A deep learning-based distribution network pole erecting positioning method and system

Abstract

The application discloses a kind of based on deep learning's network distribution erecting pole positioning method and system, method includes: with the energy response spectrum matrix of previous N depth interval as sample, the deep self-recurrence prediction network is trained online;For each subsequent new depth interval, the prediction matrix of current interval is obtained by using the previous interval matrix through prediction network, the multiscale structure dissimilarity index between prediction matrix and actual matrix is calculated;When index is greater than first mutation threshold for the first time, mark initial bottoming interval and obtain reference depth, when the change rate of index in continuous multiple intervals changes from positive to negative, mark overpressure interval and obtain overpressure depth;When the difference between overpressure depth and reference depth is greater than elastic compression threshold, determine stable bottoming, the final depth of entering soil is output after weighted average of both. Realize the accurate automatic positioning of electric pole depth of entering soil, effectively improve the anti-interference ability and positioning precision of bottoming determination.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network construction monitoring technology, and in particular relates to a power distribution network pole positioning method and system based on deep learning. Background Technology

[0002] In power distribution network construction and maintenance projects, the installation of utility poles is a fundamental and crucial construction step. Precise control of the pole's embedment depth directly affects its load-bearing capacity, overturning stability, and the operational safety of the entire power distribution line. In traditional construction, determining the pole's embedment depth relies primarily on the experience of construction workers or simple depth marker observations. This makes it difficult to accurately determine in real-time whether the pole's bottom has stably touched the bearing layer during the lowering process. Inaccurate bottom-touching judgments often lead to insufficient or excessive pole embedment depth. The former causes the risk of pole collapse, while the latter results in material waste and potential damage to underground pipelines.

[0003] To address the aforementioned issues, some improvements have been proposed in existing technologies. For example, pressure or tension sensors are installed on the hoisting equipment to monitor changes in the tension of the wire rope and indirectly infer whether the pole has touched the ground. However, this approach has significant drawbacks: during the pole's descent, the wire rope tension is affected by various factors, including the pole's own weight, mud buoyancy, sidewall friction, and the mechanical vibration of the hoisting equipment. The tension change characteristics at the moment of contact with the ground are often overwhelmed by these strong interference signals, leading to a high misjudgment rate. Other solutions use ultrasonic or laser ranging to measure the pole's height above ground, but dust, mud splashes, and irregular ground conditions in the construction environment severely affect the ranging accuracy. Furthermore, this method only obtains the position of the top of the pole and cannot reflect the actual contact state between the pole's bottom and the foundation.

[0004] In recent years, deep learning technology has demonstrated powerful feature extraction and pattern recognition capabilities in the field of structural health monitoring and state recognition. However, existing deep learning solutions are mostly designed for fault diagnosis of equipment operating in a steady state, and are difficult to adapt to the dynamic construction process of pole lifting and sinking, which is unsteady, noisy, and subject to constantly changing boundary conditions. Furthermore, the vibration response of the pole at the moment of bottom contact exhibits transient and abrupt characteristics. How to effectively separate the bottom contact characteristics from the vibration signal across a continuous depth range, and further distinguish between the two different states of "initial bottom contact" and "stable bottom contact," is a pressing technical challenge in this field. Therefore, there is an urgent need for an intelligent positioning method capable of online sensing of the pole's bottom contact state, accurate determination of the stable bottom contact moment, and output of a reliable penetration depth. Summary of the Invention

[0005] This invention provides a deep learning-based method and system for positioning power distribution network poles, which solves the technical problems in the prior art where the determination of the pole's soil penetration depth relies on manual experience, the identification of the bottom contact state is easily interfered with, and the initial bottom contact and stable bottom contact cannot be distinguished.

[0006] In a first aspect, the present invention provides a deep learning-based method for locating power distribution network poles, comprising: In multiple consecutive depth intervals during pole lifting operations, a pulse excitation with a preset frequency and preset amplitude is applied to the lifting wire rope, and vibration response signals from multiple measuring points on the wire rope in each depth interval are collected to form an energy response spectrum matrix corresponding to each depth interval. The energy response spectrum matrices of the first N consecutive depth intervals are used as normal sinking samples. The energy response spectrum matrix of a certain depth interval is used as input, and the prediction of the energy response spectrum matrix of the next depth interval is used as output to train the depth autoregressive prediction network online. For each new depth interval after the Nth depth interval, the energy response spectrum matrix of the previous depth interval is input into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval is calculated. When the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, the current depth interval is marked as the initial bottoming interval, and the depth value corresponding to the initial bottoming interval is obtained as the reference depth. The excitation is applied to the next depth interval and the response is collected. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth. Calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold. If the difference is greater than the preset elastic compression threshold, it is determined that the pole has stably reached the bottom, and the final penetration depth is output after weighted averaging of the reference depth and the overpressure depth.

[0007] Secondly, the present invention provides a power distribution network pole positioning system based on deep learning, comprising: The module is configured to apply a pulse excitation with a preset frequency and preset amplitude to the hoisting wire rope in multiple consecutive depth intervals during pole hoisting operations, and to collect the vibration response signals of multiple measuring points on the wire rope in each depth interval to form an energy response spectrum matrix corresponding to each depth interval. The training module is configured to use the energy response spectrum matrices of the first N consecutive depth intervals as normal sinking samples, take the energy response spectrum matrix of a certain depth interval as input, and take the prediction of the energy response spectrum matrix of the next depth interval as output to train the depth autoregressive prediction network online. The calculation module is configured to, for each new depth interval after the Nth depth interval, input the energy response spectrum matrix of the previous depth interval into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and calculate the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval. The marking module is configured to mark the current depth interval as the initial bottoming interval when the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, and obtain the depth value corresponding to the initial bottoming interval as the reference depth, continue to apply excitation to the next depth interval and collect the response. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth. The judgment module is configured to calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold. The output module is configured to determine that the pole has stably reached the bottom if the difference is greater than the preset elastic compression threshold, and output the final penetration depth after weighted averaging of the reference depth and the overpressure depth.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the deep learning-based power distribution pole positioning method according to any embodiment of the present invention.

[0009] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the deep learning-based power distribution pole positioning method according to any embodiment of the present invention.

[0010] This application presents a deep learning-based method and system for locating power distribution network poles. It constructs an energy response spectrum matrix that sensitively reflects changes in the contact state of the pole's bottom by combining active pulse excitation with multi-point vibration response acquisition. A deep autoregressive prediction network is used to learn the signal evolution during the normal sinking phase online, enabling accurate capture of structural response abrupt changes caused by bottom contact. Furthermore, a dual-stage marking strategy for initial bottom contact and overpressure, combined with elastic compression threshold determination, effectively distinguishes between the initial contact and stable bearing states of the pole's bottom with the bearing layer, avoiding misjudgments caused by surface loose soil or temporary resistance. Finally, a weighted fusion output of the reference depth and overpressure depth achieves high-precision positioning of the penetration depth. Compared to traditional manual judgment or single-parameter measurement methods, this invention has significant advantages such as strong anti-interference capability, objective judgment criteria, and adaptability to complex geological conditions, effectively ensuring the quality and safety of power distribution network pole construction. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating a deep learning-based method for locating power distribution network poles, as provided in an embodiment of the present invention; Figure 2 A structural block diagram of a power distribution network pole positioning system based on deep learning is provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0014] Please see Figure 1 The diagram shows a flowchart of a deep learning-based method for locating power distribution network poles according to this application.

