Power distribution network optical fiber fault positioning method and system based on time domain reflection

Abstract

The invention provides a power distribution network optical fiber fault positioning method and system based on time domain reflection, and relates to the technical field of power distribution networks, and the method comprises the steps: obtaining the basic parameter information of an optical fiber, generating and carrying out the phase modulation of an optical carrier signal, dividing the modulated signal into a chip sequence, and carrying out the orthogonal coding and time domain interleaving modulation; an enhanced detection signal with a space-time diversity characteristic is formed; signals are injected into the optical fiber, reflected signals are received, fault feature fingerprint information is extracted through multi-scale wavelet transform and adaptive threshold screening, fault type recognition and accurate position positioning are achieved, and fault detection accuracy and positioning precision are effectively improved.

Description

Technical Field

[0001] The present invention relates to distribution network technology, and in particular to a distribution network optical fiber fault location method and system based on time domain reflection. Background Art

[0002] As a vital component of the power system, the safe and stable operation of distribution networks is crucial for socioeconomic development and public well-being. With the advancement of smart grid development, fiber-optic communication systems for distribution networks are widely used in the monitoring, protection, and control of power systems. Optical fiber communication, with its advantages such as high bandwidth, immunity to electromagnetic interference, and information security, has become the primary method for communication in distribution networks. However, due to environmental factors, vandalism, and aging equipment, distribution network optical fibers are prone to various faults, such as fiber breaks, bends, and crushing. These faults directly impact the communication quality and operational reliability of the distribution network. Therefore, quickly and accurately locating optical fiber faults is crucial for shortening fault repair time and improving distribution network reliability.

[0003] Time domain reflectometry is a common method for locating optical fiber faults. Its basic principle is to transmit a probe signal into the fiber and determine the fault location by analyzing the returned reflected signal. While traditional optical time domain reflectometry has been widely used in optical fiber fault location, it still faces some challenges in the complex environment of distribution networks.

[0004] Existing distribution network fiber optic fault location technologies have the following defects and deficiencies: First, traditional detection signal designs lack spatiotemporal diversity characteristics and have limited anti-interference capabilities in complex noise environments. This results in low fault location accuracy and high false alarm rates in distribution network fiber optic systems under strong electromagnetic interference and complex environments. Second, conventional signal processing methods lack sufficient detection sensitivity for weak fault signals, making it difficult to identify early minor faults or complex situations where multiple faults coexist, and unable to accurately identify fault types and assess fault severity. Finally, existing fault feature extraction algorithms fail to fully consider the feature differences of different fault types in multi-scale domains and lack an adaptive feature screening mechanism, making it difficult to maintain stable and reliable positioning performance in dynamically changing distribution network environments. In particular, in long-distance optical fibers, positioning errors increase significantly with increasing distance. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for locating optical fiber faults in a distribution network based on time domain reflectometry, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for locating optical fiber faults in a distribution network based on time domain reflectometry, comprising: Obtaining basic parameter information of the distribution network optical fiber, generating an initial optical carrier signal according to the basic parameter information, and performing phase modulation on the initial optical carrier signal to obtain a modulated detection signal; Dividing the modulated detection signal into multiple chip sequences, encoding each of the chip sequences using an orthogonal code sequence, and interleaving and modulating the encoded multiple chip sequences in the time domain to generate an enhanced detection signal with time and space diversity characteristics; Injecting the enhanced detection signal into the distribution network optical fiber, receiving the reflected signal returned by the distribution network optical fiber, and demodulating the reflected signal to extract phase information and amplitude information from the reflected signal; Decomposing the phase information and the amplitude information into fault features based on a multi-scale wavelet transform, setting an adaptive threshold at each scale level for feature screening, mapping the screened features into a high-dimensional feature space through a nonlinear activation function for modeling, and extracting fault feature fingerprint information; Fault type identification and location are performed based on the fault feature fingerprint information to obtain a fault type determination result and fault location coordinates. The fault location is corrected based on the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result. A fault location report is generated based on the final fault location result.

[0007] Generating an initial optical carrier signal according to the basic parameter information, and performing phase modulation on the initial optical carrier signal to obtain a modulated detection signal includes: The optical fiber transmission characteristics are analyzed on a time scale corresponding to basic parameter information, and the main singular value components reflecting the optical fiber dispersion characteristics are extracted; nonlinear compensation is performed on the optical carrier signal based on the main singular value components to generate an initial optical carrier signal; the phase modulation depth is calculated based on the main singular value components, and the initial optical carrier signal is phase modulated to obtain a modulated detection signal.

[0008] Dividing the modulated detection signal into multiple chip sequences, encoding each of the chip sequences using an orthogonal code sequence, and interleaving and modulating the encoded multiple chip sequences in the time domain to generate an enhanced detection signal with time and space diversity characteristics includes: Multiplying the modulated detection signal with the transformation kernel function and performing fractional order transformation processing to obtain a transformation processing result; extracting a signal characteristic component of the transform processing result in a transform domain, determining an energy distribution threshold based on the signal characteristic component, determining a signal division position according to the energy distribution threshold, and dividing the modulated detection signal into a plurality of chip sequences based on the signal division position; Extracting characteristic frequency points from the transformation processing result, performing feature matching on the characteristic frequency points and the transformation kernel function, and generating an orthogonal code sequence group based on the result of the feature matching; Encoding each orthogonal code sequence in the orthogonal code sequence group with the corresponding chip sequence to obtain an encoded chip sequence group, and determining an interleaving parameter based on a feature enhancement effect of the transformation processing result; The encoded chip sequence group is time-rearranged according to the interleaving parameter, and the rearranged chip sequence is interleaved and modulated to generate an enhanced detection signal with time and space diversity characteristics.

[0009] The phase information and the amplitude information are decomposed into fault features based on multi-scale wavelet transform, and an adaptive threshold is set at each scale level for feature screening. The filtered features are mapped to a high-dimensional feature space through a nonlinear activation function for modeling. Extracting fault feature fingerprint information includes: Performing a multi-scale wavelet transform on the phase information and the amplitude information to obtain feature components at multiple scale levels, and calculating feature weights using a nonlinear activation function with a learnable shape parameter based on the time-frequency energy distribution of the feature components at each scale level; Dynamically quantize and encode the feature components of each scale level based on the feature weights, calculate the mapping relationship between the encoded feature components using the nonlinear activation function, and quantize and reconstruct the feature components according to the mapping relationship to obtain multi-scale reconstructed features; The multi-scale reconstruction features are improved in feature discrimination by using the learnable shape parameters of the nonlinear activation function, an adaptive threshold is determined based on the distribution of the mapped features, and the multi-scale reconstruction features are divided into main features and auxiliary features; The main feature and the auxiliary feature are mapped respectively using the nonlinear activation function, and the learnable shape parameters are iteratively optimized to obtain mapped high-dimensional features; The mapped high-dimensional features are mapped to different feature expression spaces using the nonlinear activation function, and the outputs of the different feature expression spaces are adaptively fused to generate final fault feature fingerprint information.

[0010] Dynamically quantizing and encoding the feature components of each scale level based on the feature weights, calculating the mapping relationship between the encoded feature components using the nonlinear activation function, and quantizing and reconstructing the feature components according to the mapping relationship to obtain multi-scale reconstructed features includes: constructing a dynamic quantization encoder according to feature weights, calculating a quantization step size based on the feature weights, and performing wavelet decomposition on the feature components based on the quantization step size to generate quantized feature values; Determining encoding parameters of the dynamic quantization encoder based on the accuracy of the quantization feature values, performing wavelet packet transform on the quantization feature values ​​at different scale levels by the dynamic quantization encoder to generate an encoding feature group, wherein the quantization accuracy of each encoding feature in the encoding feature group is determined by the encoding depth and the encoding base; Performing wavelet reconstruction on the coding feature group and the quantization feature value according to the coding depth and the coding basis to obtain a reconstructed feature, and adjusting the coding parameters of the dynamic quantization encoder based on the reconstructed feature; The reconstructed features of multiple scale levels are encoded using the updated encoding parameters to generate a feature encoding sequence, a fusion weight is generated based on the feature encoding sequence, and the fusion weight and the reconstructed features are jointly reconstructed to obtain a final multi-scale fusion feature.

[0011] Fault type identification and location are performed based on the fault feature fingerprint information to obtain the fault type determination result and fault location coordinates, including: Building a classification model based on fault feature fingerprint information, setting an adaptive wavelet decomposition scale for the fault feature fingerprint information, calculating a feature energy value at each wavelet decomposition scale, and obtaining a feature energy distribution; Inputting the characteristic energy distribution into the classification model, calculating the probability value of each fault type, and determining the type with the largest probability value as the fault type determination result; selecting a corresponding wavelet basis function according to the fault type determination result, and using the wavelet basis function to perform time-frequency analysis on the fault characteristic fingerprint information to obtain an energy distribution sequence on the time-frequency plane; A space mapping curve is established according to the energy distribution sequence, an energy extreme value point is searched on the space mapping curve, and a spatial position corresponding to the energy extreme value point is determined as a fault position coordinate.

[0012] Correcting the fault location according to the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result, and generating a fault location report based on the final fault location result includes: Based on the fault type determination result and the fault location coordinates, a chaotic attractor state space is constructed in combination with basic parameter information; the fault location coordinates are mapped to the chaotic attractor state space to obtain a chaotic orbit, and a state evolution sequence of the chaotic orbit is calculated based on the evolution equation of the chaotic attractor state space; characteristic change points are determined based on the time derivative of the state evolution sequence, and a state mapping matrix is ​​constructed based on the characteristic change points; The fault location coordinates are iteratively corrected using the state mapping matrix, wherein the step size of the iterative correction is related to the singular values ​​of the state mapping matrix. When the change in the singular values ​​of the state mapping matrix is ​​less than a preset change threshold, the iteration is stopped, the current fault location coordinates are determined as the final fault location result, and a fault location report is generated based on the stability analysis of the chaotic orbit and the characteristic change points.

