Method for detecting transformer vibration anomaly based on millimeter wave radar

Abstract

The present application relates to the technical field of transformer detection, and particularly relates to a transformer vibration abnormality detection method based on a millimeter wave radar, comprising the following steps: obtaining original reflection signals of a transformer surface, performing time domain and frequency domain analysis on the original reflection signals, and constructing a vibration feature vector; based on the constructed vibration feature vector, performing unsupervised learning on a normal transformer vibration mode through a deep self-encoding network, and automatically learning potential features of the normal vibration mode; converting a new reflection signal into a new vibration feature vector, reconstructing the new vibration feature vector using the deep self-encoding network, calculating a reconstruction error, and if the reconstruction error exceeds a predetermined threshold, determining that it is an abnormal vibration mode; and using an adaptive threshold strategy to automatically adjust the predetermined threshold to adapt to changes in the operating state of different transformer devices over time. The present application introduces an adaptive threshold strategy, which can adapt to the operating characteristics of different transformer devices and fluctuations in the external environment in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer detection, and in particular to a transformer vibration anomaly detection method based on millimeter wave radar. BACKGROUND

[0002] With the continuous development of intelligent devices in the power system, the transformer as an important device in the power transmission and distribution system, its operation state directly affects the stability and safety of power supply, the transformer will be subjected to vibration in varying degrees during long-term operation, and abnormal changes in vibration are often early signals of transformer failure. Therefore, timely and accurate detection of transformer vibration anomalies is of great significance. Traditional transformer vibration anomaly detection methods usually rely on sensor monitoring or mechanical structure-based analysis, which have certain limitations and are difficult to fully and accurately reflect the real state of device operation.

[0003] Currently, the research in the field of transformer vibration anomaly detection is gradually developing from traditional mechanical sensor detection to more intelligent technology. Among them, the vibration monitoring technology based on millimeter wave radar has become a research hotspot in recent years because it can penetrate the external structure of the transformer and monitor the vibration signal without contact. Millimeter wave radar can obtain real-time vibration information of the transformer surface by emitting and receiving millimeter waves. However, the reflected signal received by the millimeter wave radar contains a large amount of noise and interference information, which makes it a challenge to extract accurate vibration features from it.

[0004] Traditional vibration anomaly detection methods usually rely on manually set thresholds or simple statistical methods, such as judgment based on mean, standard deviation, etc. This method not only relies on human experience, but also is prone to misjudgment or omission when facing complex and dynamic working environments, especially when the load, temperature, environment, etc. of the transformer change, simple threshold setting is often difficult to cope with. SUMMARY

[0005] The present application provides a transformer vibration anomaly detection method based on millimeter wave radar.

[0006] The transformer vibration anomaly detection method based on millimeter wave radar comprises the following steps:

[0007] S1, using a millimeter wave radar to emit millimeter waves to the external structure of the transformer, obtaining the original reflected signal of the transformer surface, and performing time domain and frequency domain analysis on the original reflected signal to construct a vibration feature vector;

[0008] S2, training and feature learning of deep auto-encoding network: based on the constructed vibration feature vector, the normal transformer vibration mode is learned unsupervisedly through a deep auto-encoding network, which is composed of an encoder and a decoder, the encoder compresses the input vibration feature vector into a low-dimensional feature representation, and the decoder recovers the original vibration feature vector from the low-dimensional feature, through training the network, the latent features of the normal vibration mode are automatically learned;

[0009] S3, detection of abnormal vibration mode: after training, new reflection signals are collected again, the new reflection signals are converted into new vibration feature vectors through the method of S1, the new vibration feature vectors are reconstructed using the deep auto-encoding network, the reconstruction error is calculated, if the reconstruction error exceeds the predetermined threshold, it is judged as an abnormal vibration mode, and the abnormal mode different from the normal vibration mode is adaptively identified, including the changes in frequency, amplitude and period;

[0010] S4, an adaptive threshold strategy is used to automatically adjust the predetermined threshold to adapt to the changes of the running state of different transformer devices over time.

