Sick chicken abnormal cry recognition method and system

Abstract

The invention relates to a sick chicken abnormal cry recognition method, which comprises the following steps: S1, collecting continuous audio signals in a chicken house, and cutting the continuous audio signals to obtain a plurality of non-overlapping audio clips with equal duration; S2, performing preprocessing based on the non-overlapping audio clips with equal duration, and extracting differences of a plurality of set orders of audio features in the audio clips, fusing the audio features and differences of a plurality of set orders of the audio features to obtain fused features; s3, inputting the fusion features into a pre-trained sick chicken abnormal cry recognition model, and outputting a classification result of each equal-duration audio clip; and S4, according to a change condition of a classification result, judging whether to trigger an alarm, and when a trigger condition is met, giving an alarm.

Description

A method and system for identifying abnormal cries of sick chickens Technical Field

[0001] This invention relates to the field of livestock and poultry health monitoring, specifically to a method and system for identifying abnormal cries of sick chickens. Background Technology

[0002] In the process of raising chickens, if abnormal sounds such as coughing or wheezing occur, it usually indicates the early clinical symptoms of respiratory disease.

[0003] Traditional techniques rely on the experience of farmers for identification. In the nighttime auscultation process, it is easily affected by subjective experience, resulting in low efficiency, high labor costs, and a high risk of missed detections and misjudgments. Furthermore, it cannot achieve real-time and automated monitoring.

[0004] With the advancement of intelligent technology, some modeling techniques have been used to identify abnormal calls from sick chickens. However, due to the complexity of the chicken coop environment, multiple chickens may call simultaneously, accompanied by complex environmental noises. Existing technologies rely on endpoint detection, requiring lengthy processing procedures, and the models exhibit poor generalization performance. Therefore, there is an urgent need for an automatic identification method and system for abnormal calls from sick chickens that is highly accurate, robust, and capable of end-to-end processing. Summary of the Invention

[0005] This invention provides a method and system for recognizing abnormal sounds from sick chickens, addressing the problems of existing technologies that rely on manual inspections for identifying abnormal sounds in chicken coops, resulting in poor real-time performance, high false negative rates, and low recognition accuracy. Traditional methods struggle to handle complex speech scenarios in chicken coops with multiple chickens resonating and noise interference, and the models do not adequately capture the time-frequency changes of abnormal sound segments, affecting the early identification of diseases.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, this application provides a method for identifying abnormal calls from sick chickens, comprising: S1, collecting continuous audio signals from a chicken coop and segmenting the continuous audio signals to obtain multiple non-overlapping audio segments of equal duration; S2, preprocessing the non-overlapping audio segments of equal duration and extracting several predetermined order differences of the audio features therein, and then fusing the audio features and their several predetermined order differences to obtain fused features; S3, inputting the fused features into a pre-trained abnormal call recognition model for sick chickens and outputting the classification results of each audio segment of equal duration; S4, determining whether an alarm is triggered based on the changes in the classification results, and triggering an alarm when the triggering conditions are met.

[0008] In one implementation, the preprocessing in S2 includes pre-emphasis and frame-by-frame windowing; the pre-emphasis method is implemented using a first-order differential filter; during frame-by-frame windowing, the window function types include: rectangular window, Hamming window, and Haining window.

[0009] In one implementation, in S2, the LogFbank features of non-overlapping audio segments and their first and second differences are extracted by short-time Fourier transform, and then fused along the event dimension to obtain the fused features of one channel.

[0010] In one implementation, the abnormal cries of sick chickens are identified using a ResNet18-TF network as the base network.

[0011] In one implementation, the ResNet18-TF model uses the ResNet18 network as the backbone network and adds a time-frequency attention mechanism to the backbone network.

[0012] In one implementation, the time-frequency attention mechanism includes parallel time attention branches and frequency attention branches, which enhance the feature map through adaptive weighting before output.

