State monitoring method based on prior perception calibration and dynamic mapping network, storage medium and equipment

Abstract

The invention discloses a state monitoring method based on prior perception calibration and a dynamic mapping network, a storage medium and equipment, and belongs to the technical field of artificial intelligence and animal state monitoring. The problem that the recognition effect of an existing recognition method based on acoustic recognition needs to be improved is solved. A sound signal recognition network is adopted for recognition, the sound signal recognition network is a deep neural network comprising a PACConv module and based on GDDMCB, the PACConv module separates input features into a related priori subset and a context subset by simulating a mechanism of focusing high-value information of a biological system, differentiation processing and self-adaptive calibration are carried out, and a sound signal recognition result is obtained. Therefore, the model is stimulated to autonomously learn high-value prior information, and meanwhile, a potential context dependency relationship is fully utilized; and the GDDMCB module carries out dynamic mapping by utilizing a Gaussian derivative function with learnable parameters according to semantic differences among channels of the high-dimensional feature vector so as to enhance the category decision-making capability of the model.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and animal condition monitoring technology, specifically relating to a method, storage medium, and device for monitoring animal health status based on acoustic signal classification. Background Technology

[0002] In modern animal husbandry, accurate and efficient health monitoring of farmed animals such as broilers is crucial for ensuring the industry's economic benefits, maintaining animal welfare, and guaranteeing public food safety. Traditional monitoring methods, such as veterinary inspections and laboratory tests, have inherent drawbacks such as high subjectivity, delayed response, and high labor intensity, making it difficult to meet the needs of large-scale, real-time, and automated monitoring.

[0003] As a non-invasive information carrier, the sound signals emitted by broiler chickens can continuously and cost-effectively reflect their physiological and health status. Therefore, acoustic monitoring methods have become a research hotspot in this field. With the development of artificial intelligence technology, the mainstream approach has evolved from early signal physical feature analysis to a model that uses deep learning models, especially convolutional neural networks (CNNs), to classify acoustic features. These methods typically include three stages: acoustic data acquisition and preparation, feature learning and representation, and category decision. First, the raw sound signal is converted into an acoustic feature set containing prior knowledge using time-frequency analysis algorithms (such as Mel spectrum, Mel frequency-cephalic coefficients (MFCC), etc.). Then, the powerful nonlinear fitting capability of CNNs is used to learn these features, generating a high-dimensional feature vector for classification. Finally, a classifier maps this vector to a specific health status category.

[0004] Although existing CNN-based methods have improved the accuracy of monitoring to some extent, they still face a series of deep-seated technical bottlenecks in core stages such as feature learning and representation, and category decision-making, which seriously restrict their performance and practical application.

[0005] (1) Insufficient and diluted representation of high-value prior information: In actual aquaculture environments, the collected sound signals inevitably contain a large amount of redundant or irrelevant information. Existing time-frequency analysis algorithms usually cannot distinguish the value of information and will convert all information into acoustic features. When standard CNN models process these features, their inherent non-selective convolution operation will perform indiscriminate transformation on all features within the receptive field. This mechanism leads to a severe dilution of the effective representation density of weak but critical acoustic patterns (i.e., high-value prior information) that are highly correlated with health status, thus limiting the model's recognition accuracy.

[0006] (2) Insufficient utilization of potential contextual dependencies: Some studies attempt to eliminate redundant or irrelevant features at the input stage through feature selection algorithms in order to eliminate their negative impact. However, these seemingly irrelevant features often provide the necessary potential contextual dependencies for the model to understand high-value information. For example, stable patterns of background noise can help the model better identify abnormal coughs. Directly removing these features will destroy the integrity of the information, cause potential misjudgment effects, and weaken the model's generalization ability.

[0007] (3) Static and rigid category decision-making mechanism: In the category decision-making stage, existing methods generally adopt a static mapping structure composed of fully connected layers. However, the high-dimensional feature vectors extracted by multiple convolutions abstractly represent acoustic patterns with different semantic properties at different channel dimensions. The static mapping structure cannot fully consider the differences of these acoustic patterns and adopts a "one-size-fits-all" approach to mapping, which makes it impossible for the model to fully understand the complex correspondence between high-dimensional feature vectors and sound signal types, thus limiting the accuracy of the final decision.

[0008] (4) Lack of interpretability in the model decision-making process: The "black box" nature of CNN models makes their decision-making process abstract and opaque. In fields such as agriculture and veterinary medicine, which rely heavily on professional knowledge and trust, this lack of interpretability becomes a major obstacle to the promotion and application of the technology. Users cannot understand the basis for the model's diagnosis, making it difficult to build trust in automated systems.

[0009] In conclusion, there is an urgent need for an innovative method that can fundamentally solve the above problems in order to achieve more accurate, robust, and reliable intelligent broiler health monitoring. Summary of the Invention

[0010] This invention aims to address the problem that existing acoustic recognition-based recognition methods have insufficient recognition performance.

