Transformer noise detection method based on seasonal trend decomposition and time sequence prediction

Abstract

The invention discloses a transformer noise detection method based on season trend decomposition and time sequence prediction, and belongs to the technical field of transformer noise anomaly detection. Comprising the following steps: S1, acquiring noise data of a transformer, and preprocessing the noise data; s2, adopting an STL decomposition model to obtain a seasonal component, a trend component and a residual component; s3, performing Transform time sequence prediction on the seasonal components and the trend components to obtain prediction results of the seasonal components and the trend components; s4, performing anomaly detection on prediction results of the seasonal components and the trend components, and outputting a detection result; and S5, carrying out weighted composition on the detection results of all the components, and outputting a diagnosis result. According to the method, the time sequence data of the noise of the transformer substation is decomposed, predicted and subjected to abnormal point detection, and reliable technical support is provided for state monitoring and fault prevention of transformer substation equipment.

Description

Technical Field

[0001] The invention relates to the technical field of transformer noise anomaly detection, and in particular to a transformer noise detection method based on seasonal trend decomposition and time series prediction. Background Art

[0002] Traditional transformer noise monitoring methods mainly rely on physical signal processing technology, such as spectrum analysis and filtering, to decompose and classify noise data. These methods can effectively identify and analyze noise in specific scenarios, but their limitations are also obvious. First, the complexity of noise signals is high, and different types of noise may overlap or interfere with each other. Traditional methods have poor separation effects when facing such complex signals. Secondly, traditional methods usually rely on fixed models, which are difficult to adapt to the characteristics of noise signals that change with time and environmental conditions, resulting in insufficient monitoring effect and accuracy. More importantly, traditional noise monitoring and control methods often take measures to control noise after it exceeds the standard, and fail to provide early warning and intervention before the noise anomaly occurs. The present invention, by combining time series prediction and anomaly detection technology, can identify the trend of noise exceeding the standard in advance, and make predictions and interventions before the noise exceeds the standard, thereby effectively preventing the occurrence of noise problems and improving the foresight and initiative of noise management. Summary of the invention

[0003] The purpose of the present invention is to provide a transformer noise detection method based on seasonal trend decomposition and time series prediction, which uses time series decomposition technology combined with Transformer network to effectively predict substation noise and detect abnormal noise in advance. By decomposing, predicting and detecting abnormal points of substation noise time series data, it not only improves the intelligent level of noise management, but also provides reliable technical support for status monitoring and fault prevention of substation equipment.

[0004] To achieve the above object, the present invention provides a transformer noise detection method based on seasonal trend decomposition and time series prediction, comprising the following steps:

[0005] S1: Obtain the noise data of the transformer and preprocess the noise data, including data cleaning and normalization, remove outliers and fill in missing values;

[0006] S2: Use the STL decomposition model to decompose the preprocessed noise data to obtain seasonal components, trend components and residual components;

[0007] S3: Perform Transformer time series prediction on both seasonal components and trend components to obtain the prediction results of seasonal components and trend components;

[0008] S4: Perform anomaly detection on the prediction results of seasonal components and trend components, and output the detection results respectively;

[0009] S5: Perform weighted compounding on the detection results and output the recognition results.

[0010] Preferably, in step S2, the noise data decomposition process is as follows:

[0011] The STL decomposition formula is as follows:

[0012] Y t =T t +S t +R t ;

[0013] In the above formula, Y t is the original noise signal; T t is the trend component, reflecting the long-term changes of the noise signal; S t is the seasonal component, capturing the periodic fluctuations. The periodic fluctuations are the noise changes caused by the periodic load of the equipment. t is the residual component, which represents random fluctuations or non-systematic changes in noise;

[0014] The trend component and seasonality are estimated using local weighted regression Loess, and the estimation formula is as follows:

[0015] T t =Loess(Y t ,λ trend );

[0016] S t =Loess(Y t -T t ,λ seasonal );

[0017] In the above formula, λ trend represents the window size parameter for trend smoothing, λ seasonal is the smoothing parameter for the seasonal term, representing the period length of seasonal fluctuations;

[0018] The residual component is obtained by subtracting the trend term and the period term from the original signal. The formula is as follows:

[0019] R t =Y t -T t -S t ;

[0020] The residual component reflects the irregular and random part of the noise signal.

[0021] Preferably, in step S3, a Transformer model is used for prediction, wherein the seasonal component is used to capture the periodicity in the noise data, and the trend component is used to capture the long-term change trend.

