A deep learning-based power pole geometric deformation prediction method

Abstract

The application discloses a kind of based on deep learning's electric pole geometric deformation prediction method, comprising: obtaining original electric pole image data list and original electric pole multimodal data list;Original electric pole image data list and original electric pole multimodal data list are carried out data preprocessing, respectively obtain the electric pole image data list and electric pole multimodal data list after processing;Electric pole geometric deformation prediction network is constructed, and the network is handled to the electric pole image data list and electric pole multimodal data list after processing, and the prediction result of electric pole geometric deformation is obtained;The application accurately predicts inclination angle, crack width and the like parameters, provides reliable basis for power operation and maintenance, promotes the change from passive repair to active prevention, reduces accident risk, and improves power grid operation safety and economy.

Description

Technical Field

[0001] This invention belongs to the field of geometries of utility poles, and particularly relates to a deep learning-based method for predicting geometries of utility poles. Background Technology

[0002] As a core infrastructure of the power transmission network, the structural stability of utility poles directly affects the safety of the power grid and the reliability of power supply. During long-term service, utility poles are susceptible to geometric deformations such as tilting and crack propagation due to factors such as natural environment (strong winds, temperature differences, rainfall), material aging (concrete cracking, metal corrosion), foundation settlement, and external impacts. If such deformations are not addressed in time, they may lead to accidents such as short circuits and pole collapse, threatening the safe operation of the power grid and the surrounding environment. Therefore, accurate prediction of utility pole geometric deformation is of great significance for preventative maintenance of the power system and is a key link in ensuring the stable operation of the power grid.

[0003] Currently, methods for predicting the deformation of utility poles have significant limitations. Manual inspections rely on on-site observation, which is inefficient and influenced by subjective experience, making it difficult to capture subtle deformations. Single-sensor monitoring can only acquire local data such as tilt angle and strain value, failing to integrate with the overall structural features of the pole, such as crossarm position and insulator condition, as well as environmental parameters such as wind force and temperature, leading to incomplete predictions. Some machine learning methods attempt to fuse multi-source data, but they also have certain drawbacks. On the one hand, spatiotemporal features are fragmented, with image spatial information and sensor time-series data simply spliced ​​together without exploring the dynamic coupling relationship between the environment and deformation. On the other hand, they have weak anti-interference capabilities, failing to adequately suppress sensor noise and image capture errors, making it difficult to focus on key areas such as crack-prone areas. In addition, multi-step prediction in utility pole deformation prediction has low accuracy, failing to effectively handle the temporal dependence between historical and current states, with errors accumulating with the number of prediction steps. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes a deep learning-based method for predicting the geometric deformation of utility poles. This method accurately predicts parameters such as tilt angle and crack width, providing a reliable basis for power operation and maintenance. It promotes a shift from reactive emergency repairs to proactive prevention, reduces accident risks, and enhances the safety and economy of power grid operation.

[0005] To achieve the above objectives, this invention provides a deep learning-based method for predicting the geometric deformation of utility poles, comprising:

[0006] Obtain the list of raw utility pole image data and the list of raw utility pole multimodal data;

[0007] The original list of utility pole image data and the original list of utility pole multimodal data are preprocessed to obtain the processed list of utility pole image data and the processed list of utility pole multimodal data, respectively.

[0008] A network for predicting the geometric deformation of utility poles is constructed. This network is then used to process the processed list of utility pole image data and the list of utility pole multimodal data to obtain the prediction results of the geometric deformation of the utility poles.

[0009] The utility pole geometric deformation prediction network performs the following operations in sequence:

[0010] S1. The processed list of utility pole image data and the list of utility pole multimodal data are fused with spatial features and multimodal temporal features to obtain a spatiotemporal fusion feature vector of utility poles;

[0011] S2. Perform multi-scale feature extraction and enhancement on the spatiotemporal fusion feature vector of the utility pole to generate an enhanced multi-scale spatiotemporal fusion feature vector of the utility pole.

[0012] S3. Perform noise suppression and key feature focusing on the enhanced multi-scale spatiotemporal fusion feature vector of the utility pole and the processed multi-modal data list of the utility pole to generate a high-precision deformation feature vector of the utility pole.

[0013] S4. Perform gating mechanism processing and multi-step deformation prediction on the high-precision pole deformation feature vector to obtain the prediction result of the pole geometric deformation; deploy the pole geometric deformation prediction network to the monitoring system to output the prediction result.

[0014] Technical Effects of this Invention: This invention discloses a deep learning-based method for predicting the geometric deformation of utility poles. By deeply integrating spatial structural information from image data with temporal monitoring data from multimodal sensors, an end-to-end intelligent prediction model is constructed, enabling accurate capture of the dynamic correlation between environmental factors and pole deformation. Simultaneously, a physical constraint mechanism is introduced to enhance feature anti-interference capabilities, strengthen feature representation of key deformation areas, and effectively reduce error accumulation in long-term predictions by optimizing temporal dependency processing strategies. This method can accurately predict parameters such as tilt angle and crack width, providing reliable data for power operation and maintenance, promoting a shift from passive emergency repairs to proactive prevention, reducing accident risks, and improving the safety and economy of power grid operation. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1This is a flowchart illustrating a method for predicting the geometric deformation of utility poles based on deep learning, according to an embodiment of the present invention.

[0017] Figure 2 This is a flowchart illustrating the EPGDPNet network execution process according to an embodiment of the present invention.

