A real-time diagnosis method for transformer faults

Abstract

The present invention provides a method for real-time diagnosis of transformer faults. This method uses a delay estimation method based on wavelet denoising and quadratic cross-correlation to eliminate the time difference between various sound signals, thereby realizing the time synchronization of various sound signals, improving the accuracy of voiceprint feature extraction, and effectively suppressing environmental noise. This method independently processes various sound signals in the pre-emphasis, time synchronization, MFCC feature extraction, and quantum coding stages, retains the specific fault information of various sound signals, and improves the flexibility of quantum feature enhancement. Based on the MFCC feature extraction, this method further uses quantum feature coding technology. By mapping the MFCC features of each channel to quantum states and performing parametric entanglement transformation, it effectively enhances the expression ability of fault features, reduces the noise sensitivity of MFCC features, fully excavates the non-linear correlation of features within and between channels, and can improve the accuracy of fault diagnosis results.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer monitoring, and particularly to a method for real-time diagnosis of transformer faults. Background Art

[0002] When a transformer is in operation, noise is generated due to the vibration of the iron core and windings. The acoustic fingerprint signal of the transformer contains a large amount of information characteristics that can reflect the operating state of the equipment. Some existing transformer acoustic fingerprint monitoring technologies have defects. Using multiple microphone sensors in the acoustic signal acquisition link can obtain more comprehensive acoustic fingerprint characteristics, capture the correlation between acoustic signals at different positions, and enhance the ability to resist environmental noise interference. However, factors such as the different distances between the sound source and each microphone sensor, system hardware delay, and insufficient clock synchronization accuracy will all cause the acoustic signals of each channel to be asynchronous in time. The existing transformer acoustic fingerprint monitoring technology design ignores the time difference between the acoustic signals of each channel, resulting in inaccurate acoustic fingerprint feature extraction and weakened environmental noise suppression ability.

[0003] In addition, the existing transformer acoustic fingerprint monitoring technology based on MFCC (Mel Frequency Cepstral Coefficients) feature extraction also has problems such as high noise sensitivity and lack of non-linear correlation between features, resulting in inaccurate fault diagnosis results and a high false alarm rate. Summary of the Invention

[0004] In view of this, the present invention provides a method for real-time diagnosis of transformer faults to improve the accuracy of fault diagnosis results.

[0005] A method for real-time diagnosis of transformer faults includes:

[0006] Step S1, using a plurality of microphone sensors to perform real-time multi-channel acoustic fingerprint acquisition on the transformer, and performing pre-emphasis processing on the acoustic signals of each channel collected by the microphone sensors to obtain the pre-emphasized acoustic signals;

[0007] Step S2, taking the first acoustic signal in the pre-emphasized acoustic signals as a reference signal, and using the delay estimation method based on wavelet denoising and second-order cross-correlation to respectively estimate the time delay estimation values between the acoustic signals of each channel except the first acoustic signal and the first acoustic signal, and obtaining the synchronized acoustic signals of each channel based on the time delay estimation values;

[0008] Step S3, respectively performing MFCC feature extraction on the synchronized acoustic signals of each channel to obtain an MFCC feature matrix, and each row of the MFCC feature matrix stores the MFCC features of each dimension of an acoustic signal of one channel;

[0009] Step S4: Independently normalize the MFCC features of each dimension of each voice signal in the MFCC feature matrix, initialize the MFCC features of each dimension of each normalized voice signal, map them to quantum states, allocate a number of qubits for the MFCC features of each dimension of each voice signal, construct a cross-layer quantum entanglement circuit, and then input each qubit into the cross-layer quantum entanglement circuit to obtain the final quantum state. Calculate the expectation value of the Pauli-Z operator of each qubit in the final quantum state, and thus obtain the features after quantum feature enhancement.

[0010] Step S5: Input the features after quantum feature enhancement into the CNN-LSTM hybrid network and train it. Use the trained CNN-LSTM hybrid network to diagnose the faults of the transformer.

