Acoustic intelligent identification method for behavior state of penaeus vannamei boone and model building method

Abstract

The invention provides a penaeus vannamei behavior state acoustic intelligent identification method and a model building method, and belongs to the technical field of aquaculture acoustic monitoring. Comprising the following steps: acquiring audio data of penaeus vannamei in different behavior states; filtering and framing, threshold detection and feature extraction are carried out on the audio, and a standardized audio feature data set is constructed; a deep learning model ASP-SEResNet fusing attention statistical pooling and a channel attention mechanism is built, the deep learning model comprises a feature channel optimization branch based on a compression-excitation module and a time sequence feature expression branch adopting attention weighted statistic aggregation, and joint representation of acoustic features is achieved through a double-path feature fusion mechanism; training and optimizing the model to obtain a final model; experimental results show that the method shows excellent classification performance on a prawn audio feature data set, and high-precision identification of five behavior states of the penaeus vannamei can be realized.

Description

Technical Field

[0001] This invention belongs to the field of acoustic monitoring technology for aquaculture, and particularly relates to an acoustic intelligent recognition method and model building method for the behavioral status of Litopenaeus vannamei. Background Technology

[0002] The Pacific white shrimp is a key economic species in recirculating aquaculture ecosystems. It is not only an important output of the aquaculture industry but also plays a central role in the material cycle and energy flow within the aquaculture system. Accurate monitoring and identification of the behavioral status of Pacific white shrimp helps to scientifically assess population health, optimize feeding strategies and aquaculture management, and provide data support and decision-making basis for disease early warning, welfare-oriented aquaculture, and sustainable development.

[0003] However, accurate monitoring of the behavioral status of Litopenaeus vannamei remains a key technical challenge in recirculating aquaculture systems, especially in intensive, high-density environments where numerous technical difficulties arise. First, the acoustic signals emitted by shrimp are weak and highly susceptible to broadband noise interference from the continuous operation of pumps, aerators, and other equipment in the recirculating aquaculture system, thus obscuring effective acoustic characteristics. Second, the acoustic signals of different behavioral states (such as feeding, fighting, and stress) exhibit highly similar time-frequency characteristics, making it difficult for traditional signal processing methods to capture subtle differences. Furthermore, differences in individual size, physiological state, and environmental factors lead to significant variations in acoustic characteristics, further increasing the difficulty of pattern differentiation. These factors limit the applicability of traditional threshold- or spectral analysis-based acoustic recognition methods in recirculating aquaculture scenarios. Therefore, it is necessary to develop intelligent recognition models that can integrate temporal dynamic characteristics and deep feature representations, transforming the identification of Litopenaeus vannamei behavioral status into an audio classification problem based on deep learning models.

[0004] On the other hand, the audio signal preprocessing stage also faces numerous challenges. Recirculating aquaculture environments contain significant low-frequency water flow noise, equipment operating noise, and biological activity interference, resulting in extremely low signal-to-noise ratios for the raw audio signals. Effective shrimp vocalizations are often submerged in environmental noise. Traditional detection methods based on single thresholds or temporal features struggle to adapt to the large fluctuations in shrimp vocalization intensity and irregular duration, easily leading to missed detections of effective signals or false detections due to noise. Furthermore, shrimp vocalization events are discontinuous in time, and the acoustic features of different behavioral patterns overlap in the time-frequency domain, necessitating a feature representation method that balances temporal continuity and spectral characteristics. Therefore, a complete audio preprocessing workflow is required to provide high-quality, highly discriminative input features for subsequent deep learning models. Summary of the Invention

[0005] To address the aforementioned problems, the first aspect of this invention provides a method for building an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei, comprising the following steps: Step 1: In a recirculating aquaculture system environment, passive sonar is used to collect raw audio data of Litopenaeus vannamei under normal, hungry, feeding, weak, and stressed states. Step 2: Preprocess the raw audio data, and divide the data into training and test sets based on the preprocessed audio feature dataset. Step 3: An audio feature classification model for Litopenaeus vannamei (SPAV) sound signals, ASP-SEResNet, is built within a deep learning framework. First, the model takes audio features from different shrimp behavioral states as input. Through a compression-excitation module integrated into the ResNet backbone, the importance of feature channels is adaptively recalibrated to enhance key information representation and suppress noise interference, thereby extracting discriminative temporal feature representations. Subsequently, these temporal feature representations are input to an attention-based statistical pooling module. Through an attention weighting mechanism, the weighted mean and standard deviation of all frame-level temporal features are dynamically aggregated, transforming the variable-length temporal feature input sequence into a fixed-length global feature vector. Finally, the global feature vector is processed by a fully connected layer and normalized to output the classification and recognition results corresponding to each behavioral state of the SPAV. Step 4: Train and test the built audio signal classification model ASP-SEResNet using the training and test sets, and select the model with the best performance as the final model.

