Lightweight household abnormal sound recognition method and system based on transfer learning and knowledge distillation

Abstract

The application discloses a kind of based on transfer learning and knowledge distillation light-weight family abnormal sound identification method and system.Firstly, the common background sound and abnormal audio sample under the family scene are filtered, and a transfer learning audio dataset is constructed;Secondly, the dataset is preprocessed, the spectrogram is calculated, and then the mel spectrogram is generated;Then, the classification model is pre-trained, and fine-tuning is carried out to obtain a high-performance audio classification model;Then, the light-weight network model is trained, and the multi-label prediction vector of the independent occurrence probability of each abnormal event is output by the classification head network;Finally, knowledge distillation training is carried out based on the teacher and student model, and the recognition result is output.The application is suitable for the family edge computing scene, and is independent of the cloud to prevent data leakage;Lightweight structure is used to realize low-delay real-time response;Transfer learning and other mechanisms are used to balance lightweight and high precision, which can accurately identify family abnormal sound and is suitable for baby monitoring and other fields.

Description

Technical Field

[0001] This invention relates to the field of abnormal sound detection, specifically to a lightweight method and system for identifying abnormal sounds in the home based on transfer learning and knowledge distillation. Background Technology

[0002] With the increasing global trend of population aging, the number of "empty nesters" is growing explosively. Due to declining physical function and the isolation of their living environment, these elderly people living alone are highly vulnerable to sudden health crises or accidental injuries. For example, severe coughing caused by respiratory illnesses, crying or thrashing for help due to accidental falls or physical pain—these unusual sounds are often important signals that the elderly need emergency assistance. Currently, monitoring methods for elderly people living alone mainly rely on video surveillance and wearable alarm devices.

[0003] However, these existing technologies have significant limitations in real-world home scenarios. For example, while monitoring data is intuitive, it easily infringes on the privacy of the elderly, often fails to cover private areas where accidents frequently occur, such as bedrooms and bathrooms, and has blind spots. Wearable alarms require the elderly to wear them for extended periods and constantly monitor their charging status. The elderly often forget to wear them, forget to charge them, or experience discomfort while wearing them, leading to monitoring failure. Furthermore, these devices require the elderly to be able to operate them actively. When an elderly person faints due to oxygen deprivation from a violent cough, or is unable to move due to pain after a fall, they often cannot trigger the alarm themselves, resulting in missed opportunities for timely rescue.

[0004] Furthermore, existing methods for identifying unusual sounds mostly employ a server-side centralized processing architecture, requiring the collected audio data to be uploaded to a server in real time for analysis using large-scale deep learning models. While such methods can achieve high recognition accuracy, their reliance on continuous or frequent uploads of raw audio data poses risks such as privacy breaches, high bandwidth costs, and significant response latency. Therefore, there is an urgent need for a method that can achieve high-precision identification of unusual sounds in the home on resource-constrained edge devices. Summary of the Invention

[0005] To address the problems existing in current technologies, this invention proposes a lightweight method and system for identifying abnormal household sounds based on transfer learning and knowledge distillation. This method enables online identification of abnormal household sounds on resource-constrained edge devices with high reliability.

[0006] In a first aspect, the present invention provides a lightweight method for identifying abnormal household sounds based on transfer learning and knowledge distillation, comprising the following steps:

[0007] Dataset Construction: Common background noise and abnormal audio samples in home scenarios were selected to construct a transfer learning audio dataset, which was then divided into training set, test set and validation set;

[0008] Data preprocessing: The transfer learning audio data is preprocessed sequentially by resampling, segmentation or padding, feature extraction and data augmentation, and spectrograms are calculated to generate Mel spectrograms;

[0009] Transfer learning model training: The pre-trained base model network is trained using the Audioset dataset, and then fine-tuned using the pre-processed transfer learning audio dataset to obtain a high-performance audio classification model.

[0010] Lightweight network model training: Construct a lightweight network model that includes a shallow feature extraction module and a deep semantic enhancement module. The input audio is processed by the shallow feature extraction module and then input to the deep semantic enhancement module. The features are optimized by reparameterized depthwise separable convolution and lightweight attention, and the multi-label prediction vector is output through the classification head network.