[0011] According to the transformer fault real-time diagnosis method provided by the present invention, the following beneficial effects are achieved:

[0012] (1) The present invention uses the delay estimation method based on wavelet denoising and quadratic cross-correlation to eliminate the time difference between each voice signal, thereby realizing the time synchronization of each voice signal, significantly improving the accuracy of voiceprint feature extraction, and effectively suppressing environmental noise at the same time.

[0013] (2) The present invention independently processes each voice signal in the pre-emphasis, time synchronization, MFCC feature extraction, and quantum coding stages, retains the specific fault information of each voice signal, improves the flexibility of quantum feature enhancement, and can provide better input features for the CNN-LSTM hybrid network.

[0014] (3) On the basis of MFCC feature extraction, the present invention further uses quantum feature coding technology. By mapping the MFCC features of each channel to quantum states and performing parametric entanglement transformation, the expression ability of fault features is effectively enhanced, the noise sensitivity of MFCC features is reduced, the nonlinear correlation between intra-channel features and inter-channel features is fully explored, the problems of high noise sensitivity and lack of nonlinear correlation between features existing in traditional methods are solved, the accuracy of fault diagnosis results can be improved, and the diagnostic false alarm rate can be reduced. Description of the Drawings

[0015] Figure 1 It is a flowchart of the transformer fault real-time diagnosis method provided by the embodiment of the present invention. Detailed Embodiments

[0016] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where like or similar reference numerals denote like or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the embodiments of the present invention and should not be construed as limiting the present invention.

[0017] Please refer to Figure 1 , embodiments of the present invention provide a method for real-time diagnosis of transformer faults, including steps S1 to S5:

[0018] Step S1, use a plurality of microphone sensors to perform real-time multi-channel voiceprint acquisition on the transformer, and perform pre-emphasis processing on each path of sound signals collected by the microphone sensors to obtain pre-emphasized sound signals.

[0019] In this embodiment, 8 microphone sensors are installed around the transformer, so that the total number of channels is 8. The total number of channels is the total number of signal paths, and the sound signals in the working state of the transformer are collected in real-time and multiplexed. During the propagation of sound signals, high-frequency components attenuate faster than low-frequency components. Pre-emphasis can enhance high-frequency components, making the consonant details in the sound signals more obvious. Transmit each path of signals to a high-pass filter for pre-emphasis processing. The expressions of each path of pre-emphasized sound signals are as follows:

[0020] ;

[0021] ;

[0022] where represents the first path of pre-emphasized sound signal, represents the th path of pre-emphasized sound signal, , represents the total number of channels, is the first path of source signal, is the first path of noise signal, is the th path of source signal without time-domain correction, is the th path of noise signal, is time, the th path of sound signal and the time delay between the first path of sound signal. and are theoretically uncorrelated Gaussian white noises.

[0023] Step S2: Use the first voice signal in the pre-emphasized voice signal as the reference signal, and use the delay estimation method based on wavelet denoising and quadratic cross-correlation to estimate the time delay estimation values between each voice signal except the first voice signal and the first voice signal, and obtain the synchronized voice signals based on the time delay estimation values.

[0024] Among them, step S2 specifically includes:

[0025] Step S2.1, solve the autocorrelation function of the first voice signal, and the cross-correlation function between the first voice signal and the th voice signal except the first voice signal. The expressions are as follows:

[0026] ;

[0027] Among them, represents the autocorrelation function of the first voice signal, is the autocorrelation function of, is the and cross-correlation function of, is the and cross-correlation function of, is the autocorrelation function of, is the time delay;

[0028] ;

[0029] Among them, represents the cross-correlation function between the first voice signal and the th voice signal, is the and cross-correlation function of, is the and cross-correlation function of, is the and cross-correlation function of, is the and cross-correlation function of.

