Fish anesthesia recovery activity prediction method based on acoustic signal time-frequency domain feature fusion

Abstract

The invention belongs to the technical field of aquaculture and food engineering, and relates to a fish anesthesia recovery activity prediction method based on acoustic signal time-frequency domain feature fusion. The method comprises the following steps of: acquiring acoustic respiratory vibration signals in real time in a fish anesthesia recovery process, performing 0.1-50Hz band-pass filtering, denoising and other pretreatment on the acoustic respiratory vibration signals, and extracting ZCR and Gammatone characteristic spectrums; meanwhile, fish swimming data are collected by adopting a YOLOv11 visual tracking algorithm, and activity grading standards are established by calculating activity indexes; zCR and Gammatone maps are used as double-branch input data of a deep learning model, corresponding activity grading data are used as labels, and a multi-category classification prediction model is constructed. According to the method, the anesthesia recovery state of the fish can be accurately predicted, and excessive stress and death loss in the transportation process are avoided; a non-invasive and intelligent detection means is provided for activity monitoring in the aquatic product transportation process, and the market application prospect is wide.

Description

Technical Field

[0001] This invention belongs to the field of food testing technology, specifically relating to a method for predicting the recovery effect of anesthesia during fish transport based on acoustic technology. Background Technology

[0002] As a vital global source of protein, the survival rate of aquatic products during transportation directly impacts economic benefits and food safety. In modern aquaculture, long-distance fish transportation has become a crucial link in the industrial chain, but the persistently high mortality rate caused by transportation stress has become a significant bottleneck restricting the industry's development. To alleviate transportation stress, anesthesia techniques are widely used in fish transportation to improve survival rates by reducing metabolic activity and stress responses. However, assessing the recovery status of anesthetized fish remains a technical challenge in this field.

[0003] Existing methods for assessing fish activity have significant limitations, failing to effectively address the complex physiological changes and environmental disturbances encountered during transportation. This leads to frequent problems such as delayed judgments, insufficient accuracy, or misjudgments. These issues not only reduce transportation efficiency but can also cause fish mortality and economic losses. Traditional manual observation and physiological indicator detection, due to their high subjectivity, invasiveness, and poor real-time performance, cannot meet the demands of modern aquatic product transportation for precise activity monitoring. Although non-invasive technologies such as computer vision, bioelectrical impedance analysis, and infrared thermal imaging have been gradually applied to fish status monitoring in recent years, their practicality is severely limited by factors such as the complexity of the aquatic environment, individual differences among fish, changes in light intensity, and the adaptability of equipment.

[0004] Given the aforementioned technical challenges, the aquatic product transportation industry urgently needs a non-invasive, real-time, and highly accurate technical method to accurately monitor and assess the activity recovery status of anesthetized fish, in order to minimize transportation losses, improve survival rates, and ensure economic benefits. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a method for predicting the recovery activity of anesthetized fish based on the fusion of time-frequency domain features of acoustic signals. By innovatively combining acoustic signals with advanced deep learning models, a high-precision model for discriminating and predicting the recovery activity of anesthetized fish is developed, providing key technical support for intelligent aquatic product transportation systems and possessing significant theoretical and practical application value.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting fish anesthesia recovery activity based on the fusion of time-frequency domain features of acoustic signals includes the following steps:

[0008] Step (1) Transportation Environment Simulation

[0009] S1: Select fish with a body length of K cm and divide them into two groups, N and M, denoted as N1, N2...N... n M1, M2, ..., M n N and M are positive integers; group N is used for transport anesthesia simulation, and group M is used for image acquisition;

[0010] S2: Select three aquatic transport boxes, labeled A1, A2, and A3. A1 is used for anesthesia transport simulation, A2 is used for hydrophone acoustic signal detection, and A3 is used for swimming characteristic discrimination signal acquisition. Add water to the three boxes and let them stand for a period of time for later use.

[0011] S3: Mix the anesthetic with ethanol to obtain an anesthetic solution; add the anesthetic solution to the water in the aquatic transport box A1 to form a mixture; finally, put N fish into the aquatic transport box A1 and wait for D minutes to induce anesthesia in the fish.