[0015] like Figure 1 As shown, the deep learning-based method for locating power distribution network poles specifically includes the following steps: Step S101: In multiple consecutive depth intervals of the pole lifting operation, a pulse excitation with a preset frequency and preset amplitude is applied to the lifting wire rope respectively, and the vibration response signals of multiple measuring points on the wire rope in each depth interval are collected to form an energy response spectrum matrix corresponding to each depth interval.

[0016] In this step, each depth interval is divided into multiple time windows of equal duration in chronological order. The length of each time window is 1.5 times the duration of the pulse excitation signal, and there is a 50% overlap between adjacent time windows. For the vibration response signal within each time window, a fast Fourier transform is performed to obtain a frequency domain amplitude sequence, and the amplitude of each frequency component in the frequency domain amplitude sequence is squared to obtain the energy spectrum vector for each time window. Arrange the energy spectrum vectors of all time windows at the same measurement point in chronological order to form a time-frequency energy matrix for the same measurement point. The rows of the time-frequency energy matrix correspond to the frequency points, and the columns correspond to the time window numbers. The time-frequency energy matrices of all measuring points within the same depth range are sequentially spliced ​​together according to the spatial position of the measuring points on the wire rope to form a three-dimensional energy response spectrum matrix. The first dimension of the energy response spectrum matrix is ​​the number of frequency points, the second dimension is the number of time windows, and the third dimension is the number of measuring points.

[0017] In one specific embodiment, the pole lifting operation involves using lifting equipment (such as a crane or derrick) to vertically lift the pole and slowly lower it into a pre-excavated pit. Throughout the lowering process, the lowering distance is divided into continuous and equally spaced depth intervals according to the vertical displacement of the pole. The length of each depth interval is pre-set according to the construction accuracy requirements; for example, every 10 centimeters of lowering is defined as one depth interval.

[0018] At the start of each depth interval, an electromagnetic exciter fixed to the wire rope applies a pulse excitation with a preset frequency and amplitude. The preset frequency is pre-calibrated based on the inherent characteristics of the wire rope-pole system, typically selected as 0.6 to 0.8 times the system's first natural frequency to ensure sufficient excitation energy is transferred to the bottom of the pole while avoiding resonance amplification. The preset amplitude is determined based on the wire rope's load-bearing capacity and construction safety requirements, ensuring that the generated vibration response signal has a signal-to-noise ratio higher than 30 dB and does not cause nonlinear vibration of the wire rope. The duration of the pulse excitation is preset to between 50 and 200 milliseconds, with the specific value depending on the wire rope's length and damping characteristics.

[0019] Multiple vibration acceleration sensors are arranged along the length of the wire rope as measuring points, with at least three measuring points, respectively located at the connection end between the wire rope and the hook, the middle of the wire rope, and the connection end between the wire rope and the pole. Vibration response signals are synchronously acquired at each measuring point throughout the entire depth range after pulse excitation. During acquisition, the sampling rate is set to at least five times the highest frequency component of the pulse excitation signal to satisfy the Nyquist sampling theorem. The acquired raw vibration response signals are first filtered by a bandpass filter to eliminate DC offset and power frequency interference, then amplified by an amplifier before being sent to a data acquisition card for analog-to-digital conversion.

[0020] Subsequently, a time-frequency energy matrix was constructed for the vibration response signals collected at each measuring point within each depth interval. Specifically, the vibration response signals within each depth interval were divided into multiple time windows of equal duration according to their temporal order. The length of each time window was set to 1.5 times the duration of the pulse excitation signal; for example, if the pulse duration was 100 milliseconds, the time window length was set to 150 milliseconds to ensure that each window fully encompassed the main vibration decay process after the pulse excitation. Adjacent time windows had a 50% overlap rate, meaning that the starting position of each new window was shifted backward by half a window length relative to the previous window. This overlap setting helps improve the continuity and smoothness of the time-frequency analysis and avoids information loss due to window boundary effects.

[0021] For the vibration response signal within each time window, a Fast Fourier Transform (FFT) is performed to convert the time-domain signal into a frequency-domain amplitude sequence. Each element in the frequency-domain amplitude sequence corresponds to the amplitude of a discrete frequency component. The amplitude of each frequency component in the frequency-domain amplitude sequence is then squared to obtain the energy value corresponding to each frequency component. The vector composed of these energy values ​​is the energy spectrum vector for that time window. Each element in the energy spectrum vector represents the vibration energy contribution of the corresponding frequency component within that time window.

[0022] After completing the above operations, for the same measuring point, the energy spectrum vectors of all time windows (including all windows resulting from overlap) at that measuring point are arranged sequentially according to time, forming a two-dimensional time-frequency energy matrix. The row indices of this matrix correspond to frequency points (i.e., frequencies arranged from low to high), and the column indices correspond to time window numbers (i.e., times arranged from earliest to latest). Each element in the matrix represents the magnitude of vibrational energy at a specific frequency within a specific time window. Therefore, this time-frequency energy matrix comprehensively reflects the dynamic distribution characteristics of vibrational energy at the measuring point within the current depth range, considering both frequency and time dimensions.

[0023] Finally, the time-frequency energy matrices of all measuring points within the same depth range are sequentially pieced together according to their spatial position on the wire rope (e.g., from the hook end to the pole end) to form a three-dimensional energy response spectrum matrix. The first dimension of this three-dimensional matrix represents the number of frequency points, the second the number of time windows, and the third the number of measuring points. Each element (f, t, s) in the matrix represents the vibration energy value corresponding to the frequency component f of the s-th measuring point within the t-th time window. This three-dimensional energy response spectrum matrix comprehensively characterizes the energy distribution of the vibration response signal in the frequency, time, and spatial domains within the current depth range, serving as the basic input unit for subsequent deep learning prediction networks.

[0024] Through the above processing, each depth interval corresponds to a unique three-dimensional energy response spectrum matrix. This matrix not only contains the frequency composition information of the vibration signal, but also contains the dynamic characteristics of energy decay over time and the spatial distribution differences along the steel wire rope, providing a rich and structured data foundation for the bottoming state identification based on deep learning in subsequent steps.

