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

Abstract

The application discloses a mechanical equipment voiceprint feature two-dimension adaptive extraction method, system and device, relates to the voiceprint feature extraction technical field, and the method comprises the following steps: collecting the audio signal of a mechanical equipment, and extracting the MDS of the audio signal; constructing a space-spectrogram feature set based on MDS, Mel and MFCC; constructing a Gaussian pyramid by carrying out Gaussian smoothing processing and down-sampling on multiple spectrograms; carrying out pixel-level difference operation on adjacent two layers of images through the Gaussian pyramid to form a Gaussian difference pyramid; adaptively extracting the multi-granularity feature dimension of the multiple spectrograms by fusing the Gaussian difference pyramid and a multi-head attention mechanism; and combining a dynamic convolution dimension adaptive extraction method to dynamically generate adaptive convolution kernel weights, adaptively adjusting the extracted feature map, and obtaining the voiceprint feature of the mechanical equipment. The method extracts the voiceprint common features under variable working conditions, thereby improving the fault detection precision of the mechanical equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voiceprint feature extraction, and particularly relates to a mechanical equipment voiceprint feature double-dimension adaptive extraction method, system and device. BACKGROUND

[0002] Under the background of today's rapid development of science and technology, mechanical equipment is developing towards complexity and intelligence. At the same time, the diversified failure modes of mechanical equipment lead to unnecessary downtime loss and even personnel casualties. Therefore, equipment fault feature extraction is of great significance to the safe operation of the system. Domestic and foreign scholars have conducted a large number of researches on mechanical equipment operation state feature extraction based on data driving. At present, the most commonly used methods are to construct a variety of effective feature extraction methods to extract mechanical equipment fault features through vibration signals and to extract equipment operation state monitoring and feature extraction methods based on non-contact sensing technology. However, vibration signal sensing technology usually relies on contact measurement, which is particularly inconvenient in space-limited environments. Therefore, the research on equipment operation state monitoring and feature extraction methods based on non-contact sensing technology is increasingly important.

[0003] Audio sensors have the advantages of non-contact arrangement, flexible installation, easy signal acquisition, low cost, etc. The audio signal of mechanical equipment contains rich information, which can reflect its own working state and fault condition. Speech recognition and voiceprint recognition technology provides new possibilities for mechanical equipment running state audio signal monitoring and acoustic feature extraction. In recent years, researchers have obtained audio signals through non-contact voiceprint sensing technology and extracted the state features of mechanical equipment. However, the audio signal is easily affected by noise and reflection during transmission, making the feature extraction process complex and time-consuming. To solve this problem, scholars have proposed: (1) a wavelet domain moving beamforming (WMB) method is used to extract the acoustic signal processing method of train shaft bearing fault through envelope spectrum, which enhances the signal of train shaft bearing; (2) a far-field acoustic signal processing method based on uniform circular microphone array is used for gear fault detection; (3) a 2.5-dimensional (2.5D) acoustic field diagnosis method, the acoustic field model is composed of source image, holographic acoustic image and its difference, Gabor wavelet feature extraction and random forest feature reduction algorithm are used to construct its effective features; (3) a supervised adaptive spectral graph wavelet transform (SASGWT) method is used to realize the weak fault feature extraction of three-stage planetary gearbox acoustic signal. To solve the problems of insufficient audio signal samples and weak fault features in plunger pump fault diagnosis, (1) a plunger pump fault diagnosis method based on audio signal and meta-transfer learning (MTL-PAFD) is proposed; (2) a new method called adaptive graph Morlet wavelet transform (AGMWT) is used for fault feature extraction in railway trackside acoustic detection; (3) a time and frequency domain dual-domain audio feature extraction method is proposed, which combines the "scaling" characteristics of wavelet packet transform (WPT), the local feature extraction ability of one-dimensional convolutional neural network (CNN) and the self-attention mechanism, and improves the success rate of unmanned aerial vehicle detection in common scenarios; (4) a method based on MI-CNN is used to fuse vibration and acoustic features, which realizes the effective extraction of fault features of induction motor under different working conditions; (5) a new learning model based on pseudo-Doppler time domain demodulation (PDDD) and domain adaptive network (DAN) is proposed, which reduces the feature bias caused by the approximation error and Doppler frequency shift of the acoustic feature extraction layer due to Doppler time domain demodulation.

[0004] However, the existing feature extraction method is constructed by a single or a small number of features, which has poor robustness. In the actual complex working condition, the variable working condition causes the inconsistent distribution of equipment voiceprint features, and there is a domain offset. In the voiceprint feature extraction process, there may be information missing, making it difficult to extract complete voiceprint information, thereby reducing the detection accuracy of mechanical equipment faults. SUMMARY

[0005] In view of the fact that the prior art may have information missing in the process of extracting voiceprint features under variable working conditions, and it is difficult to extract complete voiceprint information, the present application provides a mechanical equipment voiceprint feature double-dimension adaptive extraction method, system and device, which constructs a space-spectrogram feature set based on MDS, Mel and MFCC, proposes a space-spectrogram feature set multi-granularity feature representation method based on a Gaussian split pyramid model, and introduces a multi-head attention mechanism to extract multi-granularity information features, realizes dynamic learning of the importance between multi-granularity information, and solves the problems existing in the prior art.

[0006] A mechanical equipment voiceprint feature double-dimension adaptive extraction method, comprising the following steps:

[0007] Acquire the audio signal of the mechanical equipment under variable working conditions, and extract the multi-domain feature spectrogram MDS, Mel frequency spectrum feature and mel frequency cepstrum coefficient MFCC of the audio signal, and construct a space-spectrogram feature set based on MDS, Mel frequency spectrum and MFCC; the extraction process of the multi-domain feature spectrogram of the audio signal comprises the following steps: assuming that the audio signal is , the first i frame audio signal obtained after preprocessing the audio signal is , and then the time domain feature is obtained; wherein the time domain feature comprises short-time energy E n ( i ), short-time amplitude M ( i ), short-time amplitude difference D i ( k ) and zero-crossing rate ; calculate the frequency domain feature of the audio signal; the frequency domain feature comprises spectral centroid FC i , sub-band energy ratio SER ij , pitch period T 0; based on the alignment relationship of the feature time dimension, the time domain and frequency domain features of the audio signal are spliced to generate a multi-domain feature spectrogram;

[0008] A Gaussian pyramid is constructed by performing multiple Gaussian smoothing and downsampling on the space-spectrogram feature set; each difference image is generated group by group and layer by layer by performing pixel-level difference operation on multiple groups of adjacent two layers of images in the Gaussian pyramid, and a Gaussian difference pyramid is formed based on all the difference images; the extreme value of each granularity space of the difference image in the Gaussian difference pyramid is detected to find the extreme points of the Gaussian difference function, and a multi-granularity voiceprint feature map is generated; the saliency features of the multi-granularity voiceprint feature map are extracted by introducing a multi-head attention mechanism;

[0009] A dynamic convolutional adaptive extraction module is used to dynamically generate adaptive convolutional kernel weights and adaptively adjust the extracted feature maps to obtain the acoustic features of mechanical equipment under varying working conditions.

