A method for fault diagnosis of rail transit transformer

Abstract

The present invention discloses a rail transit transformer fault diagnosis method, comprising obtaining noise signals emitted by the rail transit transformer under various operating conditions; using wavelet thresholding for denoising; recursively transferring linear predictive coding to the cepstral domain to obtain linear predictive cepstral coefficients; using fast Fourier transform to obtain a spectrum graph to extract Mel-type cepstral coefficients, and combining the Mel-type cepstral coefficients with first-order difference coefficients and second-order difference coefficients to obtain optimized Mel-type cepstral coefficients; combining the characteristic parameter linear predictive cepstral coefficients and the optimized Mel-type cepstral coefficients to obtain a feature set; further performing feature learning on the feature set using a deep learning architecture to train and establish a rail transit transformer fault identification model; and using the trained rail transit transformer fault identification model to detect rail transit transformer noise. The present invention can quickly, accurately, and effectively identify rail transit transformer faults.

Description

Technical Field

[0001] The present invention belongs to the technical field of rail transportation, and in particular relates to a rail transportation transformer fault diagnosis method. Background Art

[0002] In recent years, with the continuous development of rail transit systems, the safe and stable operation of rail transit power systems is an important aspect, and the requirements for it are becoming increasingly stringent. Rail transit transformers have the important functions of voltage conversion and power distribution, and are key core equipment of rail transit power systems. Therefore, the operating conditions of transformers affect the safety of the power system. Therefore, effective diagnostic monitoring technology is of great significance to the safe and reliable operation of the entire rail transit power system. Due to the characteristics of many types of transformers, complex internal structures and diverse operating conditions, it is difficult to detect fault anomalies. Intelligent diagnosis of transformer faults based on noise signal analysis is a relatively effective non-invasive monitoring method. Its non-contact signal acquisition method will not have a negative impact on the normal operation of the transformer, and can accurately and effectively identify the working status of the transformer.

[0003] Machine learning algorithms are a popular branch of pattern recognition, and deep learning-based fault diagnosis and anomaly recognition algorithms are a hot topic within the field of identification and classification. In recent years, they have been extensively applied to transformer fault diagnosis models. However, when it comes to rail transit transformer identification, existing transformer fault diagnosis models suffer from some problems: some are too large to improve accuracy, resulting in slow recognition speed, while others suffer from poor recognition accuracy. Consequently, existing rail transit transformer identification models are unable to improve recognition speed while maintaining accuracy. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a rail transit transformer fault diagnosis method, which can quickly and effectively identify rail transit transformer faults.

[0005] To achieve the above object, the present invention adopts a technical solution: a rail transit transformer fault diagnosis method, comprising the steps of:

[0006] S1: Obtain the noise signals emitted by rail transit transformers under various operating conditions;

[0007] S2: Preprocess the transformer noise signal and use wavelet threshold denoising to filter and denoise the noise signal;

[0008] S3: Using the Durbin algorithm to obtain linear prediction coding for the preprocessed transformer noise signal data, and then recursively extrapolating the obtained linear prediction coding to the cepstral domain to obtain linear prediction cepstral coefficients;

[0009] S4: For the preprocessed transformer noise signal data, use fast Fourier transform to obtain a spectrum, perform Mel filtering on the spectrum, calculate the logarithm of the filtered data, perform discrete cosine transform to extract Mel cepstral coefficients, and combine the Mel cepstral coefficients with the first-order difference coefficients and the second-order difference coefficients to obtain optimized Mel cepstral coefficients;

[0010] S5: combining the linear prediction cepstral coefficients of the feature parameters and the optimized Mel cepstral coefficients to obtain a feature set;

[0011] S6: Use the deep learning structure to further learn the feature set and train and establish a rail transit transformer fault recognition model;

[0012] S7: Use the trained rail transit transformer fault identification model to detect rail transit transformer noise, use the Softmax loss function to perform the final working condition classification, and then achieve the purpose of rail transit transformer fault diagnosis.

[0013] Furthermore, in step S1, the noise signals {s0(n)}, 1≤n≤N of the rail transit transformer collected by the same pickup device include four working conditions: normal operation, overload operation, discharge and DC bias magnetization. s0 , N s0 is the length of the signal {s0(n)}.

[0014] Furthermore, in step S2, the transformer noise signal is preprocessed, and the noise signal is filtered and denoised using wavelet threshold denoising, including the steps of:

[0015] S21: Decompose the noise signal using a wavelet transform with a wavelet base of sym8 and a decomposition layer of 2, and then obtain two decomposition value components related to the resolution; set the threshold processing Fs, the threshold selection rule is the unbiased likelihood estimation threshold, the threshold adjustment relies on the transformer noise signal level estimation obtained by the wavelet decomposition of each layer, and the threshold function is a hard threshold function;

[0016] When threshold processing is performed, the approximate component is retained, and for the detail component, a threshold is set. If the value of the wavelet coefficient is smaller than the set threshold, it is set to zero. If the value of the wavelet coefficient is larger than the set threshold, it is saved.