[0018] Figure 3 This is a structural diagram of the PSFF module according to an embodiment of the present invention;

[0019] Figure 4 This is a structural diagram of the PMFE module according to an embodiment of the present invention;

[0020] Figure 5 This is a structural diagram of the PDFC module in an embodiment of the present invention;

[0021] Figure 6 This is a structural diagram of the PGDP module in an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0024] like Figure 1 As shown, this embodiment provides a method for predicting the geometric deformation of utility poles based on deep learning, including:

[0025] Obtain the list of raw utility pole image data and the list of raw utility pole multimodal data;

[0026] The original list of utility pole image data and the original list of utility pole multimodal data are preprocessed to obtain the processed list of utility pole image data and the processed list of utility pole multimodal data, respectively.

[0027] A network for predicting the geometric deformation of utility poles is constructed. This network is then used to process the processed list of utility pole image data and the list of utility pole multimodal data to obtain the prediction results of the geometric deformation of the utility poles.

[0028] The utility pole geometric deformation prediction network performs the following operations in sequence:

[0029] S1. The processed list of utility pole image data and the list of utility pole multimodal data are fused with spatial features and multimodal temporal features to obtain a spatiotemporal fusion feature vector of utility poles;

[0030] S2. Perform multi-scale feature extraction and enhancement on the spatiotemporal fusion feature vector of the utility pole to generate an enhanced multi-scale spatiotemporal fusion feature vector of the utility pole.

[0031] S3. Perform noise suppression and key feature focusing on the enhanced multi-scale spatiotemporal fusion feature vector of the utility pole and the processed multi-modal data list of the utility pole to generate a high-precision deformation feature vector of the utility pole.

[0032] S4. Perform gating mechanism processing and multi-step deformation prediction on the high-precision pole deformation feature vector to obtain the prediction result of the pole geometric deformation; deploy the pole geometric deformation prediction network to the monitoring system to output the prediction result.

[0033] Furthermore, the data preprocessing process includes:

[0034] Invalid image samples are removed from the original list of utility pole image data to obtain the processed list of utility pole image data.

[0035] Abnormal data in the original multimodal data list of utility poles is detected and removed, and missing values ​​are filled by interpolation to obtain the processed multimodal data list of utility poles.

[0036] Furthermore, the process of fusing spatial features and multimodal temporal features in S1 includes:

[0037] The spatial features of the processed utility pole image data list are extracted using a utility pole spatial attention mechanism to obtain the utility pole spatial feature vector;

[0038] The time features of the processed multimodal data list of utility poles are extracted using a time attention mechanism for utility poles to obtain a time feature vector for utility poles.

[0039] The spatial feature vector and the temporal feature vector of the utility pole are input into the utility pole feature fusion submodule for processing to generate the spatiotemporal fusion feature vector of the utility pole.

[0040] Furthermore, the process of processing the utility pole feature fusion submodule in S1 includes:

[0041] Element-wise multiplication and element-wise addition operations are performed on the spatial feature vector and the temporal feature vector of the utility pole to generate interactive features;

[0042] The interactive features are spliced ​​and linearly transformed to calculate the utility pole feature fusion vector for dynamically adjusting the fusion ratio.

[0043] Based on the feature fusion vector of the utility pole, the interactive features are weighted and fused and nonlinearly enhanced to output the spatiotemporal fusion feature vector of the utility pole.

[0044] Furthermore, the process of multi-scale feature extraction and enhancement in S2 includes:

[0045] Convolution operations of different scales are performed on the spatiotemporal fusion feature vector of the utility pole to capture correlations of different ranges, resulting in a multi-scale feature tensor.

[0046] The multi-scale feature tensors are spliced ​​together and the channel dimension is compressed to generate a multi-scale spatiotemporal feature tensor for utility poles.

[0047] The spatiotemporal fusion feature vector of the utility pole is residually connected with the multi-scale spatiotemporal feature tensor of the utility pole, and the result after connection is nonlinearly transformed to generate the enhanced multi-scale spatiotemporal fusion feature vector of the utility pole.

[0048] Furthermore, the noise suppression and key feature focusing process in S3 includes:

[0049] Modal decomposition, modal filtering, and dynamic soft thresholding are performed on the enhanced multi-scale spatiotemporal fusion feature vector of the utility pole and the processed multi-modal data list of the utility pole to obtain the denoised feature vector of the utility pole.

[0050] A dual attention mechanism is used to perform key feature focusing operations in the spatial and temporal dimensions on the denoised feature vector of the utility pole, generating the high-precision deformation feature vector of the utility pole.

[0051] Furthermore, the gating mechanism processing and multi-step deformation prediction process in S4 includes:

[0052] A deformation feature prediction method is used to predict the deformation feature vector of the high-precision utility pole by applying a utility pole deformation gating mechanism to obtain the hidden feature state;

[0053] Multiple prediction branches are constructed based on the prediction step size, and multi-step deformation prediction is performed on the hidden feature state to obtain a multi-step deformation prediction feature matrix.

[0054] The multi-step deformation prediction feature matrix is ​​convolved to enhance temporal correlation, and physical constraints are introduced for boundary correction to obtain the final multi-step deformation prediction vector of the utility pole.

[0055] Furthermore, the predicted geometric deformation of the utility pole includes predicted values ​​for the pole's tilt angle, material strain, and crack width at various future times. Specifically, the implementation process of this embodiment includes:

[0056] Step 1: Obtain the list of raw utility pole image data and the original list of multimodal data of utility poles ;

[0057] Using drones and fixed high-definition cameras, utility poles are periodically photographed from multiple directions and heights to record the spatial geometry of components such as the pole body, crossarms, and insulators. Timestamped images of the utility poles are generated, and the spatial coordinates of the components are extracted from these images. Using velocity vectors as node features, and arranging them by timestamps, we obtain a list of original utility pole image data. ,in , For time frames, It includes node features and their corresponding timestamps. ;

[0058] Tilt sensors were installed at the top, middle, and bottom of the utility pole to monitor the tilt angle. Strain sensors were installed at the crossarm joints to monitor material strain. Crack sensors were installed on the pole surface to measure crack width. Meteorological data acquisition equipment was installed next to the pole to record external environmental parameters such as temperature, wind force, rainfall, and air pressure. Timestamped measurements of tilt angle, material strain, crack width, temperature, wind force, rainfall, and air pressure were collected at a frequency of 1Hz to obtain the original multimodal data list of the utility pole. ,in , For quantity index, It is an octet , , For timestamps, This is the tilt angle value. The material strain value. This represents the crack width value. For temperature, For wind power, For rainfall, This is the air pressure value.