[0011] A state monitoring method based on prior perception calibration and dynamic mapping network inputs acoustic features extracted from the original audio signal into a sound signal recognition network (PPCDM) for recognition. The recognition includes:

[0012] The input acoustic features are sequentially learned layer by layer through multiple AFRB units. Each AFRB unit includes multiple AFRL layers and one GMP layer. The AFRL layer is used for feature extraction. The beginning and end of the AFRL are PWConv modules, and the middle part includes at least one PACConv module and a multilayer perceptron (MLP). The processing of the PACConv module includes:

[0013] S101, based on The features input to PACConv It is divided into two sparse feature subsets, namely and Remove all zero-value feature channels from the two feature subsets, and retain the non-zero feature channels to obtain... and ; It is a set of binary vectors, and , correspond The channel dimension, Any element in the set satisfies QM is a constant between 0 and 1. express The number of 1s in the encoding;

[0014] S102. Perform two-stream information representation: The two-stream information representation adopts a two-stream architecture, in which the main branch uses convolutional layers. Feature extraction is performed to obtain acoustic representation. The identity flow branch will Transmitted directly without any transformation, as a feature Then splice them together. and Obtain features ;

[0015] S103, targeting The weight vector is obtained based on pointwise convolution. And thus obtain ;

[0016] After processing by multiple AFRB units, a GAP layer is used for spatial dimensionality reduction. Finally, a dynamic mapping classification block based on the Gaussian derivative function completes the mapping from the feature space to the class space, achieving the final classification decision.

[0017] Furthermore, the mapping learning function used by the dynamic mapping classification block of the Gaussian derivative function in mapping the input features adopts the second derivative function based on the Gaussian function.

[0018] Furthermore, the processing procedure for dynamic mapping classification blocks based on Gaussian derivative functions includes:

[0019] Let the input feature vector be denoted as , This represents the channel dimension of the feature vector; for A two-stage successive mapping is performed, namely GDDMCB1 and GDDMCB2, to complete the category decision;

[0020] The overall calculation expression for GDDMCB1 is: , for Features on the i-th channel The mapping learning function for feature vector channel i is represented by a second-order derivative function based on a Gaussian function; and BN These represent the ReLU activation function and the batch normalization layer, respectively. This indicates the output of GDDMCB1. Indicates the channel dimension;

[0021] During the second mapping, GDDMCB2 will use the output of GDDMCB1. Learning functions through mapping The layer maps the intermediate feature semantic space to the final category output space, and uses the Softmax function to obtain the probability values ​​of the decision results of PPCDM. , This indicates the number of categories of sound signals.

[0022] Furthermore, channel dimension .

[0023] Furthermore, the mapping learning function The processing procedure is as follows:

[0024] Using the following form of mapping learning function Perform mapping:

[0025]

[0026] in, Indicates standard deviation; and These are two learnable parameters, determined through the training process of PPCDM;

[0027] Furthermore, in S102, the main branch uses convolutional layers with a kernel size of 3. Feature extraction is performed.

[0028] Furthermore, targeting The weight vector is obtained based on pointwise convolution. The process includes:

[0029] First, capture via GAP. The system first generates a global channel descriptor; then, it establishes complex nonlinear dependencies between channels through two pointwise convolutions and a ReLU activation function, and finally transforms these dependencies into weight vectors using Softmax. .

[0030] Furthermore, the voice signal recognition network PPCDM includes four AFRB units, and one PACConv module is set in the middle part of each AFRL layer in the AFRB unit.

[0031] A computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the aforementioned state monitoring method based on prior perception calibration and dynamic mapping network.

[0032] A state monitoring device based on prior perception calibration and dynamic mapping network is disclosed. The device includes a processor and a memory. The memory stores at least one instruction, which is loaded and executed by the processor to implement the state monitoring method based on prior perception calibration and dynamic mapping network.

[0033] Compared with the prior art, the beneficial effects of the present invention include:

[0034] (1) Improved efficiency and robustness of feature representation: Through the PACConv module, the present invention can autonomously focus on high-value prior information without pre-removing any information, and effectively utilize potential contextual dependencies to establish a more robust and accurate feature representation, fundamentally solving the problems of information dilution and context loss.

[0035] (2) Enhanced class decision-making ability of the model: With the help of the dynamic mapping mechanism of the GDDMCB module, the present invention can perform adaptive classification based on the inherent semantic differences of high-dimensional features, deeply understand the complex nonlinear relationship between features and categories, and significantly improve classification accuracy and overall model performance.

[0036] (3) Achieving an excellent trade-off between high performance and lightweight design: This invention avoids the introduction of complex external sub-networks (such as attention modules) through innovation at the network basic computing unit level. While achieving excellent recognition accuracy, it maintains low computational overhead and model parameter quantity, enhancing the applicability and deployability of the model on resource-constrained edge devices.

[0037] (4) Improved reliability of monitoring system: The framework proposed in this invention includes a dual interpretability analysis method, which can visualize the internal feature representation and decision-making process of the model, effectively alleviate the "black box" problem of existing deep learning methods, and improve the transparency and reliability of the model in practical applications. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall process of the interpretable broiler health monitoring framework (FBPPCDM) based on prior perception calibration and dynamic mapping network proposed in this invention.

[0039] Figure 2 This is a schematic diagram of the overall structure of the sound signal recognition network (PPCDM) based on prior perception calibration and dynamic mapping proposed in this invention.

[0040] Figure 3 This is a schematic diagram of the structure of the prior perception calibration convolution (PACConv) module proposed in this invention.

[0041] Figure 4 This is a schematic diagram of the calculation of the Gaussian derivative-based dynamic mapping classification block (GDDMCB) proposed in this invention.

[0042] Figure 5 The results are experimental comparisons of the performance of the method of this invention with other existing methods on multiple datasets.

[0043] Figure 6 This is a confusion matrix of the recognition results of the method of this invention and other methods on the MSAF dataset.

[0044] Figure 7 These are the ablation experiment results of the core module of this invention.

[0045] Figure 8 This is the t-SNE visualization result corresponding to the ablation experiment.

