Infant cry detection method and system

Abstract

The invention discloses an infant cry detection method and system, and belongs to the technical field of deep machine learning, and the method comprises the steps: extracting MFCC depth features of different types of infant cry audio signals, and constructing training set data; inputting a part of training set data into the OS-ELM for initialization learning, and initializing a hidden layer output matrix and an output weight matrix of the OS-ELM to obtain the OS-ELM in an initialization stage; and inputting the other part of training set data into the OS-ELM in the initialization stage in batches for online learning, updating the OS-ELM by updating a hidden layer output matrix and an output weight matrix in real time batch by batch, training a hybrid model, and detecting different types of baby cry audio signals acquired in real time. According to the method, dynamic updating is carried out through an OS-ELM incremental learning ability support model when new data arrives, the training time of a hybrid model can be greatly shortened while the requirements for hardware storage resources and computing power can be greatly reduced, and the high-precision recognition ability for various baby crying sounds is kept all the time.

Description

Technical Field

[0001] This invention relates to the field of deep machine learning technology, and more specifically to a method and system for detecting infant cries. Background Technology

[0002] Speech is the most primitive form of human communication, and the cry of an infant is a core signal for expressing needs and physiological state, so accurate recognition is of great significance.

[0003] Existing deep learning models for infant cry recognition still have many shortcomings in practical applications and technology, which restricts their efficient popularization in this field.

[0004] In terms of resource consumption, most single deep learning models, such as traditional convolutional neural networks or recurrent neural networks, often fall into the dilemma of "high accuracy depends on high computing power." Traditional deep learning models, such as AlexNet, have a large number of parameters, reaching millions or more, requiring high-performance GPU support. Their runtime power consumption can reach several watts. The models rely on repeated backpropagation iterations, resulting in long training times. For example, processing a single sample can take more than 100ms, affecting real-time performance. Furthermore, continuous high-load operation wastes energy. To improve the ability to capture the characteristics of an infant's cry, the model needs to be designed with deeper network layers or more parameters, resulting in large storage consumption and long computation time. This makes it difficult to adapt to resource-constrained scenarios such as smart home devices and portable monitoring terminals, significantly reducing real-time performance and deployment flexibility.

[0005] In terms of feature extraction, while lightweight architectures such as a single ShuffleNetV2 network can control resource consumption, their high time cost during network training makes them unsuitable for real-time detection of infant cries. Due to the environmental noise and individual differences mixed in with infant cries, using only the ShuffleNetV2 network for feature extraction exhibits weak adaptability to the differences in cry features among infants of different ages, leading to problems such as class aliasing and high misclassification rates during detection, making it difficult to meet the requirements for accurate recognition.

[0006] Existing recognition models are mostly trained and optimized on a single dataset or in a specific scenario. When faced with complex and varied crying samples in the real environment, such as crying variations accompanied by different states like coughing, hunger, and pain, the stability of feature learning is insufficient, resulting in weak generalization ability of the model. This leads to large fluctuations in accuracy in practical applications, making it difficult to form a stable and reliable recognition effect.

[0007] In summary, existing infant cry recognition models often face the dual dilemma of "high computational power dependence" and "high energy consumption," resulting in poor infant cry recognition performance. Summary of the Invention

[0008] To address the problems existing in the aforementioned fields, this invention proposes a method and system for detecting infant cries. A hybrid model is constructed, replacing the classification output layer of the ShuffleNetV2 network with OS-ELM. OS-ELM performs batch-wise online learning on the training data, avoiding the energy waste caused by full retraining of the ShuffleNetV2 network. This improves the diagnostic speed and accuracy of the ShuffleNetV2 network. Combining the ShuffleNetV2 network with OS-ELM through a lightweight design significantly reduces the demand for hardware storage resources and computing power while achieving real-time monitoring and accurate classification of infant cries.

[0009] To address the aforementioned technical problems, this invention discloses a method for detecting an infant's cry, comprising the following steps: Acquire audio signals of different categories of infant cries; A hybrid model based on ShuffleNetV2 network and OS-ELM is constructed; the hybrid model is to replace the classification output layer of ShuffleNetV2 network with OS-ELM for detecting baby cries; The feature extraction part of the ShuffleNetV2 network extracts MFCC deep features of different categories of infant cry audio signals and constructs training set data. A portion of the training set data is input into OS-ELM for initialization learning, which initializes the hidden layer output matrix and output weight matrix of OS-ELM to obtain the OS-ELM in the initialization stage. Another portion of the training set data is input into the OS-ELM in the initialization stage in batches for online learning. The hidden layer output matrix and output weight matrix are updated in real time through batch online learning, thereby updating OS-ELM and training the hybrid model. The real-time acquired audio signals of different categories of baby cries are input into the trained hybrid model to detect baby cries and output the baby cry classification results.