[0011] Optionally, the time domain and frequency domain analysis of the original reflection signal in S1 to construct the vibration feature vector specifically includes:

[0012] S11, time domain analysis: peak detection is performed on the original reflection signal, the amplitude, periodic change and instantaneous frequency characteristics of the signal are analyzed, specifically including:

[0013] peak detection is performed by the maximum and minimum values of the signal, and the amplitude change of the signal is calculated ;

[0014] the zero-crossing point and periodic peak of the signal are used to analyze the periodic change of the vibration ;

[0015] the instantaneous frequency of the signal is calculated , reflecting the frequency change of the vibration;

[0016] S12, frequency domain analysis: the original reflection signal is converted to the frequency domain through Fourier transform, the frequency spectrum characteristics of the signal are analyzed, and the frequency components related to the vibration mode are extracted, specifically including:

[0017] fast Fourier transform is used to analyze the reflection signal in the frequency domain, the frequency spectrum of the signal is obtained, and the main frequency components in the signal frequency spectrum are extracted ;

[0018] the frequency band distribution in the frequency spectrum is analyzed, and the frequency band width of the vibration signal is calculated ;

[0019] S13, combine the time domain features and the frequency domain features (amplitude, period, frequency) into a vibration feature vector , the vibration feature vector is expressed as:

[0020] ;

[0021] wherein, is the amplitude, is the periodicity, is the main frequency, is the frequency band width, is the instantaneous frequency at the time .

[0022] Optionally, the S2 specifically comprises:

[0023] S21, based on the constructed vibration feature vector , the normal transformer vibration mode is unsupervised learned through a deep auto-encoding network, the deep auto-encoding network comprises an encoder and a decoder, wherein:

[0024] The encoder part: the encoder receives the input vibration feature vector , compresses it through a multi-layer neural network to obtain a low-dimensional feature representation , the encoder gradually reduces the dimension of the input data through several hidden layers to obtain a low-dimensional feature representation , which can capture the key features of the normal vibration mode:

[0025] ;

[0026] The decoder part: the decoder restores the low-dimensional feature representation back to the original vibration feature vector through an inverse neural network structure, that is, the input feature vector is gradually reconstructed through back propagation, and the target is to minimize the reconstruction error, so that the reconstruction result is as close as possible to the input vibration feature vector: ;

[0027] S22, training process: through unsupervised learning, a large number of vibration feature vectors of normal transformers are used for training, and the parameters of the encoder and the decoder are optimized, so that the encoder can effectively compress the normal vibration feature vector to a low-dimensional space, and the original input is accurately reconstructed through the decoder. During the training process, the network parameters are optimized by minimizing the reconstruction error function , wherein the reconstruction error is measured by using a mean square error loss function: ; wherein, represents the vibration feature vector of the th training sample, represents the vibration feature vector of the i-th training sample reconstructed by the auto-encoder network, L represents the reconstruction error of the network, and represents the difference between the input and the reconstructed signal.

[0028] Optionally, the encoder uses a fully connected layer structure to gradually reduce the dimension of the data.

[0029] Suppose the input vibration feature vector is a vector of dimension The task of the encoder is to compress it into a low-dimensional feature representation of dimension , where ;

[0030] The process of the encoder is implemented through the following fully connected layer:

[0031] , where is the weight matrix of the first layer, with a size of , used to map the input data from -dimensional to -dimensional, is the bias term, with a size of -dimensional, is the activation function, using ReLU or Sigmoid activation function, is the low-dimensional feature representation.

[0032] Optionally, the decoder is implemented through an inverse fully connected layer, which maps the low-dimensional feature back to the high-dimensional space step by step, recovering the original vibration feature:

[0033] , where is the weight matrix of the first layer of the decoder, with a size of , used to map the low-dimensional feature back to the high-dimensional space, is the bias term, with a size of dimension, is the activation function.

