Feeding amount decision-making method for fish farming based on acoustic information and water surface fluctuation information

Abstract

A method for determining fish feeding amounts based on acoustic and water surface fluctuation information divides fish feeding activity into four levels. The method collects the sounds of the fish feeding to obtain a data set of the fish feeding sound, and simultaneously collects the fluctuations in the aquaculture water surface to obtain a data set of water surface fluctuation information. The fish feeding process is videotaped during feeding. The data set is preprocessed, and eigenvalues are extracted from the preprocessed information. Principal component analysis is performed on the extracted eigenvalues to obtain a eigenvector that fuses the acoustic signals of the fish feeding and the water surface fluctuation signals. A classifier identifies the data in the eigenvector, classifies the fish feeding status into different levels of feeding activity, and issues work instructions based on the determined feeding activity. This method determines the feeding activity of the fish school by recognizing acoustic and water surface fluctuation information, and thus determines the fish's hunger level, providing technical support for accurate feeding of feeders.

Description

Technical Field

[0001] The present invention relates to the technical field of fish farming, and in particular to a method for deciding a feeding amount for fish farming based on acoustic information and water surface fluctuation information. Background Art

[0002] When raising fish, the feeding activity of fish is an important indicator of their hunger level. The higher the feeding activity, the hungrier the fish. Using changes in the activity of fish during feeding as an indicator, the operation of feeders can be more accurately controlled, feeding on demand and avoiding underfeeding or overfeeding during the feeding process. The method based on acoustic information recognition can identify the feeding activity of fish. This method can avoid the influence of light and turbidity, but is susceptible to the noise in the breeding environment. Although noise reduction can be performed after collection, when large aerators and sewage dischargers are turned on and the fish are feeding weakly, it is still difficult to accurately detect the feeding status of the fish after noise reduction. The water surface fluctuation information obtained by the tilt angle sensor is more stable and can avoid the influence of noise. Summary of the Invention

[0003] The present invention provides a method for deciding the feeding amount for fish farming based on acoustic information and water surface fluctuation information. The method determines the feeding activity of fish schools by identifying acoustic information and water surface fluctuation information, and then determines the hunger level of fish schools, providing technical support for feed feeding machines to achieve precise feeding.

[0004] The technical solution adopted by the present invention is:

[0005] The method for determining the feeding amount for fish farming based on acoustic information and water surface fluctuation information includes the following steps:

[0006] Step 1: Classify the feeding activity of the fish school into four levels: relatively active feeding, active feeding, inactive feeding, and no feeding;

[0007] Step 2: Use a hydrophone to collect the sounds of the fish feeding, obtaining a data set of the fish feeding sounds. Use a tilt angle sensor to simultaneously collect the fluctuations of the aquaculture water surface, obtaining a data set of water surface fluctuation information. Videos of the fish feeding process are taken during feeding.

[0008] Step 3: Preprocess the data set of fish feeding sound and water surface fluctuation information obtained in step 2;

[0009] Step 4: Extract eigenvalues from the sound data and water surface fluctuation information preprocessed in step 3, and perform principal component analysis on the extracted eigenvalues. Through principal component analysis, a eigenvector of the fusion of the sound signal and the water surface fluctuation signal when the fish are feeding is obtained;

[0010] Step 5: Import the feature vector obtained in step 4 into the classifier. The classifier identifies the data in the feature vector, classifies the feeding status of the fish into different feeding activities, and issues work instructions based on the determined feeding activities.

[0011] In step 1, active feeding means that the fish quickly gather in the feeding area and scramble for food;

[0012] Active feeding means that fish swim to the feeding area and eat, and the aggregation degree decreases;

[0013] Inactive feeding means: fish eat slowly but there is no remaining bait;

[0014] Not feeding means that the fish no longer eat and there is obvious remaining bait.

[0015] The second step specifically includes:

[0016] First, the system uses a Song Meter SM2+ acoustic recorder and two HTI-96-MIN standard hydrophones to collect sound. One standard hydrophone is used to capture the sounds of the fish feeding, while the other is used to collect the sounds of the same fishless environment to serve as background noise samples. Simultaneously, a digital tilt angle sensor is used to detect the frequency of water surface fluctuations and collect target data.

