An underground power transmission cable fault location method and system for smart grid

Abstract

This invention belongs to the field of power system monitoring and fault diagnosis technology, specifically relating to a method and system for determining the fault location of underground transmission cables for smart grids, aiming to solve the problems of low positioning accuracy and slow response of traditional methods. The system employs an FPGA and ARM collaborative architecture to achieve high-precision signal acquisition. It unifies the data format through cubic spline interpolation resampling, linear scaling mapping, and standardization, and generates subsequences with local labels using sliding window segmentation. A lightweight one-dimensional convolutional neural network model is constructed, utilizing adaptive pooling and diagonal fraction decomposition mechanisms to achieve end-to-end fault location regression. During training, cross-entropy loss with invalid labels, the AdamW optimizer, and an early stopping strategy are used to improve model performance. Through the above technical solutions, this application achieves fault location with second-level response and meter-level accuracy, significantly outperforming traditional traveling wave and impedance methods, effectively improving grid operation and maintenance efficiency and power supply reliability.

Description

Technical Field

[0001] This invention belongs to the field of power system monitoring and fault diagnosis technology, specifically relating to a method and system for determining the location of faults in underground transmission cables for smart grids. Background Technology

[0002] With the deepening of smart grid construction, underground transmission lines, as the core carriers of urban power transmission, play a crucial role in ensuring power supply reliability through their safe and stable operation. Underground cables are buried in complex geological environments for extended periods, making them susceptible to multiple factors such as soil corrosion, mechanical damage, insulation aging, and external construction disturbances. If a fault cannot be quickly and accurately located, it will lead to prolonged repair cycles, expanded power outage areas, and even cascading failures. Therefore, achieving high-precision and timely fault location identification for underground transmission cables has become a key technical requirement in modern power grid operation and maintenance systems.

[0003] Among these, fault location technologies for underground cables mainly rely on physical model-driven methods such as traveling wave propagation characteristics, impedance variation patterns, or wide-area synchronous phasor measurements. The traveling wave method locates the fault by capturing the arrival time difference of the voltage or current traveling waves generated at the moment of the fault at both ends of the cable. The impedance method calculates the equivalent distance based on changes in electrical quantities before and after the fault point. Wide-area measurement technology uses phasor data synchronously acquired from multiple nodes to construct global constraint equations to inversely determine the fault location. While these methods possess certain theoretical feasibility, they face significant challenges in practical engineering applications.

[0004] Existing technologies generally suffer from the dual drawbacks of insufficient positioning accuracy and response delay. On the one hand, traveling wave and impedance methods are highly sensitive to signal quality and are easily affected by high-frequency noise, signal attenuation, sampling jitter, and uncertainties in line parameters, making it difficult to achieve precise positioning at the hundred-meter or even meter level in complex electromagnetic environments. On the other hand, while wide-area phasor measurement has global sensing capabilities, it relies on large-scale data communication and centralized computing, resulting in a lengthy processing flow that cannot meet the emergency response requirements of second-level response after a fault occurs. In addition, traditional methods often rely on human experience to perform secondary analysis of monitoring data, lacking the ability to automatically mine the implicit nonlinear characteristics and multi-scale time-series patterns in fault signals, leading to highly subjective diagnostic results and weak generalization ability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for determining the location of underground power transmission cables for smart grids. This method overcomes the aforementioned deficiencies in existing technologies by integrating high-precision signal acquisition, intelligent feature extraction, and real-time reasoning capabilities for determining the location of underground power transmission cables, thus breaking through the bottlenecks in accuracy, speed, and automation levels of existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for determining the location of faults in underground power transmission cables for smart grids, characterized by including:

[0008] The fault data acquisition and labeling steps involve acquiring fault signals from underground power transmission cables to obtain raw fault data, and labeling the fault point locations corresponding to the raw fault data.

[0009] The data resampling and label matching steps involve performing cubic spline interpolation resampling on the labeled original fault data to unify the data dimensions, while linearly scaling the labeled fault point locations to achieve dimensional adaptation between the data and the fault point labels.

[0010] The data preprocessing step involves standardizing the resampled data and then using a sliding window to split it into sub-data groups containing sub-sequence data and local fault labels.

[0011] The deep learning model construction and feature extraction steps include: constructing a deep learning model for power transmission line fault location; performing end-to-end feature extraction on the sub-data set; and outputting the fault score for the window position.

[0012] The model training and optimization steps involve training the model using a cross-entropy loss function with an invalid label masking mechanism and an optimizer with weight decay, and then determining the optimal fault location model by combining an early stopping mechanism.