[0013] A second aspect of an embodiment of the present invention provides a distribution network optical fiber fault location system based on time domain reflectometry, comprising: The first unit is configured to obtain basic parameter information of the optical fiber of the distribution network, generate an initial optical carrier signal according to the basic parameter information, and perform phase modulation on the initial optical carrier signal to obtain a modulated detection signal; a second unit, configured to divide the modulated detection signal into a plurality of chip sequences, encode each of the chip sequences using an orthogonal code sequence, and interleave and modulate the encoded plurality of chip sequences in the time domain, thereby generating an enhanced detection signal having a time and space diversity characteristic; The third unit is configured to inject the enhanced detection signal into the distribution network optical fiber, receive the reflected signal returned by the distribution network optical fiber, demodulate the reflected signal, and extract phase information and amplitude information from the reflected signal; A fourth unit is configured to perform fault feature decomposition on the phase information and the amplitude information based on a multi-scale wavelet transform, set an adaptive threshold at each scale level for feature screening, map the screened features to a high-dimensional feature space through a nonlinear activation function for modeling, and extract fault feature fingerprint information; The fifth unit is used to identify the fault type and locate the position based on the fault feature fingerprint information, obtain a fault type determination result and a fault location coordinate, correct the fault location based on the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result, and generate a fault location report based on the final fault location result.

[0014] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0015] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0016] The beneficial effects of this application are as follows: The present invention generates a detection signal with enhanced anti-interference capability by phase modulating the initial optical carrier signal and performing space-time diversity interleaving modulation using an orthogonal code sequence. This effectively improves the sensitivity and accuracy of fault detection in complex distribution network environments, and solves the technical problem of inaccurate positioning of traditional detection methods under conditions of strong noise interference.

[0017] The present invention uses multi-scale wavelet transform to decompose the features of the reflected signal, and sets an adaptive threshold at each scale level for feature screening. The screened features are mapped to a high-dimensional feature space through a nonlinear activation function, which can accurately capture the subtle feature differences of different types of faults, greatly improving the accuracy of fault type identification and the accuracy of fault location.

[0018] The present invention combines the basic parameter information of the distribution network optical fiber to correct the fault location and generate a detailed fault location report, thereby achieving rapid and accurate positioning of the distribution network optical fiber fault, reducing operation and maintenance costs, and improving the reliability and safety of the distribution network, with significant economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a method for locating optical fiber faults in a distribution network based on time domain reflectometry according to an embodiment of the present invention; Figure 2 This is a flowchart of the detection signal enhancement process according to an embodiment of the present invention; Figure 3 This is a flow chart of chaotic mapping for fault location according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0021] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0022] Figure 1 FIG. 1 is a flow chart of a method for locating optical fiber faults in a distribution network based on time domain reflectometry according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtaining basic parameter information of the distribution network optical fiber, generating an initial optical carrier signal according to the basic parameter information, and performing phase modulation on the initial optical carrier signal to obtain a modulated detection signal; Dividing the modulated detection signal into multiple chip sequences, encoding each of the chip sequences using an orthogonal code sequence, and interleaving and modulating the encoded multiple chip sequences in the time domain to generate an enhanced detection signal with time and space diversity characteristics; Injecting the enhanced detection signal into the distribution network optical fiber, receiving the reflected signal returned by the distribution network optical fiber, and demodulating the reflected signal to extract phase information and amplitude information from the reflected signal; Decomposing the phase information and the amplitude information into fault features based on a multi-scale wavelet transform, setting an adaptive threshold at each scale level for feature screening, mapping the screened features into a high-dimensional feature space through a nonlinear activation function for modeling, and extracting fault feature fingerprint information; Fault type identification and location are performed based on the fault feature fingerprint information to obtain a fault type determination result and fault location coordinates. The fault location is corrected based on the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result. A fault location report is generated based on the final fault location result.

[0023] In an optional implementation, generating an initial optical carrier signal according to the basic parameter information, and performing phase modulation on the initial optical carrier signal to obtain a modulated detection signal includes: The optical fiber transmission characteristics are analyzed on a time scale corresponding to basic parameter information, and the main singular value components reflecting the optical fiber dispersion characteristics are extracted; nonlinear compensation is performed on the optical carrier signal based on the main singular value components to generate an initial optical carrier signal; the phase modulation depth is calculated based on the main singular value components, and the initial optical carrier signal is phase modulated to obtain a modulated detection signal.

[0024] Obtain basic parameter information in the optical fiber transmission system, including optical fiber length of 50 km, dispersion coefficient of 17 picoseconds / (nanometer·km), and nonlinear coefficient of 1.3 watts -1 ·km -1 , fiber loss is 0.2 dB / km, input optical power is 5 mW, central wavelength is 1550 nm, and signal bandwidth is 10 GHz. These parameters form the basis for analyzing optical fiber transmission characteristics.

[0025] The singular value decomposition (SVD) method is used to analyze the optical fiber transmission characteristics on the time scale corresponding to the acquired basic parameter information. Specifically, the optical fiber transmission equation is discretized into a matrix form, with 1024 time sampling points and a time resolution of 5 picoseconds. The optical fiber transmission matrix contains dispersion terms and nonlinear terms. The dispersion terms are implemented using Fourier transforms, while the nonlinear terms are calculated using step-by-step iterative calculations. Singular value decomposition is performed on the constructed optical fiber transmission matrix, resulting in a series of singular values ​​and their corresponding left and right singular vectors.

[0026] From the decomposition results, the first three contributing singular values ​​are selected as the principal singular value components, with corresponding values ​​of 0.923, 0.756, and 0.421. The left and right singular vectors corresponding to the first principal singular value, 0.923, are denoted as u1 and v1, respectively, and contain the main information about the fiber's dispersion characteristics. According to actual calculations, the first ten elements of the u1 vector are [0.0312, 0.0345, 0.0378, 0.0456, 0.0534, 0.0623, 0.0712, 0.0801, 0.0845, 0.0889], and the first ten elements of the v1 vector are [0.0301, 0.0332, 0.0363, 0.0425, 0.0487, 0.0549, 0.0611, 0.0673, 0.0704, 0.0735].

[0027] Based on the extracted principal singular value components, the optical carrier signal is nonlinearly compensated to generate an initial optical carrier signal. During the compensation process, the principal singular value components are used to construct an inverse transmission operator, which can pre-compensate for nonlinear distortion during fiber transmission. In the specific implementation, the complex amplitude of the initial optical carrier signal is represented by A(t), where t is a time variable. Let the power distribution function of the initial optical signal, P(t) = |A(t)| 2 The signal obeys Gaussian distribution, has a peak power of 5 mW, and a pulse width of 100 picoseconds. The signal is convolved with the main singular value component u1 to obtain the pre-compensated signal A'(t).

[0028] During the pre-compensation calculation, the initial Gaussian pulse signal is discretized into 1024 points, then convolved with the discretized u1 vector. The convolution result is then multiplied by a compensation coefficient of 0.87 (determined through optimization experiments). After this processing, the generated initial optical carrier signal exhibits a slightly compressed Gaussian shape in terms of power, with the peak power reduced to 4.35 milliwatts. However, the pulse edges exhibit slight oscillation characteristics designed to compensate for nonlinear effects in optical fiber transmission.

[0029] The phase modulation depth is calculated based on the principal singular value component, and the initial optical carrier signal is phase modulated to obtain the modulated detection signal. The calculation of the phase modulation depth φ is related to the principal singular value 0.923 and the corresponding singular vector v1. Based on experimental optimization, φ = π × (1 - 0.923) × K, where K is the modulation coefficient and is set to 2.5. The calculated phase modulation depth φ is approximately 0.1205π.

[0030] Phase modulation is implemented using an electro-optical modulator. The input is the initial optical carrier signal A'(t) generated above, and the modulation signal is cos(ω0t), where ω0 is the modulation frequency, set to 1 / 5 of the signal bandwidth, or 2 GHz. The modulated detection signal is expressed as A'(t)·exp(jφ·cos(ω0t)), where j is the imaginary unit. In the actual implementation, the amplitude of the modulation signal is controlled at 0.8 volts, corresponding to the calculated phase modulation depth.

[0031] Through the above processing, the resulting modulated detection signal appears as a phase-modulated quasi-Gaussian pulse in the time domain. Its power envelope is identical to that of the initial optical carrier signal, but its phase varies over time according to a cosine function. This modulation structure effectively detects nonlinear and dispersion properties in optical fibers. In the frequency domain, the modulated signal exhibits a main lobe near the center frequency and side lobes at intervals of the modulation frequency. The side lobe power is approximately 0.01 times that of the main lobe.

[0032] This implementation extracts fiber transmission characteristics through singular value decomposition, enabling precise pre-compensation and phase modulation of the optical carrier signal, providing technical support for high-quality fiber detection. By adjusting the modulation parameters and pre-compensation coefficients, it can adapt to the detection requirements of different fiber systems.

[0033] In an optional embodiment, dividing the modulated sounding signal into multiple chip sequences, encoding each of the chip sequences using an orthogonal code sequence, and interleaving and modulating the encoded multiple chip sequences in the time domain to generate an enhanced sounding signal with time and space diversity characteristics includes: Multiplying the modulated detection signal with the transformation kernel function and performing fractional order transformation processing to obtain a transformation processing result; extracting a signal characteristic component of the transform processing result in a transform domain, determining an energy distribution threshold based on the signal characteristic component, determining a signal division position according to the energy distribution threshold, and dividing the modulated detection signal into a plurality of chip sequences based on the signal division position; Extracting characteristic frequency points from the transformation processing result, performing feature matching on the characteristic frequency points and the transformation kernel function, and generating an orthogonal code sequence group based on the result of the feature matching; Encoding each orthogonal code sequence in the orthogonal code sequence group with the corresponding chip sequence to obtain an encoded chip sequence group, and determining an interleaving parameter based on a feature enhancement effect of the transformation processing result; The encoded chip sequence group is time-rearranged according to the interleaving parameter, and the rearranged chip sequence is interleaved and modulated to generate an enhanced detection signal with time and space diversity characteristics.

[0034] like Figure 2 As shown, the method includes: The radar system generates the original detection signal using a linear frequency modulation signal as the base waveform, with a signal bandwidth of 500 MHz, a pulse width of 10 μs, and a sampling rate of 1 GHz. The original detection signal is generated using a digital waveform generator with a signal-to-noise ratio set to 15 dB. The system applies quadrature amplitude modulation to the original detection signal, with modulation parameters including a 10 GHz carrier frequency, 16 modulation orders, and a 50 MHz symbol rate. 16QAM modulation is used, mapping the bit stream to 16 points on a constellation diagram, with each point representing 4 bits of information. The modulated detection signal contains 10,000 sampling points, and the signal duration is 10 μs.

