Poultry live body breast meat rate, leg meat rate, and sebaceous rate detection system based on acoustic detection

Abstract

The present application relates to the technical field of poultry live meat rate detection, in particular to a poultry live breast meat rate, leg meat rate and subcutaneous fat rate detection system based on acoustic detection. It comprises: an acoustic sensing unit for collecting acoustic signals emitted by different parts of a live poultry to obtain acoustic characteristic information related to meat quality; a signal processing unit for converting analog signals containing acoustic characteristic information collected by the acoustic sensing unit into digital signals and extracting acoustic characteristic parameters related to breast meat rate, leg meat rate and subcutaneous fat rate; and a data analysis unit for estimating the breast meat rate, leg meat rate and subcutaneous fat rate of a live poultry based on the acoustic characteristic parameters of the signal processing unit through a neural network model MLP. The present application designs a multilayer perceptron neural network model, combines various characteristic parameters such as frequency spectrum analysis and time domain analysis, and the system can learn and identify patterns closely related to breast meat rate, leg meat rate and subcutaneous fat rate in complex data sets.

Description

Technical Field

[0001] This invention relates to the field of poultry live meat percentage detection technology, and more specifically, to a system for detecting the percentage of breast meat, leg meat, and skin fat in live poultry based on acoustic detection. Background Technology

[0002] Traditional testing methods typically require slaughtering poultry to obtain accurate data on breast meat percentage, leg meat percentage, and subcutaneous fat percentage. This is not only costly for observing meat performance indicators in poultry breeding and for commercial poultry farming, but also fails to provide real-time monitoring of individual growth and continuous monitoring in a live animal state. Furthermore, traditional measurement methods require slaughtering poultry, failing to retain superior breeding individuals, are time-consuming, require manual operation, and can only be performed at specific times, resulting in low efficiency and difficulty in obtaining timely data feedback. These are intractable problems in poultry breeding and, for commercial poultry farming, can impact the feeding management and market decisions of farming enterprises. Therefore, this paper proposes an acoustic detection-based system for detecting breast meat percentage, leg meat percentage, and subcutaneous fat percentage in live poultry. Summary of the Invention

[0003] The purpose of this invention is to provide an acoustic detection-based system for detecting the percentage of breast meat, leg meat, and skin fat in live poultry, in order to solve the problems of high demand, low efficiency, and difficulty in monitoring the live condition of poultry mentioned in the background art.

[0004] To achieve the above objectives, the present invention aims to provide an acoustic detection-based system for detecting the percentage of breast meat, leg meat, and skin fat in live poultry, comprising:

[0005] An acoustic sensing unit is used to collect sound wave signals reflected from different parts of a live poultry carcass in order to obtain acoustic characteristic information related to meat quality.

[0006] The signal processing unit is used to convert the analog signal containing acoustic characteristic information collected by the acoustic sensing unit into a digital signal and extract acoustic feature parameters related to breast fat percentage, leg fat percentage and subcutaneous fat percentage.

[0007] The data analysis unit estimates the breast meat percentage, leg meat percentage, and subcutaneous fat percentage of live poultry using a neural network model (MLP) based on the acoustic characteristic parameters of the signal processing unit. It also designs a multi-input model architecture based on the MLP to account for the effects of poultry pecking sounds, wing flapping sounds, and feather friction sounds on the model.

[0008] As a further improvement to this technical solution, the acoustic sensing unit is used to capture and collect sound wave signals emitted by the chest muscles, leg muscles and surrounding tissues.

[0009] As a further improvement to this technical solution, the signal processing unit converts the collected analog signals into digital signals, including the following steps:

[0010] S1.1 Convert a continuous-time signal into a discrete-time signal;

[0011] S1.2 Quantize the discrete-time signal, that is, convert the amplitude of the discrete-time signal into discrete values;

[0012] S1.3 Encode the quantized discrete values ​​and convert them into binary form.

[0013] As a further improvement to this technical solution, the acoustic feature parameters extracted by the signal processing unit that are related to the percentage of breast fat, leg fat, and subcutaneous fat specifically include: time-domain features and frequency-domain features.

