Noise artifact dynamic identification method based on multi-scale feature modeling and deep learning
By combining multi-scale feature modeling with deep learning, this method solves the problem that existing noise monitoring systems cannot distinguish between noise sources and artifact signals in complex industrial scenarios, thereby improving the accuracy and reliability of noise monitoring and making it suitable for complex industrial environments.
Patent Information
- Application Number
- CN202511799639.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-04-14
AI Technical Summary
Existing vibration and noise measurement systems and audio recognition systems cannot accurately distinguish noise sources from non-noise signals in complex industrial scenarios. Traditional noise measurement systems cannot characterize non-Gaussian distributed noise and are prone to artifact signals, resulting in insufficient accuracy and reliability of noise monitoring.
This paper adopts a method combining multi-scale feature modeling and deep learning. Through multi-scale time-frequency feature analysis and deep learning algorithms, it identifies and removes artifact signals, outputs feature indicators such as sound pressure level and kurtosis of real noise signals, establishes a dynamically updatable artifact signal pattern library, and uses a convolutional-self-attention model with CNN+Attention architecture for artifact signal recognition.
It significantly improves the accuracy and reliability of noise monitoring, can adapt to complex industrial environments, accurately identify and remove artifact signals, comprehensively characterize the characteristics of non-Gaussian noise, and provide high-quality noise monitoring data.
Smart Images
Figure CN121862153A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of noise detection technology, specifically relating to a dynamic identification method for noise artifacts based on multi-scale feature modeling and deep learning. Background Technology
[0002] In production practice and scientific research, noise monitoring is a crucial means of assessing environmental quality, equipment operating status, and protecting personnel health. However, currently widely used vibration and noise measurement systems, audio recognition systems, and noise measurement systems have numerous technical limitations, severely impacting the accuracy and reliability of noise monitoring.
[0003] Existing vibration and noise measurement systems generally employ the broadband energy integration method, which can obtain equivalent continuous sound pressure levels, but lacks signal separation technology. In complex industrial scenarios, it cannot distinguish between noise sources and non-noise signals, leading to significant deviations in measurement results. Traditional audio recognition systems mainly optimize human hearing characteristics, but feature extraction of broadband mechanical noise suffers from spectral folding distortion, making it difficult to identify and distinguish non-speech signals and unsuitable for noise meter testing environments. Current noise measurement systems rely on A, C, and Z weighted sound pressure level indicators, which cannot fully characterize the characteristics of non-Gaussian distributed noise such as impact noise and intermittent noise commonly found in complex industrial scenarios.
[0004] Furthermore, noise monitoring is susceptible to artifacts such as wind noise, receiver friction noise, and noise from personnel conversations, further degrading the signal-to-noise ratio of the collected data. Traditional spectral analysis methods have limited ability to analyze the time-frequency characteristics of transient artifacts, fixed-threshold filtering easily leads to loss of effective signal, and sound pressure level-based measurements are difficult to accurately reflect the characteristics of non-Gaussian noise. All of these problems are extremely detrimental to the analysis and identification of the specificity and harmfulness of noise. Therefore, there is an urgent need for a technical solution that can accurately identify and remove artifacts and comprehensively characterize noise features to overcome the bottlenecks of existing noise monitoring technologies. Summary of the Invention
[0005] The main objective of this invention is to provide a dynamic identification method for noise artifacts based on multi-scale feature modeling and deep learning. By combining multi-scale time-frequency feature analysis with deep learning algorithms, it aims to achieve accurate identification and removal of artifact signals during industrial noise acquisition, outputting comprehensive characteristic indicators such as real noise signal, identified artifact signals, and real noise sound pressure level and kurtosis, thereby significantly improving the accuracy and reliability of noise monitoring.
[0006] To achieve the above objectives, this invention provides a method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning, comprising the following steps: Step S1: Continuously collect complete noise signals from the surrounding environment and convert the audio input signal into a digital signal; Step S2: Preprocess the digital signal by dividing it into multi-scale short-time frames through a multi-step framing operation; Step S3: Identify whether short frames contain artifact noise and remove short frames containing artifact noise; Step S4: Maintain and update the artifact signal pattern library and deep learning recognition model; Step S5: Post-process the noise signal after artifact noise identification and calculate the kurtosis value of the noise segment by segment; Step S6: Calculate the continuous equivalent sound pressure level and other indicators for the noise signal after processing, and simultaneously perform a weighted average of the kurtosis calculated segment by segment to obtain the final noise detection result.