[0035] The modulated detection signal is multiplied by the transform kernel function and subjected to a fractional transform to obtain the transform result. The transform kernel function uses a fractional Fourier transform kernel, and the transform order α is set to 0.75. This transform order is selected based on signal characteristic analysis. When α = 0.75, it has the best energy concentration characteristics for linear frequency modulation signals. The specific form of the transform kernel function is the product of a complex exponential function and a non-integer power of time. The kernel function length matches the signal length and is 10,000 points. The system multiplies the modulated detection signal by the transform kernel function point by point to obtain a weighted signal. The weighted signal is transformed using a discrete fractional Fourier transform processor, implemented using a fast algorithm with a computational complexity of O(N log N), where N is the signal length. The transform result is a complex sequence containing 10,000 complex points, each of which contains a real part and an imaginary part.

[0036] The system extracts the signal characteristic components of the transform results in the transform domain. This extraction is based on energy distribution analysis. The system calculates the energy spectrum of the transform results, which is equal to the square of the modulus of the complex sequence. For a 10,000-point transform result, the energy spectrum also contains 10,000 points. The system smoothes the energy spectrum using a 51-point Hanning window for convolution to reduce the impact of noise. The smoothed energy spectrum exhibits a multimodal distribution, containing multiple regions of concentrated energy corresponding to different frequency components of the signal. The system calculates statistical properties based on the smoothed energy spectrum, including mean, standard deviation, kurtosis, and skewness. For the signal in this example, the energy spectrum has a mean of 0.025, a standard deviation of 0.018, a kurtosis of 3.75, and a skewness of 0.82.

[0037] The energy distribution threshold is determined based on the signal's characteristic components. This threshold is set adaptively using the mean plus the standard deviation multiplied by an adjustment factor. The adjustment factor is dynamically adjusted based on kurtosis and skewness, increasing for distributions with higher kurtosis and decreasing for distributions with higher skewness. In this example, the adjustment factor is calculated to be 2.1, and the energy distribution threshold is set to 0.063. The system marks points in the energy spectrum with an energy value greater than the threshold as feature points, and points with an energy value less than the threshold as non-feature points. Feature points account for 28% of the total number of points and are primarily located in specific regions of the transform domain, representing the primary energy concentration areas of the signal.

[0038] The signal division position is determined based on the energy distribution threshold. Signal division is based on the distribution characteristics of feature points in the transform domain. The system combines continuous feature points into feature regions and identifies a total of four main feature regions, each located at a different position in the transform domain. The system calculates the center position and span for each feature region. The center position corresponds to a specific moment in the time domain signal, and the span corresponds to the duration of the time domain signal. Based on these parameters, the system determines the division position of the time domain signal, dividing the 10μs signal into four code chip sequences. Each code chip sequence lasts approximately 2.5μs and has approximately 2500 sampling points. The division positions are specifically the 0th, 2500th, 5000th, 7500th, and 10000th sampling points of the original signal.

[0039] The modulated detection signal is divided into multiple chip sequences based on the signal division positions. The original 10,000-point signal is divided into four chip sequences based on the division positions, each containing approximately 2,500 sampling points. The first chip sequence contains points 0 to 2,499 of the original signal, the second chip sequence contains points 2,500 to 4,999, the third chip sequence contains points 5,000 to 7,499, and the fourth chip sequence contains points 7,500 to 9,999. The system performs energy normalization on each chip sequence to ensure consistent energy levels for subsequent processing. After normalization, the energy mean of the four chip sequences is 1, and the standard deviation is less than 0.05.

[0040] Characteristic frequency points are extracted from the transform processing results. Characteristic frequency points correspond to the peak of the energy spectrum in the transform domain and represent the main frequency components of the signal. The system searches for local maximum points in the transform domain, setting a search window of 51 points. When the energy value of the center point is greater than that of all other points in the window and greater than the energy distribution threshold, it is marked as a characteristic frequency point. The system identifies a total of 16 characteristic frequency points, located at different positions in the transform domain, covering a frequency range of 250MHz to 450MHz. Each characteristic frequency point records its position index and energy value to form a set of characteristic frequency points. The system sorts the characteristic frequency points in descending order of energy value and selects the eight points with the highest energy as key characteristic frequency points. The position indices of these points are 1245, 3782, 5631, 6230, 7112, 7865, 8524, and 9130.

[0041] The system performs feature matching on the characteristic frequency points and the transform kernel function. Based on the matching results, an orthogonal code sequence group is generated. The feature matching process is based on the transform kernel function values ​​at the characteristic frequency points. The system calculates the transform kernel function values ​​at eight key characteristic frequency points, resulting in eight complex numbers. The system uses these complex numbers to construct a Walsh-Hadamard matrix with an order of 8, where each matrix element is either +1 or -1. This construction method involves quantizing the phase of the complex number, where phases between 0 and π are quantized to +1, and phases between π and 2π are quantized to -1. The resulting 8×8 matrix is ​​orthogonal, with the inner product between any two rows or columns being 0. The system extracts four rows from this matrix as orthogonal code sequences, each with a length of 8 and elements of either +1 or -1. The four selected sequences are rows 1, 3, 5, and 7 of the matrix. These sequences are mutually orthogonal and suitable for encoding different chip sequences.

[0042] Each orthogonal code sequence in the orthogonal code sequence group is encoded with the corresponding chip sequence to produce an encoded chip sequence group. The encoding process utilizes direct sequence spread spectrum technology, multiplying each sampling point in each chip sequence by the corresponding element in the orthogonal code sequence. Since the orthogonal code sequence length is 8, while the chip sequence length is approximately 2500, the system periodically repeats the orthogonal code sequence to equal the chip sequence length. After encoding, the original information in each chip sequence is spread across a wider frequency band, enhancing interference resistance. The four chip sequences are encoded with the four orthogonal code sequences to form an encoded chip sequence group, each of which still contains approximately 2500 sampling points.

[0043] The interleaving parameters are determined based on the feature enhancement effect of the transform processing results. This enhancement effect is evaluated by calculating the cross-correlation coefficient between the original signal and the transformed signal. The cross-correlation coefficient is 0.78, indicating that the transform processing has a good feature enhancement effect. The interleaving parameters include two key parameters: interleaving depth and interleaving pattern. The interleaving depth represents the number of units involved in interleaving and is set to 32. The interleaving pattern represents the interleaving rules and adopts a pseudo-random interleaving pattern. The system generates a pseudo-random sequence of length 32, with sequence elements ranging from 0 to 31, and no elements are repeated. The pseudo-random sequence is generated using a linear feedback shift register method, with the characteristic polynomial selected as an 8th-order irreducible polynomial and the initial state set to 0x5A. The generated pseudorandom sequence is [7, 15, 23, 31, 3, 11, 19, 27, 6, 14, 22, 30, 2, 10, 18, 26, 5, 13, 21,29, 1, 9, 17, 25, 4, 12, 20, 28, 0, 8, 16, 24].

[0044] The system reorders the encoded chip sequence groups based on the interleaving parameters, combining the four encoded chip sequences into a single long sequence of approximately 10,000 points. This long sequence is then divided into 32 equal-length data blocks, each approximately 312 points long. These 32 data blocks are then reordered according to the order specified by the pseudo-random sequence. For example, data block 0 in the original sequence is moved to position 28 in the new sequence, data block 1 in the original sequence is moved to position 20 in the new sequence, and so on. This reordering disperses previously adjacent data blocks to different positions in the sequence, enhancing the signal's ability to resist interference. After the reordering is complete, the system generates a new long sequence of approximately 10,000 points, the same length as the original sequence.

[0045] The reordered chip sequence is interleaved and modulated to generate an enhanced sounding signal with time and space diversity. Interleaved modulation utilizes orthogonal frequency division multiplexing (OFDM). The system divides the reordered long sequence into 64 subcarriers, with each subcarrier modulating 312 / 64≈5 data points. QPSK modulation is used, mapping the data points to four points in the complex plane. The system generates 64 orthogonal subcarriers with a frequency spacing of 7.8 MHz, covering a 500 MHz bandwidth. The mapped data points are modulated onto the corresponding subcarriers, and then all subcarrier signals are summed to generate an enhanced sounding signal in the time domain. The enhanced sounding signal exhibits diversity in both the time and frequency domains: the time domain diversity comes from the reordering, and the frequency domain diversity comes from OFDM. The resulting enhanced sounding signal has a length of 10,000 points, a duration of 10 μs, and a bandwidth of 500 MHz.

[0046] In practical applications, the system tested the performance of the enhanced detection signal, demonstrating its excellent performance in multipath environments. In a dual-path environment with a reflection coefficient of 0.5, the enhanced detection signal achieved a range resolution of 0.3 meters, a 20% improvement over the original detection signal. The sidelobe ratio of the waveform ambiguity function was -28 dB, a 5 dB reduction compared to the original signal. In an environment with a signal-to-noise ratio of 5 dB, the target detection probability reached 92%, a 15 percentage point improvement over the original signal. These performance improvements demonstrate the effectiveness of the proposed signal processing method, and the resulting enhanced detection signal is suitable for target detection tasks in complex electromagnetic environments.

[0047] In an optional embodiment, fault feature decomposition is performed on the phase information and the amplitude information based on a multi-scale wavelet transform, an adaptive threshold is set at each scale level for feature screening, and the screened features are mapped to a high-dimensional feature space for modeling using a nonlinear activation function. Extracting fault feature fingerprint information includes: Performing a multi-scale wavelet transform on the phase information and the amplitude information to obtain feature components at multiple scale levels, and calculating feature weights using a nonlinear activation function with a learnable shape parameter based on the time-frequency energy distribution of the feature components at each scale level; Dynamically quantize and encode the feature components of each scale level based on the feature weights, calculate the mapping relationship between the encoded feature components using the nonlinear activation function, and quantize and reconstruct the feature components according to the mapping relationship to obtain multi-scale reconstructed features; The multi-scale reconstruction features are improved in feature discrimination by using the learnable shape parameters of the nonlinear activation function, an adaptive threshold is determined based on the distribution of the mapped features, and the multi-scale reconstruction features are divided into main features and auxiliary features; The main feature and the auxiliary feature are mapped respectively using the nonlinear activation function, and the learnable shape parameters are iteratively optimized to obtain mapped high-dimensional features; The mapped high-dimensional features are mapped to different feature expression spaces using the nonlinear activation function, and the outputs of the different feature expression spaces are adaptively fused to generate final fault feature fingerprint information.