[0017] S22: A first-order FIR high-pass digital filter is used to pre-emphasize the noise signal {s0(n)} to enhance the high-frequency portion and flatten the signal spectrum, resulting in the pre-emphasized signal {s(n)}:

[0018] s(n)=s0(n+1)-μ*s0(n),1≤n≤(N s0 -1);

[0019] Where, s0(n) is the nth data of the noise signal {s0(n)}, s0(n+1) is the n+1th data of the noise signal {s0(n)}, and the value range of μ is 0.93~0.97;

[0020] S23: Frame the pre-emphasized noise signal to ensure the time resolution of the noise signal analysis; let the nth frame data after framing be x(n), and the time length corresponding to one frame be N / transformer noise acquisition frequency, where N is the length of one frame signal;

[0021] S24: Apply a window function to the framed noise data to reduce spectral leakage; set the window function to a Hamming window, the independent variable is p, the dependent variable w(p) is the amplitude of the window function, P is a constant, and the total length of the window function is L=P+1, where L is equal to the frame length N; the expression of the window function is:

[0022] w(p)=0.54-0.46*cos(2*π*p / P),0≤p≤P;

[0023] In order to avoid attenuation of data at the edge of the window function, the data frames should be overlapped.

[0024] Furthermore, in step S3: the pre-processed transformer noise signal data is subjected to a Durbin algorithm to obtain linear prediction coding, and the obtained linear prediction coding is then recursively transferred to the cepstral domain to obtain linear prediction cepstral coefficients, including the following steps:

[0025] S31: for the pre-processed transformer noise signal data, the output signal of the noise signal data n is expressed by a linear combination of the previous m samples, and is derived by linear prediction analysis coefficients of various orders to form an equation group consisting of m equations;

[0026] S32: solving the equations using the Durbin algorithm to obtain linear predictive coding;

[0027] S33: Perform fast Fourier transform calculation on the linear prediction coding, then perform logarithmic operation, and finally obtain the linear prediction cepstral coefficients through inverse fast Fourier transform operation.

[0028] Furthermore, in step S4: the preprocessed transformer noise signal data is subjected to fast Fourier transform to obtain a spectrum, the spectrum is subjected to Mel filtering, the filtered data is logarithmized, a discrete cosine transform is performed to extract Mel cepstral coefficients, and the Mel cepstral coefficients are combined with the first-order difference coefficients and the second-order difference coefficients to obtain optimized Mel cepstral coefficients, including the following steps:

[0029] S41: performing an N-point fast Fourier transform on each frame of the preprocessed transformer noise signal data, converting the noise signal from the time domain to a spectrum, and obtaining a spectrum diagram;

[0030] S42: setting a Mel triangle filter bank, filtering the spectrum graph using the Mel triangle filter bank, and calculating the logarithm of the filtered data;

[0031] S43: Use discrete cosine transform to transform the logarithmic Mel spectrum into the time domain and extract the Mel cepstral coefficients;

[0032] S44: Calculate the first-order difference coefficient and the second-order difference coefficient of the Mel-cepstral coefficient, and linearly combine the Mel-cepstral coefficient with the first-order difference coefficient and the second-order difference coefficient to obtain the optimized Mel-cepstral coefficient.

[0033] Furthermore, the calculation formulas for the first-order difference coefficients and the second-order difference coefficients of the Mel-frequency cepstral coefficients are as follows:

[0034]

[0035] d t represents the tth first-order difference, C t represents the tth Mel cepstral coefficient, Q represents the order of the cepstral coefficient; K represents the time difference of the first-order derivative, which is 1 or 2, and k is the cumulative coefficient from 1 to K;

[0036] Substitute the Mel cepstral coefficients into the above formula to obtain the same number of first-order difference coefficients d t ;

[0037] Then we can use the d we just obtained t As C t Substituting into the above formula, we can obtain the second-order difference coefficient.

[0038] Furthermore, in step S5, the feature parameter linear prediction cepstral coefficients and the optimized Mel cepstral coefficients are combined to obtain a feature set, including the steps of:

[0039] S51: Obtain the linear prediction cepstral coefficient and the optimized Mel cepstral coefficient of the characteristic parameters of the track transformer noise signal of each frame, combine the characteristic parameter sequences of the two in the dimension of each frame, construct a combined characteristic parameter sequence, and obtain the combined characteristic parameters;

[0040] S52: The length of the combined feature parameter sequence is fixed to 499, the part exceeding 499 is truncated, and the part less than 499 is supplemented with 0, and the final feature set is input into the network model.

