Cable external damage prevention signal noise reduction method, system and terminal

Abstract

The invention relates to the technical field of electric power safety monitoring, and particularly discloses a cable external damage prevention signal noise reduction method and system and a terminal, and the method comprises the steps: carrying out the preprocessing of an original disturbance phase signal, so as to eliminate a part of noise and low-frequency drift in an original signal, and dynamically extracting the disturbance information in the signal, so as to achieve the noise reduction. Namely, the wavelet basis is dynamically selected for wavelet analysis, and the mismatch problem of the traditional fixed wavelet basis is relieved in combination with a decomposition layer number self-adaptive strategy. Meanwhile, an auxiliary scheme based on deep learning is provided, and the adaptive noise reduction capability in a complex scene is enhanced.

Description

Technical Field

[0001] This application relates to the field of power safety monitoring technology, and more specifically, to a method, system and terminal for reducing the noise of cable damage signals. Background Technology

[0002] Due to the difficulty in directly locating faults in buried cables, existing technologies employ distributed optical fiber vibration sensing technology. This involves laying sensing optical fibers along the buried cable route and detecting changes in the laser phase within the sensing optical fibers to determine whether an external damage event has occurred on the cable.

[0003] However, such signals containing perturbation information are susceptible to external noise. For example, fiber optic sensing systems are easily affected by ambient temperature drift, mechanical vibration noise, and inherent optical path noise, leading to a decrease in signal-to-noise ratio (SNR), and weak perturbation signals are easily submerged by noise.

[0004] To enable subsequent analysis of disturbance signals and identification of external damage events, it is necessary to denoise the disturbance signals in order to extract useful information.

[0005] Traditional noise reduction methods rely on using a fixed wavelet basis and a fixed threshold to process the perturbation phase signal. However, different types of external disturbance events result in different time-frequency characteristics of the perturbation phase signal, which may lead to the loss of effective perturbation information. Therefore, an optimized technical solution is needed. Summary of the Invention

[0006] The technical problem to be solved by this application is to provide a method, system and terminal for reducing the noise of cable damage prevention signals, which solves the problem in the prior art that the disturbance information extraction strategy cannot be adaptively changed for different types of external damage events.

[0007] The technical problem to be solved in this application is achieved by the following technical solution:

[0008] In a first aspect, this application provides a method for reducing signal noise in cables to prevent external damage, including:

[0009] Acquire the original disturbance phase signal obtained by the distributed fiber optic vibration sensing module based on φ-OTDR;

[0010] The original perturbation phase signal is preprocessed to obtain the preprocessed perturbation phase signal;

[0011] The perturbation information of the preprocessed perturbation phase signal is dynamically extracted and denoised to obtain the restored perturbation denoised signal.

[0012] Further, the original perturbation phase signal is preprocessed to obtain a preprocessed perturbation phase signal, including: preprocessing the original perturbation phase signal using moving average filtering and / or phase difference to obtain the preprocessed perturbation phase signal.

[0013] Further, the preprocessed perturbation phase signal is subjected to dynamic perturbation information extraction and noise reduction to obtain a restored perturbation-denoised signal, including:

[0014] By dynamically selecting the wavelet basis and the number of decomposition levels, wavelet decomposition is performed on the preprocessed perturbed phase signal to obtain multiple detail coefficients.

[0015] The initial denoised signal is obtained by performing hierarchical threshold combination processing and inverse wavelet transform reconstruction processing on the multiple detail coefficients.

[0016] The initial denoised signal is decomposed and reconstructed using dual-tree complex wavelet transform, and a phase compensation factor is embedded to improve positioning accuracy, thus obtaining the restored perturbation denoised signal.

[0017] Furthermore, by dynamically selecting the wavelet basis and the number of decomposition levels, wavelet decomposition is performed on the preprocessed perturbed phase signal to obtain multiple detail coefficients, including:

[0018] The signal category of the preprocessed perturbation phase signal is determined based on the signal difference between the preprocessed perturbation phase signal and the perturbation phase signals of each group of samples in the sample library;

[0019] Based on the signal category, a wavelet basis is selected, and the preprocessed perturbed phase signal is decomposed by wavelet to obtain the multiple detail coefficients.

[0020] In the wavelet decomposition process, an adaptive strategy for the number of decomposition levels is used to control whether the wavelet decomposition is terminated.

[0021] Furthermore, the signal category includes at least one of the following: transient impact, continuous vibration, and mixed noise;

[0022] Wherein, when the signal category of the preprocessed perturbation phase signal is instantaneous impulse, the db4 wavelet and coif3 wavelet are selected as wavelet bases;

[0023] When the preprocessed perturbation phase signal belongs to the signal category of continuous vibration, the sym4 wavelet is selected as the wavelet basis.

[0024] When the preprocessed perturbation phase signal belongs to the mixed noise category, the bior3.3 wavelet is selected as the wavelet basis.

[0025] Furthermore, an adaptive strategy for the number of decomposition levels is employed during the wavelet decomposition process to control whether the wavelet decomposition terminates, including:

[0026] If the current decomposition level is less than 2, continue decomposing;

[0027] If the current decomposition level is greater than or equal to 3, then perform the following steps:

[0028] The termination threshold is determined based on the signal category to which the preprocessed perturbation phase signal belongs.

[0029] Calculate the energy change rate of the current decomposition layer;

[0030] The decomposition terminates when the energy change rate is less than the termination threshold; otherwise, the decomposition continues.

[0031] Further, the multiple detail coefficients are subjected to hierarchical threshold combination processing and wavelet inverse transform reconstruction processing to obtain the initial denoised signal, including:

[0032] Hard thresholding is applied to detail coefficients at a decomposition level of 1.

[0033] Semi-soft thresholding is applied to detail coefficients at decomposition levels 2 or 3.

[0034] Soft thresholding is applied to detail coefficients at a decomposition level of 4 or higher.

[0035] Further, the preprocessed perturbation phase signal is subjected to dynamic perturbation information extraction and noise reduction to obtain a restored perturbation-denoised signal, including:

[0036] The preprocessed perturbation phase signal is processed using a multi-granularity feature residual stripper to obtain a first-granularity perturbation feature map, a second-granularity perturbation feature map, and a third-granularity perturbation feature map.

