Wavelet analysis based vibration signal denoising and feature extraction method
By combining dual-tree complex wavelet decomposition and adaptive thresholding with energy enhancement and LSTM network optimization, the problems of noise masking and one-sided feature diagnosis in traditional wavelet transform are solved. This achieves high-precision noise reduction and multi-dimensional feature fusion of cable vibration signals, improving the accuracy and robustness of fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID BEIJING ELECTRIC POWER CO
- Filing Date
- 2026-04-24
- Publication Date
- 2026-08-04
AI Technical Summary
In existing technologies, most vibration signal noise reduction and feature extraction schemes adopt traditional discrete wavelet transform, which makes it difficult to distinguish noise from signals. In particular, weak fault features are masked by noise in low signal-to-noise ratio environments. Furthermore, they only focus on some sub-bands and ignore low-frequency sub-band information, resulting in one-sided fault diagnosis, poor noise reduction effect, and low diagnostic accuracy.
A dual-tree complex wavelet decomposition combined with a hierarchical adaptive threshold analysis model is adopted. Through adaptive threshold processing and energy enhancement, signal components that meet the impulse pulse index and frequency band energy ratio range are retained. Combined with LSTM neural network optimization, time-frequency, spatial and full subband features are extracted to achieve multi-dimensional feature fusion.
It achieves high-precision noise reduction of cable vibration signals, enhances the robustness and diagnostic accuracy of fault characteristics, solves the problems of noise masking and one-sided feature diagnosis in traditional methods, and improves the high precision and robustness of cable monitoring.
Smart Images
Figure CN122508002A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cable monitoring technology, specifically relating to a vibration signal denoising and feature extraction method based on wavelet analysis. Background Technology
[0002] In cable external damage monitoring, vibration signals are often affected by environmental noise and sensor noise, obscuring fault characteristics. Wavelet analysis is needed for noise reduction and feature extraction of the vibration signals. Wavelet analysis is a time-frequency analysis method that decomposes the signal into subspaces of different frequency components through multi-scale decomposition, while preserving localized information in both the time and frequency domains. Wavelets are functions with finite energy and rapid decay characteristics. By scaling and translating, a set of basis functions is generated for signal decomposition. Unlike the fixed sinusoidal basis of Fourier transform, the shape of the wavelet basis can adaptively adjust according to the signal characteristics, making it more suitable for the analysis of non-stationary signals.
[0003] Most existing vibration signal denoising and feature extraction schemes employ traditional discrete wavelet transform, which is sensitive to minute signal shifts and can lead to drastic changes in decomposition coefficients, making it difficult to distinguish noise from the signal. This is especially problematic in low signal-to-noise ratio environments, where subtle fault features may be completely masked by noise. Furthermore, extracting only a single type of feature cannot comprehensively characterize the signal's time-frequency, statistical, and spatial properties. For example, relying solely on energy features may fail to distinguish high-frequency noise from actual fault impacts. Traditional wavelet analysis may focus only on certain subbands, neglecting vibration information contained in low-frequency subbands, resulting in incomplete fault diagnosis, poor denoising performance, and low accuracy in fault diagnosis due to the extracted features. Consequently, these schemes struggle to meet the high-precision, high-robustness, and real-time requirements of cable vibration monitoring under complex operating conditions. Summary of the Invention
[0004] The purpose of this invention is to provide a vibration signal denoising and feature extraction method based on wavelet analysis, so as to at least solve or improve the technical problems in the prior art, which only focuses on some sub-bands and ignores the vibration information contained in the low-frequency sub-bands, resulting in one-sided fault diagnosis, poor denoising effect, and low accuracy of fault diagnosis of extracted features.
[0005] A vibration signal denoising and feature extraction method based on wavelet analysis includes the following steps: The vibration signal of the cable is acquired, and the vibration signal is decomposed into a dual-tree complex wavelet to obtain detail coefficients of multiple decomposition levels, including real part detail coefficients and imaginary part detail coefficients. A hierarchical adaptive threshold analysis model is established based on the statistical characteristics of environmental noise. The model is used to perform threshold processing on the detail coefficients of each layer. The detail coefficients whose absolute values are less than the adaptive threshold of the corresponding layer are determined as noise correlation coefficients and set to zero. Based on the impact pulse index and frequency band energy ratio of the vibration signal of the transient impact of cable external failure in historical data after dual-tree complex wavelet decomposition, the impact pulse index range and frequency band energy ratio range are determined. The vibration signal after threshold processing is filtered using the impact pulse index range and frequency band energy ratio range to retain the signal components that meet the range requirements. The detail factor after filtration is enhanced by energy to obtain the enhanced detail factor. Based on the enhanced detail coefficients and decomposition levels, the time-frequency entropy and instantaneous frequency are calculated, the time-frequency joint features are extracted, and the spatial correlation coefficient is calculated by collecting data from several deployed vibration sensors to extract spatial features. Based on the full subband of the vibration signal after dual-tree complex wavelet decomposition, energy features, entropy features, and statistical features are extracted. The energy features, entropy features, and statistical features are then concatenated with the extracted time-frequency joint features and spatial features to form a detail-enhanced feature vector. The detail-enhanced feature vectors are input into the trained neural network model based on LSTM and fully connected layers, and the output is a denoised and optimized feature vector.