[0041] Furthermore, in step S6, the feature set is further subjected to feature learning using a deep learning structure to train and establish a rail transit transformer fault recognition model, including the steps of:

[0042] S61: After the feature set is input into the network model, it enters the 3-layer BiLSTM layer. After the 3-layer BiLSTM layer, a Dropout layer is added to prevent overfitting of the network training. Dropout randomly discards some neurons according to a set probability, so that the number of input and output neurons of each layer remains consistent during the training parameter iteration process of the transformer noise signal feature information. In order to be subsequently input into the convolution block, the parameters are reshaped into three dimensions through the reshape layer.

[0043] S62: Convolutional neural network is constructed by connecting two convolution blocks in series. Each convolution block includes a two-dimensional convolution layer, a nonlinear function activation layer, a maximum pooling layer, a batch normalization layer and a Dropout layer. The two-dimensional convolution layer, the nonlinear function activation layer and the maximum pooling layer are sequentially passed, and the batch normalization layer and the Dropout layer are set at the end. In order to enhance the generalization performance of the model, the L2 regularization method is used for the weights of each layer.

[0044] Furthermore, the first convolution block contains two two-dimensional convolution layers, Conv1 and 2, using 16 large 5×5 convolution kernels with a moving step of 1. The first convolution block takes 1 channel as input, outputs 16 channels, and uses the ReLU function for nonlinear fitting; then a maximum pooling layer of size 2×2 is set to reduce the feature size, outputting the first feature set.

[0045] Furthermore, the second convolution block contains two two-dimensional convolution layers, Conv3 and 4, using 32 large 5×5 convolution kernels with a moving step of 1. The second convolution block takes 16 channels as input, outputs 32 channels, and uses the ReLU function for nonlinear fitting. Subsequently, a maximum pooling layer of size 2×2 is set to output the second feature set.

[0046] The beneficial effects of adopting this technical solution are:

[0047] The present invention obtains a transformer noise signal and performs preprocessing on it, including wavelet threshold denoising; performs feature extraction on the collected transformer noise signal, obtains characteristic parameters of the noise signal under different working conditions, and combines the linear prediction cepstral coefficients and optimized Mel cepstral coefficients to form a feature set; further features the feature set using a deep learning structure model established for combining the characteristic parameters of the linear prediction cepstral coefficients and the optimized Mel cepstral coefficients; and uses the Softmax loss function to perform final transformer working condition classification, thereby realizing the function of rapid and effective diagnosis and identification of rail transit transformer faults. The present invention can fully extract the cepstral features of the transformer noise signal in the presence of interfering noise, and then use the deep learning algorithm three-layer BiLSTM and dual-series convolutional block neural convolution hybrid network structure to realize fault diagnosis, which can reduce detection costs while ensuring the accuracy and effectiveness of detection.

[0048] The wavelet threshold denoising method established in the present invention filters and denoises the collected rail transit transformer noise signal. Based on the three optimal parameters of wavelet threshold denoising wavelet basis, number of decomposition layers and threshold selection rule, the interference noise is filtered out to the greatest extent to achieve better feature extraction effect.

[0049] The present invention uses a combination of characteristic parameter linear prediction cepstral coefficients and optimized Mel cepstral coefficients as a feature vector input to a classification and recognition module. This feature contains more time domain information and frequency domain information, which is helpful for the final working condition classification.

[0050] The proposed hybrid neural convolutional network architecture, comprised of a three-layer BiLSTM and dual concatenated convolutional blocks, allows the model to better learn and extract features from preprocessed noise sequences, enabling rail transit transformer operating condition classification and fault diagnosis. This makes the network highly suitable for identification and classification tasks within the noise analysis-based intelligent fault diagnosis model for traction transformers. Its feature extraction capabilities are demonstrated in two ways: First, the two concatenated convolutional blocks can extract the energy distribution within the two-dimensional array features of the linearly predicted cepstral coefficients and optimized Mel-cepstral coefficients of the traction transformer noise signal parameters, which are crucial for distinguishing between different transformer operating states. Second, the two concatenated convolutional blocks can fully learn the local time-frequency characteristics of the linearly predicted cepstral coefficients and optimized Mel-cepstral coefficients of the two-dimensional array features of the transformer noise signal parameters under different operating conditions, thereby obtaining the local time-frequency characteristics inherent therein. This characteristic enables more precise differentiation of noise signals generated by different transformer operating conditions. The learned feature information represents an abstract summary of the input transformer noise signal characteristics, expressing the common attributes of different noise samples under the same transformer operating condition and including discriminative information between noise signals under different transformer operating conditions. However, transformer noise signals are sequence data, and their characteristic parameters exhibit long-term dependencies between frames in the temporal dimension. This relationship contains a wealth of important information that needs to be extracted and learned. Convolutional networks have limited ability to extract feature information from previous and subsequent sequences on the temporal scale of transformer noise signals, and are unable to characterize long-term dependencies between frames. Therefore, an improved three-layer BiLSTM layer was considered. This solves the problem of vanishing or exploding gradients. Furthermore, due to its built-in state memory unit, the three-layer BiLSTM is capable of learning long-term dependencies between sequence feature information, enabling a better representation of inter-frame dependencies and achieving greater robustness.