[0048] The phase and amplitude information of the power equipment are obtained. The phase information includes the phase angle data of the three-phase voltage and current, and the amplitude information includes the amplitude data of the three-phase voltage and current. The phase and amplitude information are collected through the power monitoring device. The sampling frequency is set to 12.8kHz, and 256 points are sampled per cycle. The collection time is from 4 cycles before the fault occurs to 16 cycles after the fault occurs, a total of 20 cycles of data, forming a 5120-point sampling sequence. The collected raw data is preprocessed by a digital filter to remove power frequency interference and high-frequency noise. The filter adopts a Butterworth bandpass design with a passband range of 10Hz to 5kHz. The preprocessed phase and amplitude information are stored separately according to the three phases, forming six basic signal sequences.

[0049] A multi-scale wavelet transform is performed on the phase and amplitude information to obtain characteristic components at multiple scale levels. Multi-scale decomposition is then performed using a discrete wavelet transform, with the db4 wavelet basis function selected for a five-layer decomposition. For each phase's phase and amplitude information, the system uses a discrete wavelet transform to obtain five levels of detail coefficients and one coarsest-scale approximation coefficient. The detail coefficients represent the high-frequency components of the signal, while the approximation coefficients represent the low-frequency components of the signal. Taking the phase A voltage amplitude signal as an example, the original signal contains 5120 sampling points. After five layers of wavelet decomposition, the lengths of the detail coefficients at each layer are 2560, 1280, 640, 320, and 160 points, respectively, and the length of the coarsest-scale approximation coefficient is 160 points. Together, these coefficients constitute the multi-scale characteristic components of the signal.

[0050] Based on the time-frequency energy distribution of the feature components at each scale level, a nonlinear activation function with learnable shape parameters is used to calculate feature weights. The time-frequency energy distribution is obtained by calculating the energy value of the feature components at each level, which is equal to the sum of the squares of the coefficients in the feature components. For the five-level decomposition of the phase A voltage amplitude signal, the energy percentages of the detail coefficients at each level are 5%, 8%, 15%, 25%, and 40%, respectively, while the energy percentage of the coarsest-scale approximation coefficient is 7%. The system designed a nonlinear activation function with learnable shape parameters, which is a variant of the Swish function with learnable shape parameters. The learnable shape parameters include the slope parameter a and the offset parameter b, with initial values ​​set to 1.0 and 0.0, respectively. The system inputs the energy percentages of each level into the nonlinear activation function to obtain the corresponding feature weights. After activation function mapping, the feature weights of the detail coefficients at the five levels are 0.15, 0.22, 0.35, 0.62, and 0.85, respectively, and the feature weight of the coarsest-scale approximation coefficient is 0.18. The feature weights reflect the contribution of the feature components at each level to the fault characteristics.

[0051] Based on feature weights, the feature components at each scale level are dynamically quantized and encoded, constructing a dynamic quantization encoder that adaptively sets quantization parameters based on the feature weights. The quantization step size is inversely proportional to the feature weight: higher feature weights result in smaller quantization steps and higher quantization accuracy. The system sets a base quantization step size of 0.05 and calculates the actual quantization step size for each level based on the feature weights. For the fifth-level detail coefficients with a feature weight of 0.85, the quantization step size is 0.059; for the first-level detail coefficients with a feature weight of 0.15, the quantization step size is 0.333. Based on the quantization step size, the system performs a wavelet packet transform on the feature components to generate quantized eigenvalues. The wavelet packet transform is an extension of the wavelet transform, further subdividing the feature components at each scale level and providing a finer frequency resolution. The system sets the number of wavelet packet decomposition layers based on the feature weights; higher feature weights result in deeper decomposition layers. For the 4th and 5th layer detail coefficients with feature weights greater than 0.6, the number of wavelet packet decomposition levels is set to 3; for the 2nd and 3rd layer detail coefficients with feature weights between 0.2 and 0.6, the number of decomposition levels is set to 2; for the 1st layer detail coefficients and approximation coefficients with feature weights less than 0.2, the number of decomposition levels is set to 1.

[0052] A nonlinear activation function is used to calculate the mapping relationship between the encoded feature components. The system applies the aforementioned nonlinear activation function to the coefficients of each subband after wavelet packet transformation to obtain the activated feature expression. The shape parameters of the activation function are optimized by the back-propagation algorithm, with the goal of maximizing the feature discrimination between different fault types. The system calculates the mutual information between different subbands and constructs a feature association graph, in which the nodes represent the subbands and the edges represent the strength of the association between the subbands. Based on the feature association graph, the system analyzes the mapping relationship between the feature components and identifies highly correlated feature groups and complementary feature groups. For the A-phase voltage amplitude signal, the system found that the detail coefficients of the 3rd and 4th layers have strong complementarity, while the detail coefficients of the 5th layer have a high correlation with the coarsest scale approximation coefficients.

[0053] The feature components are quantitatively reconstructed based on the mapping relationship to obtain multi-scale reconstructed features. The reconstruction process uses two stages: inverse wavelet packet transform and inverse wavelet transform. The system first applies inverse wavelet packet transform to each subband coefficient to obtain the reconstructed feature components at each scale level. The reconstructed feature components are then weighted and combined according to the feature weights to obtain a complete multi-scale reconstructed feature. During the reconstruction process, the system considers the mapping relationship between feature components, adopts a joint reconstruction strategy for highly correlated feature groups, and adopts a strategy of separate reconstruction followed by fusion for complementary feature groups. For the phase A voltage amplitude signal, the average relative error between the reconstructed signal and the original signal is 3.5%, and the structural similarity is 0.96, retaining the key feature information of the original signal.

[0054] The system applies the learnable shape parameters of a nonlinear activation function to the multi-scale reconstructed features to improve feature discrimination. The shape parameters of the nonlinear activation function are optimized for different fault signal types. For single-phase grounding faults, the slope parameter a is optimized to 1.5, and the offset parameter b is optimized to 0.2; for two-phase short-circuit faults, a is optimized to 1.8 and b to -0.1; and for three-phase short-circuit faults, a is optimized to 2.0 and b to 0.0. The system applies the optimized nonlinear activation function to transform the multi-scale reconstructed features, enhancing the discrimination of the fault features. The transformed features form a more dispersed distribution in the feature space, facilitating subsequent fault identification.

[0055] Based on the distribution of mapped features, an adaptive threshold is determined, and the multi-scale reconstructed features are divided into primary and auxiliary features. The statistical distribution of the transformed features is analyzed, and their mean, standard deviation, and kurtosis are calculated. The adaptive threshold is set to the mean plus 1.5 times the standard deviation. Eigenvalues ​​greater than the threshold are considered primary features, while those less than or equal to the threshold are considered auxiliary features. For the phase A voltage amplitude signal, after threshold division, the primary features account for 25% of the total features, primarily distributed in the detail coefficients of the 4th and 5th layers, representing the high-frequency transient characteristics of the fault. The auxiliary features account for 75% of the total features, primarily distributed in the detail coefficients and approximation coefficients of the 1st to 3rd layers, representing the low-frequency characteristics and persistent impact of the fault.

[0056] A nonlinear activation function is used to map the primary and auxiliary features separately, and the learnable shape parameters are iteratively optimized to obtain the mapped high-dimensional features. For the primary features, the system uses a steeper activation function, increasing the slope parameter a by 20% and decreasing the offset parameter b by 50% to enhance the nonlinear representation of the features. For the auxiliary features, a flatter activation function is used, decreasing the slope parameter a by 10% and increasing the offset parameter b by 100% to maintain a smooth transition. An iterative optimization algorithm is designed to continuously adjust the shape parameters by minimizing the distance within fault samples and maximizing the distance between fault samples. The optimization process uses gradient descent, with an initial learning rate of 0.01 that decays by 0.9 times every 50 iterations. After 200 iterations, the shape parameters of the primary and auxiliary features converged to their optimal values, resulting in a stable high-dimensional feature representation. The high-dimensional features are three times the dimensionality of the original features and contain richer fault characteristic information.

[0057] The system uses nonlinear activation functions to map the mapped high-dimensional features into different feature expression spaces. Three parallel feature expression spaces are designed: the time domain feature space, the frequency domain feature space, and the joint time-frequency feature space. In the time domain feature space, the system maintains the temporal structure of the high-dimensional features and applies a one-dimensional convolution operation to extract the time domain patterns. In the frequency domain feature space, the system performs a Fourier transform on the high-dimensional features to extract spectral features. In the joint time-frequency feature space, the system applies a wavelet packet transform to extract the joint time-frequency features. These three feature spaces extract different aspects of the fault characteristics, forming complementary feature expressions.

[0058] The outputs of different feature expression spaces are adaptively fused to generate the final fault signature fingerprint. This adaptive fusion is implemented using an attention mechanism. The system calculates the importance weights of the outputs of the three feature spaces, taking into account three metrics: information entropy, sparsity, and discriminability. Feature spaces with high information entropy, moderate sparsity, and high discriminability are assigned higher fusion weights. The system then weights the outputs of the three feature spaces according to the fusion weights to generate a preliminary fused feature. A nonlinear activation function is applied to the preliminary fused feature for a final transformation to generate the fault signature fingerprint. The fault signature fingerprint is represented as a 256-dimensional vector containing the key features and identification identifiers of the fault.