[0013] In one implementation, before inputting the one-channel fusion feature into the ResNet18-TF model to obtain the abnormal call category corresponding to each audio segment output by the abnormal chicken call recognition model, the method further includes: obtaining an initial abnormal chicken call recognition model; training the abnormal chicken call recognition model based on the abnormal chicken call samples and the manually labeled results of each audio segment corresponding to the abnormal chicken call samples; during the training process of the abnormal chicken call recognition model, determining the model index of the abnormal chicken call recognition model based on the corresponding validation set data; and saving the model weight with the highest value of the model index on the validation set to obtain the abnormal chicken call recognition model.

[0014] In one implementation, model metrics include precision, recall, and F1 score.

[0015] The formula for calculating precision is as follows:

[0016] The formula for calculating recall is as follows:

[0017] The formula for calculating the F1 score is as follows:

[0018] Precision represents the proportion of correctly predicted positive data out of the total number of correctly predicted positive data. Recall represents the proportion of correctly predicted positive data out of the total number of actually positive data. The F1 score is the harmonic mean. TP stands for True Positive, which is a sample that is predicted to be positive and is also positive; FP stands for False Positive, which is a sample that is predicted to be positive but is actually negative; FN stands for False Negative, which is a sample that is predicted to be negative but is actually positive.

[0019] In one implementation, during the training process of the abnormal chicken call recognition model, four data augmentation methods—time stretching, volume adjustment, time masking, and frequency masking—are applied to the fusion features. The enhanced fusion features are then input into the ResNet18-TF model to obtain the trained abnormal chicken call recognition model.

[0020] Secondly, a system for recognizing abnormal calls from sick chickens is provided, comprising: an audio acquisition module for acquiring continuous audio signals within the chicken coop and segmenting the continuous audio signals to obtain multiple non-overlapping audio segments of equal duration; a signal processing module for preprocessing the non-overlapping audio segments of equal duration and extracting several predetermined order differences in the audio features therein, and then fusing the audio features and their several predetermined order differences to obtain fused features; a depth recognition module for inputting the fused features into a pre-trained abnormal call recognition model for sick chickens and outputting the classification results of each audio segment of equal duration; and an early warning and positioning module for determining whether an alarm is triggered based on changes in the classification results, and triggering an alarm when the triggering conditions are met.

[0021] In one implementation, the depth recognition module runs on an edge computing device or an embedded hardware platform.

[0022] In one implementation, the audio acquisition module, the signal processing module, or the depth recognition module are deployed in the inspection robot.

[0023] In one implementation, the hardware structure of the inspection robot includes: a mobile chassis, a microphone array, an edge computing module, and a communication module.

[0024] Thirdly, a computer device is provided, including a processor and a memory, the memory storing a computer program, wherein when the processor executes the computer program, it implements the method of the first aspect.

[0025] The present invention has the following advantages: 1. No endpoint detection required: The present invention provides an automatic identification method for abnormal cries of sick chickens, which uses a fixed-length slicing method to replace the traditional endpoint detection, thus simplifying the process.

[0026] 2. Rich feature representation: The present invention provides an automatic identification method for abnormal cries of sick chickens, which integrates LogFbank and its differential features to improve the model's perception capability.

[0027] 3. High recognition accuracy: The present invention provides an automatic recognition method for abnormal cries of sick chickens, which is based on the ResNet18-TF model and greatly improves the recognition accuracy.

[0028] 4. Strong system integration: The automatic identification method for abnormal cries of sick chickens in this invention is adaptable to inspection robots or embedded terminals and has engineering potential. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 is a flowchart illustrating a method for recognizing abnormal cries of sick chickens provided by the present invention; Figure 2 is a structural diagram illustrating the time-frequency attention mechanism in the method for recognizing abnormal cries of sick chickens provided by the present invention.