[0025] This step, by actively applying preset pulse excitation and combining it with synchronous acquisition from multiple measurement points, effectively overcomes the shortcomings of passive monitoring methods, such as weak signal characteristics and susceptibility to environmental noise. The energy spectrum extraction method, combining overlapping time windows with short-time Fourier transform, significantly improves temporal positioning accuracy while maintaining frequency resolution, allowing for the complete preservation of time-varying details of vibration energy. Arranging the time-frequency energy vectors of the same measurement point in chronological order to form a time-frequency energy matrix achieves a compact representation of the dynamic evolution of the vibration response. Furthermore, splicing the time-frequency energy matrices of multiple measurement points along the spatial dimension to form a three-dimensional energy response spectrum matrix further introduces the spatial distribution information of vibration propagation along the wire rope propagation path, enabling subsequent analysis to utilize complementary features from the frequency, time, and spatial dimensions simultaneously. This three-dimensional matrix structure retains the physical meaning of the original signal and possesses a regularized tensor form, directly adapting to the input data format requirements of deep learning networks. This provides a high-quality data foundation with complete information and a clear structure for subsequent anomaly detection based on deep autoregressive prediction, significantly improving the comprehensiveness and robustness of bottoming-out feature extraction.

[0026] Step S102: The energy response spectrum matrices of the first N depth intervals are continuously acquired as normal sinking samples. The energy response spectrum matrix of a certain depth interval is used as input, and the prediction of the energy response spectrum matrix of the next depth interval is used as output to train the depth autoregressive prediction network online.

[0027] In this step, the energy response spectrum matrices of the first N consecutive depth intervals are numbered E1, E2, ..., E in ascending order of depth. N For any positive integer i, E iAs input samples, with E i+1 As supervisory labels, construct N-1 training sample pairs, where 1≤i≤N-1; A deep autoregressive prediction network is constructed, comprising an encoding module and a decoding module. The encoding module is used to compress the input energy response spectrum matrix into a low-dimensional feature vector, and the decoding module is used to restore the low-dimensional feature vector into a prediction matrix with the same dimension as the input. The mean squared error loss function is used to pre-train the deep autoregressive prediction network offline using the N-1 training samples until the loss function value is less than the preset first loss threshold or the number of iterations reaches the preset maximum number of iterations, thus obtaining the initial prediction network. The actual energy response spectrum matrix E for each newly added depth interval is subsequently obtained. j Then, E j E with the previous depth interval j-1 New training sample pairs are formed, and the parameters of the decoding module of the initial prediction network are fine-tuned online using the single-step gradient descent method, while the parameters of the encoding module remain unchanged, where j>N.

[0028] In one specific embodiment, the training of the deep autoregressive prediction network is divided into two stages: an offline pre-training stage based on the first N depth interval samples, and an online update stage where parameters are fine-tuned using newly added depth interval samples as the deployment process progresses. The two stages are interconnected and together constitute a complete online training mechanism.

[0029] In the initial stage of pole lowering, the bottom of the pole has not yet touched the bearing layer at the bottom of the pit, and the pole is in a free-sinking state in the mud or loose backfill soil layer. At this time, the vibration response of the pole-wire rope system mainly depends on the combined effect of the pole's own weight, mud buoyancy, and sidewall friction. There is no abrupt reflection or energy dissipation caused by bottom contact. Therefore, the energy response spectrum matrix of each depth interval has good consistency and continuity in statistical distribution. In engineering practice, depending on the depth of the pit and the complexity of the geological conditions, the number N of the first N depth intervals is preferably set to an integer between 20 and 50. The principle for choosing the value of N is: while ensuring that the sample size is sufficient to train the initial network, it is ensured that the lowering stroke corresponding to these N depth intervals is still in the free-sinking stage before the bottom of the pole contacts the bearing layer. The selection of N can be dynamically judged by real-time monitoring of the point-by-point difference between the energy response spectrum matrices of the first few depth intervals—if the difference of multiple consecutive depth intervals remains at a stable low level, it indicates that it is still in the normal sinking stage, and it can continue to accumulate until N reaches the preset lower limit.

[0030] It should be noted that the principle for determining the number N of the first N depth intervals is as follows: First, based on the design depth of the foundation pit and the geological conditions, the length of the free sinking section is estimated. The lower limit of N is 20, and the upper limit is 50. The dynamic judgment method is as follows: calculate the point-by-point Euclidean distance between the energy response spectrum matrices of the acquired continuous depth intervals in real time, and calculate their average value. When the fluctuation range of the average distance of the most recent 5 depth intervals relative to the average distance of the previous 5 intervals is less than 10%, and this state continues for at least 3 intervals, it is determined to be in the "low level of difference". At this time, the accumulation of N can be stopped, and the number of intervals currently acquired is used as N. If the low level condition is still not met after more than 50 intervals, then N is forced to be 50.

[0031] The energy response spectrum matrices of the first N consecutive depth intervals are numbered E1, E2, ..., E in ascending order of depth. N For any positive integer i (1≤i≤N-1), the energy response spectrum matrix E of the i-th depth interval is... i As input samples, the energy response spectral matrix E of the next adjacent depth interval is used. i+1 As supervisory labels, N-1 training sample pairs are constructed. This autoregressive sample pair design enables the network to learn by predicting the energy response spectrum of the next adjacent depth interval, given the energy response spectrum of the current depth interval, thereby characterizing the evolution of the energy response spectrum with increasing depth during the normal sinking phase.

[0032] A deep autoregressive prediction network is constructed, employing an encoder-decoder symmetric architecture. The encoding module consists of multiple alternating stacks of 3D convolutional layers and pooling layers. The kernel size of the 3D convolutional layers is set to, for example, 3×3×3, with a stride of 2, allowing each convolutional operation to simultaneously extract local features in the frequency, time, and measurement dimensions. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function layer. The batch normalization layer accelerates convergence and mitigates internal covariate bias, while the ReLU activation function layer introduces nonlinear transformation capabilities into the network. The pooling layers employ 3D max pooling with a pooling window size of 2×2×2 and a stride of 2, progressively reducing the spatial resolution of the feature maps, expanding the receptive field, and reducing computational cost. Through the alternating stacking of multiple convolutional and pooling layers, the encoding module progressively compresses the input 3D energy response spectrum matrix into a low-dimensional feature vector. The dimension of this low-dimensional feature vector is much smaller than the total number of elements in the original input matrix. Its specific dimension is determined based on the size of the input matrix and the network depth. For example, when the input matrix has 64 frequency points, 32 time windows, and 4 measurement points, the dimension of the feature vector output by the encoding module can be set to 256. This feature vector is considered to capture the most essential and discriminative time-frequency-space joint features in the input energy response spectrum matrix, eliminating redundant information and noise components.

[0033] The decoding module is symmetrical to the encoding module, consisting of multiple alternating stacked 3D deconvolutional layers and upsampling layers. The deconvolutional layers (also called transposed convolutional layers) progressively restore the low-dimensional feature vectors to a 3D matrix with the original dimensions. The upsampling layers use nearest-neighbor interpolation or bilinear interpolation to progressively enlarge the feature map to the target size in three dimensions. The final layer of the decoding module is a 3D convolutional layer with the sigmoid activation function, mapping the output value to the same dynamic range as the input energy response spectrum matrix. The output of the decoding module is the predicted energy response spectrum matrix, whose dimensions are exactly the same as the input energy response spectrum matrix for the current depth interval, allowing for element-wise loss calculation between the predicted and actual matrices.