[0037] The first granularity perturbation feature map and the second granularity perturbation feature map are input into a convolutional neural network using a spatial attention mechanism to obtain a first perturbation enhancement feature map and a second perturbation enhancement feature map.

[0038] The first perturbation enhancement feature map, the second perturbation enhancement feature map, and the third granularity perturbation feature map are processed using a bidirectional dynamic fusion processor to obtain a multi-granularity perturbation feature map.

[0039] The multi-granularity perturbation feature map is processed using a decoder-based denoiser to obtain the restored perturbation denoised signal.

[0040] Secondly, this application also provides a cable damage prevention signal noise reduction system, comprising:

[0041] The disturbance signal acquisition module is used to acquire the original disturbance phase signal obtained by the distributed optical fiber vibration sensing module based on φ-OTDR;

[0042] The preprocessing module is used to preprocess the original perturbation phase signal to obtain the preprocessed perturbation phase signal;

[0043] The noise reduction module is used to dynamically extract and reduce the noise of the preprocessed perturbation phase signal to obtain the restored perturbation-denoised signal.

[0044] Thirdly, this application also provides a terminal, including: a processor for coupling with a memory, and for reading and executing instructions stored in the memory; when the processor is running, the processor executes the instructions, causing the processor to perform a cable damage prevention signal noise reduction method.

[0045] This application includes at least one of the following beneficial technical effects:

[0046] 1. By preprocessing the original perturbation phase signal to eliminate some noise and low-frequency drift in the original signal, a more accurate data source is provided for the extraction of perturbation features.

[0047] 2. During wavelet analysis, the wavelet basis and decomposition level are dynamically selected to alleviate the mismatch problem of traditional fixed wavelet basis.

[0048] 3. Provide deep learning-based auxiliary solutions to enhance adaptive noise reduction capabilities in complex scenarios. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the cable damage prevention signal noise reduction method provided in this application.

[0050] Figure 2 A schematic diagram of the structure of the distributed optical fiber vibration sensing module based on φ-OTDR provided in this application.

[0051] Figure 3A The flowchart for obtaining the restored perturbation-denoised signal provided in this application Figure 1 .

[0052] Figure 3B The flowchart for obtaining the restored perturbation-denoised signal provided in this application Figure 2 .

[0053] Figure 4 This is a structural diagram of the cable damage prevention signal noise reduction system provided in this application.

[0054] Figure 5 A schematic diagram of the terminal provided in this application.

[0055] In the picture:

[0056] Disturbance signal acquisition module 110; preprocessing module 120; noise reduction module 130;

[0057] Processor 10; Memory 20; Instruction Set 30. Detailed Implementation

[0058] To facilitate a clear understanding of the technical means, creative features, objectives, and effects of this application, the following description, in conjunction with specific illustrations, further elaborates on this application.

[0059] Underground cables, with their advantages of convenient construction and space efficiency, have become the mainstream form of cable laying in modern cities. However, as the scale of underground cable networks continues to expand, the risk of external damage is also increasing.

[0060] Cable failures caused by external factors such as mechanical construction misoperation and geological subsidence can not only cause direct economic losses, but also lead to large-scale power outages, affecting industrial and commercial operations and people's livelihoods.

[0061] Existing technologies employ distributed fiber optic vibration sensing technology, which detects changes in the laser phase within the sensing fiber to determine whether an external damage event has occurred on the cable.

[0062] However, such signals containing perturbation information are susceptible to external noise, and weak perturbation signals are easily drowned out by noise.

[0063] To enable subsequent analysis of disturbance signals and identification of external damage events, noise reduction of disturbance signals is necessary to extract useful information.

[0064] Traditional noise reduction methods rely on using a fixed wavelet basis and a fixed threshold to process the perturbation phase signal. However, different types of external disturbances result in different time-frequency characteristics of the perturbation phase signal, which may lead to the loss of effective perturbation information during the noise reduction process.

[0065] like Figure 1 As shown, this application provides a method for reducing signal noise in cables to prevent external damage, thereby solving the above-mentioned technical problems. The specific steps include:

[0066] S1. Acquire the original disturbance phase signal obtained by the distributed optical fiber vibration sensing module based on φ-OTDR;

[0067] S2. Preprocess the original perturbation phase signal to obtain the preprocessed perturbation phase signal;

[0068] S3. Dynamically extract and denoise the perturbation information of the preprocessed perturbation phase signal to obtain the restored perturbation denoised signal.

[0069] like Figure 2As shown, the distributed fiber optic vibration sensing module based on φ-OTDR includes a narrow linewidth laser source, a first coupler, an optoelectronic modulator, an acousto-optic modulator, a signal amplifier, a filter, a circulator, a sensing fiber, a second coupler, a balanced detector, a photodetector, and an analog-to-digital converter.

[0070] Among them, the distributed optical fiber vibration sensing module based on φ-OTDR (Phase-sensitive Optical Time Domain Reflectometry) detects external vibration events by acquiring the modulation signal of backscattered light in the optical fiber.

[0071] Due to the elastic-optical effect and strain effect, external vibrations can cause minute changes in the refractive index or length of optical fibers, thereby affecting the phase information of the backscattered light signal in the fiber. In the technical solution of this application, obtaining the original disturbance phase signal acquired by a distributed optical fiber vibration sensing module based on φ-OTDR can provide a data source for the dynamic extraction of disturbance information.

