Rotary machinery multi-granularity voiceprint feature extraction method and system and medium

Through the multi-particle voiceprint feature extraction method of multi-spectral graph with Gaussian differential pyramid and multi-head attention mechanism, combined with the nonlinear mapping of Mel and MFCCs, the problem of insufficient robustness of voiceprint feature extraction of rotating mechanical faults in strong noise environments is solved. , achieving higher fault identification accuracy and information characterization capabilities.

CN119068910APending Publication Date: 2024-12-03XIAN UNIV OF SCI & TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411129654.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art is difficult to effectively extract the soundprint characteristics of rotating machinery failures in a strong noise environment, resulting in slow data analysis processing speed and difficult to obtain fault characteristics, poor robustness, information loss, and difficulty in extracting comprehensive sounds adaptively. Text information.

Method used

Using a multi-grained voiceprint feature extraction method based on Gaussian differential pyramid and multi-head attention mechanism, combining the nonlinear mapping of Mel spectrum and MFCCs, an acoustic and image data sample set is constructed, and a multi-head attention module is introduced to realize Mel and MFCCs. Adaptive extraction of multi-grained vocalprint features of multi-spectral graphs.

Benefits of technology

It improves the soundprint feature characterization ability of low-frequency signals, comprehensively characterizes the soundprint features, and effectively extracts the soundprint features with relatively complete information characterization in rotating machinery fault recognition in strong noise environments, improving the accuracy of fault recognition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068910A_ABST
    Figure CN119068910A_ABST
Patent Text Reader

Abstract

The invention provides a rotating machine multi-granularity voiceprint feature extraction method and system and a medium, and relates to the technical field of fault recognition, and the method comprises the steps: obtaining an audio signal of a rotating machine, carrying out the fast Fourier transform of each frame of audio signal after pre-emphasis processing, and obtaining the spectral line energy of each frame of audio signal; enabling the spectral line energy of each frame of audio signal to pass through a Mel filter bank to obtain a Mel spectrogram and a MFCCs spectrogram, respectively constructing two Gaussian difference pyramids, and carrying out Gaussian convolution smoothing processing and down-sampling on the two Gaussian difference pyramids for multiple times to obtain a multi-granularity voiceprint feature map; respectively calculating the attention weight of each layer of voiceprint feature map by using a plurality of independent attention heads for each layer of voiceprint feature map, and constructing a weight matrix; and weighting each layer of voiceprint feature map and the weight matrix to obtain a weight so as to realize self-adaptive extraction of the multi-granularity voiceprint feature map. The method can effectively extract voiceprint features with relatively complete information representation in rotating machinery fault identification in a strong noise environment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of rotating machinery fault identification, and in particular to a rotating machinery multi-granularity voiceprint feature extraction method, system and medium. Background Art

[0002] Fault diagnosis technology is crucial to ensure the reliability, safety and maintainability of rotating machinery, and has attracted widespread attention from scholars at home and abroad. With the development and application of artificial intelligence methods, the fault feature diagnosis method of rotating machinery is developing in the direction of automation and intelligence. The data-driven fault diagnosis method mines and analyzes big data information representing the equipment status to achieve real-time and intelligent fault diagnosis of the equipment, which has become a hot topic in current research. The sound signal of mechanical equipment contains rich information and can reflect its own working status and fault conditions. Acoustic signals have attracted widespread attention in intelligent fault diagnosis due to their non-contact layout and low cost. However, voiceprint signals are easily affected by noise and other equipment sound signals, and their reflection and scattering during the propagation process will also affect the accurate extraction of voiceprint features. This makes the data analysis and processing slow and the fault characteristics difficult to obtain, so voiceprint feature extraction has become a hot topic in current research.