[0034] The mean squared error loss function is used as the objective function for network optimization. The mean squared error loss function calculates the average of the squared differences between the predicted and actual matrices element-wise, effectively measuring the degree of difference in the overall numerical distribution of the two matrices. Using the aforementioned N-1 training sample pairs, the deep autoregressive prediction network is pre-trained offline using a mini-batch stochastic gradient descent optimization algorithm. In each iteration, a small batch of samples (e.g., batch size of 8 or 16) is randomly selected from one training sample pair, the loss function value is calculated, and backpropagation updates all network parameters. During pre-training, when the loss function value is less than a preset first loss threshold, it indicates that the network's prediction accuracy has reached an acceptable level for engineering purposes, and training can be terminated early; or when the number of iterations reaches a preset maximum number of iterations (e.g., 2000 to 5000), training is stopped regardless of whether the loss function value reaches the threshold, to prevent excessively long training time from affecting the construction progress. The network obtained after pre-training is called the initial prediction network, which has initially grasped the adjacent step evolution law of the energy response spectrum during the normal sinking stage.

[0035] After the pole is lowered to a depth exceeding N depth intervals, the actual energy response spectrum matrix E is obtained for each newly added depth interval. j (where j>N), thus triggering the online update mechanism. E j E of the previous depth interval j-1 To form new training sample pairs, the input sample is E. j-1 The supervision label is E. jA single-step gradient descent method is used to update the parameters of the decoding module of the initial prediction network once. This involves calculating the loss function value for the new sample pair and updating only the parameters of each layer in the decoding module via backpropagation, while all parameters of the encoding module remain frozen. The rationale behind this design is that the encoding module has learned a relatively stable feature extraction method during the offline pre-training phase with the first N samples. Frequent changes to the encoding module parameters during subsequent online updates could cause the learned general features to drift, thus reducing prediction stability. Fine-tuning only the decoding module to adapt to the latest signal evolution trends can both track the drift in response features caused by changes in geological conditions and maintain the overall stability of the network. The learning rate for online fine-tuning is set to one-tenth of the initial learning rate during offline pre-training, allowing for precise adjustments with small step sizes to avoid disrupting the converged parameter distribution due to excessively large single-step updates.

[0036] Through the training strategy of combining offline pre-training with online fine-tuning, the deep autoregressive prediction network can continuously track the slow changes in vibration response characteristics during the lowering of the pole, providing an accurate and reliable baseline reference for bottoming detection based on prediction bias in subsequent steps.

[0037] Step S103: For each new depth interval after the Nth depth interval, input the energy response spectrum matrix of the previous depth interval into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and calculate the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval.

[0038] In this step, the element-wise difference matrix between the predicted energy response spectrum matrix and the actual energy response spectrum matrix is ​​calculated, and all non-zero elements in the difference matrix are divided into multiple amplitude levels according to their numerical values. The proportion of the number of elements in each amplitude level to the total number of all non-zero elements is counted to obtain the sparsity characteristic value of the difference distribution. The element-wise difference matrix is ​​marginalized and projected along the frequency dimension and the measurement point dimension respectively to obtain the first difference marginal vector in the frequency dimension and the second difference marginal vector in the measurement point dimension. The second central moments of the first difference marginal vector and the second difference marginal vector are calculated respectively and defined as the frequency marginal dispersion and the measurement point marginal dispersion. Calculate the first sign consistency ratio of each row element and the second sign consistency ratio of each column element of the element-wise difference matrix respectively. Calculate the average of all the first sign consistency ratios and the second sign consistency ratios to obtain the row sign consistency index and the column sign consistency index. Then, calculate the weighted average of the row sign consistency index and the column sign consistency index to obtain the total sign consistency index of the difference. Calculate the first frequency energy centroid of the predicted energy response spectrum matrix in the frequency dimension and the second frequency energy centroid of the actual energy response spectrum matrix in the frequency dimension, respectively, and calculate the absolute offset between the first frequency energy centroid and the second frequency energy centroid, which is defined as the frequency centroid offset. Calculate the first measurement point energy centroid of the predicted energy response spectrum matrix in the measurement point dimension and the second measurement point energy centroid of the actual energy response spectrum matrix in the measurement point dimension, respectively, and calculate the absolute offset between the first measurement point energy centroid and the second measurement point energy centroid, which is defined as the measurement point centroid offset. The multi-scale structural dissimilarity index is obtained by normalizing and weighting the sparsity characteristic value of the difference distribution, the frequency marginal dispersion, the measurement point marginal dispersion, the total index of difference sign consistency, the frequency centroid offset, and the measurement point centroid offset.

[0039] In a specific implementation, the sparsity eigenvalues ​​of the difference distribution are calculated: The element-by-element difference matrix between the predicted energy response spectrum matrix and the actual energy response spectrum matrix is ​​calculated by subtracting the element values ​​at the same positions (same frequency point, same time window, same measurement point) in the two matrices one by one. Each element in this difference matrix reflects the degree of deviation between the predicted and actual values ​​at the corresponding positions.

[0040] All non-zero elements in the difference matrix are divided into multiple amplitude levels according to their numerical values. For example, the numerical range of non-zero elements can be divided into five levels from the minimum to the maximum value, corresponding to extremely low amplitude, low amplitude, medium amplitude, high amplitude, and extremely high amplitude, respectively. The number of elements contained in each amplitude level is counted, and the proportion of the number of elements in each level to the total number of all non-zero elements is calculated. The proportions of each level are arranged in ascending order of amplitude level to form a sparse feature vector of the difference distribution, and feature values ​​representing the degree of distribution concentration are extracted from it. This sparse feature value reflects the clustering characteristics of the prediction deviation in the amplitude distribution: if the deviation is mainly concentrated in a few high amplitude levels, it indicates that the prediction deviation has local bursts, which may correspond to a sudden change in response caused by bottoming out; if the deviation is evenly distributed in all levels, it indicates that the deviation may come from random noise or model drift.

[0041] Calculate the frequency marginal dispersion and the measurement point marginal dispersion: The element-wise difference matrix is ​​marginally projected along the frequency dimension. This involves fixing the measurement point and time dimensions within the difference matrix, and summing or averaging all difference elements at the same frequency point to obtain the first marginal difference vector along the frequency dimension. Each element of this vector corresponds to a frequency component, representing the cumulative deviation of that frequency component across all measurement points and all time windows. Similarly, the element-wise difference matrix is ​​marginally projected along the measurement point dimension. This involves fixing the frequency and time dimensions within the difference matrix, and summing or averaging all difference elements at the same measurement point to obtain the second marginal difference vector along the measurement point dimension. Each element of this vector corresponds to a measurement point location, representing the cumulative deviation of that measurement point across all frequency components and all time windows.

[0042] Calculate the second central moments of the first and second difference marginal vectors respectively. Specifically, calculate the average of all elements in the vector, then calculate the average of the squares of the differences between each element and this average. The result is the second central moment. The second central moment of the first difference marginal vector is defined as the frequency marginal dispersion; a larger value indicates a more uneven distribution of the prediction error across different frequencies. The second central moment of the second difference marginal vector is defined as the measurement point marginal dispersion; a larger value indicates a more uneven distribution of the prediction error across different measurement points.