[0010] Preferably, the step of extracting MFCC deep features of different categories of infant cry audio signals through the feature extraction part of the ShuffleNetV2 network and constructing training set data specifically includes: Acquire different categories of baby cry signals; Different categories of infant cry signals are divided into short-time frames of infant cry time-domain signals and weighted using a Hamming window; The Hamming window-weighted time-domain signal of the baby's cry is converted to the frequency domain to obtain the spectrum of the baby's cry for each frame. A set of triangular filters based on the Mel scale distribution is applied to perform weighted summation on the spectrogram of infant cries, and the perceptual frequency band energy features of different categories of infant cries are extracted through the ShuffleNetV2 network. DCT transform is performed on the logarithmic infant cry perception frequency band energy vectors of different categories, and the first few dimensions of the Mel-frequency cepstral coefficient (MFCC) spectrum are retained. The Mel-frequency cepstral coefficient (MFCC) spectrogram is input into the feature extraction part of the ShuffleNetV2 network to extract the deep features of the MFCC spectrogram. Training set data is constructed by labeling the categories corresponding to the infant cries in the MFCC deep features.

[0011] Preferably, the step of extracting MFCC deep features of different categories of infant cry audio signals through the feature extraction part of the ShuffleNetV2 network further includes data preprocessing of the MFCC deep features. The data preprocessing includes audio clipping, dividing the audio clipped MFCC deep features into training set, validation set and test set according to proportions. The training set is used to label the categories corresponding to the infant cry of the audio clipped MFCC deep features.

[0012] Preferably, the output infant cry classification results include hunger, tiredness, laughter, no crying, and white noise.

[0013] Preferably, the different types of baby crying audio signals are obtained from resources provided by the Baidu PaddlePaddle platform and the Keggle website.

[0014] Preferably, it also includes an infant cry detection system, comprising: The infant cry data acquisition module is used to acquire different types of infant cry audio signals; A hybrid model building module for detecting infant cries is provided, which is used to construct a hybrid model based on ShuffleNetV2 network and OS-ELM. The hybrid model replaces the classification output layer of ShuffleNetV2 network with OS-ELM for detecting infant cries. The feature extraction part of ShuffleNetV2 network extracts MFCC deep features of different categories of infant cry audio signals and constructs training set data. A portion of the training set data is input into OS-ELM for initialization learning, which initializes the hidden layer output matrix and output weight matrix of OS-ELM to obtain the OS-ELM in the initialization stage. Another portion of the training set data is input into the OS-ELM in the initialization stage in batches for online learning. The hidden layer output matrix and output weight matrix are updated in real time through batch online learning, thereby updating OS-ELM and training the hybrid model. The detection module is used to input real-time acquired audio signals of different categories of baby cries into the trained hybrid model to detect baby cries and output the baby cry classification results.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The proposed infant cry detection method utilizes a hybrid model that replaces the classification output layer of the ShuffleNetV2 network with OS-ELM. OS-ELM's incremental learning capability allows for dynamic model updates as new data arrives, avoiding the energy waste caused by full retraining of the ShuffleNetV2 network. By enhancing the network's classification and generalization capabilities, it maintains high-precision recognition of various infant cries while significantly reducing hardware storage and computing power requirements. The feature extraction part of the ShuffleNetV2 network extracts MFCC deep features from different categories of infant cry audio signals and constructs a training set. The OS-ELM is then used in batches for initialization and online learning, continuously updating the hidden layer output matrix and output weight matrix learned through OS-ELM, thereby updating the OS-ELM. The OS-ELM online learning process continuously updates the model with the arrival of new data, rather than retraining the model. This method breaks the technical constraint that "high performance inevitably depends on high resources" by replacing the classification output layer of the ShuffleNetV2 network with OS-ELM. Through the efficient online learning mechanism of OS-ELM, this dual advantage of "low resource consumption + high classification performance" improves the diagnostic efficiency of the ShuffleNetV2 network, making it irreplaceable in practical applications. It can be adapted to resource-constrained scenarios such as home smart baby care devices and portable medical monitoring terminals, and realize real-time monitoring and accurate classification of baby cries. Attached Figure Description