[0030] Step S2.2, use wavelet transform to decompose and denoise the autocorrelation function of the first voice signal and the cross-correlation function between the first voice signal and the th voice signal respectively, and obtain the decomposed and denoised autocorrelation function and cross-correlation function;

[0031] Due to the noise term and is not actually ideal Gaussian white noise. Therefore, the noise correlation terms 、 and are not zero. Using the direct correlation method will result in a large error in the time delay estimation value at low signal-to-noise ratios. Therefore, wavelet transform is used to decompose and denoise the autocorrelation function and the cross-correlation function respectively, and the expressions are as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] where is the autocorrelation function of the first sound signal after wavelet denoising, is the time shift factor, is the approximation coefficient of , is the scaling function in the low-frequency band, is the scale factor, is the scaling function, is the detail coefficient of , is the wavelet function in the high-frequency band, is the wavelet function, is the cross-correlation function between the first sound signal after wavelet denoising and the th sound signal, is the approximation coefficient of , is the detail coefficient of .

[0037] By setting appropriate thresholds, the wavelet coefficients with strong noise correlation in , , , are set to zero, so as to achieve the purpose of denoising.

[0038] Step S2.3, calculate the quadratic cross-correlation function of the decomposed and denoised autocorrelation function and cross-correlation function to obtain the quadratic cross-correlation function between the first sound signal and the th sound signal;

[0039] Calculate the quadratic cross-correlation function of and to further suppress the residual noise and obtain the first signal and the Second cross-correlation function of the path signal:

[0040] ;

[0041] wherein, is the second cross-correlation function of the first voice signal and the th voice signal, is the expectation operation; is the time-shifted second cross-correlation function of the first voice signal and the th voice signal, which is the result obtained after delaying by on the time axis. By introducing the correlation at different delays is explored to improve the accuracy of time delay estimation;

[0042] Step S2.4, perform peak search on the second cross-correlation function of the first voice signal and the th voice signal, and take the time delay corresponding to the peak of the second cross-correlation function of the first voice signal and the th voice signal as the time delay estimation value of the first voice signal and the th voice signal;

[0043] Obtain , perform peak search on it, and take the time delay corresponding to the peak value as the time delay estimation value of the first signal and the th signal:

[0044] ;

[0045] wherein, represents taking the maximum value.

[0046] Step S2.5, substitute the time delay estimation value of the first voice signal and the th voice signal into the expressions of the pre-emphasized voice signals of each path, perform time domain correction on the voice signals of each path except the first voice signal, so that the voice signals of each path except the first voice signal are aligned with the first voice signal in the time domain, realize the time synchronization of the voice signals of each path, and thus obtain the synchronized voice signals of each path. The expressions are as follows:

[0047] ;

[0048] wherein, is the th voice signal after time domain correction, is the th voice signal after time domain correction, The th path source signal after time-domain correction.

[0049] Step S3: Perform MFCC feature extraction on each synchronized voice signal to obtain an MFCC feature matrix. Each row of the MFCC feature matrix stores the MFCC features of each dimension of one voice signal.

[0050] Step S3 specifically includes:

[0051] Step S3.1: Send the obtained after synchronization to a data acquisition card for analog-to-digital conversion. In this embodiment, the sampling rate ≥ 16 kHz and the resolution is 16 bit;

[0052] Step S3.2: Perform frame segmentation on the multiplexed voice signals after analog-to-digital conversion;

[0053] Step S3.3: Perform windowing on the multiplexed voice signals after frame segmentation to obtain windowed frame signals , where is the frame index,

[0054] is the sampling point index within the time domain frame. In this embodiment, the frame length is set to 25 ms, the frame shift is 15 ms, and the window function uses a Hamming window;

[0054] Step S3.4: Perform a fast Fourier transform on the windowed frame signal to obtain a spectrum , where

[0055] is the index of the frequency component in the frequency domain;

[0055] Step S3.5: Take the square of the spectrum as the energy spectrum and input it into a Mel filter bank to obtain the logarithmic energy spectrum output by the Mel filter bank of the th path voice signal;