[0043] Calculate the overall sign consistency index for the difference: Analyze each row of the element-wise difference matrix (all elements at the same frequency point), count the signs (positive, negative, or zero) of all elements in that row, count the number of positive signs pos and the number of negative signs neg, and take max(pos, neg) as the number of elements with the same sign. Divide this by the total number of non-zero elements in that row to obtain the first sign consistency ratio for that row. The higher the sign consistency ratio, the more likely the biased elements in that row are all positive or all negative, meaning there is a systematic unidirectional shift in the predicted value relative to the actual value. The lower the sign consistency ratio, the more likely the biased elements in that row are alternating between positive and negative, mainly due to random bias. After calculating the first sign consistency ratio for all rows in the difference matrix, average the first sign consistency ratios of all rows to obtain the row sign consistency index.

[0044] Similarly, analyze each column (all elements under the same measurement point) of the element-wise difference matrix, calculate the sign consistency ratio of the non-zero elements in the column as the second sign consistency ratio, and average the second sign consistency ratios of all columns to obtain the column sign consistency index.

[0045] The row and column sign consistency indices are weighted and averaged to obtain the overall difference sign consistency index. The weights of the row and column sign consistency indices can be preset based on construction experience, for example, set to equal weight (0.5 each), or assigned different weights according to the relative importance of the number of measurement points and frequency points. This overall index reflects the systematic nature of the prediction deviation in the spatial structure.

[0046] Calculate the frequency centroid offset: The energy centroids of the predicted energy response spectrum matrix and the actual energy response spectrum matrix in the frequency dimension are calculated separately. Specifically, for the predicted energy response spectrum matrix, the frequency value is used as a variable, and the total energy at each frequency point in the matrix (i.e., the sum of the energy at all measurement points and over all time windows at that frequency) is used as a weight to calculate the weighted average, obtaining the first frequency energy centroid. This centroid value represents the concentrated location of vibration energy in the spectrum of the predicted matrix. Similarly, the same calculation is performed on the actual energy response spectrum matrix to obtain the second frequency energy centroid.

[0047] The absolute difference between the first frequency energy centroid and the second frequency energy centroid is calculated; this absolute difference is the frequency centroid offset. The larger the offset, the more significant the difference between the predicted matrix and the actual matrix in the dominant frequency distribution of vibration energy. This may correspond to a change in the system's natural frequency or boundary conditions after the bottom of the pole contacts the bearing layer, causing a shift in the frequency distribution centroid of vibration energy.

[0048] Calculate the centroid offset of the measuring point: The energy centroids of the predicted and actual energy response spectrum matrices along the measurement point dimension are calculated separately. Specifically, for the predicted energy response spectrum matrix, the spatial coordinates of the measurement points are used as variables (the spatial position of the measurement point along the wire rope from the hook end to the pole end can be quantified as normalized coordinate values ​​between 0 and 1). The total energy at each measurement point (i.e., the sum of energy at all frequency points and over all time windows at that measurement point) is used as the weight to calculate the weighted average, obtaining the energy centroid of the first measurement point. This centroid value represents the concentrated distribution of vibration energy along the wire rope at each measurement point in the predicted matrix. Similarly, the same calculation is performed on the actual energy response spectrum matrix to obtain the energy centroid of the second measurement point.

[0049] Calculate the absolute difference between the energy centroid of the first measuring point and the energy centroid of the second measuring point. This absolute difference is the centroid offset of the measuring point. The larger the offset, the more significant the difference between the predicted matrix and the actual matrix in the propagation distribution of vibration energy along the wire rope. This may correspond to changes in the reflection and transmission characteristics of the vibration wave at the pole-soil interface after bottoming out, causing a displacement of the centroid of energy distribution along the wire rope.

[0050] Weighted fusion yields a multi-scale structural dissimilarity index: The sparsity eigenvalues, frequency marginal dispersion, measurement point marginal dispersion, overall index of difference sign consistency, frequency centroid offset, and measurement point centroid offset of the aforementioned difference distribution are all normalized. The purpose of normalization is to map each index to the same numerical range (e.g., the interval between 0 and 1), eliminating fusion bias caused by differences in the original dimensions and numerical ranges of each index. The normalization method can employ min-maximum standardization, where for each index, the minimum value in its historical data during the normal sinking phase (within the first N depth intervals) is used as the lower bound, and the maximum value as the upper bound, mapping the current value to the interval between 0 and 1.

[0051] After normalization, the six indicators are weighted and summed according to preset weighting coefficients. The result is the multi-scale structural dissimilarity index. The values ​​of each weighting coefficient can be pre-calibrated based on the sensitivity of each indicator to the bottoming response under different geological conditions and construction scenarios. For example, in hard foundations, the weight of frequency centroid offset can be appropriately increased because the influence of the hard bearing layer on the vibration frequency is more significant; in soft foundations, the weight of measuring point centroid offset can be appropriately increased because soft soil is more sensitive to the spatial distribution of vibration energy absorption and reflection characteristics. Once the weighting coefficients are determined, they are kept fixed during construction to ensure consistency of the judgment criteria.

[0052] When the pole is in the normal sinking phase, the difference between the predicted matrix and the actual matrix mainly consists of random noise and small model errors. All six indices remain at low levels, and the fused multi-scale structural dissimilarity index is relatively small. When the bottom of the pole begins to contact the bearing layer, the vibration response undergoes a substantial abrupt change. The predicted matrix can no longer accurately predict the current response based on the normal sinking pattern. Some or all of the six indices show a significant increase, and the fused index also increases sharply, thus enabling sensitive detection of the bottoming-out event.

[0053] The weight coefficients of the above six sub-indices (sparseness eigenvalue, frequency marginal dispersion, measuring point marginal dispersion, total index of difference sign consistency, frequency centroid offset, and measuring point centroid offset) were determined using the Analytic Hierarchy Process (AHP) combined with on-site calibration tests. Specifically, before construction, a test site with the same geological conditions as the construction site was selected, and no fewer than three simulated pole-erecting tests were conducted, recording the changes in each sub-indicator during the normal settlement stage and the bottoming-out stage. With the bottoming-out identification accuracy as the optimization objective, a genetic algorithm or grid search method was used to search for the optimal weight combination within the [0,1] interval. As a typical example, in general clay foundations, the weights can be set as follows: sparseness eigenvalue 0.10, frequency marginal dispersion 0.15, measuring point marginal dispersion 0.15, total index of difference sign consistency 0.20, frequency centroid offset 0.25, and measuring point centroid offset 0.15, with a total of 1.0. In hard foundations (such as gravelly soil), the weight of frequency centroid offset can be increased to 0.30–0.35, and the centroid offset of the measuring point can be reduced accordingly; in soft foundations, the weight of measuring point centroid offset can be increased to 0.25–0.30. The specific values ​​of the weighting coefficients can be fine-tuned based on field tests, but once determined, they remain unchanged during the construction of a single pole.