[0006] The above scheme obtains the real and imaginary detail coefficients through dual-tree complex wavelet decomposition, uses hierarchical adaptive thresholding for noise filtering, combines historical external impact characteristics for targeted filtering, and then solves the problems of traditional discrete wavelet transform being sensitive to small translations leading to drastic changes in decomposition coefficients, fault features being easily masked by noise under low signal-to-noise ratio, and fault diagnosis being one-sided due to the extraction of only a single type of feature through energy enhancement, multi-dimensional feature fusion, and robustness enhancement for cable vibration signals.
[0007] Furthermore, the hierarchical adaptive threshold analysis model calculates the adaptive threshold for each layer at the corresponding time using the following formula: T j (t)=
[0008] Among them, T j (t) represents the adaptive threshold of the j-th layer at time t. Let be the standard deviation of the local noise of the j-th layer at time t. Let be the real-time signal-to-noise ratio of the j-th layer at time t, and N be the total number of discrete sampling points of the vibration signal. and To adjust the parameters, is the scaling factor; J is the number of decomposition layers.
[0009] The above scheme introduces a hierarchical adaptive threshold calculation formula that includes real-time signal-to-noise ratio and scale adjustment factor, which enables the threshold to be dynamically adjusted according to the time-varying characteristics of noise and the number of decomposition layers. This solves the problem of over-smoothing or distortion caused by traditional global threshold processing, thereby achieving refined noise suppression of detail coefficients of each layer and improving the adaptability and accuracy of noise reduction.
[0010] Furthermore, based on the impact pulse index and frequency band energy ratio of the vibration signal from the transient impact of cable external failure in historical data after dual-tree complex wavelet decomposition, the impact pulse index range and frequency band energy ratio range are determined, specifically including: The vibration signal of transient impact from external failure in historical data is decomposed by dual-tree complex wavelet decomposition. The impact pulse index of the real part and the impact pulse index of the imaginary part of the vibration signal are calculated respectively. The fluctuation range of the impact pulse index of the real part and the imaginary part are respectively used as the impact pulse index range of the real part and the impact pulse index range of the imaginary part. Calculate the frequency band energy ratio of the real part and the frequency band energy ratio of the imaginary part of the vibration signal respectively, and take the fluctuation range of the frequency band energy ratio of the real part and the frequency band energy ratio of the imaginary part as the real part frequency band energy ratio interval and the imaginary part frequency band energy ratio interval respectively. The vibration signal after threshold processing is filtered by using the real part impact pulse index range, the imaginary part impact pulse index range, the real part frequency band energy ratio range, and the imaginary part frequency band energy ratio range.
[0011] The above scheme establishes impact pulse index ranges and frequency band energy ratio ranges for the real and imaginary parts respectively, and uses the statistical characteristics of historical external transient impacts to perform dual filtering on vibration signals, thus solving the problem of difficulty in distinguishing between broadband environmental noise and external impacts. This achieves effective screening of external transient impacts and reduces the misjudgment rate.
[0012] Furthermore, the impact pulse index is calculated using the following formula: Real impact pulse index:
[0013] in, The SPI impact pulse index represents the real part of the vibration signal. Let be the real detail coefficients of the j-th layer at time t; Let be the mean of the real part detail coefficients of the j-th layer at time t. Let be the standard deviation of the real part detail coefficients of the j-th layer at time t. As a regulating factor; Imaginary impulse pulse index:
[0014] in, The SPI impact pulse index represents the imaginary part of the vibration signal. Let be the imaginary detail coefficients of the j-th layer at time t; Let be the mean of the imaginary detail coefficients of the j-th layer at time t. Let be the standard deviation of the imaginary detail coefficients of the j-th layer at time t.
[0015] The above scheme solves the problem of inaccurate quantification of external impact intensity by using the impact pulse index formula and normalizing the real and imaginary parts to calculate the maximum value using the mean and standard deviation of the detail coefficients. This achieves accurate characterization of transient impact intensity and provides a reliable basis for setting the filter threshold.
[0016] Furthermore, the detail factor after filtering is enhanced with energy using the following formula: D j (t)′= D j (t)*(1+α2*SNR j ) Among them, D j (t)′ represents the detail coefficient of the j-th layer after energy enhancement at time t, D j (t) represents the detail coefficients of the j-th layer after wavelet decomposition at time t, and α2 is the enhancement coefficient; SNR j Let be the signal-to-noise ratio of the coefficients in the j-th layer.
[0017] The above scheme introduces an energy enhancement formula based on signal-to-noise ratio to adaptively amplify the filtered detail coefficients, solving the problem that weak fault features are easily masked by noise in low signal-to-noise ratio environments. This achieves a prominent enhancement of fault-related features, providing a clearer input signal for subsequent time-frequency analysis.
[0018] Furthermore, the signal-to-noise ratio (SNR) of the j-th layer coefficients j Represented as:
[0019] in, The total energy of the original detail coefficients at layer j; denoted as the total energy of the noise correlation coefficients filtered out after threshold processing by the hierarchical adaptive threshold analysis model for the coefficients of the j-th layer.