[0072] Specifically, a narrow-linewidth laser source is used to generate a continuous optical signal; a first coupler is connected to the narrow-linewidth laser source to split the continuous optical signal into local light and signal light; an opto-modulator is connected to the first coupler to process the signal light via an electro-optic modulator for opto-modulation; an acousto-optic modulator is connected to the opto-modulator to send the signal to the acousto-optic modulator for chopping and frequency shifting; a signal amplifier is connected to the acousto-optic modulator to enhance the optical power of the signal; a filter is connected to the signal amplifier to remove spontaneous emission noise generated during amplification; a circulator is connected to the filter and the sensing fiber respectively to inject the noise-reduced signal into the sensing fiber laid along the cable; a second coupler is connected to the first coupler and the circulator respectively to beat the scattered light with the local light; a balanced detector is connected to the second coupler to reduce the noise of the beat signal; a photodetector is connected to the balanced detector to convert the noise-reduced signal into an electrical signal; and an analog-to-digital converter is connected to the photodetector to convert the analog electrical signal into a digital electrical signal (the original perturbation phase signal).

[0073] Next, the original perturbation phase signal is preprocessed to eliminate some noise and low-frequency drift in the original signal, providing high-quality data for subsequent processing.

[0074] Specifically, the original perturbation phase signal is preprocessed using moving average filtering and / or phase difference to obtain the preprocessed perturbation phase signal.

[0075] Here, the moving average filter aims to smooth the signal and reduce the impact of high-frequency random noise (such as electronic noise and environmental interference) at the perimeter of the cable line.

[0076] Among them, the moving average filter smooths the signal by calculating the mean of the data within the window, and the formula is:

[0077]

[0078] Where L is the window length, i.e., the number of data items within the window. s[ni] is the (n+1)th data point in the signal window after the moving average filtering, and s[ni] is the (n-i+1)th data point in the original input signal window.

[0079] In practical applications, the window length can be dynamically adjusted based on the signal characteristics. Typically, the choice of window length depends on the signal's sampling rate and highest frequency.

[0080] Preferably, in the technical solution of this application, the window length is an odd number, and this odd number is greater than the quotient of the signal sampling rate and twice the highest frequency. Here, an odd number is preferred to avoid phase shift, and zero-padding is applied to the signal to handle boundary effects.

[0081] Here, phase difference aims to remove low-frequency trend terms caused by factors such as temperature drift and slow changes in the optical path, thereby highlighting high-frequency vibration signals.

[0082] In the technical solution of this application, the input signal is phase-differentially processed using the following differential formula to obtain the output signal; wherein, the differential formula is:

[0083]

[0084] Where t takes values ​​from 1 to the length of the input signal, Let be the value of the output signal at time t. Let be the value of the input signal at time t. The value of the input signal at time t-1.

[0085] in, The value of the output signal at the initial moment. This is the value of the input signal at the initial moment. In other words, the initial value is padded with zeros or the original value is retained.

[0086] It is worth mentioning that, since there is not enough historical data available for calculation to fill the window at the beginning of the signal, if no processing is performed, the length of the output signal will be shorter than that of the input signal, resulting in incomplete boundary data.

[0087] Common padding methods include zero padding and window truncation. Zero padding sets missing data points to 0 to maintain the same length as the input signal. Window truncation can use the initial value of the input signal as the initial value of the output signal, preserving the original value. Its advantage is that it avoids introducing spurious data and retains the original signal characteristics, making it suitable for scenarios with high boundary accuracy requirements.

[0088] The core purpose of boundary initialization is to ensure that the output length of the filter is consistent with the input length, and that the transition at the boundary is smooth. In practice, the specific padding method should be chosen based on the signal characteristics and application scenario.

[0089] This differential processing is equivalent to high-pass filtering, which can suppress low-frequency components (frequency f). <fs / 2)。

[0090] Then, in a specific embodiment of this application, such as Figure 3A As shown, the following steps are performed to dynamically extract perturbation information and reduce noise in the preprocessed perturbed phase signal:

[0091] S301. Dynamically select the wavelet basis and the number of decomposition levels, and perform wavelet decomposition on the preprocessed perturbed phase signal to obtain multiple detail coefficients.

[0092] S302. Perform hierarchical threshold combination processing and wavelet inverse transform reconstruction processing on the multiple detail coefficients to obtain the initial noise reduction signal;

[0093] S303. The initial noise-reduced signal is decomposed and reconstructed using dual-tree complex wavelet transform, and a phase compensation factor is embedded to improve the positioning accuracy, thereby obtaining the restored perturbation-reduced signal.

[0094] It is understandable that in cable damage monitoring, vibration signals from different damage events (such as instantaneous impact, continuous vibration, and mixed noise) exhibit significant differences in time-frequency characteristics. A dynamic wavelet decomposition method is employed to optimize the decomposition effect for different signal types through hybrid wavelet basis combination and a hierarchical strategy.

[0095] The specific implementation of step S301 includes: First, determining the signal category of the preprocessed perturbation phase signal (at least one of instantaneous impact, continuous vibration and mixed noise) based on the signal difference between the preprocessed perturbation phase signal and the perturbation phase signals of each group of samples in the sample library.

[0096] Next, a wavelet basis is selected based on the signal category, and wavelet decomposition is performed on the preprocessed perturbed phase signal to obtain multiple detail coefficients.

[0097] In the embodiments of this application, the signal difference is calculated using correlation analysis, statistical methods, or signal distance measurement methods.

[0098] More specifically, correlation analysis measures the relationship between two signals by calculating the correlation coefficient between them, such as the Pearson correlation coefficient or the Spearman correlation coefficient.

[0099] Statistical methods, such as mean, variance, and standard deviation, are used to quantify the statistical characteristics of a signal in order to compare the statistical differences between two signals.

[0100] Signal distance is a measure of the difference between signal waveforms. For example, it can be calculated by averaging the absolute values ​​of the differences between two signals at corresponding moments, or by using more complex distance metrics such as Euclidean distance or Manhattan distance.

[0101] The perturbation phase signals of each group of samples in the sample library were generated by simulation software or obtained through actual acquisition. The perturbation phase signals of the samples in the sample library correspond to at least one of instantaneous impact, continuous vibration, and mixed noise.

[0102] Among them, when the signal category of the preprocessed perturbation phase signal is instantaneous impulse, the db4 wavelet and coif3 wavelet are selected as wavelet bases;

[0103] When the preprocessed perturbation phase signal belongs to the signal category of continuous vibration, the sym4 wavelet is selected as the wavelet basis.

