Power transmission line multi-parameter integrated monitoring and fault intelligent prediction system and method

Abstract

The present application belongs to the technical field of fault detection, and discloses a power transmission line multi-parameter integrated monitoring and intelligent fault prediction system and method; comprising, acquiring multi-source heterogeneous monitoring data of a target section of a power transmission line; and after spatiotemporal alignment, extracting one-dimensional time sequence features and two-dimensional visual features respectively to obtain multi-modal fusion representation; based on the topological connection relationship and span information of the power transmission line tower, constructing a line topology graph and configuring the multi-modal fusion representation; extracting spatial topological dependency features between adjacent nodes of the power transmission line and time evolution features of the nodes themselves; introducing a power transmission line physical mechanism model as a priori constraint to physically correct the spatial topological dependency features and the time evolution features, and obtaining multi-parameter evolution trends; based on the multi-parameter evolution trends, performing multi-task joint prediction to output the fault risk index and the fault type of the target section; realizing multi-parameter integrated monitoring and intelligent fault prediction of the power transmission line.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and more specifically, to a multi-parameter integrated monitoring and intelligent fault prediction system and method for power transmission lines. Background Technology

[0002] Transmission lines are the backbone of the power system for transmitting electrical energy. They are exposed to complex natural environments for a long time and are susceptible to failures due to factors such as micro-meteorology, topography, and their own aging.

[0003] However, existing methods for monitoring and predicting transmission lines still have significant shortcomings in practical applications. First, most existing systems set independent alarm logic for single parameters (such as excessive temperature or excessive wind deflection angle), ignoring the fact that transmission line faults are often the result of spatiotemporal coupling of multiple parameters. For example, icing faults are not only related to temperature and humidity, but also dynamically intertwined with wind direction, wind speed, and conductor tension. Single-parameter alarms are prone to false alarms or missed alarms. Second, there is a lack of ability to mine the deep correlation features of multi-source heterogeneous data. Monitoring data includes one-dimensional time-series sensor data and two-dimensional image data. Existing methods usually only perform simple data splicing or feature concatenation, failing to effectively eliminate the semantic gap between different modal data, resulting in insufficient feature representation. Finally, purely data-driven prediction models lack the constraints and guidance of physical mechanisms. The operating state of transmission lines follows clear mechanical and electrical laws (such as icing growth models and wind deflection mechanics equations). Existing deep learning models rely only on statistical data patterns. When encountering extreme conditions that have not occurred before, the prediction results are prone to violating physical common sense, resulting in weak model generalization ability and poor prediction reliability.

[0004] Therefore, it is necessary to propose a multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines, comprising: The data acquisition and fusion module is used to acquire multi-source heterogeneous monitoring data of the target section of the transmission line; and after spatiotemporal alignment, it extracts one-dimensional temporal features and two-dimensional visual features respectively, and obtains multimodal fusion representation through cross-modal fusion. The line topology construction and configuration module is used to construct a line topology map based on the topological connection relationship and span information of transmission line towers, and configure multimodal fusion representation; based on the line topology map, it extracts the spatial topological dependency features between adjacent nodes of the transmission line and the temporal evolution features of the nodes themselves. The physical constraint correction module is used to introduce the physical mechanism model of the transmission line as a priori constraint to perform physical consistency correction on the spatial topology dependence characteristics and temporal evolution characteristics, and obtain multi-parameter evolution trends. The fault prediction output module is used to perform multi-task joint prediction based on multi-parameter evolution trends, and output the fault risk index and fault type of the target section.

[0006] Preferably, the step of extracting one-dimensional temporal features and two-dimensional visual features after spatiotemporal alignment, and obtaining a multimodal fusion representation through cross-modal fusion, includes: Multi-source heterogeneous monitoring data includes one-dimensional sensor time-series data and two-dimensional visual image data; among which, the one-dimensional sensor time-series data includes micro-meteorological parameters, conductor tension parameters and insulator leakage current parameters, and the two-dimensional visual image data includes images of towers and conductors covered with ice and images of the surrounding environment. Based on the monitoring timestamp, the time-dimension interpolation alignment of one-dimensional sensor time-series data and two-dimensional visual image data is performed; One-dimensional sensor time-series data is input into a long short-term memory network, and time-series dynamic dependency features are extracted as query vectors. Two-dimensional visual image data is input into a convolutional neural network to extract spatial semantic features as key vectors and value vectors. The attention distribution of the query vector and key vector is applied to the value vector to output a multimodal fusion representation.

[0007] Preferably, the step of constructing a line topology map based on the transmission line tower topology connection relationship and span information, and configuring a multimodal fusion representation, includes: Each tower on the transmission line is used as a graph node, and the conductor connection between adjacent towers is used as an edge; Use the span and elevation difference between adjacent towers as the initial weights of the edges; The multimodal fusion representation corresponding to each tower node is mapped to a node feature vector to complete the construction of the line topology map.

[0008] Preferably, the spatial topological dependency feature is obtained in the following way: In the spatial dimension, a spatial aggregation weight matrix is ​​constructed based on the adjacency matrix and degree matrix of the route topology map, combined with the edge weights that include span and elevation difference. The node feature vectors corresponding to each tower node are input into the graph convolutional network, and the node feature vectors of adjacent tower nodes are weighted and aggregated according to the spatial aggregation weight matrix. During the aggregation process, the information transmission contribution between adjacent nodes is adjusted according to the span and elevation difference, so that micro-meteorological parameters, conductor tension parameters and visual features can be spatially interacted and coupled on the edges of the line topology map, thereby extracting spatial topology dependency features.