[0014] As a further improvement to this technical solution, the data analysis unit includes a breast fat percentage module, a leg fat percentage module, and a subcutaneous fat percentage module;

[0015] Among them, the breast muscle percentage module uses a neural network model MLP to analyze the acoustic feature parameters of the breast muscle tissue to estimate the proportion of poultry breast muscle tissue.

[0016] The leg muscle percentage module estimates the proportion of poultry leg muscle tissue by analyzing the acoustic characteristic parameters of leg muscle tissue using a neural network model (MLP).

[0017] The subcutaneous fat percentage module estimates the proportion of subcutaneous fat tissue in poultry by analyzing the acoustic characteristic parameters of subcutaneous fat tissue using a neural network model (MLP).

[0018] As a further improvement to this technical solution, the data analysis unit estimates the breast meat percentage, leg meat percentage, and subcutaneous fat percentage of live poultry using a neural network model (MLP), including the following steps:

[0019] S2.1 Collect actual measurements and acoustic characteristic parameters of breast fat percentage, leg fat percentage, and subcutaneous fat percentage, and divide the data into training set and validation set;

[0020] S2.2 Define a multilayer perceptron (MLP) neural network model in the breast meat percentage module, leg meat percentage module, and subcutaneous fat percentage module. Each model contains an input layer, a hidden layer, and an output layer. Extend the functionality of the model by designing a multi-input model architecture to consider the effects of poultry pecking sounds, wing flapping sounds, and feather friction sounds on the MLP neural network model.

[0021] S2.3. The mean squared error (MSE) is used as the loss function. The gradient of the loss function with respect to the weights and biases is calculated through the backpropagation algorithm. The parameters of poultry calls and footsteps are introduced into the loss function to reduce the influence of noise on the loss function.

[0022] S2.4 Train the multilayer perceptron (MLP) neural network model on the training set;

[0023] S2.5. Use a trained multilayer perceptron (MLP) neural network model to predict breast fat percentage, leg fat percentage, and subcutaneous fat percentage.

[0024] As a further improvement to this technical solution, in S2.2, the multilayer perceptron (MLP) neural network model defined in the breast fat percentage module, leg fat percentage module, and subcutaneous fat percentage module is specifically as follows:

[0025] The chest muscle percentage module processes chest muscle tissue feature parameters extracted from acoustic signals using a forward propagation function to predict chest muscle percentage.

[0026] The leg muscle percentage module processes the leg muscle tissue feature parameters extracted from the acoustic signal using a forward propagation function to predict the leg muscle percentage.

[0027] The subcutaneous fat percentage module processes the subcutaneous adipose tissue feature parameters extracted from the acoustic signal using a forward propagation function to predict the subcutaneous fat percentage.

[0028] Among them, the characteristic parameters of chest muscle tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of chest muscle tissue; the characteristic parameters of leg muscle tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of leg muscle tissue; and the characteristic parameters of subcutaneous adipose tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of subcutaneous adipose tissue.

[0029] As a further improvement to this technical solution, in S2.2, a multi-input model architecture is designed to consider the influence of poultry pecking sounds, wing flapping sounds, and feather friction sounds on the model. Specifically:

[0030] Two parallel feature processing subnetworks were constructed: the first feature processing subnetwork takes acoustic feature parameters extracted from acoustic signals that are directly related to pectoral muscle tissue, leg muscle tissue, and subcutaneous adipose tissue as input; the second feature processing subnetwork takes real-time collected and separated poultry behavioral acoustic features as input, where poultry behavioral acoustic features include quantized parameters of pecking sounds, wing flapping sounds, and feather rubbing sounds.

[0031] Each of the two subnetworks performs forward propagation and feature transformation using its independent weights and biases;

[0032] A feature fusion layer is constructed to receive the high-order feature representations output by the two feature processing sub-networks, and to perform a weighted combination of the high-order feature representations through a fusion function.

[0033] As a further improvement to this technical solution, in step S2.3, poultry calls and footsteps are introduced into the loss function, specifically as follows:

[0034] The loss function consists of a basic error term and a noise regularization term: the basic error term is the standard mean square error, which is used to directly quantify the deviation between the model's predicted values ​​and the actual values ​​of breast meat percentage, leg meat percentage and subcutaneous fat percentage; the noise regularization term is constructed based on the poultry call and footstep features separated from the signal, and a penalty term is formed by calculating the distance between the noise feature and the model's predicted value.

