A power transmission line anomaly detection method and system

Abstract

The application discloses a kind of transmission line anomaly detection method and system, content includes data acquisition, multiscale time sequence feature engineering, training set construction, training anomaly detection model and alarm decision.The application relates to the technical field of power transmission safety management, specifically refers to a kind of transmission line anomaly detection method and system, the scheme uses topological weighted cross-correlation time alignment and multiscale window feature extraction, and with historical quantile normalization, improve cross-node consistency and transient retention, enhance short-time fault sensitivity and reduce false alarm;Fusion event, relay protection and artificial confirmation generate soft label, use conditional variational autoencoder to synthesize rare fault sample and train according to confidence weighting, improve small sample recognition and robustness;Construct joint discrimination, reconstruction and prediction framework, comprehensive multidimensional error is carried out anomaly score and graded alarm, improve detection sensitivity, early warning and fault location ability.

Description

Technical Field

[0001] This invention relates to the field of power transmission safety management technology, specifically to a method and system for detecting abnormalities in power transmission lines. Background Technology

[0002] With the continuous increase in the scale and complexity of power systems, the detection of anomalies in transmission lines has become a core link in the safe and stable operation of the power grid. Traditional methods mainly rely on threshold-based rule judgment and relay protection action records. However, the former is easily affected by noise interference and is difficult to adapt to different operating conditions, resulting in frequent false alarms and missed alarms. Although the latter can reflect serious faults, the detection is lagging and lacks sensitivity to early or minor anomalies.

[0003] In recent years, big data and machine learning methods have been gradually introduced into transmission line anomaly detection. These methods have the potential to learn complex patterns, but they still face challenges in engineering applications: labeled data is scarce and contains noise and inconsistencies; the scarcity of fault samples makes it difficult for models to identify a few types of faults; and single discrimination or reconstruction strategies cannot balance accuracy and early warning capabilities. In addition, existing methods often fail to fully utilize the topological priors of transmission lines, and the effect of cross-node information fusion is limited. Overall, current technologies still have significant shortcomings in achieving high-precision, robust, and real-time transmission line anomaly detection and hierarchical alarm. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a method and system for detecting anomalies in transmission lines. Addressing the problems of traditional transmission line anomaly detection methods relying on single-scale features and time asynchrony leading to low sensitivity, inaccurate location, and high false alarm rates in transient fault detection, this solution employs topology-weighted timestamp correction at the edges and adaptive alignment with neighboring cross-correlation peaks. It combines multi-scale window extraction of lightweight features such as time-domain root mean square, narrowband energy proportion, and neighborhood difference, and performs adaptive normalization based on node historical quantiles. Furthermore, addressing the problems of traditional transmission line anomaly detection methods relying on sparse, noisy, single-source labels, difficulty in identifying rare faults, and overfitting, this solution… By fusing event records, relay protection triggers, and manual confirmations into soft labels based on weights and time decay, a conditional variational autoencoder is used to synthesize rare fault samples in a normalized feature space. Training weights are assigned to the samples based on label confidence and confidence intervals. To address the problem that traditional transmission line anomaly detection methods' single detection strategies cannot simultaneously achieve accuracy, early warning, and location capabilities, this scheme constructs a joint framework that enables the encoder, discriminator, reconstruction branch, and time-series prediction branch to work collaboratively. The loss is trained using sample confidence weights and adaptively normalized according to confidence intervals, allowing discriminator confidence, reconstruction error, prediction error, and latent variable distance to be used synergistically for anomaly scoring and graded alarms.

[0005] The technical solution adopted by this invention is as follows: A method for detecting anomalies in power transmission lines, the method comprising the following steps:

[0006] Step S1: Data acquisition. Set the transmission line as N nodes and periodically collect transmission data for each node.

[0007] Step S2: Multi-scale temporal feature engineering. First, the node data is time-aligned. An alignment strategy based on timestamp correction and adaptive search of neighborhood cross-correlation peaks is used to generate a unified resampling sequence. Then, the time domain and frequency domain features are calculated, and the features at each scale are connected in parallel to form a multi-scale feature vector. Finally, quantile normalization is performed on each feature vector to obtain the confidence interval measure.

[0008] Step S3: Training set construction, generating soft label confidence for each sample, and using a conditional variational autoencoder to synthesize samples, finally constructing a training set containing sample features and labels;

[0009] Step S4: Train the anomaly detection model and build a joint model. The model includes an encoder, a discriminator, a predictor, and a decoder. During training, the discriminator loss, reconstruction loss, and prediction loss are minimized by weighting the sample weights, and the reconstruction and prediction errors are adaptively normalized according to the confidence interval. An adaptive optimizer is used for training optimization.