[0010] Furthermore, the short-time energy E n ( i ), short-term amplitude M ( i ), short-term amplitude difference D i ( k ) and zero-crossing rate , respectively represented as:

[0011] ;

[0012] ;

[0013] ;

[0014] ;

[0015] in, N For the first i Frame length; k This is the delay amount; sgn [ ] represents a symbolic function, when x When ≥0, sgn [ x ]=1; when x When <0, sgn [ x ]=-1;

[0016] Spectral centroid FC i Sub-band energy ratio SER ij Fundamental period T 0, respectively, represents:

[0017] ;

[0018] ;

[0019] ;

[0020] in, For the first i The Fourier transform result of the frame signal, For the first j The upper boundary frequency of each sub-band For the first j The lower boundary frequency of each subband li and h i The first i The minimum and maximum frequencies in the subband after Fourier transform of the frame signal;

[0021] Based on the alignment relationship of the feature time dimension, the time domain and frequency domain features of the audio signal are spliced ​​together to generate a multi-domain feature spectrogram.

[0022] Furthermore, the extraction process of the Mel-frequency cepstral coefficients (MFCCs) specifically includes the following steps:

[0023] Perform a Fast Fourier Transform on each frame of audio signal to obtain Calculate the spectral energy of each frame of audio signal. :

[0024] ;

[0025] ;

[0026] Will The logarithmic energy of the filter bank output can be obtained using the Mel filter bank. :

[0027] ;

[0028] in, For the Mel filter bank, the first m The response of each filter, M The number of filters;

[0029] right Perform Discrete Cosine Transform (DCT) to obtain the signal. R MFCC coefficients of order :

[0030] ;

[0031] in, For the first i The first frame of the audio signal MFCC coefficients; the first 13 MFCC coefficients with a maximum Mel frequency of 8000Hz were selected as parameters for acoustic image feature extraction.

[0032] Furthermore, the Gaussian pyramid comprises multiple image groups of different resolutions, and each image group contains several layers; wherein the original image is enlarged by a factor of 1 to form the first layer of the first group of the Gaussian pyramid, and the first layer of the first group is convolved with a Gaussian convolution function to form the second layer of the first group of the pyramid, wherein the Gaussian convolution function is:

[0033] ;

[0034] The smoothing factor is multiplied by a proportional coefficient λ, and a new smoothing factor is obtained σ = λ σ σ ; the first group of the second layer image is smoothed, and the result image is taken as the third layer; in this way, finally, L layer images are obtained; wherein the size of each layer image in the same group is consistent, and different smoothing coefficients are used, and the corresponding smoothing coefficients are respectively: 0, σ , λ σ , λ^2 σ , λ^3 σ … λ^( L -2) σ .

[0035] Further, the adaptive convolution kernel weight is represented as:

[0036] ;

[0037] ;

[0038] Wherein, is the attention weight, Q represents the number of parallel linear functions , is the activation function, g is the bias, b is the network layer output; the constraint condition of the attention weight y is:

[0039] .

[0040] The application also includes a mechanical equipment voiceprint feature two-dimensional adaptive extraction system, comprising:

[0041] The acquisition module is used for acquiring audio signals of mechanical equipment under variable working conditions, and extracting multi-domain feature spectrograms MDS, Mel spectrum and mel frequency cepstrum coefficient MFCC of the audio signals, and constructing a space-spectrogram feature set based on MDS, Mel spectrum and MFCC; the extraction process of the multi-domain feature spectrogram of the audio signal comprises the following steps: assuming that the audio signal is , the first i frame audio signal obtained after the audio signal is preprocessed is , and then the time domain feature is obtained; wherein the time domain feature includes short-time energy E n ( i ), short-time amplitude M ( i ), and short-time amplitude difference D ​​i k ) and zero-crossing rate ; calculating frequency domain features of the audio signal; the frequency domain features include spectral centroid FC i , sub-band energy ratio SER ij , pitch period T 0; splicing the time domain and frequency domain features of the audio signal based on an alignment relationship of a feature time dimension, to generate a multi-domain feature spectrogram;

[0042] The feature extraction module is configured to construct a Gaussian pyramid by performing multiple Gaussian smoothing processing and down-sampling on the spatial-spectrogram feature set; generate each difference image layer by layer and group by group by performing pixel-level difference operation on multiple groups of adjacent two layers of images in the Gaussian pyramid; form a Gaussian difference pyramid based on all the difference images; perform extreme value detection on each granularity space of the difference images in the Gaussian difference pyramid to find extreme points of the Gaussian difference function and generate multi-granularity voiceprint feature maps; and extract saliency features of the multi-granularity voiceprint feature maps by introducing a multi-head attention mechanism.

[0043] The adaptive adjustment module is configured to dynamically generate adaptive convolution kernel weights by using a dynamic convolution adaptive extraction module, to adaptively adjust the extracted feature maps and obtain voiceprint features of the mechanical equipment under variable working conditions.

[0044] The present application also includes a computer device for double-dimension adaptive extraction of mechanical equipment voiceprint features, which comprises a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the double-dimension adaptive extraction method of mechanical equipment voiceprint features when executing the computer program.

[0045] The present application also includes a readable storage medium storing a computer program, and the computer program comprises program instructions which are executed by a processor to perform the steps of the double-dimension adaptive extraction method of mechanical equipment voiceprint features.