[0035] As a further improvement to this technical solution, in S2.3, the gradient of the loss function with respect to the weights and biases is calculated using the backpropagation algorithm, specifically as follows:

[0036] Calculate the difference between the output layer activation value and the target output value, and output the error term of the output layer; for each hidden layer, its error term is calculated by multiplying the error term of the next layer by the transpose of the connection weight matrix, and then multiplying by the derivative of the activation function of this layer with respect to its weighted input;

[0037] Based on the activation output of each layer and the error term of the next layer, the gradient of the weight matrix of that layer is calculated; at the same time, the gradient of the bias of that layer is calculated by summing and averaging the error term of the next layer over all training samples.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This acoustic detection-based poultry live meat percentage, leg meat percentage, and subcutaneous fat percentage detection system utilizes a multilayer perceptron (MLP) neural network model, combined with various feature parameters such as spectral analysis and time-domain analysis. This system can learn and identify patterns closely related to breast meat percentage, leg meat percentage, and subcutaneous fat percentage in complex datasets. By estimating breast meat percentage, leg meat percentage, and subcutaneous fat percentage in live poultry through the MLP neural network model, it can handle data with complex intrinsic relationships, including the relationship between acoustic signals and the meat quality of different parts of the poultry. Acoustic signals are typically high-dimensional, and MLP can effectively process this type of high-dimensional input data and extract useful features. By introducing environmental factors (such as genetic factors, stocking density, and disease history) and poultry-specific noises (such as calls and footsteps) as additional inputs, the system can more comprehensively assess various factors affecting poultry growth, thereby improving the accuracy of prediction results.

[0040] 2. This acoustic-based system for detecting breast meat percentage, leg meat percentage, and subcutaneous fat percentage in live poultry utilizes acoustic sensing technology to acquire acoustic characteristic information from different parts of the poultry. This avoids the trauma and stress that may occur with traditional physical measurement methods, making the detection process more humane and efficient. By collecting and analyzing acoustic signals in real time, the system can quickly and accurately estimate the breast meat percentage, leg meat percentage, and subcutaneous fat percentage of poultry. For large-scale poultry farming, this not only improves production efficiency but also ensures the timeliness and reliability of the data. Attached Figure Description

[0041] Figure 1 This is an overall flowchart of the present invention;

[0042] The meanings of the labels in the diagram are as follows:

[0043] 1. Acoustic sensing unit; 2. Signal processing unit; 3. Data analysis unit. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0045] Please see Figure 1 As shown, an acoustic detection-based system for detecting the percentage of breast meat, leg meat, and skin fat in live poultry is provided, comprising:

[0046] The acoustic sensing unit 1 is used to collect sound wave signals reflected from different parts of live poultry to obtain acoustic characteristic information related to meat quality.

[0047] In this embodiment, the acoustic sensing unit 1 generates and transmits sound wave signals to different parts of the poultry via a transmitting module. These parts include the breast muscles, leg muscles, and surrounding tissues. A receiving module captures and collects the sound wave signals reflected back from these different parts of the poultry. Different types of tissues (such as muscle, fat, and bone) have different characteristics in reflecting and absorbing sound waves. Therefore, features related to breast meat, leg meat, and subcutaneous fat can be extracted from the acoustic signals. The transmitting module is used to irradiate the target area with sound wave signals of a specific frequency. These sound wave signals can penetrate different types of tissues and have different reflection characteristics depending on the nature of the tissue. The receiving module is used to capture these reflected signals.

[0048] The signal processing unit 2 is used to convert the analog signal containing acoustic characteristic information collected by the acoustic sensing unit 1 into a digital signal and extract the acoustic characteristic parameters related to the breast meat percentage, leg meat percentage and subcutaneous fat percentage.