[0104] When the preprocessed perturbation phase signal belongs to the mixed noise category, the bior3.3 wavelet is selected as the wavelet basis.

[0105] Specifically, transient shocks (such as excavation) have time-frequency characteristics of high frequency, short-duration abrupt changes, and concentrated energy. The db4 wavelet basis has the core characteristics of compact support (short filter length), fourth-order vanishing moment, and asymmetry, while the coif3 has the core characteristics of approximate symmetry, longer support (18 points), and sixth-order vanishing moment.

[0106] Thus, in the technical solution of this application, the instantaneous impact signal is first decomposed into shallow layers (db4 decomposed into 3 layers), the high-frequency transient components (D1-D3) are quickly extracted, the rising edge and peak value of the impact signal are captured, and the low-frequency approximation (A3) is retained for subsequent deep refinement.

[0107] For example, set 3 decomposition layers (covering the 0-500 Hz frequency band, assuming a sampling rate of 1 kHz). The output results are high-frequency detail coefficients D1 (500-250 Hz), D2 (250-125 Hz), and D3 (125-62.5 Hz), and low-frequency approximation coefficients: A3 (0-62.5 Hz).

[0108] Then, deep refinement is performed (coif3 decomposition of A3 to 3 layers), A3 is further decomposed to extract potentially hidden weak impact features (such as low-frequency vibration aftershocks), generating deep high-frequency details (D4-D6) and the final low-frequency approximation (A6).

[0109] For example, setting 3 decomposition layers (covering sub-bands in the 0-62.5 Hz frequency range) yields deep high-frequency details D4 (62.5-31.25 Hz), D5 (31.25-15.625 Hz), and D6 (15.625-7.8125 Hz), with the final low-frequency approximation being A6 (0-7.8125 Hz).

[0110] Finally, coefficient merging is performed, adopting a merging strategy that combines shallow high-frequency details (D1-D3) with deep high-frequency details (D4-D6) such as weighted averaging to retain all high-frequency information. The low-frequency approximation (A6) is directly used for reconstruction.

[0111] Specifically, continuous vibrations (such as those caused by vehicles running over objects) have low-frequency, periodic, and smooth, continuous time-frequency characteristics. The sym4 wavelet basis has the core characteristics of approximate symmetry, high vanishing moment (4th order), and moderate support length (8 points).

[0112] Thus, in the technical solution of this application, the continuous vibration signal is decomposed into four layers based on the sym4 wavelet basis (covering the 0-62.5 Hz frequency band, with a sampling rate of 1 kHz).

[0113] For example, the frequency bands are divided into D1 (500-250 Hz), D2 (250-125 Hz), D3 (125-62.5 Hz), D4 (62.5-31.25 Hz) and A4 (0-31.25 Hz), retaining the low-frequency periodic components.

[0114] The low-pass LPD and high-pass HPD are used for decomposition filtering, and symmetrical extension (mode='sym') reduces edge effects. This reduces phase distortion during signal decomposition and reconstruction, effectively suppresses low-frequency interference, accurately extracts the main frequency component of continuous vibration, and balances time-frequency resolution.

[0115] Specifically, mixed noise (complex environment) has time-frequency characteristics of multi-band, non-stationary, and complex interference. Bior3.3 has the core characteristics of biorthogonality (independent design of decomposition and reconstruction filters, supporting linear phase), long support (10-point decomposition / 6-point reconstruction, high-frequency band subdivision capability), and third-order vanishing moment (balancing high-frequency noise suppression and weak signal preservation).

[0116] Thus, in the technical solution of this application, the mixed noise signal is decomposed into 5 layers based on the Bior 3.3 wavelet basis (covering the low frequency band of 0-15.625 Hz, with a sampling rate of 1 kHz), separating low-frequency geological vibration and high-frequency wind and rain noise.

[0117] For example, the frequency bands are divided into D1 (500-250 Hz), D2 (250-125 Hz), D3 (125-62.5 Hz), D4 (62.5-31.25 Hz), D5 (31.25-15.625 Hz), and A5 (0-15.625 Hz). Among them, A5 contains low-frequency components of geological activity.

[0118] Specifically, the lo_d (low-pass decomposition) and hi_d (high-pass decomposition) of the decomposition filter bior3.3 and the lo_r (low-pass reconstruction) and hi_r (high-pass reconstruction) of the reconstruction filter bior3.3 are established to adapt to complex noise environments, avoid reconstruction distortion, and balance high-frequency noise suppression and weak signal preservation.

[0119] Furthermore, in the technical solution of this application, an adaptive strategy for the number of decomposition levels is adopted to control whether the wavelet decomposition is terminated during the wavelet decomposition process.

[0120] More specifically, including:

[0121] If the current decomposition level is less than 2, continue decomposing;

[0122] If the current decomposition level is greater than or equal to 3, then perform the following steps:

[0123] The termination threshold is determined based on the signal category to which the preprocessed perturbation phase signal belongs.

[0124] Calculate the energy change rate of the current decomposition layer;

[0125] Decomposition terminates when the rate of energy change is less than the termination threshold; otherwise, decomposition continues.

[0126] The energy change rate is calculated using the following formula:

[0127]

[0128] in, For the k-th detail coefficient of the j-th layer, Let be the number of detail coefficients at the j-th layer, and |·| denotes the calculation of the absolute value. For the energy of the j-th layer, The energy of the (j-1)th layer, Let be the energy change rate of the j-th layer relative to the (j-1)-th layer.

[0129] Here, an adaptive strategy for the number of decomposition levels is established using the energy gradient method to automatically adapt to the signal characteristics of different cable breakage events. In each wavelet decomposition level, the energy of the detail coefficients (high-frequency components) is used to measure the activity level of the signal at that level. In the technical solution of this application, the energy of the detail coefficients (high-frequency components) is used to measure the activity level of the signal at that level. Representing the energy of the j-th layer, using This represents the rate of energy change between adjacent decomposition layers, used to determine whether to continue decomposition.

[0130] In a specific example, when ΔE j When the energy change is below the termination threshold (e.g., 5%), it is considered that the energy change has stabilized, and the decomposition stops. The adaptive decomposition process first performs an initial decomposition, with at least three levels of decomposition, to avoid premature termination.

