Cable elbow type voiceprint signal feature extraction and recognition method and system

Abstract

The invention discloses a cable elbow type voiceprint signal feature extraction and recognition method and system, and relates to the technical field of voiceprint recognition, and the method comprises the steps: receiving a cable elbow type terminal body sound signal, and extracting an intrinsic component based on an ensemble empirical mode decomposition (EEMD) algorithm in an intrinsic component threshold range, and obtaining a reconstruction signal; obtaining a key voiceprint component from the reconstructed signal based on a sound signal blind separation algorithm of sparse representation; the key voiceprint component is subjected to improved S transformation based on energy aggregation dynamic adjustment to obtain an S transformation result, and the S transformation result is subjected to feature extraction based on a wavelet packet energy-Mel frequency cepstrum coefficient to obtain an image sample set of a voiceprint feature spectrum corresponding to a specific ponding degree; and inputting the image sample set of the voiceprint characteristic spectrum corresponding to the specific ponding degree into a pre-established convolutional neural network model for training, and outputting to obtain an optimized variable parameter state recognition model based on the actual working condition, on-line accurate recognition of the water accumulation state of the cable elbow-shaped terminal is achieved through the optimized variable parameter state recognition model based on the actual working condition.

Description

Technical Field

[0001] The present invention relates to the technical field of voiceprint recognition, and in particular to a method and system for extracting and recognizing characteristics of cable elbow voiceprint signals. Background Art

[0002] Electricity is an important pillar of modern economy and life, but with the growth of urban scale and electricity demand, traditional overhead distribution networks pose challenges to urban landscape, planning and residents' safety. Therefore, underground cable laying has gradually become a trend. However, the laying of underground cables often requires the use of elbow cables to connect cables in all directions at corners. Due to the complex environment of the cable trench, when the temperature changes, water is easily accumulated at the elbow break, which may cause insulation degradation, short circuit and other faults, and maintenance is difficult and costly. In recent years, voiceprint recognition technology, as an emerging means, has shown potential in detecting and solving water accumulation at elbow ends.

[0003] The technical principle of voiceprint recognition can be divided into two main steps: feature extraction and pattern recognition. In the feature extraction stage, the voiceprint recognition system will extract and select acoustic features that have strong separability and high stability for the voiceprint. In the pattern recognition stage, the extracted voiceprint features will be used to compare with the known voiceprint features in the database to find matches. This process can be completed using different algorithms and models, such as distance-based matching algorithms, probability-based classification algorithms, neural network models, etc.

[0004] The commonly used acoustic feature extraction method currently obtains the information carried by the acoustic signal by directly extracting the characteristic parameters of the acoustic signal waveform. However, this extraction method obtains limited information and will cause the loss of useful information. In addition, for pipeline systems with more complex working conditions, the accuracy of the obtained parameters is easily affected by external environmental factors. Summary of the invention

[0005] In order to solve the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a method and system for extracting and identifying cable elbow voiceprint signal features.

[0006] In the first aspect, the purpose of the present invention can be achieved by the following technical solution: a method for extracting and identifying characteristics of cable elbow voiceprint signals, the method comprising the following steps:

[0007] Receive the acoustic signal of the cable elbow terminal, and extract the intrinsic component based on the ensemble empirical mode decomposition algorithm EEMD of the intrinsic component threshold range to obtain the reconstructed signal; use the acoustic signal blind separation algorithm based on sparse representation to obtain the key voiceprint component;

[0008] The key voiceprint components are dynamically adjusted based on the improved S transform of energy concentration to obtain the S transform result, and the S transform result is extracted based on the characteristics of wavelet packet energy-Mel frequency cepstrum coefficient to obtain an image sample set corresponding to the voiceprint feature map under a specific degree of water accumulation;

[0009] The image sample set corresponding to the voiceprint feature map under a specific water accumulation degree is input into the pre-established convolutional neural network model for training, and the optimized variable parameter state recognition model based on actual working conditions is output. The optimized variable parameter state recognition model based on actual working conditions can be used to realize online and accurate recognition of the water accumulation state of the cable elbow terminal.