[0051] The rail transit transformer fault diagnosis method provided by the present invention is based on wavelet threshold denoising and deep learning combined linear prediction cepstral coefficients and optimized Mel cepstral coefficient features. It can fully extract the time domain information and frequency domain information characteristics of the transformer noise signal in the presence of interference noise, and then use the deep learning algorithm to realize the classification of transformer operating conditions including normal operation, overload operation, discharge and DC bias magnetization. It has a high recognition rate and good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of a rail transit transformer fault diagnosis method according to the present invention;

[0053] Figure 2 Schematic diagram of the confusion matrix of the experimental results of the embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings.

[0055] In this embodiment, see Figure 1 As shown, the present invention proposes a rail transit transformer fault diagnosis method, comprising the steps of:

[0056] S1: Obtain the noise signals emitted by rail transit transformers under various operating conditions;

[0057] In step S1, the noise signals {s0(n)}, 1≤n≤N, of the rail transit transformer collected by the same pickup device include four working conditions: normal operation, overload operation, discharge and DC bias magnetization. s0 , N s0 is the length of the signal {s0(n)}. The sampling noise parameters are: 8kHz sampling rate and 16-bit digital quantization. The aforementioned data acquisition environment and conditions should be consistent. The pickup should be mounted near the low-voltage bushing, facing the side of the oil tank of each phase winding. A horizontal position at half the height of the upper and lower ends facilitates noise signal acquisition.

[0058] S2: Preprocess the transformer noise signal and use wavelet threshold denoising to filter and denoise the noise signal;

[0059] The preprocessing process includes sampling quantization, filtering denoising, pre-emphasis and windowing and framing.

[0060] In step S2, the transformer noise signal is preprocessed and the noise signal is filtered and denoised using wavelet threshold denoising, including the following steps:

[0061] S21: Decompose the noise signal using a wavelet transform with a wavelet base of sym8 and a decomposition layer of 2, and then obtain two decomposition value components related to the resolution; set the threshold processing Fs, the threshold selection rule is the unbiased likelihood estimation threshold, the threshold adjustment relies on the transformer noise signal level estimation obtained by the wavelet decomposition of each layer, and the threshold function is a hard threshold function;

[0062] When threshold processing is performed, the approximate component is retained, and for the detail component, a threshold is set. If the value of the wavelet coefficient is smaller than the set threshold, it is set to zero. If the value of the wavelet coefficient is larger than the set threshold, it is saved.

[0063] S22: A first-order FIR high-pass digital filter is used to pre-emphasize the noise signal {s0(n)} to enhance the high-frequency portion and flatten the signal spectrum, resulting in the pre-emphasized signal {s(n)}:

[0064] s(n)=s0(n+1)-μ*s0(n),1≤n≤(Ns0 -1);

[0065] Where s0(n) is the nth data of the noise signal {s0(n)}, s0(n+1) is the n+1th data of the noise signal {s0(n)}, and the value of μ is 0.97;

[0066] S23: Framing the pre-emphasized noise signal to ensure the time resolution of the noise signal analysis; assuming that the nth frame of data after framing is x(n), the time length corresponding to one frame is N / transformer noise acquisition frequency, where N is the length of one frame signal. In the present invention, the time length corresponding to one frame is 10ms;

[0067] S24: Apply a window function to the framed noise data to reduce spectral leakage; set the window function to a Hamming window, the independent variable is p, the dependent variable w(p) is the amplitude of the window function, P is a constant, and the total length of the window function is L=P+1, where L is equal to the frame length N; the expression of the window function is:

[0068] w(p)=0.54-0.46*cos(2*π*p / P),0≤p≤P;

[0069] In order to avoid attenuation of the edge data of the window function, the data frames should be overlapped. In this embodiment, the overlap rate is 50%.

