Substation noise abnormity identification method based on non-intrusive decomposition technology

Abstract

The invention discloses a transformer substation noise abnormity identification method based on a non-intrusive decomposition technology, and belongs to the technical field of power system equipment monitoring and signal processing. Comprising the following steps: S1, arranging non-contact sound sensors around a transformer substation factory boundary, and collecting noise data of a transformer substation; s2, preprocessing the noise data; s3, performing multi-scale decomposition on the preprocessed noise data by using wavelet transform, and extracting spectrum features from each frame of noise signal; s4, forming a time sequence-frequency spectrum joint input matrix; s5, setting a convolutional network, and outputting a noise category; and S6, performing diagnosis by combining the noise category with the equipment operation parameters as a judgment basis. By adopting the method, the non-intrusive monitoring characteristic is achieved, the equipment installation and maintenance cost is reduced, the method is suitable for operation in complex environmental noise, an abnormal noise source can be accurately positioned, and the equipment operation and maintenance efficiency and the intelligent level of fault diagnosis can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of power system equipment monitoring and signal processing, and in particular to a method for identifying abnormal substation noise based on non-invasive decomposition technology. Background Art

[0002] As the scale of power systems expands and becomes more complex, the stable operation of substations is crucial to the safety and reliability of the power grid. The noise generated by equipment such as transformers and reactors during operation often contains rich information about the equipment's operating status. These noises can not only reflect the normal working status of the equipment, but may also indicate potential equipment failures. Therefore, accurately identifying and decomposing the operating noise of substation equipment is of great significance for ensuring the stable operation of the power system, improving the operation and maintenance level of equipment, and reducing the risk of failures.

[0003] At present, many noise monitoring methods for substations rely on intrusive detection methods, that is, various detection devices such as vibration sensors or noise sensors are directly installed inside the substation. Although this method can capture the operating noise of equipment such as transformers and reactors, it has many shortcomings. The intrusive monitoring method requires the modification or installation of sensors for each key equipment when installing equipment inside the substation, which not only increases the cost of the equipment, but may also interfere with normal operation and even require shutdown for installation. Moreover, relying solely on noise monitoring data inside the substation cannot accurately reflect the actual noise level at the plant boundary. Evaluating the impact of noise pollution only through internal noise detection may have large deviations and cannot accurately reflect the impact of plant boundary noise on the surrounding environment and residents. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying substation noise anomalies based on non-invasive decomposition technology, which can not only accurately separate and identify the operating noise of equipment such as transformers and reactors, but also effectively deal with background noise interference in complex environments. The present invention can adapt to the dynamic changes of noise characteristics in real time, thereby providing a more intelligent, efficient and reliable solution for noise monitoring, equipment status assessment and fault diagnosis in substations.

[0005] To achieve the above object, the present invention provides a method for identifying abnormal substation noise based on non-intrusive decomposition technology, comprising the following steps:

[0006] S1: Non-contact sound sensors are arranged around the substation boundary to collect noise data of the substation and record the noise data in chronological order to obtain the noise time series data X time ;

[0007] S2: Preprocess the noise data, including data quality verification, missing value processing, normalization, frame processing and windowing processing;

[0008] S3: Use wavelet transform to perform multi-scale decomposition on the pre-processed noise data and extract spectral features of the noise signal of each frame;

[0009] S4: stack the time series data and spectrum features of the noise along the time dimension to form a time series-spectrum joint input matrix;

[0010] S5: Set up a convolutional network with a joint time-spectrum input matrix as input and output noise categories;

[0011] S6: Use the noise category in combination with the equipment operating parameters as the basis for diagnosis.

[0012] Preferably, in step S1, the noise data of the substation specifically includes operating noise of the transformer and reactor and environmental noise.

