Power transmission line bolt state real-time monitoring method and system based on intelligent monitoring

Abstract

The invention discloses a power transmission line bolt state real-time monitoring method based on intelligent monitoring, and the method comprises the steps: deploying a piezoelectric-piezoresistive composite sensing array integrated with a temperature-strain compensation unit on the surface of a power transmission line bolt, capturing the radial and axial strain signals of the bolt, carrying out the feature decoupling and information gain of the strain signals, and carrying out the real-time monitoring of the state of the power transmission line bolt. Obtaining a structured feature vector; building a bolt state abnormity identification model based on comparative transfer learning, carrying out domain adversarial training, and carrying out cross-working-condition bolt state classification by taking the structured feature vector as input; performing online learning and dynamic updating on the bolt state abnormity identification model through a hierarchical edge computing architecture; and establishing a bolt health assessment system based on the minimum path entropy, and performing probability inference on the evolution process of the bolt state. According to the scheme, the high-precision synchronous measurement of the radial and axial strain of the bolt is realized through the design of the piezoelectric-piezoresistive composite sensing array and the arrangement of the orthogonal electrodes in combination with alternating polarization processing and a temperature compensation mechanism.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power transmission line operation and maintenance, and particularly relates to a power transmission line bolt state real-time monitoring method and system based on intelligent monitoring. BACKGROUND

[0002] As a key fastening component, the state of the power transmission line bolt directly affects the safety and reliability of line operation. At present, artificial inspection, ultrasonic detection and strain monitoring methods are mainly used to detect the bolt state. Among them, the artificial inspection method is time-consuming and labor-intensive, and is easily affected by subjective factors; although the ultrasonic detection can reflect the change of bolt pretightening force, the measurement accuracy is easily disturbed by environmental noise; the monitoring method based on strain sensing has problems such as sensor damage, signal drift and insufficient temperature compensation in actual application.

[0003] In addition, the existing bolt state monitoring system mainly relies on preset thresholds for judgment, and lacks systematic analysis of the multi-dimensional stress characteristics of the bolt. Under complex external environment, due to the coupling of factors such as vibration, temperature and stress, a single monitoring index is difficult to accurately reflect the actual state of the bolt. At the same time, the traditional state recognition algorithm is sensitive to environmental changes, and is difficult to adapt to the monitoring needs under different working conditions. SUMMARY

[0004] In view of the above problems, the purpose of the present application is to provide a power transmission line bolt state real-time monitoring method and system based on intelligent monitoring.

[0005] The specific technical scheme for realizing the purpose of the present application is as follows:

[0006] A power transmission line bolt state real-time monitoring method based on intelligent monitoring, comprising the following steps:

[0007] Deploying a piezoelectric-piezoresistive composite sensing array integrated with temperature-strain compensation units on the surface of the power transmission line bolt to capture the radial and axial strain signals of the bolt;

[0008] Using orthogonal variational modal decomposition and characteristic entropy matrix algorithm to decouple the characteristics and gain information of the strain signals, and obtaining a structured feature vector;

[0009] Constructing a bolt state anomaly recognition model based on comparative transfer learning and performing domain adversarial training, taking the structured feature vector as input, and performing cross-condition bolt state classification;

[0010] Through a hierarchical edge computing architecture, the bolt state anomaly recognition model is subjected to online learning and dynamic updating;

[0011] Establishing a bolt health evaluation system based on minimum path entropy to probabilistically infer the evolution process of the bolt state.

[0012] Compared with the prior art, the application has the beneficial effects that:

[0013] The scheme of the application first realizes high-precision synchronous measurement of bolt radial and axial strain by piezoelectric-piezoresistive composite sensing array design and orthogonal electrode arrangement, combined with alternating polarization processing and temperature compensation mechanism, solves the problem that single sensing mode cannot comprehensively obtain strain information; secondly, a method combining composite modal decomposition and characteristic entropy matrix is adopted, linear correlation between modes is eliminated through orthogonal projection decoupling algorithm, efficient characteristic decoupling and information gain of strain signal are realized, and the characteristic representation ability is improved; thirdly, a bolt state abnormality recognition model is constructed based on comparative transfer learning, domain-invariant features are obtained through domain adversarial generation network, combined with twin network comparative learning strategy and conditional entropy minimization, accurate recognition of bolt state under cross-condition scene is realized; finally, through hierarchical edge computing architecture and incremental learning mechanism, combined with state evolution analysis based on minimum path entropy, online updating of the model and quantitative evaluation of state degradation risk are realized, and the adaptability and prediction ability of the bolt state monitoring system are improved.

[0014] The application will be further described below in conjunction with the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The flowchart of the real-time monitoring method for the bolt state of the power transmission line based on intelligent monitoring.

[0016] Figure 2 The strain signal characteristic decoupling flowchart of the real-time monitoring method for the bolt state of the power transmission line based on intelligent monitoring.