[0070] S3: Using the Durbin algorithm to obtain linear prediction coding for the preprocessed transformer noise signal data, and then recursively extrapolating the obtained linear prediction coding to the cepstral domain to obtain linear prediction cepstral coefficients;

[0071] In step S3: the pre-processed transformer noise signal data is subjected to the Durbin algorithm to obtain linear prediction coding, and the obtained linear prediction coding is then recursively transferred to the cepstral domain to obtain linear prediction cepstral coefficients, including the following steps:

[0072] S31: for the pre-processed transformer noise signal data, the output signal of the noise signal data n is expressed by a linear combination of the previous m samples, and is derived by linear prediction analysis coefficients of various orders to form an equation group consisting of m equations;

[0073] S32: solving the equations using the Durbin algorithm to obtain linear predictive coding;

[0074] S33: Perform fast Fourier transform calculation on the linear prediction coding, then perform logarithmic operation, and finally obtain the linear prediction cepstral coefficients through inverse fast Fourier transform operation.

[0075] In this example, the Durbin algorithm is used to obtain linear prediction codes (LPCs), with an LPC order of 10. When performing cepstral calculations on LPCs, the LPC coefficient order is 39. To ensure consistent training sample dimensions, padding is performed on the extracted features. The LPC coefficient input model has a feature dimension of 499 × 39, where the input width (499) represents the temporal component (number of frames) of the noise signal, and the input height (39) represents the dimensional component of the LPC coefficients.

[0076] S4: For the preprocessed transformer noise signal data, use fast Fourier transform to obtain a spectrum, perform Mel filtering on the spectrum, calculate the logarithm of the filtered data, perform discrete cosine transform to extract Mel cepstral coefficients, and combine the Mel cepstral coefficients with the first-order difference coefficients and the second-order difference coefficients to obtain optimized Mel cepstral coefficients;

[0077] In step S4: the preprocessed transformer noise signal data is subjected to fast Fourier transform to obtain a spectrum, the spectrum is subjected to Mel filtering, the filtered data is logarithmized, discrete cosine transform is performed to extract Mel cepstral coefficients, and the Mel cepstral coefficients are combined with first-order difference coefficients and second-order difference coefficients to obtain optimized Mel cepstral coefficients, including the following steps:

[0078] S41: performing an N-point fast Fourier transform on each frame of the preprocessed transformer noise signal data, converting the noise signal from the time domain to a spectrum, and obtaining a spectrum diagram;

[0079] S42: setting a Mel triangle filter bank, filtering the spectrum graph using the Mel triangle filter bank, and calculating the logarithm of the filtered data;

[0080] S43: Use discrete cosine transform to transform the logarithmic Mel spectrum into the time domain and extract the Mel cepstral coefficients;

[0081] S44: Calculate the first-order difference coefficient and the second-order difference coefficient of the Mel-cepstral coefficient, and linearly combine the Mel-cepstral coefficient with the first-order difference coefficient and the second-order difference coefficient to obtain the optimized Mel-cepstral coefficient.

[0082] Preferably, the calculation formula of the first-order difference coefficient and the second-order difference coefficient of the Mel-frequency cepstral coefficient is:

[0083]

[0084] d t represents the tth first-order difference, C t represents the tth Mel cepstral coefficient, Q represents the order of the cepstral coefficient; K represents the time difference of the first-order derivative, which is 1 or 2, and k is the cumulative coefficient from 1 to K;

[0085] Substitute the Mel cepstral coefficients into the above formula to obtain the same number of first-order difference coefficients d t ;

[0086] Then we can use the d we just obtained t As C t Substituting into the above formula, we can obtain the second-order difference coefficient.

[0087] The transformer noise signal changes continuously, and each frame is not independent. The standard Mel-cepstral coefficient only reflects the static characteristics of the noise parameters, and cannot reflect the dynamic characteristics of the transformer noise signal. The present invention effectively improves the recognition performance of the system by combining dynamic and static features. The standard Mel-cepstral coefficient of the rail transit transformer noise signal is combined with the first-order difference coefficient and the second-order difference coefficient to obtain the optimized Mel-cepstral coefficient. The optimized Mel-cepstral coefficient obtained in this way can better reflect the characteristics of the noise signal, which is conducive to more accurate extraction of fault information in the later network model. The combination of dynamic and static features can effectively improve the performance of the subsequent classification and recognition system.

[0088] In this example, the Mel filter order is 24, the FFT transform length is 256, and to optimize the Mel cepstral coefficient features, the standard Mel cepstral coefficients are linearly combined with the first-order and second-order difference coefficients to obtain 39-dimensional optimized Mel cepstral coefficients. To ensure consistent training sample dimensions, the extracted features are padded. The feature dimension of the optimized Mel cepstral coefficient input model is 499×39, where the input width of 499 is the time component (number of frames) of the noise signal, and the input height of 39 is the dimensional component of the optimized Mel cepstral coefficients.