[0056] Step S3.6: Perform a discrete cosine transform on the logarithmic energy spectrum output by the Mel filter bank of the th path voice signal to obtain the MFCC features of the th path voice signal from dimension 0 to dimension . Take the logarithm of the time domain energy of the frame signal as the th dimension MFCC feature of the th path voice signal, thereby forming an MFCC feature vector of dimension for the th path voice signal , and further form an MFCC feature matrix , , where is the total number of dimensions of the MFCC features. In this embodiment, take 13; the feature vector in, the th dimensional MFCC feature of the th

[0057] Step S4: Independently normalize the MFCC features of each dimension of each channel of voice signal in the MFCC feature matrix, initialize the MFCC features of each dimension of each normalized channel of voice signal, and map them to quantum states. Allocate a number of qubits for the MFCC features of each dimension of each channel of voice signal, construct a cross-layer quantum entanglement circuit, and then input each qubit into the cross-layer quantum entanglement circuit to obtain the final quantum state. Calculate the expectation value of the Pauli-Z operator of each qubit in the final quantum state, and then obtain the features after quantum feature enhancement.

[0058] The quantum-inspired feature encoding technology effectively enhances the expression ability of fault features by mapping the MFCC features of each channel to quantum states and performing parametric entanglement transformation, and solves the problems of noise sensitivity and lack of non-linear correlation between features existing in traditional methods such as MFCC.

[0059] Step S4 specifically includes:

[0060] Step S4.1: Independently perform quantile rank normalization on the MFCC features of each channel of voice signal in the MFCC feature matrix, and then take the inverse function of the standard normal distribution for the result to convert the uniform distribution into a Gaussian distribution to eliminate the dimensional difference of different MFCC dimensions:

[0061] ;

[0062] where, is the th dimensional MFCC feature of the th channel of voice signal after normalization, is the inverse function of the standard normal distribution, is the sorting position of in ascending order of numerical value among all training samples,

[0063] is the total number of training samples. In this embodiment, the total number of training samples is expanded to 10,000 through data augmentation to cover various fault types and normal working states.

[0063] Step S4.2: Apply a rotation gate to the normalized feature to initialize the quantum state, and Mapped to the coherent superposition state of qubits, with each dimension of MFCC features assigned qubits, obtaining the initialized qubits:

[0064] ;

[0065] Among them, is the quantum state of the th dimensional MFCC feature of the th qubit of the th voice signal at the 0th layer of the quantum circuit. The quantum state of the qubit at the 0th layer is the initialized qubit; is the total number of qubits assigned to each dimension of MFCC features. In this embodiment, = 5; is the rotation gate around the Y axis, indicating that through a fixed scaling factor maps from to to cover the complete superposition state of the qubits; is the rotation gate rotating around the Z axis, introducing the phase freedom, so that the quantum state can represent more complex feature correlations in the complex space; is the dimensional MFCC feature of the th qubit of the

[0066] th voice signal at the 0th layer. The phase parameter at the 0th layer is the initialized phase parameter, which is used to dynamically adjust the phase characteristics of the quantum state and is initialized according to the following rules to avoid all quantum states from collapsing to the same direction:

[0067] Step S4.3, construct a cross-layer entangled quantum circuit with a total number of layers of ;

[0068] Step S4.4, input the initialized qubits into the cross-layer entangled quantum circuit to obtain the final quantum state:

[0069] ;

[0070] Among them, is the final quantum state, is the product symbol; is the initial quantum state of the quantum circuit, which is the initial state of all qubits before any operation after initialization, , is the tensor product symbol; is the Intra-layer entanglement controlled NOT gate, is an inter-layer entanglement controlled NOT gate; is the th dimensional MFCC feature of the th qubit of the th layer, used to dynamically adjust the amplitude of the quantum state (through backpropagation training); is the th dimensional MFCC feature of the th qubit of the th layer, the phase parameter of the

[0071] Step S4.5, calculate the expectation value of the Pauli-Z operator for each qubit, and the total number of theoretical dimensions of the generated features is , compressed to dimensions through principal component analysis to adapt to the input of the downstream network, and the features after quantum feature enhancement are as follows:

[0072] ;

[0073] ;

[0074] Among them, is the principal component analysis; is the union symbol, which concatenates the expectation values of each channel, dimension, and qubit into a high-dimensional vector; is the th dimensional MFCC feature of the th qubit of the is the final quantum state under the action of the operator expectation value; is a 2×2 identity matrix.