[0012] Preferably, the fish mentioned in step (1) S1 includes grass carp with a body length of 25-30cm.

[0013] Preferably, the aquatic product transport box in step (1) S2 has a compartment size of 60×50×35cm, the water added is river water, the water volume is 1 / 2 to 4 / 5 of the height of the water box, and it is left to stand for a period of 48-72 hours.

[0014] Preferably, the anesthetic agent in step (1) S3 is eugenol, and the volume ratio of eugenol to ethanol is 6-9:1; the final concentration of the mixture is 30-50 mg / L; and the waiting time D minutes is 3-4 minutes.

[0015] Step (2) Signal Acquisition

[0016] S1: The signal acquisition device includes an acoustic signal device and an industrial camera; the acoustic signal device consists of a host computer electrically connected to a hydrophone and a digital acquisition card to realize data transmission and control.

[0017] S2: Place the fish from group M one by one into water tank A3, and acquire images of them using an industrial camera. Collect E images of each fish, for a total of n×E images. Train the acquired images using YOLOv11. After training, call the trajectory tracking function of YOLOv11 to calculate the velocity vector based on the centroid displacement of two adjacent frames, and calculate the acceleration vector based on the velocity change at adjacent time points. Finally, sum the velocity and acceleration using weighted averages to obtain the comprehensive activity index S = α·v + β·a, where α and β are weighting coefficients, v is the velocity, and a is the acceleration. Set three threshold intervals [S1, S2] for activity determination: when S ≥ S2, it is considered highly active; when S1 ≤ S ≤ S2, it is considered moderately active; and when S ≤ S1, it is considered weakly active. S1 = 0.3·S_max, S2 = 0.7·S_max, where S_max represents the maximum value of the comprehensive activity index for all N fish during the entire anesthesia recovery period, determined by statistical analysis of all observed data.

[0018] S3: Take the fish N1, which has been anesthetized in step (1), out of the transport box A1 and put it into the test area box A2; and put the probe part of the hydrophone of the acoustic signal device into the box A2;

[0019] S4: Set the acquisition duration to F seconds on the host computer interface, acquire the fish's acoustic breathing vibration audio signal data, and record the data acquired at this time as T. 1-1 An industrial camera synchronously acquires F seconds of video data showing fish swimming. Based on the acquired video, the YOLOv11 target detection and trajectory tracking algorithm is used to calculate the instantaneous velocity and acceleration parameters of the fish in its swimming state in real time. A weighted summation is then used to obtain a comprehensive activity evaluation index. Following the discrimination method in step S2, the evaluation of the fish in this state (T seconds) can be achieved. 1-1 The classification of strong, medium, and weak activity, and the provision of signal data T 1-1 Add category tags;

[0020] S5: During the anesthesia recovery process of fish N1 in box A2, the operation S4 is repeated every t minutes, and the operation is repeated f times. The corresponding acoustic respiratory vibration audio signal data of grass carp are denoted as T. 1-2 T 1-3 ... T 1- f generates an acoustic signal dataset in the gradient recovery state;

[0021] S6: Following steps S3, S4, and S5, complete the process for grass carp N2, N3...N n The data acquisition yielded corresponding acoustic signal datasets T. 2-1 T 2-2 ...T 2- f;T 3-1 T3-2 ...T 3- f; ...; T n-1 T n-2 ...T n- f;

[0022] Preferably, in step (2) S1, the host computer is a computer configured with an Intel i7 processor and 16GB of memory; the hydrophone model is LST-DH01, the data acquisition card model is LST-DAQ, and the industrial camera model is MV-CA050-12UM.

[0023] Preferably, the value of F in step (2)S4 is 2-5.

[0024] Preferably, in step (2) S5, the value of t is 1-2 minutes, and the number of repetitions f is 4-6 times.