[0013] Preferably, in step S2, the specific process of preprocessing the noise data is as follows:

[0014] Data quality verification: perform outlier detection on noise data and use linear interpolation or median interpolation to handle outliers that are beyond the normal range;

[0015] Missing value processing: interpolation methods such as linear interpolation, forward filling, and backward filling of missing values;

[0016] Normalization processing: perform maximum-minimum normalization, Z-score normalization or decimal calibration normalization on the processed noise data;

[0017] Framing: Set the frame length to 20ms to 40ms and the frame rate to 1 / 2 or 1 / 3;

[0018] Window processing: Use Hamming window or Haining window processing on the framed data.

[0019] Preferably, the specific process of step S3 is as follows:

[0020] S31: Wavelet decomposition: Use wavelet transform to perform multi-scale decomposition on the noise signal of each frame, decompose the noise signal into different frequency components, extract low-frequency and high-frequency components, and the signal after wavelet transform is expressed as follows:

[0021]

[0022] In the above formula, A j (t) represents the low-frequency component, which represents the smooth part of the signal, D j (t) represents the high-frequency component, which represents the details of the signal, and J represents the number of decomposition levels;

[0023] S32: Extract spectrum features: Perform short-time Fourier transform on the decomposed noise signal to convert the time domain data into frequency domain data. The formula is as follows:

[0024]

[0025] In the above formula, τ is the time offset, which is used to shift the signal in the time domain; represents the center frequency of the window function, which is used to focus on a specific frequency band of the signal; w(tT) is the window function used to intercept a part of the signal for spectrum analysis; T represents the time center of the window function, which is used to adjust the position of the window function; ω is the frequency variable, which represents the frequency component of the Fourier transform;

[0026] The spectrum after short-time Fourier transform is processed by Mel filter bank, and the conversion formula of Mel frequency is as follows:

[0027]

[0028] In the above formula, Mel(·) represents Mel frequency, f represents frequency, and after converting frequency f into Mel frequency, a set of triangular filters are used to perform weighted summation on the spectrum to obtain the energy of each Mel filter and form an energy spectrum;

[0029] The energy spectrum is discrete cosine transformed to obtain the Mel frequency cepstral coefficient MFCC. The feature extraction formula of the Mel frequency cepstral coefficient of each frame is:

[0030]

[0031] In the above formula, MFCC(k) represents the kth Mel frequency cepstral coefficient; k represents the ordinal number of the cepstral coefficient, which represents the kth MFCC coefficient; n represents the frequency index, which is used to identify the nth frequency component; N represents the total number of Mel filters, that is, the number of frequency bands into which the spectrum is divided; S[n] represents the energy of the Mel filter on the nth frequency component, which represents the energy value of the signal after passing through the nth Mel filter;

[0032] The formula of spectrum characteristics is as follows:

[0033]

[0034] In the above formula, F is the dimension of the spectral feature, f T is the spectral feature vector at the Tth time step.

[0035] Preferably, in step S4, the process of forming the time series-spectrum joint feature matrix is ​​as follows: the original time series data and spectrum characteristics of the substation boundary noise are taken as input, wherein the formula of the original time series data is as follows:

[0036]

[0037] In the above formula, T is the total number of time steps, x T is the noise signal value at the Tth time step;

[0038] The original time series data and spectrum characteristics of the substation boundary noise are spliced ​​together along the time dimension to form a time series-spectrum joint input matrix X joint , the expression is as follows:

[0039]

[0040] In the above formula, each row represents the joint representation of the time series signal and the spectrum feature at a certain time step, F is the dimension of the spectrum feature, Represents the dimensional space of the data matrix.

[0041] Preferably, in step S5, the specific process of outputting the noise category is as follows:

[0042] A CNN convolutional network is used to extract time-frequency features. The convolutional network includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. Adam is used to optimize the CNN convolutional network, and the exponential weighted average of the first-order momentum and the second-order momentum is used to update the parameters. For the trained convolutional network, the time-frequency joint input matrix is ​​used as input, and the noise category is output. The noise category includes normal noise and fault noise.

[0043] Preferably, for the output noise category, combined with the equipment operating parameters, the diagnosis result is a normal state or an early fault or a fault state.