[0016] Figure 1 This is a flowchart of the infant cry detection method proposed in this invention; Figure 2 The extraction process of Mel-frequency cepstral coefficients (MFCCs) provided in this embodiment of the invention; Figure 3 The ShuffleNetV2 network structure provided in this embodiment of the invention; Figure 4 The network structure of OS-ELM provided in this embodiment of the invention; Figure 5 This is a flowchart of the OS-ELM algorithm provided in an embodiment of the present invention; Figure 6 Accuracy curves of the ShuffleNetV2 network and traditional neural network models on the validation set provided in this embodiment of the invention; Figure 7 A dimensionality reduction visualization of the test set data output from the fully connected layers of AlexNet and ShuffleNetV2 networks trained for 30 and 50 rounds, provided in an embodiment of the present invention. Figure 8 The confusion matrix of the test set data of the ShuffleNetV2 network and hybrid model provided in the embodiments of the present invention; Figure 9 This is a dimensionality reduction visualization of the test set data output of the fully connected layer of the ShuffleNetV2 network and hybrid model provided in this embodiment of the invention. Detailed Implementation

[0017] The following will refer to the appendices in the embodiments of the present invention. Figures 1-9 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0018] Example This invention innovatively integrates the lightweight ShuffleNetV2 network with the online extreme learning machine OS-ELM to construct a highly efficient and energy-saving hybrid model for detecting infant cries. Its core advantage lies in the synergistic design of "low energy consumption and high performance".

[0019] like Figure 1 As shown, this invention provides a method for detecting an infant's cry, comprising the following steps: S1: Acquire audio signals of different categories of infant cries; S2: Construct a hybrid model based on ShuffleNetV2 network and OS-ELM; the hybrid model is to replace the classification output layer of ShuffleNetV2 network with OS-ELM for detecting baby cries; The feature extraction part of the ShuffleNetV2 network extracts MFCC deep features of different categories of infant cry audio signals and constructs training set data. A portion of the training set data is input into OS-ELM for initialization learning, which initializes the hidden layer output matrix and output weight matrix of OS-ELM to obtain the OS-ELM in the initialization stage. Another portion of the training set data is input into the OS-ELM in the initialization stage in batches for online learning. The hidden layer output matrix and output weight matrix are updated in real time through batch online learning, thereby updating OS-ELM and training the hybrid model. S3: Input the real-time acquired audio signals of different categories of baby cries into the trained hybrid model to detect baby cries and output the baby cry classification results.

[0020] Specifically, in step S1, the audio signals of different categories of infant cries were obtained from resources provided by the Baidu PaddlePaddle platform and the Keggle website. To ensure data consistency, all audio clips were uniformly edited into 8-second segments. The MFCC deep feature dataset after audio editing was divided into training, validation, and test sets to support model training and evaluation. The MFCC deep feature dataset after audio editing covers four emotional states and one environmental sound (i.e., hunger, tiredness, laughter, not crying, and white noise), totaling 772 .wav audio files. The audio files were randomly divided into training, validation, and test sets in a 6:2:2 ratio. The sample size statistics are shown in Table 1.

[0021] The training set labels the categories corresponding to the baby's cries based on the MFCC deep features after audio clipping (including hunger, tiredness, laughter, no crying, and white noise).

[0022] Table 1. Number of samples of infant cries for each category To capture the time-frequency information in the crying sound, the audio data is converted into Mel-frequency cepstral coefficients (MFCCs). MFCC spectrograms visually represent the frequency distribution and temporal variations of a sound signal, making them ideal input for deep learning models. The feature extraction part of the ShuffleNetV2 network will be used to extract deep features (local spatial features) from the MFCC spectrograms, reducing reliance on hand-designed features.

[0023] Mel-frequency cepstral coefficients (MFCCs) are a widely used characteristic parameter in audio processing to describe the spectral characteristics of audio signals, such as... Figure 2 As shown. The conversion relationship between Mel frequency m and linear frequency ƒ (in Hz) is as follows:

[0024] In the formula, the constants 2595 and 700 are parameters obtained by fitting experimental data and are used to calibrate the matching degree between the Mel scale and human auditory perception.

[0025] The definition of Mel frequency m is based on the characteristics of the human auditory system and can effectively simulate the human ear's perception of sounds at different frequencies.

[0026] From a computational perspective, the calculation of Mel-Cepstral coefficients involves multiple steps.

[0027] First, the audio signal undergoes preprocessing, including pre-emphasis, framing, and windowing. The spectral information is then mapped to a Mel frequency scale. Unlike the linear frequency scale, the Mel frequency scale better reflects the frequency perception characteristics of the human auditory system, offering higher resolution in the low-frequency range and lower resolution in the high-frequency range.