[0025] Step (3) Signal Processing

[0026] S1: Based on the data T collected in step (2) 1-1 T 1-2 ...T 1- First, the F-second audio is subjected to a 4th-order Butterworth bandpass filter from 0.1 to 50 Hz. Then, the zero-crossing rate (ZCR) is calculated. Next, its variation trend is captured by first-order and second-order differences, and a 3-point smoothing process is applied to the ZCR. Finally, the feature matrix is ​​expanded to 64 frequency bins, and the ZCR spectrum (denoted as X) is obtained through Inferno color mapping. 1-1 X 1-2 ...X 1- f;

[0027] When generating the Gammatone spectrum, 64 center frequencies are first generated in the range of 0.1-50Hz using the ERB scale for the F-second audio, and a filter bank is created. After the audio is passed through the filter bank, the RMS (root mean square) energy of each frequency band is calculated. Then, the breathing characteristics are enhanced by combining the energy change rate. Finally, the Gammatone spectrum (abbreviated as R) is obtained through Viridis color mapping, and is denoted as R. 1-1 R 1-2 ...R 1- f;

[0028] S2: Repeat step (3) S1, and process fish N2, N3...N in the same way. n The data were used to obtain ZCR and Gammatone spectra, denoted as X. 2-1 X 2-2 ...X 2-f ..., X n-1 X n-2 ...X n-f and R2-1 R 2-2 ...R 2-f , ..., R n-1 R n-2 ...R n-f Finally, the obtained ZCR and Gammatone maps will be used as input data for subsequent deep learning models.

[0029] Preferably, in step (3) S1, the zero-crossing rate is calculated with a frame length of 25ms and a jump length of 10ms; the RMS energy of each frequency band is calculated with a frame length of 50ms and a jump length of 10ms.

[0030] Step (4) Model building

[0031] Based on the ZCR spectrum (X) and Gammatone spectrum (R) obtained in step (3), save each spectrum data as a .jpg file and label each spectrum with the corresponding strong activity, medium activity, and weak activity classification labels; store them in the corresponding category folders according to the classification labels;

[0032] ZCR and Gammatone maps were selected for prediction using feature fusion, and a deep learning model was implemented using Python. The learning model was trained by dividing the data into training and validation sets. Based on the input data X and R, a classification mapping relationship was constructed between the data and the strong, medium, and weak activity levels of the swimming features, realizing multi-class classification modeling of swimming states. The model performance was evaluated using accuracy (Acc) and cross-entropy loss (Loss), and the best-performing learning model was selected as the final model. The model was then saved to a file in .pth format.

[0033] Preferably, in step (4), 70%-80% of the total samples are used as the training set, and the remainder is used as the validation set.

[0034] Preferably, the deep learning models mentioned in step (4) include MobileNetV4-Mid, ConvNeXt V2-S, GhostNetV3-2.0x, and RepViT-M4.5; the criteria for judgment are as follows: Acc is used to measure the model's ability to interpret data, and the closer its value is to 1, the higher the model's fit to the data; Loss is used to measure the deviation between the model's predicted value and the true value, and the smaller its value, the higher the model's prediction accuracy.

[0035] Step (5) Practical Model Application

[0036] Select the sample to be tested, and obtain the ZCR and Gammatone spectra of the sample according to the operations of steps (1) to (3). Substitute them into the final model constructed in step (4) to quickly predict the corresponding swimming activity classification through the loaded model, and realize the prediction of fish transport anesthesia recovery.

[0037] The beneficial effects of this invention are:

[0038] This invention, based on hydrophone acoustic detection technology, innovatively establishes a platform for acquiring and predicting fish activity recovery signals. By performing multi-domain feature fusion (time-domain ZCR + frequency-domain Gammatone) processing on the acquired acoustic signals, and combining it with the YOLOv11 visual tracking algorithm to establish activity grading standards, a non-invasive real-time prediction and grading identification of the activity recovery status of anesthetized fish is achieved using deep learning models (MobileNetV4-Mid, ConvNeXtV2-S, GhostNetV3-2.0x, and RepViT-M4.5). Through optimization of the multi-domain feature fusion algorithm and deep learning model parameters, the constructed prediction model achieves an accuracy (Acc) of over 0.882 in the three-classification task of strong, medium, and weak fish activity, significantly improving the accuracy and efficiency of anesthesia recovery assessment. This overcomes the limitations of traditional manual observation and invasive detection methods, providing intelligent and standardized technical support for activity monitoring and loss control during aquatic product transportation.