[0009] Preferably, the method for obtaining the time evolution features is as follows: In the time dimension, the feature vectors of each tower node after spatial aggregation are arranged according to time steps to form a spatiotemporal feature sequence; The spatiotemporal feature sequence is input into the temporal convolutional network, and a one-dimensional dilated convolution operation is performed along the time axis. By setting convolutional kernels with different dilation rates, the dynamic evolution dependency of multiple parameters at different time scales is captured. At the same time, a causal convolution mechanism is introduced to ensure that the prediction at the current moment depends only on the state at the previous moment. Thus, the temporal evolution characteristics of micrometeorological parameters, conductor tension parameters, and visual features were extracted.

[0010] Preferably, the introduction of a transmission line physical mechanism model as a priori constraint to perform physically consistent corrections on spatial topology dependence characteristics and temporal evolution characteristics, thereby obtaining multi-parameter evolution trends, includes: Spatial topological dependency features and temporal evolution features are input into the prediction network, and the initial prediction parameters are output. Based on the meteorological parameters in the initial prediction parameters, obtain their corresponding theoretical values ​​of physical state; Calculate the physical violation between the initial predicted parameters and the theoretical values ​​of the physical state; A physical loss penalty term is constructed based on the physical violation, and the physical loss penalty term is weighted and fused with the data-driven loss function to form the total loss function. The total loss function is used to backpropagate gradient updates and correct the weight parameters of the prediction network, generating a multi-parameter evolution trend.

[0011] Preferably, the calculation of the physical violation between the initial predicted parameters and the theoretical values ​​of the physical state includes: Obtain the predicted icing thickness and wind deflection angle, calculate the difference between them and the theoretical icing thickness and theoretical wind deflection angle, and identify the sign and absolute value of the difference. The product of the sign of the difference and its absolute value is taken as the basic degree of violation. When the trend of the predicted value is opposite to that of the theoretical value, that is, when the sign of the difference is negative, the product of the basic degree of violation and the trend violation penalty coefficient is recorded as the physical degree of violation. The trend violation penalty coefficient is greater than 1.

[0012] Preferably, the step of performing multi-task joint prediction based on multi-parameter evolution trends to output the fault risk index and fault type of the target segment includes: The multi-parameter evolution trend is input into the shared feature extraction layer, which is composed of alternating fully connected layers and batch normalization layers; The shared feature extraction layer is used to branch the icing prediction branch, the wind deviation prediction branch, and the pollution flashover prediction branch. The system simultaneously outputs the probabilities of icing risk, wind deflection risk, and pollution flashover risk. Combines these with the weighting coefficients of each fault type under the current operating conditions to generate a fault risk index and the corresponding fault type.

[0013] Preferably, the step of combining the weighting coefficients of each fault type under the current operating condition to aggregate and generate a fault risk index and the corresponding fault type includes: The current micro-meteorological conditions are obtained, and the weight coefficients of each branch are dynamically adjusted based on the preset meteorological threshold mapping table. If the current temperature is lower than the preset low temperature threshold and the humidity is higher than the preset high humidity threshold, the weight coefficient of the icing prediction branch is adjusted to the first preset high weight value. If the current wind speed is higher than the preset high wind threshold, the weight coefficient of the wind deflection prediction branch will be adjusted to the second preset high weight value. The risk probabilities of each branch are weighted and summed based on the dynamically adjusted weight coefficients to output the fault risk index, and the branch with the highest risk probability is assigned to the fault type.

[0014] A method for integrated monitoring of multiple parameters and intelligent fault prediction of transmission lines, including: Step S1: Obtain multi-source heterogeneous monitoring data of the target section of the transmission line; Step S2: Spatiotemporally align the multi-source heterogeneous monitoring data, extract one-dimensional temporal features and two-dimensional visual features respectively, and obtain multimodal fusion representation through cross-modal fusion; Step S3: Based on the transmission line tower topology connection relationship and span information, construct the line topology map and configure multimodal fusion representation; Step S4: Based on the line topology map, extract the spatial topological dependency features between adjacent nodes of the transmission line and the temporal evolution features of the nodes themselves; Step S5: Introduce the physical mechanism model of the transmission line as a priori constraint, and make physical consistency corrections to the spatial topology dependence characteristics and temporal evolution characteristics to obtain the multi-parameter evolution trend. Step S6: Perform multi-task joint prediction based on multi-parameter evolution trends, and output the fault risk index and fault type of the target section.

[0015] The technical effects and advantages of the multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines of this invention are as follows: By employing a cross-modal attention mechanism for deep fusion, using temporal dynamic dependency features as the query vector (Q) and visual spatial semantic features as the key vector (K) and value vector (V), and through attention distribution calculation, the adaptive enhancement of temporal features by visual features is achieved, effectively eliminating the semantic gap between different modal data. This enables the model to automatically focus on visual regions that are strongly correlated with the current mechanical and electrical state, and the output multimodal fusion representation contains more comprehensive and accurate line status information.

[0016] By constructing a line topology map that integrates span and elevation difference, and using graph convolutional networks and temporal convolutional networks to extract spatiotemporal features, in the spatial dimension, the contribution of information transmission between adjacent nodes is adjusted according to span and elevation difference, which truly restores the physical transmission process of micrometeorological and mechanical states between adjacent towers; in the temporal dimension, convolutional kernels with different expansion rates are used to capture the dynamic evolution dependence of multiple parameters at different time scales, breaking the limitations of isolated monitoring of single points and single parameters, and fully exploring the spatiotemporal coupling characteristics of the dynamic interweaving of temperature and humidity, wind direction and speed and conductor tension, fundamentally reducing the false alarm and missed alarm rates in complex environments.

[0017] By calculating the physical deviation between the predicted parameters and the theoretical values ​​of the physical state, and especially by applying a trend deviation penalty coefficient to cases where the predicted and theoretical values ​​change in opposite directions (i.e., the difference is negative), a total loss function is constructed that fuses the physical loss penalty term with the data-driven loss. This physical consistency correction mechanism forces the model to follow mechanical and electrical laws while fitting the data, effectively avoiding prediction outputs that violate physical common sense, and greatly enhancing the model's generalization ability and the reliability of prediction results under extreme conditions.