[0044] Therefore, the present invention adopts the above-mentioned method for identifying abnormal substation noise based on non-intrusive decomposition technology, which has the following advantages:

[0045] (1) The technical solution of the present invention adopts non-intrusive monitoring, which is lower in cost: The method of the present invention adopts non-intrusive sound sensors to collect plant boundary noise data, without the need to modify specific equipment or install additional sensors, which greatly reduces the equipment maintenance and monitoring costs. Compared with the traditional method that requires sensors to be directly installed on the equipment, the present invention realizes comprehensive monitoring of the entire substation through plant boundary noise collection, further reducing the impact on the normal operation of the equipment.

[0046] (2) The technical solution of the present invention is more adaptable to complex environments: Traditional technologies mostly work in relatively single noise environments, while the present invention can effectively operate in complex environmental noise. By combining Mel-frequency cepstral coefficients (MFCC) with multi-scale wavelet decomposition, the present invention can extract the unique spectral features of the device from a noisy environment, significantly improving the recognition accuracy of different noise sources.

[0047] (3) The technical solution of the present invention can locate the source of abnormal noise: through the joint analysis based on time series data and spectrum characteristics, different noise sources in the substation boundary noise can be accurately identified and separated. For example, the present invention can not only distinguish the noise under normal working conditions, but also locate the source of abnormal noise emitted by different equipment such as transformers and reactors.

[0048] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The present invention is a flow chart of a method for identifying abnormal substation noise based on non-intrusive decomposition technology. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations. The specific model specifications need to be selected and determined according to the actual specifications of the device, and the specific selection calculation method adopts the existing technology in the field, so it will not be described in detail.

[0051] Example

[0052] like Figure 1 As shown, the present invention provides a method for identifying abnormal substation noise based on non-intrusive decomposition technology, comprising the following steps:

[0053] S1: Non-contact sound sensors are arranged around the substation boundary to collect noise data of the substation. The noise data of the substation specifically includes the operating noise of the transformer and reactor and the environmental noise. The noise data is recorded in chronological order and the collection point and time are marked to obtain the time series data of the noise X time , providing a basis for subsequent data processing and analysis;

[0054] S2: Preprocess the collected noise data. The preprocessing includes data quality verification, missing value processing, normalization, frame processing and windowing processing. The specific process is as follows:

[0055] Data quality verification: perform outlier detection on noise data and use linear interpolation or median interpolation to handle outliers that are beyond the normal range;

[0056] Missing value processing: missing values ​​are processed by linear interpolation or interpolation methods such as forward filling and backward filling;

[0057] Normalization processing: perform maximum-minimum normalization, Z-score normalization or decimal calibration normalization on the processed noise data;

[0058] Frame processing: according to the noise characteristics, the frame length is set to 20ms to 40ms and the frame rate is 1 / 2 or 1 / 3;

[0059] Windowing: To reduce spectrum leakage, the framed data is processed using window functions such as the Hamming window or the Haining window.

[0060] S3: Use wavelet transform to perform multi-scale decomposition on the pre-processed noise data, and extract the spectrum features of the noise signal of each frame; the specific process is as follows:

[0061] S31: Wavelet decomposition: Use wavelet transform to perform multi-scale decomposition on the noise signal of each frame, decompose the noise signal into different frequency components, extract low-frequency and high-frequency components, and the signal after wavelet transform is expressed as follows:

[0062]

[0063] In the above formula, A j (t) represents the low-frequency component, which represents the smooth part of the signal, D j (t) represents the high-frequency component, which represents the details of the signal, and J represents the number of decomposition levels;

[0064] S32: Extract spectrum features: Perform short-time Fourier transform on the decomposed noise signal to convert the time domain data into frequency domain data. The formula is as follows:

[0065]

