High-frequency transformer noise feature extraction method and related device
By extracting high-frequency transformer noise features using the OpenL3 model and deep learning technology, the problem of low accuracy in traditional sound quality prediction is solved, achieving higher accuracy and stronger adaptability in sound quality prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-07
AI Technical Summary
Existing technologies have low accuracy in predicting the sound quality of high-frequency transformers. Traditional objective parameters of sound quality are difficult to capture the differentiated characteristics of noise in high-frequency transformers, resulting in limited model prediction accuracy and poor adaptability.
The OpenL3 model combined with the Mel filter is used to extract noise features. The bidirectional long short-term memory network is used to process the correlation between the information of the noise samples. The nonlinear expression of the features is enhanced by fully connected layers and activation functions. Semantic features are extracted by combining deep learning models and attention mechanisms. Subjective scores and traditional objective parameters of sound quality are fused for prediction.
It significantly improves the accuracy and reliability of the high-frequency transformer sound quality prediction model, enabling it to better adapt to differences in various operating conditions and samples, and enhancing the model's generalization ability and robustness.
Smart Images
Figure CN121808346A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of high-frequency transformer noise analysis and evaluation, and in particular to a high-frequency transformer noise feature extraction method and related device. BACKGROUND
[0002] Compared with traditional power frequency transformers, high-frequency transformers have the advantages of high operating frequency and high power density, but their noise problem is very prominent. High-frequency vibration threatens the reliable operation of high-frequency transformers, and their noise also has adverse effects on the environment and human health. In order to better evaluate the impact of high-frequency transformer noise, it is necessary to obtain more adequate noise signal features.
[0003] In the field of sound quality prediction technology, sound quality objective parameters (such as loudness, roughness, etc.) are the core input basis for building sound quality prediction models, and they have good adaptability to most conventional noise samples and can meet the basic prediction needs. However, when faced with high-frequency transformer noise with special acoustic characteristics, traditional sound quality objective parameters have significant limitations: high-frequency transformer noise has obvious differences in acoustic characteristics - single-frequency pure tone components are prominent, sharpness values are extremely high, and noise energy in non-peak frequency bands is extremely low, forming a unique sound energy distribution characteristic of "peak concentration and weak non-peak". Traditional sound quality objective parameters are obtained based on fixed calculation formulas, which cannot accurately capture the above-mentioned differentiated characteristics of high-frequency transformer noise, and cannot fully reproduce the complex nonlinear mapping relationship between "sound signal input and subjective listening output" in the human auditory perception process, ultimately leading to the problem of limited prediction accuracy and poor adaptability to high-frequency transformer noise under different operating conditions for high-frequency transformer sound quality prediction models based on traditional parameters.
[0004] From the current status of the prior art, the current sound quality prediction scheme still stays at the level of using traditional sound quality objective parameters as the only or core input features, without breaking through the technical framework of "single physical acoustic indicator driving". This type of scheme has two key problems: first, the model generalization ability is weak, sensitive to changes in sample distribution, and prone to overfitting due to changes in noise characteristics caused by operating conditions; second, in the scenario where the amount of noise samples is limited, a model constructed only relying on traditional objective parameters will have the problem of "insufficient effective feature learning" due to single feature dimension and insufficient information representation, which will further lead to poor stability and reliability of the model prediction results, making it difficult to meet the engineering application requirements.
[0005] In summary, the prior art cannot solve the problem of low precision in high-frequency transformer sound quality prediction, and therefore there is an urgent need for a technical solution that can extract high-frequency transformer noise features and fuse multi-dimensional information, in order to break through the limitations of traditional parameters and improve the performance of the prediction model. SUMMARY
[0006] This invention provides a method and related apparatus for extracting noise features from high-frequency transformers, which addresses the problem of low accuracy in predicting the sound quality of high-frequency transformers in existing technologies.