[0075] Among them, in this embodiment, when constructing an inter-layer entanglement quantum circuit with a total number of layers of , the total number of layers of the quantum circuit is determined by 5-fold cross-validation, and the value ranges from 2 to 5. In this example, 3 layers are preferably selected. The construction process of the inter-layer entanglement quantum circuit includes three steps: intra-layer entanglement, inter-layer entanglement, and rotation gate operation. Specifically, step S4.3 specifically includes:

[0076] Step S4.3.1, intra-layer entanglement, apply a chained controlled NOT gate between the qubits of each dimensional MFCC feature within the same layer:

[0077] ;

[0078] Among them, CNOT represents a controlled NOT gate. is the position of the -th qubit of the -dimensional MFCC feature of the -th sound signal at the layer. is the position of the -th qubit of the -dimensional MFCC feature of the

[0079] Step S4.3.2: Since the first qubit of each layer transmits the main information of the feature, cross-layer entanglement is applied only between the first qubits of the layer and the layer, reducing the computational complexity while retaining cross-layer information transmission:

[0080] ;

[0081] Among them, is the position of the first qubit of the -dimensional MFCC feature of the -th sound signal at the layer, is the position of the first qubit of the -dimensional MFCC feature of the

[0082] Step S4.3.3: Apply trainable rotation gates to each layer:

[0083] ;

[0084] Among them is the quantum state of the -th qubit of the -dimensional MFCC feature of the -th sound signal at the layer, is the quantum state of the -th qubit of the -dimensional MFCC feature of the

[0085] Step S5: Input the features enhanced by quantum features into the CNN-LSTM hybrid network for training, and use the trained CNN-LSTM hybrid network to diagnose the faults of the transformer.

[0086] In this embodiment, the CNN-LSTM hybrid network includes: an input layer, a first convolutional layer (64 3×3 convolutional kernels, stride = 1), a max pooling layer (pooling window = 2×2, stride = 2), a second convolutional layer (128 3×3 convolutional kernels, stride = 1), a time series reshaping layer (converting the convolutional output into a time step series), a bidirectional LSTM layer, an attention layer, a fully connected layer (128 neurons, activation function ReLU), and a dual-task output layer.

[0087] During the model training process, the Adam optimizer is adopted, and hyperparameters such as the learning rate, batch size, loss weight, number of LSTM units, and Dropout ratio of the fully connected layer are optimized through grid search and cross-validation.

[0088] In summary, according to the transformer fault real-time diagnosis method of the above embodiment, the following beneficial effects are achieved:

[0089] (1) The present invention uses the delay estimation method based on wavelet denoising and quadratic cross-correlation to eliminate the time difference between various sound signals, thereby realizing the time synchronization of various sound signals, significantly improving the accuracy of voiceprint feature extraction, and effectively suppressing environmental noise at the same time.

[0090] (2) In the present invention, each sound signal is independently processed in the pre-emphasis, time synchronization, MFCC feature extraction, and quantum coding stages, retaining the specific fault information of each sound signal, improving the flexibility of quantum feature enhancement, and being able to provide better input features for the CNN-LSTM hybrid network.

[0091] (3) On the basis of MFCC feature extraction, the present invention further uses quantum feature coding technology. By mapping the MFCC features of each channel to the quantum state and performing parametric entanglement transformation, the expression ability of fault features is effectively enhanced, the noise sensitivity of MFCC features is reduced, the nonlinear correlation between intra-channel features and inter-channel features is fully mined, the problems of high noise sensitivity and lack of nonlinear correlation between features existing in traditional methods are solved, the accuracy of fault diagnosis results can be improved, and the diagnostic false alarm rate can be reduced.