[0039] The details are as follows:

[0040] The invention is highly innovative: it is the first to apply acoustic hydrophone technology to the detection of fish anesthesia recovery state, and innovatively proposes a signal processing method that fuses zero-crossing rate (ZCR) in the time domain with gammatone features in the frequency domain. This method overcomes the limitations of traditional single feature extraction and achieves multi-dimensional and accurate identification of fish respiratory vibration signals. This technical solution is groundbreaking in the field of aquaculture.

[0041] The predictive model demonstrates excellent performance: A deep learning-based prediction model for fish activity recovery was successfully constructed. Through comparative analysis of various advanced network architectures, including MobileNetV4-Mid, ConvNeXt V2-S, GhostNetV3-2.0x, and RepViT-M4.5, the optimal model configuration was selected. This model achieved an accuracy of 0.882 in predicting strong, medium, and weak activity levels, with the cross-entropy loss reduced to 0.793, enabling accurate identification and rapid response to anesthesia recovery status.

[0042] Non-invasive real-time monitoring: Utilizing hydrophone acoustic detection technology, the system enables real-time online monitoring of the anesthesia recovery process of fish without any invasive procedures on the fish, avoiding secondary stress and additional damage, maintaining the natural physiological state of the fish and the authenticity of the detection results, and significantly improving the survival rate of fish during transportation.

[0043] Strong environmental adaptability: Through optimized signal preprocessing algorithms (0.1-50Hz bandpass filtering, multi-point smoothing, etc.), the system can effectively suppress environmental noise interference such as water flow and equipment vibration, and maintain stable detection performance under complex environments such as different water quality conditions, temperature changes and transportation vehicle vibration.

[0044] Intelligent grading standard: The system integrates the YOLO11 target tracking algorithm and establishes an objective grading standard of strong, medium and weak activity by calculating the swimming speed and acceleration parameters of fish in real time, avoiding the subjectivity and low standardization of traditional manual observation.

[0045] High scalability: The technical solution has good scalability and can be applied to different kinds of freshwater fish (such as carp, crucian carp, grass carp, etc.). By adjusting the model parameters and training dataset, it can be extended to the activity detection of marine fish and other aquatic products. It has a wide range of applications and huge market potential. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the operation process of an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of a signal acquisition device.

[0048] Figure 3 The ZCR spectra of the time-domain signal are as follows: a) "strong" activity category; b) "medium" activity category; c) "weak" activity category.

[0049] Figure 4 Gammatone spectra of frequency domain signals: a) "strong" activity category; b) "medium" activity category; c) "weak" activity category.

[0050] Figure 5 Performance of the ConvNeXt V2-S model: a) cross-entropy loss; b) accuracy.

[0051] Figure 6 This is the confusion matrix for the ConvNeXt V2-S model. Detailed Implementation

[0052] The present invention will be further illustrated and explained in more detail through the following embodiments. It should be understood that the embodiments are merely illustrative examples, and the scope of protection of the present invention should not be limited or constrained by these embodiments.

[0053] It should be understood that the terminology set forth in this invention is used only to describe specific embodiments and is not intended to limit the scope of protection of this invention. Unless otherwise expressly stated, all technical and scientific terms used herein have the same meaning and connotation as commonly understood by one of ordinary skill in the art described herein.

[0054] Although this invention describes only preferred methodologies and materials, any methods and materials similar to or equivalent to those described herein may be used in the implementation or experimental verification of this invention. All references mentioned in this specification are incorporated herein by reference to disclose and describe in detail the methodologies and / or materials associated with those references. The raw materials and reagents used in this invention are all available through commercial channels or conventional means. In the event of any conflict or inconsistency between the cited references and this specification, the content of this specification shall prevail.

[0055] Example 1:

[0056] (1) Transportation Environment Simulation

[0057] S1: Purchase 20 grass carp, each 25-30cm in length, from the market, and select 10 of them, labeled N1, N2, ..., N. 10 Used for anesthesia simulation, the remaining 10 lines are used for image acquisition and are denoted as M1, M2, ..., M 10 Three aquatic product transport boxes with dimensions of 60×50×35cm were selected and labeled A1, A2, and A3. A1 was used for transport simulation, A2 for acoustic testing, and A3 for swimming characteristic identification. 80L of river water was poured into each of the three boxes, and they were left to stand for two days for later use.