[0059] Step 2: [The sentence is incomplete and requires more context.] and Data preprocessing was performed to obtain separate lists of processed utility pole image data. Multimodal data list of utility poles ;

[0060] Step 2.1: Elimination Invalid images that are blurry, severely obstructed, or taken from abnormal angles are removed. Metadata consistency is ensured through timestamp verification. Clear and complete valid samples containing key components are retained, resulting in a processed list of utility pole image data. ,in , This represents the number of time frames.

[0061] Step 2.2: Use the Z-score method to detect and eliminate. The abnormal octets in the data are defined as octets where, due to sensor malfunctions or environmental interference, at least one of the measured values ​​(timestamp, tilt angle, material strain, crack width, temperature, wind force, rainfall, air pressure, etc.) significantly deviates from the normal distribution. Then, linear interpolation is used to fill in the missing values ​​after removing the abnormal octets to ensure the continuity of the time series, resulting in a processed multimodal data list for utility poles. ,in , Indexed by quantity.

[0062] Step 3: Design an EPGDPNet network and use this network to... and The process is performed to obtain the predicted results of the geometric deformation of the utility pole;

[0063] like Figure 2 As shown, an Electric Pole Geometric Deformation Prediction Network (EPGDPNet) is designed to achieve high-precision prediction of electric pole geometric deformation, providing data-driven support for preventive maintenance of power infrastructure and ultimately obtaining the predicted results of electric pole geometric deformation. The network includes the following modules: a Pole Spatiotemporal Feature Fusion Module (PSFF), a Pole Multi-scale Feature Enhancement Module (PMFE), a Pole Deformation Feature Correction Module (PDFC), and a Pole Geometric Deformation Prediction Module (PGDP).

[0064] Step 3.1: Build the PSFF module, for and Spatial features and multimodal temporal features are fused to obtain the spatiotemporal fusion feature vector of the utility pole. .

[0065] The PSFF module employs both spatial and temporal attention mechanisms for utility poles, integrating spatial and multimodal temporal features of the image and dynamically adjusting the fusion ratio to more comprehensively capture the deformation correlation of utility poles. The structure of the PSFF module is as follows: Figure 3 As shown.

[0066] Step 3.1.1: Extracting spatial attention using utility poles The spatial characteristics of the utility pole are used to obtain its spatial feature vector. .

[0067] In the spatial attention mechanism for utility poles, multi-scale convolution and temporal stitching are used to enhance the fusion of spatiotemporal correlations, details, and global features, accurately focusing on the key structures of the utility poles and improving the targeting and effectiveness of feature extraction. The specific execution steps of the spatial attention mechanism for utility poles are as follows:

[0068] Step 3.1.1.1: [The text appears to be incomplete and contains several grammatical errors. A more The input is processed by the spatial attention mechanism of the utility pole, and any adjacent... Perform a concat concatenation operation in sequence to obtain .right The Conv3×3 convolution operation is performed using a 3×3 kernel to obtain the result. ,right Attention coefficients were calculated using the LeakyReLU activation function. And then Perform Softmax normalization to output spatial features of utility poles with attention weights. .

[0069] Step 3.1.1.2: For Two 3×3 convolutional kernels are used to perform Conv3×3 convolution operations to extract local detail features of the utility pole, thus obtaining the local features of the utility pole. ,right A Conv5×5 convolution operation with three 5×5 kernels is used to extract the global structural features of the utility pole, resulting in the global features of the utility pole. And then and Perform a concat stitching operation along the channel dimension to generate multi-scale spatial features of utility poles. .

[0070] Step 3.1.1.3: For By performing a Conv1×1 convolution operation with a 1×1 kernel, the feature channel dimension is reduced, resulting in... By using a fully connected layer to map features of different scales to the same dimensional space, spatial feature vectors of utility poles are obtained. .

[0071] Step 3.1.2: Extracting attention using the time-based attention mechanism of utility poles The time characteristics are used to obtain the time feature vector of the utility pole. .

[0072] In the time-attention mechanism for utility poles, key time steps are dynamically focused through attention weights to strengthen temporal features that significantly affect deformation, thereby improving the ability to capture temporal correlations and the accuracy of feature representation. The specific execution steps of the time-attention mechanism for utility poles are as follows:

[0073] Step 3.1.2.1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] The input to the utility pole is processed using a time-attention mechanism, for any... Softmax normalization was performed to obtain preliminary time-series features of the utility poles. Through a fully connected layer The feature vectors at each time step are mapped to a unified high-dimensional space to generate an encoded time-series feature state sequence of the utility pole. .

[0074] Step 3.1.2.2: For The LeakyReLU activation function is used to calculate the temporal attention coefficients, and then the coefficients are normalized using the Softmax function to obtain the attention weights for the temporal features of the utility poles at each time step. The weight value is positively correlated with the degree of influence of that time step on the deformation. Based on right Weighted aggregation is performed, and weighted aggregation is based on... As weight, for The features at different time steps are weighted and summed to output the time-attention features of the utility pole with attention weights. :

[0075] Step 3.1.2.3: For A Conv1×1 convolution operation is performed using a 1×1 kernel to compress the feature channel dimension and enhance the correlation between features, resulting in the compressed temporal weight features of the utility poles. ; through a fully connected layer Mapped to the feature vector of the utility pole space By matching the dimensional space, we obtain the time feature vector of the utility pole. .