[0007] In view of this, the first aspect of the present invention provides a method for extracting noise features from a high-frequency transformer, the method comprising: The noise generated during the operation of the high-frequency transformer is collected and preprocessed to obtain the subjective score value of the preprocessed noise and the objective parameters of traditional sound quality. The preprocessed noise features are extracted using an OpenL3 model configured with a Mel filter to obtain matrix features; The matrix features are input into a bidirectional long short-term memory network for processing, and a vector is output. Using the subjective rating value as the true label, the fully connected layer is trained through the vector. During the training process, the output of the fully connected layer is nonlinearly transformed using an activation function to obtain several activation features, which are then spliced and fused with the traditional objective sound quality parameters. Semantic features are extracted from the spliced and fused features using a deep learning model, and an attention mechanism is used to emphasize key features of the extracted semantic features to obtain the final feature vector, which is used to predict the sound quality of high-frequency transformers.
[0008] Optionally, the conventional objective parameters of sound quality include: loudness, roughness, sharpness, fluctuation intensity, sound-to-noise ratio, and prominence ratio.
[0009] Optionally, the step of extracting features from the preprocessed noise using an OpenL3 model configured with a Mel filter to obtain matrix features includes: The preprocessed noise is input into the OpenL3 model and filtered by a configured Mel filter to obtain matrix features; The number of Mel filters is 32 to 64, and the frequency coverage range is 20Hz to 20kHz; the configuration parameters of the OpenL3 model include: noise embedding dimension of 512, frame length of 0.1 seconds to 0.5 seconds, and frame overlap rate of 50% to 75%; the output matrix features have a dimension of 18 rows × 512 columns.
[0010] Optionally, the step of inputting the matrix features into a bidirectional long short-term memory network for processing and outputting a vector includes: The matrix features are input into the bidirectional long short-term memory network in time series for processing, and a vector of 1 row × 50 columns is output. The bidirectional long short-term memory network includes: an input layer with the number of neurons configured to match the dimension of the matrix features; one to three hidden layers, each with 64 to 256 neurons and bidirectional connections; and an output layer.
[0011] Optionally, the subjective rating value is used as the true label, and the fully connected layer is trained using the vector. During training, an activation function is used to perform a non-linear transformation on the output of the fully connected layer to obtain several activation features, including: The vector is input into a network structure containing two fully connected layers; wherein, the first fully connected layer is used to receive the vector and perform non-linear mapping; the second fully connected layer is used to perform regression training on the two fully connected layers with the subjective rating value as the true label, and the output of the second fully connected layer is processed by an activation function to generate and output several activation features; wherein, the activation function is any one of the ReLU function, Leaky ReLU function or Sigmoid function.
[0012] Optionally, the step of extracting semantic features from the concatenated and fused features using a deep learning model, and then using an attention mechanism to emphasize key features of the extracted semantic features to obtain the final feature vector, includes: Deep semantic features are extracted from the concatenated and fused features using a deep learning model; the deep learning model is at least one of the following: convolutional neural network, recurrent neural network, long short-term memory network, or Transformer model. The features output by the deep learning model are processed through an attention mechanism to assign different weights to different feature dimensions, resulting in the final feature vector.
[0013] Optionally, the noise generated during the operation of the high-frequency transformer is collected by using a mono audio method.
[0014] A second aspect of the present invention provides a high-frequency transformer noise feature extraction system, the system comprising: The acquisition unit is used to acquire the noise generated during the operation of the high-frequency transformer and perform preprocessing to obtain the subjective score value of the preprocessed noise and the traditional objective parameters of sound quality. The first extraction unit is used to extract features from the preprocessed noise using an OpenL3 model configured with a Mel filter, and obtain matrix features. The transformation unit is used to input the matrix features into a bidirectional long short-term memory network for processing and output a vector. The transformation unit is used to train the fully connected layer using the subjective rating value as the real label and the vector. During the training process, the activation function is used to perform a nonlinear transformation on the output of the fully connected layer to obtain several activation features, which are then spliced and fused with the traditional objective sound quality parameters. The second extraction unit is used to extract semantic features from the spliced and fused features through a deep learning model, and to use an attention mechanism to emphasize key features of the extracted semantic features to obtain the final feature vector, which is used to predict the sound quality of high-frequency transformers.