[0020] The above scheme solves the problem of difficulty in quantifying the enhancement coefficient in energy enhancement by calculating the signal-to-noise ratio using the ratio of the total energy of the original detail coefficients to the total energy of the filtered noise. This enables the objective determination of the energy enhancement amplitude and ensures the adaptability and stability of the enhancement effect.
[0021] Furthermore, based on the enhanced detail coefficients and the number of decomposition layers, the time-frequency entropy and instantaneous frequency are calculated, and joint time-frequency features are extracted, specifically including: Calculate the time-frequency entropy using the following formula:
[0022]
[0023] in, Let the time-frequency entropy be at the j-th layer and time t. This represents the time-frequency energy distribution, where k is the frequency index. Let the real part detail coefficients be the coefficients at the j-th level, time t, and frequency index k. These are the imaginary detail coefficients at level j, time t, and frequency index k. Instantaneous frequencies are extracted from the enhanced detail coefficients using Hilbert transform or synchronous squeeze wavelet transform.
[0024] The above scheme quantifies the disorder of time-frequency energy distribution by calculating time-frequency entropy and extracts instantaneous frequency by combining Hilbert transform or synchronous squeezing wavelet transform. This solves the problem that the time-frequency characteristics of non-stationary vibration signals are difficult to fully characterize, thereby achieving dual capture of signal complexity and frequency change law and improving the sensitivity of fault identification.
[0025] Furthermore, the spatial correlation coefficient is calculated from the data collected from several deployed vibration sensors using the following formula:
[0026] in, Let be the spatial correlation coefficient between sensor a and sensor b at time t; Let be the real part detail coefficients of sensor a at time t; Let be the real part detail coefficients of sensor b at time t; Let be the imaginary detail coefficients of sensor a at time t; Let be the imaginary detail coefficients of sensor b at time t; Let be the variance of the real part detail coefficients of sensor a at time t; Let be the variance of the detail coefficients of the imaginary part of sensor a at time t; Let be the variance of the real part detail coefficients of sensor b at time t; Let be the variance of the imaginary detail coefficients of sensor b at time t.
[0027] The above scheme obtains the spatial correlation coefficient by calculating the ratio of the joint covariance to the variance of the detail coefficients of the real and imaginary parts among different sensors. This solves the problem of misjudgment caused by environmental noise or installation error of a single sensor, thereby realizing the quantification of spatial correlation between multi-sensor data and providing effective support for the location of external damage sources and the screening of abnormal data.
[0028] Furthermore, based on the full subband of the vibration signal after dual-tree complex wavelet decomposition, energy features, entropy features, and statistical features are extracted, specifically including: Based on each sub-band signal of the vibration signal after dual-tree complex wavelet decomposition, the real part and imaginary part are extracted respectively. The subband energy of the real and imaginary parts of each subband is extracted as energy features, the subband entropy of the real and imaginary parts of each subband is extracted as entropy features, and the subband kurtosis of the real and imaginary parts of each subband is extracted as statistical features.
[0029] The above scheme extracts energy, entropy, and kurtosis features from the real and imaginary parts of each subband after dual-tree complex wavelet decomposition, solving the problem that traditional wavelet analysis only focuses on some subbands, resulting in the neglect of low-frequency vibration information. This enables the comprehensive extraction of full-band signal characteristics and provides a richer statistical feature basis for fault diagnosis.
[0030] Furthermore, training methods for neural network models based on LSTM and fully connected layers include: The noise signal and the vibration signal of transient impact from external damage are collected from the original vibration signal of the cable in the statistical historical data. Energy features, entropy features, statistical features, time-frequency joint features, and spatial features are extracted from the noise signal and concatenated into a target noise feature vector; Energy features, entropy features, statistical features, time-frequency joint features, and spatial features are extracted from the vibration signal of external transient impact and concatenated into a feature vector of external transient impact. A neural network model based on LSTM and fully connected layers is trained using the target noise feature vector and the feature vector of external transient impact. The trained network model can identify noise signals and external transient impacts in the detail-enhanced feature vector, filter the noise signals, increase the weight of the external transient impact signals, and output the optimized feature vector.
[0031] The above scheme trains an LSTM network using the feature vectors of historical noise signals and external impact signals, enabling the network to identify and distinguish between noise and impact, and to enhance the weights of the impact signals. This solves the problem that traditional methods cannot adaptively optimize feature vectors, thereby achieving end-to-end adaptive noise reduction and fault feature enhancement, and improving the model's generalization ability and diagnostic accuracy. Attached Figure Description
[0032] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating the vibration signal denoising and feature extraction method based on wavelet analysis according to an embodiment of the present invention.
[0033] Figure 2 This is a block diagram of the vibration signal denoising and feature extraction system based on wavelet analysis according to an embodiment of the present invention. Detailed Implementation
[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.
[0035] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.
[0036] Example 1 This embodiment provides a vibration signal denoising and feature extraction method based on wavelet analysis. The method achieves high-precision denoising and fault feature enhancement of cable vibration signals through dual-tree complex wavelet decomposition, hierarchical adaptive thresholding, transient impact filtering, energy enhancement, multi-dimensional feature extraction and LSTM network optimization.