[0059] In a practical application, the power monitoring system at a 110kV substation collected phase and amplitude data for a single-phase grounding fault. The data was subjected to a multi-scale wavelet transform, yielding five layers of feature components. Time-frequency energy analysis calculated the feature weights for each layer to be 0.12, 0.18, 0.31, 0.67, and 0.89, respectively, with an approximation coefficient weight of 0.15. The system dynamically quantized the feature weights, with quantization steps ranging from 0.056 to 0.417. After nonlinear activation function mapping and quantization reconstruction, the resulting multi-scale reconstructed features achieved a structural similarity of 0.97 with the original signal. The system applied an adaptive threshold of 0.42 to divide the features into primary and auxiliary features, with the primary features accounting for 22%. After mapping and iterative optimization, a high-dimensional feature representation of 768 dimensions was obtained. The system then mapped the high-dimensional features into three feature representation spaces, with fusion weights of 0.45, 0.30, and 0.25, respectively. The fault feature fingerprint information finally generated successfully identified the fault type as a phase A single-phase grounding fault, with a positioning accuracy of 98.5%, providing a reliable basis for fault diagnosis and treatment.

[0060] In an optional embodiment, dynamically quantizing and encoding the feature components of each scale level based on the feature weights, calculating the mapping relationship between the encoded feature components using the nonlinear activation function, and quantizing and reconstructing the feature components according to the mapping relationship to obtain the multi-scale reconstructed features includes: constructing a dynamic quantization encoder according to feature weights, calculating a quantization step size based on the feature weights, and performing wavelet decomposition on the feature components based on the quantization step size to generate quantized feature values; Determining encoding parameters of the dynamic quantization encoder based on the accuracy of the quantization feature values, performing wavelet packet transform on the quantization feature values ​​at different scale levels by the dynamic quantization encoder to generate an encoding feature group, wherein the quantization accuracy of each encoding feature in the encoding feature group is determined by the encoding depth and the encoding base; Performing wavelet reconstruction on the coding feature group and the quantization feature value according to the coding depth and the coding basis to obtain a reconstructed feature, and adjusting the coding parameters of the dynamic quantization encoder based on the reconstructed feature; The reconstructed features of multiple scale levels are encoded using the updated encoding parameters to generate a feature encoding sequence, a fusion weight is generated based on the feature encoding sequence, and the fusion weight and the reconstructed features are jointly reconstructed to obtain a final multi-scale fusion feature.

[0061] Obtain raw feature data, which can be images, audio, or signal sequences collected by sensors. Taking image processing as an example, the system collects high-resolution images of 1024×1024 pixels, where each pixel contains three RGB channels and the pixel value range is 0-255. The raw feature data undergoes preprocessing, including normalization to map pixel values ​​to the range of 0-1 and denoising to reduce random noise interference. The preprocessed feature data is decomposed into feature components at multiple scale levels. The system uses discrete wavelet transform for multi-scale decomposition, selecting the db4 wavelet basis function and performing a five-layer decomposition to obtain feature components at different scale levels. Each layer of decomposition produces a low-frequency approximation component and a high-frequency detail component. The five-layer decomposition produces a total of five high-frequency detail components and one low-frequency approximation component, which constitute the feature component set.

[0062] The feature weights of the feature components at each scale level are calculated based on information entropy and local contrast. For the i-th level feature component, the system divides it into 16×16 local blocks and calculates the information entropy value of each block. A higher information entropy value indicates greater information content. The standard deviation of the pixel values ​​within the block is also calculated. A larger standard deviation indicates higher local contrast. The system then weights the information entropy value and standard deviation to obtain the block importance index, with weight coefficients of 0.6 and 0.4, respectively. All blocks are ranked by importance index, and the top 30% of blocks are selected as key blocks. The average importance index of these key blocks is used as the feature weight of the feature component at that scale level. The feature component weights for the five-level decomposition are 0.85, 0.72, 0.63, 0.51, and 0.42, respectively, with the low-frequency approximation component having a weight of 0.9. The feature weights reflect the importance of feature components at different scale levels in the final reconstruction.

[0063] A dynamic quantization encoder is constructed based on feature weights. The dynamic quantization encoder consists of multiple quantization units, each of which is responsible for processing feature components at a scale level. The quantization unit contains a quantization step calculation module, a nonlinear activation function module, and a coding parameter adjustment module. The system calculates the quantization step based on feature weights. The quantization step is inversely proportional to the feature weight. The higher the feature weight, the smaller the quantization step and the higher the quantization accuracy. The specific calculation method is: the quantization step is equal to the basic step divided by the square of the feature weight, and the basic step is set to 0.05. For the first-layer feature component with a feature weight of 0.85, the calculated quantization step is 0.069; for the fifth-layer feature component with a feature weight of 0.42, the quantization step is 0.283.

[0064] Wavelet decomposition of the feature components is performed based on the quantization step size to generate quantized eigenvalues. For each scale-level feature component, the system uniformly quantizes the feature components using the corresponding quantization step size. During the quantization process, the continuous values ​​of the feature components are mapped to discrete quantization levels. The number of quantization levels is determined by the quantization step size; a smaller quantization step size results in a greater number of quantization levels. For the first-level feature components, the quantization step size is 0.069, resulting in 15 quantization levels; for the fifth-level feature components, the quantization step size is 0.283, resulting in 4 quantization levels. The quantized eigenvalues, called quantized eigenvalues, retain the key information of the original feature components while achieving data compression.

[0065] The coding parameters of the dynamic quantization encoder are determined based on the accuracy of the quantization eigenvalues. The coding parameters include two key parameters: coding depth and coding base. The coding depth indicates the number of layers of the wavelet packet transform, and the coding base indicates the type of wavelet basis function. The system dynamically adjusts these two parameters based on the accuracy of the quantization eigenvalues. The higher the accuracy of the quantization eigenvalues, the deeper the coding depth and the more complex the coding base are assigned. The accuracy is calculated as the logarithm of the number of quantization levels, and the accuracy ranges from 0.6 to 1.2. For quantization eigenvalues ​​with an accuracy greater than 1.0, the coding depth is set to 3 and the coding base is selected as db6; for quantization eigenvalues ​​with an accuracy between 0.8 and 1.0, the coding depth is set to 2 and the coding base is selected as db4; for quantization eigenvalues ​​with an accuracy less than 0.8, the coding depth is set to 1 and the coding base is selected as haar.

[0066] The dynamic quantization encoder applies wavelet packet transforms to the quantized eigenvalues ​​at different scale levels to generate coding feature groups. The wavelet packet transform is an extension of the wavelet transform, decomposing not only low-frequency components but also high-frequency components, providing a finer frequency resolution. The system applies wavelet packet transforms to the quantized eigenvalues ​​at each scale level based on the previously determined coding depth and coding base. For the first-level quantized eigenvalues ​​at a coding depth of 3, the wavelet packet transform generates 8 subbands; for the second and third-level quantized eigenvalues ​​at a coding depth of 2, 4 subbands are generated; and for the fourth and fifth-level quantized eigenvalues ​​at a coding depth of 1, 2 subbands are generated. These subbands form a coding feature group, and the quantization precision of each coding feature in the coding feature group is determined by the coding depth and coding base.

[0067] Wavelet reconstruction is performed on the coded feature group and quantized feature values ​​based on the coding depth and coding basis to obtain reconstructed features. Wavelet reconstruction is the inverse process of wavelet decomposition, recombining components from different frequency bands into a complete signal. During the reconstruction process, the system first applies an inverse wavelet packet transform to each subband in the coded feature group to obtain preliminary reconstructed quantized feature values. Dequantization is then applied to the reconstructed quantized feature values ​​to map the discrete quantization levels back to a continuous value domain. The dequantization operation takes into account the effects of the quantization step size and the nonlinear activation function, bringing the reconstructed values ​​closer to the original feature components. The nonlinear activation function uses a rectified linear unit function, truncating negative values ​​to zero while leaving positive values ​​unchanged, thereby enhancing the sparsity of the reconstructed features. For the first-level feature components, the reconstruction error is kept within 3%; for the fifth-level feature components, the reconstruction error is approximately 12%.

[0068] The coding parameters of the dynamic quantization encoder are adjusted based on the reconstructed features. The error between the original feature components and the reconstructed features is calculated. A larger error indicates that the current coding parameters are suboptimal. The system uses a gradient descent method to update the coding parameters, with the goal of minimizing the reconstruction error. During the update process, the coding depth can be increased or decreased by one layer, and the coding basis can be switched between five wavelet bases: Haar, DB2, DB4, DB6, and Sym4. The system evaluates different parameter combinations and selects the parameter combination with the smallest reconstruction error as the updated coding parameter. During the actual optimization process, the coding depth of the first-layer feature components was increased from 3 to 4, and the coding basis was changed from DB6 to Sym4. The coding depth of the fifth-layer feature components remained at 1, and the coding basis was changed from Haar to DB2.

[0069] The updated coding parameters are used to encode the reconstructed features at multiple scale levels, generating a feature code sequence. During the encoding process, the system applies an optimized wavelet packet transform to the reconstructed features at each scale level, achieving a finer subband division. Further compression is performed on each subband using entropy coding techniques. Entropy coding employs Huffman coding, assigning variable-length codes based on the frequency of occurrence of values ​​within the subband. The encoded data forms a feature code sequence, which contains both the encoded data and the coding parameters. The coding parameters record the coding depth, coding base, and quantization step size used at each scale level, which are used for subsequent decoding and reconstruction.

[0070] Fusion weights are generated based on the feature coding sequence. These weights determine the contribution of reconstructed features at different scale levels to the final multi-scale fusion. The system analyzes the statistical properties of the feature coding sequence, including information entropy, code length, and coding efficiency. Scale levels with higher information entropy, shorter code length, and higher coding efficiency are assigned higher fusion weights. The resulting fusion weights are 0.32, 0.25, 0.18, 0.15, and 0.10 for the feature components at the first to fifth levels, respectively. The system further considers the influence of the original feature weights and takes a weighted average of the fusion weights and feature weights, resulting in final fusion weights of 0.45, 0.28, 0.20, 0.12, and 0.05.

[0071] The fusion weights and reconstructed features are jointly reconstructed to obtain the final multi-scale fusion features. The joint reconstruction process uses a weighted summation method to combine the reconstructed features at different scale levels according to the fusion weights. To maintain the structural integrity of the features, the system spatially aligns the reconstructed features at each scale level before weighted summation, ensuring that features at different scale levels are superimposed in the same coordinate space. The system also applies an edge-preserving filter to enhance feature boundaries. The filter adopts a bilateral filter design with a spatial domain radius of 3 and a value range radius of 0.1. The resulting multi-scale fusion features retain the main structure and detailed information of the original features while achieving data compression and feature enhancement.