[0010] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the set empirical mode decomposition algorithm process based on the eigencomponent threshold range includes the following steps:

[0011] Construct a new time series by adding the white noise time series to the original knock signal time series:

[0012] X i (t) = x(t) + w i (t),i=1,2,...,N(4.1)

[0013] Where X i (t) is the newly constructed time series, x(t) is the time series of the knocking signal under standard working conditions, and w i (t) is the white noise time series, N is the number of times different sequences of white noise are added to the original signal and the empirical mode decomposition (EMD) is performed;

[0014] For each newly generated signal X i (t), using the standard empirical mode decomposition EMD, a series of intrinsic mode functions IMFs are obtained:

[0015]

[0016] IMF ij (t) represents the jth IMF component obtained in the i-th experiment, r i (t) represents the residual of the i-th test, and M represents the number of extracted IMF components;

[0017] Repeat the iteration N times, and for each IMF position, take the average of all test results as the final IMF; the final IMF calculation formula is:

[0018]

[0019] The spectrum of the IMF component is calculated and compared with the frequency band of the cable elbow terminal body acoustic signal. The intrinsic mode functions with small frequency band intersection and representing continuous noise interference are removed, and the remaining intrinsic mode functions are superimposed to obtain the reconstructed signal.

[0020] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of obtaining the key voiceprint component by using a blind separation algorithm of an acoustic signal based on a sparse representation of the reconstructed signal is as follows:

[0021] initialization:

[0022] Given a training set Y = [y1,y2,...,y N ], where n is an m-dimensional vector;

[0023] Initialize dictionary D = [d1, d2, ..., d K ], where d k represents the kth dictionary atom, which is an m-dimensional vector;

[0024] The initial sparse coefficient matrix is ​​X = [x1, x2, ..., x N ];

[0025] Sparse Coding:

[0026] For each training sample y n , find the sparse vector x by orthogonal matching pursuit OMP n Make y n ≈Dx n ;

[0027] Dictionary updates:

[0028] Select dictionary atom d k Update, for the selected d k , calculate the residual of all samples using this atom Updated k and the corresponding sparse coefficient x k , so that Minimize, || || F represents the Frobenius norm;

[0029] Repeat the steps until the stopping condition is met, such as reaching the maximum number of iterations or the reconstruction error is less than a certain threshold; output the final dictionary D and sparse coefficient matrix X; calculate the final training set Y, reconstruct the signal, and obtain the target voiceprint.

[0030] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the dictionary update is completed by singular value decomposition SVD:

[0031] For Ek Perform SVD to get E k =U∑V T , then the new d k Take U(:,1), the new x k Taken as ∑(1,1)V(:,1) T .

[0032] In combination with the first aspect, in certain implementations of the first aspect, the method also includes: the steps of the improved S transform based on dynamic adjustment of energy concentration include: initializing data, calculating the initial S transform, calculating energy concentration, dynamically adjusting the Gaussian window width, recalculating the S transform, and iterative optimization.

[0033] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of initializing data:

[0034] Set the initial Gaussian window width σ0 and frequency range f, and define the step size Δσ;

[0035] The process of calculating the initial S-transform:

[0036] Assume a discrete time signal x[m], and use the initial Gaussian window width σ0 to calculate the S transform S x (f,n);

[0037]

[0038] Where N is the signal length and n is the time index;

[0039] The process of calculating energy concentration:

[0040] For each frequency f, calculate the S-transform result S x Energy distribution of (f,n);

[0041] E(f,n)=|S x (f,n)| 2 (4.5)

[0042] Calculate the concentration of energy distribution, using entropy or other concentration indicators:

[0043] Entropy H(f) is used to represent the concentration of energy distribution:

[0044]

[0045] in,

[0046] The process of dynamically adjusting the Gaussian window width:

[0047] Dynamically adjust the Gaussian window width σ(f) according to the energy concentration H(f); if H(f) is large, increase the Gaussian window width; if H(f) is small, reduce the Gaussian window width;

[0048] σ(f)=σ0+α(H(f)-H0)(4.7)

[0049] Among them, α is the adjustment coefficient, H0 is the baseline entropy value;

[0050] The process of recalculating the S transform:

[0051] Recalculate the S transform S using the adjusted Gaussian window width σ(f) x (f,n);

[0052]

[0053] Iterative optimization process:

[0054] Repeat the steps until the Gaussian window width converges or reaches a predetermined number of iterations;

[0055] The final S transformation result S x In combination with the first aspect, in some implementations of the first aspect, the method further includes: extracting features of the S-transform result based on wavelet packet energy-Mel-frequency cepstral coefficients to obtain an image sample set corresponding to a voiceprint feature map under a specific water accumulation degree:

[0056] Use wavelet packet decomposition to transform S into the result S x (f,n) is decomposed into multiple sub-bands, and some wavelet basis functions and decomposition layers are selected to obtain the coefficients W of multiple sub-bands. j,k (n), where j represents the decomposition layer, k represents the subband index, and n represents the time index;