[0076] Step 3.1.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The input is processed by the utility pole feature fusion submodule to generate a spatiotemporal fusion feature vector for the utility pole. .

[0077] In the utility pole feature fusion submodule, by integrating the spatiotemporal features of the utility pole, redundant interference can be eliminated to enhance effective deformation information. The specific execution steps of the utility pole feature fusion submodule are as follows:

[0078] Step 3.1.3.1: For and The first feature of spatiotemporal interaction in generating utility poles by performing element-wise multiplication operations :

[0079] ;

[0080] in This indicates element-wise multiplication.

[0081] Step 3.1.3.2: For and Perform element addition to generate the second spatiotemporal interaction feature of the telephone pole. , The basic feature information obtained by integrating the spatial and temporal features of the utility pole is preserved:

[0082] .

[0083] Step 3.1.3.3: For and Perform the Concat concatenation operation to obtain the joint interaction features of the utility poles. And then After performing a linear transformation, the Sigmoid activation function is used to calculate the feature fusion vector of the utility pole. :

[0084] ;

[0085] in For learnable weight matrix, For bias terms, The range of values ​​for each component is: It is used to dynamically adjust the fusion ratio of the spatial and temporal characteristics of utility poles.

[0086] Step 3.1.3.4: Based on right and Perform a spatiotemporal feature weighted fusion operation to obtain the spatiotemporal adaptive feature vector of the utility pole. :

[0087] ;

[0088] in, Weighted components of spatial features are used to retain messages in the spatial features of a component that significantly contribute to the current deformation. It is a weighted component of time features, used to enhance the feature representation strength of time series patterns driven by environmental parameters.

[0089] Step 3.1.3.4: Process the fused feature vectors Perform nonlinear enhancement, After a linear transformation, the input is fed into a fully connected layer containing the LeakyReLU activation function. The following transformation enhances cross-modal feature interaction, resulting in a spatiotemporal fusion feature vector of the utility pole. :

[0090] ;

[0091] in, This is the weight matrix; For bias terms; This is a spatiotemporal fusion feature vector for utility poles, which encodes the spatial characteristics of utility pole components and the temporal patterns between the environment and deformation.

[0092] Step 3.2: Construct the PMFE module, for Multi-scale feature extraction and enhancement are performed to generate enhanced multi-scale spatiotemporal fusion feature vectors for utility poles. .

[0093] In the PMFE module, multi-scale convolution captures correlations at different ranges, fuses multi-scale features and performs residual enhancement to strengthen feature interactions, mitigate gradient vanishing, and improve feature generalization and representation capabilities. The structure of the PMFE module is as follows: Figure 4 As shown.

[0094] First of all, A Conv1×1 convolution operation is performed using a 1×1 kernel to capture the local correlations in each dimension of the feature vector, resulting in a small-scale feature tensor. ;right Conv3×3 convolution is performed using a 3×3 kernel to capture the medium-range correlation across dimensions in the feature vectors, resulting in a mesoscale feature tensor. ;right A Conv5×5 convolution operation is performed using a 5×5 kernel. By capturing global correlations in the feature vectors through a larger receptive field, a large-scale feature tensor is obtained. Then... The small-scale deformation feature tensor is obtained by using the batch normalization (BN) operation. ,right The mesoscale deformation feature tensor is obtained by performing a BN operation. ,right Small-scale deformation feature tensors are obtained using Batch Normalization (BN) operations. .

[0095] Then, , , Perform a concat concatenation operation along the channel dimension to generate a multi-scale fusion feature tensor for utility poles. Then use a 1×1 convolution kernel to pair Performing a Conv1×1 convolution operation compresses the channel dimensions, resulting in a dimensionally uniform multi-scale spatiotemporal feature tensor for utility poles. This enhances the interaction and correlation between features at different scales.

[0096] Finally, for and Element-wise addition is performed to alleviate the gradient vanishing problem in deep networks, resulting in the utility pole residual enhancement feature tensor. Then The input is fed into a fully connected layer containing the GELU (Gaussian Error Linear Unit) activation function, and the enhanced multi-scale spatiotemporal fusion feature vector of the utility pole is obtained through the following transformation to achieve nonlinear enhancement of the features. :

[0097] ;

[0098] in, For learnable weight matrix, For bias terms, This means flattening a three-dimensional tensor into a one-dimensional vector, using GELU as the activation function, and introducing random regularization to improve the generalization ability of the features.

[0099] Step 3.3: Build the PDFC module, for and Noise suppression and key feature focusing are performed to generate high-precision feature vectors of utility pole deformation. .

[0100] In the PDFC module, the modal information extraction submodule is used to... and Accurate noise reduction is achieved; key features are focused through a dual-attention mechanism, and stability is improved by combining physical constraints, enabling better identification and extraction of key features easily masked by noise. The structure of the PDFC module is as follows: Figure 5 As shown.

[0101] Step 3.3.1: Design a modal information extraction submodule, for... and Noise suppression is performed to obtain the denoised feature vector of the utility pole. .