[0054] Step S104: When the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, the current depth interval is marked as the initial bottoming interval, and the depth value corresponding to the initial bottoming interval is obtained as the reference depth. The excitation is applied to the next depth interval and the response is collected. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth.

[0055] In one specific embodiment, as the pole is continuously lowered, the system synchronously updates the multi-scale structural dissimilarity index for each depth interval as new depth intervals are acquired. The system presets a first mutation threshold, which distinguishes between index fluctuations during normal sinking and index mutations caused by bottoming out. The first mutation threshold is set as follows: based on the multi-scale structural dissimilarity index of the previous N normal sinking depth intervals, its average value and standard deviation are calculated. The first mutation threshold is set as the average value plus a preset multiple (e.g., 3 to 5 times) of the standard deviation, making the probability of the index exceeding this threshold under normal sinking conditions extremely low. When the multi-scale structural dissimilarity index of a certain depth interval first exceeds the first mutation threshold, it means that a significant and non-random mutation has occurred in the vibration response mode within that depth interval, and the system determines that the bottom of the pole has made substantial contact with the bearing layer at the bottom of the pit for the first time within that depth interval.

[0056] The current depth range is marked as the initial contact range, and the corresponding depth value is obtained and recorded as the reference depth. This reference depth marks the starting point where the bottom of the pole begins to make physical contact with the foundation soil, and serves as the basic reference point for subsequent calculations of soil compression deformation.

[0057] After marking the initial contact zone, the system does not immediately determine that the pole has stably contacted the bottom. Instead, it continues the lowering of the hoisting equipment and performs the same operations as described above for each subsequent depth zone: applying pulse excitation, acquiring vibration response signals, constructing an energy response spectrum matrix, obtaining a prediction matrix through a depth autoregressive prediction network, and calculating a multi-scale structural dissimilarity index. The core monitoring objective at this stage is the trend of the multi-scale structural dissimilarity index as a function of the depth zone number.

[0058] For each new depth interval after the initial bottoming, the system calculates the rate of change of the multi-scale structural dissimilarity index across multiple consecutive depth intervals. The specific calculation method for the rate of change is as follows: taking the current depth interval as the endpoint, take M consecutive depth intervals including the current interval (M is a preset positive integer, for example, M takes values ​​from 3 to 5), and perform a linear fit on the multi-scale structural dissimilarity index of these M depth intervals as the depth interval number changes. The slope obtained from the fit is the rate of change at the current depth interval. This rate of change reflects the upward or downward trend of the multi-scale structural dissimilarity index within a local depth range: a positive rate of change indicates that the index continues to rise with increasing depth, meaning the compression of the pole on the bearing layer is still intensifying, and the contact state has not yet stabilized; a negative rate of change indicates that the index begins to fall with increasing depth, meaning the compression of the pole on the bearing layer has reached a certain level, the contact state tends to stabilize, and the vibration response mode tends to return to a stable state.

[0059] When the rate of change of the multi-scale structural dissimilarity index changes from positive to negative in multiple consecutive depth intervals, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth.

[0060] It is important to note that the determination of a "change in rate of change from positive to negative" requires certain sustained conditions to avoid misjudgments triggered by accidental negative changes in the rate of change due to single-point noise fluctuations. Specifically, the system requires that after the rate of change changes from positive to negative, it must remain negative for at least two consecutive depth intervals before a substantial change from positive to negative can be considered to have occurred. Only when this sustained condition is met is the depth interval that first fulfills the condition marked as an overpressure interval. This design effectively eliminates instantaneous reversals in the rate of change caused by random signal fluctuations, ensuring the reliability of overpressure interval marking.

[0061] Overcompression depth represents the critical position where the compression process of the pole's base against the bearing layer transitions from "continuous compression and increasingly intense response" to "compression nearing completion and the response gradually stabilizing." The difference in travel between the reference depth and the overcompression depth represents the cumulative compressive deformation of the pole's base against the bearing layer. This deformation includes both elastic and plastic compression of the soil and serves as the direct basis for determining whether the pole has reached a stable bottoming state.

[0062] In summary, by employing a dual-stage marking strategy of initial contact and overpressure, a refined, phased identification of the evolution of the contact state between the pole's bottom and the bearing layer is achieved. This effectively solves the problem in existing technologies where a single contact judgment cannot distinguish between the two different physical states of "initial contact" and "stable bearing." In the initial contact stage, marking is based on the first time the multi-scale structural dissimilarity index exceeds the mutation threshold. This approach can capture the moment of initial contact between the pole's bottom and the bearing layer with extremely high sensitivity. Furthermore, this judgment does not depend on the specific vibration amplitude but is based on structural changes in prediction biases using deep learning, exhibiting strong anti-interference capabilities and wide adaptability. In the overpressure stage, the overpressure interval is marked by monitoring the turning point where the rate of change of the multi-scale structural dissimilarity index changes from positive to negative. This physically aligns with the evolutionary law of soil from unstable to stable during compression—a continuous increase in the index corresponds to the stage of continuous soil compression and constantly changing response, while a change in the index from increasing to decreasing corresponds to the critical state where soil compression is nearing completion and the system boundary conditions are stabilizing. The combination of dual-stage markers also allows the subsequent calculation of the soil penetration depth to simultaneously include two key reference points: the bottoming-out starting position (baseline depth) and the critical position of stable bearing capacity (overpressure depth), providing a data foundation for the high-precision fusion output of the final soil penetration depth.

[0063] Step S105: Calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than the preset elastic compression threshold.

[0064] In one specific embodiment, if the difference is not greater than a preset elastic compression threshold, it is determined that the pole has not stably touched the bottom, and the multi-scale structural dissimilarity index of the subsequent newly added depth interval is determined based on the actual energy response spectrum matrix of the subsequent newly added depth interval. The new baseline depth and new overpressure depth are determined based on the multi-scale structural dissimilarity index of the subsequently added depth range; Calculate the new difference between the new overpressure depth and the new reference depth, and determine whether the new difference is greater than the preset elastic compression threshold.

[0065] The elastic compression threshold is determined by calculation based on soil mechanics compression theory according to the bearing soil type and the cross-sectional area of ​​the pole base, or calibrated through field static cone penetration tests. The calculation formula is: δ th =α·q u ·A / (Es ·B), where δ th q is the elastic compression threshold. u The ultimate bearing capacity of the bearing stratum is given by A, where A is the cross-sectional area of ​​the pole's base, and E is the cross-sectional area of ​​the pole's base. s Where is the soil compression modulus, B is the pole diameter, and α is an empirical reduction factor with a value of 0.2 to 0.5.

[0066] Step S106: If the difference is greater than the preset elastic compression threshold, it is determined that the pole has stably reached the bottom. The final penetration depth is output after weighted averaging of the reference depth and the overpressure depth.