[0006] Preferably, the background of shrimp's normal, hungry, stressed, weak, and feeding states is artificially created. After collecting the original audio data, corresponding tags are marked. For example, audio data after fasting for one day is hunger, audio data during environmental changes is stress, and audio data after feeding is feeding.

[0007] Preferably, the preprocessing in step 2 includes: S21, Signal Filtering: Passes the original signal through a filter to retain the effective frequency band of the shrimp's sound, while suppressing out-of-band current noise and low-frequency environmental noise interference. S22, Signal Framing: Convert the filtered continuous audio signal into a continuous short-time frame sequence for analysis, ensuring a smooth transition between frames and reducing information loss that may occur in subsequent short-time analysis; S23, Threshold setting: Based on the acoustic signal characteristics of Litopenaeus vannamei, specific short-time average energy and short-time average zero-crossing rate are set as thresholds for dual-threshold detection. The setting of both parameters should clearly include the minimum threshold and the maximum threshold. S24, Dual-threshold sound frame detection: Determine whether each frame signal simultaneously meets the minimum and maximum thresholds set by two parameters. Sound frames that pass the dual-threshold detection will be determined as valid sound frames. S25, Finding and storing sound frames: By comparing the decision results of adjacent frames, find continuous sound frame regions, extract all extended effective audio segments from the original filtered signal, and store them as a new audio file for subsequent feature extraction and analysis. S26, Audio feature extraction of frame signals: The preprocessed audio segments are converted into Mel spectrograms. Mel spectrograms conform to the two-dimensional time-frequency representation of human hearing characteristics, and integrate time-domain and frequency-domain information, which can serve as ideal features for direct input and learning of deep neural networks.

[0008] Preferably, the specific calculation process for calculating the short-time average energy and short-time average zero-crossing rate of the frame signal in S23 is as follows:

[0009]

[0010] in, For the first Short-time average energy of a frame For the first The short-time average zero-crossing rate of the frame, For frame length, For the first The first frame The value of each sampling point, For the sampling point index within the frame, It is a symbolic function.

[0011] Preferably, the specific process for extracting audio features of a sound signal using the Mel spectrogram method in S26 is as follows: S261, Pre-emphasis: Increases the energy of high-frequency components. High-frequency components in audio signals have relatively low energy; pre-emphasis can compensate for this, ensuring they are not ignored in subsequent processing, and also helps eliminate the effects of radiation during sound production. S262, Framing and Windowing: Continuous time-domain signals are divided into short time segments (frames) because audio signals can be considered stationary for short periods (typically 10-30ms). To reduce spectral leakage, a window function is applied to each frame; the calculation process is as follows:

[0012]

[0013] in, For the first The signal of the frame, For frame length, For frame shift, For window functions, It is the first time after adding a window The signal of a frame; S263, Fast Fourier Transform (FFT) and Power Spectrum Calculation: Convert the time-domain signal of each frame into its frequency-domain representation, obtain its spectrum, and calculate the power spectrum:

[0014]

[0015] in It is the first The complex spectrum of the frame, It is a frequency index. It is the power spectrum; S264, through the Mel filter bank: maps a linear Hertz (Hz) frequency scale to a nonlinear Mel (Mel) frequency scale to simulate the auditory characteristics of the human ear; S265, taking the logarithm: compresses the dynamic range and highlights audio details. The calculation process is as follows:

[0016] in That is the final number Mel frequency domain coefficients of the frame It is a very small number, the purpose of which is to prevent taking the logarithm of 0.

[0017] Preferably, the ASP-SEResNet model uses ResNet as the feature extraction backbone network, and the compression-excitation module is embedded in the residual module of the backbone network. The compression-excitation module adaptively recalibrates the response weights of the feature channels through its unique channel attention mechanism, enhances the representation of key behavioral features and suppresses the generation of irrelevant noise. The recalibrated feature map is further extracted by the backbone network to generate a discriminative frame-level temporal feature representation. The temporal feature representation is then input to the attention statistical pooling module, which dynamically aggregates the weighted mean and standard deviation of all frame-level temporal features through the attention weight mechanism, transforming the variable-length temporal feature input sequence into a fixed-length global feature vector that summarizes the statistical characteristics of the input audio features. The global feature vector integrates the static statistical characteristics in the time-frequency domain with the dynamic discriminative features emphasized by the attention mechanism. Then, the global feature vector is dimensionally transformed through a fully connected layer, and the classification result is output through a normalization function. Based on this classification result, the classification and discrimination of different behavioral states of Litopenaeus vannamei is achieved.