[0102] The execution flow of the modal information extraction submodule is as follows:

[0103] Step 3.3.1.1: For and Perform a modal decomposition operation on the pole deformation to obtain the first mode set of the pole deformation. This operation can be expressed as the following optimization objective function:

[0104] ;

[0105] in, Each element contains frequency components of specific deformations such as tilt angle change and crack width; For set The center frequencies of each mode, It is a modal index used to distinguish different pole deformation modes; yes The first in One deformation mode; For Dirac function, Represents a time variable; The time partial derivative describes the rate of change of the deformation characteristics of the utility pole over time. This is a frequency domain transformation factor that converts a time-domain signal to the frequency domain. The square of the L2 norm; These are the physical constraint weighting coefficients; This is a multi-scale spatiotemporal fusion feature vector for utility poles; These are the theoretical deformation components derived from the mechanical properties of utility pole materials; This represents the minimization operation; it is obtained by minimizing the optimization objective function. This function consists of two parts: one is to let each modal component... via Dirac Functions and Combinatorial time-domain processing, time partial derivatives After the operation and frequency domain transformation factor transformation, and with Minimizing the sum of squares of the L2 norm makes the decomposition result fit the original data; secondly, through the weighting coefficients... adjust, and theoretical deformation components Minimizing the sum of squares ensures that the modes conform to physical laws. By minimizing the sum of the results from the above two parts, we finally obtain the first mode set of the pole deformation, which includes specific deformation frequency components such as tilt angle change and crack width. .

[0106] Step 3.3.1.2: Calculation The cosine similarity is used to obtain the deformation similarity. Then by right By performing mode screening of the pole deformation, the set of second modes of pole deformation was obtained. .

[0107] The formula for calculating cosine similarity is as follows:

[0108] ;

[0109] in, for and The vector dot product between them.

[0110] Depend on right By performing mode screening of the pole deformation, the set of second modes of pole deformation was obtained. The rules are as follows:

[0111] ;

[0112] in, This is the similarity threshold, with a range of values. .

[0113] Step 3.3.1.3: Calculation Dynamic soft threshold ,use After screening Dynamic soft thresholding is performed to obtain the third mode set of utility pole deformation. ,Will All modal components are linearly superimposed using element-wise addition to obtain the denoised feature vector of the utility pole. .

[0114] First, calculate using the following formula Dynamic soft threshold :

[0115] ;

[0116] in, Dynamic soft threshold; It is the set of second modes of utility pole deformation after feature filtering. Standard deviation; It is the feature length; It is the signal-to-noise ratio.

[0117] Then, by After screening Perform dynamic soft thresholding to obtain The dynamic soft threshold processing formula is as follows:

[0118] ;

[0119] Finally, All modal components are linearly superimposed using element-wise addition to obtain... , Retained The core deformation features in it, and integrated with The effective components after screening.

[0120] Step 3.3.2: Utilizing the dual attention mechanism to... Perform key feature focusing operations to generate high-precision utility pole deformation feature vectors. .

[0121] The execution flow of the dual attention mechanism is as follows:

[0122] Step 3.3.2.1: Calculation Spatial attention weight matrix, generating the spatial attention weight matrix for utility poles. ,based on right The spatial focusing features of the utility poles are obtained by weighting through element-wise multiplication. .

[0123] The formula for calculating the spatial attention weight matrix is ​​as follows:

[0124] ;

[0125] in, The normalization function maps values ​​to interval; To prevent gradient vanishing, a non-linear characteristic is introduced into the activation function. Here is the learnable weight matrix for the spatial attention branch. For bias terms; Used to highlight The feature dimensions are related to the key pole body tilt sensitive areas, crack-prone areas and deformation parts.

[0126] based on right Weighting is performed using element-wise multiplication to obtain... .

[0127] Step 3.3.2.2 Calculation The time attention weight vector is used to generate the time attention weight vector for utility poles. ,based on right The time-focused features of the utility poles are obtained by weighting through element-wise multiplication. .

[0128] The formula for calculating the time attention weight vector is as follows:

[0129] ;

[0130] in, The learnable weight matrix for the temporal attention branch. For bias terms; Indicates execution Flattening operation converts a multidimensional vector into a one-dimensional vector; Used to capture the contribution of features at different time steps to the current deformation state.

[0131] based on right Weighting is performed using element-wise multiplication to obtain... .

[0132] Step 3.3.2.3: [The text appears to be incomplete and contains several grammatical errors. A more and Perform a Concat stitching operation to obtain the joint focusing features of the utility poles. Furthermore, a high-precision feature vector of utility pole deformation is generated by performing a nonlinear transformation through a fully connected layer containing the GELU activation function. , It integrates key deformation information from both spatial and temporal dimensions.

[0133] The formula is as follows:

[0134] ;

[0135] in, The learnable weight matrix of the fully connected layer. For bias terms, Indicates feature concatenation operation; We introduce random regularization properties into the Gaussian error linear unit activation function.

[0136] Step 3.4: Construct the PGDP module to obtain the predicted results of the geometric deformation of the utility pole.

[0137] The PGDP module incorporates gating mechanisms and physical constraints, employing multi-step prediction branches to strengthen temporal correlations. This allows for the correction of prediction boundaries and improves the accuracy of capturing the deformation trend of utility poles. The structure of the PGDP module is as follows: Figure 6 As shown.

[0138] Step 3.4.1: Utilize the pole deformation gating mechanism to... Deformation feature prediction is performed to obtain the fourth hidden feature state of the utility pole. .

[0139] Step 3.4.1.1: Use a gating mechanism to... Selective integration is performed, and the filtering coefficient of the power pole deformation history information is calculated using the following formula. Input coefficient for current information on pole deformation Candidate states for continuous information on pole deformation The final output coefficient of the pole deformation information :

[0140] ;

[0141] in, Determines the current characteristics The degree of adoption; Candidate states used to capture the temporal dependencies of utility pole deformation features; Used to control the output strength of the final hidden state; As an activation function, it compresses the values ​​to interval; It is the hyperbolic tangent activation function; , , and The weight matrix is ​​a learnable weight matrix; , , and This is a bias term used to enhance the model's expressive power; The first hidden feature state of the utility pole is the gating mechanism that integrates historical information with the current input (i.e., the high-precision utility pole deformation feature vector). The key reference benchmark is used to store deformation-related information obtained by the model in previous time steps, such as historical tilt angle trends, crack development laws, and material strain laws. for and The concatenated vector.