[0092] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A real-time diagnosis method for transformer faults, characterized in that Including: Step S1: Use a number of microphone sensors to perform real-time multi-channel voiceprint acquisition on the transformer, and perform pre-emphasis processing on each path of voice signal collected by the microphone sensors to obtain the pre-emphasized voice signal. Step S2: Use the first path of voice signal in the pre-emphasized voice signal as the reference signal, and use the delay estimation method based on wavelet denoising and quadratic cross-correlation to estimate the time delay estimation values between each path of voice signal except the first path of voice signal and the first path of voice signal respectively, and obtain the synchronized voice signals based on the time delay estimation values. Step S3: Perform MFCC feature extraction on each path of the synchronized voice signals respectively to obtain an MFCC feature matrix, and each row of the MFCC feature matrix stores the MFCC features of each dimension of one path of voice signal. Step S4: Independently normalize the MFCC features of each dimension of each path of voice signal in the MFCC feature matrix, initialize the MFCC features of each dimension of each path of voice signal after normalization, and map them to quantum states, allocate a number of qubits for the MFCC features of each dimension of each path of voice signal, construct a cross-layer quantum entanglement circuit, and then input each qubit into the cross-layer quantum entanglement circuit to obtain the final quantum state, calculate the expectation value of the Pauli-Z operator of each qubit in the final quantum state, and thus obtain the features after quantum feature enhancement. Step S5: Input the features after quantum feature enhancement into the CNN-LSTM hybrid network and perform training, and use the trained CNN-LSTM hybrid network to perform fault diagnosis on the transformer.

2. The real-time transformer fault diagnosis method according to claim 1, wherein Step S1 satisfies the following formula: ； ； Among them, represents the first voice signal after pre-emphasis processing, represents the th voice signal after pre-emphasis processing, , represents the total number of channels, is the first source signal, is the first noise signal, is the th source signal without time-domain correction, is the th noise signal, is time, the time delay between the th voice signal and the first voice signal.

3. The real-time transformer fault diagnosis method according to claim 2, wherein Step S2 specifically includes: Step S2.1, solve the autocorrelation function of the first audio signal and the cross-correlation function between the first audio signal and the audio signal of the other path except the first audio signal; Step S2.2, use wavelet transform to decompose and denoise the autocorrelation function of the first audio signal, the first audio signal, and the cross-correlation function of the first audio signal and the second audio signal respectively, to obtain the decomposed and denoised autocorrelation function and cross-correlation function; Step S2.3, calculate the quadratic cross-correlation function of the decomposed and denoised autocorrelation function and cross-correlation function to obtain the quadratic cross-correlation function of the first sound signal and the second sound signal; Step S2.4, perform peak search on the second-order cross-correlation function of the first voice signal and the voice signal of the path, and use the time delay corresponding to the peak of the second-order cross-correlation function of the first voice signal and the voice signal of the path as the time delay estimation value of the first voice signal and the Step S2.5, substitute the time delay estimation values of the first audio signal and the audio signal of the other path into the expressions of the pre-emphasized audio signals of each path, and perform time domain correction on the audio signals of each path except the first audio signal, so that the audio signals of each path except the first audio signal are aligned with the first audio signal in the time domain, realizing the time synchronization of the audio signals of each path, and thus obtaining the synchronized audio signals of each path.

4. The real-time transformer fault diagnosis method according to claim 3, wherein Step S2.2 satisfies the following formula: ； ； ； ； Among them, is the autocorrelation function of the first-channel sound signal after wavelet denoising, is the time shift factor, is 's approximation coefficient, is the scaling function of the low-frequency band, is the scale factor, is the scaling function, is 's detail coefficient, is the wavelet function of the high-frequency band, is the wavelet function, is the cross-correlation function between the first-channel sound signal and the -th channel sound signal after wavelet denoising, is 's approximation coefficient, is 's detail coefficient.

5. The real-time transformer fault diagnosis method according to claim 4, characterized in that, Step S2.3 satisfies the following formula: ； Among them, is the quadratic cross-correlation function of the first audio signal and the th audio signal, is the desired operation, is the time-shifted quadratic cross-correlation function of the first audio signal and the th audio signal.