[0034] Optionally, in S3, a new reflection signal is collected again, which is the transformer vibration data to be detected, and the time domain and frequency domain analysis method in S1 is used to extract the time domain and frequency domain features of the new reflection signal , construct a new vibration feature vector , and use the trained deep auto-encoder network to reconstruct the new vibration feature vector , obtaining the reconstructed vibration feature vector ​​​, the original high-dimensional feature vector is recovered from the low-dimensional feature space by the decoder, and the error is calculated by the mean square error:

[0035] The calculated reconstruction error is compared with a predetermined threshold If the predetermined threshold is exceeded, it is determined that the abnormal vibration mode.

[0036] Optionally, the predetermined threshold in S3 is an initial threshold , which is set based on historical data and statistical analysis, and the initial threshold is set as multiple times of the reconstruction error:

[0037] , wherein is the mean of the reconstruction error of the normal vibration mode, is the standard deviation of the reconstruction error of the normal vibration mode, is a constant multiple for adjusting the sensitivity of the threshold.

[0038] Optionally, the reconstruction error of the vibration feature vector changes over time, and the adaptive threshold strategy in S4 specifically includes:

[0039] The new mean and standard deviation of the reconstruction error are calculated according to the sliding window in each time period, and a sliding window with a size of is set, and the reconstruction error of the vibration feature is calculated in real time within the sliding window:

[0040] ;

[0041] ;

[0042] , wherein represents the mean of all reconstruction errors within the sliding window at time point , which reflects the average level of the reconstruction error in the current window, is the size of the sliding window, represents the reconstruction error value at the th time point, is the index of the current time, indicating the current time point or data position when calculating the mean; represents the standard deviation of all reconstruction errors within the sliding window at time point , which measures the fluctuation amplitude of the reconstruction error in the current window;

[0043] The predetermined threshold is updated as: , wherein is an adjustable constant for dynamically controlling the sensitivity of the threshold value, is the adaptive threshold value at time t.

[0044] Optionally, it further comprises a comprehensive adjustment mechanism for ensuring that the initial threshold value is combined with the updated threshold value Smooth combination, when the system just starts running (less data), prefer to use the initial threshold value as the basis, with more real-time data accumulation, gradually let the threshold value calculated by the sliding window become dominant, dynamically adjust the detection sensitivity, the comprehensive adjustment mechanism introduces a smoothing factor, the results of the initial threshold value and the sliding window updated threshold value are weighted and averaged when adjusting, to avoid sudden changes or excessive dependence on a certain threshold value:

[0045] wherein, is a smoothing factor, and , control the balance between the initial threshold value and the sliding window calculation.

[0046] The beneficial effects of the present application are:

[0047] The present application, based on the transformer vibration anomaly detection method of millimeter wave radar, through the vibration feature learning and reconstruction error calculation of deep auto-encoding network, can identify abnormal vibration mode different from normal vibration mode, compared with traditional method, without complex signal processing process and artificial setting threshold value, but through automatic learning of vibration feature, and according to real-time data, the threshold value is adaptively adjusted, which significantly improves the accuracy of anomaly detection, especially in the complex operation state of transformer load change, fluctuation and the like, the system can flexibly cope with, avoiding false alarm and missing report.

[0048] The present application, through the unsupervised learning of normal vibration mode by deep auto-encoding network, and based on the learned low-dimensional feature representation, the new vibration feature vector is reconstructed, which can effectively capture the potential law and slight change of transformer vibration, combined with deep learning technology, using the structure of encoder and decoder, automatically extracting and compressing the key information of input vibration feature, so as to greatly improve the recognition ability of the system to complex vibration mode, compared with traditional technology based on manual feature extraction, deep auto-encoding network not only can learn the internal characteristics of vibration data more comprehensively, but also can adapt to the subtle changes of transformer state.