[0017] Figure 3 The bolt state abnormality recognition model learning flowchart of the real-time monitoring method for the bolt state of the power transmission line based on intelligent monitoring. DETAILED DESCRIPTION

[0018] EMBODIMENT

[0019] The technical solutions in the embodiments of the application will be clearly and completely described below in conjunction with the drawings in the embodiments of the application. The described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0020] As shown herein and in the claims, the words "a," "an," "one," and / or "the" as used herein are not intended to be singular only but include plural, unless otherwise indicated. Generally, the terms "including," "includes," "comprising," "comprises," and the like are used in the sense of "including but not limited to."

[0021] The relative arrangement of components and steps, numerical expressions, and numerical values set forth in the examples are not intended to limit the scope of the application unless otherwise specifically stated. It is to be understood that the drawings are not necessarily to scale, and that, unless otherwise noted, specific numerical values disclosed herein are not intended to limit the scope of the application. Techniques, methods, and devices known to those of ordinary skill in the relevant art can not be discussed in detail but can be assumed by those skilled in the art to be within the scope of the present application. In the examples shown and discussed herein, any specific values should be interpreted as merely illustrative and not as a limitation on the scope of the exemplary embodiments. Thus, other examples of the exemplary embodiments can have different values. It is to be noted that like numbers and letters refer to like elements throughout the several views of the drawings, and that the use of or insertion of a reference number and letter, once defined in one view, need not be discussed further in subsequent views.

[0022] In conjunction with Figure 1 A real-time monitoring method for the state of transmission line bolts based on intelligent monitoring, comprising the following steps:

[0023] Step 1, deploy a piezoelectric-piezoresistive composite sensing array integrated with temperature-strain compensation units on the surface of the transmission line bolt to capture radial and axial strain signals of the bolt, including:

[0024] Step 1-1, prepare a piezoelectric-piezoresistive composite sensing array, which is composed of a plurality of piezoelectric-piezoresistive composite sensing units, which are prepared by combining a piezoelectric ceramic substrate and a piezoresistive film layer;

[0025] In the preparation of the piezoelectric-piezoresistive composite sensing array, lead zirconate titanate (PZT) is selected as the piezoelectric ceramic substrate material, which has excellent piezoelectric performance and mechanical strength, and can convert mechanical strain into electric charge signals. The piezoresistive film layer is made of nickel-chromium (NiCr) alloy, which has a stable strain sensitivity coefficient, excellent linear response, and good repeatability, and can convert mechanical strain into resistance change, especially suitable for capturing small mechanical strain.

[0026] In the specific preparation process, first, a piezoelectric ceramic substrate is prepared by a solid phase sintering method, and then a piezoresistance thin film layer is deposited on the surface of the piezoelectric ceramic substrate by a magnetron sputtering process to form a single piezoelectric-piezoresistance composite sensing unit. A plurality of sensing units are prepared according to a predetermined arrangement mode to form a piezoelectric-piezoresistance composite sensing array.

[0027] Step 1-2, electrodes are prepared on the piezoelectric ceramic substrate along the first and second orthogonal directions, and the piezoresistance thin film layer is arranged in a cross-grid manner;

[0028] In the specific implementation, first, the piezoelectric ceramic substrate is surface pretreated; a first orthogonal direction electrode pattern is prepared on the surface of the piezoelectric ceramic substrate by a photolithography process; a second orthogonal direction electrode pattern is prepared according to the vertical orthogonal principle based on the first orthogonal direction electrode pattern; and the prepared electrode pattern is detected to ensure the integrity and conductivity of the electrode pattern. The electrode pattern is used to collect the charge signals generated by the piezoelectric effect.

[0029] Then, the piezoresistance thin film layer is arranged in a cross-grid manner, that is, the NiCr piezoresistance thin film is arranged along the orthogonal direction corresponding to the electrode pattern to form a grid structure for collecting the resistance change signals in the two orthogonal directions;

[0030] Step 1-3, the piezoelectric-piezoresistance composite sensing unit is subjected to an alternating polarization treatment:

[0031] The piezoelectric-piezoresistance composite sensing unit is subjected to an alternating polarization treatment under a high temperature environment, an alternating electric field is applied in the first and second orthogonal directions to polarize the grains inside the piezoelectric ceramic along the orthogonal directions to form an ordered polarization direction; when the bolt is subjected to a mechanical strain, the charge signals generated by the piezoelectric ceramic and the resistance change signals of the piezoresistance thin film are effectively transmitted through the interface coupling effect to form a complete strain-electric signal conversion channel; after the polarization is completed, annealing treatment is performed to stabilize the polarization effect;

[0032] Step 1-4, a temperature sensor is integrated on the surface of the piezoelectric-piezoresistance composite sensing unit to establish a strain-temperature mapping model for temperature compensation;

[0033] Specifically, the charge signals of the piezoelectric ceramic and the resistance change signals of the piezoresistance thin film are both affected by the change of the environmental temperature, and the temperature influence needs to be compensated, including: a known standard strain is applied to the sensing unit to obtain the output characteristics of the two kinds of electric signals under different temperature conditions to establish a strain-temperature mapping model. The model is used for actual measurement, and the two kinds of electric signals are compensated according to the real-time data of the temperature sensor to ensure the accuracy of the strain measurement.