[0131] At the same time, a constraint is imposed on the maximum number of layers:

[0132]

[0133] Where N is the total number of detail coefficients, The maximum number of floors. This represents a base-2 logarithmic function operation. This ensures that enough data points (at least 8 samples) are retained after decomposition.

[0134] In one embodiment of this application, the initial termination threshold is set to 3% to 10%, and is dynamically adjusted according to the signal type. For example, the threshold is lower (3%) for high-frequency impact signals (such as mechanical excavation), allowing for more layers of decomposition; the threshold is higher (8%) for low-frequency continuous vibrations (such as vehicle crushing), reducing excessive decomposition.

[0135] Then, in step S302, the high-frequency detail coefficients of different decomposition layers are processed by layered threshold combination, and the processed results are reconstructed by wavelet inverse transform to generate the initial denoised signal.

[0136] The specific process of using layered threshold combinations to process high-frequency detail coefficients of different decomposition layers includes: hard thresholding for detail coefficients at decomposition layer 1; semi-soft thresholding for detail coefficients at decomposition layers 2 or 3; and soft thresholding for detail coefficients at decomposition layers 4 or higher.

[0137] Here, wavelet decomposition separates the signal layer by layer from high to low frequency. The high-frequency detail coefficients of different decomposition layers correspond to noise and effective signals in different frequency bands. Layered threshold combination processing achieves aggressive denoising in the shallow layer (high frequency band) to eliminate random noise by dynamically adjusting the threshold rules and functions; balanced noise suppression and signal preservation in the middle layer (mid-high frequency band); and conservative processing in the deep layer (low frequency band) to avoid loss of effective signals.

[0138] For the high-frequency detail coefficients Dj after wavelet decomposition, the mathematical definition of hard thresholding is:

[0139]

[0140] Where Dj is the detail coefficient of the j-th level wavelet decomposition, and λ is a preset threshold, usually based on the noise standard deviation (e.g., λ=3σ). These are the coefficients after thresholding.

[0141] Among them, the hard threshold function has the characteristic of preserving abrupt changes; coefficients whose absolute values ​​exceed the threshold are completely preserved, making it suitable for detecting transient signals (such as mechanical shocks). It also has noise resistance and a significant effect on suppressing high-frequency noise.

[0142] Typical applications are in high-frequency noise-dominated layers (such as the first layer of wavelet decomposition). High-frequency noise has high energy, and a hard threshold can quickly filter out most of the noise, making it suitable for detecting the initial impact signal of cable breakage events in φ-OTDR systems. It can handle strong transient signals, preserve the amplitude integrity of pulse signals, and avoid the amplitude attenuation problem of soft thresholds.

[0143] The mathematical definition of soft thresholding is:

[0144]

[0145] Where Dj represents the detail coefficients of the j-th level wavelet decomposition, and λ is a preset threshold. These are the coefficients after thresholding.

[0146] Soft thresholding functions can effectively suppress noise components, especially in wavelet denoising signal processing. By weakening or zeroing low-intensity signal coefficients, high-frequency noise can be removed, making the main features of the signal more prominent.

[0147] Compared to hard thresholding, soft thresholding performs smoother signal processing near the threshold. Because it linearly shrinks the coefficients (rather than simply setting them to zero), it helps preserve more signal detail and characteristics.

[0148] The threshold of a soft threshold can be adjusted according to different noise levels and signal characteristics. This makes the soft thresholding method more adaptable to a variety of applications.

[0149] The mathematical definition of semi-soft thresholding is:

[0150]

[0151] Where Dj represents the detail coefficients of the j-th level wavelet decomposition. The coefficients after thresholding. and These are the first preset threshold and the second preset threshold, respectively. It is a symbolic function.

[0152] The main characteristic of a semi-soft thresholding function is that it retains a certain signal coefficient within a certain threshold, rather than setting it to zero. This allows it to preserve some low-intensity signal information while denoising, providing better smoothness and selectivity.

[0153] For signal coefficients whose boundaries are near the threshold, the semi-soft thresholding function can reduce artifacts caused by noise, thereby improving the stability of the results. This method helps to avoid complete rejection at low signal strengths and effectively suppresses the effects of noise.

[0154] Instead of setting all coefficients less than the threshold to zero, semi-soft thresholds reduce them, making them more flexible in adapting to different types of signals, especially when there is a lot of noise or uneven noise levels in the signal.

[0155] Preferably, set =2 To balance smoothness and detail retention.

[0156] In the embodiments of this application, in the shallow layer dominated by high-frequency noise (such as the first layer), an aggressive denoising threshold rule is adopted, a hard threshold function is selected, and the preset threshold is determined based on the noise standard deviation σ estimation: σ=median(|D1|) / 0.6745, λ=3σ. That is, the median of the absolute value of D1 is calculated and divided by 0.6745 to obtain the standard deviation, and the preset threshold is set to three times the standard deviation.

[0157] In layers 2-3, where mixed noise and weak signals exist, a semi-soft thresholding function is selected, and the preset threshold is determined by minimizing the Stein unbiased risk estimate. The optimal threshold is determined by minimizing the Stein unbiased risk estimate.

[0158] SURE (Stein's Unbiased Risk Estimator) is a mathematical method for estimating denoising error. Its core idea is that it can estimate the mean square error (MSE) after denoising without knowing the real signal, based solely on the observed noisy data.

[0159] In other words, SURE provides an unbiased estimator for evaluating denoising performance at different thresholds λ without knowing the true signal. By minimizing SURE, the optimal threshold λ can be found.

[0160] In the fourth and higher layers dominated by low-frequency effective signals, a conservative thresholding rule (λ=1.5σ) is adopted, and a soft thresholding function is selected. The preset threshold is determined by dynamically adjusting the energy ratio. .

[0161] in, denoted as the total signal energy, and β as the adjustment factor (values ​​range from 0.8 to 1.2).