[0089] S5: combining the linear prediction cepstral coefficients of the feature parameters and the optimized Mel cepstral coefficients to obtain a feature set;

[0090] In step S5, the feature parameter linear prediction cepstral coefficients and the optimized Mel cepstral coefficients are combined to obtain a feature set, including the steps of:

[0091] S51: Obtain the linear prediction cepstral coefficient and the optimized Mel cepstral coefficient of the characteristic parameters of the track transformer noise signal of each frame, combine the characteristic parameter sequences of the two in the dimension of each frame, construct a combined characteristic parameter sequence, and obtain the combined characteristic parameters;

[0092] S52: In order to allow the noise signal data to be input into the network model, the data structure must be uniform. In this process, the length of the combined feature parameter sequence is fixed to 499. The part exceeding 499 is truncated, and the part not exceeding is supplemented with 0. The final feature set is input into the network model [batch_size,499,78], which is the feature set composed of batch_size, where batch_size represents the data batch.

[0093] S6: Use the deep learning structure to further learn the feature set and train and establish a rail transit transformer fault recognition model;

[0094] In step S6, the feature set is further subjected to feature learning using a deep learning structure to train and establish a rail transit transformer fault recognition model, including the following steps:

[0095] S61: After the feature set is input into the network model, it enters a 3-layer BiLSTM layer with 100 neurons. After the 3-layer BiLSTM layer, a Dropout layer is added to prevent overfitting of the network training. Dropout randomly discards some neurons according to a set probability, so that the number of input and output neurons of each layer remains consistent during the training parameter iteration process. In order to be subsequently input into the convolution block, the parameters are reshaped into three dimensions through the reshape layer, and a feature set of [batch_size, 499, 100, 1] is obtained after the LSTM part.

[0096] S62: Convolutional neural network is constructed by connecting two convolution blocks in series. Each convolution block includes a two-dimensional convolution layer, a nonlinear function activation layer, a maximum pooling layer, a batch normalization layer and a Dropout layer. The two-dimensional convolution layer, the nonlinear function activation layer and the maximum pooling layer are sequentially passed, and the batch normalization layer and the Dropout layer are set at the end. In order to enhance the generalization performance of the model, the L2 regularization method is used for the weights of each layer.

[0097] The first convolution block contains two two-dimensional convolution layers, Conv1 and 2, using 16 large 5×5 convolution kernels with a moving step of 1. The first convolution block takes 1 channel as input and outputs 16 channels, and uses the ReLU function for nonlinear fitting; then a maximum pooling layer of size 2×2 is set to reduce the feature size and output the first feature set.

[0098] The second convolution block contains two two-dimensional convolution layers, Conv3 and 4, using 32 large 5×5 convolution kernels and a moving step of 1. The second convolution block takes 16 channels as input and outputs 32 channels. It uses the ReLU function for nonlinear fitting, and then sets a maximum pooling layer of size 2×2 to output the second feature set.

[0099] It is constructed by connecting two convolution blocks in series. Each convolution block includes a two-dimensional convolution layer, a nonlinear function activation layer, a maximum pooling layer, a batch normalization layer, and a Dropout layer. In the convolution layer part, the first convolution block contains two two-dimensional convolution layers, Conv1 and 2, using 16 large 5×5 convolution kernels with a moving step of 1. The first convolution block takes 1 channel as input and outputs 16 channels, and uses the ReLU function for nonlinear fitting. Then a maximum pooling layer of size 2×2 is set to reduce the feature size, and the output is [batch_size, 249, 50, 1 6] feature set, the second convolution block contains two two-dimensional convolution layers, Conv3 and 4, using 32 large 5×5 convolution kernels with a moving step of 1. The second convolution block takes 16 channels as input and outputs 32 channels, and uses the ReLU function for nonlinear fitting. Then a maximum pooling layer with a size of 2×2 is set to output the feature set of [batch_size,124,25,32]; at the end, a batch normalization layer and a Dropout layer are set. In order to enhance the generalization performance of the CNN-LSTM model, the L2 regularization method is used for the weights of each layer.

[0100] S7: Use the trained rail transit transformer fault identification model to detect rail transit transformer noise, use the Softmax loss function to perform the final working condition classification, and then achieve the purpose of rail transit transformer fault diagnosis.

[0101] The transformer noise signal feature information output by the convolution block will finally be input into the fully connected layer to reduce the dimension of the noise signal features and output a feature set of [batch_size, 3200]. The Softmax function then displays the multi-classification results of the transformer noise signals under various working conditions in the form of probabilities.