[0057] Calculate the energy distribution E of each subband j,k :

[0058]

[0059] Among them, wavelet packet decomposition includes:

[0060] Assume that the signal x(n) is decomposed by wavelet packet, and the obtained subband coefficients are represented by a matrix;

[0061] First level decomposition:

[0062] Decompose the signal x[n] into a low-frequency part A1 and a high-frequency part D1;

[0063] A1=h[n]*x[n] (4.9)

[0064] D1=g[n]*x[n] (4.10)

[0065] where h[n] and g[n] are the impulse responses of the low-pass and high-pass filters, respectively;

[0066] Second level decomposition:

[0067] Decompose the low-frequency part A1 and the high-frequency part D1 again to obtain A2, D2, A1D1, and D1A1;

[0068] A2=h[n]*A1 (4.11)

[0069] D2=g[n]*A1 (4.12)

[0070] A1D1=h[n]*D1 (4.13)

[0071] D1A1=g[n]*D1 (4.14)

[0072] Until the specified number of decomposition layers is reached;

[0073] Calculate Mel-frequency cepstral coefficients MFCCs, including:

[0074] Transform S to result S x (f,n) Convert from linear frequency to Mel frequency:

[0075]

[0076] Transform S to result S x (f,n) passes through the Mel filter bank to obtain the Mel frequency spectrum M(f m ,n):

[0077]

[0078] Mel frequency spectrum M(f m ,n) Perform discrete cosine transform DCT to obtain Mel frequency cepstrum coefficients MFCCs:

[0079]

[0080] Use feature selection algorithm to evaluate the contribution of each wavelet packet energy feature and MFCC feature to the state recognition result, select features with high contribution, form a preliminary feature set, assign a weight to each feature according to the feature contribution evaluation result, and use support vector machine SVM or random forest Random Forest model to automatically learn feature weights during training:

[0081] The weighted wavelet packet energy features and MFCC features are combined to form a comprehensive feature vector F:

[0082]

[0083] Among them, w i is the weight of the i-th feature;

[0084] The comprehensive voiceprint feature vectors of each sensor are combined according to the weights to form a multidimensional feature map represented by a matrix. The multidimensional feature map is converted into an image form to form an image sample set. Each image sample corresponds to a voiceprint feature map under a specific degree of water accumulation.

[0085] In a second aspect, in order to achieve the above-mentioned purpose, the present invention discloses a cable elbow voiceprint signal feature extraction and recognition system, comprising:

[0086] The signal processing module is used to receive the acoustic signal of the cable elbow terminal body, and extract the intrinsic component based on the EEMD algorithm of the ensemble empirical mode decomposition of the intrinsic component threshold range to obtain the reconstructed signal; the reconstructed signal is based on the acoustic signal blind separation algorithm of the sparse representation to obtain the key voiceprint component;

[0087] The transformation extraction module is used to dynamically adjust the key voiceprint components based on the improved S transformation of energy concentration to obtain the S transformation result, extract the characteristics of the S transformation result based on the wavelet packet energy-Mel frequency cepstrum coefficient, and obtain an image sample set corresponding to the voiceprint feature map under a specific water accumulation degree;

[0088] The state recognition module is used to input the image sample set corresponding to the voiceprint feature map under a specific water accumulation degree into a pre-established convolutional neural network model for training, and output an optimized variable parameter state recognition model based on actual working conditions. The optimized variable parameter state recognition model based on actual working conditions is used to realize online and accurate recognition of the water accumulation state of the cable elbow terminal.

[0089] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, a cable elbow voiceprint signal feature extraction and recognition method as described above is adopted.

[0090] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a computer-readable storage medium is disclosed, in which a computer program is stored. When the computer program is loaded and executed by a processor, the cable elbow voiceprint signal feature extraction and recognition method as described above is adopted.

[0091] Beneficial effects of the present invention:

[0092] The present invention is used to realize water ingress detection of cable elbow terminal head through two parts: an aliasing voiceprint signal key feature component extraction algorithm and a complex voiceprint signal high-resolution time-frequency characterization and feature parameter extraction and recognition algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0094] Figure 1 It is a schematic flow chart of the method of the present invention;

[0095] Figure 2 It is the EEMD time domain diagram of the voiceprint sample;

[0096] Figure 3 It is the EEMD spectrum of the voiceprint sample;

[0097] Figure 4 This is a comparison chart of sample EEMD reconstruction results and K-SVD denoising results;