[0028] The power spectrum is processed through a set of M triangular Mel filters, each filter (the m-th filter) covering a specific frequency range, with the center frequencies distributed according to the Mel scale. The filter response is defined as:

[0029] Calculate the energy of the output of each filter, where f m These are the boundary and center frequencies of the m-th filter.

[0030] Taking the logarithm of these energy values ​​and then performing a discrete cosine transform yields the Mel-frequency cepstral coefficients (MFCCs).

[0031] Mel-frequency cepstral coefficients (MFCCs) have some notable characteristics. They have a strong ability to characterize the features of audio signals, can effectively extract the spectral features of audio signals, and have a certain degree of robustness, with good adaptability to environmental noise and speech variations.

[0032] Furthermore, the Mel-frequency cepstral coefficients (MFCCs) have a relatively low dimensionality, which reduces computational and storage requirements in subsequent processing and analysis.

[0033] Mel-spectral coefficients (MFCCs) have wide applications in audio feature extraction. In the field of speech recognition, MFCCs are one of the most commonly used features. They can accurately capture the features of speech signals, helping to identify different speech content.

[0034] Mel-frequency cepstral coefficients (MFCCs) can be used to extract features from audio, enabling tasks such as audio classification and similarity calculation. For example... Figure 2 The diagram shows the MFCC extraction process, which mainly includes the following five steps:

[0035] 1) Divide the original signal (audio signals of different categories of baby cries) into short time frames and use Hamming window weighting to reduce spectral leakage; 2) Convert the time-domain frame signal of the Hamming window-weighted baby crying audio signal to the frequency domain to obtain the power spectrum of each frame; 3) Apply a set of triangular filters (usually 26-40) based on the Mel scale distribution to perform weighted summation on the power spectrum of each frame and extract the sensing frequency band energy; 4) Based on the extracted sensing frequency band energy, take the logarithm of the output of each Mel filter to obtain the logarithmic sensing frequency band energy vector, which enhances low-amplitude energy information and suppresses high-energy spikes. 5) Perform DCT transformation on the logarithmic sensing frequency band energy vector to achieve dimensionality reduction and decorrelation processing. Finally, retain the first few dimensions of the Mel-Cepstral Coefficient (MFCC) spectrum (e.g., 13-40 dimensions) as input to the feature extraction part of the ShuffleNetV2 network to perform feature extraction and output the MFCC deep features of different categories of infant cry audio signals.

[0036] The ShuffleNetV2 network is used on devices with limited computing resources. Its feature extraction component includes grouped convolutions and depthwise separable convolutions. By using grouped and depthwise separable convolutions, the number of network parameters and computational cost can be significantly reduced, making it more suitable for deployment in embedded systems and mobile devices.

[0037] Among them, grouped convolution replaces and optimizes the original convolutional neural network by using multiple independent sub-convolutions, thereby reducing the number of parameters and computational cost. However, grouped convolution can only operate within its own group, which hinders the flow of information between groups and reduces the ability to express feature information.

[0038] To overcome the above shortcomings, the ShuffleNetV2 network employs channel shuffling after each group convolution, enabling the features between different groups to be fused together and solving the problem of information isolation between groups. For example... Figure 3 The diagram shows the ShuffleNetV2 network structure provided in this embodiment.

[0039] Figure 4 The diagram shows the network structure of OS-ELM. Since the online sequential extreme learning machine OS-ELM possesses the advantages of ELM in terms of speed and generalization ability, and is also a fast, incremental online learning algorithm, OS-ELM is used as the adaptive classifier in the hybrid model. Parameters are updated through initialization learning and online learning corresponding to ELM and OS-ELM. H k and β k Furthermore, by training the hybrid model, the model can be continuously updated as new data arrives, rather than being retrained.

[0040] like Figure 5 The diagram shown is the OS-ELM algorithm flowchart. OS-ELM consists of two phases: the initialization phase and the online learning phase. During the initialization phase, a small portion of the training set data is input into OS-ELM for initialization learning, resulting in the parameter hidden layer output matrix for this initialization learning phase. H 0 and output weight matrix β 0; and output the parameter hidden layer matrix. H0 and output weight matrix β 0 Stored in ELM, we obtain the OS-ELM during the initialization phase; During the online learning phase, parameters are updated in real time based on data fed in online. H k and β k ,get H k+1 and β k+1 This is to enhance the network's classification and generalization capabilities and improve classification accuracy.