[0018] Preferably, the compression-excitation module introduces an adaptive channel feature recalibration mechanism. This module achieves adaptive calibration of the feature response values ​​of different channels by establishing a dynamic nonlinear mapping relationship between feature channels; it can explicitly model the interdependencies between channels, enabling the network to enhance the representation ability of information-rich feature channels while suppressing the feature responses of redundant or unimportant channels; this module adopts a closed-loop feedback structure of "compression-excitation-recalibration". S31, compresses global spatial information into a single channel descriptor, capturing the global distribution characteristics of each channel (compression):

[0019] in, The first result obtained after compression operation Global descriptors for each channel, These represent the height and width of the feature map, respectively. For the first input feature map The feature matrix of each channel has a dimension of . ; For the input feature map Global average pooling is performed on each channel C, and the features of each two-dimensional channel are... Compressed into a scalar The generated channel descriptor It contains global spatial information for each channel; S32, based on the compressed information, learns the weights (excitations) of each channel through a self-gating mechanism:

[0020] in, Represents the channel weight vector. This is the channel descriptor vector output by the compression operation. and These are the weight matrices for the first and second fully connected layers, respectively. Through the first fully connected layer For compression vectors Dimensionality reduction is performed. This is done using the ReLU activation function. After nonlinearization, it is then passed through a second fully connected layer. Restored to the original dimensions; finally, activated by the Sigmoid function. Generate the weight vector for each channel. Each value represents the importance of the corresponding channel; S33 applies the learned channel weights to the original feature map to achieve adaptive calibration (recalibration) of the feature channels:

[0021] in, For the recalibrated first Output feature map of each channel The first feature map of the original input feature map One channel, The output of the stimulus operation The weighting coefficients of each channel; The weight scalar output of the stimulus operation Compared with the original feature map Channel-by-channel multiplication is performed, and the features of each channel are recalibrated to finally output the calibrated feature map. .

[0022] Preferably, the attention statistical pooling module is used to generate a fixed-length global feature vector, specifically including: S41, Calculate the attention weights, first for each time step of the sequence. Calculate an importance score Then, the Softmax function is used to normalize the scores of all time steps into a probability distribution to obtain the attention weights for each frame. :

[0023]

[0024] in, It is a time step Frame-level feature vectors, and These are the weights and biases of the linear transformation. It is a learnable weight vector. It is a bias term. Weight The size of the feature directly reflects the importance of the frame feature for the current shrimp behavior classification task.

[0025] S42, calculate the weighted statistics. After obtaining the attention weights, ASP no longer treats all frames equally, but instead calculates the weighted mean vector and the weighted standard deviation vector:

[0026]

[0027] in, To calculate the weighted mean vector, This is the weighted standard deviation vector.

[0028] S43, Generate a global feature vector by concatenating the calculated weighted mean and weighted standard deviation vectors to form the final fixed-length global feature vector:

[0029] in, This is the global feature vector.

[0030] A second aspect of this invention provides an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei, comprising the following processes: S1, real-time acquisition of raw audio data of Litopenaeus vannamei under normal, hungry, feeding, weak and stress states; S2, after preprocessing the original audio data such as noise filtering, frame detection, and feature extraction, forms the standard input data for the model; S3, input the standard input data of the model into the ASP-SEResNet model built by the construction method described in the first aspect; The S4 ASP-SEResNet model identifies and classifies the current behavioral state of shrimp based on the input data and outputs the classification results in real time.

[0031] A third aspect of the present invention provides an acoustic intelligent recognition device for the behavioral state of Litopenaeus vannamei, the device comprising at least one processor and at least one memory, the processor and the memory being coupled together; the memory storing a computer executable program for an ASP-SEResNet model constructed by the construction method described in the first aspect; when the processor executes the computer executable program stored in the memory, the processor executes an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei.

[0032] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer-executable program of an ASP-SEResNet model constructed by the construction method described in the first aspect, wherein when the computer-executable program is executed by a processor, the processor executes an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei.