[0046] Figure 9 These are the results of a comparative experiment on the impact of different quota factor (QM) settings on the performance of the model of this invention.

[0047] Figure 10 This is the result of visualizing the features learned at each stage of the PPCDM network of this invention using t-SNE technology.

[0048] Figure 11 The results are visualizations of the feature vectors output by each stage of the PPCDM network of this invention. (a) represents the visualization result of the output features of AFRB 1, (b) represents the visualization result of the output features of AFRB 2, (c) represents the visualization result of the output features of AFRB 3, (d) represents the visualization result of the output features of AFRB 4, (e) represents the visualization result of the output features of GDDMCB, and (f) represents the visualization result of the output features of GDDMCB after Softmax. Detailed Implementation

[0049] This invention first proposes a sound signal recognition network based on prior perception calibration and dynamic mapping (PPCDM). The core innovation of PPCDM lies in two key technical components:

[0050] (1) Prior-Aware Calibration Convolution (PACConv) Module: This module is designed for the feature learning and representation stages. Its core idea stems from the simulation of the survival mechanism of complex biological systems, namely, focusing on high-value information with limited resources while maintaining awareness of the environmental context. PACConv introduces a "quota factor" constraint to encourage the model to autonomously identify and separate the "relevant prior feature subset" most relevant to health status from the input features, while retaining the remaining information as the "context feature subset". Subsequently, the two subsets are differentiated through a two-stream architecture—deeply refining the prior information and fully preserving the context information. Finally, through an adaptive channel calibration mechanism, the representation strength of key features is enhanced, thereby maximizing the representation density of high-value information without losing context information.

[0051] (2) Gaussian Derivative-based Dynamic Mapping Classification Block (GDDMCB): This module is designed for the category decision stage to address the limitations of traditional static mapping. GDDMCB employs a mapping learning function based on the second-order derivative of a Gaussian function and introduces two learnable parameters (α and β) for each feature channel. These parameters enable the mapping function to dynamically adjust its mapping granularity (focusing on macro trends or local details) and sensitive interval (focusing on the numerical range with the richest information) according to the semantic characteristics of the acoustic patterns represented by different channels. In this way, GDDMCB can clearly and adaptively represent the complex mapping relationship from the high-dimensional feature space to the final category space, significantly improving the model's decision-making ability.

[0052] Building upon this, the present invention further provides a complete interpretable broiler health monitoring framework (FBPPCDM), which integrates data processing methods, the aforementioned PPCDM network, and a dual interpretability analysis method (DIAM), realizing a complete solution from data input to reliable result output.

[0053] The following detailed description is provided in conjunction with specific implementation methods.

[0054] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0055] This implementation describes a state monitoring method based on prior perception calibration and dynamic mapping networks. Using broiler health monitoring as an example, this method achieves monitoring through the interpretable broiler health monitoring framework based on prior perception calibration and dynamic mapping networks (FBPPCDM), and includes the following steps:

[0056] (1) Data collection:

[0057] The raw audio signals of the broiler chickens are acquired using acoustic sensors.

[0058] (2) Dataset construction:

[0059] The original audio signal is preprocessed (e.g., segmentation, annotation, normalization), and then the prior information such as the spectrum and perceived loudness reflecting the signal type in the original audio signal is quantized into acoustic feature values ​​by the Mel-Spectrum and Dimension Truncation Strategy (MSDTS) to form a prior acoustic feature dataset containing multiple acoustic feature vectors.

[0060] (3) Use the constructed dataset to train the proposed sound signal recognition network (PPCDM) and save the optimal model parameters.

[0061] The core of this invention is a Pronunciation Processing (PPCDM) network for voice signal recognition, the detailed structure of which is as follows: Figure 2 As shown in Table 1, the specific parameter configuration is as follows. The PPCDM mainly consists of an input layer, four cascaded acoustic feature representation blocks (AFRBs), a global average pooling layer (GAP), and a Gaussian derivative-based dynamic mapping classification block (GDDMCB). It should be noted that this structure can be modified according to different types of sound signals, such as by adding other modules or increasing the number of AFRBs.

[0062] The input acoustic features (prior acoustic features) are sequentially learned and abstracted through four AFRB units. Each AFRB includes multiple acoustic feature representation layers (AFRL) and a global max pooling layer (GMP).

[0063] The AFRL is the core feature extraction unit. Its beginning and end parts use pointwise convolution (PWConv), while the middle part includes at least one PACConv module (a priori perceptual calibration convolution module) and a multilayer perceptron (MLP). Specifically, the PWConv in the beginning part transforms the dimension of the acoustic feature vector input to the AFRL, and the PWConv in the end part transforms the dimension of the acoustic feature vector output from the middle part. The GMP layer downsamples the feature map to reduce computational complexity and highlight key features. It should be noted that in this embodiment, only one PACConv module is designed in the middle part of the AFRL. This is because research in this invention has shown that for the audio signal processed by this invention, a single PACConv module can achieve very good results; adding more PACConv modules has very limited improvement in performance and increases computational overhead. However, in practice, the number of PACConv modules included in the middle part can be designed according to actual needs.

[0064] After four AFRB processes, the Global Average Pooling (GAP) layer performs spatial dimensionality reduction on the generated high-dimensional feature vectors. Finally, the GDDMCB layer completes the mapping from the feature space to the class space, achieving the final classification decision.

[0065] Table 1 Detailed PPCDM Parameter Settings

[0066]

[0067] In Table 1, In and Out represent the channel dimension settings of AFRB, "C" represents the number of sound types, and "L" represents the feature dimension. "" indicates that AFRL is stacked X times consecutively, and "QM" indicates the credit limit factor.