[0015] A third aspect of the present invention provides a high-frequency transformer noise feature extraction device, the device comprising a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the steps of the high-frequency transformer noise feature extraction method as described in the first aspect above, according to the instructions in the program code.
[0016] A fourth aspect of the present invention provides a computer-readable storage medium for storing program code for executing the high-frequency transformer noise feature extraction method described in the first aspect above.
[0017] As can be seen from the above technical solutions, the present invention has the following advantages: This invention provides a method for extracting noise features from a high-frequency transformer. First, it collects the noise generated during the operation of the high-frequency transformer. Then, it extracts noise features by introducing OpenL3 and combining it with a Mel filter to better match the characteristics of human hearing, effectively capturing the key features of the high-frequency transformer noise and obtaining richer noise information compared to traditional parameters. Next, it processes the features extracted by OpenL3 using BiLSTM, fully considering the correlation between information before and after the noise sample, which better simulates the sequential information processing in human auditory perception, improving the effectiveness of the features. Finally, the activated features obtained through fully connected layers and activation functions not only retain semantic information closely related to subjective human perception but also enhance the nonlinear expressive power of the features, making them more suitable for sound quality prediction. The task involves integrating activation features with traditional objective parameters of sound quality to construct a composite input system. This system overcomes the limitations of traditional sound quality models that rely solely on physical acoustic indicators, and more comprehensively reflects the intrinsic relationship between subjective listening perception and sound signal characteristics. This significantly improves the prediction accuracy and reliability of the high-frequency transformer sound quality prediction model, providing strong support for the accurate assessment and optimization of high-frequency transformer sound quality. Finally, the semantic features extracted through the deep learning model can more comprehensively and deeply describe the complex characteristics of high-frequency transformer noise, including the unique sound energy distribution of "concentrated peaks and weak non-peaks" and single-frequency pure tone components. The sound quality prediction model constructed based on deep semantic features has stronger generalization ability, can better adapt to the differences in different operating conditions and samples, and improves the robustness and reliability of the model. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 A flowchart illustrating a high-frequency transformer noise feature extraction method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a high-frequency transformer noise feature extraction system provided in an embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0021] Please see Figure 1 The present invention provides a method for extracting noise features from a high-frequency transformer, comprising: Step 101: Collect the noise generated during the operation of the high-frequency transformer and preprocess it to obtain the subjective score value of the preprocessed noise and the traditional objective parameters of sound quality. In one embodiment, acquiring the noise generated during the operation of a high-frequency transformer includes: acquiring the noise generated during the operation of a high-frequency transformer using a mono audio method.
[0022] It should be noted that professional noise acquisition equipment (such as a mono unit) should be used, and microphones should be placed at appropriate locations around the high-frequency transformer to ensure accurate acquisition of the noise generated during transformer operation. Under each operating condition, the noise acquisition duration should be no less than 15 seconds, and the sampling frequency no less than 25600Hz, to ensure that the acquired noise fully reflects the characteristics of the high-frequency transformer noise and meets the noise quality requirements for subsequent processing. One hundred noise samples from typical operating conditions are selected for preprocessing. The preprocessing method involves extracting 5 seconds of interference-free and stable noise data, storing them to form a noise sample set, and providing input for subsequent feature extraction.
[0023] The subjective rating values are obtained as follows: At least 20 trained evaluators are recruited to score the noise sample set described in step 1. The evaluators should be statistically representative of the sample in terms of gender, age, and occupation to ensure the validity of the scores. After obtaining the test data, the scores of the evaluators are analyzed for validity, outliers are removed, and the final rating values are obtained, providing accurate label values for subsequent steps.
[0024] In one embodiment, conventional objective parameters of sound quality include: loudness, roughness, sharpness, fluctuation intensity, sound-to-noise ratio, and prominence ratio.
[0025] It should be noted that traditional objective parameters of sound quality include at least loudness, roughness, sharpness, and fluctuation intensity. For the single-frequency tone characteristics of high-frequency transformers, single-frequency noise evaluation parameters must be included, namely sound-to-noise ratio and prominence ratio.
