Transformer fault diagnosis method based on improved MFCC and RNN

Abstract

The invention relates to the technical field of power equipment fault diagnosis, in particular to a method for performing transformer fault diagnosis by using an improved Mel-frequency cepstral coefficient (MFCC) and a recurrent neural network (RNN) through monitoring and analyzing sound signals during operation of a transformer. According to the method, a microphone array is arranged to collect voiceprint signals during operation of a transformer, and the voiceprint signals are preprocessed to eliminate noise interference. Then, spectrum features of the signals are extracted through dual-tree complex wavelet transform, and a feature parameter data set is constructed through Mel filtering and discrete transform; and the feature data sets are input into the RNN network for deep information perception and learning, so that intelligent identification of different operation states of the transformer is realized. In the diagnosis operation process of the RNN network, the advantage of processing sequence data is utilized, the accuracy and the real-time performance of fault identification are improved through updating of an internal state and the action of an activation function, and the robustness of the system is enhanced.

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis of electric power equipment, and in particular to a method for diagnosing transformer faults by monitoring and analyzing sound signals during transformer operation and utilizing improved Mel Frequency Cepstral Coefficients (MFCC) and Recurrent Neural Networks (RNN). Background Art

[0002] In the power system, the transformer is a key device for energy conversion and transmission. Its operating status directly affects the safety and stability of the entire power grid. Therefore, timely and accurate fault diagnosis of the transformer is of great significance. In recent years, with the rapid development of signal processing technology and artificial intelligence technology, transformer fault diagnosis methods based on sound signals have gradually become a research hotspot.

[0003] Traditional transformer fault diagnosis methods mainly include oil chromatography analysis, electrical quantity monitoring and manual inspection. Although these methods can reflect the operating status of the transformer to a certain extent, they have many limitations. For example, oil chromatography analysis requires regular sampling and sending it to the laboratory for analysis, which has poor timeliness and high cost; although electrical quantity monitoring can reflect the electrical characteristics of the transformer in real time, it is often difficult to directly link it to the specific fault type; manual inspection relies on the experience and judgment of the operation and maintenance personnel, and has problems such as strong subjectivity and high labor intensity.

[0004] The transformer fault diagnosis method based on sound signals uses the sound signals generated by the transformer during operation, extracts characteristic parameters through signal processing technology, and combines intelligent recognition algorithms to determine the fault type. This method has the advantages of non-intrusiveness, strong real-time performance, and low cost, so it has attracted widespread attention.

[0005] In sound signal processing technology, Mel-Frequency Cepstral Coefficient (MFCC) is a commonly used feature extraction method that can effectively describe the spectral characteristics of sound signals. However, when processing complex sound signals, the traditional MFCC method often has problems such as inaccurate feature extraction and poor robustness. At the same time, traditional recognition algorithms such as support vector machines (SVM) and hidden Markov models (HMM) also have certain limitations when processing time series data.

[0006] When extracting transformer sound signal features, the traditional MFCC method may not be able to fully capture the key information in the sound signal, resulting in inaccurate feature vectors, which in turn affects the subsequent fault identification effect. The transformer operating environment is complex and changeable, and the sound signal is easily interfered by various factors such as environmental noise and mechanical vibration. When processing these interference signals, the traditional MFCC method often finds it difficult to maintain stable performance, resulting in poor robustness of fault diagnosis.

[0007] When processing time series data, traditional recognition algorithms often fail to fully utilize the time correlation information in the data, resulting in low recognition accuracy. Especially for equipment with complex dynamic characteristics such as transformers, traditional recognition algorithms are even more difficult to meet actual needs.

[0008] Although the fault diagnosis method based on sound signals has the advantage of strong real-time performance, in practical applications, the time-consuming data processing and algorithm calculation may affect the overall real-time performance. This problem is particularly prominent when dealing with large-scale data or complex models.