[0022] Preferably, the specific process of step S4 is as follows:

[0023] Fast Fourier transform is used to detect periodic anomalies in the forecast results of seasonal components.

[0024] The seasonal component S t Convert to frequency domain representation:

[0025]

[0026] In the above formula, X(k) is the complex number representation of the kth frequency component, S(n) represents the value of the time domain signal at the nth sampling point, N is the total number of sampling points of the seasonal component, and the amplitude and phase are calculated as follows:

[0027]

[0028] In the above formula, A(k) represents the amplitude of the kth frequency component, and Φ(k) represents the phase of the kth frequency component; the mean and standard deviation of the amplitude A(k) and phase Φ(k) are calculated respectively:

[0029]

[0030] μA 为 The mean of the amplitude; σ A is the standard deviation of the amplitude; μ Φ is the mean value of the amplitude; σ Φ is the standard deviation of the amplitude; corresponding to the setting of the amplitude dynamic threshold T A and phase dynamic threshold T θ , the formula is as follows:

[0031] T A =μ A +k*σ A , T θ =μ Φ +kσ A ;

[0032] In the above formula, when A(k)>T A For amplitude anomaly, when |Φ(k)|>T θ is a phase anomaly. When both are abnormal, it is recorded as a significant anomaly.

[0033] The Bayesian inflection point detection algorithm is used to detect inflection points of the forecast results of the trend component and identify significant inflection points and abnormal inflection points:

[0034] The formula for modeling the inflection point location is as follows:

[0035] P(τ)=λe -λτ ;

[0036] In the above formula, λ represents the rate parameter of the inflection point. For each time step t, when the inflection point occurs at a certain time step, the mean and variance of the trend components before and after the inflection point are calculated respectively, and the likelihood function is constructed based on the normal distribution as follows:

[0037]

[0038] Among them, T t is the trend component in time, μ before and are the mean and variance before the inflection point; μ after and are the mean and variance after the inflection point, respectively, and τ is the time step where the inflection point is located;

[0039] The posterior probability of the inflection point indicates that when the trend data T is observed 1:n After that, the probability that a certain time step is an inflection point is as follows:

[0040]

[0041] Among them, P(T 1:n |τ) is the likelihood function, P(τ) is the prior probability, P(T 1:n ) is the marginal likelihood of the data, and the sum of the posterior probabilities at all time points is 1; for each new time step t+1, the posterior probability is recursively updated as follows:

[0042] P(τ|T 1:t+1 )∞P(T t+1 |τ)P(τ|T 1:t );

[0043] In the above formula, the threshold θ is set to determine the inflection point. When the posterior probability P(τ|T 1:n ) exceeds the threshold θ, then time t is determined to be the inflection point.

[0044] Preferably, in step S5, the results of the trend component and the seasonal component are weighted and superimposed to generate the final anomaly detection and change inflection point results, and the formula is as follows:

[0045] A final (t) = w1A trend (t)+w2A seasonal (t);

[0046] Among them, A final (t) is the final anomaly detection result at time t; A trend(t) and A seasonal (t) are the anomaly detection results of trend component and seasonal component respectively; w1 and w2 are the weighting coefficients of trend component and seasonal component respectively. In the continuous time step, the anomaly detection result A final If (t) continues to exceed the threshold, it is determined that these time steps constitute an abnormal interval.

[0047] Therefore, the present invention adopts the above-mentioned transformer noise detection method based on seasonal trend decomposition and time series prediction, which has the following advantages:

[0048] (1) In the present invention, the fast Fourier transform (FFT) and Bayesian inflection point detection are combined to perform multi-level anomaly detection on different noise components, respectively detecting periodic anomalies and trend inflection points. Compared with the existing single anomaly detection method, the multi-level detection can more comprehensively and intelligently analyze the anomalies in the noise signal, identify potential equipment problems, and have the ability to provide early warning, thereby providing operators with more intelligent maintenance suggestions.

[0049] (2) In the present invention, by using an adaptive Transformer model, new noise data can be self-adjusted and optimized. The model can continuously learn new data features and has the ability to adapt to noise changes and environmental influences, thereby improving the accuracy and robustness of prediction and anomaly detection. Compared with traditional static models, the scheme proposed in the present invention has stronger adaptability when processing complex and changeable noise signals.

[0050] (3) In the present invention, transformer noise monitoring no longer requires the establishment of complex physical models for each device, but instead uses intelligent algorithms to achieve efficient decomposition, prediction and detection of noise data. This greatly reduces the reliance on physical monitoring equipment and manual analysis, reduces the construction and maintenance costs of the monitoring system, and improves the efficiency of noise management.