[0026] Step 102: Extract features from the preprocessed noise using an OpenL3 model configured with a Mel filter to obtain matrix features; In one embodiment, step 102 specifically includes: The preprocessed noise is input into the OpenL3 model and filtered by a configured Mel filter to obtain matrix features; wherein the number of Mel filters is 32 to 64, and the frequency coverage range is 20Hz to 20kHz; the configuration parameters of the OpenL3 model include: noise embedding dimension of 512, frame length of 0.1 seconds to 0.5 seconds, and frame overlap rate of 50% to 75%; the output matrix features have a dimension of 18 rows × 512 columns.
[0027] It should be noted that in the specific implementation, the pre-configured OpenL3 library is used to input the noise sample set preprocessed in step 101 into the OpenL3 model. When calling the OpenL3 model, the Mel filter parameters are set, such as the number of filters being 40 and the frequency range being 20Hz-20kHz, to make the extracted features more closely match the characteristics of human hearing. After OpenL3 processing, each noise sample will obtain a 18×512 matrix feature representation. For example, for a specific high-frequency transformer noise sample, after processing with OpenL3 and the Mel filter, each element in the resulting feature matrix represents the characteristic value of the noise at a specific frequency and time scale. These values comprehensively reflect the acoustic characteristics of the noise.
[0028] Note: OpenL3: A deep learning model for learning general feature representations of audio and visual (video frames). Its core objective is to extract features with cross-task and cross-modal transfer capabilities through pre-training, making it suitable for various downstream audio or visual tasks (without retraining the core model).
[0029] Step 103: Input the matrix features into a bidirectional long short-term memory network for processing and output a vector; In one embodiment, step 103 specifically includes: The matrix features are input into the bidirectional long short-term memory network in time sequence for processing, and a vector of 1 row × 50 columns is output. The bidirectional long short-term memory network includes: an input layer with the number of neurons configured to match the dimension of the matrix features; 1 to 3 hidden layers with 64 to 256 neurons in each hidden layer, and bidirectional connections are used; and an output layer.
[0030] It should be noted that in the specific implementation, a bidirectional long short-term memory network, namely the BiLSTM network model, is first constructed. This model includes an input layer, hidden layers, and an output layer. The number of neurons in the input layer is determined based on the feature dimension extracted by OpenL3, i.e., 18×512. The hidden layer is set with an appropriate number of neurons, such as 128, and bidirectional connections are set to process forward and backward sequence information simultaneously. The 18×512 matrix features extracted by OpenL3 are expanded according to the time series and input into the BiLSTM network sequentially. After processing by the BiLSTM network, the features of each noise sample are converted into a fixed-dimensional vector output. For example, for the noise sample features processed by OpenL3 above, after passing through the BiLSTM network, each element in the output 1×50 vector incorporates the information of the noise sample at different time steps, better reflecting the overall features of the noise sample.
[0031] Specifically, the BiLSTM network model of this invention includes a forward LSTM and a backward LSTM. The forward LSTM traverses each row of the matrix in chronological order, learning the variation of noise features from the initial time to subsequent time (such as the rising / falling trend of noise energy, the order of appearance of single-frequency pure tones, etc.). The backward LSTM traverses each row of the matrix in reverse chronological order, learning the correlation pattern of noise features from the final time to the initial time (such as the correspondence between the noise end stage features and the previous stage). Finally, feature fusion is performed: the feature vectors output by the forward and backward sub-networks are concatenated or weighted and fused to generate a comprehensive feature vector containing bidirectional temporal information, and finally outputs a 1×50 dimension vector.
[0032] Note: BiLSTM stands for "Bidirectional Long Short-Term Memory Network," which is an improved variant of Recurrent Neural Network (RNN) specifically designed for processing sequential data (such as text, time series, audio signals, etc.).
[0033] Step 104: Using subjective ratings as the true labels, train the fully connected layer using vectors. During the training process, use activation functions to perform nonlinear transformations on the output of the fully connected layer to obtain several activation features, which are then spliced and fused with traditional objective sound quality parameters. In one embodiment, step 104 specifically includes: The vector is input into a network structure containing two fully connected layers; wherein, the first fully connected layer is used to receive the vector and perform nonlinear mapping; the second fully connected layer is used to perform regression training on the two fully connected layers with the subjective rating value as the real label, and the output of the second fully connected layer is processed by an activation function to generate and output several activation features, which are then concatenated and fused with traditional objective sound quality parameters; wherein, the activation function is any one of the ReLU function, Leaky ReLU function, or Sigmoid function.