[0058] S2: Prepare an anesthetic solution using eugenol and ethanol at a ratio of 9:1. Add the anesthetic solution to tank A1 to a final concentration of 40 mg / L. Place grass carp N1-N... 10 Place the fish in the container and wait 3 minutes to ensure it is anesthetized.

[0059] (2) Data Acquisition

[0060] S1: Signal Acquisition System Configuration: Host computer (Intel i7 processor, 16GB memory); Hydrophone model: LST-DH01; Acquisition card model: LST-DAQ; Industrial camera model: MV-CA050-12UM, such as... Figure 2 The device is shown in the schematic diagram.

[0061] S2: Purchase grass carp M1-M 10 The grass carp were placed into water tank A3 in sequence, and images were collected from them using an industrial camera. 60 images were collected from each grass carp, and a total of 600 images of grass carp dataset B were constructed.

[0062] The collected dataset B was trained using YOLOv11 for 150 epochs, and the best trained model was saved as YOLOv11-fish.pt. The trained YOLOv11-fish.pt model was used for fish trajectory tracking, which involved calculating the velocity and acceleration vectors. Finally, the velocity and acceleration were weighted and summed to obtain the comprehensive activity index S = 0.6v + 0.4a. Three threshold intervals [S1, S2] were set, with S1 = 0.3·S_max and S2 = 0.7·S_max. Statistical analysis showed that the comprehensive activity index ranged from 0.5 to 26.2 pixels / second², with a maximum value of S_max = 26.2 pixels / second. Based on S_max = 26.2 pixels / second², the theoretical calculations yielded: Weak activity threshold: S1 = 0.3 × 26.2 = 7.86 pixels / second²; Strong activity threshold: S2 = 0.7 × 26.2 = 18.34 pixels / second². After 50 sets of preliminary experiments, and considering the actual physiological characteristics of grass carp (25-30cm), the thresholds were adjusted to: Weak activity threshold: S1 = 7.86 pixels / second²; Strong activity threshold: S2 = 18.34 pixels / second². Classification criteria: S ≥ 18.34 indicates strong activity; 7.86 ≤ S < 18.34 indicates moderate activity; and S < 7.86 indicates weak activity.

[0063] S3: Take the anesthetized N1 out of the transport box A1 and place it in the test area box A2. Then put the hydrophone probe into box A2.

[0064] S4: Set the acquisition time to 3 seconds on the host computer interface to acquire acoustic breathing vibration audio signals. Record the acquired signal as T. 1-1 Next, a 3-second video was captured using an industrial camera, and the corresponding T was obtained based on step S1-1. 1-1 The recovery activity level of the audio signal is "weak activity";

[0065] S5: During the anesthesia recovery process of grass carp N1, starting from the 2nd minute, repeat operation S4 every 1 minute for 4 consecutive times, at the 2nd, 3rd, 4th, and 5th minutes respectively, to obtain the acoustic signal dataset T of grass carp N1. 1-2 T 1-3 T 1-4 T 1-5 Simultaneously, 3-second videos were collected to obtain the corresponding tags "weak activity", "medium activity", "medium activity" and "strong activity";

[0066] S3: Repeat steps S2, S2-1, and S2-2 to complete the data collection from 10 grass carp and obtain T. 2-1 ...T 2-5 T 3-1 ...T 3-5 ..., T 10-1 ...T 10-5 And the corresponding activity recovery label.

[0067] (4) Data processing

[0068] Based on the collected acoustic signal T of grass carp N1 1-1 First, a 4th-order Butterworth bandpass filter (0.1-50Hz) is applied. Then, the ZCR is calculated with a frame length of 25ms and a jump length of 10ms. After capturing the changing trend through first-order and second-order differences, the ZCR is smoothed at three points. The feature matrix is ​​then expanded to 64 frequency bins and subjected to Inferno color mapping to obtain the spectrum X. 1-1 ,like Figure 3 As shown;

[0069] When generating the Gammatone spectrum, 64 center frequencies are generated using the ERB scale within the 0.1-50Hz range, and a filter bank is created. After the audio is passed through the filter bank, the RMS energy of each frequency band is calculated with a frame length of 50ms and a jump length of 10ms. Combined with the energy change rate enhancement feature, the spectrum R is obtained through Viridis color mapping. 1-1 ,like Figure 4 As shown.