[0011] Knowledge distillation training: The lightweight network model is trained using knowledge distillation based on the high-performance audio classification model. During training, the multi-label classification loss between the output of the lightweight home abnormal sound recognition network and the real label, as well as the knowledge distillation loss between the output of the lightweight home abnormal sound recognition network and the high-performance audio classification model, are calculated simultaneously. The weight parameters of the lightweight network model are updated using the backpropagation algorithm. The recognition network is tested using the validation set and the test set, and the recognition results are output.

[0012] Secondly, the present invention provides a lightweight home noise identification system based on transfer learning and knowledge distillation, comprising:

[0013] The dataset construction module is used to filter common background noise and abnormal audio samples in home scenarios, build a transfer learning audio dataset, and divide it into training set, test set and validation set;

[0014] The data preprocessing module is used to perform resampling, segmentation or padding, feature extraction and data augmentation preprocessing on the transfer learning audio data in sequence, calculate the spectrogram, and then generate the Mel spectrogram.

[0015] The transfer learning model training module is used to complete the transfer learning training of the pre-trained base model network based on the Audioset dataset, and to fine-tune it using the pre-processed transfer learning audio dataset to obtain a high-performance audio classification model.

[0016] The lightweight network model training module is used to construct a lightweight network model that includes a shallow feature extraction module and a deep semantic enhancement module. The input audio is processed by the shallow feature extraction module and then input to the deep semantic enhancement module. The features are optimized by reparameterized depthwise separable convolution and lightweight attention, and the multi-label prediction vector is output through the classification head network.

[0017] The knowledge distillation training module is used to perform knowledge distillation training on the lightweight network model based on the high-performance audio classification model. During the training process, the multi-label classification loss between the output of the lightweight home abnormal sound recognition network and the real label, as well as the knowledge distillation loss between the output of the lightweight home abnormal sound recognition network and the high-performance audio classification model, are calculated simultaneously. The weight parameters of the lightweight network model are updated using the backpropagation algorithm. The recognition network is tested using the validation set and the test set, and the recognition result is output.

[0018] Thirdly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described lightweight home abnormal sound recognition method based on transfer learning and knowledge distillation.

[0019] The beneficial effects of this invention are:

[0020] Adapted to home edge computing scenarios: The entire abnormal sound recognition workflow of this invention is completed independently by the edge device, eliminating dependence on cloud computing servers, reducing costs, and eliminating the risk of leakage of private audio data in the home, which meets the application needs of home scenarios.

[0021] Meeting the real-time response needs of families: This invention adopts core lightweight structures such as deep separable convolution and lightweight attention mechanism, which greatly reduces the number of model parameters and computational complexity, improves the model inference speed, and realizes low-latency, real-time detection of sudden abnormal audio streams.

[0022] Ensuring accurate family response precision: This invention introduces transfer learning, knowledge distillation, and structural reparameterization mechanisms. Based on transfer learning, the amount of training information is expanded, and through knowledge distillation and reparameterization mechanisms, feature representation is enhanced, providing a powerful guarantee for the training effect of small edge models and achieving a perfect balance between lightweight and high precision.

[0023] Highly practical: This invention can accurately classify and distinguish various abnormal sounds commonly found in the home environment, and can be widely used in various smart home detection fields such as infant monitoring and care for elderly people living alone. Attached Figure Description

[0024] Figure 1 The overall flowchart of the lightweight home abnormal sound recognition method based on transfer learning and knowledge distillation provided by the present invention;

[0025] Figure 2 The overall flowchart of the audio preprocessing module provided by this invention;

[0026] Figure 3The overall flowchart of the teacher model training module based on transfer learning provided by this invention;

[0027] Figure 4 The overall framework diagram of the teacher-student knowledge distillation module provided by this invention;

[0028] Figure 5 The overall framework diagram of the lightweight network provided by this invention. Detailed Implementation

[0029] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0030] like Figure 1 As shown, this invention proposes a lightweight method for identifying abnormal household sounds based on transfer learning and knowledge distillation. It mainly includes: constructing a household scene audio dataset, preprocessing the audio data, training a teacher model based on transfer learning, and training a lightweight household abnormal sound recognition network based on knowledge distillation. Specifically, it is implemented through the following technical solutions:

[0031] S1: Construction of a Home Scene Audio Dataset

[0032] Furthermore, in this embodiment of the invention, the original audio is filtered based on common background noises and specific abnormal (i.e., attention-grabbing) audio information in a home setting, retaining audio samples that conform to the characteristics of a home setting, and constructing a transfer learning dataset. This dataset consists of dedicated samples collected on-site and general samples conforming to a home setting filtered from large-scale open-source audio datasets; wherein, the collection of on-site audio information is mainly completed by the digital microphone module of the edge device to ensure that the training data conforms to the characteristics of the actual deployment environment.

[0033] The preprocessed and multi-labeled home scene audio dataset is divided into a training set, a validation set, and a test set according to a preset ratio. In this embodiment of the invention, the preset ratio is 8:1:1. The validation set is used to evaluate the convergence of the model during training, and the test set is used to test the generalization performance of the model. The distribution of the category labels in the home scene audio dataset is shown in Table 1.

[0034] Table 1. Distribution of Attention Category Labels in the Home Scene Audio Dataset of this Invention

[0035]

[0036] Furthermore, the distribution of background category labels in the home scene audio dataset is shown in Table 2.

[0037] Table 2. Distribution of Background Category Labels in the Home Scene Audio Dataset of this Invention

[0038]

[0039] In this embodiment of the invention, the home scene audio dataset contains multiple acoustic event labels, rather than a single label dominating the event, to ensure that the model learns the co-occurrence relationship between different acoustic events.

[0040] S2: Audio Data Preprocessing

[0041] Furthermore, such as Figure 2 As shown, this embodiment of the invention preprocesses audio data in a transfer learning dataset, including resampling, segmentation or padding, feature extraction, and data augmentation.

[0042] Furthermore, in this embodiment of the invention, when extracting features from the original audio signal, frame-by-frame windowing, fast Fourier transform, and Mel frequency domain filtering are employed to ultimately generate a Mel spectrogram that meets the model input requirements. To meet the computing power requirements of edge devices, this embodiment of the invention tests and selects appropriate sampling rate, window size, step size, number of Mel frequency bands, minimum audio frequency, and maximum audio frequency, as shown in Table 3 below.

[0043] Table 3. Lightweight Network Audio Data Preprocessing Parameters of the Invention

[0044]

[0045] S21: The librosa library is used to resample the training audio, resampling it to a preset frequency to preserve the acoustic features of the original speech signal in the low-frequency band to the greatest extent, thus improving the accuracy of subsequent feature extraction. After resampling, the audio signal is segmented or padded: if the audio duration is longer than the preset duration, it is segmented; if the audio duration is shorter than the preset duration, it is padded to ensure uniform data length in the input model. Feature extraction employs Fast Fourier Transform and Mel-Filter Bank to improve real-time computing power and generate feature maps that conform to human auditory perception.

[0046] S22: Considering the limited support of deep learning framework audio processing libraries for edge embedded devices, the embodiments of the present invention are built based on the device's underlying instruction support, and are adapted to the inference environment of edge devices through Mel scale and frequency conversion, fast Fourier transform and Mel filter bank processing.

[0047] Furthermore, the formula for converting Mel scale to frequency is:

[0048]

[0049] in, For the Mel scale, For frequency.

[0050] Furthermore, the formula for the Fast Fourier Transform is:

[0051]

[0052] in, This is the index of the current sampling point. This refers to the window size.

[0053] Furthermore, the formula for constructing the Mel filter bank is:

[0054]

[0055] in, For filter index, The center frequency of the filter. For frequency indexing.

[0056] S23: The audio preprocessing employs the Mixup data augmentation method to increase the number of training samples. Specifically, in each iteration step of the high-performance audio classification model training and the lightweight model training, two different audio samples and their corresponding labels are randomly selected, mixed in a certain proportion, and a virtual training sample is generated. Subsequently, the weight parameters of the network module are updated using the virtual training sample.