[0046] The present application provides a double-dimension adaptive extraction method, system and device for mechanical equipment voiceprint features, which has the following advantages:

[0047] ​The present application aims at the domain offset problem existing in the variable working condition state, constructs a multi-domain feature vector and carries out significant feature reorganization, generates a multi-domain feature spectrogram; combining the nonlinear mapping advantages of the high and low frequency characteristics of the Mel spectrum, a space-spectrogram feature set based on MDS, Mel and MFCC is constructed, the voiceprint feature is enhanced, and the voiceprint feature representation ability of the low frequency signal is improved; a multi-granularity feature representation method of the space-spectrogram feature set based on the Gaussian split pyramid model is proposed by comprehensively considering the characteristics of the geometric information and semantic information in the graph information, the different granularity information of the space-spectrogram feature set based on MDS, Mel and MFCC is sampled, multi-granularity features are combined to realize high and low dimensional feature fusion, and the global voiceprint information extraction ability is improved; and the multi-head attention mechanism is introduced to extract multi-granularity information features, and the importance between the multi-granularity information is dynamically learned; meanwhile, a mechanical equipment voiceprint feature adaptive extraction model of all-around dynamic network is proposed, the dynamic adjustment of the depth and breadth of the internal convolution kernel and the set parameter weight is realized, the domain related features are found, the working condition influence is weakened, the voiceprint common features under the variable speed and variable load working conditions are extracted, and therefore the fault detection precision of the mechanical equipment is improved. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The mechanical equipment voiceprint feature double-dimension adaptive extraction block diagram in the embodiment of the present application;

[0049] Figure 2 The audio signal preprocessing flowchart in the embodiment of the present application;

[0050] Figure 3 The multi-domain feature spectrogram MDS construction flowchart in the embodiment of the present application;

[0051] Figure 4 The spectrogram multi-granularity feature extraction framework diagram in the embodiment of the present application;

[0052] Figure 5 The multi-head attention mechanism structure diagram in the embodiment of the present application;

[0053] Figure 6 The dynamic convolution layer basic architecture diagram in the embodiment of the present application;

[0054] Figure 7 The full-dimension dynamic convolution kernel adaptive structure diagram in the embodiment of the present application;

[0055] Figure 8 The sound image feature set diagram constructed under different rotating speeds and load states in the embodiment of the present application;

[0056] Figure 9 The multi-granularity feature diagram of the mel spectrogram under different fault forms in the embodiment of the present application;

[0057] Figure 10 Loss and accuracy curves in the four model training processes in the embodiment of the present application are shown in the following figures;

[0058] Figure 11 The box plot of the variable working condition recognition rate in the embodiment of the present application is shown in the following figure;

[0059] Figure 12 The confusion matrix result under different test working conditions in the embodiment of the present application is shown in the following figure;

[0060] Figure 13 The model training process result under different feature extraction in the embodiment of the present application is shown in the following figure;

[0061] Figure 14 The t-SNE feature visualization in the embodiment of the present application is shown in the following figure;

[0062] Figure 15 The layer-by-layer network feature visualization of MS-MGFAE under variable working conditions in the embodiment of the present application is shown in the following figure;

[0063] Figure 16 The mechanical transmission system fault diagnosis test platform in the embodiment of the present application is shown in the following figure;

[0064] Figure 17 The multi-fault mode in the embodiment of the present application is shown in the following figure;

[0065] Figure 18 The layer-by-layer feature visualization in the embodiment of the present application is shown in the following figure;

[0066] Figure 19 The mechanical equipment acoustic feature two-dimensional adaptive extraction flowchart in the embodiment of the present application is shown in the following figure. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments.

[0068] The present application proposes a mechanical equipment acoustic feature two-dimensional adaptive extraction method, as shown in Figure 1 、 Figure 19 , which contains spatial-spectrogram feature set construction of audio signals, multi-sound map multi-granularity feature adaptive extraction, and all-around dynamic convolution network acoustic feature adaptive extraction. The method specifically includes the following steps:

[0069] S1, constructing a spatial-spectrogram feature set of an audio signal: a time-domain and frequency-domain feature fusion method based on time dimension alignment is proposed, a multi-domain feature spectrogram feature (voiceprint multi-domain feature time-frequency graph) is constructed, combining the nonlinear mapping advantages of Mel time-frequency features in high and low frequencies, MDS is constructed to extract Mel spectrum voiceprint features and mel frequency cepstral coefficients (MFCCs), and the voiceprint feature representation ability of low frequency signals is improved.

[0070] S1.1, constructing a spectrogram feature based on multi-domain features.

[0071] Before performing audio feature extraction, in order to suppress the time-varying characteristics, spectral tilt and spectral leakage phenomenon of the audio signal, improve the quality of the audio signal and improve the signal features, each frame signal is preprocessed, including pre-emphasis, framing and windowing processing, as shown in Figure 2 .

[0072] (1) Time-domain features: including short-time energy, short-time amplitude, short-time amplitude difference and zero-crossing rate, etc., representing the energy distribution, amplitude change and frequency approximation of the audio signal in the time dimension and other time sequence features. The time-domain feature matrix based on time-value function analysis selected by the present application can describe the basic form and evolution law of the signal, and by extracting these time sequence features, the state information of each time period of the signal can be obtained. Assuming that the audio signal sequence is , the i frame audio signal obtained after preprocessing is , and the time-domain features and calculation formula are shown in Table 1.

[0073] Table 1: Voiceprint time-domain feature calculation

[0074]

[0075] (2) Frequency domain features: the commonly used frequency domain features of audio signals are spectral centroid, sub-band energy ratio and pitch period T 0, representing the distribution of timbre characteristics, energy distribution features, different pitch variation patterns of the audio signal in high and low frequencies, reflecting the frequency domain features of the excitation source of the audio signal, etc., as shown in Table 2.

[0076] Table 2: Voiceprint frequency domain feature calculation

[0077]

[0078] The time-frequency information feature extraction of the sound signal combines the information features of the signal time domain and frequency domain, and ignores the signal structure information reflecting the time-frequency correlation between the samples. The time domain and frequency domain feature fusion method based on time dimension alignment is proposed. The alignment relationship of various types of features in the time dimension is ensured, that is, the same number of frames are obtained, and the time resolution matching relationship is achieved. Based on this, the extracted time-frequency features are spliced in the feature dimension. Specifically, the various types of features are designed and calculated in the number of time frames and the number of data features, the data is normalized, and the data features are spliced to obtain a new feature array, and a multi-domain feature spectrogram is generated to represent the time-frequency domain features, and the specific process is as shown in Figure 3 .