[0003] Shan (Shan S, Liu J, Wu S, et al. A motor bearing fault voiceprint recognition method based on Mel-CNN model [J]. Measurement, 2023, 207: 112408.) used variational mode decomposition (VMD) to remove the high-frequency components of motor noise, extract Mel spectrum voiceprint features, and re-extracted Mel voiceprint features using convolutional neural network (CNN). Guoqiang Li (Guoqiang Li, Jun Wu, Chao Deng, Xuebing Xu, Xinyu Shao. Deep reinforcement learning-based online domain adaptation method for fault diagnosis of rotating machinery[J]. IEEE / ASME Transactions on Mechatronics.) combined the fault diagnosis method of Mel-frequency cepstral coefficient (MFCC) and designed a parallel multi-fusion convolutional neural network (MFCNN) to improve its diagnostic performance in different noise environments. Zhuoran Yu et al. (Yu Z, Wei Y, Niu B, et al. Automatic Condition Monitoring and Fault Diagnosis System for Power Transformers Based on Voiceprint Recognition[J]. IEEE Transactions on Instrumentation and Measurement, 2024.) proposed a voiceprint feature extraction method for transformers under different operating conditions based on short-time energy and zero-crossing rate characteristics, and constructed a power transformer condition monitoring and fault diagnosis system based on hidden Markov model, realizing automatic identification of fault categories. However, the above feature extraction methods are all constructed from a single or a small number of features, with poor robustness, and information representation inevitably has information missing, making it difficult to adaptively extract comprehensive voiceprint information. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a method for extracting multi-granularity voiceprint features of rotating machinery, which can effectively extract voiceprint features with relatively complete information representation in rotating machinery fault identification under strong noise environment.

[0005] To achieve the above object, the present invention provides the following technical solutions.

[0006] A method for extracting multi-granularity voiceprint features of a rotating machine comprises the following steps:

[0007] Acquire the audio signal of the rotating machinery and perform pre-emphasis processing on the audio signal;

[0008] Performing fast Fourier transform on each frame of audio signal after pre-emphasis processing to obtain the spectral line energy of each frame of audio signal;

[0009] The spectral line energy of each frame of audio signal is passed through the Mel filter bank to obtain the logarithmic energy of each Mel filter; a Mel spectrogram is constructed based on the logarithmic energy; the logarithmic energy is subjected to discrete cosine transformation to obtain MFCC coefficients; and an MFCCs spectrogram is constructed based on the MFCC coefficients;

[0010] Gaussian pyramids are constructed separately, and multiple Gaussian convolution smoothing and downsampling are performed on the Mel spectrogram and MFCCs spectrogram. Pixel-level difference operations are performed on two adjacent layers of images in the Gaussian pyramid, and two Gaussian difference pyramids are constructed according to the voiceprint feature map, and multi-granularity voiceprint feature maps are generated group by group and layer by layer.

[0011] For each layer of voiceprint feature graph of each Gaussian difference pyramid, use multiple independent attention heads to calculate the attention weight of each layer of voiceprint feature graph and construct a weight matrix;

[0012] The voiceprint feature map of each layer of each Gaussian difference pyramid is weighted with the weight matrix to obtain the weighted voiceprint feature; the two weighted voiceprint features are then weightedly fused to obtain the multi-granularity voiceprint feature.

[0013] Preferably, the pre-emphasis processing of the audio signal comprises the following steps:

[0014] Determine the frequency domain representation of the audio signal as:

[0015] H(z)=1-μz -1

[0016] Where H(z) represents the frequency domain result of the output signal; μ represents the pre-emphasis coefficient; z represents the frequency;

[0017] The audio signal is framed and windowed; the Hamming window w(n) is selected as the window function; the sound signal after the window function is:

[0018] y i (n) = x' i (n)w(n)

[0019] in:

[0020]

[0021] In the formula, x' i (n) is the pre-emphasized sound signal, y i (n) is the time domain signal of each frame after the window function is added; N is the number of sample points, and n is the current frame number.