[0013] The fault location determination step involves processing the fault data to be determined through data resampling and label matching steps and data preprocessing steps, and then inputting it into the optimal fault location model to output the global location of the fault point.

[0014] In this invention, preferably, in the fault data acquisition and labeling step, the fault signal acquisition is controlled by an FPGA+ARM microcontroller: the FPGA undertakes high-speed sampling control at a sampling rate of 200M, real-time signal preprocessing of 32nd-order FIR low-pass filtering, and FIFO data temporary storage tasks. The FIR low-pass filtering is implemented by parallel operation of the multiplier and adder resources inside the FPGA, and the signal delay after filtering is ≤20ns.

[0015] The ARM microcontroller is responsible for overall system control, data storage management, human-computer interaction, and communication with the host computer, transmitting pre-processed fault data in timestamp + waveform data format.

[0016] In this invention, preferably, in the data resampling and label matching step, the cubic spline interpolation is specifically implemented as follows: within the interval formed by any two adjacent original data points... Inside, the interpolation function is a cubic polynomial:

[0017] ,

[0018] in Indicates the first Data points, Indicates the interpolation point. Indicates the interpolation point The corresponding values, polynomial coefficients , , and Fault data points acquired through hardware, and meeting the interpolation conditions. First derivative continuous Second derivative continuity The constraints.

[0019] In this invention, preferably, in the data preprocessing and sub-data group generation steps, the sliding window segmentation rule includes: the starting position of the sliding window. ,

[0020] Where M is the total length of the standardized data, S is the sliding window step size, and W is the sliding window length; the generation rule for local fault labels is:

[0021] ,

[0022] in, This represents the location of the fault point before splitting; the sub-data group is a quadruple. , For window subsequences, Here, n is the local fault location label, n is the original sample number, and s is the window start position.

[0023] In this invention, preferably, the deep learning model construction and feature extraction steps include the following steps for the deep learning model for power transmission line fault location:

[0024] Input channel dimension expansion sub-step, for input ,in For batch size, Given the window length, the channel dimension expansion can be expressed as:

[0025] ;

[0026] The feature extraction sub-step extracts deep fault features by sequentially performing a first one-dimensional convolution with ReLU activation and a second one-dimensional convolution with ReLU activation.

[0027] The feature optimization sub-step maps deep fault features into a feature sequence of fixed length W through one-dimensional adaptive average pooling, thereby suppressing noise interference.

[0028] The dimension adaptation sub-step converts the pooled feature tensor from the dimension... Replace with ,in ;

[0029] The fault score calculation sub-step outputs the fault score for each window position through a position-by-position one-dimensional linear transformation and diagonal element extraction.

[0030] In this invention, preferably, the parameter initialization rules in the model training and optimization steps include:

[0031] The weights for the convolution linear transform are normally distributed random variables with bias initialized to zero; the weights for the ReLU nonlinear transform are variables with a mean of 0 and a variance of 0. The normal distribution, where K represents the number of input channels, and K represents the kernel size.

[0032] In this invention, preferably, in the model training and optimization steps, the cross-entropy loss function of the invalid label masking mechanism is:

[0033] ,

[0034] in, Indicates the result of loss. This indicates that the b-th batch is at the fault location. The output score is on. This represents the total number of all sample data. For window size, Let be a natural constant, when This represents the actual fault location of the b-th sample. =-1 indicates that the b-th sample does not have a fault point in this window. The masking coefficient is used to identify whether a location is a valid supervised sample. hour, =1, when hour, =0.

[0035] In this invention, preferably, the parameter update process of the AdamW optimizer with weight decay in the model training and optimization steps is as follows:

[0036] Update the first-moment exponential moving average , can be represented as: ,

[0037] in, It is the first-order moment decay factor. This represents the loss gradient for the current batch.

[0038] Update the second-order moment element-wise exponential moving average , can be represented as: ,

[0039] in, It is the second-order moment attenuation factor. The product of Hadamard;

[0040] Calculation bias correction , To compensate for the deviation caused by the initial value being zero:

[0041] ,

[0042] ,

[0043] Calculate the adaptive step size based on the modified moment estimation :

[0044] ,

[0045] By using a decoupling approach, weight decay and gradient updates are applied to the parameters. The parameter update can be expressed as: ;

[0046] in, This represents the set of trainable parameter tensors from the previous training iteration t. This represents the set of trainable parameter tensors for the current training iteration t+1. The current step size, This is the weight decay coefficient. This is an indicator function; the value is 1 when the corresponding parameter needs to be decayed, and 0 otherwise.