[0031] Figure 3 is a structural schematic diagram of an abnormal call recognition system for sick chickens provided by the present invention; Figure 4 is a flowchart of Embodiment 1 of the present invention; Figure 5 is a flowchart of Embodiment 2 of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0033] To address the shortcomings and problems of existing technologies, this application provides a method for identifying abnormal calls from sick chickens, comprising: S1, collecting continuous audio signals from within a chicken coop and segmenting the continuous audio signals to obtain multiple non-overlapping audio segments of equal duration; S2, preprocessing the non-overlapping audio segments of equal duration and extracting several predetermined order differences in the audio features therein, and then fusing the audio features and their several predetermined order differences to obtain fused features; S3, inputting the fused features into a pre-trained abnormal call recognition model for sick chickens and outputting the classification results of each audio segment of equal duration; S4, determining whether an alarm is triggered based on changes in the classification results, and triggering an alarm when the triggering conditions are met.

[0034] Detailed Embodiments The solutions of this application will be specifically described below in one or more more detailed embodiments, in conjunction with Figures 1 to 5.

[0035] In Example 1, the main symptom of sick chickens in a chicken farm is abnormal calls, such as hoarse calls, breathing rales, coughing, snoring, or strange noises. In this example, abnormal calls are defined as chicken coughing and snoring.

[0036] Please refer to Figures 1, 2, 3, and 4. Figure 1 is a flowchart illustrating an automatic identification method for abnormal calls of sick chickens provided by the present invention. In this embodiment, the automatic identification method for abnormal calls of sick chickens specifically includes steps S110 to S140, each step as follows: S110: Continuous audio data is collected inside the chicken coop using an acoustic sensor and cut into non-overlapping audio segments of equal duration. The duration of the non-overlapping audio segments can be adjusted.

[0037] S120: Preprocessing and Feature Extraction. Preprocessing methods include pre-emphasis and frame-by-frame windowing. Pre-emphasis is implemented using a first-order differential filter. A rectangular window, Hamming window, or Haining window is selected as the window function for frame-by-frame windowing. STFT is used to extract the LogFbank features, the first-order difference of the LogFbank features, and the second-order difference of the LogFbank features. These features are then fused along the time dimension to obtain the fused features.

[0038] S130: Input the fused features into the automatic identification model for abnormal calls of sick chickens to obtain the corresponding abnormal call category.

[0039] S140: Determine whether to trigger a sick chicken warning based on the identification results.

[0040] In this embodiment, the automatic recognition model for abnormal chicken calls uses a large number of fusion features of coughing sounds, snoring sounds, and other sounds (including rooster crowing, equipment noise, and environmental noise) as training samples to accurately distinguish coughing and snoring sounds from other sounds. During the training process of the abnormal chicken call recognition model, four data augmentation methods—time stretching, volume adjustment, time masking, and frequency masking—are applied to the fusion features. The enhanced fusion features are then input into the ResNet18-TF model to obtain the trained abnormal chicken call recognition model.

[0041] The ResNet18-TF model is designed for lightweight architecture. It uses the ResNet18 network as the backbone and adds a time-frequency attention mechanism (TFBlock) to the backbone.

[0042] Specifically, Figure 2 illustrates the structural diagram of the time-frequency attention mechanism in the automatic identification method for abnormal cries of sick chickens provided by this invention. TFBlock includes parallel temporal attention branches and frequency attention branches, which enhance the feature map through adaptive weighting before outputting. The temporal attention branch is used to capture specific frames where sound events occur, while the spectral attention branch is used to pay attention to different frequency bands to varying degrees. The construction of parallel branches reduces the mutual interference between temporal and spectral attention, helping to improve the robustness of the model. Specifically, for the input feature map U, a convolutional layer "f" is used... c "Obtain the global time feature map V across channels" T and global spectral feature map V F Then, the global average pooling "f" is applied. GAP The sigmoid activation function "σ" is applied to V. T and V F To obtain time direction activation W T and frequency direction activation W F Multiplying the input feature map U by the time attention and spectral attention respectively yields the feature map U after applying time attention and spectral attention. T and U F Then, X is obtained by fusing these two branches through element-wise summation. Global average pooling is then used to generate channel features S from the global information of the feature map X, and finally, a simple fully connected layer f is applied. fc Create a compact feature z to achieve accurate and adaptive selection. Guided by feature z, for U T and U F Channel attention a and b are added separately to adaptively select different information space scales. A softmax operation is applied to each channel of a and b. Finally, by applying attention to the two branches U... T and UF The weighted feature map Y is obtained after applying the time-frequency attention mechanism.