6. The real-time transformer fault diagnosis method according to claim 5, characterized in that Step S2.5 satisfies the following formula: ； Among them, is the th voice signal after time-domain correction, is the th voice signal after time-domain correction, is the th source signal after time-domain correction.

7. The real-time transformer fault diagnosis method according to claim 6, characterized in that, Step S4 specifically includes: Step S4.1, respectively perform percentile rank normalization on each MFCC feature of each voice signal in the MFCC feature matrix, and then take the inverse function of the standard normal distribution for the result to convert the uniform distribution into a Gaussian distribution. The expression is as follows: MFCC features are independently normalized, and then the inverse function of the standard normal distribution is taken for the result to convert the uniform distribution into a Gaussian distribution. The expression is: ； Among them, is the -th -dimensional MFCC feature of the -th normalized voice signal, is the inverse function of the standard normal distribution, is the sorting position arranged in ascending order of numerical values among all training samples, is the total number of training samples; Step S4.2, for the normalized features apply a rotation gate to initialize the quantum state, and map to a coherent superposition state of qubits, assign qubits to each dimension of MFCC features, and obtain the initialized qubits. The expression is: ； Among them, is the th -dimensional MFCC feature th qubit's quantum state at the 0th layer of the quantum circuit. The quantum state of the qubit at the 0th layer is the initialized qubit; is the rotation gate around the Y-axis, represents that through a fixed scaling factor the is mapped from to ; is the rotation gate rotating around the Z-axis; is the th -dimensional MFCC feature th qubit's phase parameter at the 0th layer, ; Step S4.3, construct a cross-layer entangled quantum circuit with a total number of layers being ; Step S4.4: Input the initialized qubits into the cross-layer entangled quantum circuit to obtain the final quantum state, and the expression is: ； Among them, is the final quantum state, is the product symbol; is the initial quantum state of the quantum circuit, which is the initial state of all qubits before any operation after initialization, , is the tensor product symbol; is the entanglement-controlled NOT gate within the th layer, is the cross-layer entanglement-controlled NOT gate; is the th angle parameter of the th-dimensional MFCC feature of the th qubit of the th path of voice signal in the is the th phase parameter of the th-dimensional MFCC feature of the th qubit of the th path of voice signal, is the total number of qubits assigned to each-dimensional MFCC feature; Step S4.5, calculate the expectation value of the Pauli-Z operator for each qubit, and the total number of theoretical dimensions of the generated features is , compressed to dimensions through principal component analysis, and the features after quantum feature enhancement are as follows: ； ； Among them, is the principal component analysis; is the union symbol, which concatenates the expected values of each channel, dimension, and qubit into a high-dimensional vector; is the th dimensional MFCC feature of the th Pauli-Z measurement operator of the qubit, is the final quantum state under the action of the operator expected value; is the 2×2 identity matrix.

8. The real-time transformer fault diagnosis method according to claim 7, characterized in that Step S4.3 specifically includes: Step S4.3.1, intra-layer entanglement, apply chained controlled-NOT gates between qubits for each dimension of MFCC features within the same layer, and the expression is: qubits, and the expression is: ； Among them, CNOT represents a controlled NOT gate, is the th -dimensional MFCC feature of the th qubit at the th layer, is the th -dimensional MFCC feature of the th qubit at the th layer; Step S4.3.2, cross-layer entanglement. Apply a cross-controlled NOT gate between the first qubits of the layer and the layer. The expression is: ； Among them, is the position of the first qubit of the -dimensional MFCC feature of the th path sound signal at the layer, is the position of the first qubit of the -dimensional MFCC feature of the th path sound signal at the layer; Step S4.3.3: Apply a trainable rotation gate, and the expression is: ； where is the -th quantum state of the -dimensional MFCC feature of the -th qubit at the -th layer, is the -th quantum state of the -dimensional MFCC feature of the -th qubit at the -th layer.

Images