[0066] In the above formula, τ is the time offset, which is used to shift the signal in the time domain; represents the center frequency of the window function, which is used to focus on a specific frequency band of the signal; w(tT) is the window function used to intercept a part of the signal for spectrum analysis; T represents the time center of the window function, which is used to adjust the position of the window function; ω is the frequency variable, which represents the frequency component of the Fourier transform;

[0067] The spectrum after short-time Fourier transform is processed by Mel filter bank, and the conversion formula of Mel frequency is as follows:

[0068]

[0069] In the above formula, Mel(·) represents Mel frequency, f represents frequency. After converting frequency f into Mel frequency, the spectrum is weighted and summed through a set of triangular filters to obtain the energy of each Mel filter and form an energy spectrum. The filtering steps are:

[0070] E m =∑ k |S k | 2 H m (k);

[0071] Among them, E m represents the output energy of the mth filter, |S k | 2 Represents the square of the spectrum amplitude (energy), H m (k) is the weight of the mth triangular filter;

[0072] The energy spectrum is discrete cosine transformed to obtain the Mel frequency cepstral coefficient MFCC. The feature extraction formula of the Mel frequency cepstral coefficient of each frame is:

[0073]

[0074] In the above formula, MFCC(k) represents the kth Mel-frequency cepstral coefficient; k represents the ordinal number of the cepstral coefficient, indicating the kth MFCC coefficient; n represents the frequency index, which is used to identify the nth frequency component; N represents the total number of Mel filters, that is, the number of frequency bands into which the spectrum is divided; S[n] represents the energy of the Mel filter on the nth frequency component, representing the energy value of the signal after passing through the nth Mel filter.

[0075] The formula of spectrum characteristics is as follows:

[0076]

[0077] In the above formula, F is the dimension of the spectral feature, that is, the number of frequency components calculated in each time step T. T It is the spectral feature vector at the Tth time step, reflecting the frequency distribution of the signal at that time point.

[0078] S4: Stack the time series data and spectrum characteristics of the noise along the time dimension to form a time series-spectrum joint input matrix. The process is as follows: The original time series data and spectrum characteristics of the substation boundary noise are used as input, where the formula for the original time series data is as follows:

[0079]

[0080] In the above formula, T is the total number of time steps, x T is the noise signal value at the Tth time step;

[0081] The original time series data and spectrum characteristics of the substation boundary noise are spliced ​​together along the time dimension to form a time series-spectrum joint input matrix X joint , the expression is as follows:

[0082]

[0083] In the above formula, each row represents the joint representation of the time series signal and the spectrum feature at a certain time step, F is the dimension of the spectrum feature, Represents the dimensional space of the data matrix. For example:

[0084] X time ∈R T×1 Represents the time series data matrix X time The dimension is T rows and 1 column;

[0085] X frq ∈R T×F Represents the spectrum feature matrix X freq The dimensions are T rows and F columns;

[0086] X joint ∈R T×(1+F) Denotes the combined input matrix X joint The dimension is T rows 1 + F columns;

[0087] S5: Set up a convolutional network, take the time-spectrum joint input matrix as input, and output the noise category; the specific process is as follows:

[0088] The CNN convolutional network is used to extract time-frequency features. The convolutional network includes input layer, convolution layer, pooling layer, and fully connected layer. Adam is used to optimize the CNN convolutional network, and the exponential weighted average of the first-order momentum and the second-order momentum is used to update the parameters. The specific contents are as follows:

[0089] Input layer, the input is T×(1+F) time series-spectrum joint feature matrix. Each input batch contains B samples, and the size of the input matrix is:

[0090]

[0091] Convolutional layer, using a 1D convolutional layer to process the joint input matrix, convolve the features in the time dimension, and extract local patterns of timing and spectrum. The formula for the convolution operation is:

[0092]

[0093] In the above formula, y t is the convolution output at time step t, σ(·) is the activation function, and W is the convolution kernel size k×(1+F).

[0094] Pooling layer, use pooling layer (max pooling) to reduce feature dimensions and reduce the amount of calculation. The role of pooling is to compress the feature map and retain the most important features:

[0095]

[0096] Where: p is the size of the pooling window.