[0047] In this invention, preferably, in the data preprocessing and sub-data set generation step, the data after cubic spline interpolation resampling is calculated. Mean and standard deviation:

[0048] ,

[0049] ,

[0050] in This represents the mean. Indicates standard deviation,

[0051] Standardize each data point:

[0052] ,

[0053] in, For the standardized first Data points.

[0054] A fault location determination system for underground power transmission cables for smart grids, employing the method described in any one of the above, comprising:

[0055] The data acquisition module is used to collect fault signals from underground power transmission cables;

[0056] The data preprocessing module is used to perform data resampling and label matching, data preprocessing and subgroup generation, and complete cubic spline interpolation resampling, standardization and sliding window segmentation.

[0057] The model training module is used to build a deep learning model for fault location of transmission lines and to train the model using a loss function and optimization strategy.

[0058] The fault location module receives fault data to be identified, preprocesses the data, inputs it into the optimal fault location model, and outputs the global location and confidence distribution of the fault point.

[0059] The beneficial effects of this invention are:

[0060] This invention effectively solves the problems of high noise, sparse samples, and difficult label alignment in underground cable fault signal acquisition by constructing a high-precision data acquisition hardware platform with FPGA+ARM collaboration and combining data preprocessing strategies such as cubic spline interpolation, sliding window segmentation, and invalid label masking. Based on this, a lightweight one-dimensional convolutional neural network model is designed that can automatically extract multi-scale temporal features and achieve end-to-end fault location regression through adaptive pooling and diagonal fraction extraction mechanisms. The AdamW optimizer and early stopping mechanism introduced during training significantly improve the model's convergence stability and generalization ability. Ultimately, in actual deployment, the system achieves fault location with second-level response and meter-level accuracy, overcoming the shortcomings of traditional methods in complex electromagnetic environments, such as insufficient accuracy and response delay. This significantly shortens the power grid fault repair time and improves power supply reliability and the level of operation and maintenance automation. Simultaneously, the probability distribution form of the model output provides a confidence basis for operation and maintenance decisions, enhancing the interpretability and practicality of the diagnostic results. Attached Figure Description

[0061] Figure 1 This is a schematic diagram of the overall technical solution architecture of the underground power transmission cable fault location identification system for smart grids proposed in this invention;

[0062] Figure 2 This is a schematic diagram of the overall process of the method for determining the fault location of underground power transmission cables for smart grids proposed in this invention. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0065] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0066] This invention first collects fault data of underground power transmission cables using self-developed hardware equipment. The fault data is then transmitted to the fault location module on the host computer via the system communication module for algorithm analysis. The system structure is referenced... Figure 1 The overall technical architecture of the method and system for fault location determination of underground transmission cables for smart grids proposed in this invention includes a high-precision signal acquisition hardware subsystem, a data preprocessing and label alignment module, a deep learning fault location model, a model training optimization strategy, and a host computer visualization decision support module. The entire system uses an FPGA and ARM microcontroller as the core to build a hardware acquisition platform, ensuring high-fidelity acquisition of the original reflected fault signals. Subsequently, through a series of standardization, resampling, and sliding window segmentation operations, the original signals are transformed into structured data suitable for input to the deep learning model. The model directly outputs the probability distribution of the fault location through end-to-end learning. Finally, the host computer receives the processing results and visualizes them, providing maintenance personnel with intuitive and reliable decision-making support.

[0067] The main process of the method is as follows: Figure 2 As shown, the specific implementation process includes the following steps:

[0068] Step 1: Collect fault data. The hardware device for data acquisition in this invention uses an FPGA + ARM microcontroller as the main control core. The overall architecture follows the division of labor principle of FPGA handling high-speed tasks and ARM handling low-speed control. Combined with high-speed signal transmission, high-speed sampling, data storage, clock synchronization, and communication modules, it achieves accurate data acquisition and processing at a 200MHz sampling rate. The FPGA is responsible for 200MHz high-speed sampling control, real-time signal preprocessing, and high-speed data temporary storage. The ARM microcontroller is responsible for overall system control, data storage management, human-machine interaction, and communication with the host computer. All modules work collaboratively through standardized interfaces to ensure the real-time performance and accuracy of fault data acquisition.