[0034] It should be noted that the 1×50 vector output from the BiLSTM in step 103 is input into a fully connected layer, which includes a first fully connected layer and a second fully connected layer. The first fully connected layer has 50 neurons, mapping the 50-dimensional vector output from the upper layer to a new 50-dimensional feature space through a non-linear transformation, enhancing the non-linear expressive power of the features. The second fully connected layer has 10 neurons, consistent with the number of output activation features; here, the 50-dimensional features are compressed into 10 dimensions. In the fully connected layer, the weights are trained using the true label values (subjective rating values) of the training set for regression training. During training, a suitable activation function, such as the ReLU function, is selected. After processing by the fully connected layer and the activation function, each noise sample yields 10 activation features. For example, for a certain noise sample, after processing by the activation function in the fully connected layer, each element in the resulting activation feature vector undergoes a non-linear transformation, enhancing the expressive power of the features and better reflecting semantic information related to human subjective perception. Furthermore, the calculated 10 activation features are concatenated and fused with traditional objective sound quality parameters (i.e., loudness, roughness, sharpness, fluctuation intensity, sound-to-noise ratio, and prominence ratio obtained in step 101) to form a new feature vector, which serves as the input feature for the subjective sound quality prediction model. For example, for a noise sample, its 10 activation features are concatenated sequentially with the calculated parameters such as loudness, roughness, sharpness, and fluctuation intensity to obtain an input vector containing multiple feature information.
[0035] Step 105: Extract semantic features from the spliced and fused features using a deep learning model, and use an attention mechanism to emphasize key features of the extracted semantic features to obtain the final feature vector, which is used to predict the sound quality of high-frequency transformers.
[0036] In one embodiment, step 105 specifically includes: Deep semantic features are extracted from the concatenated and fused features using a deep learning model. The deep learning model is at least one of the following: convolutional neural network, recurrent neural network, long short-term memory network, or Transformer model. The features output by the deep learning model are then processed through an attention mechanism to assign different weights to different feature dimensions, resulting in the final feature vector.
[0037] It should be noted that the spliced and fused features are input into a deep learning model, which utilizes the automatic feature learning capability of the deep learning model to extract deep semantic features. These features not only contain traditional acoustic information, but also contain complex noise patterns, nonlinear relationships, and implicit information related to operating conditions. The features output by the deep learning model are further processed, including using attention mechanisms to emphasize key features, or combining the outputs of multiple deep learning models for feature fusion, to obtain a more comprehensive and discriminative feature representation.
[0038] It's important to note that when using different deep learning models to extract deeper semantic features from the concatenated and fused features, it's necessary to consider the input characteristics and feature transformation logic of each model. The following example is based on common knowledge of deep learning. We know that the concatenated and fused features are a combination of activation features (e.g., a 10-dimensional vector) and traditional sound quality parameters (e.g., a 6-dimensional vector), forming a fixed-length one-dimensional vector (e.g., 16-dimensional). This vector needs to be converted into an input format compatible with various models. For example, the input requirements for a Convolutional Neural Network (CNN) are: a two-dimensional tensor (e.g., [number of samples, feature length, number of channels]) or a one-dimensional sequence. The adaptation method is to convert the concatenated one-dimensional vector (e.g., length K) into a two-dimensional tensor: [1, K, 1] (1 sample, K feature points, 1 channel). Local features are extracted through a one-dimensional convolutional layer (e.g., using a 3×1 convolutional kernel), followed by pooling and fully connected layers to obtain deeper semantic features. Example: The concatenated feature is 16-dimensional, converted to [1, 16, 1], input to a one-dimensional CNN, and the output is semantic features. For other deep learning models, those skilled in the art can design different adaptation methods according to the actual situation, which will not be elaborated here.