[0022] Preferably, the acquisition of the spectral line energy of each frame of the audio signal comprises the following steps:

[0023] Perform fast Fourier transform on each frame of audio signal obtained after pre-emphasis processing to obtain Y i (k), which can be used to calculate the spectral line energy E of each frame of audio signal i (k):

[0024]

[0025] E i (k) = [Y i (k)] 2

[0026] In the formula, y i (n) is the time domain signal of each frame after the window function is added; N is the number of sample points.

[0027] Preferably, the solution of the logarithmic energy is shown as follows:

[0028]

[0029] In the formula, H m (k) is the response of the mth filter in the Mel filter bank, and M is the number of filters.

[0030] Preferably, the solution of the MFCC coefficients comprises the following steps:

[0031] For S i (m) is subjected to discrete cosine transform to obtain the MFCC coefficient M(l) of the L-order signal, as shown in the following formula:

[0032]

[0033] Where M(i,l) is the l-th dimension MFCC coefficient of the i-th frame audio signal;

[0034] Among them, the maximum Mel frequency of 8000 Hz and the first 13 MFCC coefficients are selected as the feature extraction parameters of the acoustic image data.

[0035] Preferably, the construction of the Gaussian difference pyramid includes the following steps:

[0036] The Mel spectrogram or MFCCs spectrogram is doubled and used as the first group and first layer of the Gaussian pyramid. After being smoothed by Gaussian convolution, it is used as the second group and the Gaussian convolution function is:

[0037]

[0038] Where σ is the smoothing factor, which takes a fixed value of 1.6; x and y represent the pixel positions of the image; x0 and y0 represent the center coordinates (0,0);

[0039] Multiply σ by a proportional coefficient K to obtain a new smoothing factor; smooth the second layer image of the first group by the new smoothing factor, and the obtained result is used as the third layer;

[0040] Repeat the process to obtain L-layer images and construct a Gaussian pyramid. In the same group, the smoothing coefficients corresponding to the images are: 0, σ, Kσ, K*2σ, K*3σ…K*(L-2)σ.

[0041] A pixel-level difference operation is performed on two adjacent layers of images in the Gaussian pyramid, and each difference image is generated group by group and layer by layer. All difference images constitute a Gaussian difference pyramid.

[0042] A rotating machinery multi-granularity voiceprint feature extraction system, the system comprising:

[0043] processor;

[0044] a memory having stored thereon a computer program executable on the processor;

[0045] Wherein, when the computer program is executed by the processor, the steps of the rotating machinery multi-granularity voiceprint feature extraction method are implemented.

[0046] A computer-readable storage medium stores a data processing program, and when the data processing program is executed by a processor, the steps of the rotating machinery multi-granularity voiceprint feature extraction method are implemented.

[0047] Beneficial effects of the present invention:

[0048] The present invention proposes a multi-spectrogram multi-granularity voiceprint feature extraction method based on Gaussian difference pyramid and multi-head attention mechanism, combines the nonlinear mapping advantages of high and low frequency features of Mel spectrum, constructs an audio-visual data sample set based on Mel and MFCCs, and improves the voiceprint feature representation capability of low frequency signals; the method comprehensively considers the geometric and semantic information features in the graph information, proposes a multi-granularity representation method for multi-voiceprint graph information based on Gaussian split pyramid, and introduces a multi-head attention mechanism module to realize the adaptive extraction of multi-granularity voiceprint features of multi-spectrograms of Mel and MFCCs, comprehensively characterizes voiceprint features, and effectively extracts voiceprint features with relatively complete information representation in rotating machinery fault identification under strong noise environment, further improving the fault identification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flowchart of multi-granularity voiceprint feature extraction of a rotating machine according to an embodiment of the present invention;

[0050] Figure 2 It is a framework diagram of a method for adaptively extracting multi-granularity voiceprint features of rotating machinery by fusing multi-spectrogram information according to an embodiment of the present invention;