[0069] In a specific embodiment, step 1 further includes:

[0070] Step 1.1: The pulse generator module generates a narrow pulse excitation signal and injects it into the faulty cable line, providing a source for the acquisition of reflected signals. First, the FPGA outputs a synchronization control signal to drive the high-speed pulse generator to generate a narrow pulse signal. The pulse width can be adjusted through the FPGA configuration register to ensure that the pulse rise edge steepness is ≤1ns, meeting the accuracy requirements. Second, the pulse signal is amplified to 5V-12V by a high-speed power amplifier, and then connected to the cable line through an impedance matching circuit to reduce signal reflection loss and ensure that the excitation signal is effectively injected into the cable.

[0071] Step 1.2: High-speed sampling by the analog-to-digital converter (ADC). First, the cable reflection signal is amplified by a differential amplifier circuit composed of a high-speed operational amplifier, then filtered by a low-pass filter to remove high-frequency noise, and finally normalized to the ADC input range by a level conversion circuit. Then, a high-speed ADC with 10-bit resolution and a 200MHz sampling rate is selected. Its sampling clock is generated by the phase-locked loop (PLL) module inside the FPGA. The PLL input clock is a 25MHz external high-precision crystal oscillator, which is multiplied by 8 to output a 200MHz sampling clock, ensuring that the sampling clock jitter is ≤50ps to avoid sampling timing errors. The FPGA is connected to the ADC through a low-voltage differential signal interface to receive the 10-bit parallel sampling data output by the ADC in real time. The sampling mode and gain parameters of the ADC are configured through the SPI interface. At the same time, the sampling clock generated by the FPGA and the pulse trigger signal of the signal transmission module are synchronized and calibrated to ensure that the time deviation between the sampling start time and the pulse transmission time is ≤1ns, ensuring the accuracy of the time axis of the reflected signal.

[0072] Step 1.3, noise suppression, implements 32nd order FIR low-pass filter with a cutoff frequency of 80MHz, uses the FPGA's multiplier and adder resources for parallel operation to filter out high-frequency noise during the sampling process, while ensuring that the delay of the filtered signal is ≤20ns, without affecting the waveform time axis accuracy.

[0073] Step 1.4, valid data extraction and temporary storage: The peak range of the reflected signal is identified by threshold comparison logic. Only the sampled data before and after the peak is retained, and invalid data without reflection information is removed. Then, the data is written to the FIFO. When the FIFO storage reaches the half-full threshold, the FPGA sends an interrupt request to the ARM microcontroller to trigger the ARM to read the data, thus avoiding data loss due to FIFO overflow. At the same time, the timing of high-speed writing and low-speed reading is matched by FIFO read and write clock isolation, with a write clock of 200MHz and a read clock of 100MHz.

[0074] Step 1.5: After the ARM responds to the interrupt request from the FPGA, it reads the preprocessed sampled data from the FPGA's FIFO through a 16-bit parallel interface at a reading rate of 100MHz, and sends the data to the fault location module of the host computer in the format of "timestamp + waveform data" via the communication module.

[0075] Step 2: Label the fault location for each fault data point. Label the physical location of the corresponding fault point for each fault data point collected in Step 1. Specifically, this labeling process is completed in a laboratory simulation environment or a historical fault database. For laboratory simulations, a short circuit or open circuit fault is artificially created at a specific location on an underground cable of known length L, and the reflected waveform data collected at this time is recorded. This physical distance of 2350 meters is used as the label for this data point. For historical data, the labeling is performed retrospectively based on the fault location confirmed in the repair records. All labeled data is stored in a waveform data file + physical location format, forming the basic dataset for subsequent model training.

[0076] Step 3: To ensure consistent model input dimensions, each fault waveform needs to be resampled from the original number of sampling points to a uniform length M. In this embodiment, M=1024, so cubic spline interpolation is used to achieve high-fidelity resampling of the fault data acquired by the hardware. The specific steps are as follows:

[0077] 3.1 Fault data sample points acquired by hardware , For sampling times in the time series, For the corresponding The value is taken, so the cubic spline is in each The up-interpolation function is expressed as a cubic polynomial:

[0078] ,

[0079] in Indicates the first Data points, Indicates the interpolation point. Indicates the interpolation point The corresponding values, polynomial coefficients , , and Fault data points and interpolation conditions acquired through hardware , First derivative continuous Second derivative continuity The constraints.

[0080] 3.2, Since the original time points were sampled at equal intervals, a set of target equal-interval interpolation points was generated. :

[0081] ,

[0082] For each Within the corresponding interval Calculate using polynomial 3.1 The resampled sequence is obtained. .