[0057] The formula for generating virtual training samples is:

[0058]

[0059] in, For the input feature tensor, For input batch even index samples, To input an odd number of samples in a batch, The mixing coefficient, The total number of samples after Mixup data augmentation.

[0060] S3: Teacher Model Training Based on Transfer Learning

[0061] like Figure 3As shown in the embodiment of the present invention, the teacher model training module based on transfer learning is used to construct a teacher model for a knowledge distillation system. The pre-trained high-performance home audio classification model is based on the Cnn14_16k network under the PANNs (Large-Scale Pretrained Audio Neural Networks for Audio Pattern Recognition) framework. Based on the pre-trained Cnn14_16k model, fine-tuning is performed using a home scene audio dataset to generate a high-performance audio classification model for home scenes.

[0062] During training, the weight parameters of the pre-trained model remain frozen, and only the weight parameters of the newly added home scene target classifier are updated. The calculation process of the target classifier is as follows:

[0063]

[0064] in, As weight, This is the bias value. This is the input feature vector.

[0065] Furthermore, since multiple acoustic events may occur simultaneously in a home setting, this invention uses the Sigmoid activation function to independently calculate the activation probability of each category. The multi-label prediction vector calculation formula is as follows:

[0066]

[0067] in, For category indexing, Predict scores for the teacher model.

[0068] S4: Lightweight Home Abnormal Sound Recognition Network Training

[0069] like Figure 5 As shown, the lightweight model architecture of this embodiment of the invention consists of a shallow feature extraction module and a deep semantic enhancement module. The shallow feature extraction module uses reparameterization depthwise separable convolution blocks (Rep_DW block) as its main component, while the deep semantic enhancement module uses reparameterization depthwise separable convolution blocks and lightweight attention blocks (Efficient Block) as its main components.

[0070] Furthermore, the relevant parameters of the lightweight home abnormal sound recognition network are shown in Table 4.

[0071] Table 4. Audio Data Preprocessing Parameters for the Teacher Model of this Invention

[0072]

[0073] S41: The reparameterized depthwise separable convolutional block of this embodiment of the invention improves the expressive power of the model while maintaining the inference computation speed of depthwise separable convolution by constructing an asymmetric training and inference architecture.

[0074] Specifically, the reparameterized depthwise separable convolution consists of multiple parallel branches of depthwise separable convolutional layers and identity mapping branches during the training phase, and the audio sample feature maps are simultaneously passed through the parallel branches; the reparameterized depthwise separable convolution consists of a single-path structure during the inference phase, and the structures of the two are determined through mathematical equivalent transformations.

[0075] Furthermore, in this embodiment of the invention, the reparameterized depthwise separable convolutional structure includes multiple parallel branches of depthwise separable convolutional layers and an identity mapping branch, with all branches accepting input feature maps in parallel. The specific calculation process of the reparameterized depthwise separable convolution is as follows:

[0076]

[0077] in, For the first Depth-separable convolutional modules for each branch. For the first Convolution operations in each branch, This indicates the batch normalization operation. This indicates the training phase.

[0078] Furthermore, the depthwise separable convolution module can be divided into depthwise convolution (DW) and pointwise convolution (PW). Depthwise convolution performs independent convolution operations on each channel of the input. The specific calculation process is as follows:

[0079]

[0080] in, For the first One input channel, This is a convolution operation.

[0081] Furthermore, pointwise convolution receives the output of depthwise convolution and interacts with information from each channel. The specific calculation process is as follows:

[0082]

[0083] in, For the first One output channel, This is a convolution operation.

[0084] Furthermore, the output of the shallow feature extraction module is input into the deep semantic enhancement module for processing to obtain a semantically enhanced audio feature map.

[0085] S42: Input the output of the shallow feature extraction module into the heavily parameterized depthwise separable convolution module in the lightweight attention block for processing, and obtain the local feature information output.

[0086] S43: The local feature information output is input into the lightweight attention block. The lightweight attention block performs global context encoding on the local feature information output to generate global feature information output. The calculation process of the lightweight attention block is as follows:

[0087]

[0088] in, This refers to the Softmax function.