[0049] The adaptive threshold strategy is introduced, the detection threshold is automatically adjusted by dynamic calculation based on the sliding window reconstruction error mean and standard deviation, combined with the running state of the transformer and the change of external environment, the adaptive threshold mechanism can adapt to the running characteristics of different transformer devices and the fluctuation of external environment in real time, effectively cope with the influence of load change, temperature fluctuation and other factors on the vibration mode, ensure the stability and high efficiency under different working conditions, avoid the problems of excessive sensitivity or insensitivity caused by fixed threshold by adjusting the threshold flexibly. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0051] Fig. 1 The flowchart of the anomaly detection method of the embodiment of the present application is shown.

[0052] Fig. 2 The threshold updating schematic diagram of the embodiment of the present application is shown. DETAILED DESCRIPTION

[0053] The present application will be described in detail below in combination with the drawings and specific embodiments. It should be noted that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement them; and the drawings are only used to describe the embodiments in more detail, and are not intended to specifically limit the present application.

[0054] It should be noted that in the specification, "one embodiment", "embodiment", "exemplary embodiment", "some embodiments" and the like indicate that the described embodiments can include specific features, structures or characteristics, but not necessarily every embodiment includes the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in combination with an embodiment, it should be within the knowledge of those skilled in the art to realize this feature, structure or characteristic in combination with other embodiments (whether or not explicitly described).

[0055] In general, terms can be understood, at least partly, from usage in the context. For example, depending at least in part on the context, the term "one or more" as used herein can describe any feature, structure, or characteristic in the singular or can describe combinations of features, structures, or characteristics, in the plural, both singular and plural. Further, the term "based on" can be understood as not necessarily being refrained to a set of exclusive factors, but, instead, can allow for additional factors not necessarily explicitly described, again, depending at least in part on the context.

[0056] As shown in Figs. 1-2 The transformer vibration anomaly detection method based on millimeter wave radar comprises the following steps:

[0057] S1, using millimeter wave radar to emit millimeter wave to the external structure of the transformer, obtaining the original reflection signal of the transformer surface, performing time domain and frequency domain analysis on the original reflection signal, and constructing a vibration feature vector;

[0058] S2, training and feature learning of deep auto-encoding network: based on the constructed vibration feature vector, the deep auto-encoding network is used for unsupervised learning of normal transformer vibration mode, the deep auto-encoding network is composed of an encoder and a decoder, the encoder compresses the input vibration feature vector into a low-dimensional feature representation, and the decoder restores the original vibration feature vector from the low-dimensional feature, and the potential features of the normal vibration mode are automatically learned through training the network;

[0059] S3, detection of abnormal vibration mode: after training, new reflection signals are collected again, the new reflection signals are converted into new vibration feature vectors through the method of S1, the deep auto-encoding network is used for reconstruction of the new vibration feature vectors, the reconstruction error is calculated, and if the reconstruction error exceeds a predetermined threshold, it is judged as an abnormal vibration mode, and the abnormal mode different from the normal vibration mode is adaptively identified, including changes in frequency, amplitude and period;

[0060] S4, as the running environment and state of the transformer change, the predetermined threshold needs to be dynamically adjusted, an adaptive threshold strategy is adopted to automatically adjust the predetermined threshold to adapt to the change of the running state of different transformer devices over time.

[0061] The time domain and frequency domain analysis of the original reflection signal in S1 to construct the vibration feature vector specifically comprises:

[0062] S11, time domain analysis: peak detection is performed on the original reflection signal, and the amplitude, periodic change and instantaneous frequency characteristics of the signal are analyzed, specifically including:

[0063] The peak value detection is performed by the maximum value and the minimum value of the signal, and the amplitude change of the signal is calculated ;

[0064] Using zero-crossing points and periodic peaks of the signal, analyze the periodic variation of the vibration ;

[0065] Calculate the instantaneous frequency of the signal , reflecting the frequency variation of the vibration;

[0066] S12, frequency domain analysis: convert the original reflection signal to the frequency domain by Fourier transform, analyze the spectral characteristics of the signal, extract the frequency components related to the vibration mode, including:

[0067] Using fast Fourier transform to analyze the reflection signal in the frequency domain, obtaining the frequency spectrum of the signal, extracting the main frequency components in the signal spectrum ;

[0068] Analyze the frequency band distribution in the spectrum, calculate the frequency band width of the vibration signal ;

[0069] S13, combine the time domain features and frequency domain features (amplitude, period, frequency) into a vibration feature vector , the vibration feature vector is represented as:

[0070] ;

[0071] Where, is the amplitude, is the periodicity, is the main frequency, is the frequency band width, is the instantaneous frequency at time .