[0098] Figure 5 It is a diagram of the implementation scheme of the algorithm for extracting key characteristic components of aliased voiceprint signals according to the present invention;

[0099] Figure 6 It is a diagram of the implementation scheme of the present invention for realizing high-resolution time-frequency representation of complex voiceprint signals and feature parameter extraction and recognition algorithm;

[0100] Figure 7 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0101] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0102] Embodiment 1:

[0103] like Figure 1 As shown, a cable elbow voiceprint signal feature extraction and recognition method comprises the following steps:

[0104] S101: receiving the acoustic signal of the cable elbow terminal body, and extracting the intrinsic component based on the ensemble empirical mode decomposition algorithm EEMD of the intrinsic component threshold range to obtain a reconstructed signal; applying the reconstructed signal to the acoustic signal blind separation algorithm based on sparse representation to obtain the key voiceprint component;

[0105] Implement the ensemble empirical mode decomposition algorithm based on the intrinsic component threshold range, including:

[0106] Construct a new time series by adding the white noise time series to the original knock signal time series:

[0107] X i (t) = x(t) + w i (t),i=1,2,...,N(4.1)

[0108] Where X i (t) is the newly constructed time series, x(t) is the time series of the knocking signal under standard working conditions, and w i (t) is the white noise time series, and N is the number of times different sequences of white noise are added to the original signal and empirical mode decomposition (EMD) is performed.

[0109] For each newly generated signal X i (t), using standard empirical mode decomposition (EMD), a series of intrinsic mode functions (IMFs) are obtained:

[0110]

[0111] IMF ij (t) represents the jth IMF component obtained in the i-th experiment, r i (t) represents the residual of the ith test, and M represents the number of extracted IMF components.

[0112] Repeat the iteration N times, and for each IMF position, take the average of all test results as the final IMF. This can eliminate the deviation introduced by random noise while retaining the characteristics of the signal itself. The final IMF calculation formula is:

[0113]

[0114] The spectrum of the IMF component is calculated and compared with the frequency band of the cable elbow terminal body acoustic signal. The intrinsic mode functions with very small frequency band intersection representing continuous noise interference are removed, and the remaining intrinsic mode functions are superimposed to obtain the reconstructed signal to achieve denoising.

[0115] The EEMD time domain diagram and spectrum diagram of the voiceprint sample are as follows: Figure 2 , Figure 3 shown.

[0116] Implement the acoustic signal blind separation algorithm based on sparse representation, including:

[0117] (initialization):

[0118] Given a training set Y = [y1, y2, ..., y N ], where n is an m-dimensional vector.

[0119] Initialize dictionary D = [d1, d2, ..., d K ], where d k Represents the kth dictionary atom, which is an m-dimensional vector.

[0120] (Sparse Coding):

[0121] For each training sample y n , find a sparse vector x by orthogonal matching pursuit (OMP) n Make y n ≈Dx n .

[0122] (Dictionary update):

[0123] Select a dictionary atom d k Update. For the selected d k , calculate the residual of all samples using this atom Updated k and the corresponding sparse coefficient x k , so that Minimize, here || || F represents the Frobenius norm.

[0124] Among them, dictionary update can be completed through singular value decomposition (SVD):

[0125] For E k Perform SVD to get E k =U∑V T , then the new d k Take U(:,1), the new x k Taken as ∑(1,1)V(:,1) T .

[0126] (iteration):

[0127] Repeat steps 2 and 3 until the stopping condition is met (such as reaching the maximum number of iterations or the reconstruction error is less than a certain threshold).

[0128] (Output):

[0129] Output the final dictionary D and sparse coefficient matrix X.

[0130] Will Figure 2 , Figure 3 The comparison between the EEMD reconstruction results and the K-SVD denoising results of the samples used is shown in Figure 4 shown.

[0131] S102: performing an improved S transform based on dynamic adjustment of key voiceprint components based on energy concentration to obtain an S transform result, extracting features of the S transform result based on wavelet packet energy-Mel frequency cepstrum coefficients to obtain an image sample set corresponding to a voiceprint feature map under a specific water accumulation degree;

[0132] Implement an improved S-transform based on dynamic adjustment of energy concentration, including:

[0133] (initialization):

[0134] Set the initial Gaussian window width σ0 and frequency range f, and define a step size Δσ for dynamically adjusting the Gaussian window width.

[0135] (Calculate the initial S-transform):

[0136] Assume a discrete time signal x[m], and use the initial Gaussian window width σ0 to calculate the S transform S x (f,n).

[0137]

[0138] Where N is the signal length and n is the time index.