[0089] S44: The local feature information output and the global feature information output are concatenated and fused, and then merged with the initial input through the residual structure to obtain the final feature map.

[0090] S45: The final feature map is converted into a one-dimensional data vector through a fully connected layer, and then output to the Sigmoid layer to obtain the multi-label prediction vector.

[0091] S5: Conduct knowledge distillation training on a lightweight home noise detection network

[0092] S51: As Figure 4 As shown, the converged teacher model Cnn14_16K is loaded into the teacher model training module based on transfer learning, and the lightweight student model to be trained is initialized. During training, the weight parameters of the teacher model remain frozen, and only the parameters of the student model are updated.

[0093] Furthermore, the selection of the teacher model follows the principle of selecting the best, and specific performance evaluation and screening are performed on the candidate fine-tuned high-performance audio classification models.

[0094] Specifically, based on the transfer learning audio test set, a teacher model is obtained from multiple fine-tuned high-performance audio classification models. Among them, the classification teacher model is the fine-tuned model with the highest accuracy among the multiple fine-tuned high-performance audio classification models.

[0095] S52: Input each audio in the transfer learning audio training set into the teacher model with the highest accuracy, obtain the multi-class score of each audio in the teacher model, label the transfer learning audio training set according to the label category table, and obtain the multi-label prediction vector of each audio in the training set as the teacher guidance soft label.

[0096] Furthermore, based on the teacher model, labels are generated and filtered for unlabeled and single-labeled audio in the family scene audio training set to construct hard labels for guiding student model training. The label generation and filtering process mainly includes filtering the original multi-label data and generating hard labels for single-labeled or unlabeled data.

[0097] Specifically, for audio samples in the transfer learning audio training set that already carry multi-label classification information, a label validity verification and filtering operation is performed based on a predefined family category table. This includes: traversing each label originally carried by the audio sample and matching it one by one with a predefined family category table containing target family scene categories; if the original label exists in the family category table, it is determined to be a valid label and retained; if the original label does not exist in the family category table, it is determined to be a redundant label and filtered out, thereby ensuring that the label space in the training set remains consistent with the target classification space of the lightweight student model.

[0098] S53: Further, for audio samples in the transfer learning audio training set that only have single-label annotations or are collected on-site but have not yet been labeled, hard labels based on confidence thresholds are generated using the teacher model with the highest accuracy.

[0099] Specifically, audio samples are input into the teacher model with the highest accuracy for forward inference calculation, obtaining the predicted probability score for each category in the family category table output by the teacher model. Then, a preset confidence threshold is set, and the predicted probability score for each category output by the teacher model is numerically compared with the preset confidence threshold. If and only if the predicted probability score of a category is greater than or equal to the preset confidence threshold, that category is determined as a deterministic hard label for the audio sample; otherwise, if the predicted probability score is less than the preset confidence threshold, the label is filtered out. Combining the results of label validity verification and filtering operations with the hard label generation results based on the confidence threshold, the final label for each audio sample in the transfer learning audio training set is obtained.

[0100] Furthermore, targeting Figure 5 In the training process, this embodiment of the invention employs a multi-label classification loss function, and the specific calculation process is as follows:

[0101]

[0102]

[0103] in, The total number of categories, For the current index, For real labels, Output for the student model. This is the activation function.

[0104] Furthermore, the formula for calculating the knowledge distillation loss function is as follows:

[0105]

[0106]

[0107] in, For temperature coefficient, Input for the teacher model, Output for the student model. The total number of categories, For category indexing.

[0108] Furthermore, the formula for calculating the joint loss function is as follows:

[0109]

[0110] in, This is the preset mixing ratio coefficient.

[0111] Furthermore, to verify the advantages of the method proposed in this embodiment, the algorithm was analyzed and compared using the Audioset dataset and a home scene audio dataset, and the results are listed in Tables 5 and 6. The structure of this invention is shown in Table 7, and the comparison with the MobileNetV2 structure is shown in Table 8.

[0112] Table 5 Performance Comparison Based on Audioset

[0113]

[0114] Table 6 Performance Comparison Table Based on Home Scene Audio Dataset

[0115]

[0116] Table 7 Structural Design of the Invention

[0117]

[0118] The number of layers in the lightweight network model of this invention can be set according to the actual situation, and is not limited here.