[0018] By adopting a multi-task architecture with a shared feature extraction layer and independent prediction of three branches (icing, wind deflection, and pollution flashover), and dynamically adjusting the weight coefficients based on a meteorological threshold mapping table, the model can automatically focus on the dominant risks according to the current micro-meteorological conditions. This avoids the masking of weak links caused by equalization calculations, resulting in output fault risk indices and fault types that are more in line with actual operating scenarios, significantly improving the accuracy and practicality of fault warnings. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of the multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines of the present invention; Figure 2 This is a schematic diagram of the method for integrated monitoring of multiple parameters and intelligent fault prediction of transmission lines according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1, please refer to Figure 1 and Figure 2In this embodiment of the invention, a multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines includes: The data acquisition and fusion module is used to acquire multi-source heterogeneous monitoring data of the target section of the transmission line; and after spatiotemporal alignment, it extracts one-dimensional temporal features and two-dimensional visual features respectively, and obtains multimodal fusion representation through cross-modal fusion. The line topology construction and configuration module, connected to the data acquisition and fusion module, is used to construct a line topology map based on the transmission line tower topology connection relationship and span information, and configure multimodal fusion representation; based on the line topology map, it extracts the spatial topological dependency features between adjacent nodes of the transmission line and the temporal evolution features of the nodes themselves. The physical constraint correction module, connected to the line topology construction and configuration module and the fault prediction output module, is used to introduce the physical mechanism model of the transmission line as a priori constraint, perform physical consistency correction on the spatial topology dependence characteristics and temporal evolution characteristics, and obtain multi-parameter evolution trends. The fault prediction output module is used to perform multi-task joint prediction based on multi-parameter evolution trends, and output the fault risk index and fault type of the target section.

[0022] Its specific implementation methods include: Step S1: Obtain multi-source heterogeneous monitoring data of the target section of the transmission line; Step S2: Spatiotemporally align the multi-source heterogeneous monitoring data, extract one-dimensional temporal features and two-dimensional visual features respectively, and obtain multimodal fusion representation through cross-modal fusion; Step S3: Based on the transmission line tower topology connection relationship and span information, construct the line topology map and configure multimodal fusion representation; Step S4: Based on the line topology map, extract the spatial topological dependency features between adjacent nodes of the transmission line and the temporal evolution features of the nodes themselves; Step S5: Introduce the physical mechanism model of the transmission line as a priori constraint, and make physical consistency corrections to the spatial topology dependence characteristics and temporal evolution characteristics to obtain the multi-parameter evolution trend. Step S6: Perform multi-task joint prediction based on multi-parameter evolution trends, and output the fault risk index and fault type of the target section.

[0023] Step S2 includes: Multi-source heterogeneous monitoring data includes one-dimensional sensor time-series data and two-dimensional visual image data; among which, the one-dimensional sensor time-series data includes micro-meteorological parameters, conductor tension parameters and insulator leakage current parameters, and the two-dimensional visual image data includes images of towers and conductors covered with ice and images of the surrounding environment. It should be noted that micro-meteorological parameters refer to the meteorological state parameters of the local micro-environment of transmission lines, mainly including ambient temperature, ambient humidity, wind speed, wind direction, atmospheric pressure, and precipitation. These parameters are the direct causes of faults such as icing and wind deflection. Real-time data is collected by micro-meteorological monitoring stations (integrating temperature and humidity sensors, anemometers, barometers, rain gauges, etc.) installed on transmission line towers and uploaded to the monitoring platform via a wireless communication network at a preset sampling frequency (e.g., once per minute), forming one-dimensional sensor time-series data.

[0024] Conductor tension parameters refer to the axial tensile force exerted on the tower suspension point by the transmission conductor under its own weight, icing load, and wind load. Abnormal changes in this parameter directly reflect the risk of mechanical overload of the line (such as icing overload). The conductor tension changes are measured in real time by tension sensors (or tension sensors connected in series within the tension section) installed at the connection between the insulator string and the tower crossarm. The collected data is then signal-conditioned and converted to digital (A / D) values ​​before being uploaded to form one-dimensional sensor time-series data.

[0025] Insulator leakage current parameters refer to the current value flowing along the surface of the insulator in polluted and humid environments. Its amplitude and pulse frequency are key characteristic quantities for measuring the pollution status of the insulator and predicting flashover faults. By installing a leakage current sensor (such as a microcurrent acquisition ring) at the grounding terminal of the insulator string, the microampere to milliampere level current signal flowing through the surface of the insulator is collected in real time. The peak value, average value, and pulse count are extracted and uploaded to form one-dimensional sensor time-series data.

[0026] Pole and conductor icing images refer to two-dimensional visual images that include the shape, thickness, and type of ice accumulation on conductors, as well as the icing status of insulator strings. These images are used to visually reflect the severity of icing on the line. High-definition pan-tilt cameras or fixed bullet cameras mounted on the crossarms of poles capture close-up images of conductors and insulators at regular intervals (e.g., every 10 minutes) or triggered actions. These images are then compressed and uploaded to the monitoring platform to form two-dimensional visual image data.

[0027] Surrounding environment images refer to two-dimensional visual images that include the terrain, vegetation growth, weather visibility (such as heavy fog), and potential external damage (such as construction machinery intrusion) around the tower. These images are used to help assess the impact of the external environment on line safety. Wide-angle cameras installed high on the tower capture images of the line corridor and surrounding environment with a wide field of view, acquiring macroscopic environmental images and uploading them to form two-dimensional visual image data.