[0079] S1.2, Mel and MFCC-based acoustic image data feature construction.

[0080] MDS combines the time domain and frequency domain features of the audio signal in the time dimension, effectively ensuring the robustness of the audio features, but fails to extract the saliency of the audio features from the high and low frequency dimensions. The Mel frequency representation simulates the nonlinear perception of the human ear to different frequency speech, which is reflected in that the Mel filter bank is more dense near the low frequency, and the filter is more sparse near the high frequency, and more attention is paid to the resolution of the low frequency signal, so as to realize the difference representation of the high and low frequency spectral resolution features, and therefore has a significant advantage in low frequency signal voiceprint feature extraction, and is commonly used in basic analysis of audio signals, such as audio feature extraction, speech recognition, voiceprint recognition, etc.

[0081] Based on this, the saliency feature extraction method of the mechanical equipment high and low frequency audio signal based on Mel and MFCCs is proposed to fuse different types of feature information and enhance the performance of the model.

[0082] The traditional spectrogram is usually obtained by performing short-time Fourier transform (STFT) on the audio signal, and each time window contains information of the entire frequency range. The frequency axis is represented by linear frequency scale, which is constant in the entire frequency range, that is, the same resolution is obtained at high and low frequencies, and the frequencies are displayed at equal intervals. However, the human perception of frequency is not linear, and the perception of low frequency signals is more sensitive than that of high frequency signals. The relationship between the Mel scale and the frequency of the signal frequency f is as follows.

[0083] (1);

[0084] wherein, F mel is the perceived frequency in Mel, f is the actual frequency in Hz.

[0085] The preprocessed each frame of audio signal is subjected to fast Fourier transform to obtain Calculate the spectral energy of each frame of audio signal. :

[0086] (2);

[0087] (3);

[0088] Will The Mel filter bank is used, and the logarithmic energy of the filter bank output is calculated. This enables nonlinear Mel mapping in the frequency dimension, enhancing the characterization of low-frequency signals.

[0089] (4);

[0090] in, For the Mel filter bank, the first m The response of each filter, M This represents the number of filters.

[0091] right By performing a Discrete Cosine Transform (DCT), the signal can be obtained. R MFCC coefficients of order As shown in the following formula:

[0092] (5);

[0093] In the formula For the first i The first frame of the audio signal MFCC coefficient.

[0094] The frequency distribution range of audio signals of most mechanical equipment is between a few hertz and several thousand hertz. The amount of information carried by higher-order coefficients can be ignored. This invention intends to select the first 13 MFCC coefficients with a maximum Mel frequency of 8000 Hz as the parameters for sound image feature extraction.

[0095] S2, Multi-voice image multi-granularity feature adaptive extraction, such as Figure 4 As shown.

[0096] Feature extraction aims to extract and select acoustic or linguistic features that possess strong separability and stability in the voiceprint of a specific object. In feature extraction algorithms, deep networks have large receptive fields and strong semantic information representation capabilities, but their feature maps have low resolution, weak geometric information representation capabilities, and lack spatial geometric detail. Low-level networks have smaller receptive fields, strong geometric detail representation capabilities, high resolution, and strong localization features, but weak semantic information representation capabilities. Therefore, feature extraction at the single-image granularity often misses some feature information.

[0097] The present application comprehensively considers the multi-granularity information existing in image features, and proposes a multi-domain multi-granularity voiceprint feature extraction method based on Gaussian difference pyramid (Gaussian Difference Pyramid) and multi-head attention mechanism fusion. Different granularities of feature maps (spatial-spectrogram feature sets based on MDS, Mel and MFCC) are sampled, and high and low dimensional feature fusion is realized by combining multi-granularity features, which improves the global voiceprint information extraction capability.

[0098] S2.1, multi-granularity voiceprint feature representation method based on Gaussian difference pyramid.

[0099] The Gaussian difference pyramid is a kind of image pyramid, which is constructed by calculating the difference between the Gaussian filtered images at different layers of the Gaussian pyramid. First, the constructed spectrogram feature map is processed by multiple Gaussian smoothing and down-sampling to construct a Gaussian pyramid. The Gaussian pyramid is composed of a series of images with different resolutions (Octave), and each group of pyramid contains several layers (Interval). The original image is enlarged by one time and used as the first layer of the first group of Gaussian pyramid, and the first layer image of the first group is used as the second layer of the first group of Gaussian pyramid after Gaussian convolution (which is actually Gaussian smoothing or Gaussian filtering). The Gaussian convolution function is:

[0100] (6)

[0101] The σ is multiplied by a scale factor λ, then the new smoothing factor σ =λ* σ , the result image is used as the third layer. In this way, L layers of images are obtained, and the image size of each layer in the same group is consistent, but different smoothing coefficients are used. Their corresponding smoothing coefficients are: 0, σ , λ σ , λ^2 σ , λ^3 σ ……λ^(L-2) σ Each image is obtained by Gaussian filtering and down-sampling of the previous level image. The important information of the image at different granularities is retained.

[0102] The pixel-level difference operation is performed on two adjacent layers of images, that is, the first layer of the first group is obtained by subtracting the first layer of the first group from the second layer of the first group of the Gaussian pyramid. In this way, each difference image is generated layer by layer and group by group, and all the difference images constitute a Gaussian difference pyramid. The construction process of the Gaussian difference pyramid enables the detailed information of the image at different granularities to be fully extracted, providing more comprehensive and diverse feature information. On the basis of constructing the Gaussian difference pyramid, extreme value detection is performed on each granularity space to find the extreme points of the DOG function, and a significant voiceprint feature map under multiple granularity spaces is generated.

[0103] On this basis, a multi-granularity voiceprint feature fusion method based on a multi-head attention mechanism is proposed: the multi-head attention mechanism is a variant of the self-attention mechanism, aiming to enhance the expression ability and generalization ability of the model. The multi-granularity mechanical equipment voiceprint image features are divided into multiple subspaces, and multiple independent attention heads are used to calculate the attention weights, and the results are spliced or weighted summed, thereby enhancing the generalization and robustness of the model. The structure is as shown in Figure 5 First, it calculates attention in four different attention heads. X is multiplied by the weight matrix. Then, the obtained [Q, K, V] matrix is used to calculate attention. Finally, the obtained Z matrix is multiplied by the weight matrix to produce the output of the layer. The multi-head attention mechanism structure is as shown in Figure 2 .