[0051] Figure 3 This is a flow chart of a multi-granularity voiceprint feature fusion method based on a multi-head attention mechanism according to an embodiment of the present invention;

[0052] Figure 4 is a sound image feature set constructed under different bearing fault modes in an embodiment of the present invention;

[0053] Figure 5 is a multi-spectrogram multi-granularity feature graph of an embodiment of the present invention;

[0054] Figure 6 t-SNE feature visualization diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] Example 1

[0057] The present invention provides a method for adaptively extracting multi-granularity voiceprint features of rotating machinery by fusing multi-spectrogram information. The specific process is as follows: Figure 1 As shown, the adaptation extraction framework is shown in Figure 2 As shown, the following steps are included:

[0058] S1: Acquire the audio signal of the rotating machinery and perform pre-emphasis processing on the audio signal.

[0059] S2: Perform fast Fourier transform on each frame of audio signal after pre-emphasis processing to obtain spectral line energy of each frame of audio signal.

[0060] S3: Pass the spectral line energy of each frame of audio signal through the Mel filter group to obtain the logarithmic energy of each Mel filter; construct a Mel spectrogram based on the logarithmic energy; perform discrete cosine transform on the logarithmic energy to obtain MFCC coefficients; construct an MFCCs spectrogram based on the MFCC coefficients.

[0061] S4: Construct Gaussian pyramids respectively, perform multiple Gaussian convolution smoothing and downsampling on the Mel spectrogram and MFCCs spectrogram; perform pixel-level difference operations on two adjacent layers of images in the Gaussian pyramid, construct two Gaussian difference pyramids based on the voiceprint feature map, and generate multi-granularity voiceprint feature maps group by group and layer by layer.

[0062] S5: For each layer of the voiceprint feature map of each Gaussian difference pyramid, use multiple independent attention heads to calculate the attention weight of each layer of the voiceprint feature map and construct a weight matrix.

[0063] S6: Weight each layer of the voiceprint feature map of each Gaussian difference pyramid with the weight matrix to obtain the weighted voiceprint feature; perform weighted fusion on the two weighted voiceprint features to obtain the multi-granularity voiceprint feature. The fusion method flow chart is as follows: Figure 3 shown.

[0064] Further, the pre-emphasis processing of the audio signal in S1 includes the following steps:

[0065] Determine the frequency domain representation of the audio signal as:

[0066] H(z)=1-μz -1

[0067] Where H(z) represents the frequency domain result of the output signal; μ represents the pre-emphasis coefficient; z represents the frequency;

[0068] The audio signal is framed and windowed; the Hamming window w(n) is selected as the window function; the sound signal after the window function is:

[0069] y i (n) = x' i (n)w(n)

[0070] in:

[0071]

[0072] In the formula, x' i(n) is the pre-emphasized sound signal, y i (n) is the time domain signal of each frame after the window function is added; N is the number of sample points.

[0073] Furthermore, the spectral line energy acquisition in S2, the logarithmic energy acquisition in S3, and the solution of the MFCC coefficients include the following steps:

[0074] S3.1: Acquisition of spectral line energy of each frame of audio signal, comprising the following steps:

[0075] Perform fast Fourier transform on each frame of audio signal obtained after pre-emphasis processing to obtain Y i (k), which can be used to calculate the spectral line energy E of each frame of audio signal i (k):

[0076]

[0077] E i (k) = [Y i (k)] 2

[0078] In the formula, y i (n) is the time domain signal of each frame after the window function is added; N is the number of sample points.

[0079] S3.2: The solution for the logarithmic energy is as follows:

[0080]

[0081] In the formula, H m (k) is the response of the mth filter in the Mel filter bank, and M is the number of filters.

[0082] S3.3: Solving the MFCC coefficients includes the following steps:

[0083] For S i (m) is subjected to discrete cosine transform to obtain the MFCC coefficient M(l) of the L-order signal, as shown in the following formula:

[0084]

[0085] Where M(i,l) is the l-th dimension MFCC coefficient of the i-th frame audio signal;

[0086] Among them, the maximum Mel frequency of 8000 Hz and the first 13 MFCC coefficients are selected as the feature extraction parameters of the acoustic image data.