[0041] This invention constructs a hybrid model for infant cry recognition based on the lightweight convolutional neural network model of ShuffleNetV2 and OS-ELM. For infant cry detection, features are extracted through the feature extraction part of the ShuffleNetV2 network to extract MFCC deep features of different categories of infant cry audio signals, and training set data is constructed. OS-ELM is then used to perform online updates and learning with a very small amount of training set data to improve the diagnostic speed and accuracy of the ShuffleNetV2 network.

[0042] The specific operating steps of the infant cry detection method proposed in this invention are as follows: Replace the softmax classification output layer of the ShuffleNetV2 network with OS-ELM, and use OS-ELM as the adaptive classifier for the hybrid model; Feature extraction is performed using the feature extraction part of the ShuffleNetV2 network, and the MFCC deep features of different categories of baby crying audio signals are used as input to the adaptive classifier.

[0043] The training set data is divided into two parts. OS-ELM is used to initialize and learn one part of the training set data, while the other part of the training set data is learned online in batches. The hybrid model is trained by updating OS-ELM.

[0044] The present invention also proposes an infant cry detection system, comprising: The infant cry data acquisition module is used to acquire different types of infant cry audio signals; A hybrid model building module for detecting infant cries is provided, which is used to construct a hybrid model based on ShuffleNetV2 network and OS-ELM. The hybrid model replaces the classification output layer of ShuffleNetV2 network with OS-ELM for detecting infant cries. The feature extraction part of ShuffleNetV2 network extracts MFCC deep features of different categories of infant cry audio signals and constructs training set data. A portion of the training set data is input into OS-ELM for initialization learning, which initializes the hidden layer output matrix and output weight matrix of OS-ELM to obtain the OS-ELM in the initialization stage. Another portion of the training set data is input into the OS-ELM in the initialization stage in batches for online learning. The hidden layer output matrix and output weight matrix are updated in real time through batch online learning, thereby updating OS-ELM and training the hybrid model. The detection module is used to input real-time acquired audio signals of different categories of baby cries into the trained hybrid model to detect baby cries and output the baby cry classification results.

[0045] This method utilizes a lightweight deep learning hybrid model built upon the ShuffleNetV2 network and OS-ELM, achieving low-energy operation for infant cry recognition through lightweight design. An MFCC optimization extraction strategy is employed to reduce computational energy consumption in the preprocessing stage. The online incremental learning mechanism of OS-ELM avoids the energy waste caused by full retraining of the ShuffleNetV2 network. A model deployment scheme adapted to low-power hardware (such as microcontrollers and embedded chips) further reduces overall device energy consumption.

[0046] Table 2 shows the analysis of the number of parameters and computational cost of the ShuffleNetV2 network and traditional neural network models.

[0047] Table 2. Number of parameters and computational cost of ShuffleNetV2 network and traditional neural network models In terms of the number of parameters, the ShuffleNetV2 network has a significant reduction compared to traditional neural network models (AlexNet, VGG-16, ResNet-50 and EfficientNet-B0 models), which means that the ShuffleNetV2 network occupies less space during storage and can effectively reduce the dependence on hardware storage resources.

[0048] In terms of computational requirements, the ShuffleNetV2 network is also at the lowest level, a characteristic that significantly reduces its demand for computing power during operation.

[0049] This dual advantage significantly reduces the hardware requirements for running programs. For some hardware devices with relatively limited performance, such as small-scale integrated microcontroller circuits, traditional neural network models often struggle to run smoothly or even fail to run due to excessive parameters and computational load. The ShuffleNetV2 network greatly expands the application scenarios of traditional neural network models. With reduced hardware requirements, circuit power consumption also decreases significantly. In many applications, power consumption is a critical consideration, especially in portable devices, embedded systems, or environments with strict energy consumption constraints. Lower power consumption not only extends device battery life but also reduces design pressure related to heat dissipation, lowering the overall system cost and complexity.

[0050] In summary, ShuffleNetV2, with its advantages in the number of parameters and computational cost, not only lowers the hardware threshold and enables its application in small-scale integrated microcontroller circuits, but also performs excellently in power consumption control, laying a solid foundation for the popularization and application of traditional neural network models in more fields.