[0009] In view of the shortcomings of traditional transformer fault diagnosis methods and the limitations of sound signal processing technology, it is necessary to propose a transformer fault diagnosis method based on improved MFCC and RNN to improve the accuracy, robustness and real-time performance of fault diagnosis. Summary of the invention

[0010] The present invention aims to propose a transformer fault diagnosis method based on improved Mel Frequency Cepstral Coefficient (MFCC) and Recurrent Neural Network (RNN) to achieve efficient and accurate monitoring and fault diagnosis of transformer operating status. By optimizing the feature extraction algorithm of sound signals and building advanced recognition models, the accuracy of fault diagnosis is improved, the false alarm rate and missed alarm rate are reduced, and a strong guarantee is provided for the stable operation of the power system.

[0011] The traditional MFCC method may not be able to fully capture the key information in the sound signal when extracting the transformer sound signal features, resulting in inaccurate feature vectors. The present invention improves the accuracy and robustness of feature extraction by improving the MFCC algorithm and introducing more sophisticated spectrum analysis and feature selection strategies.

[0012] The technical solution of the present invention is as follows:

[0013] A transformer fault diagnosis method based on improved MFCC and RNN includes the following steps:

[0014] S1: Arrange a microphone array to collect transformer voiceprint signals in different operating states and pre-process the voiceprint signals;

[0015] S2: The spectrum characteristics of the signal are obtained by using the dual-tree complex wavelet transform;

[0016] S3: Establish feature parameter data set through Mel filtering and discrete transformation;

[0017] S4: Input the feature data set into the RNN network for deep information perception to achieve intelligent identification of different states of the transformer.

[0018] Preferably, the diagnostic operation process of the RNN in step S4 is as follows:

[0019]

[0020] In the formula, α, β, γ, d, c are the parameters of the model, H t is the hidden state of RNN, f(x) represents the activation function, which is the hyperbolic tangent function, and softmax(x) is the normalized exponential function, which is used to convert the input real number vector into a probability distribution, as follows:

[0021]

[0022] Where i = 1, 2, ..., n, x i The real number vector x is the input value. 1 ,x 2 ,…,x n ), softmax(x)i is the i-th element of the output vector softmax(x), is the exponent of the i-th element in the input vector, and the denominator is the sum of the exponents of all input elements;

[0023] In the diagnostic operation of RNN, the parameters α, β, γ, d, and c have the following meanings:

[0024] α: Input weight matrix, which determines the current input x t For the hidden state h t the extent of the impact;

[0025] β: recursive weight matrix, which determines the previous hidden state H t-1 For the current hidden state h t the extent of the impact;

[0026] γ: Output weight matrix, which determines the hidden state H t For output y t the extent of the impact;

[0027] d: bias vector, which adds a constant term to the calculation of the hidden state;

[0028] c: output bias vector, which is the output y t A constant term is added to the calculation of .

[0029] Preferably, the parameters α, β, γ, d, and c are obtained through a training process of the RNN model, and the training process includes the following steps:

[0030] S41: Collect sound signal data when the transformer is running, and divide it into a training set and a test set;

[0031] S42: Extract features from sound signals using the improved MFCC method;

[0032] S43: Initialize the parameters α, β, γ, d, c of the RNN model;

[0033] S44: Input the extracted features into the RNN model to calculate the hidden state and output of each time step;

[0034] S45: Calculate the loss function based on the output of the model and the actual fault label;

[0035] S46: Use the loss function to back-propagate the model parameters and calculate the gradient of the parameters;

[0036] S47: Update model parameters using gradient descent;

[0037] S48: Repeat steps S44 to S47 until the performance of the model on the training set no longer improves;

[0038] Through this process, the model parameters α, β, γ, d, and c are adjusted to minimize the loss function, thereby improving the accuracy of fault diagnosis.