[0051] (4) In the present invention, through the prediction and detection system of the present invention, the operator can monitor the noise status of the equipment in real time, accurately identify noise anomalies and potential equipment failure problems, and take maintenance measures in advance to avoid equipment operation failure or more serious consequences. This system can significantly improve the reliability of transformer equipment operation.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The present invention is a flow chart of a transformer noise detection method based on seasonal trend decomposition and time series prediction. DETAILED DESCRIPTION

[0054] 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.

[0055] Example

[0056] like Figure 1 As shown, the present invention provides a transformer noise detection method based on seasonal trend decomposition and time series prediction, comprising the following steps:

[0057] S1: Obtain the noise data of the transformer and preprocess the noise data, including data cleaning and normalization, remove outliers and fill in missing values;

[0058] Data cleaning: Check missing values ​​in noisy data and use interpolation to fill in missing data; use Z-score method to detect outliers in noisy data. When some data points deviate significantly from the normal range, they can be corrected through historical data.

[0059] Data normalization: Since the noise data may span different numerical ranges, the data needs to be normalized. Normalization compresses the data into a uniform range to avoid affecting the model training results due to different data scales in subsequent analysis.

[0060] S2: Use the STL decomposition model to decompose the preprocessed noise data to obtain seasonal components, trend components and residual components;

[0061] The STL decomposition formula is as follows:

[0062] Y t =T t +S t +R t ;

[0063] In the above formula, Y t is the original noise signal; T t is the trend component, reflecting the long-term changes of the noise signal; S t is the seasonal component, capturing the periodic fluctuations. The periodic fluctuations are the noise changes caused by the periodic load of the equipment. t is the residual component, which represents random fluctuations or non-systematic changes in noise;

[0064] The trend component and seasonality are estimated using local weighted regression Loess, and the estimation formula is as follows:

[0065] T t =Loess(Y t ,λ trend );

[0066] S t =Loess(Y t -T t ,λ seasonal );

[0067] In the above formula, λ trend represents the window size parameter for trend smoothing, λ seasonal is the smoothing parameter for the seasonal term, representing the period length of seasonal fluctuations;

[0068] The residual component is obtained by subtracting the trend term and the period term from the original signal. The formula is as follows:

[0069] R t =Y t -T t -S t ;

[0070] The residual component reflects the irregular and random part of the noise signal.

[0071] S3: Transformer time series prediction is performed on both seasonal components and trend components to obtain the prediction results of seasonal components and trend components; the structure of Transformer is as follows:

[0072] 1. Multi-head Self-attention: It can simultaneously focus on the relationship between different time steps in the time series and capture complex dependency patterns.

[0073] 2. Positional Encoding: Since the Transformer model cannot inherently perceive the order of time series data, it is necessary to introduce position information to maintain the sequential dependency of the time series.

[0074] 3) In order for the model to understand the time steps in the sequence, positional encoding encodes the time step position into a vector in the following way:

[0075]

[0076] Where pos is the position, i is the index of the dimension, and d modelis the dimension size of the model. This formula uses sine and cosine functions so that the distance between time steps can be represented by the difference between their position encodings, thereby enhancing the time perception ability of the model.

[0077] The self-attention mechanism calculates the dependencies of each time step in the sequence through three matrices Q (Query), K (Key), and V (Value):

[0078]

[0079] in: is the query matrix, K is the key matrix, V is the value matrix, d k is the dimension of the key vector;

[0080] pass The similarity between each time step is calculated, and then normalized by the softmax function to generate attention weights, and then these weights are used to perform weighted summation on V.

[0081] 5) Residual connection and layer normalization: After each self-attention layer and feed-forward network, residual connection and layer normalization are applied:

[0082] LayerNorm(x+SubLayer(x))

[0083] Residual connections prevent gradient vanishing and improve the training stability of deep networks by adding the input directly to the output.

[0084] Through the above Transformer time series prediction, the prediction results of seasonal components and trend components are obtained; among them, the seasonal component is used to capture the periodic law in the noise data, the trend component is used to capture the long-term change trend, and the residual component is used to capture short-term fluctuations.

[0085] S4: Perform anomaly detection on the prediction results of seasonal components and trend components, and output the detection results respectively;

[0086] Fast Fourier transform is used to detect periodic anomalies in the forecast results of seasonal components.