[0033] Compared with the prior art, the present invention has the following beneficial effects: This invention designs an acoustic intelligent recognition method and model building method for the behavioral state of Litopenaeus vannamei, effectively overcoming the limitations of traditional acoustic analysis methods in recognizing subtle behavioral states of Litopenaeus vannamei. Addressing the characteristics of low signal-to-noise ratio and subtle feature differences in shrimp vocal signals, this invention proposes an ASP-SEResNet model that integrates channel attention and attention weighting mechanisms. Through multi-level feature optimization and temporal statistical feature fusion, it achieves a joint representation of global information and local detailed features, further focusing on the realization and accuracy improvement of multi-class classification tasks for the behavioral state of Litopenaeus vannamei. In the data preprocessing stage, this invention designs a comprehensive audio preprocessing workflow tailored to the acoustic characteristics of recirculating aquaculture environments, effectively enhancing the model's adaptability to complex underwater acoustic environments. Targeting the characteristics of the acoustic signals emitted by Litopenaeus vannamei, this invention proposes an acoustic event extraction method based on a dual-threshold detection mechanism. By combining short-time energy and zero-crossing rate adaptive threshold determination, the accuracy of capturing effective acoustic events is significantly improved. Furthermore, this invention utilizes Mel-spectrum feature transformation technology to convert the preprocessed audio signal into a representation conforming to auditory characteristics, providing highly discriminative feature input for the deep learning model. In the model architecture design, the compression-excitation (SE) module enhances the representation ability of key frequency band features through a channel attention mechanism. The attention statistical pooling (ASP) module dynamically weights and aggregates temporal feature representations through an attention weight mechanism, effectively extracting discriminative global feature representations. The two modules work synergistically in the feature extraction and feature aggregation stages to jointly improve the feature representation ability of weak acoustic signals of shrimp and the classification accuracy of the model in complex environments. The method and ideas proposed in this invention can not only be used for behavior detection of aquatic organisms in RAS, but can also be extended to behavior recognition, state classification and intelligent monitoring of other aquaculture species. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall process of the model building method proposed in this invention.

[0035] Figure 2 This is a schematic diagram of dual-threshold detection for frame signals.

[0036] Figure 3 This is a schematic diagram of audio feature extraction.

[0037] Figure 4 The image shows the waveform and Mel spectrum of the preprocessed audio signal.

[0038] Figure 5 This is a schematic diagram of the attention statistics pooling module.

[0039] Figure 6 This is a schematic diagram of the ASP-SEResNet model structure of the present invention.

[0040] Figure 7 Confusion matrix classification results.

[0041] Figure 8 This is a simplified structural diagram of the intelligent detection device of the present invention. Detailed Implementation

[0042] The invention will be further described below with reference to specific embodiments.

[0043] This embodiment further illustrates the method proposed in this invention through specific experimental procedures. The overall process is as follows: Figure 1 As shown.

[0044] 1. Raw data collection This experiment was conducted at the experimental base of the West Coast Campus of Ocean University of China in Qingdao. The Litopenaeus vannamei used in the experiment were in good health, with an average length of 10±2cm. The base was equipped with a recirculating aquaculture system. The culture pond was made of PP plastic, 1m high and 1.3m in diameter. The filter box had an outer diameter of 60*40*23cm and was divided into three layers. The top layer contained green and white filter cotton, the middle layer contained ceramic rings and multi-effect filter media activated carbon, and the bottom layer contained long strips of bacterial houses. The filter box was equipped with a water pump, model CLB-8000, with a maximum head of 6m, a maximum flow rate of 6000L / h, and an outlet diameter of 20mm. The aerator was model LY45S, with a power of 50W, a maximum exhaust volume of 50L / min, and 6 air outlets. The hydrophone used in the experiment was from the OCEAN SONICS product series icListen-900m24-bitSmart Hydrophone-200kHz, model RB9-ETH. The hydrophone sampling rate was set to 128kHz / s, and it could capture a maximum sound frequency of 50kHz, which meets the frequency range of Litopenaeus vannamei (2kHz-20kHz). The experiment lasted for 3 months, from April 15, 2025 to July 15, 2025. The water temperature was controlled at 25℃ by a heater, the salinity was 15, the pH was between 8.0 and 8.3, the water depth was 70cm, and the stocking density was 800g / m³. 3 Feed was delivered via an automatic feeder, normally once every 4 hours. A hydrophone was submerged 15cm in the water, and data was collected at the center of the pool. The background conditions were manually controlled before the hydrophone was activated, with each data collection session lasting 5 or 10 minutes. A total of 175 complete audio samples were obtained by the end of the experiment, with each behavioral state evenly distributed.

[0045] 2. Data Preprocessing and Dataset Construction Based on the original audio files collected above, this embodiment designs a complete audio preprocessing workflow: First, the original audio signal is filtered using an elliptic bandpass filter to retain the effective frequency band of shrimp vocalization in the range of 2000-20000Hz, while suppressing low-frequency environmental noise and high-frequency interference; then, the filtered signal is divided into frames with a frame length of 960 sampling points and a frame shift of 480 sampling points, and a Hamming window function is used to ensure smooth transition between frames; based on the shrimp vocalization characteristics, dual threshold detection conditions of short-time energy (threshold range 1e-10 to 2e-4) and zero-crossing rate (threshold range 0.1 to 0.3) are set, and each frame signal is jointly judged; continuous vocalization segments are obtained through the decision results of adjacent frames, and the forward and backward intervals of 5ms are extended to completely preserve the acoustic events; finally, the effective audio segments are converted into Mel spectrograms, and a two-dimensional feature representation with both time and frequency characteristics is generated through a 128-dimensional Mel filter bank, a 1024-point FFT, and logarithmic compression, providing optimized input for subsequent deep learning models. This process is used to provide the data foundation for accurate classification by the model.