[0139] (Calculate energy concentration):

[0140] For each frequency f, calculate the S-transform result S x Energy distribution of (f,n).

[0141] E(f,n)=|S x (f,n)| 2 (4.5)

[0142] To calculate the concentration of energy distribution, entropy or other concentration indicators can be used. For example, entropy H(f) is used to represent the concentration of energy distribution:

[0143]

[0144] in,

[0145] (Dynamically adjust the Gaussian window width):

[0146] The Gaussian window width σ(f) is dynamically adjusted according to the energy concentration H(f). If the energy distribution is more dispersed (H(f) is larger), the Gaussian window width is increased; if the energy distribution is more concentrated (H(f) is smaller), the Gaussian window width is reduced.

[0147] σ(f)=σ0+α(H(f)-H0)(4.7)

[0148] Among them, α is the adjustment coefficient and H0 is the baseline entropy value.

[0149] (Recalculate the S transform):

[0150] Recalculate the S transform S using the adjusted Gaussian window width σ(f) x (f,n).

[0151]

[0152] (Iterative Optimization):

[0153] Repeat the steps until the Gaussian window width converges or reaches a predetermined number of iterations. After each iteration, check whether the energy concentration has been significantly improved. If not, stop the iteration.

[0154] The final S transformation result S x (f,n) has higher time-frequency resolution and is used for further signal analysis.

[0155] Implement feature extraction based on wavelet packet energy-Mel frequency cepstral coefficients, including:

[0156] Use wavelet packet decomposition to transform S into the result S x (f,n) is decomposed into multiple sub-bands, and the appropriate wavelet basis function and decomposition layer number are selected to obtain the coefficients W of multiple sub-bands. j,k (n), where j represents the decomposition layer, k represents the subband index, and n represents the time index.

[0157] Calculate the energy distribution E of each subband j,k :

[0158]

[0159] Among them, wavelet packet decomposition includes:

[0160] Assuming that the signal x(n) is decomposed by wavelet packets, the obtained subband coefficients can be represented by a matrix.

[0161] First level decomposition:

[0162] Decompose the signal x[n] into a low-frequency part A1 and a high-frequency part D1.

[0163] A1=h[n]*x[n] (4.9)

[0164] D1=g[n]*x[n] (4.10)

[0165] where h[n] and g[n] are the impulse responses of the low-pass and high-pass filters, respectively.

[0166] Second level decomposition:

[0167] The low-frequency part A1 and the high-frequency part D1 are decomposed again to obtain A2, D2, A1D1, and D1A1.

[0168] A2=h[n]*A1 (4.11)

[0169] D2=g[n]*A1 (4.12)

[0170] A1D1=h[n]*D1 (4.13)

[0171] D1A1=g[n]*D1 (4.14)

[0172] And so on, until the specified number of decomposition levels is reached.

[0173] Compute Mel-frequency cepstral coefficients (MFCCs), including:

[0174] Transform S to result S x (f,n) Convert from linear frequency to Mel frequency:

[0175]

[0176] Construct a set of Mel filters to cover the entire frequency range. Each filter is a triangle on the Mel frequency axis, and adjacent filters partially overlap.

[0177] Transform S to result S x (f,n) passes through the Mel filter bank to obtain the Mel frequency spectrum M(f m ,n):

[0178]

[0179] Mel frequency spectrum M(f m ,n) performs discrete cosine transform (DCT) to obtain Mel frequency cepstral coefficients (MFCCs):

[0180]

[0181] Use feature selection algorithms (such as mutual information, recursive feature elimination RFE, etc.) to evaluate the contribution of each wavelet packet energy feature and MFCC feature to the state recognition result, select features with high contribution, and form a preliminary feature set. According to the feature contribution evaluation results, assign a weight to each feature. Use support vector machine (SVM) or random forest (RandomForest) and other models to automatically learn feature weights during training

[0182] The weighted wavelet packet energy features and MFCC features are combined to form a comprehensive feature vector F:

[0183]

[0184] Among them, w i is the weight of the ith feature.

[0185] The comprehensive voiceprint feature vectors of each sensor are combined according to the weights to form a multi-dimensional feature map represented by a matrix. The multi-dimensional feature map is converted into an image form to form an image sample set, each of which corresponds to a voiceprint feature map under a specific degree of water accumulation.

[0186] S103: Input the image sample set corresponding to the voiceprint feature map under a specific water accumulation degree into a pre-established convolutional neural network model for training, and output an optimized variable parameter state recognition model based on actual working conditions. The optimized variable parameter state recognition model based on actual working conditions is used to realize online and accurate recognition of the water accumulation state of the cable elbow terminal.