[0097] The fully connected layer expands the pooled feature map into a vector and passes it to the fully connected layer for further feature fusion and classification. The fusion expression is:

[0098] z=σ(W fc y+b fc );

[0099] In the above formula, W fc is the weight matrix of the fully connected layer, y is the feature vector after convolution and pooling, σ(·) is the activation function, and b fc Represents the bias term of the fully connected layer, which helps to adjust the output of the model to improve the model's ability to fit complex data.

[0100] Use Adam to optimize CNN network In order to improve training efficiency and recognition accuracy, the Adam algorithm is used to optimize the CNN network. The specific parameter update rules are as follows:

[0101] m t =β1m t-1 +(1-β1)g t

[0102]

[0103] In the above formula, g t is the gradient of the current time step, m t and v t are the first-order and second-order momentum of the gradient, β1 and β2 are the decay coefficients of the momentum, α is the learning rate, and θ t is the parameter of the model, ∈ is a constant (usually 10 -8 ).

[0104] The weights and biases of each layer are updated according to the Adam algorithm. After each training cycle, the performance of the model on the validation set is evaluated and the learning rate is dynamically adjusted to avoid overfitting.

[0105] The mean square error (MSE) is used to calculate the error between the predicted value and the actual value, and the classification or prediction performance of the model is evaluated on the test set. The formula for calculating the mean square error is as follows:

[0106]

[0107] For the trained convolutional network, the time series-spectrum joint input matrix is ​​used as input and the noise category is output. The noise category includes normal noise and fault noise.

[0108] S6: The noise category is combined with the equipment operating parameters, and the diagnosis result is normal state or early fault or fault state. The judgment process is as follows:

[0109] Normal operation: When the noise category is "normal noise" and the operating status and parameters of the device are normal, the device is in a healthy state.

[0110] Incipient failure: When the noise category is "fault noise", but the operating parameters of the equipment have not obviously exceeded the preset range, it indicates that the equipment may be in an incipient failure state and preventive maintenance is required.

[0111] Fault status: When the noise category is "fault noise" and the operating parameters of the equipment are also in an abnormal state, it is considered that the equipment has a fault and emergency measures must be taken.

[0112] Therefore, the present invention adopts a method for identifying substation noise anomalies based on non-invasive decomposition technology, which can not only accurately separate and identify the operating noise of equipment such as transformers and reactors, but also effectively deal with background noise interference in complex environments. The present invention can adapt to the dynamic changes of noise characteristics in real time, thereby providing a more intelligent, efficient and reliable solution for noise monitoring, equipment status assessment and fault diagnosis of substations.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for identifying abnormal substation noise based on non-intrusive decomposition technology, characterized in that: The following steps are involved: S1: Non-contact sound sensors are arranged around the substation boundary to collect noise data of the substation and record the noise data in chronological order to obtain the noise time series data X time ; S2: Preprocess the noise data, including data quality verification, missing value processing, normalization, frame processing and windowing processing; S3: Use wavelet transform to perform multi-scale decomposition on the pre-processed noise data and extract spectral features of the noise signal of each frame; S4: stack the time series data and spectrum features of the noise along the time dimension to form a time series-spectrum joint input matrix; S5: Set up a convolutional network with a joint time-spectrum input matrix as input and output noise categories; S6: Use the noise category in combination with the equipment operating parameters as the basis for diagnosis.

2. According to claim 1, a method for identifying abnormal substation noise based on non-intrusive decomposition technology is characterized by: In step S1, the noise data of the substation specifically includes the operating noise of the transformer and the reactor and the environmental noise.

3. The method for identifying abnormal substation noise based on non-intrusive decomposition technology according to claim 2 is characterized in that: In step S2, the specific process of preprocessing the noise data is as follows: Data quality verification: perform outlier detection on noisy data and use linear interpolation or median interpolation to handle outliers that are beyond the normal range; Missing value processing: interpolation methods such as linear interpolation, forward filling, and backward filling of missing values; Normalization processing: perform maximum-minimum normalization, Z-score normalization or decimal calibration normalization on the processed noise data; Framing: Set the frame length to 20ms to 40ms and the frame rate to 1 / 2 or 1 / 3; Windowing processing: Use Hamming window or Haining window processing on the framed data.