[0087] Furthermore, the construction of the Gaussian difference pyramid in S4 includes the following steps:

[0088] S4.1: After doubling the Mel spectrogram or MFCCs spectrogram, it is used as the first group and the first layer of the Gaussian pyramid, and after being smoothed by Gaussian convolution, it is used as the second layer of the first group, where the Gaussian convolution function is:

[0089]

[0090] Where σ is the smoothing factor, which is fixed at 1.6; x and y represent the pixel positions of the image; x0 and y0 represent the center coordinates, which are generally (0,0);

[0091] S4.2: multiply σ by a proportional coefficient K to obtain a new smoothing factor; smooth the second layer image of the first group by the new smoothing factor, and the obtained result is used as the third layer;

[0092] S4.3: Repeat the execution to obtain L-layer images and construct a Gaussian pyramid; in the same group, the smoothing coefficients corresponding to the images are: 0, σ, Kσ, K*2σ, K*3σ…K*(L-2)σ;

[0093] S4.4: Perform pixel-level difference operation on two adjacent layers of images in the Gaussian pyramid, generate each difference image group by group and layer by layer, and all difference images constitute a Gaussian difference pyramid.

[0094] In this embodiment, the above feature extraction method is verified as follows:

[0095] 1. Construct a rotating machinery fault diagnosis dataset for rolling bearings

[0096] Based on the rotor comprehensive fault simulation test bench data set provided by iFlytek and the National Intelligent Speech Center, the soundprint features under different fault modes are extracted. The load torque range of the rotor test bench is 0-6N·m, and the bearing models used are all NSK 6200. The audio data in the data set is collected by a pickup with a sampling rate of 48000Hz and a length of 4 seconds for each sample. The data label types include normal, pitting of the outer ring of the bearing, pitting of the inner ring of the bearing, pitting of the bearing rolling element, and fracture of the bearing cage. The data set includes three parts: training set, verification set and test set. This embodiment uses the same batch of training set samples to verify the soundprint feature extraction method. The specific sample situation is shown in Table 1 below: In order to expand the data samples under each rolling bearing fault state, the sliding window is used to overlap the vibration data for sampling, the sampling time is 1s, the overlap rate is 40%, and noise processing is performed. Considering the actual multi-noise intensity environment, 3600 data samples are obtained, and the labels and sample information are shown in Table 1.

[0097] Table 1 Samples of training set and test set

[0098]

[0099]

[0100] 2. Acoustic image information sample construction

[0101] In the time-frequency image, the horizontal axis represents time, the vertical axis represents frequency, and the color or brightness of the image represents the energy or power distribution of the signal at different frequencies and times. This embodiment processes the sound data and constructs a multi-spectrogram information sample set based on Mel and MFCCs. The generated RGB image has three-channel information, which represents the spectral characteristics and chromaticity characteristics contained in the sound image information under different fault modes, such as Figure 4 As shown in Figure 2, there are differences in the Mel and MFCCs acoustic image information representation based on sound signals under different bearing fault modes. Gaussian pyramid is used to extract multi-spectrogram multi-granularity samples of cages and rollers. The results are shown in Figure 2. Figure 5 shown.

[0102] Under multi-level noise conditions, a comparative analysis of the representation capabilities of different feature extraction schemes for voiceprint information is conducted. The proposed comparison scheme is shown in Table 2, where multi-spectrogram feature extraction is information fusion of Mel and MFCCs, and voiceprint features are reflected in the adaptive extraction of multi-granularity information based on the fusion of multi-spectrogram information features.