[0017] When recording the sound information and water surface fluctuation information of the fish school feeding, the video of the fish school feeding process during feeding is shot synchronously as a reference for the feeding activity of the fish school.

[0018] The step three comprises the following steps:

[0019] S3.1: Using the fish-free background noise samples collected in step 2, noise reduction is performed on the feeding sound signals of the school of fish;

[0020] S3.2: removing singular values from the target data information collected by the tilt angle sensor;

[0021] S3.3: Based on the video footage of the school feeding, distinguish the obtained fish feeding sound signal segments and water surface fluctuation signal segments from a single feeding process, and obtain the fish feeding sound signal and water surface fluctuation signal of the single feeding process;

[0022] S3.4: Based on the feeding patterns of fish in the videos captured during feeding, the acoustic signatures of active feeding and water surface fluctuations during a single feeding bout are distinguished, categorized as relatively active feeding bouts, active feeding bouts, inactive feeding bouts, and no feeding bouts. This provides a reference for determining the data range for the relevant acoustic signatures during the feeding process.

[0023] The step 4 includes the following steps:

[0024] S4.1: Extract the time domain features of the audio data when the fish are feeding - short-term energy:

[0025] Assume that the speech waveform time domain signal is x(n), the window function is w(n), and the i-th frame speech signal obtained after frame processing is y i (n), then y i (n) Satisfy:

[0026] y i (n)=w(n)*x((i-1)*inc+n),1≤n≤L,1≤i≤fn

[0027] Where w(n) is the window function, usually a rectangular window or a Hanning window; y i (n) is the value of a frame, n = 1, 2, ... L, i = 1, 2, ... fn, L is the frame length; inc is the frame shift length; fn is the total number of frames after framing; the formula for calculating the short-time energy of the i-th frame speech signal yi(n) is:

[0028]

[0029] Where: E(i) is the short-term energy of the i-th frame;

[0030] S4.2: Extracting frequency domain features of audio data from fish feeding — Wavelet packet decomposition:

[0031] Based on the principle of extracting multi-scale spatial energy features through wavelet packet decomposition, the signal energy at different decomposition scales is solved and these energy values are arranged in scale order to form feature vectors for identification. If a 4-layer wavelet packet decomposition is performed, a 16-dimensional frequency band energy feature can be obtained, and its calculation formula is as follows:

[0032] E(j,i)=∑ k∈z [p s (n,j,k)] 2

[0033] Where: E(j,i) represents the energy value of each node i at decomposition level j; k = 1, 2, ..., n, Z represents an integer; Ps(n, j, k) is the wavelet packet decomposition coefficient;

[0034] Each dimension has a frequency band energy, so after wavelet packet decomposition, 16 frequency band energy eigenvalues are obtained.

[0035] S4.3: Extracting the short-time energy of the water surface fluctuation information from the information measured by the tilt angle sensor; the extraction method is the same as the short-time energy feature extraction in step S4.1;

[0036] S4.4: Perform principal component analysis on the 18 eigenvalues of the short-time energy of the sound information extracted in steps S1), S2, and S3), the energy of the 16-dimensional frequency bands after wavelet packet decomposition, and the short-time energy of the water surface fluctuation information. The mathematical model of principal component analysis is as follows:

[0037]

[0038] Among them, u jp ———standardization coefficient;

[0039] x ip ———the pth normalized feature of the i-th underwater acoustic signal sample;

[0040] z ij ———the jth principal component variable of the i-th underwater acoustic signal sample;

[0041] The characteristic vector of the fusion of the acoustic signal and the water surface fluctuation signal when the fish school is feeding is obtained and used to identify and analyze the feeding activity level.

[0042] In step 5, the obtained feature vector is imported into the classifier, and the classifier further classifies and judges it and the controller issues a working instruction, including the following steps:

[0043] S5.1: The feature vector is imported into the classifier. The classifier identifies the data in the feature vector and classifies the feeding status of the fish into different feeding activity levels.

[0044] S5.2: If it is determined that the feeding is relatively active, the controller issues an instruction to continue normal operation;

[0045] If it is identified as active feeding, the controller will issue a command to reduce the feeding amount to 60% of the original;

[0046] If it is identified as inactive feeding, the controller will issue a command to reduce the feeding amount to 20% of the original;

[0047] If it is determined that the animal is not eating, a stop instruction is issued through the controller.