[0102] The experiment adopts a four-category experiment. The operating conditions of rail transit transformers used are normal operation, overload operation, discharge and DC bias magnetization. The experimental results generate a confusion matrix and obtain relevant evaluation indicators such as accuracy, precision, recall rate and F1 score. Figure 2 In the figure, the labels of the four working conditions of rail transit transformers, namely normal operation, overload operation, discharge and DC bias magnetization, are 0, 1, 2 and 3 respectively. It can be seen that the true positive and true negative examples correctly classified in the noise categories of each working condition of the transformer account for the vast majority, indicating that the model has high reliability in the classification and identification of rail transit transformer noise.

[0103] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A rail transit transformer fault diagnosis method, characterized in that: Including steps: S1: Obtain the noise signals emitted by rail transit transformers under various operating conditions; S2: Preprocess the transformer noise signal and use wavelet threshold denoising to filter and denoise the noise signal; S3: Using the Durbin algorithm to obtain linear prediction coding for the preprocessed transformer noise signal data, and then recursively extrapolating the obtained linear prediction coding to the cepstral domain to obtain linear prediction cepstral coefficients; S4: For the preprocessed transformer noise signal data, use fast Fourier transform to obtain a spectrum, perform Mel filtering on the spectrum, calculate the logarithm of the filtered data, perform discrete cosine transform to extract Mel cepstral coefficients, and combine the Mel cepstral coefficients with the first-order difference coefficients and the second-order difference coefficients to obtain optimized Mel cepstral coefficients; S5: combining the linear prediction cepstral coefficients of the feature parameters and the optimized Mel cepstral coefficients to obtain a feature set; S6: Use the deep learning structure to further learn the feature set and train and establish a rail transit transformer fault recognition model; S7: Use the trained rail transit transformer fault identification model to detect rail transit transformer noise, use the Softmax loss function to perform the final working condition classification, and then achieve the purpose of rail transit transformer fault diagnosis.

2. A rail transit transformer fault diagnosis method according to claim 1, characterized in that: In step S1, the noise signals {s0(n)}, 1≤n≤N, of the rail transit transformer collected by the same pickup device include four working conditions: normal operation, overload operation, discharge and DC bias magnetization. s0 , N s0 is the length of the signal {s0(n)}.

3. A rail transit transformer fault diagnosis method according to claim 2, characterized in that: In step S2, the transformer noise signal is preprocessed and the noise signal is filtered and denoised using wavelet threshold denoising, including the following steps: S21: Decompose the noise signal using a wavelet transform with a wavelet base of sym8 and a decomposition layer of 2, and then obtain two decomposition value components related to the resolution; set the threshold processing Fs, the threshold selection rule is the unbiased likelihood estimation threshold, the threshold adjustment relies on the transformer noise signal level estimation obtained by the wavelet decomposition of each layer, and the threshold function is a hard threshold function; When threshold processing is performed, the approximate component is retained, and for the detail component, a threshold is set. If the value of the wavelet coefficient is smaller than the set threshold, it is set to zero. If the value of the wavelet coefficient is larger than the set threshold, it is saved. S22: A first-order FIR high-pass digital filter is used to pre-emphasize the noise signal {s0(n)} to enhance the high-frequency portion and flatten the signal spectrum, resulting in the pre-emphasized signal {s(n)}: s(n)=s0(n+1)-μ*s0(n),1≤n≤(N s0 -1); Where, s0(n) is the nth data of the noise signal {s0(n)}, s0(n+1) is the n+1th data of the noise signal {s0(n)}, and the value range of μ is 0.93~0.97; S23: Frame the pre-emphasized noise signal to ensure the time resolution of the noise signal analysis; let the nth frame data after framing be x(n), and the time length corresponding to one frame be N / transformer noise acquisition frequency, where N is the length of one frame; S24: Apply a window function to the framed noise data to reduce spectral leakage; set the window function to a Hamming window, the independent variable is p, the dependent variable w(p) is the amplitude of the window function, P is a constant, and the total length of the window function is L=P+1, where L is equal to the frame length N; the expression of the window function is: w(p)=0.54-0.46*cos(2*π*p / P),0≤p≤P; In order to avoid attenuation of data at the edge of the window function, the data frames should be overlapped.

4. A rail transit transformer fault diagnosis method according to claim 1, characterized in that: In step S3: the pre-processed transformer noise signal data is subjected to the Durbin algorithm to obtain linear prediction coding, and the obtained linear prediction coding is then recursively transferred to the cepstral domain to obtain linear prediction cepstral coefficients, including the following steps: S31: for the pre-processed transformer noise signal data, the output signal of the noise signal data n is expressed by a linear combination of the previous m samples, and is derived by linear prediction analysis coefficients of various orders to form an equation group consisting of m equations; S32: solving the equations using the Durbin algorithm to obtain linear predictive coding; S33: Perform fast Fourier transform calculation on the linear prediction coding, then perform logarithmic operation, and finally obtain the linear prediction cepstral coefficients through inverse fast Fourier transform operation.