[0037] like Figure 1 As shown, the vibration signal denoising and feature extraction method based on wavelet analysis includes the following steps: S1. Obtain the vibration signal of the cable, perform dual-tree complex wavelet decomposition on the vibration signal, and obtain detail coefficients of multiple decomposition levels, including real part detail coefficients and imaginary part detail coefficients.
[0038] Specifically, the collected cable vibration signal is decomposed into multiple scales using Dual-Tree Complex Wavelet Transform (DT-CWT). The decomposition formula is as follows: f(t) =
[0039] Where f(t) is the vibration signal of the j-th layer at time t, j=1,2,…,J; J is the number of decomposition layers; The approximate coefficients of the j-th layer after DT-CWT decomposition at time t are: Let be the real part detail coefficients of the j-th layer at time t. Let be the imaginary detail coefficients of the j-th layer at time t. and These correspond to vibration modes in different directions; Corresponding to the vibration mode in the horizontal direction, The vibration mode corresponds to the vertical direction; in this embodiment, the real part detail coefficient corresponds to the horizontal direction detail coefficient, and the imaginary part detail coefficient corresponds to the vertical direction detail coefficient.
[0040] Dual-tree complex wavelet decomposition achieves decomposition in different directions through two independent wavelet trees, and can simultaneously extract the characteristics of vibration signals in the horizontal and vertical directions. For example, the impact signal of cable external rupture may contain vibration modes in specific directions. Dual-tree complex wavelets can separate these directional components and quantify the impact intensity through subsequent frequency band energy ratios, thereby improving the identifiability of fault characteristics.
[0041] S2. Establish a hierarchical adaptive threshold analysis model based on the statistical characteristics of environmental noise. Use the hierarchical adaptive threshold analysis model to perform threshold processing on the detail coefficients of each layer. Detail coefficients with absolute values less than the corresponding layer's adaptive threshold are determined as noise correlation coefficients and set to zero.
[0042] Specifically, dual-tree complex wavelets map signals to different frequency bands through multi-scale decomposition. Combining this with the statistical characteristics of environmental noise, adaptive thresholds can be designed for wavelet coefficients at each scale. For example, hard thresholds are used to suppress noise in the high-frequency sub-band, while soft thresholds are used in the low-frequency sub-band to preserve signal details, avoiding over-smoothing or distortion caused by traditional global thresholds. For transient impact signals from cable breakage, the tightly supported wavelet basis of dual-tree complex wavelets can accurately capture the transient characteristics of the impact, while layered thresholding can suppress background noise while preserving the impact amplitude and time-domain waveform, providing a reliable basis for fault location. The layered adaptive thresholding analysis model yields the first... j The adaptive threshold of the layer at time t, for the real part detail coefficients of the j-th layer at time t. and the detail coefficient of the imaginary part The coefficients whose absolute values are less than the corresponding adaptive threshold are set to zero, and the corresponding coefficients are determined to be noise correlation coefficients.
[0043] S3. Based on the vibration signal of the transient impact of cable external damage in historical data, after dual-tree complex wavelet decomposition, determine the impact pulse index range and the frequency band energy ratio range. Use the impact pulse index range and the frequency band energy ratio range to filter the vibration signal after threshold processing, and retain the signal components that meet the range requirements.
[0044] Specifically, after performing dual-tree complex wavelet decomposition on the vibration signals of transient impacts from external cable failures in historical data, the intensity and frequency band energy ratio of the transient impact vibration signals are analyzed. Based on the analysis results, the cable vibration signals are filtered to distinguish between external damage and broadband environmental noise. By calculating the energy distribution of each frequency band, a frequency band energy ratio index can be constructed to distinguish between normal vibration and abnormal impacts, achieving automatic fault type classification. Traditional wavelet transform is sensitive to small signal shifts, which may lead to drastic changes in decomposition coefficients. However, the dual-tree complex wavelet, through the redundant design of two sets of filters, keeps the transform results stable after signal shifts. This characteristic is particularly important in cable vibration monitoring, as it can avoid misjudgments caused by small sensor displacements or installation errors. The complex domain decomposition of the dual-tree complex wavelet can simultaneously extract the real and imaginary parts of the signal, enhancing the noise suppression capability. For example, in a strong electromagnetic interference environment, the complex wavelet coefficients can further distinguish between signal and noise through phase information, improving the signal-to-noise ratio of feature extraction.
[0045] S4. Energy enhancement is performed on the filtered detail coefficient to obtain the enhanced detail coefficient.
[0046] Specifically, energy enhancement can amplify weak fault-related features in detail coefficients while suppressing random fluctuations in background noise. For example, in low signal-to-noise ratio environments, energy enhancement can make the amplitude characteristics of impulse signals more prominent, preventing them from being masked by noise. The tightly supported wavelet basis of dual-tree complex wavelets can preserve the transient characteristics of the signal, while energy enhancement further enhances the performance of these details in the frequency domain, providing clearer input for subsequent time-frequency analysis.
[0047] S5. Based on the enhanced detail coefficient and the number of decomposition layers, calculate the time-frequency entropy and instantaneous frequency, extract the time-frequency joint features, and collect data from several deployed vibration sensors to calculate the spatial correlation coefficient and extract spatial features.