[0043] Specifically, the ResNet18-TF model consists of four stages, each containing {2, 2, 2, 2} TF units. Each TF unit comprises a series of 3×3 convolutions and a TFBlock. Different architectures can be obtained by changing the number of TF units in each stage. ResNet18-TF adopts a deep residual network structure based on ResNet18 and introduces a time-frequency attention mechanism (TFBlock) to enhance the extraction of audio time-frequency features. The model mainly consists of four residual stages (TF units), each consisting of a convolutional layer (Conv), batch normalization (BatchNorm), a ReLU activation function, and a TFBlock. The number of TFBlocks in each stage is 2, 2, 2, 2. The model is structured in four stages: Stage 1 extracts low-level time-frequency features from the audio, such as edges and basic patterns; Stage 2 refines these features to capture more complex local information; Stage 3 further learns high-level time-frequency relationships to enhance class discriminative power; and Stage 4 extracts deep features to obtain a global representation. The model then compresses the feature maps using Global AvgPool to reduce computation and connects them to a fully connected layer (FCN) for final classification. Finally, Softmax outputs three classes (coughing, snoring, and other sounds). Compared to ResNet18, ResNet18-TF, through TFBlock, enables the network to more effectively focus on key time-frequency regions of abnormal sounds, improving the accuracy and robustness of abnormal chicken sound classification.

[0044] In this embodiment, before inputting a channel fusion feature into the ResNet18-TF model to obtain the abnormal call category corresponding to each audio segment output by the abnormal call recognition model for sick chickens, the method further includes: obtaining an initial abnormal call recognition model for sick chickens; training the abnormal call recognition model for sick chickens based on the sick chicken call samples and the audio recognition results of each audio segment corresponding to the sick chicken call samples; during the training process of the abnormal call recognition model for sick chickens, determining the precision, recall, and F1 score of the abnormal call recognition model for sick chickens based on the corresponding validation set data; and storing the model weight with the highest F1 score on the validation set to obtain the abnormal call recognition model for sick chickens.

[0045] The F1 score is the harmonic mean of precision and recall, taking into account both. It is a metric used in models for recognizing abnormal calls from sick chickens, reflecting the model's overall performance in identifying these calls. Precision measures the accuracy of the model in predicting positive samples, while recall measures the proportion of correctly detected positive samples out of all positive samples.

[0046] Understandably, for the three-class classification model in this embodiment, the precision, recall, and F1 score for each class are calculated separately, and then macro-averaging is used to evaluate the overall performance.

[0047] In this embodiment, the formula for calculating precision is:

[0048] The formula for calculating recall is:

[0049] The formula for calculating the F1 score is:

[0050] Wherein, TP represents a true positive, which is a sample that is predicted to be positive and is actually positive; FP represents a false positive, which is a sample that is predicted to be positive but is actually negative; and FN represents a false negative, which is a sample that is predicted to be negative but is actually positive.

[0051] Furthermore, the abnormal chicken call recognition model uses a large number of fusion features of sick chicken coughs, snoring sounds, and other sounds (including rooster crowing, equipment noise, and environmental noise) as training samples to enable it to identify sound types. The collected continuous audio signal is segmented into non-overlapping audio segments of equal duration. After preprocessing the non-overlapping audio segments, the LogFbank features and their first and second differences are extracted using STFT. Then, the LogFbank features and their first and second differences are fused to obtain fusion features. These fusion features are input into the trained abnormal chicken call recognition model, which outputs the classification results for each audio segment. Based on the recognition results, it is determined whether to trigger a sick chicken warning. If the classification results include chicken coughs or snoring sounds, a sick chicken warning is issued.