[0034] Step 1-5, using a cross-scale signal conditioning circuit, separating the orthogonal components of the piezoelectric-piezoresistive composite sensor array output signal, obtaining temperature-compensated radial and axial strain signals:

[0035] Since the piezoelectric-piezoresistive composite sensor array outputs a composite electrical signal containing radial and axial strain information, it is necessary to separate and condition the signal to obtain strain time-domain signals reflecting the bolt state. Specifically, a cross-scale signal conditioning circuit is designed, including a low-noise amplifier, a digital filter and a signal demodulation module; first, the composite electrical signal is processed by orthogonal demodulation algorithm to separate and obtain the radial and axial orthogonal components; then the separated orthogonal components are subjected to low-noise amplification, digital filtering and other signal conditioning; then the strain-temperature mapping model is used for temperature compensation, and finally the time-domain signal sequence representing the radial and axial strain state of the bolt is obtained.

[0036] Generally speaking, existing bolt monitoring sensors often use a single sensing method, which can only obtain single-direction strain information, and have signal interference and temperature drift problems, making it difficult to accurately reflect the actual state of the bolt; Step 1 of the present scheme designs a piezoelectric-piezoresistive composite sensor array, which combines PZT piezoelectric ceramic and NiCr piezoresistive film, first proposes orthogonal electrode arrangement and cross-grid piezoresistive layer design, and uses alternating polarization treatment to form a strain coupling channel, while establishing a strain-temperature mapping model and a cross-scale signal conditioning circuit to separate orthogonal components. Through the synergistic effect of the above technical solutions, high-precision synchronous measurement of radial and axial strain of the bolt is realized. The sensor array has high strain sensitivity, excellent temperature stability and good signal separation effect, providing a reliable data basis for subsequent bolt state recognition.

[0037] Step 2, combining Figure 2 , using orthogonal variational modal decomposition and characteristic entropy matrix algorithm, decoupling the characteristics and gaining information of the strain signal, to obtain a structured feature vector, including:

[0038] Step 2-1, constructing a composite modal decomposition operator, performing multi-scale decomposition on the strain signal through joint transformation of wavelet transform and empirical mode decomposition, to obtain intrinsic modal components;

[0039] Specifically, a wavelet basis function suitable for strain characteristics is selected to establish a multi-scale wavelet transform framework; discrete wavelet transform is performed on the time-domain signal sequence to obtain approximate coefficients and detail coefficients of multi-resolution decomposition; based on the empirical mode decomposition algorithm, intrinsic mode functions (IMF) are extracted from the transformed coefficients; an adaptive screening criterion is designed to evaluate the oscillation characteristics and physical meaning of each IMF, and modal components with characteristic representation ability are selected; finally, a set of intrinsic modal components with multi-scale characteristics are generated through joint transformation reconstruction.

[0040] Step 2-2, calculate the complexity features of the intrinsic modal components according to the information theory entropy gain criterion, and construct a feature entropy matrix:

[0041] Firstly, select information entropy, fuzzy entropy, sample entropy and other multi-dimensional information theory entropy evaluation indexes; based on the maximum mutual information criterion, evaluate the representation ability of different entropy indexes for the strain state features, and determine the weights of each index through a weight dynamic allocation algorithm; then calculate the entropy values of each type for each intrinsic modal component, and quantify the complexity, randomness and irregularity of the signal; finally, integrate the comprehensive complexity features to generate a multi-dimensional feature entropy matrix, wherein each element represents the complexity feature vector of the intrinsic modal component;

[0042] Step 2-3, perform eigenvalue decomposition and orthogonal projection on the feature entropy matrix using an orthogonal projection decoupling algorithm:

[0043] A covariance matrix of the feature entropy matrix is constructed to analyze the internal correlation structure between the modal components; the covariance matrix is subjected to eigenvalue decomposition to extract the eigenvectors and eigenvalues; based on the eigenvectors, an orthogonal projection transformation matrix is designed, and through a Schmidt orthogonalization process, an orthogonal basis for modal decoupling is established; the orthogonal projection transformation matrix is used to map the feature entropy matrix to an orthogonal feature space, eliminating the linear correlation between the modes, and obtaining the decoupled feature matrix.

[0044] Step 2-4, select a subset of feature vectors associated with the bolt state through the minimum redundancy maximum correlation criterion:

[0045] The mutual information method is used to construct a correlation measure index between the feature vectors and the bolt state, and the Pearson correlation coefficient is used to construct a redundancy measure index between the feature vectors; based on the minimum redundancy maximum correlation criterion, an incremental search strategy is used to iteratively select feature vectors to obtain an optimal feature vector subset.

[0046] Step 2-5, reconstruct the feature vector subset to generate a multi-dimensional decoupled structured feature vector.