5. A rail transit transformer fault diagnosis method according to claim 1, characterized in that: In step S4: the preprocessed transformer noise signal data is subjected to fast Fourier transform to obtain a spectrum, the spectrum is subjected to Mel filtering, the filtered data is logarithmized, discrete cosine transform is performed to extract Mel cepstral coefficients, and the Mel cepstral coefficients are combined with first-order difference coefficients and second-order difference coefficients to obtain optimized Mel cepstral coefficients, including the following steps: S41: performing an N-point fast Fourier transform on each frame of the preprocessed transformer noise signal data, converting the noise signal from the time domain to a spectrum, and obtaining a spectrum diagram; S42: setting a Mel triangle filter bank, filtering the spectrum graph using the Mel triangle filter bank, and calculating the logarithm of the filtered data; S43: Use discrete cosine transform to transform the logarithmic Mel spectrum into the time domain and extract the Mel cepstral coefficients; S44: Calculate the first-order difference coefficient and the second-order difference coefficient of the Mel-cepstral coefficient, and linearly combine the Mel-cepstral coefficient with the first-order difference coefficient and the second-order difference coefficient to obtain the optimized Mel-cepstral coefficient.

6. A rail transit transformer fault diagnosis method according to claim 5, characterized in that: The calculation formulas for the first-order difference coefficient and the second-order difference coefficient of the Mel-frequency cepstral coefficient are as follows: d t represents the tth first-order difference, C t represents the tth Mel cepstral coefficient, Q represents the order of the cepstral coefficient; K represents the time difference of the first-order derivative, which is 1 or 2, and k is the cumulative coefficient from 1 to K; Substitute the Mel cepstral coefficients into the above formula to obtain the same number of first-order difference coefficients d t ; Then we can use the d we just obtained t As C t Substituting into the above formula, we can obtain the second-order difference coefficient.

7. A rail transit transformer fault diagnosis method according to claim 1, characterized in that: In step S5, the feature parameter linear prediction cepstral coefficients and the optimized Mel cepstral coefficients are combined to obtain a feature set, including the steps of: S51: Obtain the linear prediction cepstral coefficient and the optimized Mel cepstral coefficient of the characteristic parameters of the track transformer noise signal of each frame, combine the characteristic parameter sequences of the two in the dimension of each frame, construct a combined characteristic parameter sequence, and obtain the combined characteristic parameters; S52: The length of the combined feature parameter sequence is fixed to 499, the part exceeding 499 is truncated, and the part less than 499 is supplemented with 0, and the final feature set is input into the network model.

8. A rail transit transformer fault diagnosis method according to claim 1, characterized in that: In step S6, the feature set is further subjected to feature learning using a deep learning structure to train and establish a rail transit transformer fault recognition model, including the following steps: S61: After the feature set is input into the network model, it enters the 3-layer BiLSTM layer. After the 3-layer BiLSTM layer, a Dropout layer is added to prevent overfitting of the network training. Dropout randomly discards some neurons according to a set probability, so that the number of input and output neurons of each layer remains consistent during the training parameter iteration process of the transformer noise signal feature information. In order to be subsequently input into the convolution block, the parameters are reshaped into three dimensions through the reshape layer. S62: Convolutional neural network is constructed by connecting two convolution blocks in series. Each convolution block includes a two-dimensional convolution layer, a nonlinear function activation layer, a maximum pooling layer, a batch normalization layer and a Dropout layer. The two-dimensional convolution layer, the nonlinear function activation layer and the maximum pooling layer are sequentially passed, and the batch normalization layer and the Dropout layer are set at the end. In order to enhance the generalization performance of the model, the L2 regularization method is used for the weights of each layer.

9. A rail transit transformer fault diagnosis method according to claim 8, characterized in that: The first convolution block contains two two-dimensional convolution layers, Conv1 and 2, using 16 large 5×5 convolution kernels with a moving step of 1. The first convolution block takes 1 channel as input and outputs 16 channels, and uses the ReLU function for nonlinear fitting; then a maximum pooling layer of size 2×2 is set to reduce the feature size and output the first feature set.

10. A rail transit transformer fault diagnosis method according to claim 8 or 9, characterized in that: The second convolution block contains two two-dimensional convolution layers, Conv3 and 4, using 32 large 5×5 convolution kernels and a moving step of 1. The second convolution block takes 16 channels as input and outputs 32 channels. It uses the ReLU function for nonlinear fitting, and then sets a maximum pooling layer of size 2×2 to output the second feature set.

Images