[0039] Preferably, the preprocessing of the voiceprint signal in S1 includes three steps: pre-emphasis, framing, and windowing. Pre-emphasis is used to enhance the high-frequency sound signal so that the acoustic model can more accurately capture the characteristics of high-frequency oscillations. The expression of the pre-emphasis function is as follows:

[0040] y(t)=x(t)-σx(t-1)

[0041] Where x(t) is the original sound signal; σ is the pre-emphasis coefficient, which is between 0.9 and 1.0;

[0042] Framing is to divide the long-time signal into short-time frames, and the window length device is 25 milliseconds;

[0043] Frame shift is to move the window frame by frame by the specified frame shift amount after framing. The frame shift determines the degree of overlap between adjacent frames. The frame shift is set to 10 milliseconds.

[0044] Preferably, the step of extracting signal spectrum features using dual-tree complex wavelet transform described in S2 is:

[0045] S21: Select a pair of complex wavelet functions, namely "complex wavelet" and "conjugate wavelet";

[0046] S22: construct two wavelet trees, each tree uses a different wavelet function, one tree uses a complex wavelet function, and the other tree uses its conjugate wavelet function. The two trees perform wavelet decomposition on the signal to obtain sub-band signals of different scales and positions, and each sub-band signal represents a component of the signal within a specific frequency range;

[0047] S23: Calculate the time-frequency characteristics of each sub-band signal;

[0048] S24: combining the energy and phase information features of all sub-bands to form a complete spectrum feature of the signal;

[0049] S25: selecting those sub-band features that are most relevant to a specific sound event;

[0050] S26: Normalize, denoise and fuse the extracted features to improve the robustness and discrimination of the features.

[0051] Preferably, the step of establishing a feature parameter data set by Mel filtering and discrete transformation in S3 includes: processing the spectrum obtained in S2 through a Mel filter bank to obtain a Mel spectrum; performing cepstrum analysis on the Mel spectrum, first taking the logarithm of the Mel spectrum, and then performing discrete cosine transform on it, wherein the calculation formula of the Mel frequency in the Mel spectrum is as follows:

[0052]

[0053] In the formula, f is the linear frequency. This formula converts the linear frequency into a scale that is more consistent with the human ear perception characteristics through a nonlinear transformation, making it more suitable for audio analysis.

[0054] Preferably, step S1 further includes customizing the Mel filter bank according to the specific characteristics and fault mode of the transformer, as follows:

[0055] S11. Use a high-sensitivity microphone to record the sound of the transformer in normal operation and different fault conditions at different locations;

[0056] S12. Determine the frequency range of the sound generated under different faults through spectrum analysis;

[0057] S13. Design a Mel filter bank according to the frequency spectrum characteristics of the fault sound, and ensure that its Mel scale can cover the range of the fault sound;

[0058] S14. Determine the fault type with greater impact, and design a denser filter distribution for the fault frequency range of this type, so as to better capture the sound characteristics of this fault type.

[0059] Preferably, the fault type with greater impact described in S14 is determined by constructing a scoring model for comparison and scoring, specifically as follows: assuming that the impact score of fault type f is S(f),

[0060] S(f)=w y y(f)+w R R(f)+w I I(f)

[0061] y(f) is the recovery difficulty score of fault type f, R(f) is the occurrence frequency score of fault type f, and I(f) is the impact range score of fault type f on system operation. The weights of the three scores are w, y , w R and w I ;w y , w R and w I Obtained using the analytic hierarchy process.

[0062] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a transformer fault diagnosis method based on improved MFCC and RNN according to any one of claims 1 to 8 are implemented.

[0063] Traditional recognition algorithms such as SVM and HMM often fail to fully utilize the time correlation information in the data when processing time series data, resulting in low recognition accuracy. RNN, as a neural network structure designed specifically for processing sequence data, can well capture the temporal dependency in the data. Therefore, the present invention uses RNN as a fault recognition model to improve recognition accuracy.

[0064] Although the sound signal processing method has the advantage of strong real-time performance in theory, in actual application, data processing and algorithm calculation may consume a lot of time, affecting the overall real-time performance. The present invention reduces the amount of calculation and improves the processing speed by optimizing the algorithm structure and implementation method to meet the real-time requirements.