[0047] That is, the tensor decomposition method is used to reconstruct the feature vector subset to generate a multi-dimensional decoupled structured feature vector;

[0048] Preferably, the existing bolt state monitoring method often directly uses the original strain signal for feature extraction, which is difficult to effectively process the nonlinear characteristics and modal coupling problems of the signal, resulting in insufficient feature representation capability. This step proposes a method combining composite modal decomposition and feature entropy matrix. The intrinsic modal component is extracted by joint transformation of wavelet transform and empirical mode decomposition. The feature entropy matrix is constructed based on the information entropy gain criterion, and the orthogonal projection decoupling algorithm is used to eliminate the linear correlation between modes. Finally, the feature vector is selected by the minimum redundancy maximum correlation criterion. Through the synergistic effect of the above technical solutions, efficient feature decoupling and information gain of the strain signal are realized. This method significantly improves the feature representation capability and discrimination, laying a foundation for accurate identification of bolt state.

[0049] Step 3, constructing a bolt state anomaly recognition model based on comparative transfer learning and performing domain adversarial training, taking the structured feature vector as input, and performing cross-condition bolt state classification, including:

[0050] Step 3-1, organizing bolt state samples under different conditions in time sequence to construct a cross-condition bolt state sample set;

[0051] Specifically, the bolt state classification results are labeled, and original bolt strain signal samples under different load and temperature conditions are collected. According to the feature decoupling process of steps 2-1 to 2-5, the strain signal original sample is converted into a structured feature vector. The structured feature vector and the corresponding bolt state classification label are organized in time sequence, and a hierarchical sampling strategy is used to divide the data set into a training set and a validation set for model training and performance evaluation, respectively, to finally form a cross-condition bolt state sample set mixed with source and target domains.

[0052] Step 3-2, constructing a bolt state anomaly recognition model, using a domain adversarial generation network architecture, including a feature extraction unit, a domain discrimination unit, and a time sequence state classification unit, wherein the feature extraction unit obtains domain invariant feature representation based on a gradient reversal layer;

[0053] Specifically, the feature extraction unit adopts a multi-layer perception structure, including: an input layer receiving a structured feature vector; two hidden layers using LeakyReLU activation functions for nonlinear transformation; a batch normalization layer after each layer to improve training stability; a gradient inversion layer connected after the output layer to send the features to the domain discrimination unit for adversarial learning, realizing domain-invariant feature extraction. The domain discrimination unit is designed as a discrimination network structure, including: an input layer connected to the gradient inversion layer, receiving the feature representation of the feature extraction unit; two fully connected networks for feature transformation; the LeakyReLU activation function and Dropout regularization are used in the intermediate layer; the output layer uses the Sigmoid function to judge the domain attribute of the features. The time series state classification unit adopts a time series attention network structure, specifically including: receiving the domain-invariant features output by the feature extraction unit; a time series feature extraction layer to capture state sequence features; an attention layer to calculate feature importance weights; a fully connected layer and a softmax function to output timestamped bolt state classification results.

[0054] In addition, the design of the gradient inversion layer includes: keeping the feature representation unchanged in forward propagation; performing sign inversion and amplitude scaling on the gradient during back propagation; dynamically controlling the domain invariance constraint strength through an adjustable domain adaptive strength hyperparameter, which is self-adaptively adjusted with the training process. The network architecture design realizes efficient extraction of domain-invariant features through the synergistic effect of the feature extraction unit, the domain discrimination unit, and the time series state classification unit, providing a reliable feature basis for cross-condition bolt state recognition.

[0055] Step 3-3, through domain adversarial training, using adversarial loss function and orthogonal regularization constraint, reducing the difference between the feature distributions of different domains, obtaining domain adaptive features;

[0056] Batch sampling and organizing source domain and target domain data, using binary classification cross-entropy to calculate domain discrimination loss, and using multi-classification cross-entropy to calculate classification loss.

[0057] During parameter updating, the gradient direction of the domain discrimination loss is reversed through the gradient inversion layer and its amplitude is modulated, so that the feature extraction unit is simultaneously constrained by the gradients of the two losses, while the domain discrimination unit and the classification unit update parameters based on their respective losses. During training, the feature distribution distance between the source domain and the target domain is measured by the maximum average difference, and when the distribution distance increases, the domain adaptive strength is increased, and vice versa.

[0058] Step 3-4, using a twin network contrast learning strategy, through similarity measurement between samples and interval maximization criteria, to learn bolt state discrimination features;

[0059] Further, the feature extraction unit is copied to construct a twin branch, the same state and different state bolt sample pairs in the training set are input into the twin branch, and a cosine similarity is used to calculate the similarity score of the feature representation; a contrast loss function is constructed based on the triplet margin criterion, the similarity between the anchor sample and the positive sample is maximized, and the similarity between the anchor sample and the negative sample is minimized, so as to maximize the interval and optimize the class separability of the feature space. The contrast learning strategy significantly improves the discriminability of the features and enhances the model's ability to recognize similar states under different working conditions.