[0119] Table 8 MobileNetV2 Structure Table

[0120]

[0121] In one exemplary embodiment, a lightweight home noise identification system based on transfer learning and knowledge distillation is provided, comprising:

[0122] The dataset construction module is used to filter common background noise and abnormal audio samples in home scenarios, build a transfer learning audio dataset, and divide it into training set, test set and validation set;

[0123] The data preprocessing module is used to perform resampling, segmentation or padding, feature extraction and data augmentation preprocessing on the transfer learning audio data in sequence, calculate the spectrogram, and then generate the Mel spectrogram.

[0124] The transfer learning model training module is used to complete the transfer learning training of the pre-trained base model network based on the Audioset dataset, and to fine-tune it using the pre-processed transfer learning audio dataset to obtain a high-performance audio classification model.

[0125] The lightweight network model training module is used to construct a lightweight network model that includes a shallow feature extraction module and a deep semantic enhancement module. The input audio is processed by the shallow feature extraction module and then input to the deep semantic enhancement module. The features are optimized by reparameterized depthwise separable convolution and lightweight attention, and the multi-label prediction vector is output through the classification head network.

[0126] The knowledge distillation training module is used to perform knowledge distillation training on the lightweight network model based on the high-performance audio classification model. During the training process, the multi-label classification loss between the output of the lightweight home abnormal sound recognition network and the real label, as well as the knowledge distillation loss between the output of the lightweight home abnormal sound recognition network and the high-performance audio classification model, are calculated simultaneously. The weight parameters of the lightweight network model are updated using the backpropagation algorithm. The recognition network is tested using the validation set and the test set, and the recognition result is output.

[0127] In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the lightweight home anomaly sound identification method based on transfer learning and knowledge distillation described above.

[0128] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A lightweight method for identifying anomalous household sounds based on transfer learning and knowledge distillation, characterized in that, Includes the following steps: Dataset Construction: Common background noise and abnormal audio samples in home scenarios were selected to construct a transfer learning audio dataset, which was then divided into training set, test set and validation set; Data preprocessing: The transfer learning audio data is preprocessed sequentially by resampling, segmentation or padding, feature extraction and data augmentation, and spectrograms are calculated to generate Mel spectrograms; Transfer learning model training: The basic model network is trained using the Audioset dataset, and then fine-tuned using a preprocessed transfer learning audio dataset to obtain a high-performance audio classification model. Lightweight network model training: Construct a lightweight network model that includes a shallow feature extraction module and a deep semantic enhancement module. The input audio is processed by the shallow feature extraction module and then input to the deep semantic enhancement module. The features are optimized by reparameterized depthwise separable convolution and lightweight attention, and the multi-label prediction vector is output through the classification head network. Knowledge distillation training: The lightweight network model is trained using knowledge distillation based on the high-performance audio classification model; During training, the multi-label classification loss between the output of the lightweight home noise recognition network and the real labels, as well as the knowledge distillation loss between the output of the lightweight home noise recognition network and the high-performance audio classification model, are calculated simultaneously. The weight parameters of the lightweight network model are updated using the backpropagation algorithm. The recognition network is tested using the validation set and the test set, and the recognition results are output.

2. The lightweight home noise identification method based on transfer learning and knowledge distillation according to claim 1, characterized in that, The transfer learning audio dataset includes: Specialized samples collected on-site were acquired using digital microphones on edge devices, closely matching the characteristics of the actual deployment environment. A general sample selected from a large-scale open-source audio dataset that fits the home environment.

3. A lightweight method for identifying abnormal household sounds based on transfer learning and knowledge distillation, as described in claim 1 or 2, characterized in that... The resampling process uses the librosa library to resample the training audio data to a preset frequency. The resampled audio data is then cut or padded as follows: if the audio duration is longer than a preset duration, then the audio is cut. If the audio duration is less than the preset duration, padding will be performed. The feature extraction process employs frame-by-frame windowing, Fourier transform, and Mel filtering to generate feature maps. The data augmentation uses the Mixup method to increase the number of training samples.