[0049] In this embodiment, the collected analog signal containing acoustic characteristic information is converted into a digital signal. The digital signal can be easily subjected to Fourier Transform (FFT) to analyze its spectral characteristics. This is very useful for identifying the acoustic features of different tissue types. Digital signal processing allows the use of advanced algorithms to extract time-domain and frequency-domain features, such as short-time energy and zero-crossing rate. These features help distinguish different tissues such as muscle, fat, and bone. Digital signal processing enables real-time processing, instantly analyzing acoustic signals and drawing conclusions, which is crucial for real-time monitoring of poultry. The process includes the following steps:

[0050] S1.1 Convert the continuous-time signal into a discrete-time signal; according to the Nyquist sampling theorem, in order to reconstruct the original signal without distortion, the sampling frequency must be at least twice the highest frequency component of the signal;

[0051] S1.2 Quantizing discrete-time signals involves converting the amplitude of a discrete-time signal into discrete values; this means mapping continuous amplitude values ​​to a finite set of discrete values.

[0052] S1.3 Encode the quantized discrete values ​​and convert them into binary form.

[0053] Among them, the acoustic feature parameters extracted that are related to breast meat percentage, leg meat percentage and subcutaneous fat percentage specifically include: time domain features and frequency domain features.

[0054] Furthermore, time-domain features are extracted directly from the time series of the signal, and these features include mean, variance, and peak value.

[0055] Frequency domain features are extracted by performing a Fourier transform on the signal to obtain the spectrum. Frequency domain features include the center frequency and the frequency band energy ratio.

[0056] Based on the acoustic characteristic parameters of the analysis signal processing unit 2, the data analysis unit 3 estimates the breast meat percentage, leg meat percentage, and subcutaneous fat percentage of live poultry through the neural network model MLP.

[0057] In this embodiment, the data analysis unit 3 includes a breast fat percentage module, a leg fat percentage module, and a subcutaneous fat percentage module;

[0058] Among them, the breast muscle percentage module uses a neural network model MLP to analyze the acoustic feature parameters of the breast muscle tissue to estimate the proportion of poultry breast muscle tissue.

[0059] Furthermore, acoustic feature parameters represent time-domain and frequency-domain features;

[0060] The leg muscle percentage module estimates the proportion of poultry leg muscle tissue by analyzing the acoustic characteristic parameters of leg muscle tissue using a neural network model (MLP).

[0061] The subcutaneous fat percentage module estimates the proportion of subcutaneous fat tissue in poultry by analyzing the acoustic characteristic parameters of subcutaneous fat tissue using a neural network model (MLP).

[0062] MLP (Multi-Level Processing) is a type of feedforward neural network capable of learning complex nonlinear relationships. This makes it well-suited for processing data with complex intrinsic correlations, including the relationship between acoustic signals and the meat quality of different parts of poultry. MLPs can automatically learn useful features from raw data, reducing the need for manual feature engineering, which is particularly useful for high-dimensional data such as acoustic signals. For the poultry farming industry, a large amount of sample data can be collected over time, allowing for the training of more accurate models. The structure of an MLP can be adjusted according to actual needs, including increasing the number of hidden layers or changing the number of nodes in each layer, to adapt to different task complexities. Estimating the breast meat percentage, leg meat percentage, and subcutaneous fat percentage of live poultry using the MLP neural network model includes the following steps:

[0063] S2.1 Collect actual measurements and acoustic characteristic parameters of breast fat percentage, leg fat percentage, and subcutaneous fat percentage. The actual measurements are obtained through traditional physical measurement methods, and the data are divided into training set and validation set.

[0064] S2.2 Define a multilayer perceptron (MLP) neural network model for the breast meat percentage, leg meat percentage, and subcutaneous fat percentage modules. Each model contains an input layer, a hidden layer, and an output layer. The input layer receives acoustic signal data from poultry, while the output layer provides predicted values ​​for breast meat percentage, leg meat percentage, and subcutaneous fat percentage. To extend the functionality of the model, design a multi-input model architecture to consider the impact of poultry pecking sounds, wing flapping sounds, and feather friction sounds on the MLP neural network model.

[0065] The chest muscle percentage module processes chest muscle tissue feature parameters extracted from acoustic signals using a forward propagation function to predict chest muscle percentage.

[0066] The leg muscle percentage module processes the leg muscle tissue feature parameters extracted from the acoustic signal using a forward propagation function to predict the leg muscle percentage.

[0067] The subcutaneous fat percentage module processes the subcutaneous adipose tissue feature parameters extracted from the acoustic signal using a forward propagation function to predict the subcutaneous fat percentage.

[0068] Among them, the characteristic parameters of chest muscle tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of chest muscle tissue; the characteristic parameters of leg muscle tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of leg muscle tissue; and the characteristic parameters of subcutaneous adipose tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of subcutaneous adipose tissue.