[0072] In a specific image processing case, a 1024×1024 pixel medical image was processed. The original image contains important detailed structures and background information, requiring precise preservation of details while simultaneously compressing the data volume. The system calculated the feature weights for the five scale levels to be 0.88, 0.75, 0.62, 0.48, and 0.39, respectively. The quantization step sizes calculated based on these weights were 0.065, 0.089, 0.130, 0.217, and 0.328, respectively. The encoded feature set generated after wavelet packet transform contains 21 subbands, with a total data volume of 35% of the original image. After parameter optimization and encoding reconstruction, the final multi-scale fusion feature data volume is only 28% of the original image, while maintaining 97% structural similarity and a peak signal-to-noise ratio of 45dB, meeting the dual requirements of high-quality compression and feature preservation.

[0073] In an optional embodiment, the fault type identification and location are performed based on the fault feature fingerprint information to obtain the fault type determination result and the fault location coordinates, including: Building a classification model based on fault feature fingerprint information, setting an adaptive wavelet decomposition scale for the fault feature fingerprint information, calculating a feature energy value at each wavelet decomposition scale, and obtaining a feature energy distribution; Inputting the characteristic energy distribution into the classification model, calculating the probability value of each fault type, and determining the type with the largest probability value as the fault type determination result; selecting a corresponding wavelet basis function according to the fault type determination result, and using the wavelet basis function to perform time-frequency analysis on the fault characteristic fingerprint information to obtain an energy distribution sequence on the time-frequency plane; A space mapping curve is established according to the energy distribution sequence, an energy extreme value point is searched on the space mapping curve, and a spatial position corresponding to the energy extreme value point is determined as a fault position coordinate.

[0074] Fault signature fingerprint information in the power system is collected by the fault detection device and includes three-phase voltage and current waveform data. The sampling frequency is set to 12.8kHz, with 256 points sampled per cycle. Each phase data is converted to a digital signal via a 16-bit A / D converter and stored. The fault signature fingerprint information also includes basic data such as load current, voltage level, and system impedance parameters before the fault occurs. The waveform data is collected from 4 cycles before the fault to 16 cycles after the fault occurs, for a total of 20 cycles of data, forming a 5120-point sampling sequence. The collected raw data is preprocessed using a digital filter to remove power frequency interference and high-frequency noise. The filter adopts a bandpass design with a passband range of 10Hz to 5kHz.

[0075] A classification model was constructed based on the preprocessed fault fingerprint information. The model uses a support vector machine architecture and considers common power system fault types, classifying them into 10 categories: single-phase ground fault, two-phase short circuit fault, two-phase ground fault, three-phase short circuit fault, three-phase ground fault, line break fault, high-resistance fault, lightning strike fault, intermittent fault, and equipment fault. The classification model was trained on a historical fault database containing 5,000 labeled fault cases. The training process used a cross-validation method, splitting the dataset into training and validation sets in an 8:2 ratio. The training accuracy reached 97.5%. The kernel function of the classification model was a radial basis function, with a penalty factor C set to 100 and a kernel parameter γ set to 0.01.

[0076] An adaptive wavelet decomposition scale is set for the fault feature fingerprint information, and the number of wavelet decomposition layers J is automatically adjusted according to the frequency characteristics of the fault waveform. The calculation method is an adaptive algorithm based on Shannon entropy. In specific implementation, the system first performs wavelet decomposition on the fault waveform from 1 to 10 layers, calculates the Shannon entropy value of each decomposition layer, and determines the optimal number of decomposition layers when the rate of change of the entropy value is less than the preset threshold of 0.05. For faults with obvious high-frequency characteristics, such as lightning strikes, the number of decomposition layers is usually 7 to 8 layers; for faults with mainly low-frequency characteristics, such as line breaks, the number of decomposition layers is usually 4 to 5 layers. The system uses the db4 wavelet as the initial decomposition wavelet basis to obtain the wavelet coefficients of each layer.

[0077] The characteristic energy value is calculated at each wavelet decomposition scale to obtain the characteristic energy distribution. The energy value Ej is calculated for the jth wavelet decomposition coefficient layer. The energy value is calculated as the square sum of all wavelet coefficients in that layer. For a 5120-point sampling sequence, the number of coefficients obtained at each layer after decomposition is 2560, 1280, 640, 320, 160, 80, 40, and 20, respectively. Taking an 8-layer decomposition as an example, the energy distribution vector [E1, E2, E3, E4, E5, E6, E7, E8] is formed. The system normalizes the energy distribution vector so that the sum of the energy values ​​at each layer is 1, forming the characteristic energy distribution. For a phase A single-phase grounding fault, the normalized characteristic energy distribution is [0.05, 0.08, 0.12, 0.25, 0.30, 0.15, 0.03, 0.02], indicating that the energy is primarily concentrated in layers 4 and 5.

[0078] The characteristic energy distribution is input into the classification model to calculate the probability of each fault type. The softmax function is used to convert the classification model's output into probability values ​​for each fault type, forming a probability distribution vector P = [P1, P2, ..., P10], where Pi represents the probability of belonging to the i-th fault type. The system identifies the type with the highest probability as the fault type. For the above case, the calculated probability distribution is [0.92, 0.03, 0.02, 0.01, 0.01, 0.01, 0.00, 0.00, 0.00, 0.00], indicating a single-phase ground fault with a confidence level of 92%.

[0079] Based on the fault type determination, the corresponding wavelet basis function is selected, and a mapping relationship between fault type and optimal wavelet basis is established: single-phase ground fault corresponds to the db4 wavelet, two-phase short-circuit fault corresponds to the db6 wavelet, two-phase ground fault corresponds to the db8 wavelet, three-phase short-circuit fault corresponds to the sym4 wavelet, three-phase ground fault corresponds to the sym6 wavelet, line break fault corresponds to the coif3 wavelet, high-resistance fault corresponds to the bio3.5 wavelet, lightning strike fault corresponds to the rbio2.8 wavelet, intermittent fault corresponds to the dmey wavelet, and equipment fault corresponds to the haar wavelet. The optimal wavelet basis is selected based on extensive fault case analysis to ensure optimal characterization of the time-frequency characteristics of specific fault types. For cases determined to be single-phase ground faults, the system selects the db4 wavelet as the wavelet basis function for subsequent analysis.

[0080] Selected wavelet basis functions are used to perform time-frequency analysis on the fault signature fingerprint information, obtaining an energy distribution sequence on the time-frequency plane. The continuous wavelet transform (CWT) method is then used to transform the fault waveform at different scales. The scale parameter a ranges from 1 to 64, with a total of 32 discrete scale points. The translation parameter b ranges across the entire time window, with a step size set to twice the sampling interval. For each time-frequency point (a, b), the wavelet coefficients W(a, b) are calculated, and the square of the coefficients represents the energy value E(a, b) at that point. The system constructs a time-frequency energy distribution matrix with a size of 32 × 2560, where 32 represents the number of scales and 2560 represents the number of time points. For each column in the matrix, the maximum energy value and its corresponding scale are extracted to form the energy distribution sequence {(t1, E1, a1), (t2, E2, a2), ..., (t2560, E2560, a2560)}.

[0081] A spatial mapping curve is established based on the energy distribution sequence, mapping time t to spatial distance d. The mapping relationship is based on the propagation speed v of electromagnetic waves in the line. For overhead lines, v is 98% of the speed of light, which is approximately 2.94×10 8 m / s; for cable lines, v is 60% of the speed of light, which is about 1.8×10 8m / s. The conversion formula for time t and distance d is d = v × t. For the measurement data at both ends, the system uses the time difference positioning principle, while also accounting for wave reflection and refraction. The system creates a three-dimensional spatial mapping curve with coordinates (d, E, a), where d is the distance coordinate, E is the energy value, and a is the scale parameter. For the single-phase grounding fault case, the spatial mapping curve shows a significant energy concentration between 35 km and 40 km.

[0082] The energy extreme point is searched on the spatial mapping curve, and a search algorithm based on particle swarm optimization is designed. 50 search particles are initialized and distributed in the fault area. The fitness function of each particle is defined as the energy value E at that position. The particles iteratively update their positions, and the maximum number of iterations is set to 100. The convergence condition is that the optimal value changes by less than 0.1% after 10 consecutive iterations. During the search process, the system considers both the energy value E and the scale parameter a, giving priority to areas with high energy and moderate scale parameters. For complex line structures, the system introduces topological constraints to ensure that the search range conforms to the actual line direction. Through iterative search, the system finds the energy extreme point (d*, E*, a*), where d* represents the fault distance.

[0083] The spatial location corresponding to the energy extreme point is determined as the fault location coordinate, and the distance d* is converted into actual physical coordinates (x, y, z). This conversion process is based on the line's geographic information data. The system queries the line tower database to determine the section where the fault point is located and the adjacent tower numbers. The system also calculates the confidence interval for the fault location by extracting all points near the energy extreme point with energy values ​​no less than 80% of the extreme value, calculating the distance distribution of these points, and obtaining a 95% confidence interval for the fault location. For the single-phase ground fault case, the system ultimately determined the fault location coordinates to be 37.8 kilometers from the substation, corresponding to towers 72 and 73, with a 95% confidence interval of [37.5 km, 38.1 km].

[0084] To validate the effectiveness of the proposed method, an application test was conducted on a 110 kV transmission line with a total length of 85 kilometers, 156 towers, and LGJ-240 / 30 conductors. The simulated fault location was located 37.6 kilometers from the substation, and the fault type was a metallic ground fault on phase A. After the fault detection device acquired the fault waveform, the system processed the fault signature fingerprint information. Adaptive wavelet decomposition determined the optimal number of decomposition layers to be six, with a characteristic energy distribution of [0.04, 0.07, 0.13, 0.26, 0.32, 0.18]. The classification model determined the fault type to be a single-phase ground fault with a confidence level of 94.6%. The system used the db4 wavelet for time-frequency analysis, obtaining a time-frequency energy distribution matrix. After establishing a spatial mapping curve, the system found the energy extreme point at a distance of 37.8 kilometers. The final fault location error from the actual fault location was 0.2 kilometers, with a relative error of 0.24%, meeting the engineering application requirements.