[0028] Using the monitoring timestamp as a benchmark, time-dimensional interpolation alignment is performed on one-dimensional sensor time-series data and two-dimensional visual image data. Specifically, due to the significant difference in sampling rate between one-dimensional sensor time-series data (high sampling frequency, such as 1Hz) and two-dimensional visual image data (low sampling frequency, such as 1 frame / 10 minutes), time alignment is required: using the monitoring timestamp as a benchmark, the timestamp of low-frequency two-dimensional visual image data is selected as the alignment target node; for high-frequency one-dimensional sensor time-series data, linear interpolation (or extracting statistical feature values ​​such as mean and variance within the time window) is used to calculate the sensor interpolation at the corresponding time between two adjacent image timestamps, so that one-dimensional time-series data and two-dimensional image data form a strict one-to-one correspondence on the time axis, eliminating semantic misalignment caused by time offset.

[0029] One-dimensional sensor time-series data is input into a Long Short-Term Memory (LSTM) network, and time-series dynamic dependency features are extracted as query vectors. Specifically, the LSTM network employs a two-layer stacked LSTM, with the input layer dimension determined by the feature dimension of the one-dimensional sensor time-series data (e.g., the dimension obtained by concatenating micrometeorological parameters, conductor tension parameters, and insulator leakage current parameters). The hidden layer dimension is set to 128. The activation function is tanh, and the recurrent gate activation is sigmoid. To prevent overfitting, dropout layers (with a dropout rate of 0.2) are added between layers.

[0030] Execution steps: Input the interpolated and aligned one-dimensional sensor time-series data into the LSTM network step by step to capture the temporal dynamic dependency features between temperature and humidity changes and tension fluctuations; take the hidden state output hT of the last time step of the LSTM, and perform linear mapping through a fully connected layer (Linear layer, input dimension 128, output dimension d) to obtain the temporal dynamic dependency feature representation, which is used as the query vector Q of the cross-modal attention mechanism.

[0031] Two-dimensional visual image data is input into a convolutional neural network to extract spatial semantic features as key vectors and value vectors. Specifically, the convolutional neural network uses a lightweight ResNet-18 as the backbone network (removing the original last fully connected classification layer). The input layer receives normalized two-dimensional visual image data (e.g., 224×224×3). Multi-scale features are extracted through four residual convolution stages, and finally, a high-dimensional visual feature vector (dimension 512) is output through a global average pooling layer.

[0032] Execution steps: Input the two-dimensional visual image data into the CNN to extract the two-dimensional spatial semantic feature vector containing the icing morphology and environmental state; input the feature vector into two parallel fully connected layers for mapping: one is mapped to a vector of dimension d as the key vector K, and the other is mapped to a vector of dimension dv as the value vector V.

[0033] By applying the attention distribution of the query vector and key vector to the value vector, visual features are adaptively enhanced to temporal features, resulting in a multimodal fusion representation. Specifically, the query vector Q and key vector K are dot-producted and divided by a scaling factor. After scaling, the data is normalized using the Softmax function to obtain the attention weight distribution of visual features on temporal features: S = Softmax(Q·K). T / The attention weight distribution S is applied to the value vector V, and a weighted summation operation is performed to obtain the fused feature vector F=S·V. The above operation enables the highly concerned physical states in the time series data (such as a sharp increase in tension) to adaptively call the corresponding spatial semantics in the visual image (such as the increased icing of the conductor), eliminate the semantic gap between different modal data, and finally output a multimodal fusion representation with sufficient feature expression.

[0034] Step S3 includes: Each tower on the transmission line is used as a graph node, and the conductor connection between adjacent towers is used as an edge; The span and elevation difference between adjacent towers are used as the initial weights of the edges. It should be explained that the span refers to the horizontal distance between the suspension points of the conductors of two adjacent towers on the transmission line. The span directly determines the conductor's own gravity load, wind load bearing area, and the force span of the conductor when it is covered with ice. The larger the span, the more sensitive the conductor is to the tension transmission of the tower, and the more significant the mechanical response caused by micro-meteorological changes.

[0035] Elevation difference refers to the difference in altitude between the suspension points of the conductors of two adjacent towers. Transmission lines often cross complex terrains such as mountains and hills. Elevation difference causes adjacent towers to be in different micro-meteorological zones (such as higher altitudes being more prone to icing). At the same time, elevation difference will change the static stress distribution and wind deflection trajectory of the conductors, which is a key factor in causing asymmetric coupling of mechanical states.

[0036] The multimodal fusion representation corresponding to each tower node is mapped to a node feature vector to complete the construction of the line topology map.

[0037] The method for obtaining the spatial topology dependency features is as follows: In the spatial dimension, a spatial aggregation weight matrix is ​​constructed based on the adjacency matrix and degree matrix of the route topology graph, combined with edge weights including span distance and elevation difference. Specifically, based on the route topology graph, a basic adjacency matrix A (Aij=1 if nodes i and j are directly connected, otherwise 0) and a degree matrix D (diagonal elements are node degrees) are constructed. Physical priors including span distance Lij and elevation difference ΔHij are introduced to construct an edge weight matrix E. The span distance and elevation difference are then normalized to obtain... ij and Δ ij, and design a weight mapping function (e.g., eij=α·exp(- ij)+β·exp(-Δ (ij), where α and β are adjustable coefficients; the smaller the span and the smaller the elevation difference between adjacent towers, the closer the coupling between micrometeorology and mechanics, and the larger the assigned edge weight eij; multiply the adjacency matrix A and the edge weight matrix E element-wise (A ⊙ E) to obtain the weighted adjacency matrix. Then, the symmetric normalized Laplace matrix is ​​calculated by combining the degree matrix, which finally forms the spatial aggregation weight matrix. .

[0038] The node feature vectors corresponding to each tower node are input into the graph convolutional network, and the node feature vectors of adjacent tower nodes are weighted and aggregated according to the spatial aggregation weight matrix. It needs to be explained that the graph convolutional network uses a two-layer graph convolutional network (2-layer GCN). The input dimension of the first graph convolutional layer is the dimension of the node feature vector (determined by the multimodal fusion representation dimension), and the output dimension is set to 128. The input dimension of the second graph convolutional layer is 128, and the output dimension is set to 64. The ReLU activation function is used between layers to introduce non-linearity, and a Dropout layer (dropout rate of 0.3) is added to prevent overfitting. The network propagation formula is H(l+1)=σ( H(l)W(l)), where H(l) is the feature matrix of the l-th layer, W(l) is the learnable weight parameter, and σ is the activation function.