[0142] Step 3.4.1.2: For , , and The candidate features are updated using element-wise multiplication and addition operations to obtain the core message state of the utility pole deformation. .

[0143] .

[0144] Step 3.4.1.3: For Using the tanh function for nonlinear transformation, we obtain ,right and Performing element-wise multiplication yields the second hidden feature state of the utility pole. .right The attention weights for the telephone pole message state at each time step are calculated using the Score attention scoring function. It is used to capture the evolutionary trends hidden in features.

[0145] Step 3.4.1.4: For The deformable normalized attention weights are obtained using the softmax function. ,right and The deformation attention context vector is obtained by performing element-wise multiplication. ,Will and Perform the Concat concatenation operation to obtain the third hidden feature state of the utility pole. ,right The fourth hidden feature state of the utility pole at the predicted time is calculated by performing a nonlinear transformation using the LeakyReLU function. .

[0146] Step 3.4.2: For Perform multi-step deformation prediction to obtain the multi-step prediction vector of the utility pole deformation. .

[0147] Step 3.4.2.1: Based on prediction step size Construct a fully connected prediction branch for time, at each time step Corresponding to a separate fully connected prediction branch, ,Will The input is fed into the fully connected prediction branch and transformed using the following prediction formula to obtain the multi-step deformation prediction feature matrix. :

[0148] ;

[0149] in, and For the first Learnable parameters for each time step prediction branch; This is the fourth hidden feature state of the utility pole; for The single-step deformation prediction feature matrix at time step. The predicted values ​​at each time step are integrated to generate a multi-step deformation prediction feature matrix. .

[0150] Step 3.4.2.2: Use a 1×1 convolution kernel. Convolution operations are performed to enhance the temporal correlation between multi-step predictions, outputting a multi-step correlation prediction matrix with consistent dimensions. Based on the mechanical properties and structural design parameters of the utility pole material, reasonable ranges for each deformation index are determined, and physical constraints on the actual deformation parameters are introduced. Boundary corrections are performed to obtain the multi-step deformation prediction correction matrix. The corrected formula is:

[0151] ;

[0152] in, and These are the minimum and maximum reasonable values ​​for the deformation index, respectively; Correction matrix for multi-step deformation prediction; for The correlation prediction matrix at time point, .

[0153] Step 3.4.2.3: For and Weighted aggregation using element-wise multiplication is used to generate the final multi-step prediction vector for utility pole deformation. , It includes predicted values ​​for key deformation parameters of the utility pole at various times, including tilt angle, material strain, and crack width.

[0154] Step 4: and The training and validation sets are loaded into the EPGDPNet network for training to update the network parameters, and then... and The test set in the dataset is used as input to validate the network performance;

[0155] First, initialize all neural network parameters and related hyperparameters, including but not limited to convolutional kernel weights, bias terms, fully connected layer weight matrices, training epochs, batch size, learning rate, activation function, and loss function. Divide the dataset into three subsets in an 8:1:1 ratio: 80% as the training set, 10% as the validation set, and 10% as the test set, maintaining time series continuity across all subsets. The data is divided in the same 8:1:1 ratio: 80% is used as the multimodal sequence training set, 10% as the multimodal sequence validation set, and 10% as the multimodal sequence test set. This division maintains consistency with previous data. The temporal correspondence.

[0156] During the model training phase, and The training set is loaded into the EPGDPNet network in batches. During training, actual deformation data of utility poles are used as labels, and the learnable parameters in the network, including the weight matrices and bias terms of each module, are continuously adjusted through backpropagation to make the network's predicted values ​​gradually approach the actual deformation values. and The validation set is used to evaluate model performance and calculate prediction error. When the validation error fails to improve for several consecutive rounds, an early stopping mechanism is triggered to save the current optimal parameters in order to balance the model's fitting ability and generalization ability and avoid overfitting.

[0157] After entering the verification phase, and The trained network is input into a test set to generate the final deformation prediction results. The model accuracy is quantitatively evaluated by comparing the prediction results with the actual deformation data in the test set. Simultaneously, the prediction deviations of deformation indices such as tilt angle, material strain, and crack width are analyzed to verify the network's reliability in capturing subtle deformation trends and deformation patterns under extreme weather conditions.

[0158] Step 5: Apply the trained EPGDPNet network to the intelligent prediction task of utility pole geometric deformation.

[0159] The trained EPGDPNet network is deployed into the intelligent monitoring system for utility pole deformation, receiving real-time image data and multimodal sensor data, including tilt angle values. Material strain value Crack width value ,temperature Wind power Rainfall , is the air pressure value Output the multi-step prediction vector of the utility pole deformation. , It includes predicted values ​​for key deformation parameters of the utility pole at various future times, including tilt angle, material strain, and crack width.

[0160] The aforementioned detection system incorporates safety thresholds for pole deformation, specifically including safety thresholds for tilt angle, material strain, and crack width. When the network outputs a multi-step prediction vector for pole deformation... When any component exceeds the corresponding threshold, the system automatically triggers a tiered alarm based on the number of items exceeding the threshold: a Level 1 alarm is triggered when only one component exceeds the safety threshold, indicating a minor risk, and displays the value and trend of the parameter exceeding the threshold and pushes it to the minor warning module; a Level 2 alarm is triggered when two components exceed the safety threshold, indicating a moderate risk, and displays the value and trend of the parameter exceeding the threshold, and pushes it to the key warning module for comparison with historical deformation data over a recent period; a Level 3 alarm is triggered when all three components exceed the safety threshold, indicating a severe risk.