[0052] This embodiment provides an automatic identification method for abnormal cries of sick chickens. It uses equal-length windows to segment audio segments, eliminating the need for additional endpoint detection and feature engineering. The model can automatically learn the optimal feature representation, thereby improving detection robustness and adaptability. By adding a time-frequency attention mechanism to the ResNet18 network architecture, the model can more effectively focus on task-relevant key signal components and ignore irrelevant information, thus enhancing its ability to capture key features and maintaining high detection performance under different environmental and seasonal conditions.

[0053] The present invention also provides an automatic identification system for abnormal calls of sick chickens. Please refer to Figure 3. Figure 3 is a structural schematic diagram of an automatic identification system for abnormal calls of sick chickens provided by the present invention. In this embodiment, an automatic identification system for abnormal calls of sick chickens includes an audio acquisition module 310, an audio processing module 320, a feature fusion module 330, a depth recognition module 340, and an early warning positioning module 350.

[0054] The audio acquisition module 310 is used to acquire audio signals inside the chicken coop, including coughing, snoring and other sounds.

[0055] The audio processing module 320 is used for audio signal preprocessing and feature extraction. First, the acquired continuous audio signal is segmented into audio segments of equal duration, and these segments are pre-emphasized and windowed. Then, STFT is used to extract the LogFbank features of the audio signal, along with its first and second differences.

[0056] The feature fusion module 330 is used to fuse the LogFbank features of the audio signal and its first-order and second-order differences in the time dimension to obtain fused features.

[0057] The deep recognition module 340 is used to input the fused features into the abnormal chicken call recognition model to obtain the abnormal sound recognition result output by the abnormal chicken call recognition model.

[0058] The early warning and positioning module 350 is used to trigger a warning for sick chickens based on the recognition results obtained from the abnormal call recognition model of sick chickens, and to locate and alarm the abnormal position in conjunction with the corresponding device.

[0059] Figure 4 illustrates the flowchart of this embodiment. Audio signals are collected from multiple fixed points, edge computing devices are deployed to perform model calculations, and a sick chicken warning is triggered based on the recognition results, with an alarm being set up at the corresponding collection point.

[0060] In this embodiment, by deploying edge computing devices (such as Jetson Xavier, Raspberry Pi 4B, or domestic AI Box) and microphone arrays in multiple key areas inside the chicken house, an automatic identification system for abnormal calls of sick chickens with full coverage and real-time monitoring is achieved.

[0061] Optionally, depending on the chicken coop structure (such as stacked or floor-raised), one node can be deployed every 10-20 meters. Each node includes a dustproof and noise-reducing microphone, an embedded computing terminal, a power supply and network adapter, and an optional wireless communication module (such as Wi-Fi or LoRa). The devices are networked via a local area network, and data can be synchronously uploaded to a local control center or cloud platform.

[0062] Optionally, the system can acquire the audio stream in real time at a sampling rate of 44.1kHz and segment it in 3-second increments. All audio data is cached locally and processed using a sliding mechanism to ensure continuity and seamless recognition.

[0063] Optionally, the system can pre-emphasize the audio signal to boost high-frequency signals; perform frame-by-frame windowing processing with a window length of 25ms and a frame shift of 10ms, using a Hamming window; calculate LogFbank features and their first and second-order differences, and fuse them in the time dimension to obtain fused features. The fused features are input into the ResNet18-TF model, outputting a 3-class classification result: "coughing sound," "snoring sound," and "other sounds." If the "coughing sound" or "snoring sound" label appears consecutively, the system will: record an alarm log locally; set a flashing control indicator or a buzzer at the corresponding microphone node; and synchronously upload the data to the cloud platform to provide an alarm notification for the management end.

[0064] This embodiment can run offline, ensuring continuous monitoring even in network outage scenarios; it allows for parallel processing between nodes, resulting in fast system response and flexible deployment; and it can be expanded as needed to adapt to chicken house environments of different sizes.

[0065] Example 2 differs from Example 1 in that it uses an automated inspection robot as a carrier, integrating audio acquisition and edge computing modules to achieve dynamic patrol-style abnormal sound acquisition and recognition.