[0070] Following this process, for N1's T 1-2 To T 1-5 The signal is processed sequentially to generate the spectrum X. 1-2 To X 1-5 R 1-2 To R 1-5 This corresponds to the activity tags "weak activity," "medium activity," "medium activity," and "strong activity." Similarly, the T... (likely referring to a specific assay or assay) was completed on 10 grass carp... 2-1 To T 10-5 Signal processing to generate the corresponding ZCR spectrum X 2-1 To X 10-5 With Gammatone spectrum R 2-1 To R 10-5All images were saved as 224×224 pixel .jpg files. The ZCR spectrum obtained after converting the 3-second audio signal was denoted as X, and the Gammatone spectrum as R. The collected and processed data X and R are denoted as TD1, containing 10 grass carp × 5 collections × 2 types of spectra = 100 spectra. The complete data collection was repeated every 3 days thereafter, for a total of 10 collections (TD1 to TD...). 10 A complete dataset of 1000 maps (10 times × 100 maps) was constructed.

[0071] Finally, save the graph in the following example path format: D:\data\Weak\ZCR\X 1-1 .jpg、D:\data\Weak\Gammatone\R 1-1 .jpg.

[0072] (5) Model building

[0073] S1: A deep learning model using time-domain and frequency-domain feature fusion is employed. A dual-branch convolutional neural network is constructed to process ZCR and Gammatone maps, enabling multi-class classification of fish activity states. Specifically, a parallel branching structure is used to extract deep features from the ZCR map (time-domain feature visualization) and the Gammatone map (frequency-domain feature visualization), respectively, and information complementarity is achieved through a cross-modal feature fusion layer.

[0074] S2: Network Framework Design: The input layer receives 224×224 pixel ZCR and Gammatone maps. A lightweight backbone network extracts temporal zero-crossing rate features and frequency-domain energy distribution features, generating feature vectors with consistent dimensions. The fusion layer employs a progressive multi-scale design, first connecting the dual-branch features dimensionally through channel concatenation, then achieving cross-modal information interaction through a three-level convolutional module (low-level texture fusion, mid-level pattern fusion, and high-level semantic fusion). Learnable feature importance weights and cross-scale jump connections are introduced during this process to enhance the extraction efficiency of temporal-frequency domain related features. The classification layer, based on the fused features, outputs "strong," "medium," and "weak" three-class classification probability values ​​through a fully connected layer combined with the Softmax function. This architecture ensures input size consistency through adaptive pooling, utilizes a feature selector to accommodate multimodal inputs, and finally achieves accurate classification of the activity recovery state of anesthetized fish through joint optimization of cross-entropy loss and accuracy. Each model underwent 50 rounds of training and testing, yielding results shown in Table 1.

[0075] Table 1 Results of the Model

[0076]

[0077] The results show that the dataset combining ZCR and Gammatone maps achieves relatively good prediction performance under all four network models, with R... 2 All exceeded 0.8, with the ConvNeXt V2-S network model having the highest ACC at 0.882 and the lowest Loss at 0.693, demonstrating strong predictive performance. Figure 5 The curves showing the cross-entropy loss versus accuracy of the model demonstrate that the model gradually converges. Figure 6 The confusion matrix for classification prediction shows the distribution of correct and incorrect predictions.

[0078] (7) Model Application

[0079] Save the selected best model, ConvNeXt V2-S, in .pth format to the file directory, and you can use the model inference function to complete the call.

[0080] This invention innovatively constructs a predictive system for the recovery effect of fish anesthesia based on hydrophone acoustic detection technology. By collecting acoustic signals of fish respiratory vibrations, extracting and fusing time-domain zero-crossing rate (ZCR) and frequency-domain gammatone features, and combining them with a swimming activity grading standard established by the YOLOv11 visual tracking algorithm, non-invasive real-time prediction of the fish's anesthesia recovery state is achieved using deep learning models such as ConvNeXt V2-S. This method achieves an accuracy of over 0.882 in a three-class classification task of strong, medium, and weak fish activity, with a cross-entropy loss as low as 0.693. It overcomes the limitations of traditional manual observation and invasive detection, providing a high-precision and intelligent technical solution for activity monitoring in aquatic product transportation, effectively improving fish survival rates and reducing transportation losses, and has significant practical application value for the intelligent development of aquaculture.