[0048] The present invention provides a method for determining the amount of fish to be fed based on acoustic information and water surface fluctuation information. The technical effects are as follows:

[0049] 1) The fish farming feeding amount decision-making method proposed in this invention combines water surface fluctuation information with acoustic information to achieve precise feeding operations through real-time detection of fish feeding needs, providing fish with appropriate amounts of food and nutrition and creating favorable environmental conditions.

[0050] 2) In step 1 of the present invention, the fish are divided into four levels, namely: relatively active feeding, active feeding, inactive feeding, and no feeding. The feeding status of the fish is classified more finely, providing a decision-making basis for fine feeding of the fish school.

[0051] 3) In step 2 of the present invention, when performing signal collection, the water sound signal of the fish feeding and the water surface fluctuation signal are collected at the same time. The two collected signals can avoid the influence of light and turbidity in the breeding environment, and the information after the fusion of the two signals is more accurate and can avoid the influence of noise.

[0052] 4) In step 3 of the present invention, preprocessing the collected signal can effectively avoid the influence of most noises in the breeding environment during signal collection, thereby making the subsequent decision-making more accurate.

[0053] 5) In step 4 of the present invention, principal component analysis is performed after feature extraction of the collected signal, which can reduce the dimension of the feature vector, avoid overfitting during classification training, and make subsequent decision-making more accurate.

[0054] 6) The present invention uses acoustic information recognition to identify fish feeding activity, which can avoid the influence of light and turbidity in the aquaculture environment. However, it is susceptible to the influence of noise in the aquaculture environment. Although noise reduction can be performed after collection, when large aerators and sewage pumps are turned on and the fish are feeding weakly, it is still difficult to accurately detect the fish feeding status after noise reduction. However, the water surface fluctuation information obtained by the tilt angle sensor is more stable and can avoid the influence of noise. Through this information fusion method, the feeding time and feeding amount can be accurately controlled according to the fish's feeding appetite, providing a good growth environment for the fish. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a system structure diagram of the intelligent control system of an aquaculture feed feeding machine based on passive acoustic information and water surface fluctuation information.

[0056] Figure 2 This is a flow chart of the method for deciding the feeding amount for fish farming according to the present invention.

[0057] Figure 3(a) shows the original signal waveforms of the feeding process of five consecutive largemouth black bass groups;

[0058] Figure 3(b) is the short-time energy waveform corresponding to Figure 3(a).

[0059] Figure 4 This is the rendering of background noise in Audition 2020 when there are no fish.

[0060] Figure 5 This is the rendering of the fish feeding sound signal in Audition 2020.

[0061] Figure 6(a) shows the water surface fluctuation signal in the absence of fish feeding;

[0062] Figure 6(b) shows the water surface fluctuation signal when the fish are feeding.

[0063] Among them, the red, green and blue curves represent the angular velocities in the x-axis direction, the y-axis direction and the z-axis direction respectively. Figure 7 It is the short-time energy diagram of water surface fluctuation information. DETAILED DESCRIPTION

[0064] The purpose of the present invention is to provide a method for determining the amount of fish to be fed in aquaculture by integrating acoustic information and water surface fluctuation information. The method comprises: collecting the feeding sounds of fish in the current environment and state and the water surface fluctuation state, extracting the fluctuation information and sound information of the aquaculture water surface, extracting data from the obtained information, comparing and comprehensively analyzing the data, and using a trained classifier to determine the feeding state of the fish, thereby determining the feeding activity of the fish and providing a decision basis for the next feeding of the feeding machine. The method can effectively extract the relevant data features of the fish feeding, and through these features, the feeding activity of the fish can be more accurately determined, thereby achieving more accurate feeding of the feeding machine.

[0065] (1) Collection device:

[0066] like Figure 1 As shown, this acquisition device is connected to a signal analysis device. In this example, the acquisition device consists of an SM2+ acoustic recorder from the Song Meter series and two HTI-96-MIN standard hydrophones. One standard hydrophone is used to collect sounds from a school of farmed fish during feeding, while the other is used to collect sounds from the same environment without fish, serving as background noise samples. A digital tilt angle sensor simultaneously detects the frequency of water surface fluctuations and obtains target data. Simultaneously, a video recording device captures video of the fish throughout their feeding process to assist in determining their feeding activity. In this example, 300 data sets were collected from largemouth bass cultured in a high-density captive environment. These 100 data sets were divided into two groups: 100 used as a training set for the classifier, and the other 200 used to evaluate the accuracy of this method. Water surface fluctuation information was also collected using the tilt angle sensor.