[0065] The operating status and environmental conditions of the power system will change over time. Traditional fault diagnosis methods often lack adaptive capabilities and are difficult to adjust in time to adapt to new environmental conditions. The present invention trains the RNN model to enable it to have learning capabilities, and can adjust model parameters according to historical data and real-time data to improve adaptive capabilities.

[0066] The operating environment of the transformer is complex and changeable, and the sound signal is easily interfered by various factors such as environmental noise and mechanical vibration. The present invention reduces the influence of interference factors on fault diagnosis and improves the anti-interference ability of the system by introducing a denoising preprocessing step and an optimized feature extraction algorithm.

[0067] In summary, the present invention aims to solve the problems existing in the traditional transformer fault diagnosis method, such as inaccurate feature extraction, insufficient recognition algorithm processing ability of time series data, poor real-time performance, insufficient adaptability, and interference from environmental noise and mechanical vibration. By introducing improved MFCC and RNN technology, the accuracy and reliability of transformer fault diagnosis are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a transformer fault diagnosis flow chart of the present invention.

[0069] Figure 2 This is a diagram of transformer fault diagnosis accuracy and loss value according to the present invention. DETAILED DESCRIPTION

[0070] See also Figure 1-2 .

[0071] Transformer fault diagnosis process based on improved MFCC and RNN

[0072] The present invention proposes a transformer fault diagnosis method based on improved MFCC and RNN, which captures the frequency locality of the signal through dual-tree complex wavelet transform, and then extracts the Mel spectrum coefficients through Mel filtering to improve the diagnostic recognition ability of the RNN system.

[0073] The specific process is as follows:

[0074] 1) Arrange a microphone array to collect transformer voiceprint signals in different operating states and pre-process the voiceprint signals.

[0075] 2) Obtain the spectral characteristics of the signal using dual-tree complex wavelet transform

[0076] 3) Establish feature parameter data set through Mel filtering and discrete transformation.

[0077] 4) Input the feature data set into the RNN network for deep information perception to realize intelligent identification of different states of the transformer.

[0078] The diagnostic operation process of RNN in step S4 is as follows:

[0079]

[0080] In the formula, α, β, γ, d, c are the parameters of the model, H t is the hidden state of RNN, f(x) represents the activation function, which is the hyperbolic tangent function, and softmax(x) is the normalized exponential function, which is used to convert the input real number vector into a probability distribution, as follows:

[0081]

[0082] Where i = 1, 2, ..., n, x i The real number vector x is the input value. 1 ,x 2 ,…,x n ), softmax(x)i is the i-th element of the output vector softmax(x), is the exponent of the i-th element in the input vector, and the denominator is the sum of the exponents of all input elements;

[0083] In the diagnostic operation of RNN, the parameters α, β, γ, d, and c have the following meanings:

[0084] α: Input weight matrix, which determines the current input x t For the hidden state h t the extent of the impact;

[0085] β: recursive weight matrix, which determines the previous hidden state H t-1 For the current hidden state h t the extent of the impact;

[0086] γ: Output weight matrix, which determines the hidden state H t For output y t the extent of the impact;

[0087] d: bias vector, which adds a constant term to the calculation of the hidden state;

[0088] c: output bias vector, which is the output y t A constant term is added to the calculation of ;

[0089] The parameters α, β, γ, d, and c are obtained through the training process of the RNN model. The training process includes the following steps:

[0090] S41: Collect sound signal data when the transformer is running, and divide it into a training set and a test set;

[0091] S42: Extract features from sound signals using the improved MFCC method;

[0092] S43: Initialize the parameters α, β, γ, d, c of the RNN model;

[0093] S44: Input the extracted features into the RNN model to calculate the hidden state and output of each time step;

[0094] S45: Calculate the loss function based on the output of the model and the actual fault label;