[0087] The seasonal component S t Convert to frequency domain representation:

[0088]

[0089] In the above formula, X(k) is the complex number representation of the kth frequency component, S(n) represents, N is the total number of sampling points of the seasonal component, and the amplitude and phase are calculated as follows:

[0090]

[0091] In the above formula, A(k) represents the amplitude of the kth frequency component, and Φ(k) represents the phase of the kth frequency component; the mean and standard deviation of the amplitude A(k) and phase Φ(k) are calculated respectively:

[0092]

[0093] μ A is the mean value of the amplitude; σ A is the standard deviation of the amplitude; μ Φ is the mean value of the amplitude; σ Φ is the standard deviation of the amplitude; corresponding to the setting of the amplitude dynamic threshold T A and phase dynamic threshold T θ , the formula is as follows:

[0094] T A =μ A +k*σ A , T θ =μ Φ +kσ A ;

[0095] In the above formula, when A(k)>T A For amplitude anomaly, when |Φ(k)|>T θ is a phase anomaly. When both are abnormal, it is recorded as a significant anomaly.

[0096] The Bayesian inflection point detection algorithm is used to detect inflection points of the forecast results of the trend component and identify significant inflection points and abnormal inflection points:

[0097] The formula for modeling the inflection point location is as follows:

[0098] P(τ)=λe -λτ ;

[0099] In the above formula, λ represents the rate parameter of the inflection point. For each time step t, when the inflection point occurs at a certain time step, the mean and variance of the trend components before and after the inflection point are calculated respectively, and the likelihood function is constructed based on the normal distribution as follows:

[0100]

[0101] Among them, T t is the trend component in time, μ before and are the mean and variance before the inflection point; μ after and are the mean and variance after the inflection point, respectively, and τ is the time step where the inflection point is located;

[0102] The posterior probability of the inflection point indicates that when the trend data T is observed 1:nAfter that, the probability that a certain time step is an inflection point is as follows:

[0103]

[0104] Among them, P(T 1:n |τ) is the likelihood function, P(τ) is the prior probability, P(T 1:n ) is the marginal likelihood of the data, and the sum of the posterior probabilities at all time points is 1; for each new time step t+1, the posterior probability is recursively updated as follows:

[0105] P(τ|T 1:t+1 )∞P(T t+1 |τ)P(τ|T 1:t );

[0106] In the above formula, the threshold θ is set to determine the inflection point. When the posterior probability P(τ|T 1:n ) exceeds the threshold θ, then time t is determined to be the inflection point.

[0107] S5: Perform weighted superposition on the results of trend component and seasonal component to generate the final anomaly detection and change inflection point results. The formula is as follows:

[0108] A final (t) = w1A trend (t)+w2A seasonal (t);

[0109] Among them, A final (t) is the final anomaly detection result at time t; A trend (t) and A seasonal (t) are the anomaly detection results of trend component and seasonal component respectively; w1 and w2 are the weighting coefficients of trend component and seasonal component respectively;

[0110] In this embodiment, the trend term shows significant predictive power when reflecting the device status. Through comparative experiments, the model using the trend term as the main feature improves the accuracy by 20% on average in anomaly detection. Because it reflects the long-term noise changes, the period term is second, and the residual term is negligible. Therefore, the weight of the long-term trend term in operation is defined as 0.7, and the period term is 0.3.

[0111] A final (t)=0.7A trend (t)+0.3A seasonal (t);

[0112] At consecutive time steps, the anomaly detection result A final If (t) continues to exceed the threshold, it is determined that these time steps constitute an abnormal interval.

[0113] Therefore, the present invention adopts a transformer noise detection method based on seasonal trend decomposition and time series prediction, uses time series decomposition technology combined with Transformer network to effectively predict substation noise, and detects abnormal noise in advance. By decomposing, predicting and detecting abnormal points of substation noise time series data, it not only improves the intelligent level of noise management, but also provides reliable technical support for status monitoring and fault prevention of substation equipment.

[0114] 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 transformer noise detection method based on seasonal trend decomposition and time series prediction, characterized in that: The following steps are involved: S1: Obtain the noise data of the transformer and preprocess the noise data, including data cleaning and normalization, remove outliers and fill in missing values; S2: Use the STL decomposition model to decompose the preprocessed noise data to obtain seasonal components, trend components and residual components; S3: Perform Transformer time series prediction on both seasonal components and trend components to obtain the prediction results of seasonal components and trend components; S4: Perform anomaly detection on the prediction results of seasonal components and trend components, and output the detection results respectively; S5: Perform weighted compounding on the detection results and output the recognition results.

