feeding action is defined, and a time domain weight is calculated through the ratio of the energy in the interval to the energy in the full frequency band ; S21.3, receiving the near-infrared fish activity image transmitted in S1, extracting fish spatial distribution information through denoising and target segmentation, and calculating the spatial weight based on the spatial aggregation degree of the fish school .

5. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 4, characterized in that, In the S22, the process of calculating the dynamic sub-coefficient based on the collected parameters includes the following steps: S22.1, based on the turbidity parameter collected in S1 , calculating the frequency domain sub-coefficients according to the inverse correlation relationship , correcting the interference of turbidity on the extraction of acoustic signal features; S22.2, based on the water flow parameters collected in S1 , calculating the time-domain sub-coefficients in positive correlation , matching the influence of water flow velocity on the feeding action timing characteristics of fish S22.3, based on the turbidity parameter , the inverse of the positive correlation relationship calculation of spatial sub-coefficient , compensation turbidity on near-infrared image recognition accuracy.

6. The dynamic feeding control method of the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 5, characterized in that, In the S23, the comprehensive weight The calculation process of the comprehensive weight includes the following steps: S23.1, the frequency domain weight with the frequency domain sub-coefficients , the time domain weight with the time domain sub-coefficients , the spatial domain weight with the spatial domain sub-coefficients are respectively weighted and coupled to obtain three weighted components of the frequency-time-space dimensions; S23.2, normalizing the three weighted components to eliminate the dimensional difference of different dimension parameters to obtain a comprehensive weight .

7. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 6, characterized in that, In the S24, the process of extracting the feeding intensity gradient includes the following steps: S24.1, wavelet packet time-frequency decomposition is performed on the feeding swallowing sound signal to obtain a multi-band sub-signal covering the feeding characteristic frequency band; S24.2, based on the integrated weight assigning attention coefficients to the frequency band sub-signals , reinforcing the signal weight of the frequency band related to the feeding behavior S24.3, based on attention coefficient The multi-band sub-signals are weighted and reconstructed, the energy change rate of the reconstructed signal is extracted, and the feeding intensity gradient reflecting the feeding activity is obtained .

8. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 7, characterized in that, In the S25, a continuous fish school hunger state parameter is calculated The process includes the following steps: S25.1, calling the feeding intensity gradient output by S24.3 , integrating the comprehensive weight output by S23.2 , synchronously extracting the variety characteristic parameters of the feeding swallowing sound signals ; the method further comprises the following steps: The characteristic frequency band distribution width and peak interval law of the signal are analyzed to match the feeding behavior differences of different fish species. S25.2, to the feeding intensity gradient The moving average processing is performed according to the variety characteristic parameters The dynamic adjustment of the sliding window parameters eliminates the transient signal fluctuations to obtain the smoothed gradient value ; S25.3, based on the comprehensive weight , smoothing gradient and variety characteristic parameters Construct the correlation logic, get the initial starvation parameter through the mapping relationship between signal characteristics and starvation state ; S25.4, to perform a standardization correction, when clipped to the boundary values if outside the interval [0, 1], and finally output a continuous fish hunger state parameter .

9. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 8, characterized in that, In the S3, the process of dynamic feeding decision generation includes the following steps: S31, receiving the average body weight data of the fish in the current breeding cycle, and combining the growth stage characteristics of the farmed fish, the baseline feeding amount that meets the basic nutritional needs is calculated; S32, calling the continuous fish population hunger state parameter output by S2 , on the basis of the reference feeding amount, establishing the continuous fish population hunger state parameter , the correlation logic of the feeding amount and the feeding interval, realizing the increase of the feeding amount with the continuous fish population hunger state parameter , the decrease of the feeding amount with the continuous fish population hunger state parameter , and the reverse linkage adjustment of the feeding interval; S33, receiving the real-time GPS positioning signal of the unmanned feeding ship and the preset feeding area coordinate information, calculating the offset between the current position of the ship body and the target feeding area, and outputting the correction angle of the throwing direction; S34, synchronously receiving the baseline feeding amount, the adjusted feeding parameters and the correction angle of the throwing direction, integrating them into standardized control instructions, and transmitting them to S4.

10. The dynamic feeding control method for the unmanned feeding ship based on the fish school feeding behavior signal analysis according to claim 9, characterized in that, In the S4, the process of precise feeding execution and silo state feedback includes the following steps: S41, receiving the standardized control instructions output by S3 to obtain the target feeding amount, feeding interval and correction angle of the throwing direction; S42, based on the analyzed target feeding amount, the feeding speed of the stepping motor feeding mechanism is adjusted; S43, according to the correction angle of the throwing direction, the steering and opening degree of the electric throwing nozzle are controlled; S44, real-time collection of silo remaining amount data is performed through the level sensor, and the remaining amount information is fed back to S3.

Images