[0104] S3, feature adaptive extraction of omnidirectional dynamic convolution.

[0105] Variable working conditions are reflected in speed changes, load changes and noise environments, which are more complex in actual situations. However, for feature extraction, variable working conditions are consistent in different speed and load spaces, and the feature spaces are different but have correlations. The conventional shared convolution kernel cannot effectively extract unique features combined with input features. The present application establishes a dynamic convolution adaptive extraction method (DCAE) to discover domain-related features, largely eliminate the influence of working conditions, and realize adaptive extraction of equipment common voiceprint features in a domain shift state.

[0106] The dynamic convolution basic architecture is as shown in Figure 6 , combined with attention weights Dynamic aggregation Q of a plurality of parallel linear functions , and the weights are adaptively adjusted according to the input data to complete linear weighted calculation, that is, the convolution kernel is related to the input, so that the dynamic convolution has input dependence. The weight calculation is as shown below:

[0107] (7)

[0108] (8)

[0109] wherein g is an activation function, y is a network layer output, is a linear function, W denotes a weight, b denotes a bias.

[0110] Attention weight constraint in dynamic convolution The constraint condition of the attention weight is shown in formula 9:

[0111] (9)

[0112] Since the multi-sound map space dimension and the multi-channel information contain the soundprint visual features under different working conditions, they have strong correlation with the input. Therefore, the application proposes multi-sound map features for different rotating speeds and load working conditions, builds a convolution kernel space and channel dimension attention dynamic interaction learning module, and realizes the depth and breadth dynamic adaptation of the convolution kernel. The full-dimensional dynamic convolution kernel adaptive structure is shown in Figure 7, which is embodied as a convolution kernel attention set on the spatial dimension, and a channel dimension attention set . The dynamic generation of adaptive convolution kernel weight set extracts common visual information and realizes the adaptive adjustment of the feature map.

[0113] This mechanism can mine more time series and spatial changes of data, so that the network can capture different features at different positions, and through the dynamic mechanism, the weights of the neural network have sample adaptability, which shows great potential in improving the generalization of the model. The representation ability of the model facing complex scenes and dynamic changes is improved.

[0114] Experimental analysis:

[0115] As a key component of mechanical equipment, the health status of bearings directly affects the safe and reliable operation of mechanical equipment. Comparative experiments are carried out on the mechanical equipment fault diagnosis data set to verify the soundprint feature extraction ability of the method proposed in the application.

[0116] The soundprint features under different fault modes are extracted based on the rotor comprehensive fault simulation experiment table data set. The load torque range of the rotor experiment table is 0-6 N·m, the bearing type used is NSK 6200, the audio data in the data set is collected by a pickup, the sampling rate is 48000 Hz, and the length of each sample is 4 seconds. The data label types include normal, bearing outer ring pitting, bearing inner ring pitting, bearing rolling element pitting, and bearing retainer fracture, and the label information is shown in Table 3. Each label type contains 16 different rotating speeds and load working conditions, and the simulation working condition information is shown in Table 4. Table 4a is a multi-speed condition state description, Table 4b is a multi-load condition state description, and Table 4c is a sample number of the multi-condition state of the data set. 19200 sound learning samples are generated, as shown in Table 3.

[0117] Table 3 Data label type description

[0118]

[0119] Table 4 Multi-condition information description

[0120] a Multi-speed condition state description

[0121]

[0122] b Multi-load condition state description

[0123]

[0124] c Sample number of multi-condition state of data set

[0125]

[0126] Multi-sound image sample set construction:

[0127] The time-frequency domain features of the audio signal are extracted, and the MDS feature map is constructed. The shape of the extracted time-frequency domain soundprint feature array is shown in Table 5, and the MDS soundprint feature time-frequency map is built as shown in Figure 5 .

[0128] Table 5 Shape of time-frequency domain soundprint feature array

[0129]

[0130] The soundprint features of the training data set are processed to realize time-frequency image visualization, analyze the soundprint features of the audio signal, and the audio signal under multiple variable working conditions is processed to construct a sound image feature set. The maximum frequency is 8000 Hz.

[0131] The sound image feature samples of the audio signal under five health conditions are as follows Figure 8As shown, the Mel feature representation of the audio signal under different fault modes is obviously different. Among them, the sound pattern samples under S2_L1, S3_L1 and S5_L1 multi-speed conditions are compared, the intra-class similarity is high, and the inter-class difference is obvious. The data is comparable under different working condition modes, the intra-class cohesion and the separability between classes are realized. The multi-granularity information of the multi-sound pattern sample is extracted to fuse the sound signal representation under different multi-granularity information and the multi-granularity features of the mel spectrum under different fault forms such as Figure 9 As shown.

[0132] Experiment 1: Model feature extraction ability experiment under variable working condition

[0133] Combined with the working condition information in Table 5, the model extraction ability under variable speed and load is verified. The variable speed working condition test design is shown in Table 6, wherein ST1 is the basic test task and ST2, ST3, ST4 and ST5 are variable speed tasks. ST1 divides the sample set into a training set, a verification set and a test set, and the division ratio is 0.7, 0.15 and 0.15, wherein the test set does not participate in model training and verification. The variable speed task selects a speed working condition as the test set, and the other speed working conditions are divided into the test set and the verification set according to the ratio of 7:3. The variable speed working condition test and the variable speed design principle are consistent, as shown in Table 7. Combined with three mainstream audio feature extraction methods, the data set analysis comparison experiment under different speed conditions is completed. The comparison model parameter information is shown in Table 8, and the CWT+CNN, MFCC+PMFCNN and PWPE+ECAPA-TDNN models are compared to verify the recognition performance of the method.

[0134] Table 6 Data set verification design under different speed conditions

[0135]

[0136] Table 7 Data set verification design under different load working condition

[0137]

[0138] Table 8 Comparison method of sound feature extraction model

[0139]

[0140] The sample set is processed in batches, and the batch size is 150. In order to prevent overfitting, dropout is added to the Dense layer of the network here, and the value is set to 0.5, and the sample label space is mapped by combining the softmax classifier. Select sparse class cross entropy as the loss function, and Adam algorithm as the optimizer. The learning rate is 10 -3epoch is 50 and batchsize is 64.