[0060] The capture 3-5 optimizes the discriminability of the classification features based on the minimum conditional entropy and the interval learning theory, and generates a classification decision boundary;

[0061] The conditional probability distribution of the target domain samples on different categories is calculated, the conditional entropy loss is constructed based on the minimum conditional entropy principle, the interval learning strategy is used to increase the separation interval of the features of different categories of samples, and the interval loss is constructed by combining the intra-class compactness constraint and the inter-class separation constraint. By minimizing the conditional entropy loss and maximizing the interval loss, the classification decision boundary is optimized.

[0062] Step 3-6, according to the similarity of the feature distribution of the source domain and the target domain, adaptively adjust the classification decision threshold:

[0063] Specifically, the distribution distance of the feature representation of the source domain and the target domain is calculated, and an adaptive adjustment function of the classification decision threshold is constructed based on the distribution distance; when the distribution distance is large, a higher classification threshold is used to improve the classification reliability, and when the distribution distance is small, a lower classification threshold is used to improve the classification efficiency. The adaptive threshold adjustment mechanism combines the minimum conditional entropy and the interval learning strategy, effectively balances the accuracy and adaptability of classification, and improves the recognition performance of the model under different distribution differences.

[0064] Step 3-7, repeat the training process until the bolt state anomaly recognition model converges;

[0065] Wherein, the domain discrimination loss, the classification loss, the contrast loss, the conditional entropy loss and the interval loss constitute the total loss function; the total loss function value is calculated on the validation set, including the domain discrimination loss, the classification loss, the contrast loss, the conditional entropy loss and the interval loss; when the total loss function value of continuous multiple iterations changes less than a preset threshold, it is determined that the model converges, and the training process is completed.

[0066] Preferably, the existing bolt state recognition method mainly models single working condition, and it is difficult to process the difference of data distribution under different working conditions, resulting in insufficient generalization performance of the model under cross-working condition scene. The bolt state abnormality recognition method based on comparative transfer learning is proposed in step 3, the domain-invariant feature representation is obtained through the domain adversarial generative network architecture, the feature discrimination ability is enhanced by combining the twin network comparative learning strategy, and the classification decision boundary is optimized by using the conditional entropy minimization and interval learning theory, and the adaptive threshold adjustment mechanism is introduced to cope with the distribution difference. Through the synergistic effect of the above technical solutions, the accurate recognition of the bolt state under the cross-working condition scene is realized, and the problem of insufficient model generalization is effectively solved. The method significantly improves the cross-domain adaptability and classification reliability of the bolt state recognition, and provides strong support for the engineering application of bolt state monitoring.

[0067] Step 4, combining Figure 3 Through the hierarchical edge computing architecture, the bolt state abnormality recognition model is learned and dynamically updated online, including:

[0068] Step 4-1, the hierarchical edge computing architecture is divided into a perception layer, an edge computing layer and a cloud management layer, and a lightweight abnormality recognition model is deployed in the edge computing layer:

[0069] Among them, the perception layer is responsible for data acquisition and preprocessing, the edge computing layer executes lightweight model inference, and the cloud management layer is responsible for global model management and optimization, and defines the standardized data interface protocol between layers.

[0070] In addition, the feature extraction unit and the time sequence state classification unit are extracted from the bolt state abnormality recognition model, and the core network structure of the model is retained; based on sensitivity analysis and channel pruning strategy, the core network structure is pruned and quantized, the model parameter size and calculation complexity are reduced, a lightweight version suitable for edge computing layer is constructed, and a model performance reduction tolerance threshold and a traceable model compression record are set; a multi-scale feature adaptive compression strategy is designed, a learnable feature selection module is used to retain key identification information of the model, and a feature importance evaluation index and a feature reconstruction verification mechanism are established; an adaptive inference engine is configured for the lightweight abnormality recognition model, a resource occupation monitoring mechanism is established, dynamic regulation of accuracy and efficiency is realized, and an inference state snapshot is saved to support state rollback. The lightweight model design significantly reduces the resource consumption of edge computing while ensuring the core identification performance.

[0071] Step 4-2, constructing a model parameter incremental learning unit in the edge computing layer, setting a learning trigger mechanism based on classification loss, and triggering the online learning process when the classification loss of the new sample exceeds the learning threshold;

[0072] Specifically, the incremental learning unit includes a parameter gradient calculation module, a parameter updating module, and a parameter verification module; a learning threshold is adaptively calculated based on the distribution characteristics of historical samples, and a dynamic adjustment mechanism is set; a multi-level classification loss evaluation system is constructed, and a multi-classification cross-entropy is used to calculate the classification loss of a single sample; when the sample loss exceeds a first threshold, the average loss and the standard deviation of the batch sample set are further calculated; if the batch statistics exceeds a second threshold and the time series cumulative loss based on the exponential moving average shows an upward trend, online learning is triggered; a progressive learning strategy is designed to trigger parameter updates of different intensities according to the loss level; a parameter update effectiveness verification mechanism is established to ensure the stability of incremental learning.