[0072] As input data for subsequent deep auto-encoding network training.

[0073] 1. Time domain analysis part.

[0074] 1.1, amplitude calculation: define the reflection signal as the original reflection signal varying with time, after peak detection, take the maximum value and the minimum value of the signal, then the amplitude is defined as: , where, is the maximum value of the signal , and is the minimum value of the signal ;

[0075] 1.2, periodicity analysis and period calculation: periodicity analysis is based on the zero-crossing points of the signal, which are the turning points of the signal from positive to negative or from negative to positive;

[0076] Period calculation formula: define the zero-crossing points of the signal , where the zero-crossing point represents the time point when the signal crosses the zero axis, and the time difference between consecutive zero-crossing points is calculated:

[0077] , where is the period of the signal, representing the average time difference between consecutive zero-crossing points, the time of the ith zero-crossing point.

[0078] 1.3, Instantaneous frequency calculation: the instantaneous frequency reflects the frequency change of the signal at a certain time, which is usually calculated by the Hilbert transform,

[0079] Using the Hilbert transform , we get the analytic signal of the signal:

[0080] ;

[0081] The instantaneous frequency of the analytic signal is obtained by the phase derivative:

[0082] ; where is the Hilbert transform of the signal , is the complex signal composed of the original signal and its Hilbert transform, is the phase of the signal , is the instantaneous frequency, representing the frequency of the signal at a certain time.

[0083] 2. Frequency domain analysis section.

[0084] 2.1, Fourier transform: convert the original reflection signal into a frequency domain signal by fast Fourier transform, and analyze the frequency components of the signal:

[0085] , where is the frequency domain representation of the signal , representing the intensity of different frequency components, is the Fourier transform operator, is the frequency.

[0086] 2.2, frequency spectrum analysis and frequency component extraction: frequency spectrum analysis is used to extract the main frequency components in the signal, especially the fundamental frequency and harmonics of vibration;

[0087] Calculate the frequency corresponding to the maximum value of the frequency spectrum : , where​ is the dominant frequency of the signal, representing the fundamental frequency of the vibration, is the amplitude of the signal at frequency .

[0088] 2.3, Bandwidth and Amplitude Analysis: Bandwidth is represented by the half-peak width in the frequency spectrum, reflecting the frequency distribution of the signal, and the bandwidth is calculated as: , where, is the highest frequency in the signal spectrum, is the lowest frequency in the signal spectrum.

[0089] S2 specifically includes:

[0090] S21, based on the constructed vibration feature vector , unsupervised learning of normal transformer vibration patterns through a deep auto-encoding network, the deep auto-encoding network includes an encoder and a decoder, wherein:

[0091] Encoder part: the encoder receives the input vibration feature vector , which is compressed through a multi-layer neural network to obtain a low-dimensional feature representation , the encoder gradually reduces the dimension of the input data through several hidden layers to obtain a low-dimensional feature representation , which can capture the key features of normal vibration patterns:

[0092] ;

[0093] Decoder part: the decoder restores the low-dimensional feature representation to the original vibration feature vector through an inverse neural network structure, that is, it gradually reconstructs the input feature vector through backpropagation, the goal is to minimize the reconstruction error so that the reconstruction result is as close as possible to the input vibration feature vector: ;

[0094] S22, training process: through unsupervised learning, a large number of vibration feature vectors of normal transformers are used for training to optimize the parameters of the encoder and decoder, so that the encoder can effectively compress the normal vibration feature vector to a low-dimensional space, and accurately reconstruct the original input through the decoder. During the training process, the network parameters are optimized by minimizing the reconstruction error function , where the reconstruction error is measured by the mean square error loss function: ; where, represents the vibration feature vector of the th training sample, represents the vibration feature vector of the The vibration feature vector of a training sample is reconstructed by the auto-encoder network, L represents the reconstruction error of the network, and represents the difference between the input and the reconstructed signal.