[0085] This method also features adaptive adjustment capabilities, dynamically adjusting wavelet decomposition parameters and search strategies based on line parameters and operating conditions. For different line and fault types, the system automatically selects the most appropriate processing parameters, improving fault location accuracy and adaptability. Practical applications have demonstrated that this method achieves an average location error of less than 0.5% of the line length for all types of faults, significantly improving fault handling efficiency and power system reliability.

[0086] In an optional embodiment, correcting the fault location according to the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result, and generating a fault location report based on the final fault location result includes: Based on the fault type determination result and the fault location coordinates, a chaotic attractor state space is constructed in combination with basic parameter information; the fault location coordinates are mapped to the chaotic attractor state space to obtain a chaotic orbit, and a state evolution sequence of the chaotic orbit is calculated based on the evolution equation of the chaotic attractor state space; characteristic change points are determined based on the time derivative of the state evolution sequence, and a state mapping matrix is ​​constructed based on the characteristic change points; The fault location coordinates are iteratively corrected using the state mapping matrix, wherein the step size of the iterative correction is related to the singular values ​​of the state mapping matrix. When the change in the singular values ​​of the state mapping matrix is ​​less than a preset change threshold, the iteration is stopped, the current fault location coordinates are determined as the final fault location result, and a fault location report is generated based on the stability analysis of the chaotic orbit and the characteristic change points.

[0087] like Figure 3 As shown, the method includes: Obtain basic power line parameter information, including line length, tower location coordinates, conductor type, line impedance parameters, line load, and meteorological conditions. Parameter collection involves extracting static parameters from the line management database and obtaining dynamic operating parameters through the SCADA system. Basic parameter information is stored in a structured data table. Each record contains the parameter ID, parameter name, parameter value, unit, and timestamp.

[0088] The fault detection device collects voltage and current waveforms during line faults at a sampling frequency of 12.8kHz, sampling 256 points per cycle. The waveform data is digitized using a 16-bit A / D converter and transmitted to the processing unit. The system preprocesses the collected waveform data, including denoising, baseline calibration, and data normalization. Denoising is performed using the wavelet thresholding method, using the db4 wavelet with a five-level decomposition and a soft threshold value of 3.5 standard deviations.

[0089] The preprocessed waveform data is decomposed using the db6 wavelet algorithm at eight levels to extract the energy, phase, and polarity characteristics of each frequency band. The extracted feature vector contains 24 elements, corresponding to the energy percentage, phase difference, and polarity identifier of each frequency band. The system inputs the feature vector into the fault identification model, which uses a support vector machine algorithm with a radial basis function kernel function, a penalty factor C set to 100, and a kernel parameter γ set to 0.01. The model outputs the fault type determination, including the fault type code and confidence level.

[0090] The system calculates preliminary fault location coordinates based on the traveling wave method. The arrival time of the traveling wave head in the fault waveform is detected, and the wavelet transform modulus maximum method is used to determine the arrival time of the traveling wave, with a time resolution of 5 microseconds. For single-ended data, the system calculates the fault distance using the reflected wave method, with a wave velocity of 2.9 × 10^8 m / s. For dual-ended data, the system calculates the fault distance using the arrival time difference method. The preliminary fault location coordinates are expressed as (d, 0, h), where d is the distance from the fault point to the monitoring end and h is the conductor height.

[0091] A chaotic attractor state space is constructed to correct the fault location. An appropriate chaotic system model is selected based on the fault type. For single-phase grounding faults, the Lorenz chaotic system is used; for phase-to-phase short-circuit faults, the Rössler chaotic system is used; and for line-break faults, the Chen chaotic system is used. Taking the Lorenz system as an example, the control parameters are set to σ = 10, ρ = 28, and β = 8 / 3.

[0092] Combined with the basic parameter information, the chaotic system parameters are adjusted. The line length affects the parameter σ, and the adjustment formula is σ'=σ×(L / 100) 0.5 ; Line load level affects parameter ρ, and the adjustment formula is ρ'=ρ×(P / Prated) 0.3; The conductor model affects the parameter β, and the adjustment formula is β'=β×(S / 240) 0.2 , where S is the conductor cross-sectional area. For 110kV lines, the typical parameter adjustment range is σ'=832 and β'=2.5~3.0.

[0093] The initial fault location coordinates are mapped to the chaotic attractor state space using the following mapping relationship: x = d / L, y = 0, z = h / hmax, where L is the total line length and hmax is the maximum conductor height. For example, if the fault location is 45.7 km away and the conductor height is 25 m, the initial point after mapping is (0.457, 0, 0.833). Assume the line length is 100 km and the maximum conductor height is 30 m.

[0094] Chaotic orbits are calculated based on the chaotic attractor evolution equation. The fourth-order Runge-Kutta method is used for this calculation, with a time step of 0.01 and 2000 iterations. Each iteration calculates a state point (xi, yi, zi), forming a state sequence. The system stores the results in a three-dimensional array, where each element contains the time point ti and the corresponding state value. After the orbit calculation is complete, the system checks whether the orbit fills the characteristic region of the chaotic attractor and whether the number of orbit points meets the statistical analysis requirements (>1000 points).

[0095] The time derivatives of the state evolution sequence are calculated to identify characteristic change points. For each state point (xi, yi, zi), the derivative values ​​(dxi / dt, dyi / dt, dzi / dt) are calculated. The derivatives are calculated using the central difference method, with the formula dxi / dt=(xi+1-xi-1) / (2Δt). The system sets derivative threshold vectors (Tx, Ty, Tz). When the absolute value of the derivative in any dimension exceeds the corresponding threshold, the point is marked as a characteristic change point. The threshold setting is related to the system parameters, with typical values ​​being Tx=15, Ty=20, and Tz=25.

[0096] Cluster analysis is performed on the identified characteristic change points to remove outliers. The density-based clustering algorithm, DBSCAN, is used with parameters set to ε=0.05 and MinPts=4. After clustering, characteristic points in the main clusters are retained to form a streamlined set of characteristic change points {(xj, yj, zj)|j=1,2,...,k}. The system calculates the distribution characteristics of the characteristic points along each dimension, including mean, standard deviation, and skewness. These statistics are used to assess the uncertainty of the fault location.

[0097] Based on the feature change points, the state mapping matrix M is constructed. The feature change point coordinates are organized into a k×3 matrix, with each row corresponding to the three-dimensional coordinates of a feature point. The matrix is ​​centered and the mean of each column is subtracted. The singular value decomposition of the centered matrix is ​​performed to obtain M=UΣV T, where Σ is a diagonal matrix whose diagonal elements are singular values ​​s1 ≥ s2 ≥ s3 ≥ 0. The system extracts the singular values ​​(s1, s2, s3) and the right singular vector matrix V for subsequent iterative correction.

[0098] The fault location is iteratively corrected based on the state mapping matrix, and the iteration step is set to α=0.1 / s1 to ensure stable convergence of the correction process. The coordinate update formula for each iteration is (xnew, ynew, znew)=(xold, yold, zold)+α·V·(s1, 0, 0) T , where V is the right singular vector matrix, indicating the main direction of change. After each iteration, the system recalculates the chaotic orbit and state mapping matrix to obtain new singular values ​​(s1', s2', s3').

[0099] The singular value change Δs = |s1'-s1| + |s2'-s2| + |s3'-s3| is calculated and compared with a preset change threshold δ = 0.01. If the singular value change for three consecutive iterations is less than the threshold δ, the system stops the iteration process. The fault location coordinates at this point are determined as the final fault location result. The system then maps the state space coordinates back to physical space, obtaining the actual fault point distance d' = x·L and height h' = z·hmax.

[0100] The stability of the chaotic orbit is analyzed to evaluate the reliability of the positioning results. The maximum Lyapunov exponent λ is calculated. The small perturbation method is used with an initial perturbation size of 10 -10 The evolution time is 10 time units, and the average value is calculated from 20 initial points. When λ < 0.5, the chaotic orbit has good stability and the positioning result is highly reliable. When 0.5 ≤ λ < 1.0, the positioning result has medium reliability. When λ ≥ 1.0, the positioning result has low reliability and needs to be verified by other methods.

[0101] The spatial distribution of characteristic change points is analyzed to determine the fault range. The standard deviation σx of the characteristic points in the x-dimension is calculated, and the 95% confidence interval for the fault point is [d'-1.96·σx·L, d'+1.96·σx·L]. The system optimizes the fault range representation based on actual line conditions, such as tower location and terrain characteristics, and generates recommended inspection sections.

[0102] Generate a fault location report, which includes the following contents: basic fault information (occurrence time, duration, affected equipment); fault type and judgment confidence level; initial fault location coordinates and location method; chaotic system parameter settings and adjustment basis; number and distribution characteristics of characteristic change points; iterative correction process records, including coordinate changes and singular value changes for each iteration; final fault location results, including distance value, range and confidence level; description of the physical location corresponding to the fault point, such as tower number, GPS coordinates, etc.; chaotic orbit stability analysis results; recommended inspection range and route; preliminary analysis of the cause of the fault and treatment suggestions.

[0103] Fault location reports are sent to the operation and maintenance center and on-site mobile terminals via the data communication network. Reports are available in both PDF and HTML formats and include text descriptions, data tables, and visual charts. The mobile terminal application displays the fault location on an electronic map and provides navigation to guide maintenance personnel to the fault location. The system records the discrepancy between the fault location results and the actual fault location for subsequent optimization of chaotic system parameters and improved location accuracy.

[0104] In a real-world case, a single-phase grounding fault occurred on a 110kV transmission line during a thunderstorm. The system acquired line parameters: total length 87.5 km, conductor type LGJ-240 / 30, and a total of 154 towers. The fault detection device captured the fault waveform and determined it to be a single-phase grounding fault on phase A with a 98.7% confidence level. The preliminary fault location was 45.7 km from the substation. The system constructed a Lorenz chaotic attractor with parameters set to σ = 9.35, ρ = 29.4, and β = 2.78. Twenty-seven significant characteristic change points were identified, primarily distributed between x = 0.45 and 0.47. After nine iterative corrections, the final fault location was determined to be 46.2 km from the substation, corresponding to towers 84 and 85. The 95% confidence interval for the fault location was [45.8 km, 46.6 km]. The maximum Lyapunov exponent λ = 0.38, indicating high reliability of the location result. The maintenance personnel inspected according to the recommended inspection range and found a branch touching the wire at 46.3 km, which was only 0.1 km away from the system positioning result.