[0161] After integrating real-time data, historical curves, and safety ranges, the data is pushed to the emergency warning module, simultaneously notifying the operation and maintenance manager and marking it as requiring urgent inspection. At the same time, the system combines historical data to generate a visual report, presenting a comparison between the predicted curves of the three deformation parameters and the safety thresholds, providing data support for inspection plans and reinforcement schemes, assisting in the formulation of preventive maintenance plans, and timely identifying risks to improve the safety and stability of power infrastructure.

[0162] An application example of this invention:

[0163] During the data collection phase, a six-month data collection process was conducted. Using quadcopter drones equipped with cameras and three fixed high-definition monitoring devices along the line, images of the utility poles were captured every three hours from five angles: 0°, 45°, 90°, 135°, and 180°. The spatial coordinates of components such as the pole body (bottom, middle, and top), crossarms, and insulators were extracted. Using velocity vectors as node features, a list of original utility pole image data with timestamps is generated. ,in =14400 (8 frames per day × 180 days), each All data includes node characteristics and time sequence information. Additionally, XW-200 tilt sensors are installed at the top, middle, and bottom of the pole; BX120-3AA strain sensors are deployed at the crossarm connections; CWF-1 crack sensors are installed on the pole surface; and a PH-100 weather station is set up 5 meters away from the pole. Parameters such as tilt angle, material strain, crack width, temperature, wind force, rainfall, and air pressure are collected at a frequency of 1Hz to form a raw multimodal data list of the utility pole. .

[0164] During data preprocessing, remove The 2100 frames of blurred images and abnormal shooting angle samples caused by heavy rain and backlighting were processed after being verified by timestamp. The dataset was divided in an 8:1:1 ratio, with 9840 frames used as the training set, 1230 frames as the validation set, and 1230 frames as the test set, maintaining temporal continuity among the subsets. The Z-score method was used, with a threshold of 3.2, to detect and remove 1.8% of outliers. Missing data of 3.1% were filled using linear interpolation, and the data was also partitioned in an 8:1:1 ratio. Ensure with The spatiotemporal correspondence.

[0165] During the model training phase, and The training set is input into the EPGDPNet network, which is processed by the PSFF module. and Generate spatiotemporal fusion feature vectors of utility poles that integrate spatial structure and temporal changes. Spatial features Focusing on the geometric relationships of key parts of the shaft. The PMFE module... Perform multi-receptive field feature extraction to generate enhanced features. The PDFC module uses noise suppression and key feature focusing to filter out effective modes with a cosine similarity greater than 0.72 to the theoretical deformation components, and obtains high-precision deformation features after removing noise. The PGDP module is based on To achieve deformation prediction over the next 5 steps (12 hours each), outputting parameters including tilt angle, material strain, and crack width. The initial learning rate was set to 0.001, the batch size to 64, the loss function to MSE, and the initial training epochs to 200. After training, the validation set error tended to stabilize, and the early stopping mechanism saved the optimal parameters. These parameters need to be fine-tuned based on the results of multiple training epochs until the model achieves the best performance.

[0166] The trained EPGDPNet network was integrated into a smart pole deformation monitoring system. This system has real-time data reception and processing capabilities, continuously acquiring image data and multimodal sensor data from the site and inputting this data into the model for real-time computation. The system internally presets safety thresholds for tilt angle, material strain, and crack width. The safety threshold for tilt angle is set to 3°, the safety threshold for material strain is set to 1500 με, and the safety threshold for crack width is set to 0.3 mm. When the model outputs... When the predicted values ​​of key deformation parameters of the utility pole exceed the corresponding thresholds at any given time, the system will immediately activate a tiered early warning mechanism: If only one parameter exceeds the standard, such as a tilt angle of 3.5°, a Level 1 alarm is triggered, indicating a slight risk, and displaying the value and trend of the parameter exceeding the threshold; if two parameters exceed the standard, such as a tilt angle of 3.2° and a crack width of 0.4mm, a Level 2 alarm is triggered, indicating a moderate risk, and in addition to displaying the value and trend of the parameter exceeding the threshold, it also includes a comparison of historical deformation data from the past month; if all three parameters exceed the standard, such as a tilt angle of 3.8°, a material strain of 1600με, and a crack width of 0.5mm, a Level 3 alarm is triggered, indicating a severe risk, and real-time monitoring data, historical trend curves, and safety threshold ranges are simultaneously pushed to the system. These alarm messages are pushed to the operation and maintenance management platform and staff's mobile terminals in real time via the network. The system also automatically retrieves the historical monitoring data of the utility pole and generates a visual report that includes predicted trend curves and historical data comparison analysis. Based on early warning information and reports, maintenance personnel can promptly arrange on-site inspections, further assess potential risk poles, and develop targeted maintenance and reinforcement plans to achieve dynamic monitoring and early intervention of pole geometric deformation.

[0167] This invention proposes a spatial attention mechanism, a temporal attention mechanism, and a feature fusion submodule for utility poles. By dynamically adjusting the fusion ratio of spatial features and multimodal temporal features, it achieves accurate capture of the coupling relationship between environmental factors and pole deformation. This invention also proposes noise suppression and key feature focusing operations, utilizing a modal information extraction submodule and a dual attention focusing mechanism to effectively suppress noise interference and enhance features in key deformation regions, thereby improving feature robustness. Finally, this invention proposes a utility pole deformation gating mechanism and a multi-step deformation prediction operation. By selectively integrating historical information and dynamically weighted multi-step prediction results, it reduces error accumulation and achieves high-precision long-term prediction of utility pole geometric deformation.