[0095] The encoder uses a fully connected layer structure to gradually reduce the dimensionality of the data;

[0096] Suppose the input vibration feature vector is a -dimensional vector , the task of the encoder is to compress it into a -dimensional low-dimensional feature representation , where ;

[0097] The process of the encoder is implemented through the following fully connected layer:

[0098] where, is the weight matrix of the first layer, with a size of , used to map the input data from -dimensional to -dimensional, is the bias term, with a size of -dimensional, is the activation function, using ReLU or Sigmoid activation function, is the low-dimensional feature representation;

[0099] The dimensionality of the data can also be further reduced step by step through multiple fully connected layers:

[0100] Here, and are the weight matrix and bias term of the second layer, further mapping the data.

[0101] The decoder is implemented through the reverse fully connected layer, which maps the low-dimensional feature back to the high-dimensional space step by step, recovering the original vibration feature:

[0102] where, is the weight matrix of the first layer of the decoder, with a size of , used to map the low-dimensional feature back to the high-dimensional space, is the bias term, with a size of dimension, is the activation function, and similarly, the decoder can also use multiple reverse fully connected layers:

[0103] Here, and are the second layer weight and bias term of the decoder.

[0104] S3, a new reflection signal is collected again , which is the transformer vibration data to be detected, the time domain and frequency domain analysis method in S1 is used to perform time domain and frequency domain feature extraction on the new reflection signal , and a new vibration feature vector is constructed , the trained deep auto-encoding network is used to reconstruct the new vibration feature vector , and the reconstructed vibration feature vector is obtained , the original high-dimensional feature vector is recovered from the low-dimensional feature space through the decoder, and the error is calculated by the mean square error:

[0105] , the calculated reconstruction error is compared with the predetermined threshold , if exceeds the predetermined threshold , it is judged as an abnormal vibration mode.

[0106] The predetermined threshold in S3 is the initial threshold , which is set based on historical data and statistical analysis, and the initial threshold is set as multiple times of the reconstruction error:

[0107] , wherein is the mean of the reconstruction error of the normal vibration mode, is the standard deviation of the reconstruction error of the normal vibration mode, is a constant multiple, which is used to adjust the sensitivity of the threshold.

[0108] The reconstruction error of the vibration feature vector will change over time, and the adaptive threshold strategy in S4 specifically includes:

[0109] According to the sliding window in each period of time, the new reconstruction error mean and standard deviation are calculated, and a sliding window with a size of is set, and the reconstruction error of the vibration feature is calculated in real time in the sliding window:

[0110] ;

[0111] ;

[0112] , wherein represents the mean value based on all reconstruction errors in the sliding window at time point , which reflects the average level of the reconstruction error in the current window, is the size of the sliding window, represents the reconstruction error value at the th time point, It is the index of the current moment, indicating the current time point or data position when calculating the mean; Indicates at a point in time At any given time, the fluctuation range of the reconstruction error within the current window is measured based on the standard deviation of all reconstruction errors within the sliding window.

[0113] The pre-defined threshold has been updated to: ,in It is an adjustable constant used to dynamically control the sensitivity of the threshold. It is the adaptive threshold at time t.

[0114] It also includes a comprehensive adjustment mechanism to ensure the initial threshold. With the updated threshold A balanced approach is adopted: when the system is just starting out (with limited data), the initial threshold is used as the basis. As more real-time data accumulates, the threshold calculated by the sliding window gradually becomes dominant, dynamically adjusting the detection sensitivity. A smoothing factor is introduced into the comprehensive adjustment mechanism. During adjustment, the results of the initial threshold and the threshold updated by the sliding window are weighted and averaged to avoid abrupt changes or over-reliance on any single threshold.