[0067] (2) Signal analysis device:

[0068] This device, connected to a controller and signal acquisition device, processes and analyzes acquired fish feeding sound signals and water surface fluctuations. The analysis results are then input into the controller, which then determines appropriate operating instructions based on the analysis and sends them to the feeder. These instructions include continuing normal operation, reducing the feeding amount to a specified percentage of the normal feeding amount, and stopping operation.

[0069] like Figure 1 As shown, this signal analysis device is connected to a corresponding server for calculating and comparing corresponding characteristic values. Its calculation and analysis process is the control method of the aquaculture feed feeding machine based on passive acoustic information and aquaculture water surface fluctuation information in the present invention, and its technical route is as follows Figure 2 As shown, the following steps are included:

[0070] (1): The background noise sample of the fish-free feeding state collected by the signal acquisition device, wherein the background noise of the fish-free state is presented in Audition2020 as follows Figure 4 shown.

[0071] The noise reduction process is performed on the feeding sound signals of the fish school. At the same time, the water surface fluctuation information when the fish school feeds is collected through the tilt angle sensor and the singular values in the fluctuation information are removed.

[0072] (2): Based on the video of the fish feeding situation, the obtained fish feeding sound signal segments and the aquaculture water surface fluctuation signal segments are distinguished from the single feeding process, and the fish feeding sound signals and water surface fluctuation signals of the single feeding process are obtained; the fish feeding sound signals are presented in Audition2020 as follows Figure 5 The water surface fluctuation signal in the absence of fish feeding state is shown in Figure 6(a), and the water surface fluctuation signal in the presence of fish feeding state is shown in Figure 6(b).

[0073] (3): Extract the time domain characteristics of the audio data when the fish are in the feeding state - short-time energy. Let the speech waveform time domain signal be x(n), the window function w(n), and the i-th frame speech signal obtained after frame processing be y i (n), then y i (n) Satisfy:

[0074] y i (n)=w(n)*x((i-1)*inc+n),1≤n≤,1≤i≤fn;

[0075] Where w(n) is the window function, usually a rectangular window or a Hanning window; y i(n) is the value of a frame, n = 1, 2, ... L, i = 1, 2, ... fn, L is the frame length; inc is the frame shift length; fn is the total number of frames after framing. The formula for calculating the short-time energy of the i-th frame speech signal yi(n) is:

[0076]

[0077] Where: E(i) is the short-time energy of the i-th frame. The short-time energy of the collected fish feeding sound data was analyzed. Figure 3 shows the original signal waveforms and corresponding short-time energy of five consecutive largemouth black bass feeding processes. As shown in Figures 3(a) and 3(b), the short-time energy of each largemouth black bass feeding process varies significantly. As the feeding process progresses, the feeding activity of the largemouth black bass school decreases, and the short-time energy of the fish feeding process also decreases.

[0078] Table 1 Analysis of significant differences in short-term energy consumption of largemouth black bass at different feeding activity states

[0079]

[0080] Table 1 is an analysis table of significant differences in short-time energy of largemouth black bass groups in different feeding activity states. There are significant differences in the short-time energy of largemouth black bass groups during the feeding process at the significance level of 0.05. The average short-time energy of the feeding process when the fish group is actively feeding is greater than the average short-time energy of the feeding process when the fish group is inactive.

[0081] (4): Extract the frequency domain features of the audio data when the fish are in the feeding state - wavelet packet decomposition. According to the principle of wavelet packet decomposition to extract multi-scale spatial energy features, the signal energy at different decomposition scales is solved, and these energy values are arranged in scale order to form feature vectors for identification. The calculation formula is as follows:

[0082] E(j,i)=∑ k∈Z [p s (n,j,k)] 2 ;

[0083] Where: E(j,i) represents the energy value of each node i at decomposition level j; Ps(n,j,k) is the wavelet packet decomposition coefficient. In this case, a four-layer wavelet packet decomposition is performed, resulting in a 16-dimensional frequency band energy signature.