[0073] The first threshold is 2 times the average value of the historical sample classification loss; the second threshold is the average loss of the last N samples plus 3 times the standard deviation. This multi-level triggering mechanism avoids frequent updates of the model and improves the reliability of online learning.

[0074] Step 4-3, the local model of the edge node adopts a parameter-level incremental learning strategy, and the local model parameter incremental information is transmitted to the cloud management layer:

[0075] Further, the parameter-level incremental learning includes: calculating the parameter gradient for the new sample, setting an adaptive learning rate adjustment strategy, and updating differentially according to the sensitivity of different parameter layers; introducing a momentum term and a weight decay term, setting a gradient accumulation threshold, and controlling the parameter update amplitude; using a hierarchical parameter pruning method, sparse processing of non-critical parameters to reduce computational overhead; setting parameter constraints to maintain the stability of the model structure and avoid parameter update divergence. This incremental learning strategy realizes efficient updating of model parameters while ensuring the stability of model performance;

[0076] An encryption transmission mechanism is used in the process of transmitting the local model parameter incremental information to the cloud management layer to protect the parameter incremental information, including block encryption and digital signature of the parameter data; setting a transmission bandwidth adaptive control strategy to dynamically adjust the data transmission rate according to the network state; constructing a parameter compression encoding scheme to reduce the amount of transmission data; establishing a transmission confirmation and retransmission mechanism to ensure the integrity of the parameter incremental information; implementing the breakpoint resume function to support batch transmission of parameter incremental information.

[0077] Step 4-4, the cloud management layer aggregates the local model increments of each edge node and reconstructs the global model, uses a weighted average strategy to aggregate the local model increments, and the weights are determined according to the sample quantity and model performance of each node; setting an anomaly detection mechanism to filter abnormal local model updates; performing incremental aggregation verification to ensure the convergence of the global model; using a progressive model reconstruction method to maintain the consistency of the model structure.

[0078] Step 4-5, the reconstructed global model parameters are distributed to each edge computing node to complete the hierarchical collaborative update of the model;

[0079] Generally, a batch-by-batch and hierarchical parameter distribution strategy is adopted to ensure the synchronous update of each node model; a parameter version control mechanism is established to track the model update history; an update confirmation mechanism is set to ensure the consistency of the global model at each node; an incremental update rollback function is implemented to support model recovery in abnormal situations.

[0080] This step adopts a hierarchical edge computing architecture, which realizes efficient online learning and dynamic updating of the model through the cooperation of the perception layer, the edge computing layer, and the cloud management layer. The lightweight model design of the edge computing layer adopts sensitivity analysis and channel pruning strategies to reduce the model complexity while retaining the core recognition ability; the learning trigger mechanism based on classification loss realizes precise control of model updating; the parameter-level incremental learning strategy ensures efficient evolution of the model through differential update and sparsification; the model aggregation and reconstruction mechanism of the cloud management layer realizes collaborative optimization of distributed nodes. Through the organic combination of the above technical solutions, the real-time performance, adaptability, and intelligent evolution capability of the bolt condition monitoring system are improved.

[0081] Step 5, a bolt health evaluation system based on minimum path entropy is established to probabilistically infer the evolution process of the bolt state, including:

[0082] Step 5-1, arrange the bolt state classification results in chronological order to form a state evolution sequence;

[0083] Specifically, the state classification results (including normal, loose, missing, etc. state labels) and their corresponding timestamp information at each time are obtained from the bolt state anomaly recognition model, and are sorted in chronological order; duplicate data points and obviously abnormal state labels (such as frequent jumps within a short period of time) are removed; and finally a bolt state evolution sequence is generated.

[0084] Step 5-2, based on the state evolution sequence, a state transition probability matrix is established to quantify the transition probability between adjacent states:

[0085] Discretize and encode different state types in the state evolution sequence to construct a unique and mutually exclusive state space; use the sliding window method to count the number of transitions between adjacent states in the state evolution sequence to establish a state transition frequency matrix based on time series; based on the state transition frequency matrix, use the maximum likelihood estimation method to calculate the transition probability between states, and normalize to obtain the state transition probability matrix; verify that the row sum of the transition probability matrix is 1 and the Markov chain steady-state distribution test to ensure the effectiveness of the state transition probability matrix.

[0086] Step 5-3: Based on the state transition probability matrix, construct a minimum path entropy evaluation model. By calculating the probability entropy of different state evolution paths, identify and quantify the potential evolution paths of the bolt state.

[0087] Based on the state transition probability matrix, a path transition graph is constructed to represent the possible evolution paths of the bolt state. A path probability calculation method is defined, which calculates the path transition probability by the probability value at the corresponding position in the state transition probability matrix and performs path length standardization. A path entropy calculation criterion is designed, which quantifies the uncertainty of path evolution after standardization based on Shannon's entropy formula in information theory. Based on the principle of minimum path entropy, the path sequence with the minimum entropy value is selected as the most likely potential evolution path of the bolt state by comparing the entropy values ​​of different paths.