[0051] like Figure 6 The figure shows the accuracy curves of the ShuffleNetV2 network on the validation set compared to traditional neural network models (AlexNet, VGG-16, ResNet-50, and EfficientNet-B0). By comparing the performance of each model, the ShuffleNetV2 network (in...) Figure 6 The ShuffleNetV2 network (replacing AlexNet with ShuffleNetV2) demonstrates significant overall advantages. As seen in the accuracy curves on the training and validation sets over 30 rounds, the ShuffleNetV2 network consistently outperforms the traditional AlexNet, exhibiting a more stable convergence trend in the later stages of training. Furthermore, throughout the entire 50-round training cycle, the ShuffleNetV2 network maintains a comprehensive lead over the traditional AlexNet in both training and validation set accuracy metrics.

[0052] It's worth noting that the ShuffleNetV2 network excels in training efficiency. Its accuracy rapidly climbs to peak levels after only 30 training iterations, with virtually no significant fluctuations in subsequent iterations. This means the model can converge in a shorter training period. In contrast, the traditional AlexNet requires more training iterations to approach a similar accuracy level, and its performance is significantly lower than the ShuffleNetV2 network with the same number of iterations.

[0053] This "dual advantage" fully demonstrates that the ShuffleNetV2 network can maintain high-precision recognition while significantly reducing the time cost of model training and effectively improving the engineering efficiency of algorithm deployment.

[0054] In summary, the ShuffleNetV2 network model demonstrates significant advantages in both training speed and recognition accuracy, and its advanced technical solution and feasibility for engineering applications have been fully verified, providing strong empirical support for model selection in related fields.

[0055] Table 3 shows the accuracy and time consumption of the ShuffleNetV2 network and the traditional AlexNet model on the test set after different training iterations.

[0056] Table 3. Accuracy and time consumption of each model on the test set As shown in Table 3, the data clearly presents the performance differences between AlexNet and ShuffleNetV2 networks on the 30 and 50 test sets. ShuffleNetV2 network shows a significant advantage. Whether in the early stage of rapid improvement or the later stage of stable convergence, its test set accuracy is consistently higher than that of AlexNet. Moreover, this advantage is maintained stably as the number of training rounds increases.

[0057] This result further confirms the advanced nature of the ShuffleNetV2 network in feature extraction and pattern recognition. Its unique network structure design (such as channel shuffling mechanism and lightweight convolution strategy) enables it to capture key information in the data more efficiently, thereby achieving better classification results.

[0058] In terms of time consumption, the total training time of the two models is basically the same, and the fluctuation range of training time per batch is also highly consistent. This phenomenon indicates that although the ShuffleNetV2 network has achieved a breakthrough in accuracy, it has not come at the cost of increased time cost. Combined with the previous training efficiency analysis, the "high accuracy, same time consumption" design of the ShuffleNetV2 network is more reasonable and provides a more valuable technical solution for model deployment in practical engineering applications.

[0059] like Figure 7 The image shows the t-SNE clustering diagrams obtained by visualizing the output data of the fully connected layers of AlexNet and ShuffleNetV2 networks after 30 and 50 training rounds, respectively, followed by dimensionality reduction. It can be seen that after different training rounds of the two models, the classification results for various types of infant cries are as follows: Figure 7As shown in the sub-figure, the clustering effect of the ShuffleNetV2 network is significantly better than that of AlexNet. Moreover, the clustering effect is improved with the increase of training epochs. There are fewer misclassifications of various types of data, the aliasing between different classes is significantly reduced, the distance between classes is larger, and the intra-class clustering becomes more compact with the change of model and the increase of training epochs. Among them, the ShuffleNetV2 network performs best with 50 training epochs, the aliasing between classes is basically eliminated, the distance between classes is the largest, and the intra-class clustering is more compact, which further illustrates the advantages of the ShuffleNetV2 network.

[0060] Table 4. Accuracy and time consumption of validation and test sets for each model. Table 4 shows the analysis of the recognition results of the hybrid model built based on ShuffleNetV2 network and OS-ELM.

[0061] When the hybrid model based on ShuffleNetV2 and OS-ELM is trained, the ShuffleNetV2 network alone achieves an accuracy of 75.2% on the validation set and 70.33% on the test set when there are 5 training epochs. This model has a certain learning ability and can correctly classify the data to a certain extent, but there is still considerable room for improvement in overall performance.

[0062] When the softmax classification output layer of the ShuffleNetV2 network was replaced with OS-ELM, the resulting hybrid model achieved a significant improvement in accuracy to 93.6% on the validation set and 92.8% on the test set. This remarkable improvement demonstrates that the hybrid model can effectively enhance the model's ability to extract and classify data features, allowing the model to exhibit superior performance with the same number of training epochs.