[0039] This invention provides a method for extracting noise features from a high-frequency transformer. First, it collects the noise generated during the operation of the high-frequency transformer. Then, it extracts noise features by introducing OpenL3 and combining it with a Mel filter to better match the characteristics of human hearing, effectively capturing the key features of the high-frequency transformer noise and obtaining richer noise information compared to traditional parameters. Next, it processes the features extracted by OpenL3 using BiLSTM, fully considering the correlation between information before and after the noise sample, which better simulates the sequential information processing in human auditory perception, improving the effectiveness of the features. Finally, the activated features obtained through fully connected layers and activation functions not only retain semantic information closely related to subjective human perception but also enhance the nonlinear expressive power of the features, making them more suitable for sound quality prediction. The task involves integrating activation features with traditional objective parameters of sound quality to construct a composite input system. This system overcomes the limitations of traditional sound quality models that rely solely on physical acoustic indicators, and more comprehensively reflects the intrinsic relationship between subjective listening perception and sound signal characteristics. This significantly improves the prediction accuracy and reliability of the high-frequency transformer sound quality prediction model, providing strong support for the accurate assessment and optimization of high-frequency transformer sound quality. Finally, the semantic features extracted through the deep learning model can more comprehensively and deeply describe the complex characteristics of high-frequency transformer noise, including the unique sound energy distribution of "concentrated peaks and weak non-peaks" and single-frequency pure tone components. The sound quality prediction model constructed based on deep semantic features has stronger generalization ability, can better adapt to the differences in different operating conditions and samples, and improves the robustness and reliability of the model.
[0040] The above is a method for extracting noise features from a high-frequency transformer provided in an embodiment of the present invention. The following is a system for extracting noise features from a high-frequency transformer provided in an embodiment of the present invention.
[0041] Please see Figure 2 The present invention provides a high-frequency transformer noise feature extraction system, comprising: The acquisition unit 201 is used to acquire the noise generated during the operation of the high-frequency transformer and perform preprocessing to obtain the subjective score value of the preprocessed noise and the traditional objective parameters of sound quality. The first extraction unit 202 is used to extract features from the preprocessed noise using an OpenL3 model configured with a Mel filter to obtain matrix features; Transformation unit 203 is used to input matrix features into a bidirectional long short-term memory network for processing and output a vector. Transformation unit 204 is used to train the fully connected layer with subjective rating values as the real labels through vectors. During the training process, the output of the fully connected layer is nonlinearly transformed using activation functions to obtain several activation features, which are then spliced and fused with traditional objective sound quality parameters. The second extraction unit 205 is used to extract semantic features from the spliced and fused features through a deep learning model, and to use an attention mechanism to emphasize key features of the extracted semantic features to obtain the final feature vector, which is used to predict the sound quality of high-frequency transformers.
[0042] Furthermore, this embodiment of the invention also provides a high-frequency transformer noise feature extraction device, the device including a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the steps of the high-frequency transformer noise feature extraction method as described in the above method embodiments, according to the instructions in the program code.
[0043] Furthermore, this embodiment of the invention also provides a computer-readable storage medium for storing program code, which is used to execute the high-frequency transformer noise feature extraction method described in the above method embodiments.
[0044] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0045] In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units 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, indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.
[0046] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0047] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0048] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of 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.
[0049] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the present invention.
Claims
1. A method for extracting noise features from a high-frequency transformer, characterized in that, include: The noise generated during the operation of the high-frequency transformer is collected and preprocessed to obtain the subjective score value of the preprocessed noise and the objective parameters of traditional sound quality. The preprocessed noise features are extracted using an OpenL3 model configured with a Mel filter to obtain matrix features; The matrix features are input into a bidirectional long short-term memory network for processing, and a vector is output. Using the subjective rating value as the true label, the fully connected layer is trained through the vector. During the training process, the output of the fully connected layer is nonlinearly transformed using an activation function to obtain several activation features, which are then spliced and fused with the traditional objective sound quality parameters. Semantic features are extracted from the spliced and fused features using a deep learning model, and an attention mechanism is used to emphasize key features of the extracted semantic features to obtain the final feature vector, which is used to predict the sound quality of high-frequency transformers.