[0095] S46: Use the loss function to back-propagate the model parameters and calculate the gradient of the parameters;

[0096] S47: Update model parameters using gradient descent;

[0097] S48: Repeat steps S44 to S47 until the performance of the model on the training set no longer improves;

[0098] Through this process, the model parameters α, β, γ, d, and c are adjusted to minimize the loss function, thereby improving the accuracy of fault diagnosis;

[0099] The preprocessing of the voiceprint signal described in S1 includes three steps: pre-emphasis, framing, and windowing. Pre-emphasis is used to enhance the high-frequency sound signal so that the acoustic model can more accurately capture the characteristics of high-frequency oscillations. The expression of the pre-emphasis function is as follows:

[0100] y(t)=x(t)-σx(t-1)

[0101] Where x(t) is the original sound signal; σ is the pre-emphasis coefficient, which is between 0.9 and 1.0;

[0102] Framing is to divide the long-time signal into short-time frames, and the window length device is 25 milliseconds;

[0103] Frame shift is to move the window frame by frame by the specified frame shift amount after framing. The frame shift determines the overlap between adjacent frames. The frame shift is set to 10 milliseconds.

[0104] The steps of using dual-tree complex wavelet transform to extract signal spectrum features in S2 are:

[0105] S21: Select a pair of complex wavelet functions, namely "complex wavelet" and "conjugate wavelet";

[0106] S22: construct two wavelet trees, each tree uses a different wavelet function, one tree uses a complex wavelet function, and the other tree uses its conjugate wavelet function. The two trees perform wavelet decomposition on the signal to obtain sub-band signals of different scales and positions, and each sub-band signal represents a component of the signal within a specific frequency range;

[0107] S23: Calculate the time-frequency characteristics of each sub-band signal;

[0108] S24: combining the energy and phase information features of all sub-bands to form a complete spectrum feature of the signal;

[0109] S25: selecting those sub-band features that are most relevant to a specific sound event;

[0110] S26: normalizing, denoising and fusing the extracted features to improve the robustness and discrimination of the features;

[0111] The step of establishing a feature parameter data set by Mel filtering and discrete transformation in S3 includes: processing the spectrum obtained in S2 through a Mel filter bank to obtain a Mel spectrum; performing cepstrum analysis on the Mel spectrum, first taking the logarithm of the Mel spectrum, and then performing discrete cosine transformation on it, wherein the calculation formula of the Mel frequency in the Mel spectrum is as follows:

[0112]

[0113] In the formula, f is the linear frequency. This formula converts the linear frequency into a scale that is more in line with the human ear's perception characteristics through a nonlinear transformation, making it more suitable for audio analysis.

[0114] Step S1 also includes customizing the Mel filter bank according to the specific characteristics and failure modes of the transformer, as follows:

[0115] S11. Use a high-sensitivity microphone to record the sound of the transformer in normal operation and different fault conditions at different locations;

[0116] S12. Determine the frequency range of the sound generated under different faults through spectrum analysis;

[0117] S13. Design a Mel filter bank according to the frequency spectrum characteristics of the fault sound, and ensure that its Mel scale can cover the range of the fault sound;

[0118] S14. Determine a fault type with a greater impact, and design a denser filter distribution for the fault frequency range of this type, so as to better capture the sound characteristics of this fault type;

[0119] The fault type with greater impact as described in S14 is determined by constructing a scoring model for comparison and scoring, as follows: Assume that the importance score of fault type f is S(f),

[0120] S(f)=w y y(f)+w R R(f)+w I I(f)

[0121] y(f) is the identification difficulty score of fault type f, R(f) is the occurrence frequency score of fault type f, and I(f) is the impact range score of fault type f on system operation. The weights of the three scores are w, y , w R and w I ;w y , w R and w I Obtained using the analytic hierarchy process.

[0122] 1. Transformer abnormal noise fault type

[0123] Usually, components such as windings, cores and oil tanks constitute the sound-generating parts of the transformer during operation. The noise of the above components during operation will indicate the occurrence of transformer failure. Common fault types are shown in Table 1.