[0069] Specifically, the multilayer perceptron (MLP) neural network model is defined for the chest fat percentage, thigh fat percentage, and subcutaneous fat percentage modules as follows:

[0070] Chest muscle percentage module:

[0071] ;

[0072] Leg fat percentage module:

[0073] ;

[0074] Subcutaneous fat percentage module:

[0075] ;

[0076] in, Indicates the predicted breast meat percentage; This indicates the predicted leg meat percentage; This indicates the predicted subcutaneous fat percentage; This represents the characteristic parameters of the chest muscle tissue extracted from the acoustic signal. , This represents the mean value of the chest muscle tissue. This represents the variance of the chest muscle tissue. This indicates the peak value of the pectoral muscle tissue. Indicates the central frequency of the pectoral muscle tissue. This indicates the frequency band energy ratio of the chest muscle tissue. This represents the transpose operation of the eigenvector; This represents the characteristic parameters of leg muscle tissue extracted from the acoustic signal. , This represents the mean value of leg muscle tissue. Represents variance. This indicates the peak value of the leg muscle tissue. This indicates the central frequency of the leg muscles. This indicates the frequency band energy ratio of leg muscle tissue; This represents the characteristic parameters of subcutaneous adipose tissue extracted from the acoustic signal. , This represents the mean value of subcutaneous adipose tissue. Represents variance. This indicates the peak value of subcutaneous adipose tissue. Indicates the central frequency of subcutaneous adipose tissue. This indicates the frequency band energy ratio of subcutaneous adipose tissue; Represents the forward propagation function of the model; Represents the set of weight matrices; A set of weight matrices.

[0077] If the input layer has n features, the output layer has m outputs (here m=3, corresponding to breast fat percentage, thigh fat percentage and subcutaneous fat percentage respectively); the hidden layer has h neurons;

[0078] The forward propagation expression from the input layer to the hidden layer is:

[0079] ;

[0080] ;

[0081] here, It is a weighted input to the hidden layer. It is the activation output of the hidden layer. It is the weight matrix from the input layer to the hidden layer. It is the bias vector from the input layer to the hidden layer. It is the sigmoid activation function. These parameters represent the mean, variance, peak value, center frequency, and frequency band energy ratio of chest muscle tissue, leg muscle tissue, or subcutaneous adipose tissue.

[0082] The forward propagation expression from the hidden layer to the output layer is:

[0083] ;

[0084] ;

[0085] here, It is a weighted input to the output layer. It is the activation output of the output layer. It is the weight matrix from the input layer to the hidden layer. It is the bias vector from the input layer to the hidden layer;

[0086] Furthermore, by incorporating multiple acoustic features into the model, noise can be better identified and processed, thereby improving the model's resistance to environmental noise. By analyzing wing-beating sounds, the model can distinguish which sounds are caused by the natural behavior of poultry, rather than other interfering factors. Different types of acoustic features carry different information; combining multiple acoustic features such as pecking sounds, wing-beating sounds, and feather rubbing sounds can provide richer input information, helping the model capture more details about the numerical characteristics of poultry. Each sound may be related to the body structure of poultry; pecking sounds may reflect the activity level of poultry, while wing-beating sounds may be related to muscle development. By fusing information from these different perspectives, breast meat percentage, leg meat percentage, and subcutaneous fat percentage can be predicted more accurately. Introducing multiple factors makes the model more robust, reducing prediction bias caused by changes in a single factor, which allows the model to maintain good prediction performance under different conditions. A multi-input model architecture is designed to consider the impact of poultry pecking sounds, wing-beating sounds, and feather rubbing sounds on the model, specifically:

[0087] Design a multi-input model architecture to consider the effects of poultry pecking sounds, wing flapping sounds, and feather rubbing sounds on the model. Specifically:

[0088] Two parallel feature processing subnetworks were constructed: the first feature processing subnetwork takes acoustic feature parameters extracted from acoustic signals that are directly related to pectoral muscle tissue, leg muscle tissue, and subcutaneous adipose tissue as input; the second feature processing subnetwork takes real-time collected and separated poultry behavioral acoustic features as input, where poultry behavioral acoustic features include quantized parameters of pecking sounds, wing flapping sounds, and feather rubbing sounds.