[0066] Please refer to Figures 1, 2, 3, and 5. Figure 5 shows a flowchart of this embodiment. Audio signals are collected by the movement of the inspection robot, dynamic model calculations are performed on the inspection robot, and warnings for sick chickens and abnormal location are triggered based on the recognition results.

[0067] The robot's hardware structure in this embodiment includes a mobile chassis, a microphone array, an edge computing module, and a communication module. The inspection robot supports obstacle avoidance and path planning; it is equipped with a Jetson Nano or Xavier module for local inference; and the communication module uses 5G or WiFi.

[0068] In this embodiment, the automatic recognition model for abnormal chicken calls uses a large number of coughs, wheezing sounds, and other sounds (including rooster crowing, equipment noise, and environmental noise) collected by inspection robots as training samples, enabling it to accurately distinguish coughs and wheezing sounds from other sounds. During the training process of the abnormal chicken call recognition model, four data augmentation methods—time stretching, volume adjustment, time masking, and frequency masking—are applied to the fusion features. The enhanced fusion features are then input into the ResNet18-TF model to obtain the trained abnormal chicken call recognition model.

[0069] In this embodiment, before inputting a channel fusion feature into the ResNet18-TF model to obtain the abnormal call category corresponding to each audio segment output by the abnormal call recognition model for sick chickens, the method further includes: obtaining an initial abnormal call recognition model for sick chickens; training the abnormal call recognition model for sick chickens based on the sick chicken call samples and the audio recognition results of each audio segment corresponding to the sick chicken call samples; during the training process of the abnormal call recognition model for sick chickens, determining the precision, recall, and F1 score of the abnormal call recognition model for sick chickens based on the corresponding validation set data; and storing the model weight with the highest F1 score on the validation set to obtain the abnormal call recognition model for sick chickens.

[0070] The F1 score is the harmonic mean of precision and recall, taking into account both. It is a metric used in models for recognizing abnormal calls from sick chickens, reflecting the model's overall performance in identifying these calls. Precision measures the accuracy of the model in predicting positive samples, while recall measures the proportion of correctly detected positive samples out of all positive samples.

[0071] Understandably, for the three-class classification model in this embodiment, the precision, recall, and F1 score for each class are calculated separately, and then macro-averaging is used to evaluate the overall performance.

[0072] In this embodiment, the formula for calculating precision is:

[0073] The formula for calculating recall is:

[0074] The formula for calculating the F1 score is:

[0075] Wherein, TP represents a true positive, which is a sample that is predicted to be positive and is actually positive; FP represents a false positive, which is a sample that is predicted to be positive but is actually negative; and FN represents a false negative, which is a sample that is predicted to be negative but is actually positive.

[0076] Optionally, the inspection robot will conduct a full-area inspection 1 to 3 times a day at set times; it will move slowly along the set route at a speed of 0.1 m / s and collect audio clips every 3 seconds at a sampling rate of 44.1 kHz.

[0077] Optionally, the system pre-emphasizes the audio signal to boost the high-frequency signal; performs frame-by-frame windowing processing with a window length of 25ms and a frame shift of 10ms, using a Hamming window.

[0078] Specifically, STFT is used to extract the LogFbank features, the first difference of the LogFbank features, and the second difference of the LogFbank features of the audio segment. The LogFbank features, the first difference of the LogFbank features, and the second difference of the LogFbank features of the audio segment are then fused in the time dimension to obtain the fused features.

[0079] Optionally, the Jetson Nano or Xavier module can be used to call the pre-deployed automatic identification method for abnormal sounds of sick chickens, input the fused features into the ResNet18-TF model, and output a 3-classification result: "coughing sound", "snoring sound", and "other sounds".

[0080] Furthermore, the output of the abnormal chicken noise recognition model is bound to the location information of the inspection robot, forming a "sound source-time-location information" triple. If an abnormal sound is detected, the robot system will: automatically report to the management platform and generate an event report; highlight the abnormal area on its visual interface; call the cloud camera to assist in recognition; and simultaneously upload to the cloud platform to provide management alarm prompts.