2. The transformer noise detection method based on seasonal trend decomposition and time series prediction according to claim 1 is characterized in that: In step S2, the noise data decomposition process is as follows: The STL decomposition formula is as follows: Y t =T t +S t +R t ; In the above formula, Y t is the original noise signal; T t is the trend component, reflecting the long-term changes of the noise signal; S t is the seasonal component, capturing the periodic fluctuations. The periodic fluctuations are the noise changes caused by the periodic load of the equipment. t is the residual component, which represents random fluctuations or non-systematic changes in noise; The trend component and seasonality are estimated using local weighted regression Loess, and the estimation formula is as follows: T t =Loess(Y t ,λ trend ); S t =Loess(Y t -T t ,λ seasonal ); In the above formula, λ trend represents the window size parameter for trend smoothing, λ seasonal is the smoothing parameter for the seasonal term, representing the period length of seasonal fluctuations; The residual component is obtained by subtracting the trend term and the period term from the original signal. The formula is as follows: R t =Y t -T t -S t ; The residual component reflects the irregular and random part of the noise signal.

3. The transformer noise detection method based on seasonal trend decomposition and time series prediction according to claim 1 is characterized in that: In step S3, a Transformer model is used for prediction, wherein the seasonal component is used to capture the periodicity in the noise data, and the trend component is used to capture the long-term change trend.

4. The transformer noise detection method based on seasonal trend decomposition and time series prediction according to claim 1 is characterized in that: The specific process of step S4 is as follows: Fast Fourier transform is used to detect periodic anomalies in the forecast results of seasonal components. The seasonal component S t Convert to frequency domain representation: In the above formula, X(k) is the complex number representation of the kth frequency component, S(n) represents the value of the time domain signal at the nth sampling point, N is the total number of sampling points of the seasonal component, and the amplitude and phase are calculated as follows: In the above formula, A(k) represents the amplitude of the kth frequency component, and Φ(k) represents the phase of the kth frequency component; the mean and standard deviation of the amplitude A(k) and phase Φ(k) are calculated respectively: μ A is the mean value of the amplitude; σ A is the standard deviation of the amplitude; μ Φ is the mean value of the amplitude; σ Φ is the standard deviation of the amplitude; corresponding to the setting of the amplitude dynamic threshold T A and phase dynamic threshold T θ , the formula is as follows: T A =μ A +k*s A ,T θ =μ Φ +kσ A ; In the above formula, when A(k)>T A For amplitude anomaly, when |Φ(k)|>T θ is a phase anomaly. When both are abnormal, it is recorded as a significant anomaly. The Bayesian inflection point detection algorithm is used to detect inflection points of the forecast results of the trend component and identify significant inflection points and abnormal inflection points: The formula for modeling the inflection point location is as follows: P(τ)=λe -λτ ; In the above formula, λ represents the rate parameter of the inflection point. For each time step t, when the inflection point occurs at a certain time step, the mean and variance of the trend components before and after the inflection point are calculated respectively, and the likelihood function is constructed based on the normal distribution as follows: Among them, T t is the trend component in time, μ before and are the mean and variance before the inflection point; μ after and are the mean and variance after the inflection point, respectively, and τ is the time step where the inflection point is located; The posterior probability of the inflection point indicates that when the trend data T is observed 1:n After that, the probability that a certain time step is an inflection point is as follows: Among them, P(T 1:n |τ) is the likelihood function, P(τ) is the prior probability, P(T 1:n ) is the marginal likelihood of the data, and the sum of the posterior probabilities at all time points is 1; for each new time step t+1, the posterior probability is recursively updated as follows: P(τ|T 1:t+1 )∞P(T t+1 |τ)P(τ|T 1:t ); In the above formula, the threshold θ is set to determine the inflection point. When the posterior probability P(τ|T 1:n ) exceeds the threshold θ, then time t is determined to be the inflection point.

5. The transformer noise detection method based on seasonal trend decomposition and time series prediction according to claim 4 is characterized in that: In step S5, the results of the trend component and the seasonal component are weighted and superimposed to generate the final anomaly detection and change inflection point results, and the formula is as follows: A final (t)=w1A trend (t)+w2A seasonal (t); Among them, A final (t) is the final anomaly detection result at time t; A trend (t) and A seasonal (t) are the anomaly detection results of trend component and seasonal component respectively; w1 and w2 are the weighting coefficients of trend component and seasonal component respectively. In the continuous time step, the anomaly detection result A final If (t) continues to exceed the threshold, it is determined that these time steps constitute an abnormal interval.

Images