[0068] The prior perception calibration convolution module, PACConv module, is as follows:

[0069] To address the shortcomings of existing methods in the feature learning and representation stages, such as insufficient representation of high-value prior information and inadequate utilization of rich potential contextual dependencies, this invention proposes a novel Prior-Aware Calibration Convolution (PACConv) as the core component for PPCDM feature learning and representation, thereby enhancing its feature representation capabilities. Its overall structure is as follows: Figure 3 As shown. The overall processing of PACConv includes steps such as information perception and separation, dual-stream information representation, and representation calibration and enhancement.

[0070] Let the acoustic feature vector of the input PACConv be denoted as ,in denotes the channel dimension, denotes the dimension of the feature vector. It should be noted that for the first PACConv in the first AFRL, its input is the prior acoustic feature vector, and the subsequent PACConv are all the feature vectors output by the previous layer. For the convenience of description, the acoustic feature vectors are uniformly denoted as .

[0071] It should also be noted that for the convenience of representation, the bias term is omitted in the subsequent formula description of the convolution process, but it is actually included in the actual process.

[0072] Step 1: Information Perception and Separation (Information Perception Module, IPM). IPM aims to identify and separate high-value prior information and retain context information. The design idea of IPM comes from the survival mechanism of complex biological systems. Specifically, in the case of limited resources, biological systems tend to focus on high-value information and context information that most affect survival or reproduction, interpret and judge high-value information more precisely, and dynamically adjust behavior or physiological state according to this information. For example, in the case of hunger or threat, organisms may quickly respond by identifying signals of food sources or potential threats, and decide the best strategy for subsequent movement according to the specific context of the environment (such as light, terrain, activities of surrounding animals and plants, etc.), so as to improve the chance of survival. In addition, existing deep learning research shows that the channel dimension of the feature vector is the basis for the model to understand high-value information. Inspired by this, IPM simulates the survival mechanism of biological systems focusing on high-value information in complex environments by setting a limited feature representation quota, enabling the model to actively identify and separate high-value prior information most relevant to the health status of broilers, while structurally retaining context information that can reveal potential context dependencies, providing an important basis for fully mining potential context dependencies in subsequent calculation steps.

[0073] Specifically, the feature representation quota is generated by the quota mask and is , a set of binary vectors. Any element in satisfies Equation (2), indicating that the model actively selects the feature channels to be represented in the current feature representation stage according to the quota factor QM (Quota Metric), encodes them as 1, and encodes the remaining channels as 0. Here, QM (0 < QM < 1) is a hyperparameter used to stimulate the autonomy of the model, a constant between 0 and 1, and the relationship with is shown in Equation (2), denotes The channels are selected autonomously by the model. For example, when QM is 0.7, it means that 70% of the vectors in Q are 1, and the remaining 30% of the vectors are 0.

[0074]

[0075] Where Count is a counting function.

[0076] Next, the multichannel is multiplied element by element. The features are divided into two sparse feature subsets. Specifically, features actively selected by the model and subsequently learned are assigned to the relevant prior feature subset. In this context, the remaining feature vectors are used in subsequent computation steps to provide potential contextual dependencies and are assigned to the contextual feature subset. The calculation process is shown in equations (3) and (4). Since there are 0 values ​​in both feature subsets at this time, considering the computational efficiency, we remove all 0-value feature channels from these two sparse feature subsets and only retain the non-zero feature channels. This operation is denoted as... The final output of IPM is: the prior feature vector (or simply the prior vector). and context feature vectors (or simply context vectors) The calculation process is shown in equations (5) and (6). For a PPCDM composed of multiple PACConvs, the IPM will be determined by a mask. The generated feature characterization limit (i.e., it can only process features represented by the limit factor) The limited feature information in the feature channels allows the model to proactively select high-value feature channels that can represent the health status of broilers in the current feature representation stage, while the feature vectors corresponding to the remaining channels are used to provide potential contextual dependencies. Furthermore, during the forward propagation process, the prior acoustic feature vectors, after passing through multiple PACConvs, are masked... This allows the local neighborhoods of each feature channel to continuously reorganize, ensuring that the model cannot rely on the acoustic features of certain fixed channels, which contributes to the robustness of PPCDM feature representation.

[0077]

[0078]

[0079] In the formula, This indicates element-wise multiplication.

[0080] Through Obtained by performing element-wise multiplication with the binary matrix Q, retaining the channels with a value of 1 in Q and setting the channels with a value of 0 to zero. This means swapping the 0s and 1s in the Q elements; that is, elements that were originally 1 become 0, and elements that were originally 0 become 1. Then, element-wise multiplication is performed to obtain... Because of the zeroing operation, and Since it contains a value of 0, we have the extraction operations in equations (5) and (6).

[0081] Step 2: Dual-Stream Information Representation (DSIR). DSIR employs a dual-stream architecture (mainstream branch and identity stream branch) to simultaneously refine prior features deeply and preserve the underlying contextual relationships within contextual features. Specifically,

[0082] The mainstream branch uses a convolutional kernel of size 3 for efficient extraction. More refined acoustic patterns most relevant to broiler health were identified, and these acoustic patterns were characterized as .

[0083] For characterizing potential contextual relationships Unlike existing approaches that first filter features and then directly eliminate them, we use an identity flow branch to directly pass the features without any transformation. This approach represents the density without sparsity of prior information, while simultaneously integrating the implied contextual dependencies with... The fusion of prior acoustic modes results in more diverse and robust acoustic characteristics. The specific process is shown in equations (7) and (8).