[0046] The flowchart for determining whether an audio frame signal is a valid sound frame signal using the dual-threshold detection method is as follows: Figure 2 As shown, the specific formulas for calculating the short-time average energy and the short-time average zero-crossing rate are as follows:

[0047]

[0048] in, For the first Short-time average energy of a frame For the first The short-time average zero-crossing rate of the frame, For frame length, For the first The first frame The value of each sampling point, For the sampling point index within the frame, It is a symbolic function; The flowchart for extracting audio features using the Mel spectrogram method is as follows: Figure 3 As shown, the specific process is as follows: (1) Preprocessing and framing: The continuous audio signal is divided into short time frames (usually each frame is 20-40 milliseconds long), assuming that the signal is stationary in a short time; (2) Windowing: A window function (such as a Hamming window or Hanning window) is applied to each frame of the signal to reduce spectral leakage; (3) Fourier transform: A short-time Fourier transform (STFT) is performed on each frame of the signal to transform it from the time domain to the frequency domain, resulting in a linear spectrum; (4) Mel filter bank: A set of triangular filters (usually 40-80) are applied to the power spectrum of the linear spectrum. These filters are of equal width on the Mel scale, but are dense in the low-frequency region and sparse in the high-frequency region on the linear frequency; (5) Logarithmic transformation: The logarithmic Mel spectrum is obtained by taking the logarithm of the output energy of each filter; The original vocal waveform of the whiteleg shrimp after preprocessing and its corresponding Mel spectrum are shown in the figure. Figure 4 As shown.

[0049] After preprocessing, 175 audio samples were filtered to obtain a total of 4,000 audio feature files containing five states of Litopenaeus vannamei shrimp, with a uniform duration of 0.1 seconds. These files were then divided into training and testing sets in an 8:2 ratio to form an audio feature dataset that can be used for training and testing deep learning models.

[0050] 3. Building the ASP-SEResNet model Based on the aforementioned self-made audio feature dataset, this embodiment designs and builds a deep learning model, ASP-SEResNet, whose overall structure is as follows: Figure 5 As shown, the model first takes the audio features of shrimp under different behavioral states as input, and then adaptively recalibrates the importance of feature channels through the compression-excitation module integrated into the ResNet backbone network to enhance the representation of key information and suppress noise interference, thereby extracting discriminative temporal feature representations.

[0051] The compression-excitation module introduces an adaptive channel feature recalibration mechanism. This module establishes a dynamic nonlinear mapping relationship between feature channels to achieve adaptive calibration of the feature response values ​​of different channels. It can explicitly model the interdependencies between channels, enabling the network to enhance the representation ability of information-rich feature channels while suppressing the feature responses of redundant or unimportant channels. This module adopts a closed-loop feedback structure of "compression-excitation-recalibration". First, the global spatial information is compressed into a single channel descriptor, capturing the global distribution characteristics of each channel (compression):

[0052] in, The first result obtained after compression operation Global descriptors for each channel, These represent the height and width of the feature map, respectively. For the first input feature map The feature matrix of each channel has a dimension of . .

[0053] For the input feature map Global average pooling is performed on each channel C, and the features of each two-dimensional channel are... Compressed into a scalar The generated channel descriptor It contains global spatial information for each channel.

[0054] Secondly, based on the compressed information, the weights (excitations) of each channel are learned through a self-gating mechanism:

[0055] in, Represents the channel weight vector. This is the channel descriptor vector output by the compression operation. and These are the weight matrices for the first and second fully connected layers, respectively.

[0056] Through the first fully connected layer For compression vectors Dimensionality reduction is performed. This is done using the ReLU activation function. After nonlinearization, it is then passed through a second fully connected layer. Restored to the original dimensions; finally, activated by the Sigmoid function. Generate the weight vector for each channel. Each value represents the importance of the corresponding channel.

[0057] Finally, the learned channel weights are applied to the original feature map to achieve adaptive calibration (recalibration) of the feature channels:

[0058] in, For the recalibrated first Output feature map of each channel The first feature map of the original input feature map One channel, The output of the stimulus operation The weighting coefficients of each channel.

[0059] The weight scalar output of the stimulus operation Compared with the original feature map Channel-by-channel multiplication is performed, and the features of each channel are recalibrated to finally output the calibrated feature map. .

[0060] Subsequently, the temporal feature representation is input into the attention statistical pooling module, which dynamically aggregates the weighted mean and standard deviation of all frame-level features through the attention weight mechanism, transforming the variable-length temporal feature input sequence into a fixed-length global feature vector.