[0007] As a further preferred technical solution to the above technical solution, step S2 is specifically implemented as follows: Step S2.1: First, the digital signal is uniformly divided into equal-length frames with relatively large time intervals; Step S2.2: Perform secondary framing according to the percentile of the duration of all labeled samples in the artifact signal pattern library, and obtain multi-scale framing results of extremely short time frames, medium short time frames, and long short time frames in each equal-length frame segment; Step S2.3: During the framing process, the overlap between each window is a preset value of the window length, and a hybrid window function that is dynamically adjusted according to the kurtosis value of the framing signal is added to suppress spectral leakage.
[0008] As a further preferred technical solution to the above technical solution, for step S3, the input is all short-time frame data, and the output is all short-time frame data identified as not containing artifact noise. The specific implementation is as follows: Step S3.1: Use various audio feature statistical extraction methods to extract time-frequency features including zero-crossing rate, energy entropy, MFCC and Mel spectrum from short frames, thereby obtaining audio feature vectors of short frame data. The original data of a noisy signal will obtain three sets of short frame audio feature vectors with different frame lengths, denoted as long short frame audio feature vector, medium short frame audio feature vector and very short frame audio feature vector. Step S3.2: Based on the cosine similarity algorithm of feature vectors, recall a batch of labeled samples that are most similar to the target short-time frame from the pseudo-signal pattern library, and record the similarity score of the target short-time frame; the similarity scores of the audio feature vectors of the three groups of short-time frames with different frame lengths are recursively averaged upwards according to their temporal relationship, and finally the recursive weighted similarity scores of the long and short-time frames are obtained. Step S3.3: Simultaneously input the long and short time frame audio feature vectors from step S3.1 into the deep learning recognition model to identify whether they contain artifact noise, and output the probability score of judging that they contain artifact noise. Step S3.4: Calculate the weighted average of the similarity score and the probability score. Based on the weighted score, retain the short and long time frames whose weighted scores are less than the set threshold and proceed to the next processing step. Otherwise, determine that the frame contains artifact noise and remove it.
[0009] As a further preferred technical solution to the above technical solution, for step S4: Step S4.1: For the pseudo-signal pattern library, the pseudo-signal pattern library is used to store known pseudo-noise signal data. By using existing database indexing, search and update techniques, the data stored in the pseudo-signal pattern library can be searched and matched, and real-time addition, deletion and modification operations can be performed. Step S4.2: The deep learning-based artifact signal recognition model is a convolutional-self-attention model with a CNN+Attention architecture, which effectively captures long-range and short-range dependencies in audio features. The deep learning recognition model is pre-trained using known noise and artifact noise data to obtain the model parameter weights of the artifact signal recognition algorithm, and the model is updated through subsequent fine-tuning training. The model training process includes feature extraction, data augmentation, and optimization training.
[0010] As a further preferred technical solution to the above technical solution, for step S5, the short-time frame data identified as not containing artifact noise are merged and reassembled into complete information in time order through kurtosis calculation and frame merging process. For the uncovered parts, interpolation is performed according to the situation of the preceding and following frames. Furthermore, the equivalent sound pressure level and kurtosis information of the equal-length frames in step S2.1 are calculated. The calculation method of the equivalent sound pressure level is as follows: ; in, It is the instantaneous sound pressure in the equal-length framed data. It is a reference sound pressure level. The duration of frames of equal length, The duration of the combined noise signal. It is the equivalent sound pressure level of equal-length frames, used to calculate the overall equivalent sound pressure level of the noise signal; The method for calculating noise kurtosis is as follows: ; in, For the first frame of equal-length data The amplitude at each location The amplitude mean is given in the equally segmented frame data. The kurtosis of the overall noise signal is calculated using both arithmetic and geometric mean methods.