[0048] Specifically, by calculating the time-frequency distribution entropy of the enhanced detail coefficients, the complexity of the signal in the time-frequency domain can be quantified. For example, the time-frequency entropy of a cable rupture impact signal is usually higher than that of a normal vibration signal because it contains more high-frequency transient components. Extracting the instantaneous frequency from the detail coefficients using Hilbert transform or synchronous squeezing wavelet transform allows for precise capture of the signal frequency variation over time. For example, the instantaneous frequency of an impact signal exhibits brief high-frequency jumps, while the frequency of a normal vibration signal is relatively stable. Multi-scale decomposition of dual-tree complex wavelets, combining time-frequency entropy and instantaneous frequency, can simultaneously analyze the local and global characteristics of a signal, suitable for non-stationary vibration signals, such as the dynamic response of cables subjected to wind loads or mechanical impacts. By collecting data from multiple vibration sensors and calculating the spatial correlation coefficient between signals, the vibration synchronicity of sensors at different locations can be quantified. For example, a cable rupture impact can cause a high correlation between signals from adjacent sensors, while a local mechanical fault may only affect a specific sensor. Spatial feature fusion can suppress misjudgments caused by environmental noise or installation errors in a single sensor. For example, if a sensor's data is affected by electromagnetic interference, its spatial correlation coefficient will deviate significantly from other sensors; threshold filtering can exclude abnormal data.
[0049] S6. Based on the full subband of the vibration signal after dual-tree complex wavelet decomposition, extract energy features, entropy features, and statistical features. Then, concatenate the energy features, entropy features, and statistical features with the extracted time-frequency joint features and spatial features to form a detail-enhanced feature vector.
[0050] Specifically, based on the full sub-bands after dual-tree complex wavelet decomposition of the vibration signal, energy features, entropy features, and statistical features are extracted and concatenated with the corresponding time-frequency joint features and spatial features of the sub-bands to form multi-dimensional features. The multi-dimensional features of all sub-bands are then concatenated to form a detail-enhanced feature vector. The energy distribution of each sub-band is calculated to reflect the degree of frequency domain energy concentration of the signal. The complexity of the signal is quantified using time-frequency entropy or wavelet entropy to distinguish between normal vibrations and abnormal impacts. Statistical quantities such as the mean, variance, and skewness of the sub-band coefficients are extracted to capture the distribution characteristics of the signal. The time-frequency joint features, spatial features, and full-sub-band features are concatenated to form a multi-dimensional feature vector. Detail energy enhancement further highlights key features, making the feature vector more sensitive to faults.
[0051] S7. Input the detail-enhanced feature vector into the trained neural network model based on LSTM and fully connected layers, and output the denoised and optimized feature vector.
[0052] Specifically, adaptive denoising and feature optimization are performed using a neural network model based on LSTM and fully connected layers. The LSTM network, through memory units and gating mechanisms, can capture the long-term temporal dependencies of vibration signals. For example, the periodic patterns of cable vibration signals or the temporal evolution of transient impacts can be effectively modeled using LSTM. The hidden layers of the LSTM can learn the difference between noise and signal, automatically adjusting weights through backpropagation to achieve end-to-end denoising. The fully connected layers map the high-order features extracted by the LSTM to the output space, enabling fault classification or regression prediction.
[0053] In one embodiment, the hierarchical adaptive threshold analysis model calculates the adaptive threshold for each layer at the corresponding time using the following formula: T j (t)=
[0054] Among them, T j (t) is the adaptive threshold of the j-th layer at time t; Let be the local noise standard deviation of the j-th layer at time t. It is estimated by the neighborhood window and reflects the fluctuation of the local noise in this layer. The noise standard deviation may be different at different times and at different layers, which reflects the time-varying characteristics of the noise. Let be the real-time signal-to-noise ratio of the j-th layer at time t; N is the total number of discrete sampling points of the vibration signal, which is the length of the signal, i.e., the number of data points. It is determined by directly counting the number of sampling points contained in the input vibration signal. For example, if the vibration signal is transmitted through a sensor at a sampling frequency f... ns T was collected s If the result is obtained in seconds, then N=f ns ×T s If the signal is a pre-acquired digital sequence, then N is the number of elements in the sequence. β1 and β2 are adjustment parameters used to adjust the sensitivity of the threshold to the real-time signal-to-noise ratio. They can be selected and adjusted according to the actual situation, for example, they can be set to 0.4 and 0.2 respectively. is the scaling factor; J is the number of decomposition layers.
[0055] In one embodiment, based on the impact pulse index and frequency band energy ratio of the vibration signal from the transient impact of cable external breakage in historical data after dual-tree complex wavelet decomposition, the impact pulse index range and frequency band energy ratio range are determined, specifically including: The vibration signal of transient impact from external failure in historical data is subjected to dual-tree complex wavelet decomposition. The impact pulse index of the real part and the impact pulse index of the imaginary part of the vibration signal are calculated separately. The fluctuation range of the impact pulse index of the real part and the imaginary part are respectively used as the impact pulse index interval of the real part and the impact pulse index interval of the imaginary part. The frequency band energy ratio of the real part and the frequency band energy ratio of the imaginary part of the vibration signal are calculated separately. The fluctuation range of the frequency band energy ratio of the real part and the imaginary part are respectively used as the frequency band energy ratio interval of the real part and the frequency band energy ratio interval of the imaginary part. The vibration signal after threshold processing is filtered using the impact pulse index interval of the real part, the impact pulse index interval of the imaginary part, the frequency band energy ratio interval of the real part, and the frequency band energy ratio interval of the imaginary part.