[0084]

[0085] in, This represents a convolutional filter of size 3. This represents the weights of the convolution kernel.

[0086] Step 3: Representation Calibration and Enhancement (RCE). The changes in channel dependencies generated in the first two computational steps cause confusion in the model regarding the semantic importance of prior vector and context vector representations, which is detrimental to the accuracy of feature learning and representation. Therefore, we first capture the channel dependencies through GAP. The global channel descriptor is then used. Next, complex nonlinear dependencies between channels are established using two pointwise convolutions (PWConv) and a ReLU activation function, and these dependencies are transformed into weight vectors using Softmax. . Each element in represents The importance and contribution of each channel's information are shown in Equation (9). Finally, With weight vector Perform point-by-point calculations and adaptive calibration based on contribution. The strength of the representation of prior information and contextual information is increased to enhance the accuracy of the model's feature representation and form the output of PACConv. After the model is trained, a dynamic calibration process is performed when faced with a new input feature vector, as shown in equation (10). and represents pointwise convolution, and GAP represents global average pooling.

[0087]

[0088] In summary, PACConv efficiently learns features from prior acoustic features through a three-stage collaborative computation process, enabling the model to autonomously and dynamically re-evaluate and combine features during feature learning and representation. This provides health monitoring models with more robust and accurate feature extraction capabilities, effectively improving their accuracy and generalization in health monitoring.

[0089] Step 4: Processing of Gaussian derivative-based dynamic mapping classification blocks (GDDMCB).

[0090] To address the limitations of existing methods in failing to adequately consider the semantic hierarchy differences between different channels of high-dimensional feature vectors during the category decision stage, this invention proposes a Gaussian Derivative-based Dynamic Mapping Classification Block (GDDMCB) and integrates it into PPCDM to improve its decision-making capabilities. The overall structure is as follows: Figure 4 As shown. We denote the high-dimensional feature vector of the input GDDMCB as... , This represents the channel dimension of a high-dimensional feature vector. For GDDMCB completes the category decision through a two-stage successive mapping (GDDMCB1 and GDDMCB2).

[0091] Specifically, GDDMCB1 learns functions through mapping. use The semantic differences of different channels are dynamically adjusted to modify the abstract representation of their mapping relationship with sound signal categories. The high-dimensional feature space is initially mapped to an intermediate feature semantic space to capture more fine-grained feature mapping relationships and enhance the model's understanding of semantic hierarchical differences. Then, batch normalization is used to normalize the mean and variance of the features to reduce the bias of the feature distribution. ReLU is used for global nonlinear transformation to improve the ability to represent complex mapping relationships. The overall calculation expression of GDDMCB1 is shown in Equation (11).

[0092]

[0093] in, The mapping learning function represents the high-dimensional feature vector channel i; for Features on the i-th channel. and BN These represent the ReLU activation function and the batch normalization layer, respectively. This indicates the output of GDDMCB1. Indicates the channel dimension, and The quantitative relationship is shown in equation (12).

[0094] To achieve the above goals, a mapping learning function is used. The second derivative function based on the Gaussian function is defined as shown in equation (13). This represents the standard deviation. Since the original audio data has already been normalized, this... In order to give The ability to adaptively adjust based on channel specificity introduces two learnable parameters. and As shown in equation (14). Specifically, It is used to control whether the model focuses more on capturing coarse-grained macro trends or fine-grained local details, which makes each Become an expert for specific acoustic semantic granularity. Considering that the feature values ​​corresponding to the acoustic semantics carried by a feature channel often concentrate within a specific range, therefore... Will The most sensitive activation points are moved to the center of this specific range, ensuring the model responds to the unique characteristics of each channel within its most informative numerical range. Through backpropagation, the model learns the optimal activation points for the current input information. and Values ​​are abstract representations that enable dynamic capture and adjustment of the mapping relationship between the target category and the target category.

[0095]

[0096] During the second mapping, GDDMCB2 will use the output of GDDMCB1. Learning functions through mapping The layer maps the intermediate feature semantic space to the final category output space, and uses the Softmax function to obtain the probability values ​​of the decision results of PPCDM. ,here This indicates the number of sound signal categories that reflect the health status of broiler chickens.

[0097] There is a close collaborative relationship between the two core modules, PACConv and GDDMCB. PACConv, through its unique three-stage processing, generates a semantically rich and well-structured high-dimensional feature vector, where different channels clearly carry different levels of prior and contextual information. Meanwhile, GDDMCB's dynamic mapping mechanism can precisely capture and utilize the subtle semantic differences between these channels, something standard static classifiers cannot achieve. Working together, they form a complete and self-consistent information processing chain from efficient feature representation to accurate dynamic decision-making.

[0098] Performance testing and interpretability analysis: The performance of the optimal model is evaluated on the test set, and the decision-making process of the model is analyzed using the dual interpretability analysis method (DIAM) to output monitoring results with both high accuracy and reliability.

[0099] The dual interpretability analysis consists of two parts, as follows:

[0100] DIAM 1: Visual Analysis of Feature Representation Capabilities Based on t-Distributed Stochastic Neighborhood Embedding Algorithm. To reveal and analyze the ability of PPCDM to extract and distinguish acoustic features in the feature learning, representation, and category decision stages, the features learned at each stage of PPCDM are visualized as a low-dimensional feature space using the t-distributed stochastic neighborhood embedding algorithm (t-SNE), intuitively explaining its feature extraction and classification capabilities.