[0011] The beneficial effects of this invention are as follows: 1. This invention innovatively adopts a technical solution that combines multi-scale feature modeling with deep learning recognition model. Addressing the technical challenge of the high similarity of the spectral features of noise and artifact signals, it significantly improves the recognition accuracy of artifact signals through multi-scale framing, dynamic window function optimization, and CNN+Attention architecture model, effectively solving the problem that traditional methods have difficulty distinguishing artifacts from target noise. 2. A dynamically updatable artifact signal pattern library has been established. Combined with the fine-tuning mechanism of the deep learning model, it can adapt to the recognition needs of new artifact signals, has good scalability and adaptability, and is suitable for complex and ever-changing industrial noise environments. 3. By introducing higher-order statistical analysis, the kurtosis calculation is used to accurately characterize non-Gaussian noise (such as impulse noise and intermittent noise), which makes up for the limitations of traditional sound pressure level measurement and provides more comprehensive data support for the analysis and identification of noise characteristics and hazards. 4. The entire technical process realizes the automatic identification and removal of artifact signals and the comprehensive calculation of noise indicators. The output results are accurate and comprehensive, which significantly improves the accuracy and reliability of noise monitoring and provides high-quality data support for applications such as environmental noise control and equipment fault diagnosis. Attached Figure Description
[0012] Figure 1 This is a flowchart illustrating the present invention.
[0013] Figure 2 This is a schematic diagram of noise signal preprocessing according to the present invention.
[0014] Figure 3 This is a flowchart of the artifact noise recognition technology of the present invention.
[0015] Figure 4 This is a schematic diagram of the pseudo-signal pattern library and deep learning recognition model of the present invention. Detailed Implementation
[0016] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.
[0017] In a preferred embodiment of the present invention, those skilled in the art should note that noise signals and the like involved in the present invention can be considered as prior art.
[0018] Preferred embodiment.
[0019] like Figure 1As shown, this invention discloses a method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning, comprising the following steps: Step S1: Continuously collect complete noise signals from the environment and convert the audio input signal into a digital signal (use a microphone or other device to continuously collect all noise information from the environment). Step S2: Preprocess the digital signal by dividing it into multi-scale short-time frames through a multi-step framing operation; Step S3: Identify whether short frames contain artifact noise and remove short frames containing artifact noise; Step S4: Maintain and update the artifact signal pattern library and deep learning recognition model; Step S5: Post-process the noise signal after artifact noise identification and calculate the kurtosis value of the noise segment by segment; Step S6: Calculate the continuous equivalent sound pressure level and other indicators for the noise signal after processing, and simultaneously perform a weighted average of the kurtosis calculated segment by segment to obtain the final noise detection result.
[0020] Specifically, such as Figure 2 As shown, step S2 is specifically implemented as follows: Step S2.1: First, the digital signal is uniformly divided into equal-length frames with relatively large time intervals (e.g., divided into minute-level durations). Step S2.2: Perform secondary framing according to the percentile of the duration of all labeled samples in the artifact signal pattern library, and obtain multi-scale framing results of extremely short time frames (0.1 seconds), medium short time frames (seconds), and long short time frames (10 seconds) in each equal-length frame segment; Step S2.3: During the framing process, the overlap between each window is the preset value of the window length (25%). At the same time, a hybrid window function that is dynamically adjusted according to the kurtosis value of the framing signal is added to suppress spectral leakage (the main window function is the Blackman-Harris window, and the Kaiser window is enabled when the kurtosis value of the framing signal is >15).
[0021] More specifically, such as Figure 3 As shown, for step S3, the input is all short-time frame data, and the output is all short-time frame data identified as not containing artifact noise. The specific implementation is as follows: Step S3.1: Use various audio feature statistical extraction methods to extract time-frequency features including zero-crossing rate, energy entropy, MFCC and Mel spectrum from short frames, thereby obtaining audio feature vectors of short frame data. The original data of a noisy signal will obtain three sets of short frame audio feature vectors with different frame lengths, denoted as long short frame audio feature vector, medium short frame audio feature vector and very short frame audio feature vector. Step S3.2: Based on the cosine similarity algorithm of feature vectors A batch of labeled samples most similar to the target short-time frames (i.e., the audio feature vectors of long short-time frames, medium short-time frames, and very short short-time frames that need to be identified) are retrieved from the artifact signal pattern library, and the similarity scores of the target short-time frames are recorded. The similarity scores of the audio feature vectors of the three groups of short-time frames with different frame lengths are recursively averaged upwards according to their temporal relationship, and finally the recursively weighted similarity scores of the audio feature vectors of long and short frames are obtained. ,in Weighting weights between different frame levels; Step S3.3: Simultaneously input the long and short time frame audio feature vectors from step S3.1 into the deep learning recognition model to identify whether they contain artifact noise, and output the probability score of judging that they contain artifact noise. Step S3.4: Perform a weighted average of the similarity score and the probability score (the weights can be adjusted according to the completeness of the fake signal pattern library and the training status of the deep learning recognition model; the default is equal weighting). Based on the weighted score, retain the short and long time frames with weighted scores less than the set threshold and proceed to the subsequent processing stage; otherwise, determine that they contain fake noise and remove the short and long time frames.