[0056] The formula for calculating the band energy ratio is as follows: Band energy ratio of the real part of the vibration signal = total energy of the real part in a specific frequency band / total energy of the real part at all frequencies; Band energy ratio of the imaginary part of the vibration signal = total energy of the imaginary part in a specific frequency band / total energy of the imaginary part at all frequencies.
[0057] In one embodiment, the impact pulse index is calculated using the following formula: Real impact pulse index:
[0058] in, The SPI impact pulse index represents the real part of the vibration signal. Let be the real detail coefficients of the j-th layer at time t; Let be the mean of the real part detail coefficients of the j-th layer at time t. Let be the standard deviation of the real part detail coefficients of the j-th layer at time t. As a regulating factor; Imaginary impulse pulse index:
[0059] in, The SPI impact pulse index represents the imaginary part of the vibration signal. Let be the imaginary detail coefficients of the j-th layer at time t; Let be the mean of the imaginary detail coefficients of the j-th layer at time t. Let be the standard deviation of the imaginary detail coefficients of the j-th layer at time t.
[0060] In one embodiment, the detail factor after filtering is enhanced with energy using the following formula: D j (t)′= D j (t)*(1+α2*SNR j ) Among them, D j(t)′ represents the detail coefficient of the j-th layer after energy enhancement at time t, D j (t) represents the detail coefficients of the j-th layer after wavelet decomposition at time t; α2 is the enhancement coefficient, which can be set to 0.1 for example; SNR j Let be the signal-to-noise ratio of the coefficients in the j-th layer.
[0061] In one embodiment, the signal-to-noise ratio (SNR) of the j-th layer coefficients j Represented as:
[0062] in, The total energy of the original detail coefficients at layer j; denoted as the total energy of the noise correlation coefficients filtered out after threshold processing by the hierarchical adaptive threshold analysis model for the coefficients of the j-th layer.
[0063] In one embodiment, based on the enhanced detail coefficients and the number of decomposition layers, the time-frequency entropy and instantaneous frequency are calculated, and joint time-frequency features are extracted, specifically including: Calculate the time-frequency entropy using the following formula:
[0064]
[0065] in, Let be the time-frequency entropy at layer j and time t, which describes the degree of disorder in the time-frequency energy distribution at that time and scale. The larger the time-frequency entropy, the less concentrated the energy distribution at that time and scale, and the more complex the time-varying characteristics of the signal; conversely, the smaller the time-frequency entropy, the more concentrated the energy distribution, and the more stable the signal characteristics. The time-frequency energy distribution reflects the combined energy in the horizontal and vertical directions at time t in the j-th layer, with a frequency index of k, representing the proportion of the total energy at that time and scale. This energy distribution can be used to calculate time-frequency entropy and other parameters to analyze the energy distribution characteristics of the signal in the time-frequency domain and capture the joint time-frequency features such as impulses in the signal. k is the frequency index, and in the discrete frequency space, different integer values of k correspond to different frequency points. For the real part detail coefficients at layer j, time t, and frequency index k, the square of the modulus represents the energy value corresponding to the coefficient; For the imaginary detail coefficients at layer j, time t, and frequency index k, the square of the modulus represents the energy value corresponding to the coefficient. Instantaneous frequencies are extracted from the enhanced detail coefficients using Hilbert transform or synchronous squeeze wavelet transform.
[0066] In one embodiment, the spatial correlation coefficient is calculated from data collected from several deployed vibration sensors using the following formula:
[0067] in, Let be the spatial correlation coefficient between sensor a and sensor b at time t. When multiple vibration sensors are deployed, the spatial correlation coefficient is used to calculate the location of the external failure. The area with high correlation may be the source of the external failure. Let be the real part detail coefficients of sensor a at time t; Let be the real part detail coefficients of sensor b at time t; Let be the imaginary detail coefficients of sensor a at time t; Let be the imaginary detail coefficients of sensor b at time t; Let be the variance of the real part detail coefficients of sensor a at time t; Let be the variance of the detail coefficients of the imaginary part of sensor a at time t; Let be the variance of the real part detail coefficients of sensor b at time t; Let be the variance of the imaginary detail coefficients of sensor b at time t.
[0068] The spatial correlation coefficient between the sensor and the vibration signals collected by the sensor within a threshold range is calculated using the spatial correlation calculation formula. The obtained spatial correlation coefficients are arranged in ascending order of distance from the corresponding sensor to form a spatial correlation coefficient array, which serves as the spatial feature of the vibration signals collected by the corresponding sensor.
[0069] In one embodiment, based on the full subband of the vibration signal after dual-tree complex wavelet decomposition, energy features, entropy features, and statistical features are extracted, specifically including: Based on the vibration signal after dual-tree complex wavelet decomposition, the real and imaginary parts of each sub-band signal are extracted. The sub-band energy of each sub-band's real and imaginary parts is extracted as energy features, the sub-band entropy of each sub-band's real and imaginary parts is extracted as entropy features, and the sub-band kurtosis of each sub-band's real and imaginary parts is extracted as statistical features. The extracted energy features, entropy features, and statistical features are concatenated with the corresponding time-frequency joint features and spatial features of the sub-band to form multi-dimensional features.