[0101] DIAM 2: Visualization and Analysis of Output Feature Vectors at Each Stage of the Acoustic Model. This section visualizes the output feature vectors of the PPCDM feature learning and representation stage and the category decision stage, and conducts interpretability analysis on the features extracted and learned at each stage of the model. This further enhances the understanding of the model's interpretability analysis of its prior information learning and representation of broiler sound signals and its category decision-making mechanism.

[0102] The trained sound signal recognition network PPCDM is obtained through the above steps.

[0103] (5) Based on the trained sound signal recognition network PPCDM, the audio signals of broilers are identified to determine the health status of broilers.

[0104] It should be noted that although this embodiment is used for the identification of broiler audio signals to monitor the health status of broilers, it can also be used for the identification of audio signals of other animals to monitor the health status of animals.

[0105] Example

[0106] This embodiment provides a non-limiting example for verifying the method of the present invention, describing the complete process from data acquisition to dataset construction.

[0107] Data Collection: The study subjects were arbor broilers raised on a farm in Heilongjiang Province, China. A total of 200 healthy broilers, both those diagnosed with bronchitis and those in the vocal organ maturity stage (week 2) and immune system transition stage (week 3), were selected. High-precision microphones were used to collect the sound signals of individual broilers in a controlled environment, recording 5 minutes per broiler to avoid sound aliasing.

[0108] Data preprocessing: Through collaboration between human and veterinary experts, eight types of sound signals that clearly reflect the health status of broilers were segmented and labeled from the original recordings. These included coughing, wing flapping, snoring, clucking, and croaking and whine sounds from two different growth stages. The segmented sound segments were then normalized to zero mean.

[0109] Dataset Construction: An acoustic feature dataset (named MSAF) was constructed using a Mel Spectrogram and Dimensionality Truncation Strategy (MSDTS) approach. To ensure consistent dimensionality in the input model, the dataset was truncated using the lowest dimensionality sample among all samples. Finally, the dataset was divided into training, validation, and test sets in a 7:2:1 ratio.

[0110] A series of experiments have demonstrated the effectiveness and advancement of the method (PPCDM) described in this invention.

[0111] 1. Efficacy verification of the core module (ablation experiment):

[0112] Ablation experiments were conducted to independently verify the necessity of PACConv and GDDMCB. Figure 7 As shown, Figure 7In the diagram, (a)-(d) correspond to the ablation experiments for the FAF, SAF, MSAF, and MCAF datasets, respectively. Without changing the other network structures, PACConv was replaced with a standard convolutional layer (denoted as PPCDM-P), and GDDMCB was replaced with a traditional fully connected classification block (denoted as PPCDM-G). Experimental results show that removing any core module leads to a significant decrease in model performance. On the MSAF dataset, the accuracy of the complete PPCDM is 99.14%, while it drops to 96.32% after removing PACConv, and to 97.39% after removing GDDMCB. Figure 8 The t-SNE visualizations shown in (a)-(c) further reveal its underlying mechanism:

[0113] After removing PACConv (retaining only GDDMCB), the feature space still maintains a certain degree of inter-class separability, indicating that GDDMCB can still play a role in the decision-making stage. However, due to the deterioration in the quality of the input features, the class boundaries become blurred.

[0114] After removing GDDMCB (retaining only PACConv), although PACConv outputs high-quality features, the lack of dynamic mapping capabilities leads to significant aliasing and disorder of sample points from different classes in the feature space, resulting in a significant reduction in inter-class separability. This result clearly demonstrates that PACConv plays an indispensable and crucial role in the feature learning and representation stage, while GDDMCB plays a crucial role in the class decision stage. Together, they achieve excellent classification performance. Furthermore, Figure 9 The experimental results show that different QM values ​​have varying impacts on model performance. This indicates that the constraint of feature representation quota plays a crucial role in the performance of PACConv and PPCDM. The model performs best when QM is 0.6, indicating that PPCDM can effectively represent high-value prior information while fully utilizing latent contextual dependencies to maintain high feature representation density. To ensure experimental consistency, we set the QM value to 0.6. Furthermore, compared to model recognition performance at other QM values, the model performance is lowest when QM is 1. In fact, when QM is 1, due to the lack of feature representation quota constraint, PACConv degenerates into a standard convolution, only performing indiscriminate transformations on all prior acoustic features within the receptive field, failing to perceive the high-value prior information contained in the prior acoustic features, and unable to fully utilize latent contextual dependencies to build more robust feature representations, thus leading to a decrease in PPCDM performance. In summary, the experimental results in this section further validate the effectiveness of PACConv.

[0115] 2. Performance comparison analysis with existing technologies:

[0116] The PPCDM of this invention was comprehensively compared with seven methods, including machine learning models (SVM, Random Forest) and various advanced deep learning models (LiConformer, MWANet, RMT, Transformer, etc.). Figure 5 As shown in Table 2, in Figure 5 On the four datasets shown in (a)-(d) – FAF Dataset, SAF Dataset, MSAF Dataset, and MCAF Dataset – PPCDM achieved the best performance across all evaluation metrics, including accuracy (Acc), precision (MP), recall (MR), and F1 score (MF1). Taking the MSAF dataset as an example, PPCDM achieved an accuracy of 99.14%, which is 0.79% higher than the second-best RMT model and significantly higher than other methods. Figure 6 The confusion matrix shown further indicates that (a)-(h) correspond to the confusion matrices of the present invention, Liconformer, MWANet, Transformer, SVM, RMT, and LSTM hybrid RF models, respectively. PPCDM achieves near-perfect classification in all eight sound categories with almost no inter-class confusion, while other methods exhibit varying degrees of misclassification.