[0141] Experiment 2: Verification of multi-speex multi-granularity feature expression ability:

[0142] An experiment of voiceprint feature expression ability under multi-search mode is designed to comprehensively evaluate the feature expression ability of the MSODDAE. As shown in Table 9, seven feature extraction methods are constructed by using MSFE, MGAE and DCAE and other combinations. Among them, MSMGAE is to adaptively extract multi-granularity information on the basis of MSFE. Under the condition of consistent network structure and hyperparameters, 19200 multi-condition data samples are processed. The sample set is divided into a training set, a validation set and a test set, with a proportion of 0.7, 0.15 and 0.15 respectively. The test set does not participate in model training and validation.

[0143] Table 9 Experiment of voiceprint feature expression ability under multi-search mode

[0144]

[0145] Experiment 3: Deep network feature visualization experiment under variable working condition mode:

[0146] The current deep learning model has the problem of poor explainability of feature extraction process. In order to reveal the representation ability of the model, the hierarchical feature of the method proposed in the application is visualized by combining the Tsne method.

[0147] Analysis of experimental results:

[0148] (1) Experimental results of model feature extraction ability under variable working condition mode:

[0149] ① Comparison results of model training stability: The superior performance of MSODDAE is verified under ST2, and the stability of the training process of the four methods shown in Table 9 is compared comprehensively. The standard deviation (SD) evaluation standard is used to determine the standard error of each iteration in the training process. The smaller the SD value, the better the model stability. The recognition loss and accuracy curves in the model training process are shown in Figure 10 The shaded area in the figure represents the standard error band.

[0150] As shown in Figure 10As shown, the loss value loss and the recognition accuracy Accuacy in the model training process gradually converge and stabilize with the increase of the iteration round number. The recognition rates of the four model training sets are steadily improved and reach more than 90%. It is worth noting that the recognition rates of PWPE_ECAPA_TDNN and MSODDAE reach 98%, and the voiceprint feature extraction capability is better than that of the other two models. In combination with the error band and the convergence condition, the MSODDAE has obvious advantages in the convergence speed and stability in the training stage.

[0151] ②Model test effectiveness comparison experiment results: further analysis of the voiceprint feature extraction performance in the model test stage under variable speed and variable load conditions, the present application selects 10 test set results for scheme comparison and analysis to avoid accidental errors in the model test stage. For example, Figure 11 is a box plot of the variable condition recognition rate, (a) represents a box plot under different speed conditions, and (b) represents a box plot under different load conditions; under variable conditions ST2, ST5, LT2 and LT5, the recognition accuracy of the MSODAE proposed by the present application is the highest and the recognition error is the smallest. The recognition stability under the variable speed condition is better than that under the variable load condition, considering that the number of training set samples under the variable speed condition is greater than that under the load condition, and another reason is that the characteristics of different speed and load conditions are different. The overall recognition stability and recognition rate of the MSODDAE are better than those of the three kinds of models.

[0152] As shown in Table 11. In the basic test task ST1, the test accuracies of the four models are 94.39%, 98.83%, 99.39% and 99.61% respectively. In the variable speed tasks ST2, ST3, ST4 and ST5, due to the existence of domain offset, there is a large distribution difference between the test set samples, and the accuracy of the model decreases, wherein the recognition accuracy of ST2 is the highest, which is 98.17%±0.0042, and the minimum recognition accuracy is ST5, the recognition accuracy is 95.73±0.0100. The reason for this phenomenon is that when the device speed is higher, the non-linear characteristics of the sample and the noise pollution are more serious. The average accuracy of the MSODDAE model proposed by the present application under the variable speed condition is 96.92%±0.0090, and good diagnostic effect is obtained. On the other hand, as shown in Table 12, the maximum accuracy of the MSODDAE under the variable load condition is 97.48%±0.0096, and the average accuracy is 96.06%±0.0106. The results show that the model proposed by the present application has the highest accuracy and stability in the eight variable condition experiment tasks, and Tables 10 and 11 show the best metrics of the proposed model, which proves the excellent performance of the model in the variable condition fault diagnosis.

[0153] Table 10 Test set accuracy of different models under variable speed conditions

[0154]

[0155] Table 11 Accuracy of different model test sets under variable load conditions

[0156]

[0157] To better confirm the performance of the proposed method for voiceprint feature extraction under different operating conditions. The confusion matrix is used to intuitively represent the accuracy of model classification. Dark squares represent accurately predicted samples, while light squares represent incorrectly predicted samples. Figure 12 Figures (a), (b), (c) and (d) in the confusion matrix results of different test conditions. And it performs well under variable speed and variable load conditions.

[0158] Feature expression ability under multiple search modes: The excellent performance of MSODDAE under multiple conditions is verified, and the stability of the training process of the seven feature extraction methods shown in Table 12 is comprehensively compared. The standard deviation of each eopch in the training process is solved through the standard deviation evaluation criterion. As shown in Figure 13 When the network structure and hyperparameters are consistent, the proposed MSODDAE has obvious advantages in convergence speed, accuracy and stability in the retraining stage.

[0159] Table 4 shows a further comparative analysis of the results of 10 tests. The results show that the recognition rate of OASFE is 61.12%, and the recognition rate of MSODDAE is 98.93%, which is improved by 37.81%, and the recognition efficiency is obviously improved. On the other hand, compared with MFFE, its recognition efficiency is slightly reduced, but the recognition accuracy is improved by 15.10%. It shows that under the condition of small time difference, the accuracy of the model is significantly improved, which can better capture and distinguish subtle features, although it may sacrifice certain processing speed.

[0160] Combined with the T-distributed Stochastic Neighbor Embedding (T-SNE) method, the model test set results are visualized. The multi-dimensional tensor extracted by the model in the FC layer of the model is used as the sample feature vector, and is flattened into a 2-dimensional tensor, as shown in Figure 14 The results clearly show the high classification accuracy of the proposed model. The distribution between different fault categories shows obvious separation and high category discrimination, which confirms that the proposed model has strong feature extraction ability.

[0161] Table 12 Accuracy of different model validation sets

[0162]

[0163] Feature extraction performance as Figure 15As shown, after t-SNE processing, the red circle areas 1, 2 and 4 exist in the voiceprint feature interlaced area, the areas 3, 5 and 6 have poor clustering effect among the same type of voiceprint features, which leads to a relatively large potential error possibility of fault diagnosis, the fault feature obtained by the MS-MGFAE method has clear boundary, optimal intra-class aggregation effect and minimum diagnosis error, which shows that the MS-MGFAE method can effectively improve the voiceprint feature representation ability.