[0039] During the aggregation process, the contribution of information transmission between adjacent nodes is adjusted according to the span and elevation difference, so that micro-meteorological parameters, conductor tension parameters and visual features can be spatially interacted and coupled on the edge of the line topology map, thereby extracting spatial topological dependence features that reflect the spatial coupling effect of multiple parameters and the state dependence of adjacent nodes.

[0040] Specifically, the multimodal fusion representations corresponding to each tower node are mapped to an initial node feature vector matrix and input into a graph convolutional network. In the first layer of graph convolution, the feature vector of each tower node and its neighboring node feature vectors are weighted and averaged according to the spatial aggregation weight matrix. Since the spatial aggregation weight matrix incorporates span and elevation difference, adjacent nodes that are physically close and have the same elevation contribute more information transmission weight. After two layers of graph convolution, micro-meteorological parameters, conductor tension parameters, and visual features complete spatial interaction and coupling transmission based on physical distance on the edges of the line topology map, outputting the final node feature matrix, which is the extracted spatial topology dependency feature.

[0041] The method for obtaining the time evolution features: In the time dimension, the feature vectors of each tower node after spatial aggregation are arranged according to time steps to form a spatiotemporal feature sequence; The spatiotemporal feature sequence is input into the temporal convolutional network and a one-dimensional dilated convolution operation is performed along the time axis. By setting convolutional kernels with different dilation rates, the dynamic evolution dependence of multiple parameters at different time scales is captured to reflect the temporal lead and lag coupling relationship between temperature and humidity changes and icing growth, and wind speed changes and wind displacement. It should be explained that the temporal convolutional network adopts a residual temporal convolutional network composed of multiple residual blocks. Each residual block contains two layers of one-dimensional dilated causal convolutions. After each convolution layer, weight normalization, ReLU activation function and Dropout layer are connected in sequence. The kernel size is set to 3, and the dilation factor increases exponentially with the network depth. The number of input channels is the dimension of the spatial topological dependency feature (64), the number of hidden channels is 64, and the number of output channels is 64. In order to solve the input and output dimension matching problem, 1x1 convolution is used at the residual connection to align the dimensions.

[0042] At the same time, the introduction of a causal convolution mechanism ensures that the prediction at the current moment depends only on the state at the previous moment; Thus, the temporal evolution characteristics of the dynamic interweaving and evolution of micrometeorological parameters, conductor tension parameters, and visual features over time can be extracted.

[0043] Specifically, in the time dimension, the feature vectors of each tower node after spatial aggregation are arranged according to their corresponding timestamps to form a spatiotemporal feature sequence of shape [time step, number of nodes, feature dimension]. The spatiotemporal feature sequence is input into the TCN network, and one-dimensional convolution operation is performed along the time axis. By setting convolution kernels with different dilation rates, the receptive field is expanded exponentially without increasing the number of parameters and pooling loss. Convolution kernels with small dilation rates capture short-term fluctuations (such as instantaneous changes in conductor tension caused by gusts of wind and leakage current pulses), while convolution kernels with large dilation rates capture long-term evolution (such as the increase in icing thickness caused by a slow decrease in temperature and humidity), thereby extracting the dynamic evolution dependence of multiple parameters at different time scales. Introducing a causal convolution mechanism into convolution operations means that in the convolution calculation at the current time step, only the historical state data of the current time step and the time steps before the current time step are allowed to be used, and zero-padding masking is performed on future time steps. This ensures that the prediction at the current time step depends only on the state at the historical time step, which conforms to the physical time causal relationship and prevents data leakage. After processing by a multi-layer residual temporal convolutional network, the temporal evolution characteristics of micrometeorological parameters, conductor tension parameters, and visual features are extracted and dynamically intertwined over time.

[0044] Step S5 includes: Spatial topological dependency features and temporal evolution features are input into the prediction network, and the initial prediction parameters are output. It should be noted that the prediction network refers to a deep learning regression network that uses a data-driven approach to directly map the future operating state parameters of transmission lines based on extracted deep spatiotemporal features. A multilayer perceptron (MLP) is used as the backbone of the prediction network, which consists of one input layer, two hidden layers, and one output layer. The dimension of the input layer is determined by the concatenation of spatial topological dependency features and temporal evolution features; the dimension of each hidden layer is set to 128, the activation function is ReLU, and Dropout (dropout rate of 0.2) is introduced between layers to prevent overfitting; the output layer dimension is the number of line state parameters to be predicted, and no activation function is set (linear output).

[0045] Based on the meteorological parameters in the initial prediction parameters, obtain their corresponding theoretical values ​​of physical state; Specifically, the spatial topological dependency features output by the graph convolutional network and the temporal evolution features output by the temporal convolutional network are concatenated along the feature dimension to form a complete spatiotemporal joint feature vector. This spatiotemporal joint feature vector is then input into a prediction network (MLP) for forward computation, outputting initial prediction parameters. These parameters include multi-dimensional prediction results, specifically: predicted icing thickness, predicted wind deflection angle, predicted conductor tension, and predicted micrometeorological parameters (temperature, humidity, wind speed, etc.). Meteorologically relevant parameters (predicted temperature, humidity, wind speed, etc.) are extracted from the initial prediction parameters and used as input conditions. These parameters are then substituted into pre-built icing growth mechanics models (such as the Makkonen model) and wind deflection mechanics models (such as the rigid straight bar wind deflection model). Through analytical calculation, the theoretical icing thickness and theoretical wind deflection angle under the current predicted meteorological conditions are obtained, i.e., the theoretical values ​​of the physical state.

[0046] Ice growth mechanics models (such as the Makkonen model) and wind deflection mechanics models are both open-source models and can be used directly.