[0187] In terms of neural network structure design, this study divides the learning process of the CNN model into two parts: CNN training and CNN recognition. In the CNN training process, the model parameters and training steps are set in advance; then, through the data forward propagation process and the error backward propagation process, the model parameters are continuously corrected until the convolution model meets the requirements. In CNN recognition, the high-dimensional features extracted by convolution and pooling operations are matched with the trained model to output the recognition results.

[0188] Before training CNN, parameters such as learning rate and batch size are selected more carefully. The learning rate determines the step size for adjusting weights and reducing errors during training, and the batch size refers to the size of the dataset subset input into the CNN network model in each training iteration. The percussion soundprint feature maps of a typical cable elbow terminal under six water accumulation conditions are used as the input features of the CNN recognition model, and the learning rate and batch size are optimized.

[0189] During the training process of the CNN model, 70 sets of image samples of the knocking sound signal under each waterlogging condition were randomly selected as training sets and input into the CNN. The remaining 30 sets of image samples were input into the trained CNN model as the validation set to complete the recognition of the water accumulation of the cable elbow terminal. The appropriate CNN model optimizer was selected according to the characteristics of the study.

[0190] Construct an optimized variable parameter state identification model based on actual working conditions, including:

[0191] Data collection, ensuring that the collected data covers voiceprint signals under different size parameters (such as diameter, length, etc.) and different operating environments (such as temperature, humidity, etc.);

[0192] Extract comprehensive feature vectors from the voiceprint signal of each sensor using the above-mentioned wavelet packet decomposition, MFCC extraction and other methods;

[0193] The model constructed under standard working conditions is used as the basic model for subsequent variable parameter optimization;

[0194] Conduct sensitivity tests on different size parameters and operating environments to evaluate the impact of these parameter changes on model performance;

[0195] Use reinforcement learning to dynamically adjust the model parameters based on the feedback of the parameters' impact on model performance. The specific steps include:

[0196] Define the state of the system, including current size parameters, operating environment, etc.;

[0197] Define the range of adjustable model parameters;

[0198] Design reward functions to encourage the model to achieve optimal performance under different conditions;

[0199] By interacting with the environment, the optimal parameter adjustment strategy is learned.

[0200] Conduct model validation and testing, including:

[0201] Use cross-validation methods to ensure the generalization ability of the model under different working conditions;

[0202] Evaluate the performance indicators of the model under different working conditions, such as accuracy, recall, etc.;

[0203] Test the stability and robustness of the model over long periods of time.

[0204] The optimized variable parameter model is deployed in practical applications to achieve real-time waterlogging status identification.

[0205] Continuously monitor the performance of the model and record indicators such as recognition accuracy and false alarm rate;

[0206] Based on the monitoring results, the model is regularly updated and optimized to maintain its high-precision recognition capability under different working conditions.

[0207] Specifically, the present invention is further described below by way of embodiments:

[0208] Embodiment 2: In the second aspect, as Figure 7 As shown, in order to achieve the above purpose, the present invention discloses a cable elbow voiceprint signal feature extraction and recognition system, comprising:

[0209] The signal processing module 11 is used to receive the acoustic signal of the cable elbow terminal body, and extract the intrinsic component based on the ensemble empirical mode decomposition algorithm EEMD of the intrinsic component threshold range to obtain the reconstructed signal; the reconstructed signal is based on the acoustic signal blind separation algorithm of sparse representation to obtain the key voiceprint component;

[0210] The transformation extraction module 12 is used to dynamically adjust the key voiceprint components based on the improved S transformation of energy concentration to obtain the S transformation result, extract the characteristics of the S transformation result based on the wavelet packet energy-Mel frequency cepstrum coefficient, and obtain an image sample set corresponding to the voiceprint feature map under a specific water accumulation degree;

[0211] The state recognition module 13 is used to input the image sample set corresponding to the voiceprint feature map under a specific water accumulation degree into a pre-established convolutional neural network model for training, and output an optimized variable parameter state recognition model based on actual working conditions. The optimized variable parameter state recognition model based on actual working conditions is used to realize online and accurate recognition of the water accumulation state of the cable elbow terminal.