[0115] ,in, It is a smoothing factor, and This controls the balance between the initial threshold and the sliding window calculation.

[0116] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0117] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting transformer vibration anomaly based on millimeter wave radar, characterized in that, The method comprises the following steps: S1, using millimeter wave radar to emit millimeter wave to the external structure of the transformer, obtaining the original reflection signal of the transformer surface, performing time domain and frequency domain analysis on the original reflection signal, and constructing a vibration feature vector; S2, training and feature learning of a deep auto-encoding network: based on the constructed vibration feature vector, the deep auto-encoding network is used to perform unsupervised learning on the normal transformer vibration mode, the deep auto-encoding network is composed of an encoder and a decoder, the encoder compresses the input vibration feature vector into a low-dimensional feature representation, and the decoder restores the original vibration feature vector from the low-dimensional feature, and through training the network, the latent features of the normal vibration mode are automatically learned; S3, detection of abnormal vibration mode: after training, new reflection signals are collected again, the new reflection signals are converted into new vibration feature vectors through the method of S1, the deep auto-encoding network is used to reconstruct the new vibration feature vectors, the reconstruction error is calculated, and if the reconstruction error exceeds a predetermined threshold, it is judged as an abnormal vibration mode, and an abnormal mode different from the normal vibration mode is adaptively identified, including changes in frequency, amplitude and period; S4, an adaptive threshold strategy is used to automatically adjust the predetermined threshold to adapt to the change of the running state of different transformer devices over time; The reconstruction error of the vibration feature vector changes over time, and the adaptive threshold strategy in S4 specifically comprises: The new reconstruction error mean and standard deviation are calculated according to the sliding window in each time period, and a sliding window with a size of is set to calculate the reconstruction error of the vibration feature in real time within the sliding window. ； ； wherein, represents at time point , based on the mean of all reconstruction errors within the sliding window, this value reflects the average level of reconstruction errors within the current window, is the size of the sliding window, represents at the time point, the corresponding reconstruction error value, is the index of the current time, indicating the current time point or data position when calculating the mean; represents at time point , based on the standard deviation of all reconstruction errors within the sliding window, measures the fluctuation amplitude of reconstruction errors within the current window; The predetermined threshold is updated as: wherein is an adjustable constant for dynamically controlling the sensitivity of the threshold, is the adaptive threshold at time t; Also included is a comprehensive adjustment mechanism to ensure the initial threshold is updated Smooth combination, in the early stage of operation, the initial threshold is used as the basis, and as more real-time data accumulates, the threshold calculated by the sliding window gradually becomes dominant, dynamically adjusting the detection sensitivity. The comprehensive adjustment mechanism introduces a smoothing factor, which weights the results of the initial threshold and the updated threshold of the sliding window during adjustment, avoiding sudden changes or excessive reliance on a single threshold: wherein, is a smoothing factor, and , controls the degree of balance between the initial threshold and the sliding window calculation.

2. The millimeter-wave radar-based transformer vibration anomaly detection method according to claim 1, characterized by, The time domain and frequency domain analysis of the original reflection signal in S1 to construct the vibration feature vector specifically comprises: S11, time domain analysis: peak detection is performed on the original reflection signal, and the amplitude, periodic change and instantaneous frequency characteristics of the signal are analyzed, specifically comprising: Peak detection by maximum and minimum of the signal, calculating the amplitude variation of the signal ; Using zero-crossings and periodic peaks of the signal, periodic variations of the vibration are analyzed ; Computing the instantaneous frequency of a signal reflecting a frequency change of the vibration; S12, frequency domain analysis: the original reflection signal is converted to the frequency domain through Fourier transform, the frequency spectrum characteristics of the signal are analyzed, and the frequency components related to the vibration mode are extracted, specifically comprising: The frequency spectrum of the signal is obtained by using fast Fourier transform to perform frequency domain analysis on the reflection signal, and main frequency components in the signal frequency spectrum are extracted ; analyzing the distribution of frequency bands in the spectrum, calculating the frequency band width of the vibration signal ; S13, combine the time domain feature and the frequency domain feature into a vibration feature vector vibration feature vector is expressed as: ； wherein, is the amplitude, is the periodicity, is the fundamental frequency, is the frequency bandwidth, is the time instant is the instantaneous frequency.