[0047] Calculate the physical violation between the initial predicted parameters and the theoretical values ​​of the physical state; A physical loss penalty term is constructed based on the physical violation degree. The physical loss penalty term is then weighted and fused with the data-driven loss function to form the total loss function. The total loss function is used to backpropagate gradient updates and correct the weight parameters of the prediction network, so as to constrain the prediction output to approximate the physical law and generate a multi-parameter evolution trend.

[0048] Specifically, based on the initial prediction parameters and real label data (historical measured ice thickness, wind deflection angle, etc.), the mean square error (MSE) is calculated and used as the data-driven loss function L_data. Based on the obtained physical violation degree V_phy, the mean of the physical violation degrees of all samples is calculated, and a physical loss penalty term L_phy is constructed. ∑V_phy; By introducing a physical constraint weight hyperparameter λ (0 < λ < 1), the data-driven loss and the physical loss penalty term are weighted and fused to form the total loss function L_total = (1-λ)L_data + λL_phy.

[0049] The gradient is calculated based on the total loss function L_total. The weight parameters of the prediction network (MLP) are updated and corrected by the backpropagation algorithm. The corrected network fits the statistical law of the data and is constrained by the physical law when making predictions. Finally, the output is a multi-parameter evolution trend that conforms to the common sense of physics.

[0050] The calculation of the physical deviation between the initial predicted parameters and the theoretical values ​​of the physical state includes: Obtain the predicted icing thickness and wind deflection angle, calculate the difference between them and the theoretical icing thickness and theoretical wind deflection angle, and identify the sign and absolute value of the difference. It should be noted that theoretical icing thickness refers to the theoretical value of the icing thickness on the conductor surface calculated by a physical model of icing growth (such as the Makkonen model) based on the conservation laws of fluid mechanics and thermodynamics under given micrometeorological conditions (temperature, humidity, wind speed, liquid water content, etc.). This value reflects the physical upper limit and inevitable trend of icing growth under specific meteorological conditions.

[0051] The theoretical wind deflection angle refers to the theoretical angle at which the insulator string deviates from the vertical position of the conductor, calculated by a wind deflection mechanical model based on the rigid body static equilibrium equation, given wind speed, wind direction, and the conductor's own gravity parameters. This value reflects the mechanical response under the static equilibrium of wind load and gravity load.

[0052] The product of the sign and absolute value of the difference is used as the basic degree of violation. When the trend of the predicted value is opposite to that of the theoretical value, that is, when the sign of the difference is negative, the product of the basic degree of violation and the trend violation penalty coefficient is recorded as the physical degree of violation. The trend violation penalty coefficient is greater than 1 to strengthen the penalty for the prediction result that violates the physical evolution trend.

[0053] It should be noted that the trend violation penalty coefficient is a preset hyperparameter greater than 1, and its specific value is determined through grid search or Bayesian optimization on the validation set. During the model training and validation phase, candidate intervals (such as [1.5, 2.0, 2.5, 3.0]) are traversed, and the value that results in the lowest physical violation rate and the highest prediction accuracy on the validation set is selected as the final penalty coefficient to ensure that sufficient penalty is applied to prediction results that violate the physical trend.

[0054] Specifically, obtain the predicted icing thickness y_ice and wind deflection angle y_wind, as well as the corresponding theoretical physical state values ​​(theoretical icing thickness y_ity, theoretical wind deflection angle y_wty), and calculate the difference between the two respectively: Δice=y_ice-y_ity, Δwind=y_wind-y_wty; Extract the sign function S (S=1 when Δ≥0; S=-1 when Δ<0) and the absolute value |Δ| for each difference; Analyzing the trends of predicted and theoretical values, when the trends are opposite, i.e., the predicted value is less than the theoretical value resulting in a negative difference (S=-1) (e.g., theoretically, icing should increase continuously with weather conditions, but the predicted value shows icing melting and decreasing), the violation of physical evolution laws is extremely serious. In this case, the absolute value is multiplied by the trend violation penalty coefficient μ (μ>1), and the product is recorded as the final physical violation degree V_phy=|Δ|×μ; if the difference is positive (S=1), i.e., the predicted value is larger but does not violate the growth trend, then the physical violation degree V_phy=|Δ|. Through this mechanism, predictions that violate the physical evolution trend (S=-1) will generate a positive penalty loss magnified by μ times, while predictions that are only larger in value but have a correct trend (S=1) will only generate a positive penalty of the normal magnitude, thus achieving a penalty of several times greater than the normal deviation for predictions that violate physical common sense.

[0055] Step S6 includes: The multi-parameter evolution trend is input into the shared feature extraction layer, which consists of alternating fully connected layers and batch normalization layers. This layer is used to extract the underlying hazardous features common to all fault types. Specifically, the generated multi-parameter evolution trend (including spatiotemporal feature vectors corrected for physical consistency) is input into the shared feature extraction layer. The shared feature extraction layer adopts a two-layer structure with alternating connections of "fully connected layer (FC) - batch normalization layer (BN) - ReLU activation function". Specifically, firstly, the first fully connected layer maps the high-dimensional multi-parameter evolution trend to a unified 256-dimensional hidden space. After the first normalization layer, the feature distribution is stabilized and convergence is accelerated. Subsequently, the second fully connected layer further compresses the features to 128 dimensions. After processing by the second normalization layer, this alternating structure can extract the underlying hazardous features common to the three fault types of icing, wind deflection, and pollution flashover (such as common precursors like abnormal tension fluctuations and sharp temperature drops), avoiding redundant learning in each branch and enhancing the generalization ability of feature representation.

[0056] The 128-dimensional shared feature vector is copied from the output of the shared feature extraction layer and fed into three parallel independent prediction branches: the icing prediction branch, the wind deviation prediction branch, and the pollution flashover prediction branch. Each prediction branch consists of a two-layer fully connected network (64 hidden layer dimensions and 1 output layer dimension) and a Sigmoid activation function. Each branch specifically decodes the shared features for a specific fault mechanism, and finally maps the output to the [0,1] interval through the Sigmoid function.