[0088] Step 5-4: Based on the evolution path quantification results of the minimum path entropy assessment model, quantitatively analyze the risk of bolt condition degradation.

[0089] The evolution path quantification results obtained from the minimum path entropy assessment model are subjected to probability normalization to ensure that the sum of the probability distributions of each state is 1. Based on the state transition probability and time series characteristics in the evolution path, the risk level and evolution rate of bolt state degradation are quantitatively assessed. Specifically, this includes: calculating the cumulative probability of transitioning to a more severe state to assess the risk level, and analyzing the frequency of state deterioration to determine the evolution rate.

[0090] This step employs a minimum path entropy-based assessment method. By constructing state sequences and establishing state transition probability matrices, the fundamental laws governing bolt state evolution are obtained. The most probable state evolution path is identified using the minimum path entropy principle, and the risk of bolt state degradation is quantitatively assessed by combining state transition probabilities and time series characteristics. This method, through probabilistic modeling of the state evolution process and path entropy analysis, achieves a quantitative assessment of bolt state degradation risk, transforming bolt state monitoring from a simple state judgment to a quantifiable risk assessment.

[0091] This invention also provides a real-time monitoring system for the status of transmission line bolts based on intelligent monitoring, comprising the following modules:

[0092] The signal acquisition module is used to deploy a piezoelectric-piezoresistive composite sensor array with integrated temperature-strain compensation units on the surface of transmission line bolts to capture radial and axial strain signals of the bolts.

[0093] The feature extraction module is used to perform feature decoupling and information gain on the strain signal using orthogonal variational mode decomposition and feature entropy matrix algorithm to obtain structured feature vectors;

[0094] A state recognition module is configured to construct a bolt state anomaly recognition model based on comparative transfer learning and perform domain adversarial training, input the structured feature vector, and perform cross-condition bolt state classification.

[0095] A model updating module is configured to perform online learning and dynamic updating of the bolt state anomaly recognition model through a hierarchical edge computing architecture.

[0096] A health assessment module is configured to establish a bolt health assessment system based on minimum path entropy and perform probabilistic inference on the evolution process of the bolt state.

[0097] The application also provides a computer device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the following steps when executing the computer program:

[0098] A piezoelectric-piezoresistive composite sensing array integrated with temperature-strain compensation units is deployed on the surface of a bolt of a power transmission line to capture radial and axial strain signals of the bolt.

[0099] Orthogonal variational modal decomposition and characteristic entropy matrix algorithm are adopted to decouple the characteristics and gain information of the strain signals, and a structured feature vector is obtained.

[0100] A bolt state anomaly recognition model is constructed based on comparative transfer learning and domain adversarial training, the structured feature vector is input, and cross-condition bolt state classification is performed.

[0101] Online learning and dynamic updating of the bolt state anomaly recognition model are performed through a hierarchical edge computing architecture.

[0102] A bolt health assessment system based on minimum path entropy is established, and probabilistic inference is performed on the evolution process of the bolt state.

[0103] The application also provides a computer storage medium, which stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0104] A piezoelectric-piezoresistive composite sensing array integrated with temperature-strain compensation units is deployed on the surface of a bolt of a power transmission line to capture radial and axial strain signals of the bolt.

[0105] Orthogonal variational modal decomposition and characteristic entropy matrix algorithm are adopted to decouple the characteristics and gain information of the strain signals, and a structured feature vector is obtained.

[0106] A bolt state anomaly recognition model is constructed based on comparative transfer learning and domain adversarial training, the structured feature vector is input, and cross-condition bolt state classification is performed.

[0107] Through the layered edge computing architecture, the bolt state anomaly identification model is learned and dynamically updated online;

[0108] A bolt health evaluation system based on minimum path entropy is established to probabilistically infer the evolution process of the bolt state.

[0109] The above-described embodiments only express one implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for real-time monitoring of the status of transmission line bolts based on intelligent monitoring, characterized in that, Includes the following steps: A piezoelectric-piezoresistive composite sensor array with integrated temperature-strain compensation units is deployed on the surface of transmission line bolts to capture radial and axial strain signals of the bolts; The orthogonal variational mode decomposition and feature entropy matrix algorithm are used to perform feature decoupling and information gain on the strain signal to obtain a structured feature vector; A bolt condition anomaly identification model is constructed based on contrastive transfer learning and subjected to domain adversarial training. The structured feature vector is used as input to classify bolt conditions across working conditions. The bolt state anomaly identification model is learned and dynamically updated online through a hierarchical edge computing architecture. A bolt health assessment system based on minimum path entropy is established to probabilistically infer the evolution of bolt states.