[0103] Table 2 Feature extraction comparison scheme

[0104]

[0105] Combined with the LeNet-5 model, the model was tested for fault diagnosis. The diagnostic efficiency of the model was verified by using the mean square error loss function, iterating 50 rounds, and taking 10 averages. During the training and verification of the model, the number of epochs was set to 50, the batch size was set to 64, the learning rate was set to 0.001, and the optimizer was set to Adam. The recognition rate of the test set with different feature extraction schemes was ablated, and the results are shown in Table 3.

[0106] Table 3 Ablation experiment of recognition rate of test set with different feature extraction schemes

[0107]

[0108] As shown in Table 3, the recognition rate of the original audio signal plus noise is 60.28%, the recognition rate of multi-domain feature fusion is 93.89%, and the recognition rate of the multi-domain multi-granularity feature adaptive extraction scheme can be improved to 97.64%, which is 37.36% higher than the original data scheme and 3.75% higher than the multi-domain feature. This feature extraction scheme effectively improves the representation ability of voiceprint information.

[0109] In order to further verify the multi-domain and multi-granularity voiceprint feature extraction capability, this embodiment uses t-SNE technology to analyze the voiceprint features under the noise state, and projects the output features of the FC layer (including the extracted sample feature information) of the LeNet-5 model onto a two-dimensional plane for visualization, as shown in Figure 2. Figure 6 As shown, each point in the figure represents a sample, different colors represent different sample labels, and different types of samples are scattered in relatively independent local areas, indicating that the feature extraction method can accurately extract the abstract representation of the fault features in the selected data set.

[0110] From the comparison of feature extraction performance, Figure 6 As shown, after t-SNE processing, there are interlaced areas of voiceprint features in the red circle areas 1, 2, 3, 4 and 5, which leads to a high possibility of potential errors in fault diagnosis. The fault feature interface obtained by the MS-MGFAE method proposed in this embodiment is clearer, the aggregation effect is optimal and the diagnostic error is minimized, which shows that the MS-MGFAE method proposed in this embodiment can effectively improve its own voiceprint feature characterization capability.

[0111] The above is a rotating machinery multi-granularity voiceprint feature extraction method provided by one embodiment of this embodiment. Based on the same idea, this embodiment also provides a corresponding rotating machinery multi-granularity voiceprint feature extraction system. For the specific definition of the rotating machinery multi-granularity voiceprint feature extraction system, please refer to the definition of the rotating machinery multi-granularity voiceprint feature extraction method in the above text, which will not be repeated here. Each module in the above-mentioned rotating machinery multi-granularity voiceprint feature extraction system can be fully or partially implemented by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0112] This embodiment also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A multi-granularity voiceprint feature extraction method for rotating machinery is provided.

[0113] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0114] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for extracting multi-granularity voiceprint features from rotating machinery. It is characterized in that The following steps are involved: Acquire the audio signal of the rotating machinery and perform pre-emphasis processing on the audio signal; Performing fast Fourier transform on each frame of audio signal after pre-emphasis processing to obtain the spectral line energy of each frame of audio signal; The spectral line energy of each frame of audio signal is passed through the Mel filter bank to obtain the logarithmic energy of each Mel filter; a Mel spectrogram is constructed based on the logarithmic energy; the logarithmic energy is subjected to discrete cosine transformation to obtain MFCC coefficients; and an MFCCs spectrogram is constructed based on the MFCC coefficients; Gaussian pyramids are constructed separately, and multiple Gaussian convolution smoothing and downsampling are performed on the Mel spectrogram and MFCCs spectrogram. Pixel-level difference operations are performed on two adjacent layers of images in the Gaussian pyramid, and two Gaussian difference pyramids are constructed according to the voiceprint feature map, and multi-granularity voiceprint feature maps are generated group by group and layer by layer. For each layer of voiceprint feature graph of each Gaussian difference pyramid, use multiple independent attention heads to calculate the attention weight of each layer of voiceprint feature graph and construct a weight matrix; Weighting each layer of the voiceprint feature map of each Gaussian difference pyramid with the weight matrix to obtain the weighted voiceprint feature; The two weighted voiceprint features are then weightedly fused to achieve adaptive extraction of multi-granularity voiceprint feature maps.