[0057] The three branches simultaneously calculate and output the icing risk probability P_ice, the wind deviation risk probability P_wind, and the pollution flashover risk probability P_flash, respectively.

[0058] The system simultaneously outputs the probabilities of icing risk, wind deflection risk, and pollution flashover risk. Combines these with the weighting coefficients of each fault type under the current operating conditions to generate a fault risk index and the corresponding fault type.

[0059] The process of combining the weighting coefficients of each fault type under the current operating condition to generate the fault risk index and the corresponding fault types includes:

[0060] Obtain the current micro-meteorological conditions and dynamically adjust the weight coefficients of each branch based on the preset meteorological threshold mapping table; If the current temperature is lower than the preset low temperature threshold and the humidity is higher than the preset high humidity threshold, the weight coefficient of the icing prediction branch will be adjusted to the first preset high weight value. If the current wind speed is higher than the preset high wind threshold, the weight coefficient of the wind deflection prediction branch will be adjusted to the second preset high weight value. The risk probabilities of each branch are weighted and summed based on the dynamically adjusted weight coefficients to output the fault risk index, and the branch with the highest risk probability is assigned to the fault type.

[0061] Specifically, the micro-meteorological conditions (temperature, humidity, wind speed) at the current moment are extracted from the multi-parameter evolution trend. Query the preset meteorological threshold mapping table: If the current temperature is lower than the preset low temperature threshold (e.g., 0℃) and the humidity is higher than the preset high humidity threshold (e.g., 85%RH), then adjust the weight coefficient of the icing prediction branch to the first preset high weight value (0.7), and scale the weight of the wind deflection branch and the weight of the pollution flashover branch in a proportional manner (e.g., adjust to 0.15). If the current wind speed is higher than the preset high wind threshold (e.g., 15 m / s), the weight coefficient of the wind deflection prediction branch will be adjusted to the second preset high weight value (0.75), and the weight of the icing branch and the weight of the pollution flashover branch will be scaled proportionally (e.g., adjusted to 0.125). If the above extreme operating conditions are not met, the default equal-weight configuration will be used; Based on the dynamically adjusted weighting coefficients, the risk probabilities of each branch are weighted and summed to output the failure risk index. By comparing the probabilities of icing risk (P_ice), wind deviation risk (P_wind), and pollution flashover risk (P_flash), the fault type (icing, wind deviation, or pollution flashover) corresponding to the branch with the highest risk probability is determined as the fault type of the current target section, thus achieving accurate and targeted early warning of faults.

[0062] It should be noted that the meteorological threshold mapping table is a rule-based lookup table established in advance based on the operation procedures of transmission lines (such as the "Guidelines for Power Grid Anti-icing and Anti-icing Work" and the "Guidelines for Power Grid Wind Prevention Work"), historical fault statistics, and meteorological induction mechanisms. Specifically, micro-meteorological conditions (temperature, humidity, wind speed) are divided into different intervals and mapped to the adjustment strategies of the weight coefficients of each branch. The mapping table stores the specific threshold parameters and the corresponding weight adjustment assignments.

[0063] Both the first and second preset high-weight values ​​are set to constants greater than 0.5 and less than 1. Their specific values ​​are determined through cross-validation optimization on historical fault datasets. When setting these values, it is necessary to ensure that the weight of the dominant fault type has an absolute advantage, while preserving the fault tolerance of other branches to avoid false alarms caused by errors from a single sensor. Specifically:

[0064] The first preset high weight value is set to 0.7. When in a low temperature and high humidity environment, icing is the absolute dominant risk. The weight coefficient of the icing prediction branch is adjusted to 0.7, and the remaining 0.3 is equally divided between wind deviation and pollution flashover (0.15 each).

[0065] The second preset high weight value is set to 0.75. When in a high-wind environment, wind deviation is the absolute dominant risk and develops extremely rapidly. The weight coefficient of the wind deviation prediction branch is adjusted to 0.75, and the remaining 0.25 is equally divided between icing and flashover (0.125 each).

[0066] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines, characterized in that, include: The data acquisition and fusion module is used to acquire multi-source heterogeneous monitoring data of the target section of the transmission line; After spatiotemporal alignment, one-dimensional temporal features and two-dimensional visual features are extracted respectively, and multimodal fusion representation is obtained through cross-modal fusion. The line topology construction and configuration module is used to construct a line topology map based on the topological connection relationship and span information of transmission line towers, and configure multimodal fusion representation; based on the line topology map, it extracts the spatial topological dependency features between adjacent nodes of the transmission line and the temporal evolution features of the nodes themselves. The physical constraint correction module is used to introduce the physical mechanism model of the transmission line as a priori constraint to perform physical consistency correction on the spatial topology dependence characteristics and temporal evolution characteristics, and obtain multi-parameter evolution trends. The fault prediction output module is used to perform multi-task joint prediction based on multi-parameter evolution trends, and output the fault risk index and fault type of the target section.

2. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 1, characterized in that, The process of extracting one-dimensional temporal features and two-dimensional visual features after spatiotemporal alignment, and obtaining a multimodal fusion representation through cross-modal fusion, includes: Multi-source heterogeneous monitoring data includes one-dimensional sensor time-series data and two-dimensional visual image data; among which, the one-dimensional sensor time-series data includes micro-meteorological parameters, conductor tension parameters and insulator leakage current parameters, and the two-dimensional visual image data includes images of towers and conductors covered with ice and images of the surrounding environment. Based on the monitoring timestamp, the time-dimension interpolation alignment of one-dimensional sensor time-series data and two-dimensional visual image data is performed; One-dimensional sensor time-series data is input into a long short-term memory network, and time-series dynamic dependency features are extracted as query vectors. Two-dimensional visual image data is input into a convolutional neural network to extract spatial semantic features as key vectors and value vectors. The attention distribution of the query vector and key vector is applied to the value vector to output a multimodal fusion representation.

3. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 2, characterized in that, The process involves constructing a line topology map based on the transmission line tower topology connection relationships and span information, and configuring a multi-modal fusion representation, including: Each tower on the transmission line is used as a graph node, and the conductor connection between adjacent towers is used as an edge; Use the span and elevation difference between adjacent towers as the initial weights of the edges; The multimodal fusion representation corresponding to each tower node is mapped to a node feature vector to complete the construction of the line topology map.

4. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 3, characterized in that, The method for obtaining the spatial topology dependency features is as follows: In the spatial dimension, a spatial aggregation weight matrix is ​​constructed based on the adjacency matrix and degree matrix of the route topology map, combined with the edge weights that include span and elevation difference. The node feature vectors corresponding to each tower node are input into the graph convolutional network, and the node feature vectors of adjacent tower nodes are weighted and aggregated according to the spatial aggregation weight matrix. During the aggregation process, the information transmission contribution between adjacent nodes is adjusted according to the span and elevation difference, so that micro-meteorological parameters, conductor tension parameters and visual features can be spatially interacted and coupled on the edges of the line topology map, thereby extracting spatial topology dependency features.

5. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 4, characterized in that, The method for obtaining the time evolution features: In the time dimension, the feature vectors of each tower node after spatial aggregation are arranged according to time steps to form a spatiotemporal feature sequence; The spatiotemporal feature sequence is input into the temporal convolutional network, and a one-dimensional dilated convolution operation is performed along the time axis. By setting convolutional kernels with different dilation rates, the dynamic evolution dependence of multiple parameters at different time scales is captured. At the same time, a causal convolution mechanism is introduced to ensure that the prediction at the current moment depends only on the state at the previous moment. Thus, the temporal evolution characteristics of micrometeorological parameters, conductor tension parameters, and visual features were extracted.

6. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 5, characterized in that, The introduction of a transmission line physical mechanism model as a priori constraint performs physical consistency correction on spatial topology dependence characteristics and temporal evolution characteristics to obtain multi-parameter evolution trends, including: Spatial topological dependency features and temporal evolution features are input into the prediction network, and the initial prediction parameters are output. Based on the meteorological parameters in the initial prediction parameters, obtain their corresponding theoretical values ​​of physical state; Calculate the physical violation between the initial predicted parameters and the theoretical values ​​of the physical state; A physical loss penalty term is constructed based on the physical violation, and the physical loss penalty term is weighted and fused with the data-driven loss function to form the total loss function. The weight parameters of the prediction network are then updated and corrected by backpropagation using the total loss function, generating a multi-parameter evolution trend.

7. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 6, characterized in that, The calculation of the physical deviation between the initial predicted parameters and the theoretical values ​​of the physical state includes: Obtain the predicted icing thickness and wind deflection angle, calculate the difference between them and the theoretical icing thickness and theoretical wind deflection angle, and identify the sign and absolute value of the difference. The product of the sign of the difference and its absolute value is taken as the basic degree of violation. When the trend of the predicted value is opposite to that of the theoretical value, that is, when the sign of the difference is negative, the product of the basic degree of violation and the trend violation penalty coefficient is recorded as the physical degree of violation. The trend violation penalty coefficient is greater than 1.

8. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 7, characterized in that, The multi-task joint prediction based on multi-parameter evolution trends outputs the fault risk index and fault type of the target segment, including: The multi-parameter evolution trend is input into the shared feature extraction layer, which is composed of alternating fully connected layers and batch normalization layers; The shared feature extraction layer is used to branch the icing prediction branch, the wind deviation prediction branch, and the pollution flashover prediction branch. The system simultaneously outputs the probabilities of icing risk, wind deflection risk, and pollution flashover risk. Combines these with the weighting coefficients of each fault type under the current operating conditions to generate a fault risk index and the corresponding fault type.

9. The multi-parameter integrated monitoring and intelligent fault prediction system for transmission lines according to claim 8, characterized in that, The process of combining the weighting coefficients of each fault type under the current operating conditions to generate a fault risk index and corresponding fault types includes: The current micro-meteorological conditions are obtained, and the weight coefficients of each branch are dynamically adjusted based on the preset meteorological threshold mapping table. If the current temperature is lower than the preset low temperature threshold and the humidity is higher than the preset high humidity threshold, the weight coefficient of the icing prediction branch is adjusted to the first preset high weight value. If the current wind speed is higher than the preset high wind threshold, the weight coefficient of the wind deflection prediction branch will be adjusted to the second preset high weight value. The risk probabilities of each branch are weighted and summed based on the dynamically adjusted weight coefficients to output the fault risk index, and the branch with the highest risk probability is assigned to the fault type.

10. A method for integrated monitoring of multiple parameters and intelligent fault prediction of transmission lines, characterized in that, The transmission line multi-parameter integrated monitoring and intelligent fault prediction system according to any one of claims 1-9 includes: Step S1: Obtain multi-source heterogeneous monitoring data of the target section of the transmission line; Step S2: Spatiotemporally align the multi-source heterogeneous monitoring data, extract one-dimensional temporal features and two-dimensional visual features respectively, and obtain multimodal fusion representation through cross-modal fusion; Step S3: Based on the transmission line tower topology connection relationship and span information, construct the line topology map and configure multimodal fusion representation; Step S4: Based on the line topology map, extract the spatial topological dependency features between adjacent nodes of the transmission line and the temporal evolution features of the nodes themselves; Step S5: Introduce the physical mechanism model of the transmission line as a priori constraint, and make physical consistency corrections to the spatial topology dependence characteristics and temporal evolution characteristics to obtain the multi-parameter evolution trend. Step S6: Perform multi-task joint prediction based on multi-parameter evolution trends, and output the fault risk index and fault type of the target section.

Images