[0083] In step 3, the target interpolation point set for cubic spline interpolation is equidistantly distributed. The interpolation process is calculated independently within each original data interval, ensuring high waveform fidelity after resampling and avoiding the introduction of false features due to interpolation. This process unifies waveforms of varying original lengths into a fixed length of 1024 points. For example, due to different cable lengths, the number of original sampling points n may vary between 800 and 1200, providing a standardized data format for subsequent batch processing and model input.

[0084] Step 4: Scale the original labeled fault locations to align with the interpolated fault data. (In the original sample...) Where n is the original number of samples, the index space changes after the cubic spline interpolation of the labeled fault locations is reused, and the original labels are mapped to the new index space. M is the length of the resampled data. Therefore, a linear scaling mapping is used to ensure endpoint consistency, which can be expressed as:

[0085] ,

[0086] This maintains endpoint consistency, allowing fault labels to be accurately aligned to the resampled waveform data.

[0087] Step 5: Perform zero-mean, unit-variance standardization on the interpolated and resampled data from Step 3 to improve the stability of subsequent data processing. The standardization process includes the following steps:

[0088] 5.1 Calculate the data after cubic spline interpolation resampling The mean and standard deviation can be expressed as:

[0089] ,

[0090] ,

[0091] in, Indicates the number of resampled samples. M represents the total number of data points in the resampled signal sequence. This represents the mean. It represents the standard deviation.

[0092] 5.2 Standardize each data point:

[0093] ,

[0094] in, For the standardized first Data points, Let it be a very small positive number, and here we take... This standardization process is used to prevent division by zero and numerical instability. By performing the above standardization process, the data distribution can be made to have zero mean and unit variance, thereby avoiding calculation errors caused by excessive differences in the amplitude of the original signal.

[0095] The smallest positive number used for standardization in step 5 is 1e-8. This value is sufficient to prevent numerical instability and has a negligible impact on the standardization result, ensuring numerical stability in the subsequent neural network training process. This standardization operation eliminates the influence of amplitude differences between different signal samples, concentrating the data distribution near zero, which is beneficial for gradient stability and rapid convergence of the neural network.

[0096] Step 6: For each sequence data in the dataset processed in Step 5, a sliding window is further used to segment it to generate multiple sub-data entries. This increases the number of samples and prevents the data from being too sparse, which could lead to difficulties in gradient propagation. The sliding window segmentation method includes the following steps:

[0097] 6.1. Divide the standardized sequence data M from step 5 into segments based on the window length W and step size S. Then, the starting position s of the sliding window satisfies:

[0098] ,

[0099] At each starting position s, a window data fragment is obtained:

[0100] ,

[0101] in, This represents the subsequence corresponding to the s-th sliding window.

[0102] 6.2, For the fault location before splitting Calculate its relative position label within the current window. :

[0103] ,

[0104] Among them, when the fault location When the fault falls within the current window, its relative index is calculated as the local label; when the fault location is not within the current window, the local label is recorded as -1, and an ignoring strategy is adopted in subsequent training to avoid interfering with model learning.

[0105] 6.3, Each window obtained from the segmentation generates a four-element data set:

[0106] ,

[0107] This includes subsequence data within the window. Local labels within a window The original sample number n to which the data belongs, and the starting position s of the window.

[0108] In step 6, the window length W is set to 128 and the step size S is set to 32, resulting in 75% overlap between adjacent windows. This ensures that the fault point is fully covered, and even if the fault point happens to be at the window boundary, it will be captured by the adjacent window. Furthermore, data augmentation enhances the model's generalization ability. After the original sequence is segmented by the sliding window, multiple subsequences are generated, and local labels are assigned to each subsequence.

[0109] Step 7: Construct a transmission line fault location model, which includes the following steps:

[0110] 7.1, The model input channel dimension has been expanded to facilitate batch data processing. For input... ,in For batch size, Given the window length, the channel dimension expansion can be expressed as:

[0111] ,

[0112] 7.2, performing a one-dimensional convolution operation on the data output from 7.1, can be represented as:

[0113] ,

[0114] in, The output of the first convolution of the model is in the [number]th [time]. Each batch of index samples, channel index Location index The value, =1 represents the number of input channels. For input channel index, =5 represents the size of the convolution kernel. The index of the convolution kernel. It is the first The output channel, the first The input channel, the first The weights at each convolutional kernel position, Is the input at the ? The sample, the first The first channel, the first The value at each position, The padding size ensures that the output length is the same as the input length. For the first The bias of each output channel.