[0084] (5): Extract the short-term energy of the water surface fluctuation information from the information measured by the tilt angle sensor, such as Figure 7 As shown in Figure 2, the short-time energy map of water surface fluctuation information. Its extraction method is consistent with the short-time energy feature extraction in (1).

[0085] (6): Principal component analysis is performed on the 18 eigenvalues of the short-time energy of the sound information extracted in (3), (4), and (5), the 16-dimensional frequency band energy after wavelet packet decomposition (each dimension has a frequency band energy, so 16 frequency band energy eigenvalues are obtained after wavelet packet decomposition), and the short-time energy of the water surface fluctuation information. The mathematical model of principal component analysis is as follows:

[0086]

[0087] Among them, u jp ———standardization coefficient;

[0088] x ip ———the pth normalized feature of the i-th underwater acoustic signal sample;

[0089] z ij ———the jth principal component variable of the i-th underwater acoustic signal sample;

[0090] The characteristic vector of the fusion of the sound signal and the water surface fluctuation signal when the fish school feeds is obtained.

[0091] The fish sound signal feature vector X=(x1, x2, ..., x 18 ), where: x1 is the short-time average energy of the fish feeding sound signal, x2 is the short-time energy of the water surface fluctuation signal, x3~x 18 The frequency band energy of the fish feeding sound signal is decomposed by wavelet packets.

[0092] (7): The feature vector is imported into the classifier, and the classifier identifies the feeding status of the fish into different feeding activity levels by identifying the data in the feature vector. If it is identified as relatively active feeding, the controller issues an instruction to continue normal operation; if it is identified as active feeding, the controller issues an instruction to reduce the feeding amount to 60% of the original; if it is identified as inactive feeding, the controller issues an instruction to reduce the feeding amount to 20% of the original; if it is identified as not feeding, the controller issues an instruction to stop working.

[0093] (3) Controller:

[0094] The controller is used to receive the sound signal data processing results sent by the signal analysis and processing device, determine the corresponding work instructions based on the results, and let the feed feeding machine execute the corresponding instructions.

[0095] In this embodiment, 200 pieces of sound information and synchronized water surface fluctuation information when the fish are feeding are collected, each representing the feeding process information of a meal of the fish. Among them: 100 pieces of sound information and water surface fluctuation information are used as the training set of the classifier, and the other 100 pieces are used as the test set of the classifier. The accuracy of the test set reaches 96.7%, which has a high accuracy rate.

Claims

1. A method for determining the amount of fish to be fed based on acoustic information and water surface fluctuation information, characterized in that The following steps are involved: Step 1: Classify the feeding activity of the fish school into four levels: relatively active feeding, active feeding, inactive feeding, and no feeding; Step 2: Collect the sounds of the fish feeding to obtain a data set of the fish feeding sounds, and simultaneously collect the fluctuations of the aquaculture water surface to obtain a data set of water surface fluctuation information; and shoot a video of the fish feeding process during feeding; Step 3: Preprocess the data set of fish feeding sound and water surface fluctuation information obtained in step 2; Step 4: Extract eigenvalues from the sound data and water surface fluctuation information preprocessed in step 3, and perform principal component analysis on the extracted eigenvalues. Through principal component analysis, a eigenvector of the fusion of the sound signal and the water surface fluctuation signal when the fish are feeding is obtained; Step 5: Import the feature vector obtained in step 4 into the classifier. The classifier identifies the data in the feature vector, classifies the feeding status of the fish into different feeding activities, and issues work instructions based on the determined feeding activities.

2. The method for determining the amount of fish to be fed based on acoustic information and water surface fluctuation information according to claim 1, wherein: In step 1, active feeding means that the fish quickly gather in the feeding area and scramble for food; Active feeding means that fish swim to the feeding area and eat, and the aggregation degree decreases; Inactive feeding means: fish eat slowly but there is no remaining bait; Not feeding means that the fish no longer eat and there is obvious remaining bait.

3. The method for determining the amount of fish to be fed based on acoustic information and water surface fluctuation information according to claim 1, wherein: The second step specifically includes: First, sound is collected using an acoustic recorder and two hydrophones. One hydrophone is used to capture the sounds of the farmed fish feeding, while the other is used to capture the sounds of the same fishless environment, serving as background noise samples. Simultaneously, a digital tilt angle sensor detects the frequency of water surface fluctuations and collects target data. When recording the sound information and water surface fluctuation information of the fish school feeding, the video of the fish school feeding process during feeding is shot synchronously as a reference for the feeding activity of the fish school.