[0089] Each of the two subnetworks performs forward propagation and feature transformation using its independent weights and biases;

[0090] A feature fusion layer is constructed to receive the high-order feature representations output from the two feature processing sub-networks and to perform a weighted combination of the high-order feature representations through a learnable fusion function (another fully connected layer). The output of this fusion layer is then subjected to a final linear transformation and activation to generate more robust predictions of breast meat percentage, leg meat percentage, or subcutaneous fat percentage that simultaneously consider the acoustic characteristics of the target tissue and the noise interference from poultry behavior.

[0091] The specific formula is as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] in, This represents the predicted breast meat percentage after taking into account the sounds of poultry pecking, wing flapping, and feather rubbing. This represents the predicted leg meat percentage after taking into account the sounds of poultry pecking, wing flapping, and feather rubbing. This represents the subcutaneous fat percentage predicted after taking into account the sounds of poultry pecking, wing flapping, and feather rubbing. The characteristics of pecking sounds, wing flapping sounds, and feather rubbing sounds in poultry. , It's the sound of poultry pecking at food. It was the sound of wings flapping. It's the sound of feathers rubbing together; A sub-model representing the processing of acoustic signal features; This represents a sub-model that handles the sounds of poultry pecking, wing flapping, and feather rubbing. This represents the function that ultimately merges the outputs of these two sub-models.

[0096] S2.3. The mean squared error (MSE) is used as the loss function, which measures the difference between the model's predicted value and the actual label. The gradient of the loss function with respect to the weights and biases is calculated through the backpropagation algorithm. The parameters of poultry calls and footsteps are introduced into the loss function to reduce the impact of noise on the loss function.

[0097] Environmental noise from poultry is unavoidable; their calls and footsteps can be considered background noise. This noise may interfere with the effective acoustic features extracted by the signal processing unit. By introducing these parameters into the loss function, this noise can be better identified and filtered out, thereby improving the model's sensitivity to effective signals. The calls and footsteps of poultry may vary at different times. By considering these parameters in the loss function, the model's noise processing method can be dynamically adjusted, allowing the model to adapt to different noise environments. Introducing noise parameters helps the model distinguish between useful information and useless noise in the signal, thereby enhancing the model's ability to differentiate features. Specifically, the parameters for introducing poultry calls and footsteps into the loss function are as follows:

[0098] The loss function consists of a basic error term and a noise regularization term. The basic error term is the standard mean squared error (MSE), which is used to directly quantify the deviation between the model's predicted values ​​and the actual values ​​of breast meat percentage, leg meat percentage, and subcutaneous fat percentage. The noise regularization term is constructed based on the poultry calls and footsteps features separated from the signal. It forms a penalty term by calculating the distance between the noise feature and the model's predicted value. Its core purpose is to constrain the model's sensitivity to such highly variable environmental noise during training and prevent noise interference from causing the model to overfit or reduce its generalization ability.

[0099] The specific formula is as follows:

[0100] ;

[0101] ;

[0102] ;

[0103] in, Indicates the first The true breast fat percentage, thigh fat percentage, or subcutaneous fat percentage of each sample; Indicates the first The predicted values ​​for breast fat percentage, thigh fat percentage, or subcutaneous fat percentage for each sample; Represents the loss function; This indicates a penalty item, used to measure the impact of poultry calls and footsteps; This represents a hyperparameter used to control the strength of the penalty term; Indicates the first Characteristics of poultry calls and footsteps in a sample; Indicates the number of samples.

[0104] Furthermore, backpropagation is an efficient method for calculating gradients. It utilizes the chain rule to calculate gradients from intermediate results obtained during forward propagation, avoiding recalculation of identical values ​​and significantly reducing computational cost. By precisely calculating the gradient of each weight and bias, backpropagation ensures that the gradient of the loss function with respect to the model parameters is correct. This ensures that each gradient descent proceeds in the direction of minimizing the loss function, thereby improving the model's training performance. Specifically, the backpropagation algorithm calculates the gradient of the loss function with respect to the weights and biases as follows:

[0105] Calculate the difference between the output layer activation value and the target output value, and output the error term of the output layer; for each hidden layer, its error term is calculated by multiplying the error term of the next layer by the transpose of the connection weight matrix, and then multiplying by the derivative of the activation function of this layer with respect to its weighted input;

[0106] Based on the activation output of each layer and the error term of the next layer, the gradient of the weight matrix of that layer is calculated; at the same time, the gradient of the bias of that layer is calculated by summing and averaging the error term of the next layer over all training samples.