[0067] In this step, the weights of the reference depth and overcompression depth are determined based on the soil rebound characteristics. Specifically, in soft soil with a large rebound amount, the overcompression depth better reflects the final stable bottoming position, so the weight of the overcompression depth should be 0.6 to 0.7, and the weight of the reference depth should be 0.3 to 0.4. In hard soil with a small rebound amount, the weights of the two can be equal or the weight of the reference depth can be slightly larger. As a typical value, for moderately compressible soil, the weight of the reference depth is 0.45, and the weight of the overcompression depth is 0.55.

[0068] In summary, the method of this application applies pulse excitation to the wire rope and collects the vibration response at multiple measurement points in multiple consecutive depth intervals during pole lifting operations, constructing an energy response spectrum matrix for each depth interval. Using the energy response spectrum matrices of the previous N depth intervals as samples, an online depth autoregressive prediction network is trained. For each newly added depth interval, the prediction matrix of the current interval is obtained through the prediction network using the matrix of the previous interval, and the multi-scale structural dissimilarity index between the prediction matrix and the actual matrix is ​​calculated. When the index is greater than the first mutation threshold for the first time, the initial contact interval is marked and the reference depth is obtained. When the rate of change of the index in multiple consecutive intervals changes from positive to negative, the overpressure interval is marked and the overpressure depth is obtained. When the difference between the overpressure depth and the reference depth is greater than the elastic compression threshold, stable contact is determined, and the two are weighted averaged to output the final penetration depth. This achieves accurate and automatic positioning of the pole penetration depth, effectively improving the anti-interference capability and positioning accuracy of contact determination.

[0069] Please see Figure 2 The diagram shows a structural block diagram of a deep learning-based power distribution pole positioning system according to this application.

[0070] like Figure 2 As shown, the power distribution pole positioning system 200 includes a component module 210, a training module 220, a calculation module 230, a marking module 240, a judgment module 250, and an output module 260.

[0071] The module 210 is configured to apply a pulse excitation of a preset frequency and preset amplitude to the hoisting wire rope in multiple consecutive depth intervals during pole hoisting operations, and to collect vibration response signals from multiple measuring points on the wire rope in each depth interval to form an energy response spectrum matrix corresponding to each depth interval. The training module 220 is configured to use the energy response spectrum matrices of the first N consecutively acquired depth intervals as normal sinking samples, take the energy response spectrum matrix of a certain depth interval as input, and use the prediction of the energy response spectrum matrix of the next depth interval as output to train an online depth autoregressive prediction network. The calculation module 230 is configured to, for each new depth interval after the Nth depth interval, input the energy response spectrum matrix of the previous depth interval into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and calculate the relationship between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval. A multi-scale structural dissimilarity index; a marking module 240, configured to mark the current depth interval as the initial bottoming interval when the multi-scale structural dissimilarity index is greater than a preset first mutation threshold for the first time, and obtain the depth value corresponding to the initial bottoming interval as the reference depth, continue to apply excitation to the next depth interval and collect the response, and when the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, mark the depth interval that first meets the condition as the overpressure interval, and record the depth value corresponding to the overpressure interval as the overpressure depth; a judgment module 250, configured to calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold; an output module 260, configured to determine that the pole has stably bottomed out if the difference is greater than the preset elastic compression threshold, and output the final soil penetration depth after weighted averaging of the reference depth and the overpressure depth.

[0072] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0073] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the deep learning-based power distribution pole positioning method in any of the above method embodiments. In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows: In multiple consecutive depth intervals during pole lifting operations, a pulse excitation with a preset frequency and preset amplitude is applied to the lifting wire rope, and vibration response signals from multiple measuring points on the wire rope in each depth interval are collected to form an energy response spectrum matrix corresponding to each depth interval. The energy response spectrum matrices of the first N consecutive depth intervals are used as normal sinking samples. The energy response spectrum matrix of a certain depth interval is used as input, and the prediction of the energy response spectrum matrix of the next depth interval is used as output to train the depth autoregressive prediction network online. For each new depth interval after the Nth depth interval, the energy response spectrum matrix of the previous depth interval is input into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval is calculated. When the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, the current depth interval is marked as the initial bottoming interval, and the depth value corresponding to the initial bottoming interval is obtained as the reference depth. The excitation is applied to the next depth interval and the response is collected. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth. Calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold. If the difference is greater than the preset elastic compression threshold, it is determined that the pole has stably reached the bottom, and the final penetration depth is output after weighted averaging of the reference depth and the overpressure depth.

[0074] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the deep learning-based power distribution pole positioning system, etc. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, and these remote memories may be connected to the deep learning-based power distribution pole positioning system via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0075] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the deep learning-based power distribution pole positioning method described in the above embodiment. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the deep learning-based power distribution pole positioning system. The output device 340 may include a display screen or other display device.

[0076] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0077] In one implementation, the above-described electronic device is applied to a deep learning-based power distribution pole positioning system for a client, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: In multiple consecutive depth intervals during pole lifting operations, a pulse excitation with a preset frequency and preset amplitude is applied to the lifting wire rope, and vibration response signals from multiple measuring points on the wire rope in each depth interval are collected to form an energy response spectrum matrix corresponding to each depth interval. The energy response spectrum matrices of the first N consecutive depth intervals are used as normal sinking samples. The energy response spectrum matrix of a certain depth interval is used as input, and the prediction of the energy response spectrum matrix of the next depth interval is used as output to train the depth autoregressive prediction network online. For each new depth interval after the Nth depth interval, the energy response spectrum matrix of the previous depth interval is input into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval is calculated. When the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, the current depth interval is marked as the initial bottoming interval, and the depth value corresponding to the initial bottoming interval is obtained as the reference depth. The excitation is applied to the next depth interval and the response is collected. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth. Calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold. If the difference is greater than the preset elastic compression threshold, it is determined that the pole has stably reached the bottom, and the final penetration depth is output after weighted averaging of the reference depth and the overpressure depth.

[0078] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for locating power distribution network poles, characterized in that, include: In multiple consecutive depth intervals during pole lifting operations, a pulse excitation with a preset frequency and preset amplitude is applied to the lifting wire rope, and vibration response signals from multiple measuring points on the wire rope in each depth interval are collected to form an energy response spectrum matrix corresponding to each depth interval. The energy response spectrum matrices of the first N consecutive depth intervals are used as normal sinking samples. The energy response spectrum matrix of a certain depth interval is used as input, and the prediction of the energy response spectrum matrix of the next depth interval is used as output to train the depth autoregressive prediction network online. For each new depth interval after the Nth depth interval, the energy response spectrum matrix of the previous depth interval is input into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval is calculated. When the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, the current depth interval is marked as the initial bottoming interval, and the depth value corresponding to the initial bottoming interval is obtained as the reference depth. The excitation is applied to the next depth interval and the response is collected. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth. Calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold. If the difference is greater than the preset elastic compression threshold, it is determined that the pole has stably reached the bottom, and the final penetration depth is output after weighted averaging of the reference depth and the overpressure depth.