[0022] Furthermore, such as Figure 4 As shown, for step S4: Step S4.1: For the artifact signal pattern library, the artifact signal pattern library is used to store known artifact noise signal data (which can be implemented as a vector database table). By using existing database indexing, search and update techniques, the data stored in the artifact signal pattern library can be searched and matched, and real-time addition, deletion and modification operations can be performed (in the implementation process, the feature vector of the labeled sample is obtained according to the feature extraction method in step S3.1, and stored in the vector database together with other static attributes of the labeled sample in index order). Step S4.2: The deep learning-based artifact signal recognition model is a convolutional-self-attention model with a CNN+Attention architecture, which effectively captures long-range and short-range dependencies in audio features. The deep learning recognition model is pre-trained using known noise and artifact noise data to obtain the model parameter weights of the artifact signal recognition algorithm, and then updated through subsequent fine-tuning training. The model training process includes feature extraction, data augmentation, and optimization training (where feature extraction follows the same method as in step S3.1; data augmentation mainly utilizes the CNN network to efficiently capture local time-frequency patterns in the feature vector; the optimization training objective is to identify whether the input sample is an artifact signal, and contrastive loss and cross-entropy loss are used to optimize the model recognition effect. Finally, dropout regularization and early stopping methods are used during training to avoid data overfitting).
[0023] It is worth mentioning that the specific architecture of the deep learning recognition model is as follows: convolutional layer (1×32 convolutional kernel) - max pooling layer 1 - convolutional layer 2 (32×64 convolutional kernel) - max pooling layer 2 - fully connected layer - self-attention module - fully connected layer - activation function.
[0024] Furthermore, in step S5, the short-time frame data identified as free of artifact noise are merged and reassembled into complete information through kurtosis calculation and frame merging. For the uncovered parts, interpolation is performed based on the preceding and following frames. The equivalent sound pressure level and kurtosis information are then calculated for the equal-length frames in step S2.1. The calculation method for the equivalent sound pressure level is as follows: ; in, It is the instantaneous sound pressure in the equal-length framed data. It is a reference sound pressure level. The duration of frames of equal length, The duration of the combined noise signal. It is the equivalent sound pressure level of equal-length frames, used to calculate the overall equivalent sound pressure level of the noise signal; The method for calculating noise kurtosis is as follows: ; in, For the first frame of equal-length data The amplitude at each location The amplitude mean is calculated for the equal-length frame data. The kurtosis value of the overall noise signal is calculated using the arithmetic mean and geometric mean methods respectively. (The input of this step is all the short-time frame noise data, and the output is the processed complete noise audio information and noise kurtosis information.)
[0025] For the present invention: 1. In terms of technical implementation, this invention addresses the technical challenge of the high similarity in spectral features between noise and artifact signals (their difference is significantly smaller than that between noise and speech signals). It innovatively employs a discrimination process combining multi-scale feature modeling and a deep learning recognition model. Compared to traditional noise power spectrum estimation methods, this scheme achieves a significant improvement in the accuracy of artifact signal recognition. 2. At the data level, this invention establishes a comprehensive and detailed library of industrial noise artifact signal patterns. This library enables rapid identification and location of industrial noise artifacts, providing reliable data support for subsequent signal processing and noise reduction.
[0026] It is worth mentioning that the technical features such as noise signals involved in this patent application should be regarded as prior art. The specific structure, working principle, and possible control methods and spatial arrangement of these technical features can be conventionally selected in the field and should not be regarded as the inventive point of this patent. This patent will not be further elaborated in detail.
[0027] For those skilled in the art, 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. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.