[0061] The specific implementation process of the attention statistics pooling module is as follows: First, the attention weights are calculated, starting with each time step in the sequence. Calculate an importance score Then, the Softmax function is used to normalize the scores of all time steps into a probability distribution to obtain the attention weights for each frame. :

[0062]

[0063] in, It is a time step Frame-level feature vectors, and These are the weights and biases of the linear transformation. It is a learnable weight vector. It is a bias term. Weight The size of the feature directly reflects the importance of the frame feature for the current shrimp behavior classification task.

[0064] Secondly, after calculating the weighted statistics and obtaining the attention weights, ASP no longer treats all frames equally, but instead calculates a weighted mean vector and a weighted standard deviation vector:

[0065]

[0066] in, To calculate the weighted mean vector, This is the weighted standard deviation vector.

[0067] Finally, a global feature vector is generated by concatenating the calculated weighted mean and weighted standard deviation vectors to form the final fixed-length global feature vector.

[0068] in, This is the global feature vector.

[0069] Finally, the global feature vector is processed by a fully connected layer and normalized to output the classification and recognition results corresponding to each behavioral state of the Litopenaeus vannamei.

[0070] 4. Model Training In this embodiment, the implementation platform for an acoustic intelligent recognition method of the behavior status of Litopenaeus vannamei is based on the Windows 10 operating system, using Python 3.9 as the programming language, PyTorch 1.12.0 as the deep learning framework, and CUDA version 11.7. The model was trained for 100 epochs on a system equipped with an NVIDIA GeForce RTX 4060 GPU (8GB of VRAM), lasting approximately 2 hours. The relevant settings of each training parameter were also recorded in detail during the experiment, as shown in Table 1.

[0071] Table 1. Detailed Experiment Configuration

[0072] The learning rate adjustment strategy employed in this invention is a preheating cosine annealing scheduler with 5 preheating rounds, a maximum learning rate of 0.001, and a minimum learning rate of 1e-5. This strategy uses a linearly increasing preheating mechanism in the initial training phase, gradually increasing the learning rate from zero to the maximum, effectively preventing the model from getting trapped in local optima due to gradient instability in the early stages. Subsequently, in the cosine annealing phase, the learning rate smoothly decays from its maximum value to its minimum value according to the cosine function. This non-linear decay method ensures that the model maintains strong gradient update capability in the early stages of convergence while achieving fine-tuning through an extremely low learning rate in the later stages of training.

[0073] During model training, the cross-entropy loss function is used to calculate the average cross-entropy loss for each batch of samples. For samples containing... The average cross-entropy loss function for each batch of samples is calculated as follows:

[0074] in Indicates the first The sample at the th The true label of the class, It is the first The sample was predicted by the model to be the th sample. The probability of a class.

[0075] 5. Evaluation of Model Performance Indicators The model proposed in this invention is evaluated using accuracy ( ), The accuracy value is used as the performance evaluation index for the model; the classification evaluation uses the accuracy ( ), recall rate ( )and The value is used as a classification performance evaluation index, and is also used as the performance index of this model. The calculation formula is as follows:

[0076]

[0077]

[0078]

[0079] in, The number of samples in the positive samples that correctly predict shrimp behavior; The number of samples that predict incorrect shrimp behavior in the positive samples; The number of samples in the negative samples that predict the correct shrimp behavior; This represents the number of samples in the negative sample that predict incorrect shrimp behavior.

[0080] 6. Experimental Results To fully verify the effectiveness of the proposed method, two parts of experiments were designed and completed. The first part was a feature extraction method comparison experiment, which compared ASP-SEResNet with commonly used sound feature extraction methods to evaluate the effectiveness of the Mel-spectrum mapping method for extracting sound features from shrimp. The second part was a model comparison experiment, which compared ASP-SEResNet with advanced sound recognition models to verify the effectiveness of ASP-SEResNet in recognizing the behavioral states of shrimp.

[0081] Comparison of feature extraction methods: This example compares commonly used sound feature extraction methods, namely spectrograms and Mel frequency cepstral coefficients. The experiment is based on the ASP-SEResNet model to demonstrate the advantages of the Mel spectrogram method in shrimp sound feature extraction. The specific comparison results are shown in Table 2.

[0082] Table 2 Performance comparison of different feature extraction methods

[0083] Experimental results show that the Mel spectrogram method is more suitable for extracting audio feature information of Litopenaeus vannamei and is significantly better than other feature extraction methods.

[0084] Model comparison experiment: This example compares the ASP-SEResNet model with state-of-the-art sound recognition models to demonstrate its effectiveness in recognizing shrimp behavior and states. The specific comparison results are shown in Table 3.

[0085] Table 3 Performance Comparison of Different Models

[0086] Experimental results show that the ASP-SEResNet model proposed in this invention is superior to existing mainstream sound recognition models and can more accurately distinguish the behavioral state of shrimp.