2. The method for real-time monitoring of transmission line bolt status based on intelligent monitoring according to claim 1, characterized in that, The process of capturing the radial and axial strain signals of the bolt includes: A piezoelectric-piezoresistive composite sensing array is prepared, wherein the piezoelectric-piezoresistive composite sensing array is composed of multiple piezoelectric-piezoresistive composite sensing units, and the piezoelectric-piezoresistive composite sensing units are prepared by combining a piezoelectric ceramic substrate and a piezoresistive thin film layer; Electrode patterns are prepared on the piezoelectric ceramic substrate along a first orthogonal direction and a second orthogonal direction, and the piezoresistive thin film layer is arranged in a cross-grid pattern. The piezoelectric-piezoresistive composite sensing unit is subjected to alternating polarization processing; A temperature sensor is integrated on the surface of the piezoelectric-piezoresistive composite sensing unit, and a strain-temperature mapping model is established for temperature compensation. By using a cross-scale signal conditioning circuit, the orthogonal components of the output signal of the piezoelectric-piezoresistive composite sensing array are separated to obtain the temperature-compensated radial and axial strain signals.

3. The method for real-time monitoring of transmission line bolt status based on intelligent monitoring according to claim 1, characterized in that, The process of performing feature decoupling and information gain on the strain signal is as follows: A composite mode decomposition operator is constructed, and the strain signal is decomposed into intrinsic mode components by combining wavelet transform and empirical mode decomposition. Based on the information theory entropy gain criterion, the complexity characteristics of the intrinsic mode components are calculated, and the feature entropy is constructed. matrix; An orthogonal projection decoupling algorithm is used to perform eigenvalue decomposition and orthogonal projection on the feature entropy matrix; The feature vector subset associated with bolt status is selected using the minimum redundancy and maximum relevance criterion. The feature vector subset is reconstructed to generate multidimensional decoupled structured feature vectors.

4. The method for real-time monitoring of transmission line bolt status based on intelligent monitoring according to claim 1, characterized in that, The process of classifying bolt conditions across working conditions is as follows: Bolt state samples under different working conditions are organized in time sequence to construct a cross-working-condition bolt state sample set. A bolt state anomaly identification model is constructed, which adopts a domain adversarial generative network architecture, including a feature extraction unit, a domain discrimination unit, and a temporal state classification unit. The feature extraction unit obtains domain-invariant feature representations based on a gradient inversion layer. By using domain adversarial training, adversarial loss function and orthogonal regularization constraint, the difference in feature distribution between domains is reduced, and domain-adaptive features are obtained. A contrastive learning strategy using Siamese networks is employed to learn bolt state discrimination features through sample similarity measurement and margin maximization criteria. Based on minimum conditional entropy and interval learning theory, the classification feature discrimination performance is optimized and a classification decision boundary is generated. The classification decision threshold is adaptively adjusted based on the similarity of feature distributions between the source and target domains. Repeat the training process until the bolt state anomaly recognition model converges.

5. The method for real-time monitoring of transmission line bolt status based on intelligent monitoring according to claim 1, characterized in that, The process of online learning and dynamic updating of the bolt condition anomaly identification model is as follows: The layered edge computing architecture is divided into a perception layer, an edge computing layer, and a cloud management layer, and a lightweight anomaly detection model is deployed in the edge computing layer. Incremental learning units for model parameters are built in the edge computing layer, and a learning triggering mechanism based on classification loss is set up to trigger the online learning process when the classification loss of a new sample exceeds the learning threshold. A parameter-level incremental learning strategy is adopted for the local model of edge nodes, and the incremental information of local model parameters is transmitted to the cloud management layer. The cloud management layer aggregates the local model increments of each edge node and reconstructs the global model. The reconstructed global model parameters are distributed to each edge computing node to complete the hierarchical collaborative update of the model.

6. The method for real-time monitoring of transmission line bolt status based on intelligent monitoring according to claim 1, characterized in that, The process of probabilistically inferring the evolution of the bolt state is as follows: The bolt state classification results are arranged in chronological order to form a state evolution sequence; Based on the state evolution sequence, a state transition probability matrix is ​​established to quantify the transition probability between adjacent states; Based on the state transition probability matrix, a minimum path entropy evaluation model is constructed. By calculating the probability entropy of different state evolution paths, the potential evolution paths of the bolt state are identified and quantified. Based on the evolution path quantification results of the minimum path entropy assessment model, the risk of bolt condition degradation is quantitatively analyzed.

7. A real-time monitoring system for the status of transmission line bolts based on intelligent monitoring, characterized in that, Includes the following modules: The signal acquisition module is used to deploy a piezoelectric-piezoresistive composite sensor array with integrated temperature-strain compensation units on the surface of transmission line bolts to capture radial and axial strain signals of the bolts. The feature extraction module is used to perform feature decoupling and information gain on the strain signal using orthogonal variational mode decomposition and feature entropy matrix algorithm to obtain structured feature vectors; The state recognition module is used to construct a bolt state anomaly recognition model based on contrastive transfer learning and perform domain adversarial training, and to perform cross-working-condition bolt state classification using the structured feature vector as input. The model update module is used to perform online learning and dynamic updates on the bolt state anomaly identification model through a hierarchical edge computing architecture. The health assessment module is used to establish a bolt health assessment system based on minimum path entropy and to make probabilistic inferences about the evolution of bolt states.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

9. A computer-storable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images