4. A lightweight method for identifying abnormal home sounds based on transfer learning and knowledge distillation according to claim 3, characterized in that, The base model network is the CNN14_16k network. During training, the pre-trained weight parameters are frozen, and only the weight parameters of the newly added home scene target classifier are updated.

5. A lightweight method for identifying abnormal household sounds based on transfer learning and knowledge distillation according to claim 1, characterized in that, The following steps are performed when training the lightweight network model to obtain multi-label prediction vectors: Shallow feature extraction: Initial feature maps are generated by reparameterizing depthwise separable convolutions; Local feature capture: The initial feature map is input into the heavily parameterized depthwise separable convolutional module in the lightweight attention block to extract local semantic features; Global context encoding: Global context encoding is performed on local semantic features within a lightweight attention block to generate global semantic features; Feature fusion enhancement: Local semantic features and global semantic features are concatenated and fused with the initial input through residual connections to obtain the final feature map; Multi-label prediction output: The final feature map is input into the classification head network to obtain the final multi-label prediction vector.

6. A lightweight method for identifying abnormal household sounds based on transfer learning and knowledge distillation according to claim 5, characterized in that, The reparameterized depthwise separable convolution employs a multi-branch parallel architecture during the training phase, containing multiple depthwise separable convolutional layers and identity mapping branches, with all branches sharing the same input feature map; and adopts a single-path structure during the inference phase.

7. A lightweight method for identifying anomalous household sounds based on transfer learning and knowledge distillation, as described in claim 5 or 6, characterized in that... The classification head network converts the final feature map into a one-dimensional data vector through a fully connected layer, which is then input into the Sigmoid layer to obtain a multi-label prediction vector.

8. A lightweight method for identifying abnormal household sounds based on transfer learning and knowledge distillation according to claim 1, characterized in that, When training the lightweight network model using knowledge distillation based on a high-performance audio classification model and a hybrid loss function, the following steps are performed: Obtain the teacher model that has converged in the transfer learning training, and initialize the student model at the same time; The teacher model generates multi-label prediction vectors as soft labels from the multi-class scores of the training set; then hard labels are generated for unlabeled and single-labeled audio, while redundant labels of the original multi-label samples are filtered according to the preset family category table, so that the label space of the training set matches the target classification space of the lightweight student model. For single-label or unlabeled samples in the training set, the teacher model is used to perform forward inference to obtain the predicted probability score of each target category. Then, the categories that meet the score threshold are set as deterministic hard labels through a pre-set reliability threshold. By combining the label verification and filtering results of multi-label samples, the final labels of all audio samples are obtained, and the multi-label classification loss function is used to guide the training of student models.

9. A lightweight home noise identification system based on transfer learning and knowledge distillation, characterized in that, include: The dataset construction module is used to filter common background noise and abnormal audio samples in home scenarios, build a transfer learning audio dataset, and divide it into training set, test set and validation set; The data preprocessing module is used to perform resampling, segmentation or padding, feature extraction and data augmentation preprocessing on the transfer learning audio data in sequence, calculate the spectrogram, and then generate the Mel spectrogram. The transfer learning model training module is used to complete the transfer learning training of the pre-trained base model network based on the Audioset dataset, and to fine-tune it using the pre-processed transfer learning audio dataset to obtain a high-performance audio classification model. The lightweight network model training module is used to construct a lightweight network model that includes a shallow feature extraction module and a deep semantic enhancement module. The input audio is processed by the shallow feature extraction module and then input to the deep semantic enhancement module. The features are optimized by reparameterized depthwise separable convolution and lightweight attention, and the multi-label prediction vector is output through the classification head network. The knowledge distillation training module is used to perform knowledge distillation training on the lightweight network model based on the high-performance audio classification model. During training, the multi-label classification loss between the output of the lightweight home noise recognition network and the real labels, as well as the knowledge distillation loss between the output of the lightweight home noise recognition network and the high-performance audio classification model, are calculated simultaneously. The weight parameters of the lightweight network model are updated using the backpropagation algorithm. The recognition network is tested using the validation set and the test set, and the recognition results are output.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images