[0162] Inverse wavelet transform reconstruction (IWT) is a key step in wavelet analysis. It recombines the coefficients obtained from wavelet decomposition to recover an approximate representation of the original signal. Its core idea is to reconstruct the signal inversely through a linear combination of wavelet basis functions to achieve signal denoising.

[0163] Specifically, it can be expressed as a formula: .

[0164] in, The initial denoised signal is represented by IWT, which indicates inverse wavelet transform reconstruction, and J represents the dynamically selected decomposition level in the above steps. These are the detail coefficients after processing with a combination of layered thresholds. is the approximation coefficient for the deepest layer.

[0165] However, it is worth noting that although the effective vibrational components are retained in the initial noise reduction signal, some noise remains, such as cross-band coupling noise caused by factors such as temperature drift and mechanical resonance.

[0166] In response, in step S303, this application employs dual-tree complex wavelet transform (DTCWT) to perform secondary decomposition and reconstruction of the initial noise-reducing signal in order to reduce residual noise.

[0167] The dual-tree complex wavelet transform uses two parallel real wavelet trees (Tree A and Tree B) to generate the real and imaginary parts of the wavelet coefficients, respectively, which are then combined to form complex coefficients. This structure simulates the characteristics of complex analytic wavelets, exhibiting approximate translation invariance and richer directional information.

[0168] Preferably, a dual-tree complex wavelet decomposition with a layer of 4 is selected to concentrate the residual noise in the low frequency (<10Hz). The frequency band division of the 4-layer decomposition (62.5-125Hz, etc.) can completely cover the target frequency band, and the amount of computation is controllable.

[0169] Furthermore, in a φ-OTDR system, the accuracy of the phase signal directly affects the location of disturbance points. Phase errors are caused by various factors, such as the nonlinear phase shift introduced by changes in the fiber's refractive index with temperature / stress; phase deviation caused by temperature drift of electronic components during long-term operation; and low-frequency phase fluctuations caused by wind, rain, and geological activity. Therefore, compensation is needed to reduce phase errors and further improve the location accuracy of disturbance points.

[0170] Under undisturbed conditions, a known phase signal is injected, and the system output phase is recorded. Perform calibration and calculate the phase deviation: Slowly varying phase drift is fitted using low-frequency approximation coefficients, and the trend term is extracted using moving average or Kalman filtering. .

[0171] Compensation factor embedding reconstruction, for the complex coefficients C after DTCWT decomposition j =D j real +jD j imag Applying a phase compensation factor, we obtain .

[0172] in, Using the corrected complex coefficients, perform inverse DTCWT to obtain the phase-corrected signal: s corrected (t)=IDTCWT({C j corrected}), and perform inverse transformation reconstruction.

[0173] In long-distance monitoring, an adaptive filtering algorithm (LMS) is used to update θ(t) in real time. The actual phase signal can be expressed as... , where θ(t) is the time-varying phase deviation and n(t) is the noise.

[0174] Phase error θ(t) can cause distance calculation deviation Δx∝θ(t), especially in long-distance (>10 km) monitoring, where even a small phase drift (e.g., 0.1 rad) can cause meter-level positioning errors. By iteratively updating the filter weights, the mean square error between the output signal and the desired signal is minimized, and θ(t) is estimated in real time.

[0175] Establish a signal model where the reference signal d(t) is a known or estimated phase-bias-free signal (such as a calibration signal or a low-frequency trend term), and the input signal x(t) is the actual measured signal containing phase bias. The filter output is Where w(t) is the adaptive weight, corresponding to the phase compensation factor. The error is calculated as follows: The weights have been updated to... Where μ is the step size factor, which controls the convergence speed and stability. The conjugate of x(t) (to be considered for complex signals).

[0176] The approximation coefficients are low-pass filtered (cutoff frequency 0.1Hz) to obtain... This completes the low-frequency trend extraction. A wave with a known calibration frequency is periodically injected, at which point d(t) is in ideal phase.

[0177] Initialize the weight w(0) = 1 (no compensation), with a step size μ = 0.01, and dynamically adjust it according to the signal power: μ(t) = 0.1 / (1 + ||x(t)|| 2 ), ensuring 0 < μ < 1 / tr(R), where R is the autocorrelation matrix of the input signal. Update w(t) point by point according to the sampling points, and calculate the instantaneous phase deviation θ(t) = arg(w(t)). For the complex wavelet coefficients C j (t) Apply compensation factor Output reconstructed signal .

[0178] However, even if different wavelet bases can be used for wavelet analysis on different perturbation phase signals, this approach still has some drawbacks. For example, it is limited by the preset finite wavelet base types and the preset hierarchical thresholding strategy, which may make it difficult to handle unforeseen complex noise reduction scenarios.

[0179] In this regard, such as Figure 3B As shown, in step S310, this application aims to utilize a multi-granularity feature residual stripper to extract shallow features (first-granularity perturbation feature map, second-granularity perturbation feature map) and deep semantic features (third-granularity perturbation feature map) from the preprocessed perturbed phase signal. This allows for the simultaneous extraction of local mutation information and global patterns in the signal (which helps in understanding the event type macroscopically).

[0180] The multi-granularity feature residual stripper includes a convolutional processing layer, a first residual module, a second residual module, and a third residual module.

[0181] Specifically, the convolutional processing layer of the multi-granularity feature residual stripper first performs convolution, batch normalization and activation processing on the preprocessed perturbation phase signal to generate a first feature map, and then uses the first feature map as the input of the first residual module, and the first residual module outputs the first granularity perturbation feature map.

[0182] Simultaneously, the first granularity perturbation feature map serves as the input to the second residual module, which outputs a second granularity perturbation feature map. Furthermore, the second granularity perturbation feature map serves as the input to the third residual module, which outputs a third granularity perturbation feature map.

[0183] Here, convolution, batch normalization, and activation processing (such as ReLU and LeakyReLU) initially extract local time-frequency features of the signal. Normalization stabilizes the training, and activation introduces nonlinear expressive power. Specifically, the first-granularity perturbation feature map extracts local details and describes minute changes in the signal; the second-granularity perturbation feature map captures medium-granularity information and perturbation features with a wider time domain; and the third-granularity perturbation feature map characterizes the overall trend of the signal.