[0164] The multi-spectrogram information fusion method has advantages in model training stability, test recognition stability and feature separation effectiveness.

[0165] Deep network feature visualization test under variable working condition mode: T-distribution stochastic neighbor embedding (T-SNE) takes the multi-dimensional tensor output by the intermediate layer of the model as a feature vector, and visualizes the feature extraction process of the model. The voiceprint feature interlaced area in network layers 0, 1, 3 and 4 is shown as color overlapping in the red area 1245. The voiceprint feature clustering effect in network layers 2 and 5 is poor, which is shown as color feature dispersion in the red areas 3 and 6. With the continuous deepening of the model, the clustering effect is obvious, and the feature label boundary of each class is clear.

[0166] To verify the effectiveness of the method, a mechanical transmission system fault diagnosis comprehensive experiment platform is built, which mainly consists of an alternating current motor, a gear reducer, a magnetic powder brake, a coupling and various sensors. The real signals during the operation of the reducer are collected by installing speed, sound, torque and other sensors, and the specific types and installation positions of the sensors are shown in Figure 16 The motor used in the experiment is QABP100L4A, the data acquisition card is MCC USB-1608FS-Plus, the torque and speed sensor is HCNJ-101, the sampling frequency is 1kHz, the sound sensor is AWA14423, and the sampling frequency is 20kHz. The experiment realizes the load change of the magnetic powder brake by adjusting the programmable acquisition card, and realizes the multi-working condition data acquisition of the reducer by changing the motor speed controlled by the frequency converter. The fault types are inner ring fault, outer ring fault, rolling body fault, retainer fault and full fault, as shown in Figure 17 The table 13 shows the type of each data label.

[0167] Table 13 Type of data label

[0168]

[0169] Table 14 Multi-working condition information description

[0170] a Multi-speed working condition description

[0171]

[0172] b Multi-load working condition state description

[0173]

[0174] c Number of multi-working condition state samples of the data set

[0175]

[0176] Table 15: Design of data set under variable working conditions

[0177]

[0178] The voiceprint feature extraction capability of the model under different rotating speeds and load conditions is verified. The variable speed working condition test design is shown in Table 14. The model test experiment is carried out under T1, T2, T3 and T4 working conditions, and the test results are shown in Table 15. The recognition accuracy under variable rotating speed T2 and T3 working conditions is 95.25%±0.0082, and the recognition accuracy under variable load working condition T4 is 96.17%±0.0067, as shown in Table 16.

[0179] Table 16: Accuracy of verification set of different models

[0180]

[0181] Through experimental verification on the device sound data set under different fault forms in the laboratory state, the deep network features are visually analyzed as shown in Table 17, and the excellent performance of the model in the variable working condition voiceprint feature extraction is verified. Figure 18

[0182] The voiceprint feature double-dimension adaptive extraction method considering the dynamic relationship between multi-sound map information features and convolution parameter variables in multiple variable working conditions is proposed. The training stability, feature separation effectiveness and feasibility of domain feature learning of the feature extraction method are verified on the data set of the Xunfei mechanical equipment challenge, and the average accuracy under variable rotating speed is 96.06%±0.0106, and the average accuracy under variable load working condition is 96.92%±0.0090. The applicability of the model is verified by building a bench test, and the voiceprint feature extraction method proposed in the application has outstanding advantages under multiple unknown rotating speed and load conditions by comparing the model test verification.

[0183] Based on the same inventive concept, the application also proposes a mechanical equipment voiceprint feature double-dimension adaptive extraction system, which comprises:

[0184] ​The collection module is configured to collect audio signals of the mechanical equipment under the variable working condition, extract a multi-domain feature spectrogram MDS, a Mel spectrum and a mel frequency cepstral coefficient MFCC of the audio signals, and construct a space-spectrogram feature set based on the MDS, the Mel spectrum and the MFCC.

[0185] The feature extraction module is configured to construct a Gaussian pyramid by performing multiple Gaussian smoothing processing and down-sampling on the space-spectrogram feature set, generate each difference image by performing pixel-level difference operation on multiple groups of adjacent two-layer images in the Gaussian pyramid, and construct a Gaussian difference pyramid based on all the difference images. The feature extraction module is further configured to perform extreme value detection on each granularity space of the difference images in the Gaussian difference pyramid, find extreme points of the Gaussian difference function, and generate a multi-granularity voiceprint feature map. The feature extraction module is further configured to extract saliency features of the multi-granularity voiceprint feature map by introducing a multi-head attention mechanism.

[0186] The adaptive adjustment module is configured to dynamically generate adaptive convolution kernel weights by using a dynamic convolution adaptive extraction module, adaptively adjust the extracted feature map, and obtain the voiceprint features of the mechanical equipment under the variable working condition.

[0187] Based on the same inventive concept, the present application further provides a computer device for double-dimension adaptive extraction of mechanical equipment voiceprint features, which comprises a memory, a processor and a computer program stored in the memory, and the processor implements the steps of the double-dimension adaptive extraction method of mechanical equipment voiceprint features when executing the computer program.

[0188] Based on the same inventive concept, the present application further provides a readable storage medium, which stores a computer program, and the computer program comprises program instructions. When the program instructions are executed by a processor, the steps of the double-dimension adaptive extraction method of mechanical equipment voiceprint features are executed.

[0189] The above description is only a preferred embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical solution and inventive concept of the present application within the technical scope disclosed by the present application, and all of them should be covered within the protection scope of the present application.