[0107] The specific calculation logic formula is as follows:

[0108] S2.31 Calculate the error term of the output layer :

[0109] ;

[0110] S2.32, Calculate the error term of the hidden layer:

[0111] For the Layer, error term Represented as:

[0112] ;

[0113] S2.33, Calculate the gradients of the weights and biases:

[0114] ;

[0115] ;

[0116] in, This represents the activation vector of the output layer; This represents the target output vector, i.e., the actual chest fat percentage, thigh fat percentage, or subcutaneous fat percentage. Indicates the output layer; Indicates the first layer to the first Transpose of the layer weight matrix; Indicates the first Layer weighted input About activation functions The derivative; Represents the loss function Compared to the first Layer weight matrix The partial derivatives; Indicates the number of training samples; It indicates that it is the first The training sample at the th ... Layer error terms;

[0117] S2.4 Train the multilayer perceptron (MLP) neural network model on the training set;

[0118] S2.5. Use a trained multilayer perceptron (MLP) neural network model to predict breast fat percentage, leg fat percentage, and subcutaneous fat percentage.

[0119] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A system for detecting the percentage of breast meat, leg meat, and skin fat in live poultry based on acoustic detection, characterized in that, include: Acoustic sensing unit (1), the acoustic sensing unit (1) is used to collect sound wave signals reflected from different parts of live poultry in order to obtain acoustic characteristic information related to meat quality; Signal processing unit (2), the signal processing unit (2) is used to convert the analog signal containing acoustic characteristic information collected by the acoustic sensing unit (1) into a digital signal and extract the acoustic characteristic parameters related to the breast meat percentage, leg meat percentage and subcutaneous fat percentage; The data analysis unit (3) estimates the breast meat percentage, leg meat percentage and subcutaneous fat percentage of live poultry through the neural network model MLP based on the acoustic characteristic parameters of the analysis signal processing unit (2). It also designs a multi-input model architecture based on the neural network model MLP to consider the influence of poultry pecking sounds, wing flapping sounds and feather friction sounds on the model.

2. The acoustic detection-based poultry live meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 1, characterized in that: The acoustic sensing unit (1) generates and sends sound wave signals to different parts of the poultry through the transmitting module, including the chest muscles, leg muscles and surrounding tissues, and captures and collects the sound wave signals reflected back from different parts of the poultry through the receiving module.

3. The acoustic detection-based poultry live meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 1, characterized in that: The signal processing unit (2) converts the collected analog signals into digital signals, including the following steps: S1.1 Convert a continuous-time signal into a discrete-time signal; S1.2 Quantize the discrete-time signal, that is, convert the amplitude of the discrete-time signal into discrete values; S1.3 Encode the quantized discrete values ​​and convert them into binary form.

4. The acoustic detection-based poultry breast meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 3, characterized in that: The acoustic feature parameters extracted by the signal processing unit (2) related to the percentage of breast meat, leg meat, and subcutaneous fat include: time-domain features and frequency-domain features.

5. The acoustic detection-based poultry live meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 1, characterized in that: The data analysis unit (3) includes a breast fat percentage module, a leg fat percentage module, and a subcutaneous fat percentage module; Among them, the breast muscle percentage module uses a neural network model MLP to analyze the acoustic feature parameters of the breast muscle tissue to estimate the proportion of poultry breast muscle tissue. The leg muscle percentage module estimates the proportion of poultry leg muscle tissue by analyzing the acoustic characteristic parameters of leg muscle tissue using a neural network model (MLP). The subcutaneous fat percentage module estimates the proportion of subcutaneous fat tissue in poultry by analyzing the acoustic characteristic parameters of subcutaneous fat tissue using a neural network model (MLP).