[0087] The specific classification and recognition results of the five behavioral states of Litopenaeus vannamei are shown in Table 4. Figure 7 As shown, the results indicate that the overall performance of the ASP-SEResNet model in identifying and classifying different behavioral states of Litopenaeus vannamei is excellent, demonstrating that the ASP-SEResNet model has good performance in the behavioral state identification task of Litopenaeus vannamei.

[0088] Table 4. Results of behavioral status identification of Litopenaeus vannamei.

[0089] like Figure 8 As shown, this invention also provides an acoustic intelligent recognition device for the behavioral state of Litopenaeus vannamei. The device includes at least one processor and at least one memory, as well as a communication interface and an internal bus. The memory stores a computer-executable program. The memory also stores a computer-executable program for the ASP-SEResNet model constructed using the method described above. When the processor executes the computer-executable program stored in the memory, it can execute an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei. The internal bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus in the accompanying drawings is not limited to only one bus or one type of bus. The memory may include high-speed RAM memory, and may also include non-volatile memory (NVM), such as at least one disk storage device, or it can be a USB flash drive, portable hard drive, read-only memory, disk, or optical disk, etc.

[0090] The device may be provided as a terminal, server, or other type of device. In an exemplary embodiment, the electronic device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0091] The present invention also provides a computer-readable storage medium storing a computer executable program of an ASP-SEResNet model constructed by the construction method described above. When the computer executable program is executed by a processor, it enables the processor to execute an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei.

[0092] Specifically, a system, apparatus, or device may be provided equipped with a readable storage medium on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system, apparatus, or device reads and executes the instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the embodiments described above, therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0093] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0094] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for constructing an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei, characterized in that, The process includes the following: Step 1: In a recirculating aquaculture system environment, passive sonar is used to collect raw audio data of Litopenaeus vannamei under normal, hungry, feeding, weak, and stressed states. Step 2: Preprocess the raw audio data, and divide the data into training and test sets based on the preprocessed audio feature dataset. Step 3: ASP-SEResNet, an audio feature classification model for Litopenaeus vannamei acoustic signals, is built within a deep learning framework. First, the audio features of shrimp under different behavioral states are used as input. The model uses ResNet as the feature extraction backbone network and leverages its residual structure to effectively learn the deep audio features of the sound signal. The compression-excitation module integrated into the ResNet backbone network adaptively recalibrates the importance of feature channels, enhances the representation of key information, suppresses noise interference, and extracts discriminative temporal feature representations. Subsequently, the temporal feature representation is input into the attention statistical pooling module, which dynamically aggregates the weighted mean and standard deviation of all frame-level features through the attention weight mechanism, transforming the variable-length temporal feature input sequence into a fixed-length global feature vector. Finally, the global feature vector is processed by a fully connected layer and normalized to output the classification and recognition results corresponding to each behavioral state of the Litopenaeus vannamei. Step 4: Train and test the built audio signal classification model ASP-SEResNet using the training and test sets, and select the model with the best performance as the final model.

2. The method for building an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei as described in claim 1, characterized in that, The preprocessing process includes: S21, Signal Filtering: Passes the original signal through a filter to retain the effective frequency band of the shrimp's sound, while suppressing out-of-band current noise and low-frequency environmental noise interference. S22, Signal Framing: Convert the filtered continuous audio signal into a continuous short-time frame sequence for analysis, ensuring a smooth transition between frames and reducing information loss that may occur in subsequent short-time analysis; S23, Threshold setting: Based on the acoustic signal characteristics of Litopenaeus vannamei, specific short-time average energy and short-time average zero-crossing rate are set as thresholds for dual-threshold detection. The setting of both parameters includes a minimum threshold and a maximum threshold. S24, Dual-threshold sound frame detection: Determine whether each frame signal simultaneously meets the minimum and maximum thresholds set by two parameters. Sound frames that pass the dual-threshold detection will be determined as valid sound frames. S25, Finding and storing sound frames: By comparing the decision results of adjacent frames, find the continuous sound frame region, extract all extended effective audio segments from the original filtered signal, and store them as a new audio file. S26, Audio feature extraction of frame signal: Convert audio segments in audio file into Mel spectrograms.

3. The method for building an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei as described in claim 2, characterized in that, In step S23, it is necessary to calculate the short-time average energy and short-time average zero-crossing rate of the frame signal. The specific calculation processes are as follows: in, For the first Short-time average energy of a frame For the first The short-time average zero-crossing rate of the frame, For frame length, For the first The first frame The value of each sampling point, For the sampling point index within the frame, It is a symbolic function.