[0212] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0213] It needs to be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0214] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0215] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. A cable elbow voiceprint signal feature extraction and recognition method, characterized in that: The method comprises the following steps: Receive the acoustic signal of the cable elbow terminal, and extract the intrinsic component based on the ensemble empirical mode decomposition algorithm EEMD of the intrinsic component threshold range to obtain the reconstructed signal; use the acoustic signal blind separation algorithm based on sparse representation to obtain the key voiceprint component; The key voiceprint components are dynamically adjusted based on the improved S transform of energy concentration to obtain the S transform result, and the S transform result is extracted based on the characteristics of wavelet packet energy-Mel frequency cepstrum coefficient to obtain an image sample set corresponding to the voiceprint feature map under a specific degree of water accumulation; The image sample set corresponding to the voiceprint feature map under a specific water accumulation degree is input into the pre-established convolutional neural network model for training, and the optimized variable parameter state recognition model based on actual working conditions is output. The optimized variable parameter state recognition model based on actual working conditions can be used to realize online and accurate recognition of the water accumulation state of the cable elbow terminal.

2. According to claim 1, a cable elbow voiceprint signal feature extraction and recognition method is characterized in that: The ensemble empirical mode decomposition algorithm process based on the intrinsic component threshold range includes the following steps: Construct a new time series by adding the white noise time series to the original knock signal time series: X i (t)=x(t)+w i (t),i=1,2,...,N(4.1) Where X i (t) is the newly constructed time series, x(t) is the time series of the knocking signal under standard working conditions, and w i (t) is the white noise time series, N is the number of times different sequences of white noise are added to the original signal and the empirical mode decomposition (EMD) is performed; For each newly generated signal X i (t), using the standard empirical mode decomposition EMD, a series of intrinsic mode functions IMFs are obtained: IMF ij (t) represents the jth IMF component obtained in the i-th experiment, r i (t) represents the residual of the i-th test, and M represents the number of extracted IMF components; Repeat the iteration N times, and for each IMF position, take the average of all test results as the final IMF; the final IMF calculation formula is: The spectrum of the IMF component is calculated and compared with the frequency band of the cable elbow terminal body acoustic signal. The intrinsic mode functions with small frequency band intersection and representing continuous noise interference are removed, and the remaining intrinsic mode functions are superimposed to obtain the reconstructed signal.

3. The cable elbow voiceprint signal feature extraction and recognition method according to claim 1 is characterized in that: The process of obtaining the key voiceprint components by using the acoustic signal blind separation algorithm based on the sparse representation of the reconstructed signal is as follows: initialization: Given a training set Y = [y1,y2,...,y N ], where n is an m-dimensional vector; Initialize dictionary D = [d1, d2, ..., d K ], where d k represents the kth dictionary atom, which is an m-dimensional vector; The initial sparse coefficient matrix is ​​X = [x1, x2, ..., x N ]; Sparse Coding: For each training sample y n , find the sparse vector x by orthogonal matching pursuit OMP n Make y n ≈Dx n ; dictionary renew: Select dictionary atom d k Update, for the selected d k , calculate the residual of all samples using this atom Updated k and the corresponding sparse coefficient x k , so that Minimize, || || F Represents the Frobenius norm; repeat the steps until the stopping condition is met, such as reaching the maximum number of iterations or the reconstruction error is less than a certain threshold; output the final dictionary D and sparse coefficient matrix X; calculate the final training set Y, reconstruct the signal, and obtain the target voiceprint.

4. The cable elbow voiceprint signal feature extraction and recognition method according to claim 3 is characterized in that: The dictionary update is done by singular value decomposition SVD: For E k Perform SVD to get E k =U∑V T , then the new d k Take U(:,1), the new x k Taken as ∑(1,1)V(:,1) T .

5. The cable elbow voiceprint signal feature extraction and recognition method according to claim 1 is characterized in that: The steps of the improved S transform based on dynamic adjustment of energy concentration include: initializing data, calculating initial S transform, calculating energy concentration, dynamically adjusting Gaussian window width, recalculating S transform and iterative optimization.

6. A cable elbow voiceprint signal feature extraction and recognition method according to claim 5, characterized in that: The process of initializing data: Set the initial Gaussian window width σ0 and frequency range f, and define the step size Δσ; The process of calculating the initial S-transform: Assume a discrete time signal x[m], and use the initial Gaussian window width σ0 to calculate the S transform S x (f,n); Where N is the signal length and n is the time index; The process of calculating energy concentration: For each frequency f, calculate the S-transform result S x Energy distribution of (f,n); E(f,n)=|S x (f,n)| 2 (4.5) Calculate the concentration of energy distribution, using entropy or other concentration indicators: Entropy H(f) is used to represent the concentration of energy distribution: in, The process of dynamically adjusting the Gaussian window width: The Gaussian window width σ(f) is dynamically adjusted according to the energy concentration H(f); if H(f) is large, the Gaussian window width is increased; if H(f) is small, the Gaussian window width is reduced; σ(f)=σ0+α(H(f)-H0)(4.7) Among them, α is the adjustment coefficient, H0 is the baseline entropy value; The process of recalculating the S transform: Recalculate the S transform S using the adjusted Gaussian window width σ(f) x (f,n); Iterative optimization process: Repeat the steps until the Gaussian window width converges or reaches a predetermined number of iterations; The final S transformation result S x (f,n).