[0063] When the number of training epochs increased to 10, the performance of the standalone ShuffleNetV2 network improved, with validation set accuracy rising from 75.2% to 82.3% and test set accuracy increasing from 70.33% to 76.3%. This is because the increased number of training epochs allows the model more opportunities to learn the underlying patterns and features in the data, thereby continuously optimizing its parameters and improving classification accuracy. Similarly, the performance of the hybrid model also improved further, with validation set accuracy reaching 94.0% and test set accuracy reaching 93.1%. This further demonstrates the superiority of lightweight deep learning hybrid models.

[0064] The hybrid model achieved significantly higher accuracy than the ShuffleNetV2 network on both the validation and test sets, and the difference in accuracy between the two sets was relatively small, indicating that the model has good stability and does not exhibit serious overfitting issues.

[0065] Furthermore, in 5 and 10 training epochs, the hybrid model showed very similar accuracy on the validation and test sets. This result clearly demonstrates that replacing the classification output layer of the ShuffleNetV2 network with OS-ELM can significantly improve the accuracy and efficiency of model diagnosis.

[0066] Replacing the classification output layer of the ShuffleNetV2 network with OS-ELM can effectively compensate for the shortcomings of the ShuffleNetV2 network alone in feature processing and classification decision-making, making the hybrid model have stronger generalization ability and classification accuracy. This shows that the hybrid model proposed in this invention has great potential in enhancing the performance of deep learning models.

[0067] like Figure 8 The image shows the confusion matrix of the test set for the ShuffleNetV2 network and the hybrid model (ShuffleNetV2+OS-ELM). The training time of the hybrid model is significantly related to the number of training epochs, and the overall time cost is mainly determined by the number of training epochs. Its accuracy on both the validation and test sets is significantly improved compared to the ShuffleNetV2 network, resulting in a significant improvement in classification accuracy and a significant reduction in time cost. From a practical engineering application perspective, this characteristic of the hybrid model provides a more valuable solution for deployment. In the scenario of infant cry classification, hospital neonatal monitoring systems and home infant care devices have dual requirements for model real-time performance and accuracy. The hybrid model can train quickly and predict efficiently with limited hardware resources, meeting the need for accurate judgment while avoiding system latency or excessive resource consumption caused by high time consumption. The hybrid model integrates the strengths of different models, possessing both the high efficiency of ShuffleNetV2 network feature extraction and the advantage of OS-ELM in controlling time consumption. Through reasonable module integration and parameter tuning, synergistic performance improvement is achieved. This design approach not only has great potential in infant cry classification, but also provides a reference for optimization in similar classification tasks, promoting the development of algorithm models towards a more efficient and practical direction.

[0068] On the infant cry data diagnostic test set, through... Figure 8Analyzing the four confusion matrices clearly reveals the recognition performance of different models. The hybrid models with 5 and 10 training epochs respectively demonstrate particularly outstanding performance, maintaining extremely high accuracy in diagnosing and recognizing various types of infant cries. Both hybrid models achieved a perfect 100% recognition rate for the first type of sound, and even in categories with relatively higher recognition difficulty, the minimum accuracy remained consistently above 86%. This comprehensive and excellent recognition capability means they can accurately capture the information conveyed by infant cries in different contexts, providing a reliable basis for clinical diagnosis or childcare monitoring.

[0069] It is worth noting that further analysis of the model runtime data reveals that these two hybrid models maintain high recognition accuracy without excessively consuming computational resources and time. Compared to other models that may suffer from "high accuracy accompanied by high time consumption" or "fast operation but insufficient recognition accuracy," they successfully achieve a dual optimization of accuracy and time efficiency. This good balance between performance and cost gives them significant advantages in practical applications. Whether used in real-time monitoring systems or batch data processing, they can efficiently complete diagnostic tasks while ensuring the reliability of results, providing strong technical support for intelligent applications in related fields.

[0070] Figure 9 t-SNE dimensionality reduction visualization was performed on the fully connected layer output features of the ShuffleNetV2 network and the hybrid model. The hybrid models with training epochs of 5 and 10 showed a significant advantage in the distribution of infant cries in the low-dimensional space. Compared to the ShuffleNetV2 network model, the hybrid models exhibited significantly reduced class overlap and a significantly lower number of misclassified points. The features extracted by the hybrid model exhibited a distinct characteristic of "compact within classes and discrete between classes": data points of the same class clustered tightly in space, indicating a high degree of consistency in feature extraction for the same type of cry; while the clear boundaries and large spacing between different classes reflected the model's ability to accurately capture the class-specific features of infant cries. This structured clustering effect essentially reflects the superiority of the hybrid model in the feature learning process. When making decisions, the hybrid model can achieve classification based on more stable feature differences, echoing the previous confusion matrix analysis results and further validating the core advantages of the hybrid model in the infant cry classification task from the perspective of feature distribution.