6. The acoustic detection-based poultry live meat breast meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 5, characterized in that: The data analysis unit (3) estimates the breast meat percentage, leg meat percentage, and subcutaneous fat percentage of live poultry using a neural network model (MLP), including the following steps: S2.1 Collect actual measurements and acoustic characteristic parameters of breast fat percentage, leg fat percentage, and subcutaneous fat percentage, and divide the data into training set and validation set; S2.2 Define a multilayer perceptron (MLP) neural network model in the breast meat percentage module, leg meat percentage module, and subcutaneous fat percentage module. Each model contains an input layer, a hidden layer, and an output layer. Extend the functionality of the model by designing a multi-input model architecture to consider the effects of poultry pecking sounds, wing flapping sounds, and feather friction sounds on the MLP neural network model. S2.

3. The mean squared error (MSE) is used as the loss function. The gradient of the loss function with respect to the weights and biases is calculated through the backpropagation algorithm. The parameters of poultry calls and footsteps are introduced into the loss function to reduce the influence of noise on the loss function. S2.4 Train the multilayer perceptron (MLP) neural network model on the training set; S2.

5. Use a trained multilayer perceptron (MLP) neural network model to predict breast fat percentage, leg fat percentage, and subcutaneous fat percentage.

7. The acoustic detection-based poultry live meat breast meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 6, characterized in that: In S2.2, the multilayer perceptron (MLP) neural network model is defined in the breast fat percentage module, leg fat percentage module, and subcutaneous fat percentage module as follows: The chest muscle percentage module processes chest muscle tissue feature parameters extracted from acoustic signals using a forward propagation function to predict chest muscle percentage. The leg muscle percentage module processes the leg muscle tissue feature parameters extracted from the acoustic signal using a forward propagation function to predict the leg muscle percentage. The subcutaneous fat percentage module processes the subcutaneous adipose tissue feature parameters extracted from the acoustic signal using a forward propagation function to predict the subcutaneous fat percentage. Among them, the characteristic parameters of chest muscle tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of chest muscle tissue; the characteristic parameters of leg muscle tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of leg muscle tissue; and the characteristic parameters of subcutaneous adipose tissue include the mean, variance, peak value, center frequency, and frequency band energy ratio of subcutaneous adipose tissue.

8. The acoustic detection-based poultry live meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 7, characterized in that: In step S2.2, a multi-input model architecture is designed to consider the effects of poultry pecking sounds, wing flapping sounds, and feather rubbing sounds on the model. Specifically: Two parallel feature processing sub-networks are constructed: the first feature processing sub-network takes acoustic feature parameters extracted from acoustic signals that are directly related to pectoral muscle tissue, leg muscle tissue, and subcutaneous adipose tissue as input; The second feature processing subnetwork takes the real-time collected and separated acoustic features of poultry behavior as input, which include the quantitative parameters of pecking sounds, wing flapping sounds and feather rubbing sounds. Each of the two subnetworks performs forward propagation and feature transformation using its independent weights and biases; A feature fusion layer is constructed to receive the high-order feature representations output by the two feature processing sub-networks, and to perform a weighted combination of the high-order feature representations through a fusion function.

9. The acoustic detection-based poultry live meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 8, characterized in that: In S2.3, poultry calls and footsteps are introduced into the loss function, specifically as follows: The loss function consists of the basic error term and the noise regularization term: The basic error term is the standard mean square error, which is used to directly quantify the deviation between the model's predicted values ​​and the actual values ​​of breast meat percentage, thigh meat percentage, and subcutaneous fat percentage. The noise regularization term is constructed based on the poultry calls and footsteps features separated from the signal, and a penalty term is formed by calculating the distance between the noise feature and the model prediction.

10. The acoustic detection-based poultry live meat percentage, leg meat percentage, and skin fat percentage detection system according to claim 9, characterized in that: In step S2.3, the gradient of the loss function with respect to the weights and biases is calculated using the backpropagation algorithm, specifically as follows: Calculate the difference between the output layer activation value and the target output value, and output the error term of the output layer; for each hidden layer, its error term is calculated by multiplying the error term of the next layer by the transpose of the connection weight matrix, and then multiplying by the derivative of the activation function of this layer with respect to its weighted input; Based on the activation output of each layer and the error term of the next layer, the gradient of the weight matrix of that layer is calculated; at the same time, the gradient of the bias of that layer is calculated by summing and averaging the error term of the next layer over all training samples.

Images