2. The high-frequency transformer noise feature extraction method according to claim 1, characterized in that, The traditional objective parameters of sound quality include: loudness, roughness, sharpness, fluctuation intensity, sound-to-noise ratio, and prominence ratio.
3. The high-frequency transformer noise feature extraction method according to claim 1, characterized in that, The method involves extracting features from the preprocessed noise using an OpenL3 model configured with a Mel filter to obtain matrix features, including: The preprocessed noise is input into the OpenL3 model and filtered by a configured Mel filter to obtain matrix features; The number of Mel filters is 32 to 64, and the frequency coverage range is 20Hz to 20kHz; the configuration parameters of the OpenL3 model include: noise embedding dimension of 512, frame length of 0.1 seconds to 0.5 seconds, and frame overlap rate of 50% to 75%; the output matrix features have a dimension of 18 rows × 512 columns.
4. The high-frequency transformer noise feature extraction method according to claim 1, characterized in that, The step of inputting the matrix features into a bidirectional long short-term memory network for processing and outputting a vector includes: The matrix features are input into the bidirectional long short-term memory network in time series for processing, and a vector of 1 row × 50 columns is output. The bidirectional long short-term memory network includes: an input layer with the number of neurons configured to match the dimension of the matrix features; one to three hidden layers, each with 64 to 256 neurons and bidirectional connections; and an output layer.
5. The high-frequency transformer noise feature extraction method according to claim 1, characterized in that, The subjective rating value is used as the true label, and the fully connected layer is trained using the vector. During training, an activation function is used to perform a non-linear transformation on the output of the fully connected layer to obtain several activation features, including: The vector is input into a network structure containing two fully connected layers; wherein, the first fully connected layer is used to receive the vector and perform non-linear mapping; the second fully connected layer is used to perform regression training on the two fully connected layers with the subjective rating value as the true label, and the output of the second fully connected layer is processed by an activation function to generate and output several activation features; wherein, the activation function is any one of the ReLU function, Leaky ReLU function or Sigmoid function.
6. The high-frequency transformer noise feature extraction method according to claim 1, characterized in that, The process involves extracting semantic features from the concatenated and fused features using a deep learning model, and then using an attention mechanism to emphasize key features of the extracted semantic features to obtain the final feature vector, including: Deep semantic features are extracted from the concatenated and fused features using a deep learning model; the deep learning model is at least one of the following: convolutional neural network, recurrent neural network, long short-term memory network, or Transformer model. The features output by the deep learning model are processed through an attention mechanism to assign different weights to different feature dimensions, resulting in the final feature vector.
7. The high-frequency transformer noise feature extraction method according to claim 1, characterized in that, The noise generated during the operation of the high-frequency transformer is collected, including the noise generated during the operation of the high-frequency transformer using a mono audio method.
8. A high-frequency transformer noise feature extraction system, characterized in that, include: The acquisition unit is used to acquire the noise generated during the operation of the high-frequency transformer and perform preprocessing to obtain the subjective score value of the preprocessed noise and the traditional objective parameters of sound quality. The first extraction unit is used to extract features from the preprocessed noise using an OpenL3 model configured with a Mel filter, and obtain matrix features. The transformation unit is used to input the matrix features into a bidirectional long short-term memory network for processing and output a vector. The transformation unit is used to train the fully connected layer using the subjective rating value as the real label and the vector. During the training process, the activation function is used to perform a nonlinear transformation on the output of the fully connected layer to obtain several activation features, which are then spliced and fused with the traditional objective sound quality parameters. The second extraction unit is used to extract semantic features from the spliced and fused features through a deep learning model, and to use an attention mechanism to emphasize key features of the extracted semantic features to obtain the final feature vector, which is used to predict the sound quality of high-frequency transformers.
9. A high-frequency transformer noise feature extraction device, characterized in that, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the high-frequency transformer noise feature extraction method according to any one of claims 1-7 according to the instructions in the program code.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for executing the high-frequency transformer noise feature extraction method according to any one of claims 1-7.