[0081] Note: The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Therefore, although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. All technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting fish anesthesia recovery activity based on the fusion of time-frequency domain features of acoustic signals, characterized in that, Includes the following steps: Step (1) Transportation Environment Simulation S1: Select fish with a body length of K cm and divide them into two groups, N and M, denoted as N1, N2...N... n M1, M2, ..., M n N and M are positive integers; group N is used for transport anesthesia simulation, and group M is used for image acquisition; S2: Select three aquatic transport boxes, labeled A1, A2, and A3. A1 is used for anesthesia transport simulation, A2 is used for hydrophone acoustic signal detection, and A3 is used for swimming characteristic discrimination signal acquisition. Add water to the three boxes and let them stand for a period of time for later use. S3: Mix the anesthetic with ethanol to obtain an anesthetic solution; add the anesthetic solution to the water in the aquatic transport box A1 to form a mixture; finally, put N fish into the aquatic transport box A1 and wait for D minutes to induce anesthesia in the fish. Step (2) Signal Acquisition S1: The signal acquisition device includes an acoustic signal device and an industrial camera; the acoustic signal device is composed of a host computer electrically connected to a hydrophone and a digital acquisition card to realize data transmission and control. S2: Place the fish from group M one by one into water tank A3, and acquire images of them using an industrial camera. Collect E images of each fish, for a total of n×E images. Train the acquired images using YOLOv11. After training, call the trajectory tracking function of YOLOv11 to calculate the velocity vector based on the centroid displacement of two adjacent frames, and calculate the acceleration vector based on the velocity change at adjacent time points. Finally, sum the velocity and acceleration using weighted averages to obtain the comprehensive activity index S = α·v + β·a, where α and β are weighting coefficients, v is the velocity, and a is the acceleration. Set three threshold intervals [S1, S2] for activity determination: when S ≥ S2, it is considered highly active; when S1 ≤ S ≤ S2, it is considered moderately active; and when S ≤ S1, it is considered weakly active. S1 = 0.3·S_max, S2 = 0.7·S_max, where S_max represents the maximum value of the comprehensive activity index for all N fish during the entire anesthesia recovery period, determined by statistical analysis of all observed data. S3: Take the fish N1, which has been anesthetized in step (1), out of the transport box A1 and put it into the test area box A2; and put the probe part of the hydrophone of the acoustic signal device into the box A2; S4: Set the acquisition duration to F seconds on the host computer interface, acquire the fish's acoustic breathing vibration audio signal data, and record the acquired data as T. 1-1 Industrial cameras synchronously acquire F seconds of video data of fish swimming; based on the acquired swimming video, the YOLOv11 target detection and trajectory tracking algorithm is used to calculate the instantaneous velocity and acceleration parameters of the fish in the swimming state in real time, and the comprehensive activity evaluation index is obtained by weighted summation. And according to the discrimination method in step S2, it is possible to realize the discrimination of T in this state. 1-1 The classification of strong, medium, and weak activity, and the application of signal data T 1-1 Add category tags; S5: During the anesthesia recovery process of fish N1 in box A2, the operation S4 is repeated every t minutes, and the operation is repeated f times. The corresponding acoustic respiratory vibration audio signal data of grass carp are denoted as T. 1-2 T 1-3 ... T 1- f generates an acoustic signal dataset in the gradient recovery state; S6: Following steps S3, S4, and S5, complete the process for grass carp N2, N3...N n The data acquisition yielded corresponding acoustic signal datasets T. 2-1 T 2-2 ...T 2- f;T 3-1 T 3-2 ...T3f; ...;T n-1 T n-2 ...T n- f; Step (3) Signal Processing S1: Based on the data T collected in step (2) 1-1 T 1-2 ...T 1- First, the F-second audio is subjected to a 4th-order Butterworth bandpass filter from 0.1 to 50 Hz. Then, the zero-crossing rate is calculated. Next, its variation trend is captured by first-order and second-order differences, and a 3-point smoothing process is applied to the ZCR. Finally, the feature matrix is ​​expanded to 64 frequency bins, and the ZCR spectrum (X) is obtained through Inferno color mapping, denoted as X. 1-1 X 1-2 ...X 1- f; When generating the Gammatone spectrum, 64 center frequencies are first generated using the ERB scale within the range of 0.1-50Hz for the F-second audio, and a filter bank is created. After the audio is passed through the filter bank, the RMS energy of each frequency band is calculated. Then, the breathing characteristics are enhanced by combining the energy change rate. Finally, the Gammatone spectrum (R) is obtained through Viridis color mapping, denoted as R0. 1-1 R 1-2 ...R 1- f; S2: Repeat step (3) S1, and process fish N2, N3...N in the same way. n The data were used to obtain ZCR and Gammatone spectra, denoted as X. 2-1 X 2-2 ...X 2-f ..., X n-1 X n-2 ...X n-f and R 2-1 R 2-2 ...R 2-f , ..., R n-1 R n-2 ...R n-f Finally, the obtained ZCR and Gammatone maps will be used as input data for subsequent deep learning models. Step (4) Model building Based on the ZCR and Gammatone spectra obtained in step (3), save each spectra data as a .jpg file and label each spectra with the corresponding strong activity, medium activity, and weak activity classification labels; Store them in folders corresponding to their respective categories, categorized by label. ZCR and Gammatone maps were selected for prediction using feature fusion, and a deep learning model was implemented using Python. The learning model was trained by dividing the data into training and validation sets. Based on the input data X and R, a classification mapping relationship was constructed between the data and the strong, medium, and weak activity levels of the swimming features, realizing multi-class classification modeling of swimming state activity. The model performance was evaluated using accuracy and cross-entropy loss, and the best-performing learning model was selected as the final model. The model was then saved in .pth format. Step (5) Practical Model Application Select the fish sample to be tested, and obtain the ZCR and Gammatone spectra of the sample according to the operation of steps (1) to (3). Substitute them into the final model constructed in step (4) to quickly predict the corresponding swimming activity classification through the loaded model, and realize the prediction of fish transport anesthesia recovery.

2. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, The fish mentioned in step (1)S1 include grass carp, with a body length of 25-30cm.

3. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, The aquatic product transport box described in step (1) S2 has a compartment size of 60×50×35cm. The water added is river water, and the water volume is 1 / 2 to 4 / 5 of the height of the water box. The box is left to stand for a period of time of 48-72 hours.

4. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, The anesthetic agent mentioned in step (1) S3 is eugenol, which is mixed with ethanol at a volume ratio of 6-9:1; the final concentration of the mixture is 30-50 mg / L; the waiting time D minutes is 3-4 minutes.

5. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, In step (2), the host computer in S1 is a computer with an Intel i7 processor and 16GB of memory; the hydrophone model is LST-DH01, the data acquisition card model is LST-DAQ, and the industrial camera model is MV-CA050-12UM.

6. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, In step (2)S4, the value of F is 2-5.

7. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, In step S5, t takes the value of 1-2 minutes, and the number of repetitions f takes the value of 4-6 times.

8. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, In step (3) S1, the zero-crossing rate is calculated with a frame length of 25ms and a jump length of 10ms; the RMS energy of each frequency band is calculated with a frame length of 50ms and a jump length of 10ms.

9. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, In step (4), 70%-80% of the total samples are used as the training set, and the remainder is used as the validation set.

10. The method for predicting fish anesthesia recovery activity based on acoustic signal time-frequency domain feature fusion according to claim 1, characterized in that, The deep learning models mentioned in step (4) include MobileNetV4-Mid, ConvNeXt V2-S, GhostNetV3-2.0x, and RepViT-M4.5; the criteria for judgment are as follows: Acc is used to measure the model's ability to interpret data, and the closer its value is to 1, the higher the model's fit to the data; Loss is used to measure the deviation between the model's predicted value and the true value, and the smaller its value, the higher the model's prediction accuracy.

Images