[0184] It is worth mentioning that in traditional CNNs, ordinary convolution stacking may cause shallow features (such as low-frequency information) to be lost in deeper layers, while residual structures can retain the corresponding feature information after each processing step. This feature information can serve as input to subsequent models, providing an important data source for strengthening key shallow features.

[0185] Then in step S320, the key features of the shallow features are enhanced based on the spatial attention mechanism, so as to automatically suppress non-key areas (such as frequency bands where environmental noise is concentrated) through attention weights.

[0186] Here, the spatial attention mechanism highlights key regions related to external disturbances and suppresses irrelevant background noise by dynamically calculating the weight of each spatial location in the feature map. In other words, it can automatically identify high-probability disturbance regions through attention weights (such as 0~1 values ​​for Sigmoid activation).

[0187] It is worth mentioning that the perturbation features in the shallow layer (such as transient time-varying characteristics in impact scenes) are more important than the low-frequency stable background noise. The spatial attention mechanism simulates the focusing ability of the human eye to apply more attention to important perturbation features from the background noise, making the subsequent model express the comprehensive perturbation features more delicately.

[0188] Next, in step S330, a bidirectional dynamic fusion machine is used to adaptively adjust the fusion weights between multi-granularity perturbation features, thereby alleviating the fusion conflict caused by the direct mixing of multi-layer perturbation features and obtaining a comprehensive perturbation feature representation.

[0189] In existing technologies, Feature Pyramid Networks (FPNs) preserve detailed information in shallow features and semantic information in deep features by fusing shallow and deep features. However, FPNs fuse data in a relatively simple way, directly adding and fusing deep and shallow features. This fusion method may lead to problems such as information conflict and the introduction of noise.

[0190] Specifically, deep features typically contain high-level semantic information (such as object category and overall structure), while shallow features retain more low-level details (such as edges and textures). The two differ significantly at the semantic level, and directly adding them together can lead to a mismatch in the feature space, potentially introducing noise rather than useful information.

[0191] Furthermore, shallow features have a small receptive field and focus on local details, while deep features have a large receptive field and focus on global context. Directly adding them together may cause the fused features to lose the advantage of multi-scale information due to scale differences, especially a decrease in the ability to capture details.

[0192] At the same time, if there is a contradiction between shallow features and deep semantic features, such as shallow features detecting an edge while deep features ignore the region, adding them together may blur important features.

[0193] In the technical solution of this application, it is expected that a bidirectional dynamic fusion device will be used to first perform attention-based feature extraction on shallow features in order to retain and enhance the weak features contained therein; and to adjust and guide deep features in order to balance the fusion between shallow and deep features.

[0194] More specifically, the second perturbation enhancement feature map and the third granularity perturbation feature map are input into the first bidirectional dynamic fusion layer to obtain the perturbation fusion feature map; then the first perturbation enhancement feature map and the perturbation fusion feature map are input into the second bidirectional dynamic fusion layer to obtain the multi-granularity perturbation feature map.

[0195] The first bidirectional dynamic fusion layer performs pointwise convolution, batch normalization, and activation processing on the second perturbation enhancement feature map, and then performs elementwise multiplication between the processed feature map and the third-granularity perturbation feature map to obtain the first fused feature map; the formula is expressed as follows:

[0196]

[0197] In this system, F2 is the second perturbation enhancement feature map, F3 is the third granularity perturbation feature map, LR is the Leaky ReLU function, sigmoid is the sigmoid function, BN is batch normalization, and PConv is point-wise convolution. Element-wise dot product, F r1 This is the first fused feature map;

[0198] Simultaneously, the third-granularity perturbation feature map is subjected to global average pooling, full convolution, batch normalization, and activation processing. The processed feature map is then multiplied element-wise with the second perturbation enhancement feature map to obtain the second fused feature map; the formula is expressed as follows:

[0199]

[0200] Where GAP stands for Global Average Pooling, FC stands for Fully Connected Layer, and F... r2 This is the second fusion feature map;

[0201] Finally, the first fused feature map and the second fused feature map are added together according to their weights to obtain the perturbation fused feature map.

[0202] Similarly, the second bidirectional dynamic fusion layer fuses the first perturbation-enhanced feature map and the perturbation-fused feature map through similar processing. In this way, by preserving and enhancing shallow features with attention, regulating and guiding deep semantic features, and setting weight parameters to balance the information fusion ratio between features, multi-granular feature representations can be better integrated.

[0203] Finally, in step S340, a decoder is used to construct a denoiser, which automatically restores and reconstructs the comprehensive disturbance feature representation by progressively upsampling through deconvolutional layers. In this way, the restored disturbance-denoised signal is intelligently generated, shifting from a highly manual reliance to a data-driven adaptive denoising method, thus better assisting professional technicians in cable signal analysis. This approach, which transforms expert experience in signal processing into learnable network parameters, is particularly suitable for monitoring cable external damage in complex scenarios.

[0204] It is worth mentioning that the reason for not directly identifying external damage events but instead denoising the signal is that the model's judgment is uninterpretable, and denoising the signal allows for visualization, making it easier for technicians to use.

[0205] like Figure 4 As shown, this application also provides a cable damage prevention signal noise reduction system, which includes:

[0206] The disturbance signal acquisition module 110 is used to acquire the original disturbance phase signal obtained by the distributed optical fiber vibration sensing module based on φ-OTDR.

[0207] The preprocessing module 120 is used to preprocess the original perturbation phase signal to obtain the preprocessed perturbation phase signal;

[0208] The noise reduction module 130 is used to dynamically extract and reduce noise from the preprocessed perturbation phase signal to obtain the restored perturbation-denoised signal.

[0209] like Figure 5 As shown, this application also provides a terminal, which includes:

[0210] Processor 10 is used to couple with memory 20 and to read and execute instructions 30 stored in memory;

[0211] When the processor 10 is running, it executes instruction 30, causing the processor 10 to perform a cable damage prevention signal noise reduction method.