[0115] 7.3, output from 7.2 Based on this, an element-wise nonlinear transformation is performed to improve the model's ability to fit complex nonlinear relationships. The nonlinear activation uses the modified linear unit (ReLU), which can be expressed as:

[0116] ,

[0117] The output of the first nonlinear transformation of the model is in the... Each batch of index samples, channel index Location index The activation value is truncated to zero by using negative values, which deactivates some results given the input, strengthens sparse representation, helps reduce overfitting, and improves the discriminative power of features and the generalization ability of the model.

[0118] 7.4, Performing a second one-dimensional convolution operation on the result after the nonlinear transformation in 7.3, can be expressed as:

[0119] ,

[0120] in, The output of the second convolution of the model is in the [number]th [time]. Each batch of index samples, channel index Location index The value, =16 is the number of input channels. For input channel index, =3 is the size of the convolution kernel. The index of the convolution kernel. It is the first The output channel, the first The input channel, the first The weights at each convolutional kernel position, Is the input at the ? The sample, the first The first channel, the first The value of the position, The padding size ensures that the output length is still the same as the input length. For the first The bias of each output channel.

[0121] 7.5, the output of 7.4 The second modified linear unit (ReLU) nonlinear transformation can be expressed as:

[0122] ,

[0123] in, The output of the second nonlinear transformation of the model is in the... Each batch of index samples, channel index Location index The activation value.

[0124] Step 7.6, based on the output of step 7.5, performs one-dimensional adaptive average pooling to ensure the output always maintains the fixed window length W from step 6. Simultaneously, averaging across intervals smooths local features and suppresses noise. The length of the input sequence is then adjusted. The mapping is a fixed output length W, for the j-th output position. Define the corresponding input range , can be represented as:

[0125] ,

[0126] Wherein, input range The length is Then the result of one-dimensional adaptive average pooling is:

[0127] ,

[0128] in, The output of the one-dimensional adaptive average pooling model is in the [missing information]. Each batch of samples, channel index Interval The value inside, The output result of the nonlinear transformation in step 7.5 is in the first... Each batch of index samples, channel index Interval The internal index is the value of t.

[0129] 7.7 To facilitate the application of linear transformations to each time point, a dimension substitution is performed, shifting the channel dimension to the last dimension. Let the tensor shape of the one-dimensional adaptive average pooling output in step 7.6 be:

[0130] ,

[0131] Where B represents the batch size, C represents the number of channels, W represents the sequence length, and the tensor components are:

[0132] ,

[0133] Tensor shape of dimension permutation for:

[0134] ,

[0135] The relationship between its components and the original tensor is as follows:

[0136] ,

[0137] 7.8. Perform a position-by-position one-dimensional linear transformation on the output of 7.7, mapping the position within each window to a vector with the same size as the window. This one-dimensional linear transformation can be expressed as:

[0138] ,

[0139] in, Each batch of samples The transformation value for each position i with respect to j, This is the weight matrix. For bias, This is the result after dimensional permutation in step 7.7.

[0140] 7.9, The three-dimensional tensor output from step 7.8 Extracting the diagonal elements along the position and output dimensions yields the position score for each window, which can be represented as:

[0141] ,

[0142] in, Represents each batch of samples In each location The score value, This means that the value is 1 when i=j, and 0 otherwise.

[0143] Step 8: Initialize the parameters of the transmission line fault location model obtained in Step 7. The initialization process can be divided into the following steps:

[0144] 8.1 To ensure input-output balance and avoid gradient vanishing, the weight parameters of the one-dimensional convolution linear transformation in steps 7.2 and 7.4 are initialized with a mean of 0 and a variance of . A normally distributed random variable, with bias parameters initialized to zero. It can be represented as:

[0145] ,

[0146] in, The size of the convolution kernel. These represent the number of input and output channels, respectively.

[0147] 8.2 To maintain variance stability under the nonlinear transformation and avoid layer-by-layer decay, the weight parameters of the modified linear unit (ReLU) nonlinear transformation in steps 7.3 and 7.5 are initialized with a mean of 0 and a variance of . The normal distribution, where Input the number of channels. This represents the size of the convolution kernel.