Claims
1. A method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning, characterized in that, Includes the following steps: Step S1: Continuously collect complete noise signals from the surrounding environment and convert the audio input signal into a digital signal; Step S2: Preprocess the digital signal by dividing it into multi-scale short-time frames through a multi-step framing operation; Step S3: Identify whether short frames contain artifact noise and remove short frames containing artifact noise; Step S4: Maintain and update the artifact signal pattern library and deep learning recognition model; Step S5: Post-process the noise signal after artifact noise identification and calculate the kurtosis value of the noise segment by segment; Step S6: Calculate the continuous equivalent sound pressure level and other indicators for the noise signal after processing, and simultaneously perform a weighted average of the kurtosis calculated segment by segment to obtain the final noise detection result.
2. The method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning according to claim 1, characterized in that, Step S2 is specifically implemented as follows: Step S2.1: First, the digital signal is uniformly divided into equal-length frames with relatively large time intervals; Step S2.2: Perform secondary framing according to the percentile of the duration of all labeled samples in the artifact signal pattern library, and obtain multi-scale framing results of extremely short time frames, medium short time frames, and long short time frames in each equal-length frame segment; Step S2.3: During the framing process, the overlap between each window is a preset value of the window length, and a hybrid window function that is dynamically adjusted according to the kurtosis value of the framing signal is added to suppress spectral leakage.
3. The method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning according to claim 2, characterized in that, For step S3, the input is all short-time frame data, and the output is all short-time frame data identified as not containing artifact noise. The specific implementation is as follows: Step S3.1: Use various audio feature statistical extraction methods to extract time-frequency features including zero-crossing rate, energy entropy, MFCC and Mel spectrum from short frames, thereby obtaining audio feature vectors of short frame data. The original data of a noisy signal will obtain three sets of short frame audio feature vectors with different frame lengths, denoted as long short frame audio feature vector, medium short frame audio feature vector and very short frame audio feature vector. Step S3.2: Based on the cosine similarity algorithm of feature vectors, recall a batch of labeled samples that are most similar to the target short-time frame from the pseudo-signal pattern library, and record the similarity score of the target short-time frame; the similarity scores of the audio feature vectors of the three groups of short-time frames with different frame lengths are recursively averaged upwards according to their temporal relationship, and finally the recursive weighted similarity scores of the long and short-time frames are obtained. Step S3.3: Simultaneously input the long and short time frame audio feature vectors from step S3.1 into the deep learning recognition model to identify whether they contain artifact noise, and output the probability score of judging that they contain artifact noise. Step S3.4: Calculate the weighted average of the similarity score and the probability score. Based on the weighted score, retain the short and long time frames whose weighted scores are less than the set threshold and proceed to the next processing step. Otherwise, determine that the frame contains artifact noise and remove it.
4. The method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning according to claim 3, characterized in that, For step S4: Step S4.1: For the pseudo-signal pattern library, the pseudo-signal pattern library is used to store known pseudo-noise signal data. By using existing database indexing, search and update techniques, the data stored in the pseudo-signal pattern library can be searched and matched, and real-time addition, deletion and modification operations can be performed. Step S4.2: The deep learning-based artifact signal recognition model is a convolutional-self-attention model with a CNN+Attention architecture, which effectively captures long-range and short-range dependencies in audio features. The deep learning recognition model is pre-trained using known noise and artifact noise data to obtain the model parameter weights of the artifact signal recognition algorithm, and the model is updated through subsequent fine-tuning training. The model training process includes feature extraction, data augmentation, and optimization training.
5. The method for dynamic identification of noise artifacts based on multi-scale feature modeling and deep learning according to claim 4, characterized in that, For step S5, the short-time frame data identified as free of artifact noise are merged and reassembled into complete information through kurtosis calculation and frame merging. For the uncovered parts, interpolation is performed based on the situation of the preceding and following frames. Furthermore, the equivalent sound pressure level and kurtosis information of the equal-length frames in step S2.1 are calculated. The method for calculating the equivalent sound pressure level is as follows: ； in, It is the instantaneous sound pressure in the equal-length framed data. It is a reference sound pressure level. The duration of frames of equal length, The duration of the combined noise signal. It is the equivalent sound pressure level of equal-length frames, used to calculate the overall equivalent sound pressure level of the noise signal; The method for calculating noise kurtosis is as follows: ； in, For the first frame of equal-length data The amplitude at each location The amplitude mean is given in the equally segmented frame data. The kurtosis of the overall noise signal is calculated using both arithmetic and geometric mean methods.