4. The method for identifying abnormal substation noise based on non-intrusive decomposition technology according to claim 3 is characterized in that: The specific process of step S3 is as follows: S31: Wavelet decomposition: Use wavelet transform to perform multi-scale decomposition on the noise signal of each frame, decompose the noise signal into different frequency components, extract low-frequency and high-frequency components, and the signal after wavelet transform is expressed as follows: In the above formula, A j (t) represents the low-frequency component, which represents the smooth part of the signal, D j (t) represents the high-frequency component, which represents the details of the signal, and J represents the number of decomposition levels; S32: Extract spectrum features: Perform short-time Fourier transform on the decomposed noise signal to convert the time domain data into frequency domain data. The formula is as follows: In the above formula, τ is the time offset, which is used to shift the signal in the time domain; represents the center frequency of the window function, which is used to focus on a specific frequency band of the signal; w(tT) is the window function used to intercept a part of the signal for spectrum analysis; T represents the time center of the window function, which is used to adjust the position of the window function; ω is the frequency variable, which represents the frequency component of the Fourier transform; The spectrum after short-time Fourier transform is processed by Mel filter bank, and the conversion formula of Mel frequency is as follows: In the above formula, Mel(·) represents Mel frequency, f represents frequency, and after converting frequency f into Mel frequency, a set of triangular filters are used to perform weighted summation on the spectrum to obtain the energy of each Mel filter and form an energy spectrum; The energy spectrum is discrete cosine transformed to obtain the Mel frequency cepstral coefficient MFCC. The feature extraction formula of the Mel frequency cepstral coefficient of each frame is: In the above formula, MFCC(k) represents the kth Mel frequency cepstral coefficient; k represents the ordinal number of the cepstral coefficient, which represents the kth MFCC coefficient; n represents the frequency index, which is used to identify the nth frequency component; N represents the total number of Mel filters, that is, the number of frequency bands into which the spectrum is divided; S[n] represents the energy of the Mel filter on the nth frequency component, which represents the energy value of the signal after passing through the nth Mel filter; The formula of spectrum characteristics is as follows: In the above formula, F is the dimension of the spectral feature, f T is the spectral feature vector at the Tth time step.

5. The method for identifying abnormal substation noise based on non-intrusive decomposition technology according to claim 4 is characterized in that: In step S4, the process of forming the time series-spectrum joint feature matrix is ​​as follows: the original time series data and spectrum features of the substation boundary noise are taken as input, wherein the formula of the original time series data is as follows: In the above formula, T is the total number of time steps, x T is the noise signal value at the Tth time step; The original time series data and spectrum characteristics of the substation boundary noise are spliced ​​together along the time dimension to form a time series-spectrum joint input matrix X joint , the expression is as follows: In the above formula, each row represents the joint representation of the time series signal and the spectrum feature at a certain time step, F is the dimension of the spectrum feature, Represents the dimensional space of the data matrix.

6. The method for identifying abnormal substation noise based on non-intrusive decomposition technology according to claim 5 is characterized in that: In step S5, the specific process of outputting the noise category is as follows: A CNN convolutional network is used to extract time-frequency features. The convolutional network includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. Adam is used to optimize the CNN convolutional network, and the exponential weighted average of the first-order momentum and the second-order momentum is used to update the parameters. For the trained convolutional network, the time-frequency joint input matrix is ​​used as input, and the noise category is output. The noise category includes normal noise and fault noise.

7. The method for identifying abnormal substation noise based on non-intrusive decomposition technology according to claim 6 is characterized by: For the output noise category, combined with the equipment operating parameters, the diagnosis result is normal state or early fault or fault state.

Images