2. The deep learning-based method for locating power distribution network poles according to claim 1, characterized in that, The process of collecting vibration response signals from multiple measuring points on the wire rope within each depth interval to construct an energy response spectrum matrix corresponding to each depth interval includes: Each depth interval is divided into multiple time windows of equal duration in chronological order. The length of each time window is 1.5 times the duration of the pulse excitation signal, and there is a 50% overlap between adjacent time windows. For the vibration response signal within each time window, a fast Fourier transform is performed to obtain a frequency domain amplitude sequence, and the amplitude of each frequency component in the frequency domain amplitude sequence is squared to obtain the energy spectrum vector for each time window. Arrange the energy spectrum vectors of all time windows at the same measurement point in chronological order to form a time-frequency energy matrix for the same measurement point. The rows of the time-frequency energy matrix correspond to the frequency points, and the columns correspond to the time window numbers. The time-frequency energy matrices of all measuring points within the same depth range are sequentially spliced ​​together according to the spatial position of the measuring points on the wire rope to form a three-dimensional energy response spectrum matrix. The first dimension of the energy response spectrum matrix is ​​the number of frequency points, the second dimension is the number of time windows, and the third dimension is the number of measuring points.

3. The deep learning-based method for locating power distribution network poles according to claim 1, characterized in that, The online training of the depth autoregressive prediction network, which uses the energy response spectrum matrices of the first N consecutive depth intervals as normal sinking samples, takes the energy response spectrum matrix of a certain depth interval as input, and outputs the prediction of the energy response spectrum matrix of the next depth interval, includes: The energy response spectrum matrices of the first N consecutive depth intervals are numbered E1, E2, ..., E in ascending order of depth. N For any positive integer i, E i As input samples, with E i+1 As supervisory labels, construct N-1 training sample pairs, where 1≤i≤N-1; A deep autoregressive prediction network is constructed, comprising an encoding module and a decoding module. The encoding module is used to compress the input energy response spectrum matrix into a low-dimensional feature vector, and the decoding module is used to restore the low-dimensional feature vector into a prediction matrix with the same dimension as the input. The mean squared error loss function is used to pre-train the deep autoregressive prediction network offline using the N-1 training samples until the loss function value is less than the preset first loss threshold or the number of iterations reaches the preset maximum number of iterations, thus obtaining the initial prediction network. The actual energy response spectrum matrix E for each newly added depth interval is subsequently obtained. j Then, E j E with the previous depth interval j-1 New training sample pairs are formed, and the parameters of the decoding module of the initial prediction network are fine-tuned online using the single-step gradient descent method, while the parameters of the encoding module remain unchanged, where j>N.

4. The deep learning-based method for locating power distribution network poles according to claim 3, characterized in that, The calculation of the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the actual energy response spectrum matrix acquired in the current depth range includes: Calculate the element-wise difference matrix between the predicted energy response spectrum matrix and the actual energy response spectrum matrix collected in the current depth range, and divide all non-zero elements in the difference matrix into multiple amplitude levels according to their numerical values. Calculate the proportion of the number of elements in each amplitude level to the total number of all non-zero elements to obtain the sparsity characteristic value of the difference distribution. The element-wise difference matrix is ​​marginalized and projected along the frequency dimension and the measurement point dimension respectively to obtain the first difference marginal vector in the frequency dimension and the second difference marginal vector in the measurement point dimension. The second central moments of the first difference marginal vector and the second difference marginal vector are calculated respectively and defined as the frequency marginal dispersion and the measurement point marginal dispersion. Calculate the first sign consistency ratio of each row element and the second sign consistency ratio of each column element of the element-wise difference matrix respectively. Calculate the average of all the first sign consistency ratios and the second sign consistency ratios to obtain the row sign consistency index and the column sign consistency index. Then, calculate the weighted average of the row sign consistency index and the column sign consistency index to obtain the total sign consistency index of the difference. Calculate the first frequency energy centroid of the predicted energy response spectrum matrix in the frequency dimension and the second frequency energy centroid of the actual energy response spectrum matrix in the frequency dimension, respectively, and calculate the absolute offset between the first frequency energy centroid and the second frequency energy centroid, which is defined as the frequency centroid offset. Calculate the first measurement point energy centroid of the predicted energy response spectrum matrix in the measurement point dimension and the second measurement point energy centroid of the actual energy response spectrum matrix in the measurement point dimension, respectively, and calculate the absolute offset between the first measurement point energy centroid and the second measurement point energy centroid, which is defined as the measurement point centroid offset. The multi-scale structural dissimilarity index is obtained by normalizing and weighting the sparsity characteristic value of the difference distribution, the frequency marginal dispersion, the measurement point marginal dispersion, the total index of difference sign consistency, the frequency centroid offset, and the measurement point centroid offset.

5. The deep learning-based method for locating power distribution network poles according to claim 1, characterized in that, After determining whether the difference is greater than a preset elastic compression threshold, the method further includes: If the difference is not greater than the preset elastic compression threshold, it is determined that the pole has not stably touched the bottom, and the multi-scale structural dissimilarity index of the subsequent newly added depth range is determined based on the actual energy response spectrum matrix of the subsequent newly added depth range. The new baseline depth and new overpressure depth are determined based on the multi-scale structural dissimilarity index of the subsequently added depth range; Calculate the new difference between the new overpressure depth and the new reference depth, and determine whether the new difference is greater than the preset elastic compression threshold.

6. A deep learning-based power distribution network pole positioning system, characterized in that, include: The module is configured to apply a pulse excitation with a preset frequency and preset amplitude to the hoisting wire rope in multiple consecutive depth intervals during pole hoisting operations, and to collect the vibration response signals of multiple measuring points on the wire rope in each depth interval to form an energy response spectrum matrix corresponding to each depth interval. The training module is configured to use the energy response spectrum matrices of the first N consecutive depth intervals as normal sinking samples, take the energy response spectrum matrix of a certain depth interval as input, and take the prediction of the energy response spectrum matrix of the next depth interval as output to train the depth autoregressive prediction network online. The calculation module is configured to, for each new depth interval after the Nth depth interval, input the energy response spectrum matrix of the previous depth interval into the depth autoregressive prediction network to obtain the predicted energy response spectrum matrix of the current depth interval, and calculate the multi-scale structural dissimilarity index between the predicted energy response spectrum matrix and the energy response spectrum matrix actually collected in the current depth interval. The marking module is configured to mark the current depth interval as the initial bottoming interval when the multi-scale structural dissimilarity index is greater than the preset first mutation threshold for the first time, and obtain the depth value corresponding to the initial bottoming interval as the reference depth, continue to apply excitation to the next depth interval and collect the response. When the rate of change of the multi-scale structural dissimilarity index in multiple consecutive depth intervals changes from positive to negative, and the rate of change remains negative in at least two consecutive depth intervals after the change, the depth interval that first meets the condition is marked as the overpressure interval, and the depth value corresponding to the overpressure interval is recorded as the overpressure depth. The judgment module is configured to calculate the difference between the overpressure depth and the reference depth, and determine whether the difference is greater than a preset elastic compression threshold. The output module is configured to determine that the pole has stably reached the bottom if the difference is greater than the preset elastic compression threshold, and output the final penetration depth after weighted averaging of the reference depth and the overpressure depth.

7. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 5.

Images