[0168] Compared with the prior art, the advantages of this invention are:

[0169] (1) By using the spatial attention mechanism, the temporal attention mechanism and the feature fusion submodule of the utility pole, the deep coupling of image spatial features and multimodal temporal features is realized, which solves the problem of single-modal information fragmentation in traditional methods. The fusion capability is stronger and can accurately capture the correlation between environmental dynamics and pole deformation.

[0170] (2) Combining physical constraint mode decomposition and dual attention mechanism, it not only suppresses sensor noise and image interference, but also focuses on key parts such as crack-prone areas, resulting in higher feature robustness and significantly better feature accuracy and anti-interference ability than traditional methods.

[0171] (3) By using the pole deformation gating mechanism and multi-step deformation prediction operation, historical information is effectively screened and the accumulation of multi-step prediction error is reduced. The accuracy of long-term deformation trend prediction is significantly improved compared with the traditional time series model, and the prediction accuracy is better, which can provide a more reliable decision basis for power operation and maintenance.

[0172] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based power pole geometric deformation prediction method, characterized in that, The method comprises the following steps: obtaining a list of original power pole image data and a list of original power pole multi-modal data; performing data preprocessing on the list of original power pole image data and the list of original power pole multi-modal data to obtain a list of processed power pole image data and a list of processed power pole multi-modal data; constructing a power pole geometric deformation prediction network, and processing the list of processed power pole image data and the list of processed power pole multi-modal data by using the network to obtain a prediction result of power pole geometric deformation; wherein the power pole geometric deformation prediction network sequentially performs the following operations: S1, performing spatial feature and multi-modal time sequence feature fusion on the list of processed power pole image data and the list of processed power pole multi-modal data to obtain a power pole spatio-temporal fusion feature vector; S2, performing multi-scale feature extraction and enhancement on the power pole spatio-temporal fusion feature vector to generate a reinforced power pole multi-scale spatio-temporal fusion feature vector; S3, performing noise suppression and key feature focusing on the reinforced power pole multi-scale spatio-temporal fusion feature vector and the list of processed power pole multi-modal data to generate a high-precision power pole deformation feature vector; S4, performing gating mechanism processing and multi-step deformation prediction on the high-precision power pole deformation feature vector to obtain a prediction result of power pole geometric deformation; and deploying the power pole geometric deformation prediction network to a monitoring system to output the prediction result; the process of spatial feature and multi-modal time sequence feature fusion in S1 comprises: extracting spatial features of the list of processed power pole image data by using a power pole spatial attention mechanism to obtain a power pole spatial feature vector; extracting time features of the list of processed power pole multi-modal data by using a power pole time attention mechanism to obtain a power pole time feature vector; inputting the power pole spatial feature vector and the power pole time feature vector into a power pole feature fusion submodule for processing to generate the power pole spatio-temporal fusion feature vector; the process of noise suppression and key feature focusing in S3 comprises: performing modal decomposition, modal screening and dynamic soft threshold processing on the reinforced power pole multi-scale spatio-temporal fusion feature vector and the list of processed power pole multi-modal data to obtain a power pole denoising feature vector; performing key feature focusing operations in spatial dimension and time dimension on the power pole denoising feature vector by using a double attention mechanism to generate the high-precision power pole deformation feature vector; the process of gating mechanism processing and multi-step deformation prediction in S4 comprises: performing deformation feature prediction on the high-precision power pole deformation feature vector by using a power pole deformation gating mechanism to obtain a hidden feature state; constructing multiple prediction branches based on a prediction step to perform multi-step deformation prediction on the hidden feature state to obtain a multi-step deformation prediction feature matrix; performing convolution operation on the multi-step deformation prediction feature matrix to strengthen time sequence correlation, and introducing physical constraints for boundary correction to obtain a final power pole deformation multi-step prediction vector.

2. The deep learning-based utility pole geometric deformation prediction method of claim 1, wherein, the process of data preprocessing comprises: Pruning invalid image samples in the original power pole image data list to obtain a processed power pole image data list; Detecting and pruning abnormal data in the original power pole multi-modal data list and interpolating missing values to obtain a processed power pole multi-modal data list.

3. The deep learning-based utility pole geometric deformation prediction method of claim 1, wherein, The process performed by the power pole feature fusion submodule in S1 includes: Performing element multiplication and element addition operations on the power pole spatial feature vector and the power pole time feature vector to generate interaction features; Concatenating and linearly transforming the interaction features to calculate a power pole feature fusion vector for dynamically adjusting the fusion ratio; Performing weighted fusion and non-linear enhancement on the interaction features based on the power pole feature fusion vector to output a power pole spatio-temporal fusion feature vector.

4. The deep learning-based utility pole geometric deformation prediction method of claim 1, wherein, The process of multi-scale feature extraction and enhancement in S2 includes: Performing convolution operations of different scales on the power pole spatio-temporal fusion feature vector to capture correlations of different ranges to obtain a multi-scale feature tensor; Concatenating and compressing the channel dimension of the multi-scale feature tensor to generate a power pole multi-scale spatio-temporal feature tensor; Performing residual connection on the power pole spatio-temporal fusion feature vector and the power pole multi-scale spatio-temporal feature tensor and performing non-linear transformation on the connected result to generate the reinforced power pole multi-scale spatio-temporal fusion feature vector.

5. The deep learning-based utility pole geometric deformation prediction method of claim 1, wherein, The power pole geometric deformation prediction result includes predicted values of the inclination angle, material strain, and crack width of the power pole at each future time.

Images