[0148] Step 9: In step 6, a sliding window is used to split the data into sub-data entries, increasing the number of samples and solving the problem of excessively sparse input data. However, after splitting, the window data contains a large number of fault-free locations that are forced to participate in supervision, causing invalid gradients to interfere with parameter updates during backpropagation, resulting in the optimization direction deviating from the actual target. Based on this, a cross-entropy loss function with an invalid label masking mechanism is proposed. This mechanism ensures that parameter optimization is driven only by valid supervision samples by masking invalid locations, thereby maintaining the accuracy of gradients and the stability of the optimization path during training. The cross-entropy loss function of the invalid label masking mechanism can be expressed as:

[0149] ,

[0150] in, Indicates the result of loss. This indicates that the b-th batch is at the fault location. The output score is on. This represents the total number of all sample data. For window size, Let be a natural constant, when This represents the actual fault location of the b-th sample. =-1 indicates that the b-th sample does not have a fault point in this window. The masking coefficient is used to identify whether a location is a valid supervised sample. hour, =1, when hour, =0.

[0151] Step 10: Building upon Step 9, an iterative method with an early stopping mechanism is employed. If the mean absolute error index (MAE) decreases after the first T training iterations, the model parameters are updated using the AdamW optimizer with weight decay; otherwise, model parameter updates are stopped. In this example, T is 100. This process is repeated for each training iteration. The calculation for each parameter follows these steps:

[0152] 10.1, Update the first-order moment exponential moving average , can be represented as:

[0153] ,

[0154] in, It is the first-order moment decay factor. This represents the loss gradient for the current batch.

[0155] 10.2, Update the second-order moment element-wise exponential moving average. , can be represented as:

[0156] ,

[0157] in, It is the second-order moment attenuation factor. It is the Hadamard product.

[0158] 10.3, Calculation bias correction , To compensate for the deviation caused by the initial value being zero:

[0159] ,

[0160] ,

[0161] 10.4 Calculate the adaptive step size based on the modified moment estimation :

[0162] ,

[0163] in, It is a numerically stable term.

[0164] 10.5, using a decoupling approach, weight decay and gradient updates are applied to the parameters. The parameter update can be expressed as:

[0165] ,

[0166] in, This represents the set of trainable parameter tensors from the previous training iteration t. This represents the set of trainable parameter tensors for the current training iteration t+1. The current step size, This is the weight decay coefficient. This is an indicator function; the value is 1 when the corresponding parameter needs to be decayed, and 0 otherwise.

[0167] Step 11: After the model training is completed in Step 10, in order to evaluate its fault location performance, the local position of each window prediction result of the input model is restored to the global position in the original sequence, and then compared with the actual fault point. Finally, the mean absolute error (MAE) and the location deviation accuracy index are calculated. .

[0168] Suppose there are N original samples in the test set, each sample The system is divided into K windows. The actual global fault location is in the Kth window. With prediction of global location They are respectively recorded as:

[0169] ,

[0170] The true location of the sample is obtained by averaging the values ​​within a window and then aggregating them. With predicted location :

[0171] ,

[0172] The mean absolute error (MAE) and positioning deviation accuracy indicators It can be represented as:

[0173] ,

[0174] ,

[0175] in, This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. This is the limit value for positioning deviation.

[0176] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for determining the location of faults in underground power transmission cables for smart grids, characterized in that, include The fault data acquisition and labeling steps involve acquiring fault signals from underground power transmission cables to obtain raw fault data, and labeling the fault point locations corresponding to the raw fault data. The data resampling and label matching steps involve performing cubic spline interpolation resampling on the labeled original fault data to unify the data dimensions, while linearly scaling the labeled fault point locations to achieve dimensional adaptation between the data and the fault point labels. The data preprocessing step involves standardizing the resampled data and then using a sliding window to split it into sub-data groups containing sub-sequence data and local fault labels. The deep learning model construction and feature extraction steps include: constructing a deep learning model for power transmission line fault location; performing end-to-end feature extraction on the sub-data set; and outputting the fault score for the window position. The model training and optimization steps involve training the model using a cross-entropy loss function with an invalid label masking mechanism and an optimizer with weight decay, and then determining the optimal fault location model by combining an early stopping mechanism. The fault location determination step involves processing the fault data to be determined through data resampling and label matching steps and data preprocessing steps, and then inputting it into the optimal fault location model to output the global location of the fault point. The deep learning model construction and feature extraction steps for power transmission line fault localization include the following steps: Input channel dimension expansion sub-step, for input ,in For batch size, Given the window length, the channel dimension expansion can be expressed as: ； The feature extraction sub-step extracts deep fault features by sequentially performing a first one-dimensional convolution with ReLU activation and a second one-dimensional convolution with ReLU activation. The feature optimization sub-step maps deep fault features into a feature sequence of fixed length W through one-dimensional adaptive average pooling. The dimension adaptation sub-step converts the pooled feature tensor from the dimension... Replace with ,in , where C represents the number of channels; The fault score calculation sub-step outputs the fault score for each window position through a position-by-position one-dimensional linear transformation and diagonal element extraction.

2. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the fault data acquisition and labeling steps, the acquisition of fault signals is controlled by FPGA+ARM microcontroller: FPGA undertakes high-speed sampling control at a sampling rate of 200M, real-time signal preprocessing of 32nd order FIR low-pass filtering, and FIFO data temporary storage tasks. The FIR low-pass filtering is implemented by parallel operation of the multiplier and adder resources inside the FPGA, and the signal delay after filtering is ≤20ns. The ARM microcontroller is responsible for overall system control, data storage management, human-computer interaction, and communication with the host computer, transmitting pre-processed fault data in timestamp + waveform data format.

3. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the data resampling and label matching steps, the specific implementation of cubic spline interpolation is as follows: within the interval formed by any two adjacent original data points... Inside, the interpolation function is a cubic polynomial: ， in Indicates the first Data points, Indicates the interpolation point. Indicates the interpolation point The corresponding values, polynomial coefficients , , and Fault data points acquired through hardware, and meeting the interpolation conditions. First derivative continuity Second derivative continuity The constraints.

4. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the data preprocessing and subgroup generation steps, the sliding window segmentation rules include: the starting position of the sliding window. , Where M is the total length of the standardized data, S is the sliding window step size, and W is the sliding window length; the generation rule for local fault labels is: ， in This represents the location of the fault point before splitting; the sub-data group is a quadruple. , For window subsequences, Here, n is the local fault location label, n is the original sample number, and s is the starting position of the window.

5. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the model training and optimization steps, the parameter initialization rules include: The weights for the convolution linear transform are normally distributed random variables with bias initialized to zero; the weights for the ReLU nonlinear transform are variables with a mean of 0 and a variance of 0. The normal distribution, where K represents the number of input channels, and K represents the kernel size.

6. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the model training and optimization steps, the cross-entropy loss function of the invalid label masking mechanism is: ， in, Indicates the result of loss. This indicates that the b-th batch is at the fault location. The output score is on. This represents the total number of all sample data. For window size, Let be a natural constant, when This represents the actual fault location of the b-th sample. =-1 indicates that the b-th sample does not have a fault point in this window. The masking coefficient is used to identify whether a location is a valid supervised sample. hour, =1, when hour, =0.

7. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the model training and optimization steps, the parameter update process for the AdamW optimizer with weight decay is as follows: Update the first-moment exponential moving average , can be represented as: , in, It is the first-order moment decay factor. This represents the loss gradient for the current batch. Update the second-order moment element-wise exponential moving average , can be represented as: , in, It is the second-order moment attenuation factor. The product of Hadamard; Calculation bias correction , To compensate for the deviation caused by the initial value being zero: ， ， Calculate the adaptive step size based on the modified moment estimation : ， in This represents the preset numerical stability term. By using a decoupling approach, weight decay and gradient updates are applied to the parameters. The parameter update can be expressed as: ; in, This represents the set of trainable parameter tensors from the previous training iteration t. This represents the set of trainable parameter tensors for the current training iteration t+1. The current step size, This is the weight decay coefficient. This is an indicator function; the value is 1 when the corresponding parameter needs to be decayed, and 0 otherwise.

8. The method for determining the fault location of underground power transmission cables for smart grids according to claim 1, characterized in that, In the data preprocessing and subgroup generation steps, the data after cubic spline interpolation resampling is calculated. Mean and standard deviation: ， ， in This represents the mean. Let M represent the standard deviation, and M represent the total number of data points in the resampled signal sequence. Standardize each data point: ， in, For the standardized first Data points, It represents a very small positive number.

9. A fault location determination system for underground power transmission cables for smart grids, characterized in that, The method described in any one of claims 1-8 includes: The data acquisition module is used to collect fault signals from underground power transmission cables; The data preprocessing module is used to perform data resampling and label matching, data preprocessing and subgroup generation, and complete cubic spline interpolation resampling, standardization and sliding window segmentation. The model training module is used to build a deep learning model for fault location of transmission lines and to train the model using a loss function and optimization strategy. The fault location module receives fault data to be identified, preprocesses the data, inputs it into the optimal fault location model, and outputs the global location and confidence distribution of the fault point.

Images