[0105] A second aspect of an embodiment of the present invention provides a distribution network optical fiber fault location system based on time domain reflectometry, comprising: The first unit is configured to obtain basic parameter information of the optical fiber of the distribution network, generate an initial optical carrier signal according to the basic parameter information, and perform phase modulation on the initial optical carrier signal to obtain a modulated detection signal; a second unit, configured to divide the modulated detection signal into a plurality of chip sequences, encode each of the chip sequences using an orthogonal code sequence, and interleave and modulate the encoded plurality of chip sequences in the time domain, thereby generating an enhanced detection signal having a time and space diversity characteristic; The third unit is configured to inject the enhanced detection signal into the distribution network optical fiber, receive the reflected signal returned by the distribution network optical fiber, demodulate the reflected signal, and extract phase information and amplitude information from the reflected signal; A fourth unit is configured to perform fault feature decomposition on the phase information and the amplitude information based on a multi-scale wavelet transform, set an adaptive threshold at each scale level for feature screening, map the screened features to a high-dimensional feature space through a nonlinear activation function for modeling, and extract fault feature fingerprint information; The fifth unit is used to identify the fault type and locate the position based on the fault feature fingerprint information, obtain a fault type determination result and a fault location coordinate, correct the fault location based on the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result, and generate a fault location report based on the final fault location result.

[0106] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0107] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0108] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than 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 they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating optical fiber faults in distribution networks based on time domain reflectometry, characterized in that: include: Obtaining basic parameter information of the distribution network optical fiber, generating an initial optical carrier signal according to the basic parameter information, and performing phase modulation on the initial optical carrier signal to obtain a modulated detection signal; Dividing the modulated detection signal into multiple chip sequences, encoding each of the chip sequences using an orthogonal code sequence, and interleaving and modulating the encoded multiple chip sequences in the time domain to generate an enhanced detection signal with time and space diversity characteristics; Injecting the enhanced detection signal into the distribution network optical fiber, receiving the reflected signal returned by the distribution network optical fiber, and demodulating the reflected signal to extract phase information and amplitude information from the reflected signal; Decomposing the phase information and the amplitude information into fault features based on a multi-scale wavelet transform, setting an adaptive threshold at each scale level for feature screening, mapping the screened features into a high-dimensional feature space through a nonlinear activation function for modeling, and extracting fault feature fingerprint information; Fault type identification and location are performed based on the fault feature fingerprint information to obtain a fault type determination result and fault location coordinates. The fault location is corrected based on the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result. A fault location report is generated based on the final fault location result.

2. The method according to claim 1, characterized in that Generating an initial optical carrier signal according to the basic parameter information, and performing phase modulation on the initial optical carrier signal to obtain a modulated detection signal includes: The optical fiber transmission characteristics are analyzed on a time scale corresponding to basic parameter information, and the main singular value components reflecting the optical fiber dispersion characteristics are extracted; nonlinear compensation is performed on the optical carrier signal based on the main singular value components to generate an initial optical carrier signal; the phase modulation depth is calculated based on the main singular value components, and the initial optical carrier signal is phase modulated to obtain a modulated detection signal.

3. The method according to claim 1, characterized in that Dividing the modulated detection signal into multiple chip sequences, encoding each of the chip sequences using an orthogonal code sequence, and interleaving and modulating the encoded multiple chip sequences in the time domain to generate an enhanced detection signal with time and space diversity characteristics includes: Multiplying the modulated detection signal with the transformation kernel function and performing fractional order transformation processing to obtain a transformation processing result; extracting a signal characteristic component of the transform processing result in a transform domain, determining an energy distribution threshold based on the signal characteristic component, determining a signal division position according to the energy distribution threshold, and dividing the modulated detection signal into a plurality of chip sequences based on the signal division position; Extracting characteristic frequency points from the transformation processing result, performing feature matching on the characteristic frequency points and the transformation kernel function, and generating an orthogonal code sequence group based on the result of the feature matching; Encoding each orthogonal code sequence in the orthogonal code sequence group with the corresponding chip sequence to obtain an encoded chip sequence group, and determining an interleaving parameter based on a feature enhancement effect of the transformation processing result; The encoded chip sequence group is time-rearranged according to the interleaving parameter, and the rearranged chip sequence is interleaved and modulated to generate an enhanced detection signal with time and space diversity characteristics.

4. The method according to claim 1, wherein The phase information and the amplitude information are decomposed into fault features based on multi-scale wavelet transform, and an adaptive threshold is set at each scale level for feature screening. The filtered features are mapped to a high-dimensional feature space through a nonlinear activation function for modeling. Extracting fault feature fingerprint information includes: Performing a multi-scale wavelet transform on the phase information and the amplitude information to obtain feature components at multiple scale levels, and calculating feature weights using a nonlinear activation function with a learnable shape parameter based on the time-frequency energy distribution of the feature components at each scale level; Dynamically quantize and encode the feature components of each scale level based on the feature weights, calculate the mapping relationship between the encoded feature components using the nonlinear activation function, and quantize and reconstruct the feature components according to the mapping relationship to obtain multi-scale reconstructed features; The multi-scale reconstruction features are improved in feature discrimination by using the learnable shape parameters of the nonlinear activation function, an adaptive threshold is determined based on the distribution of the mapped features, and the multi-scale reconstruction features are divided into main features and auxiliary features; The main feature and the auxiliary feature are mapped respectively using the nonlinear activation function, and the learnable shape parameters are iteratively optimized to obtain mapped high-dimensional features; The mapped high-dimensional features are mapped to different feature expression spaces using the nonlinear activation function, and the outputs of the different feature expression spaces are adaptively fused to generate final fault feature fingerprint information.

5. The method according to claim 4, characterized in that Dynamically quantizing and encoding the feature components of each scale level based on the feature weights, calculating the mapping relationship between the encoded feature components using the nonlinear activation function, and quantizing and reconstructing the feature components according to the mapping relationship to obtain multi-scale reconstructed features includes: constructing a dynamic quantization encoder according to feature weights, calculating a quantization step size based on the feature weights, and performing wavelet decomposition on the feature components based on the quantization step size to generate quantized feature values; Determining encoding parameters of the dynamic quantization encoder based on the accuracy of the quantization feature values, performing wavelet packet transform on the quantization feature values ​​at different scale levels by the dynamic quantization encoder to generate an encoding feature group, wherein the quantization accuracy of each encoding feature in the encoding feature group is determined by the encoding depth and the encoding base; Performing wavelet reconstruction on the coding feature group and the quantization feature value according to the coding depth and the coding basis to obtain a reconstructed feature, and adjusting the coding parameters of the dynamic quantization encoder based on the reconstructed feature; The reconstructed features of multiple scale levels are encoded using the updated encoding parameters to generate a feature encoding sequence, a fusion weight is generated based on the feature encoding sequence, and the fusion weight and the reconstructed features are jointly reconstructed to obtain a final multi-scale fusion feature.

6. The method according to claim 1, characterized in that Fault type identification and location are performed based on the fault feature fingerprint information to obtain the fault type determination result and fault location coordinates, including: Building a classification model based on fault feature fingerprint information, setting an adaptive wavelet decomposition scale for the fault feature fingerprint information, calculating a feature energy value at each wavelet decomposition scale, and obtaining a feature energy distribution; Inputting the characteristic energy distribution into the classification model, calculating the probability value of each fault type, and determining the type with the largest probability value as the fault type determination result; selecting a corresponding wavelet basis function according to the fault type determination result, and using the wavelet basis function to perform time-frequency analysis on the fault characteristic fingerprint information to obtain an energy distribution sequence on the time-frequency plane; A space mapping curve is established according to the energy distribution sequence, an energy extreme value point is searched on the space mapping curve, and a spatial position corresponding to the energy extreme value point is determined as a fault position coordinate.

7. The method according to claim 1, characterized in that Correcting the fault location according to the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result, and generating a fault location report based on the final fault location result includes: Based on the fault type determination result and the fault location coordinates, a chaotic attractor state space is constructed in combination with basic parameter information; the fault location coordinates are mapped to the chaotic attractor state space to obtain a chaotic orbit, and a state evolution sequence of the chaotic orbit is calculated based on the evolution equation of the chaotic attractor state space; characteristic change points are determined based on the time derivative of the state evolution sequence, and a state mapping matrix is ​​constructed based on the characteristic change points; The fault location coordinates are iteratively corrected using the state mapping matrix, wherein the step size of the iterative correction is related to the singular values ​​of the state mapping matrix. When the change in the singular values ​​of the state mapping matrix is ​​less than a preset change threshold, the iteration is stopped, the current fault location coordinates are determined as the final fault location result, and a fault location report is generated based on the stability analysis of the chaotic orbit and the characteristic change points.

8. A distribution network optical fiber fault location system based on time domain reflectometry, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is configured to obtain basic parameter information of the optical fiber of the distribution network, generate an initial optical carrier signal according to the basic parameter information, and perform phase modulation on the initial optical carrier signal to obtain a modulated detection signal; a second unit, configured to divide the modulated detection signal into a plurality of chip sequences, encode each of the chip sequences using an orthogonal code sequence, and interleave and modulate the encoded plurality of chip sequences in the time domain, thereby generating an enhanced detection signal having a time and space diversity characteristic; The third unit is configured to inject the enhanced detection signal into the distribution network optical fiber, receive the reflected signal returned by the distribution network optical fiber, demodulate the reflected signal, and extract phase information and amplitude information from the reflected signal; A fourth unit is configured to perform fault feature decomposition on the phase information and the amplitude information based on a multi-scale wavelet transform, set an adaptive threshold at each scale level for feature screening, map the screened features to a high-dimensional feature space through a nonlinear activation function for modeling, and extract fault feature fingerprint information; The fifth unit is used to identify the fault type and locate the position based on the fault feature fingerprint information, obtain a fault type determination result and a fault location coordinate, correct the fault location based on the fault type determination result and the fault location coordinates in combination with the basic parameter information to obtain a final fault location result, and generate a fault location report based on the final fault location result.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images