Claims

1. A two-dimensional adaptive extraction method for voiceprint features of mechanical equipment, characterized in that, Includes the following steps: Audio signals from mechanical equipment under varying operating conditions are collected, and the multi-domain feature spectrogram (MDS), Mel spectral features, and Mel cepstral coefficients (MFCC) of the audio signals are extracted to construct a spatial-spectral feature set based on MDS, Mel spectrum, and MFCC. The extraction process of the multi-domain feature spectrogram of the audio signal includes the following steps: assuming the audio signal is... The first result obtained after preprocessing the audio signal i The frame audio signal is This leads to the acquisition of time-domain features; whereby time-domain features include short-time energy. E n ( i ), short-term amplitude M ( i ), short-term amplitude difference D i ( k ) and zero-crossing rate ; Calculate the frequency domain characteristics of the audio signal; frequency domain characteristics include the spectral centroid. FC i Sub-band energy ratio SER ij Fundamental period T 0; Based on the alignment relationship of the feature time dimension, the time domain and frequency domain features of the audio signal are spliced ​​together to generate a multi-domain feature spectrogram; A Gaussian pyramid is constructed by performing multiple Gaussian smoothing and downsampling processes on the spatial-spectral image feature set. Pixel-level differencing is then performed on multiple adjacent pairs of images within the Gaussian pyramid, generating a difference image layer by layer. A Gaussian difference pyramid is then constructed based on all these difference images. Extrema detection is performed on each granularity space of the difference images within the Gaussian difference pyramid to find the extreme points of the Gaussian difference function, generating a multi-granularity voiceprint feature map. A multi-head attention mechanism is then introduced to extract the salient features of the multi-granularity voiceprint feature map. A dynamic convolutional adaptive extraction module capable of generating adaptive convolutional kernel weights is used to adaptively adjust the extracted feature maps, thereby obtaining the acoustic features of mechanical equipment under varying working conditions.

2. The method for dual-dimensional adaptive extraction of voiceprint features of mechanical equipment according to claim 1, characterized in that, The short-time energy E n ( i ), short-term amplitude M ( i ), short-term amplitude difference D i ( k ) and zero-crossing rate , respectively represented as: ； ； ； ； in, N For the first i Frame length; k This is the delay amount; sgn [ ] represents a symbolic function, when x When ≥0, sgn [ x ]=1; when x When <0, sgn [ x ]=-1; Spectral centroid FC i Sub-band energy ratio SER ij Fundamental period T 0, respectively, represents: ； ； ； in, For the first i The Fourier transform result of the frame signal, For the first j The upper boundary frequency of each sub-band For the first j The lower boundary frequency of each subband l i and h i The first i The minimum and maximum frequencies in the subband after Fourier transform of the frame signal; Based on the alignment relationship of the feature time dimension, the time domain and frequency domain features of the audio signal are spliced ​​together to generate a multi-domain feature spectrogram.

3. The method for dual-dimensional adaptive extraction of voiceprint features of mechanical equipment according to claim 1, characterized in that, The extraction process of the Mel-frequency cepstral coefficients (MFCCs) specifically includes the following steps: Perform a Fast Fourier Transform on each frame of audio signal to obtain Calculate the spectral energy of each frame of audio signal. : ； ； Will The logarithmic energy of the filter bank output can be obtained using the Mel filter bank. : ； in, For the Mel filter bank, the first m The response of each filter, M The number of filters; right Perform Discrete Cosine Transform (DCT) to obtain the signal. R MFCC coefficients of order : ； in, For the first i The first frame of the audio signal MFCC coefficients; the first 13 MFCC coefficients with a maximum Mel frequency of 8000Hz were selected as parameters for acoustic image feature extraction.

4. The method for dual-dimensional adaptive extraction of voiceprint features of mechanical equipment according to claim 1, characterized in that, The Gaussian pyramid comprises multiple groups of images at different resolutions, and each group contains several layers. The original image is enlarged by a factor of two to form the first layer of the first group of the Gaussian pyramid. The first layer of the first group is then convolved with a Gaussian convolution function to form the second layer of the first group of the pyramid. ； Smoothing factor σ Multiplying by a scaling factor λ yields a new smoothing factor. σ =λ* σ The second layer image of the first group is smoothed, and the resulting image is used as the third layer; this process is repeated until the Lth layer image is obtained. Within the same group, each layer image has the same size and uses a different smoothing coefficient, which are: 0, ... σ , λ σ , λ^2 σ ,λ^3 σ …λ^( L -2) σ .

5. The method for dual-dimensional adaptive extraction of voiceprint features of mechanical equipment according to claim 1, characterized in that, The adaptive convolutional kernel weights are represented as follows: ； ； in, For attention weights, Q Represents parallel linear functions Quantity, g For activation function, b Indicates bias. y The attention weights are the output of the network layer. The constraints are: 。 6. A dual-dimensional adaptive extraction system for voiceprint features of mechanical equipment, characterized in that, include: The acquisition module is used to acquire audio signals from mechanical equipment under varying operating conditions and extract the multi-domain feature spectrogram (MDS), Mel spectral features, and Mel cepstral coefficients (MFCC) of the audio signals, constructing a spatial-spectral feature set based on MDS, Mel spectrum, and MFCC. The extraction process of the multi-domain feature spectrogram of the audio signal includes the following steps: assuming the audio signal is... The first result obtained after preprocessing the audio signal i The frame audio signal is This leads to the acquisition of time-domain features; whereby time-domain features include short-time energy. E n ( i ), short-term amplitude M ( i ), short-term amplitude difference D i ( k ) and zero-crossing rate ; Calculate the frequency domain characteristics of the audio signal; frequency domain characteristics include the spectral centroid. FC i Sub-band energy ratio SER ij Fundamental period T 0; Based on the alignment relationship of the feature time dimension, the time domain and frequency domain features of the audio signal are spliced ​​together to generate a multi-domain feature spectrogram; The feature extraction module constructs a Gaussian pyramid by performing multiple Gaussian smoothing and downsampling processes on the spatial-spectral map feature set. It then generates a difference image layer by layer by performing pixel-level difference operations on multiple adjacent pairs of images within the Gaussian pyramid, constructing a Gaussian difference pyramid based on all the difference images. Finally, it performs extremum detection on each granular space of the difference images within the Gaussian difference pyramid to find the extreme points of the Gaussian difference function, generating a multi-granularity voiceprint feature map. A multi-head attention mechanism is introduced to extract the salient features of the multi-granularity voiceprint feature map. The adaptive adjustment module is used to dynamically generate adaptive convolution kernel weights using the dynamic convolution adaptive extraction module, and adaptively adjust the extracted feature map to obtain the acoustic features of mechanical equipment under varying working conditions.

7. A computer device for dual-dimensional adaptive extraction of voiceprint features from mechanical equipment, characterized in that, include: The memory, the processor, and the computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the two-dimensional adaptive extraction method for voiceprint features of mechanical equipment according to any one of claims 1-5.

8. A readable storage medium, characterized in that, The readable storage medium stores a computer program, which includes program instructions that, when executed by a processor, are used to perform the steps of the two-dimensional adaptive extraction method for voiceprint features of mechanical equipment as described in any one of claims 1-5.

Images