[0081] Compared to Example 1, Example 2 is more suitable for medium to large-scale smart farms with existing automated equipment. The mobile platform reduces the number of nodes required and lowers wiring and maintenance costs. Its disadvantage is that it cannot perform continuous, uninterrupted monitoring; it is suitable for supplementary monitoring or periodic inspections.

[0082] For implementation steps not detailed in this embodiment two, please refer to embodiment one.

[0083] In the several embodiments provided by this invention, it should be understood that the disclosed methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0084] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] The above are merely preferred embodiments of the present invention and are not intended to limit 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.

Claims

1. A method for identifying abnormal cries of sick chickens, characterized in that, include: S1. Collect continuous audio signals from inside the chicken coop and segment the continuous audio signals to obtain multiple non-overlapping audio segments of equal duration. S2. Preprocess the non-overlapping audio segments of equal duration and extract the differences of several predetermined orders of audio features. Then, fuse the audio features and the differences of several predetermined orders to obtain fused features. S3. Input the fused features into a pre-trained abnormal chicken call recognition model and output the classification results of each audio segment of equal duration. S4. Determine whether to trigger an alarm based on the changes in the classification results, and trigger an alarm if the triggering conditions are met.

2. The method for identifying abnormal cries of sick chickens according to claim 1, characterized in that, The preprocessing in S2 includes pre-emphasis and frame windowing; the pre-emphasis method is implemented using a first-order differential filter. When adding windows in frames, the types of window functions include: rectangular window, Hamming window, and Heining window.

3. The method for identifying abnormal cries of sick chickens according to claim 1, characterized in that, In S2, the LogFbank features of non-overlapping audio segments and their first and second differences are extracted by short-time Fourier transform, and then fused in the event dimension to obtain the fused features of one channel.

4. The method for identifying abnormal cries of sick chickens according to claim 3, characterized in that, The abnormal cries of sick chickens recognition model uses the ResNet18-TF network as its base network.

5. The method for identifying abnormal cries of sick chickens according to claim 4, characterized in that, Before inputting the fused features of one channel into the ResNet18-TF model to obtain the abnormal call category corresponding to each audio segment output by the abnormal call recognition model of sick chickens, the method further includes: obtaining an initial abnormal call recognition model of sick chickens; training the abnormal call recognition model of sick chickens based on the sick chicken call samples and the artificial labeling results of each audio segment corresponding to the sick chicken call samples; during the training process of the abnormal call recognition model of sick chickens, determining the model index of the abnormal call recognition model of sick chickens based on the corresponding validation set data; and saving the model weight with the highest value of the model index on the validation set to obtain the abnormal call recognition model of sick chickens.

6. A system for recognizing abnormal cries of sick chickens, characterized in that, include: The audio acquisition module is used to acquire continuous audio signals in the chicken coop and cut the continuous audio signals into multiple non-overlapping audio segments of equal duration. The signal processing module is used to preprocess the non-overlapping audio segments of equal duration, extract the differences of several set orders of audio features, and then fuse the audio features and the differences of several set orders to obtain fused features. The deep recognition module is used to input the fused features into a pre-trained abnormal chicken call recognition model and output the classification results of each audio segment of equal duration. The early warning and positioning module is used to determine whether an alarm should be triggered based on changes in the classification results, and to issue an alarm when the triggering conditions are met.

7. The abnormal cries of sick chickens recognition system according to claim 6, characterized in that, The depth recognition module runs on an edge computing device or an embedded hardware platform.

8. The abnormal cries of sick chickens recognition system according to claim 6, characterized in that, The audio acquisition module, the signal processing module, or the depth recognition module are deployed in the inspection robot.

9. The abnormal cries of sick chickens recognition system according to claim 8, characterized in that, The hardware structure of the inspection robot includes: a mobile chassis, a microphone array, an edge computing module, and a communication module.

10. A computer device, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method of claims 1 to 5.

Images