[0117] 3. Computational efficiency analysis:

[0118] As shown in Table 2, the PPCDM model of this invention achieves the highest recognition accuracy while having only 2.77M model parameters, and also boasts significant advantages in training and inference time. In contrast, the second-best performing RMT model has a whopping 78.39M parameters, 28.3 times that of PPCDM, and its inference time is 4 times longer. This is thanks to the innovation of this invention at the basic convolutional operation level, which fundamentally improves the feature extraction efficiency of the network without relying on a massive model structure. This result proves that this invention is not only superior in performance but also highly competitive in terms of computational resource consumption, fully meeting the needs of practical application scenarios for high accuracy, low cost, and easy deployment.

[0119] Table 2 Comparison of Model Complexity

[0120]

[0121] Wherein, "T-Time" and "I-Time" represent training time and inference time, respectively. "Params" represents the number of model parameters. Bold text indicates optimal performance.

[0122] 4. Interpretability Analysis:

[0123] Taking the MSAF dataset as an example, Figure 10 Figures (a)-(f) demonstrate the feature representation capabilities of PPCDM across four AFRBs, GAP, and GDDMCB. It can be seen that in the AFRB 1 stage, acoustic features exhibit a disordered and chaotic state in the feature space, lacking intra-class clustering and inter-class separability. As acoustic features are learned and represented block-by-block from AFRB 1 to AFRB 4, feature points in the feature space begin to show a trend of intra-class clustering and possess a certain degree of inter-class separability. GAP compresses features into more compact feature representations, at which point the features exhibit relatively obvious inter-class separability. Finally, GDDMCB dynamically maps the feature representations based on the differences in feature representations abstractly representing the health status of different broiler chickens, reconstructing the abstract representation of the mapping relationship between high-dimensional feature vectors. Therefore, at this point, feature points in the feature space have clear intra-class clustering and inter-class separability, and feature points between different categories have clear decision boundaries, forming independent feature clusters. This demonstrates that GDDMCB clearly represents the mapping relationship from the feature space to the category space, enabling the model to accurately identify the type of sound signal representing the health status of broilers. Combined with the ablation experiment mentioned earlier, this effectively improves the model's recognition performance.

[0124] Similarly, using the MSAF dataset as an example, we randomly selected some feature vectors from each stage of PPCDM for visualization, and conducted interpretability analysis of the model's feature extraction and decision-making mechanisms from a micro perspective. Figure 11 As shown. Starting from the top left corner of the feature map, the horizontal direction represents the feature dimension, and the vertical direction represents the feature channel dimension. Figure 11 The experimental results in (a) show that acoustic features with the same or similar representational meaning are encoded as texture trends with the same or similar colors, while features with different or significantly different representational meanings are encoded as edge trends with significant color changes. Both trends appear simultaneously in the feature dimension and the channel dimension, initially exhibiting horizontal and vertical stripe textures (hereinafter referred to as horizontal texture and vertical texture), and possessing a certain degree of edge characteristics. This indicates that PPCDM, with the help of PACConv, effectively extracts the edge and texture information of prior acoustic features in the AFRB 1 stage, providing an information basis for subsequent feature representation. Compared to AFRB 1, the output feature map of AFRB 2 ( Figure 11 (b) exhibits more semantically detailed characteristics, representing the prior information in the prior acoustic features in a more abstract and detailed way. AFRB 3 and AFRB 4 represent the features in a more abstract way, re-representing the broiler health information reflected by the prior acoustic features, and the output feature map is more advanced, abstract and detailed.

[0125] GDDMCB dynamically adjusts the mapping relationship based on the differences between different channels of high-dimensional acoustic features, transforming the prior sound features extracted in the feature extraction stage into a global feature representation with discriminative capabilities, and mapping it to the category space for complete sound signal category decision, outputting a feature map as shown below. Figure 11 As shown in (e), the arrangement of different colored squares in the figure represents the model's classification of the sample into a certain type of sound signal. To more clearly demonstrate GDDMCB's classification of sound signals, the authors used Softmax to probabilistically process the feature map, such as... Figure 11 As shown in (f), the mapping relationship between the input sound signal sample and its corresponding type is now clearer, with the brown square indicating the category of the sample. The above experimental results fully validate the structural rationality of PPCDM and further verify the effectiveness of core components such as PACConv and GDDMCB.

[0126] This embodiment describes a specific deployment scheme for applying the framework of the present invention to large-scale commercial farms for population health monitoring, namely the Abnormal Acoustic Event Frequency Monitoring System (AAEFMS).

[0127] System Deployment: An independent monitoring unit is deployed in each sub-area (SR, e.g., a unit housing approximately 30 broilers) of the farm. Each unit consists of an acoustic sensor, an embedded computing device equipped with a pre-trained PPCDM model, a counting module, and a wireless transmission module.

[0128] Workflow:

[0129] (1) Continuous monitoring: The sensor continuously captures sound signals within the area.

[0130] (2) Real-time classification: The local PPCDM model efficiently and accurately classifies the captured sound segments, especially identifying key sound events related to respiratory diseases such as cough and rales.

[0131] (3) Frequency statistics: The counting module counts the frequency of the disease sound type detected within a unit observation time (e.g., every 10 minutes).

[0132] (4) Warning trigger: When the frequency of the diseased sound in a certain area exceeds the preset normal fluctuation threshold, the system will automatically trigger a warning.