4. The method for building an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei as described in claim 2, characterized in that, The specific process of S26 is as follows: S261, Pre-emphasis: By pre-emphasing, the energy of high-frequency components in the audio signal is increased; S262, Framing and Windowing: The continuous time-domain signal is divided into short time segments. To reduce spectral leakage, a window function is applied to each frame; the calculation process is as follows: in, For the first The signal of the frame, For frame length, For frame shift, For window functions, It is the first time after adding a window The signal of a frame; S263, Fast Fourier Transform and Power Spectrum Calculation: Convert the time-domain signal of each frame into its frequency-domain representation, obtain its spectrum, and calculate the power spectrum: in It is the first The complex spectrum of the frame, It is a frequency index. It is the power spectrum; S264, through the Mel filter bank: maps a linear Hertz frequency scale to a nonlinear Mel frequency scale to simulate the auditory characteristics of the human ear; S265, taking the logarithm: compresses the dynamic range and highlights audio details. The calculation process is as follows: in For the final Mel frequency domain coefficients of the frame It is a very small number, to prevent taking the logarithm of 0.

5. The method for building an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei as described in claim 1, characterized in that: The compression-excitation module introduces an adaptive channel feature recalibration mechanism. By establishing a dynamic nonlinear mapping relationship between feature channels, it achieves adaptive calibration of the feature response values ​​of different channels. Specifically: S31 compresses global spatial information into a single channel descriptor, capturing the global distribution characteristics of each channel: in, The first result obtained after compression operation Global descriptors for each channel, These represent the height and width of the feature map, respectively. For the first input feature map The feature matrix of each channel has a dimension of . ; For the input feature map Global average pooling is performed on each channel C, and the features of each two-dimensional channel are... Compressed into a scalar The generated channel descriptor It contains global spatial information for each channel; S32, based on the compressed information, learns the weight of each channel through a self-gating mechanism: in, Represents the channel weight vector. This is the channel descriptor vector output by the compression operation. and These are the weight matrices for the first and second fully connected layers, respectively. Through the first fully connected layer For compression vectors Dimensionality reduction is performed; activation is achieved via ReLU function. After nonlinearization, it is then passed through a second fully connected layer. Restored to the original dimensions; finally activated by the Sigmoid function. Generate the weight vector for each channel. Each value represents the importance of the corresponding channel; S33 applies the learned channel weights to the original feature map to achieve adaptive calibration of the feature channels: in, For the recalibrated first Output feature map of each channel The first feature map of the original input map One channel, The first output of the stimulus operation The weighting coefficients of each channel; The weight scalar output of the stimulus operation Compared with the original feature map Channel-by-channel multiplication is performed, and the features of each channel are recalibrated to finally output the calibrated feature map. .

6. The method for building an acoustic intelligent recognition model of the behavioral state of Litopenaeus vannamei as described in claim 1, characterized in that: The attention statistical pooling module is used to generate a fixed-length global feature vector, specifically including: S41, Calculate attention weights; first, for each time step in the sequence... Calculate an importance score Then, the Softmax function is used to normalize the scores of all time steps into a probability distribution to obtain the attention weights for each frame. : in, It is a time step Frame-level feature vectors, and These are the weights and biases of the linear transformation. It is a learnable weight vector. It is a bias term; weight The size of the feature directly reflects the importance of the frame feature for the current shrimp behavior classification task; S42, calculate the weighted statistics, obtain the attention weights, and then calculate the weighted mean vector and weighted standard deviation vector: in, To calculate the weighted mean vector, This is the weighted standard deviation vector; S43, Generate a global feature vector by concatenating the calculated weighted mean and weighted standard deviation vectors to form the final fixed-length global feature vector: in, This is the global feature vector.

7. An acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei, characterized in that, The process includes the following: S1, real-time acquisition of raw audio data of Litopenaeus vannamei under normal, hungry, feeding, weak and stress states; S2, after preprocessing the original audio data such as noise filtering, frame detection, and feature extraction, forms the standard input data for the model; S3, input the standard input data of the model into the ASP-SEResNet model constructed by the construction method as described in any one of claims 1 to 6; The S4 ASP-SEResNet model identifies and classifies the current behavioral state of shrimp based on input data and provides classification results in real time.

8. An acoustic intelligent recognition device for the behavioral state of Litopenaeus vannamei, characterized in that: The device includes at least one processor and at least one memory, the processor and the memory being coupled together; the memory stores a computer executable program for an ASP-SEResNet model constructed by the construction method as described in any one of claims 1 to 6; when the processor executes the computer executable program stored in the memory, the processor executes an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer executable program for an ASP-SEResNet model constructed by the construction method as described in any one of claims 1 to 6. When the computer executable program is executed by a processor, it causes the processor to execute an acoustic intelligent recognition method for the behavioral state of Litopenaeus vannamei.

Images