7. The cable elbow voiceprint signal feature extraction and recognition method according to claim 1 is characterized in that: The process of extracting the features of the S-transform results based on the wavelet packet energy-Mel frequency cepstrum coefficients to obtain an image sample set corresponding to the voiceprint feature map under a specific water accumulation degree: Use wavelet packet decomposition to transform S into the result S x (f,n) is decomposed into multiple sub-bands, wavelet basis functions and decomposition levels are selected, and the coefficients W of multiple sub-bands are obtained. j,k (n), where j represents the decomposition layer, k represents the subband index, and n represents the time index; Calculate the energy distribution E of each subband j,k : Among them, wavelet packet decomposition includes: Assume that the signal x(n) is decomposed by wavelet packet, and the obtained subband coefficients are represented by a matrix; First level decomposition: Decompose the signal x[n] into a low-frequency part A1 and a high-frequency part D1; A1=h[n]*x[n](4.9) D1=g[n]*x[n](4.10) where h[n] and g[n] are the impulse responses of the low-pass and high-pass filters, respectively; Second level decomposition: Decompose the low-frequency part A1 and the high-frequency part D1 again to obtain A2, D2, A1D1, and D1A1; A2=h[n]*A1(4.11) D2=g[n]*A1(4.12) A1D1=h[n]*D1(4.13) D1A1=g[n]*D1(4.14) Until the specified number of decomposition layers is reached; Calculate Mel-frequency cepstral coefficients MFCCs, including: Transform S to result S x (f,n) Convert from linear frequency to Mel frequency: Transform S to result S x (f,n) passes through the Mel filter bank to obtain the Mel frequency spectrum M(f m ,n): Mel frequency spectrum M(f m ,n) Perform discrete cosine transform DCT to obtain Mel frequency cepstrum coefficients MFCCs: Use feature selection algorithm to evaluate the contribution of each wavelet packet energy feature and MFCC feature to the state recognition result, select features with high contribution, form a preliminary feature set, assign a weight to each feature according to the feature contribution evaluation result, and use support vector machine SVM or random forest Random Forest model to automatically learn feature weights during training: The weighted wavelet packet energy features and MFCC features are combined to form a comprehensive feature vector F: Among them, w i is the weight of the i-th feature; The comprehensive voiceprint feature vectors of each sensor are combined according to the weights to form a multidimensional feature map represented by a matrix. The multidimensional feature map is converted into an image form to form an image sample set. Each image sample corresponds to a voiceprint feature map under a specific degree of water accumulation.

8. A cable elbow voiceprint signal feature extraction and recognition system, characterized in that: include: The signal processing module is used to receive the acoustic signal of the cable elbow terminal body, and extract the intrinsic component based on the EEMD algorithm of the ensemble empirical mode decomposition of the intrinsic component threshold range to obtain the reconstructed signal; the reconstructed signal is based on the acoustic signal blind separation algorithm of the sparse representation to obtain the key voiceprint component; The transformation extraction module is used to dynamically adjust the key voiceprint components based on the improved S transformation of energy concentration to obtain the S transformation result, extract the characteristics of the S transformation result based on the wavelet packet energy-Mel frequency cepstrum coefficient, and obtain an image sample set corresponding to the voiceprint feature map under a specific water accumulation degree; The state recognition module is used to input the image sample set corresponding to the voiceprint feature map under a specific water accumulation degree into a pre-established convolutional neural network model for training, and output an optimized variable parameter state recognition model based on actual working conditions. The optimized variable parameter state recognition model based on actual working conditions is used to realize online and accurate recognition of the water accumulation state of the cable elbow terminal.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, a cable elbow voiceprint signal feature extraction and recognition method according to any one of claims 1 to 7 is adopted.

10. A computer-readable storage medium having a computer program stored therein, characterized in that: When the computer program is loaded and executed by the processor, a cable elbow voiceprint signal feature extraction and recognition method according to any one of claims 1 to 7 is adopted.

Images