[0133] (5) Information reporting: The early warning information, along with the location information and frequency data of the area, is sent to the central management terminal of the farm via a wireless module to remind veterinarians and breeders to conduct timely professional inspections and interventions.

[0134] (6) To verify the effectiveness of this application scheme, an experiment was conducted in a simulated environment, and the cough frequency counted by this scheme was compared with the actual frequency manually labeled. The results are shown in Table 3.

[0135] Table 3. Accuracy of Regional Broiler Health Monitoring Program Based on FBPPCDM

[0136]

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Specific Implementation Method Two:

[0139] This embodiment is a computer storage medium that stores at least one instruction, which is loaded and executed by a processor to implement the state monitoring method based on prior perception calibration and dynamic mapping network.

[0140] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0142] This embodiment is a state monitoring device based on prior perception calibration and dynamic mapping network. The device includes a processor and a memory. It should be understood that it includes any device including a processor and a memory described in this invention. The device may also include other units and modules that perform display, interaction, processing, control and other functions through signals or instructions.

[0143] The memory stores at least one instruction, which is loaded and executed by the processor to implement the state monitoring method based on prior perception calibration and dynamic mapping network.

[0144] Those skilled in the art will understand that at least one stored instruction is a computer program product corresponding to a method or system. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0145] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0149] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0150] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A state monitoring method based on prior perception calibration and dynamic mapping network, characterized in that, The acoustic features extracted from the original audio signal are fed into the PPCDM (Programmed PCD) sound signal recognition network for recognition. The recognition includes: The input acoustic features are sequentially learned layer by layer through multiple AFRB units. Each AFRB unit includes multiple AFRL layers and one GMP layer. The AFRL layer is used for feature extraction. The beginning and end of the AFRL are PWConv modules, and the middle part includes at least one PACConv module and a multilayer perceptron (MLP). The processing of the PACConv module includes: S101, based on The features input to PACConv It is divided into two sparse feature subsets, namely and Remove all zero-value feature channels from the two feature subsets, and retain the non-zero feature channels to obtain... and ; It is a set of binary vectors, and , correspond The channel dimension, Any element in the set satisfies QM is a constant between 0 and 1. express The number of times the code is 1; S102. Perform two-stream information representation: The two-stream information representation adopts a two-stream architecture, in which the main branch uses convolutional layers. Feature extraction is performed to obtain acoustic representation. The identity flow branch will Transmitted directly without any transformation, as a feature Then splice them together. and Obtain features ; S103, targeting The weight vector is obtained based on pointwise convolution. And thus obtain ; After processing by multiple AFRB units, a GAP layer is used for spatial dimensionality reduction. Finally, a dynamic mapping classification block based on the Gaussian derivative function completes the mapping from the feature space to the class space, achieving the final classification decision.

2. The state monitoring method based on prior perception calibration and dynamic mapping network according to claim 1, characterized in that, The mapping learning function used in the dynamic mapping classification block of the Gaussian derivative function to map input features is a second-order derivative function based on the Gaussian function.

3. The state monitoring method based on prior perception calibration and dynamic mapping network according to claim 2, characterized in that, The processing steps for dynamic mapping classification blocks based on Gaussian derivative functions include: Let the input feature vector be denoted as , This represents the channel dimension of the feature vector; for A two-stage successive mapping is performed, namely GDDMCB1 and GDDMCB2, to complete the category decision; The overall calculation expression for GDDMCB1 is: , for Features on the i-th channel The mapping learning function for feature vector channel i is represented by a second-order derivative function based on a Gaussian function; and BN These represent the ReLU activation function and the batch normalization layer, respectively. This indicates the output of GDDMCB1. Indicates the channel dimension; During the second mapping, GDDMCB2 will use the output of GDDMCB1. Learning functions through mapping The layer maps the intermediate feature semantic space to the final category output space, and uses the Softmax function to obtain the probability values ​​of the decision results of PPCDM. , This indicates the number of categories of sound signals.

4. The state monitoring method based on prior perception calibration and dynamic mapping network according to claim 3, characterized in that, Channel Dimension .

5. The state monitoring method based on prior perception calibration and dynamic mapping network according to claim 3, characterized in that, The mapping learning function The processing procedure is as follows: Using the following form of mapping learning function Perform mapping: in, Indicates standard deviation; and These are two learnable parameters, determined through the training process of PPCDM.

6. The state monitoring method based on prior perception calibration and dynamic mapping network according to claim 3, characterized in that, In S102, the main branch uses convolutional layers with a kernel size of 3. Feature extraction is performed.

7. The state monitoring method based on prior perception calibration and dynamic mapping network according to claim 3, characterized in that, against The weight vector is obtained based on pointwise convolution. The process includes: First, capture via GAP. The system first generates a global channel descriptor; then, it establishes complex nonlinear dependencies between channels through two pointwise convolutions and a ReLU activation function, and finally transforms these dependencies into weight vectors using Softmax. .

8. A state monitoring method based on prior perception calibration and dynamic mapping network according to any one of claims 1 to 7, characterized in that, The sound signal recognition network PPCDM includes four AFRB units, and one PACConv module is set in the middle part of each AFRL layer in the AFRB unit.

9. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement the state monitoring method based on prior perception calibration and dynamic mapping network as described in any one of claims 1 to 8.

10. A state monitoring device based on prior perception calibration and dynamic mapping network, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the state monitoring method based on prior perception calibration and dynamic mapping network as described in any one of claims 1 to 8.

Images