[0212] Figure 5 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0213] In some embodiments, memory 20 may be an internal storage unit of the terminal, such as the terminal's hard disk or memory.

[0214] In other embodiments, the memory 20 may also be an external storage device for the terminal, such as a plug-in hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc.

[0215] Furthermore, the memory 20 may include both internal storage units of the terminal and external storage devices.

[0216] The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code for installing the terminal.

[0217] The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores instructions 30, which can be executed by the processor 10 to perform a cable damage prevention signal noise reduction method.

[0218] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run instructions 30 stored in the memory 20 or process data, such as performing a cable damage prevention signal noise reduction method.

[0219] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments, and various changes and modifications can be made without departing from the spirit and scope of this application; all such changes and modifications fall within the scope of protection claimed in this application. The scope of protection of this application is defined by the appended claims and their equivalents.

Claims

1. A method for reducing signal noise in cables to prevent external damage, characterized in that, include: Acquire the original disturbance phase signal obtained by the distributed fiber optic vibration sensing module based on φ-OTDR; The original perturbation phase signal is preprocessed to obtain the preprocessed perturbation phase signal; The perturbation information of the preprocessed perturbation phase signal is dynamically extracted and denoised to obtain the restored perturbation denoised signal; The process includes dynamically extracting and denoising the perturbation information from the preprocessed perturbation phase signal to obtain a restored perturbation-denoised signal, including: By dynamically selecting the wavelet basis and the number of decomposition levels, wavelet decomposition is performed on the preprocessed perturbed phase signal to obtain multiple detail coefficients. The initial denoised signal is obtained by performing hierarchical threshold combination processing and inverse wavelet transform reconstruction processing on the multiple detail coefficients. The initial denoised signal is decomposed and reconstructed using dual-tree complex wavelet transform, and a phase compensation factor is embedded to improve positioning accuracy, thus obtaining the restored perturbation denoised signal.

2. The cable damage prevention signal noise reduction method according to claim 1, characterized in that, The original perturbation phase signal is preprocessed to obtain a preprocessed perturbation phase signal, including: The original perturbation phase signal is preprocessed by using moving average filtering and / or phase difference to obtain the preprocessed perturbation phase signal.

3. The cable damage prevention signal noise reduction method according to claim 2, characterized in that, By dynamically selecting the wavelet basis and the number of decomposition levels, wavelet decomposition is performed on the preprocessed perturbed phase signal to obtain multiple detail coefficients, including: The signal category of the preprocessed perturbation phase signal is determined based on the signal difference between the preprocessed perturbation phase signal and the perturbation phase signals of each group of samples in the sample library; Based on the signal category, a wavelet basis is selected, and the preprocessed perturbed phase signal is decomposed by wavelet to obtain the multiple detail coefficients. In the wavelet decomposition process, an adaptive strategy for the number of decomposition levels is used to control whether the wavelet decomposition is terminated.

4. The cable damage prevention signal noise reduction method according to claim 3, characterized in that, The signal category includes at least one of the following: transient impact, continuous vibration, and mixed noise; Wherein, when the signal category of the preprocessed perturbation phase signal is instantaneous impulse, the db4 wavelet and coif3 wavelet are selected as wavelet bases; When the preprocessed perturbation phase signal belongs to the signal category of continuous vibration, the sym4 wavelet is selected as the wavelet basis. When the preprocessed perturbation phase signal belongs to the mixed noise category, the bior3.3 wavelet is selected as the wavelet basis.

5. The cable damage prevention signal noise reduction method according to claim 4, characterized in that, In the wavelet decomposition process, an adaptive strategy for the number of decomposition levels is used to control whether the wavelet decomposition terminates, including: If the current decomposition level is less than 2, continue decomposing; If the current decomposition level is greater than or equal to 3, then perform the following steps: The termination threshold is determined based on the signal category to which the preprocessed perturbation phase signal belongs. Calculate the energy change rate of the current decomposition layer; The decomposition terminates when the energy change rate is less than the termination threshold; otherwise, the decomposition continues.

6. The cable damage prevention signal noise reduction method according to claim 5, characterized in that, The multiple detail coefficients are subjected to hierarchical threshold combination processing and wavelet inverse transform reconstruction processing to obtain the initial denoised signal, including: Hard thresholding is applied to detail coefficients at a decomposition level of 1. Semi-soft thresholding is applied to detail coefficients at decomposition levels 2 or 3. Soft thresholding is applied to detail coefficients at a decomposition level of 4 or higher.

7. The cable damage prevention signal noise reduction method according to claim 1, characterized in that, The preprocessed perturbation phase signal is subjected to dynamic perturbation information extraction and noise reduction to obtain a restored perturbation-denoised signal, including: The preprocessed perturbation phase signal is processed using a multi-granularity feature residual stripper to obtain a first-granularity perturbation feature map, a second-granularity perturbation feature map, and a third-granularity perturbation feature map. The first granularity perturbation feature map and the second granularity perturbation feature map are input into a convolutional neural network using a spatial attention mechanism to obtain a first perturbation enhancement feature map and a second perturbation enhancement feature map. The first perturbation enhancement feature map, the second perturbation enhancement feature map, and the third granularity perturbation feature map are processed using a bidirectional dynamic fusion processor to obtain a multi-granularity perturbation feature map. The multi-granularity perturbation feature map is processed using a decoder-based denoiser to obtain the restored perturbation denoised signal.

8. A cable damage prevention signal noise reduction system, characterized in that, include: The disturbance signal acquisition module is used to acquire the original disturbance phase signal obtained by the distributed optical fiber vibration sensing module based on φ-OTDR; The preprocessing module is used to preprocess the original perturbation phase signal to obtain the preprocessed perturbation phase signal; The noise reduction module is used to dynamically extract and reduce the noise of the preprocessed perturbation phase signal to obtain the restored perturbation-denoised signal.

9. A terminal, characterized in that, include: A processor is used to couple with memory and to read and execute instructions stored in memory; When the processor is running, it executes instructions that enable the processor to perform a method for reducing the noise of cable damage prevention signals.

Images