[0124] Table 1 Common abnormal noise faults of transformers

[0125]

[0126] During the specific implementation process, high-sensitivity microphones and spectrum analysis tools are used to record and analyze the specific frequency ranges of abnormal noises in the transformer core, windings, tank walls, and corona flashover discharge sounds.

[0127] 2. Transformer operating status assessment

[0128] In this experiment, the model iteration training times were set to 200 times so that the algorithm results tended to converge. Figure 2 The curves of training accuracy and loss rate are depicted. It can be observed from the figure that during the model training process, the training loss rate gradually decreases, while the training accuracy gradually approaches 100%

[0129] 3. Comparison of classification accuracy of different fault models

[0130] This experiment selects CNN and LSTM (Long Short Term Memory) two fault diagnosis algorithms for effect comparison.

[0131] Table 2 Comparison of classification accuracy of different fault models

[0132]

[0133] From Table 2, it can be seen that the classification accuracy of CNN neural network and LSTM neural network for different discharge models is low. This is because the traditional shallow neural network is prone to fall into local minimum or overfitting when processing high-dimensional data samples. In contrast, the improved MFCC-RNN model proposed in this paper has good recognition performance. By replacing Fourier transform with dual-tree complex wavelet transform in the step of extracting MFCC feature parameters, the signal processing capability is improved, and the LSTM layer is integrated on the basis of extracting feature parameters to improve data processing capability, thereby improving the classification and recognition capabilities of various abnormal noise fault models, and the overall correct recognition rate can reach 98%.

[0134] Based on the analysis of the above experimental results, the improved model successfully overcomes the challenges of high-dimensional data processing, ensures accurate extraction of key information in fault audio data, and shows significant effectiveness in retaining fault audio data information. In particular, when processing samples with high-dimensional features, the model still shows excellent recognition results.

Claims

1. A transformer fault diagnosis method based on improved MFCC and RNN, characterized in that: The following steps are involved: S1: Arrange a microphone array to collect transformer voiceprint signals in different operating states and pre-process the voiceprint signals; S2: The spectrum characteristics of the signal are obtained by using the dual-tree complex wavelet transform; S3: Establish feature parameter data set through Mel filtering and discrete transformation; S4: Input the feature data set into the RNN network for deep information perception to achieve intelligent identification of different states of the transformer.

2. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 1, characterized in that: The diagnostic operation process of RNN in step S4 is as follows: In the formula, α, β, γ, d, c are the parameters of the model, H t is the hidden state of RNN, f(x) represents the activation function, which is the hyperbolic tangent function, and softmax(x) is the normalized exponential function, which is used to convert the input real number vector into a probability distribution, as follows: Where i = 1, 2, ..., n, x i The real number vector x is input as (x1, x2, ..., x n ), soffmax(x)i is the i-th element of the output vector soffmax(x), is the exponent of the i-th element in the input vector, and the denominator is the sum of the exponents of all input elements; In the diagnostic operation of RNN, the parameters α, β, Y, d, and c have the following meanings: α: Input weight matrix, which determines the current input x t For the hidden state h t the extent of the impact; β: recursive weight matrix, which determines the previous hidden state H t-1 For the current hidden state h t the extent of the impact; γ: Output weight matrix, which determines the hidden state H t For output y t the extent of the impact; d: bias vector, which adds a constant term to the calculation of the hidden state; c: output bias vector, which is the output y t A constant term is added to the calculation of .

3. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 2, characterized in that: The parameters α, β, γ, d, and c are obtained through the training process of the RNN model. The training process includes the following steps: S41: Collect sound signal data when the transformer is running, and divide it into a training set and a test set; S42: Extract features from sound signals using the improved MFCC method; S43: Initialize the parameters α, β, γ, d, c of the RNN model; S44: Input the extracted features into the RNN model to calculate the hidden state and output of each time step; S45: Calculate the loss function based on the output of the model and the actual fault label; S46: Use the loss function to back-propagate the model parameters and calculate the gradient of the parameters; S47: Update model parameters using gradient descent; S48: Repeat steps S44 to S47 until the performance of the model on the training set no longer improves; Through this process, the model parameters α, β, γ, d, and c are adjusted to minimize the loss function, thereby improving the accuracy of fault diagnosis.

4. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 1, characterized in that: The preprocessing of the voiceprint signal described in S1 includes three steps: pre-emphasis, framing, and windowing. Pre-emphasis is used to enhance the high-frequency sound signal so that the acoustic model can more accurately capture the characteristics of high-frequency oscillations. The expression of the pre-emphasis function is as follows: y(t)=x(t)-σx(t-1) Where x(t) is the original sound signal; σ is the pre-emphasis coefficient, which is between 0.9 and 1.0; Framing is to divide the long-time signal into short-time frames, and the window length device is 25 milliseconds; Frame shift is to move the window frame by frame by the specified frame shift amount after framing. The frame shift determines the degree of overlap between adjacent frames. The frame shift is set to 10 milliseconds.

5. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 1, characterized in that: The steps of using dual-tree complex wavelet transform to extract signal spectrum features in S2 are: S21: Select a pair of complex wavelet functions, namely "complex wavelet" and "conjugate wavelet"; S22: construct two wavelet trees, each tree uses a different wavelet function, one tree uses a complex wavelet function, and the other tree uses its conjugate wavelet function. The two trees perform wavelet decomposition on the signal to obtain sub-band signals of different scales and positions, and each sub-band signal represents a component of the signal within a specific frequency range; S23: Calculate the time-frequency characteristics of each sub-band signal; S24: combining the energy and phase information features of all sub-bands to form a complete spectrum feature of the signal; S25: selecting those sub-band features that are most relevant to a specific sound event; S26: Normalize, denoise and fuse the extracted features to improve the robustness and discrimination of the features.

6. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 1, characterized in that: The step of establishing a feature parameter data set by Mel filtering and discrete transformation in S3 includes: processing the spectrum obtained in S2 through a Mel filter bank to obtain a Mel spectrum; performing cepstrum analysis on the Mel spectrum, first taking the logarithm of the Mel spectrum, and then performing discrete cosine transformation on it, wherein the calculation formula of the Mel frequency in the Mel spectrum is as follows: In the formula, f is the linear frequency. This formula converts the linear frequency into a scale that is more consistent with the human ear perception characteristics through a nonlinear transformation, making it more suitable for audio analysis.

7. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 1, characterized in that: Step S1 also includes customizing the Mel filter bank according to the specific characteristics and failure modes of the transformer, as follows: S11. Use a high-sensitivity microphone to record the sound of the transformer in normal operation and different fault conditions at different locations; S12. Determine the frequency range of the sound generated under different faults through spectrum analysis; S13. Design a Mel filter bank according to the frequency spectrum characteristics of the fault sound, and ensure that its Mel scale can cover the range of the fault sound; S14. Determine the fault type with greater impact, and design a denser filter distribution for the fault frequency range of this type, so as to better capture the sound characteristics of this fault type.

8. A transformer fault diagnosis method based on improved MFCC and RNN according to claim 7, characterized in that: The fault type with greater impact as described in S14 is determined by constructing a scoring model for comparison and scoring, as follows: Assume that the impact score of fault type f is S(f), S(f)=w y y(f)+w R R(f)+w I I(f) y(f) is the recovery difficulty score of fault type f, R(f) is the occurrence frequency score of fault type f, and I(f) is the impact range score of fault type f on system operation. The weights of the three scores are w, y , w R and w I ;w y , w R and w I Obtained using the analytic hierarchy process.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a transformer fault diagnosis method based on improved MFCC and RNN according to any one of claims 1 to 8 are implemented.

Images