[0070] In one embodiment, the training method for a neural network model based on LSTM and fully connected layers includes: This study analyzes historical data on cable vibration signals, including noise and transient impact signals. Energy, entropy, statistical, time-frequency combined, and spatial features are extracted from the noise signal and concatenated into a target noise feature vector. Similarly, energy, entropy, statistical, time-frequency combined, and spatial features are extracted from the transient impact vibration signal and concatenated into a transient impact feature vector. The target noise feature vector and the transient impact feature vector are used to train a neural network model based on LSTM and fully connected layers. This trained network model can identify noise and transient impact signals within the enhanced detail feature vector, filter the noise signal, increase the weight of the transient impact signal, and output an optimized feature vector.
[0071] For example, when increasing the weight of an external transient impact signal, the weight increase can be set to 15% of the original weight.
[0072] Example 2 like Figure 2 As shown, based on the same inventive concept as Embodiment 1, Embodiment 2 provides a vibration signal denoising and feature extraction system based on wavelet analysis, comprising: The signal acquisition and decomposition module is used to acquire the vibration signal of the cable, perform dual-tree complex wavelet decomposition on the vibration signal, and obtain detail coefficients of multiple decomposition levels, including real part detail coefficients and imaginary part detail coefficients. The hierarchical adaptive threshold processing module is used to establish a hierarchical adaptive threshold analysis model based on the statistical characteristics of environmental noise. The hierarchical adaptive threshold analysis model is used to perform threshold processing on the detail coefficients of each layer. The detail coefficients whose absolute values are less than the adaptive threshold of the corresponding layer are judged as noise correlation coefficients and set to zero. The transient impact filtering module is used to determine the impact pulse index range and the frequency band energy ratio range based on the impact pulse index and frequency band energy ratio of the vibration signal of the transient impact of cable external damage in historical data after dual-tree complex wavelet decomposition. The impact pulse index range and the frequency band energy ratio range are used to filter the vibration signal after threshold processing, and retain the signal components that meet the range requirements. The energy enhancement module is used to enhance the detail factor after filtering to obtain an enhanced detail factor. The time-frequency and spatial feature extraction module is used to calculate the time-frequency entropy and instantaneous frequency based on the enhanced detail coefficient and the number of decomposition layers, extract the time-frequency joint features, and collect data from several deployed vibration sensors to calculate the spatial correlation coefficient and extract spatial features. The multi-dimensional feature stitching module is used to extract energy features, entropy features, and statistical features from the full subband of the vibration signal after dual-tree complex wavelet decomposition. The energy features, entropy features, and statistical features are stitched together with the extracted time-frequency joint features and spatial features to form a detail-enhanced feature vector. The deep denoising optimization module is used to input the detail-enhanced feature vector into the trained neural network model based on LSTM and fully connected layers, and output the denoised and optimized feature vector.
[0073] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A vibration signal denoising and feature extraction method based on wavelet analysis, characterized in that, Includes the following steps: The vibration signal of the cable is acquired, and the vibration signal is decomposed into a dual-tree complex wavelet to obtain detail coefficients of multiple decomposition levels, including real part detail coefficients and imaginary part detail coefficients. A hierarchical adaptive threshold analysis model is established based on the statistical characteristics of environmental noise. The model is used to perform threshold processing on the detail coefficients of each layer. The detail coefficients whose absolute values are less than the adaptive threshold of the corresponding layer are determined as noise correlation coefficients and set to zero. Based on the impact pulse index and frequency band energy ratio of the vibration signal of the transient impact of cable external failure in historical data after dual-tree complex wavelet decomposition, the impact pulse index range and frequency band energy ratio range are determined. The vibration signal after threshold processing is filtered using the impact pulse index range and frequency band energy ratio range to retain the signal components that meet the range requirements. The detail factor after filtration is enhanced by energy to obtain the enhanced detail factor. Based on the enhanced detail coefficients and decomposition levels, the time-frequency entropy and instantaneous frequency are calculated, the time-frequency joint features are extracted, and the spatial correlation coefficient is calculated by collecting data from several deployed vibration sensors to extract spatial features. Based on the full subband of the vibration signal after dual-tree complex wavelet decomposition, energy features, entropy features, and statistical features are extracted. The energy features, entropy features, and statistical features are then concatenated with the extracted time-frequency joint features and spatial features to form a detail-enhanced feature vector. The detail-enhanced feature vectors are input into the trained neural network model based on LSTM and fully connected layers, and the output is a denoised and optimized feature vector.
2. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, The hierarchical adaptive threshold analysis model calculates the adaptive threshold for each layer at the corresponding time using the following formula: T j (t)= Among them, T j (t) represents the adaptive threshold of the j-th layer at time t. Let be the standard deviation of the local noise of the j-th layer at time t. Let be the real-time signal-to-noise ratio of the j-th layer at time t, and N be the total number of discrete sampling points of the vibration signal. and To adjust the parameters, is the scaling factor; J is the number of decomposition layers.
3. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, Based on the impact pulse index and frequency band energy ratio obtained from the vibration signal of transient impact of external cable failure in historical data after dual-tree complex wavelet decomposition, the impact pulse index range and frequency band energy ratio range are determined, specifically including: The vibration signal of transient impact from external failure in historical data is decomposed by dual-tree complex wavelet decomposition. The impact pulse index of the real part and the impact pulse index of the imaginary part of the vibration signal are calculated respectively. The fluctuation range of the impact pulse index of the real part and the imaginary part are respectively used as the impact pulse index range of the real part and the impact pulse index range of the imaginary part. Calculate the frequency band energy ratio of the real part and the frequency band energy ratio of the imaginary part of the vibration signal respectively, and take the fluctuation range of the frequency band energy ratio of the real part and the frequency band energy ratio of the imaginary part as the real part frequency band energy ratio interval and the imaginary part frequency band energy ratio interval respectively. The vibration signal after threshold processing is filtered by using the real part impact pulse index range, the imaginary part impact pulse index range, the real part frequency band energy ratio range, and the imaginary part frequency band energy ratio range.
4. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 3, characterized in that, The impact pulse index is calculated using the following formula: Real impact pulse index: in, The SPI impact pulse index represents the real part of the vibration signal. Let be the real detail coefficients of the j-th layer at time t; Let be the mean of the real part detail coefficients of the j-th layer at time t. Let be the standard deviation of the real part detail coefficients of the j-th layer at time t. As a regulating factor; Imaginary impulse pulse index: in, The SPI impact pulse index represents the imaginary part of the vibration signal. Let be the imaginary detail coefficients of the j-th layer at time t; Let be the mean of the imaginary detail coefficients of the j-th layer at time t. Let be the standard deviation of the imaginary detail coefficients of the j-th layer at time t.
5. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, Energy enhancement of the detail factor after filtering is achieved using the following formula: D j (t)′= D j (t)*(1+α2*SNR j ) Among them, D j (t)′ represents the detail coefficient of the j-th layer after energy enhancement at time t, D j (t) represents the detail coefficients of the j-th layer after wavelet decomposition at time t, and α2 is the enhancement coefficient; SNR j Let be the signal-to-noise ratio of the coefficients in the j-th layer.
6. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 5, characterized in that, The signal-to-noise ratio (SNR) of the j-th layer coefficients j Represented as: in, The total energy of the original detail coefficients at layer j; denoted as the total energy of the noise correlation coefficients filtered out after threshold processing by the hierarchical adaptive threshold analysis model for the coefficients of the j-th layer.
7. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, Based on the enhanced detail coefficients and the number of decomposition layers, the time-frequency entropy and instantaneous frequency are calculated, and joint time-frequency features are extracted, specifically including: Calculate the time-frequency entropy using the following formula: in, Let the time-frequency entropy be at the j-th layer and time t. This represents the time-frequency energy distribution, where k is the frequency index. Let the real part detail coefficients be the coefficients at the j-th level, time t, and frequency index k. These are the imaginary detail coefficients at level j, time t, and frequency index k. Instantaneous frequencies are extracted from the enhanced detail coefficients using Hilbert transform or synchronous squeeze wavelet transform.
8. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, The spatial correlation coefficient is calculated from data collected from several deployed vibration sensors using the following formula: in, Let be the spatial correlation coefficient between sensor a and sensor b at time t; Let be the real part detail coefficients of sensor a at time t; Let be the real part detail coefficients of sensor b at time t; Let be the imaginary detail coefficients of sensor a at time t; Let be the imaginary detail coefficients of sensor b at time t; Let be the variance of the real part detail coefficients of sensor a at time t; Let be the variance of the detail coefficients of the imaginary part of sensor a at time t; Let be the variance of the real part detail coefficients of sensor b at time t; Let be the variance of the imaginary detail coefficients of sensor b at time t.
9. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, Based on the full subband of the vibration signal after dual-tree complex wavelet decomposition, energy features, entropy features, and statistical features are extracted, specifically including: Based on each sub-band signal of the vibration signal after dual-tree complex wavelet decomposition, the real part and imaginary part are extracted respectively. The subband energy of the real and imaginary parts of each subband is extracted as energy features, the subband entropy of the real and imaginary parts of each subband is extracted as entropy features, and the subband kurtosis of the real and imaginary parts of each subband is extracted as statistical features.
10. The vibration signal denoising and feature extraction method based on wavelet analysis according to claim 1, characterized in that, Training methods for neural network models based on LSTM and fully connected layers include: The noise signal and the vibration signal of transient impact from external damage are collected from the original vibration signal of the cable in the statistical historical data. Energy features, entropy features, statistical features, time-frequency joint features, and spatial features are extracted from the noise signal and concatenated into a target noise feature vector; Energy features, entropy features, statistical features, time-frequency joint features, and spatial features are extracted from the vibration signal of external transient impact and concatenated into a feature vector of external transient impact. A neural network model based on LSTM and fully connected layers is trained using the target noise feature vector and the feature vector of external transient impact. The trained network model can identify noise signals and external transient impacts in the detail-enhanced feature vector, filter the noise signals, increase the weight of the external transient impact signals, and output the optimized feature vector.