4. The method for determining the amount of fish to be fed based on acoustic information and water surface fluctuation information according to claim 1, wherein: The step three comprises the following steps: S3.1: Using the fish-free background noise samples collected in step 2, noise reduction is performed on the feeding sound signals of the school of fish; S3.2: removing singular values from the target data information collected by the tilt angle sensor; S3.3: Based on the video footage of the school feeding, distinguish the obtained fish feeding sound signal segments and water surface fluctuation signal segments from a single feeding process, and obtain the fish feeding sound signal and water surface fluctuation signal of the single feeding process; S3.4: Based on the feeding activity of fish in the video of fish feeding, the sound state of active feeding and the water surface fluctuation state of the fish feeding in a single feeding process are distinguished, and divided into single feeding relatively active, single feeding active, single feeding inactive and single feeding no feeding.

5. The method for determining the amount of fish to be fed based on acoustic information and water surface fluctuation information according to claim 1, wherein: The step 4 includes the following steps: S4.1: Extract the time domain features of the audio data when the fish are feeding - short-term energy: Assume that the speech waveform time domain signal is x(n), the window function is w(n), and the i-th frame speech signal obtained after frame processing is y i (n), then y i (n) Satisfy: y i (n)=w(n)*x((i-1)*inc+n),1≤n≤L,1≤i≤fn Where w(n) is the window function, usually a rectangular window or a Hanning window; y i (n) is the value of a frame, n = 1, 2, ... L, i = 1, 2, ... fn, L is the frame length; inc is the frame shift length; fn is the total number of frames after framing; the formula for calculating the short-time energy of the i-th frame speech signal yi(n) is: Where: E(i) is the short-term energy of the i-th frame; S4.2: Extracting frequency domain features of audio data from fish feeding — Wavelet packet decomposition: According to the principle of extracting multi-scale spatial energy features by wavelet packet decomposition, the signal energy at different decomposition scales is solved, and these energy values are arranged in scale order to form feature vectors for identification. If a 4-layer wavelet packet decomposition is performed, a 16-dimensional frequency band energy feature can be obtained, and its calculation formula is as follows: E(j,i)=∑ k∈Z [p s (n,j,k)] 2 Where: E(j,i) represents the energy value of each node i at decomposition level j; k = 1, 2, ..., n, Z represents an integer; Ps(n, j, k) is the wavelet packet decomposition coefficient; Each dimension has a frequency band energy, so after wavelet packet decomposition, 16 frequency band energy eigenvalues are obtained; S4.3: Extract the short-term energy of the water surface fluctuation information from the information measured by the tilt angle sensor; S4.4: Principal component analysis is performed on the 18 eigenvalues of the short-time energy of the sound information extracted in steps S1), S2, and S3), the energy of the 16-dimensional frequency bands after wavelet packet decomposition, and the short-time energy of the water surface fluctuation information. The mathematical model of the principal component analysis is as follows: Among them, u jp ———standardization coefficient; x ip ———the pth normalized feature of the i-th underwater acoustic signal sample; z ij ———the jth principal component variable of the i-th underwater acoustic signal sample; The characteristic vector of the fusion of the acoustic signal and the water surface fluctuation signal when the fish school is feeding is obtained and used to identify and analyze the feeding activity level.

6. The method for determining the amount of feed to be fed to aquaculture fish based on acoustic information and water surface fluctuation information according to claim 1, wherein: In step 5, the obtained feature vector is imported into the classifier, and the classifier further classifies and judges it and the controller issues a working instruction, including the following steps: S5.1: The feature vector is imported into the classifier. The classifier identifies the data in the feature vector and classifies the feeding status of the fish into different feeding activity levels. S5.2: If it is determined that the feeding activity is relatively active, the controller issues an instruction to continue normal operation; If it is identified as active feeding, the controller will issue a command to reduce the feeding amount to 60% of the original; If it is identified as inactive feeding, the controller will issue a command to reduce the feeding amount to 20% of the original; If it is determined that the animal is not eating, a stop instruction is issued through the controller.

Images