3. The millimeter-wave radar-based transformer vibration anomaly detection method according to claim 2, characterized by, S2 specifically comprises: S21, based on the constructed vibration feature vector by unsupervised learning of normal transformer vibration patterns through a deep auto-encoder network, the deep auto-encoder network comprising an encoder and a decoder, wherein: Encoder part: the encoder receives the input vibration feature vector , and compresses it through a multi-layer neural network to obtain a low-dimensional feature representation The encoder gradually reduces the dimension of the input data through several hidden layers to obtain a low-dimensional feature representation The low-dimensional feature representation can capture the key features of normal vibration patterns Decoder part: The decoder reconstructs the low-dimensional feature representation by an inverse neural network structure i.e. by backpropagation step by step, the goal is to minimize the reconstruction error so that the reconstruction result is close to the input vibration feature vector: ; S22, training process: through unsupervised learning, the vibration feature vector of the normal transformer is used for training, the parameters of the encoder and the decoder are optimized, so that the encoder compresses the normal vibration feature vector to a low-dimensional space, and accurately reconstructs the original input through the decoder, in the training process, the network parameters are optimized by minimizing the reconstruction error function , wherein the reconstruction error is measured by a mean square error loss function: ; wherein, represents the vibration feature vector of the th training sample, represents the vibration feature vector of the th training sample reconstructed by the auto-encoding network, L represents the reconstruction error of the network, and represents the difference between the input and the reconstructed signal.

4. The millimeter-wave radar-based transformer vibration anomaly detection method according to claim 3, characterized by, The encoder uses a fully connected layer structure to gradually reduce the dimensionality of the data; Let the input be a -dimensional vector , the task of the encoder is to compress it into a -dimensional low-dimensional feature representation where ; The process of the encoder is implemented through the following fully connected layers: wherein, is a weight matrix of the first layer of size for mapping input data from -dimensional to -dimensional, is a bias term of size -dimensional, is an activation function, is a low-dimensional feature representation.

5. The millimeter-wave radar-based transformer vibration anomaly detection method according to claim 4, characterized by, The decoder is implemented by a reverse fully connected layer, which maps back the low-dimensional features to the original vibrational features in the high-dimensional space step by step. wherein, is a weight matrix of the first layer of the decoder of size , used to map the low-dimensional features back to the high-dimensional space, is a bias term of size dimensions, is an activation function.

6. The millimeter-wave radar-based transformer vibration anomaly detection method according to claim 5, characterized by, The S3 again collects a new reflection signal The signal is transformer vibration data to be detected, and the time domain and frequency domain analysis method in S1 is used to perform time domain and frequency domain feature extraction on the new reflection signal to construct a new vibration feature vector The trained deep auto-encoding network is used to reconstruct the new vibration feature vector to obtain a reconstructed vibration feature vector The original high-dimensional feature vector is recovered from the low-dimensional feature space through the decoder, and the error is calculated by the mean square error: The calculated reconstruction error is compared with a predetermined threshold value If the predetermined threshold value is exceeded an abnormal vibration pattern is determined.

7. The millimeter-wave radar-based transformer vibration anomaly detection method according to claim 6, characterized by, The predetermined threshold in S3 is an initial threshold , set based on historical data and statistical analysis, the initial threshold is set as a multiple of the reconstruction error: wherein, is the mean of the reconstruction error of the normal vibration mode, is the standard deviation of the reconstruction error of the normal vibration mode, is a constant factor for adjusting the sensitivity of the threshold.

Images