[0071] In summary, the hybrid model (ShuffleNetV2+OS-ELM) proposed in this invention provides a highly innovative solution for the task of infant cry recognition and classification. The core of this hybrid model lies in successfully breaking the technical constraint that "high performance inevitably depends on high resources." Through the lightweight architecture design of the ShuffleNetV2 network and the efficient learning mechanism of OS-ELM, and by leveraging OS-ELM's incremental learning capability to support dynamic model updates as new data arrives, it significantly reduces the demand for hardware storage resources and computing power while greatly reducing the training time of the hybrid model, maintaining a consistently high accuracy in recognizing various types of infant cries. This dual advantage of "low resource consumption + high classification performance" makes it irreplaceable in practical applications. It is adaptable to resource-constrained scenarios such as home smart baby care devices and portable medical monitoring terminals, enabling real-time monitoring and accurate classification of infant cries; it also provides reliable technical support for areas such as infant health status assessment and early demand response, promoting the implementation of intelligent infant care.

[0072] From a technical perspective, this hybrid model verifies the feasibility and superiority of heterogeneous model fusion in lightweight tasks through the collaborative optimization of "feature extraction-classification decision". It provides a valuable reference paradigm for similar audio recognition or small sample classification problems. Its technical ideas and practical results have important theoretical significance and application value.

[0073] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0074] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method of detecting a baby's cry, characterized by, The method comprises the following steps: acquiring infant cry audio signals of different categories; constructing a hybrid model based on a ShuffleNetV2 network and an OS-ELM; the hybrid model is obtained by replacing the classification output layer of the ShuffleNetV2 network with the OS-ELM, and is used for detecting infant cries; extracting MFCC deep features of the infant cry audio signals of different categories through a feature extraction part of the ShuffleNetV2 network, and constructing training set data; inputting part of the training set data into the OS-ELM for initialization learning, initializing the hidden layer output matrix and the output weight matrix of the OS-ELM, and obtaining the OS-ELM in the initialization stage; inputting another part of the training set data into the OS-ELM in the initialization stage in batches for online learning, and updating the hidden layer output matrix and the output weight matrix in real time through batch-by-batch online learning to update the OS-ELM, and training the hybrid model; inputting real-time acquired infant cry audio signals of different categories into the trained hybrid model for infant cry detection, and outputting infant cry classification results.

2. The infant cry detection method according to claim 1, characterized in that, The method of extracting MFCC deep features of the infant cry audio signals of different categories through the feature extraction part of the ShuffleNetV2 network further comprises data preprocessing of the MFCC deep features, and the data preprocessing comprises audio clipping. The output infant cry classification results comprise hunger, tiredness, laughter, no crying and white noise. The infant cry audio signals of different categories are acquired from resources provided by a Baidu PaddlePaddle platform and a Keggle website. The method comprises the following steps: an infant cry data acquisition module, configured to acquire infant cry audio signals of different categories; ​ ​ ​ ​ 3. The infant cry detection method of claim 1, wherein ​ 4. The infant cry detection method of claim 1, wherein ​ 5. The infant cry detection method of claim 1, wherein, ​ 6. A baby cry detection system characterized in that, ​ ​ The mixed model construction module for detecting infant crying sound is used for constructing a mixed model based on a ShuffleNetV2 network and an OS-ELM; the mixed model is used for detecting infant crying sound by replacing the classification output layer of the ShuffleNetV2 network with the OS-ELM; the MFCC deep features of different categories of infant crying sound signals are extracted through the feature extraction part of the ShuffleNetV2 network, and training set data is constructed; a part of the training set data is input into the OS-ELM for initialization learning, the hidden layer output matrix and the output weight matrix of the OS-ELM are initialized, and the OS-ELM in the initialization stage is obtained; another part of the training set data is input into the OS-ELM in the initialization stage in batches for online learning, the hidden layer output matrix and the output weight matrix are updated in real time through batch-by-batch online learning, and the OS-ELM is updated, and the mixed model is trained; The detection module is used for inputting real-time acquired different categories of infant crying sound signals into the trained mixed model for infant crying sound detection, and outputting infant crying sound classification results.

Images