2. The method for extracting multi-granularity voiceprint features of rotating machinery according to claim 1, It is characterized in that The pre-emphasis processing of the audio signal comprises the following steps: Determine the frequency domain representation of the audio signal as: H(z)=1-μz -1 Where H(z) represents the frequency domain result of the output signal; μ represents the pre-emphasis coefficient; z represents the frequency; The audio signal is framed and windowed; the Hamming window w(n) is selected as the window function; the sound signal after the window function is: y i (n)=x' i (n)w(n) in: In the formula, x' i (n) is the pre-emphasized sound signal, y i (n) is the time domain signal of each frame after the window function is added; N is the number of sample points, and n is the current frame number.

3. The method for extracting multi-granularity voiceprint features of rotating machinery according to claim 2, It is characterized in that The acquisition of the spectral line energy of each frame of the audio signal comprises the following steps: Perform fast Fourier transform on each frame of audio signal obtained after pre-emphasis processing to obtain Y i (k), k is the current sample point; use this value to calculate the spectral line energy E of each frame of audio signal i (k): AND i (k)=[And i (k)] 2 In the formula, y i (n) is the time domain signal of each frame after the window function is added; N is the number of sample points.

4. The method for extracting multi-granularity voiceprint features of rotating machinery according to claim 3, It is characterized in that The solution of the logarithmic energy is shown as follows: In the formula, H m (k) is the response of the mth filter in the Mel filter bank, and M is the number of filters.

5. The method for extracting multi-granularity voiceprint features of rotating machinery according to claim 4, It is characterized in that The solution of the MFCC coefficients includes the following steps: For S i (m) is subjected to discrete cosine transform to obtain the MFCC coefficient M(l) of the L-order signal, as shown in the following formula: Where M(i,l) is the l-th dimension MFCC coefficient of the i-th frame audio signal; Among them, the maximum Mel frequency of 8000 Hz and the first 13 MFCC coefficients are selected as the feature extraction parameters of the acoustic image data.

6. The method for extracting multi-granularity voiceprint features of rotating machinery according to claim 1, It is characterized in that The construction of the Gaussian difference pyramid includes the following steps: The Mel spectrogram or MFCCs spectrogram is doubled and used as the first group and first layer of the Gaussian pyramid. After being smoothed by Gaussian convolution, it is used as the second group and the Gaussian convolution function is: In the formula, σ is the smoothing factor, which takes a fixed value of 1.6; x and y represent the pixel positions of the image; x 0 ,y 0 It is represented as the center sitting (0,0); Multiply σ by a proportional coefficient K to obtain a new smoothing factor; smooth the second layer image of the first group by the new smoothing factor, and the obtained result is used as the third layer; Repeat the process to obtain L-layer images and construct a Gaussian pyramid. In the same group, the smoothing coefficients corresponding to the images are: 0, σ, Kσ, K*2σ, K*3σ…K*(L-2)σ. A pixel-level difference operation is performed on two adjacent layers of images in the Gaussian pyramid, and each difference image is generated group by group and layer by layer. All difference images constitute a Gaussian difference pyramid.

7. A multi-granularity voiceprint feature extraction system for rotating machinery, It is characterized in that The system comprises: processor; a memory having stored thereon a computer program executable on the processor; Wherein, when the computer program is executed by the processor, the steps of the method for extracting multi-granularity voiceprint features of rotating machinery as described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a data processing program, and when the data processing program is executed by the processor, the steps of the method for extracting multi-granularity voiceprint features of rotating machinery according to any one of claims 1 to 6 are implemented.

Citation Information

Cited By

  • Mechanical equipment voiceprint feature two-dimensional adaptive extraction method, system and device

    CN120032649A