[0010] Step S5: Alarm decision-making, calculate the comprehensive anomaly score, classify alarms, and generate and send a diagnostic package containing the triggering node, triggering time and comprehensive anomaly score when an alarm is triggered.

[0011] Further, in step S1, the data acquisition specifically includes the following steps:

[0012] Step S11: Define the sampling fields. Set the transmission line as N nodes and define the transmission information fields collected by the nodes. Each field is accompanied by a node identifier n and a timestamp t during collection. The transmission information includes: phase voltage, current, phase angle, switch status, ambient temperature, wind speed, vibration amplitude, electromagnetic noise, event records, relay protection records, and node metadata of node n at time t. The node metadata includes the node's unique identifier, coordinates, line ID, last maintenance timestamp, and a list of adjacent nodes.

[0013] Step S12: Set the sampling period. For each node N, collect power transmission information according to a fixed sampling period. For transient sensitive channels that include phase voltage, current, phase angle, vibration amplitude and electromagnetic noise, set the sampling period to 0.5 milliseconds; after alignment, the unified resampling period is 5 milliseconds; for environmental slow variables that include ambient temperature and wind speed, set the sampling period to 10 seconds; for switch status and relay protection records, set the sampling period to 100 milliseconds.

[0014] Furthermore, in step S2, the multi-scale temporal feature engineering specifically includes the following steps:

[0015] Step S21: Time alignment. Time alignment is performed on the node data. An adaptive alignment strategy based on timestamp correction and local cross-correlation maximum value is adopted to achieve low-complexity cross-correlation peak search at the edges and output a unified resampling sequence, as shown below:

[0016] ;

[0017] in, This represents the time series data after alignment and resampling. Indicates that node n at time n The original data, This indicates resampling, which downsamples the time series data to the sampling period. , Represents the time offset, in the candidate time offset set Select the optimal value to align the adjacent node sequences. This represents the cross-correlation confidence value, which measures the sequence of node n shifted by the offset. During alignment, with the set of adjacent nodes The overall similarity of the sequences; a higher value indicates better alignment; m represents the index of adjacent nodes. This represents an entry in the node topological adjacency matrix. If nodes n and m have a direct conductive topological connection, then... ,otherwise ; This represents the vector dot product operation; This represents the set of all nodes adjacent to node n;

[0018] Step S22: Construct a feature vector. Calculate a set of time-domain and frequency-domain features for each uniformly sampled sequence, using a multi-scale window set, and concatenate the features at each scale into a feature vector, as shown below:

[0019] ;

[0020] in, This represents the multi-scale eigenvector of node n at time t; , and These represent the time windows. , and The time series data after alignment and resampling of the nth node. Calculate the root mean square value; Indicates within the time window Upper calculation subband Spectral energy percentage, subband , This represents the fundamental frequency, with a value of 50 Hz. The value is 2 Hz; Indicates a pair of sub-bands Short-time Fourier transform coefficient extraction; Indicates time window A set of time points; Indicates in window The average difference between the features of node n and its neighboring nodes is calculated. This indicates that the mean of the sequence is calculated within the window. This indicates taking the absolute value. Represents a set The number of nodes;

[0021] Step S23: Multi-scale normalization. Adaptive quantile normalization is performed on the feature vector of each node. Linear normalization is performed based on the historical quantiles of the nodes, as shown below:

[0022] ;

[0023] in, This represents the normalized eigenvector of node n. This represents the vector formed by calculating the p-th percentile of the multi-scale feature vector of node n within the historical window, with the historical window length being the past day. Let represent the vector formed by calculating the qth percentile of the multi-scale feature vector of node n within the historical window, where , ; This indicates a measure of the width of the confidence interval.

[0024] Further, in step S3, the construction of the training set specifically includes the following steps:

[0025] Step S31: Multi-source label fusion. Initial labels for training samples are generated by weighted fusion of event records and relay protection records, as shown below:

[0026] ;

[0027] in, This represents the overall label confidence score of node n at time t, with a value range of [value range missing]. , as soft labels for training samples; , and These represent three confidence weight constants, corresponding to the confidence level of the event record, the confidence level in the relay protection record, and the confidence level of manual confirmation, respectively. This represents the event indicator function. If all event records except for manually confirmed events occur at time t, the value is 1; otherwise, it is 0. The time-decay confidence function representing manual confirmation is derived from the event log. The event type is manual confirmation. It is composed of and decays exponentially with respect to time difference; This indicates a relay protection record; the value is 1 if the record appears, and 0 otherwise. Indicates the time decay parameter. This indicates an indicator function; it takes a value of 1 if the condition is true, and a value of 0 otherwise. Represents an exponential function with the natural constant as its base;

[0028] Step S32: Rare sample synthesis. For rare fault types, a conditional variational autoencoder is used to synthesize samples in the feature space. The conditional variable is the soft label of the fault category, as shown below:

[0029] ;

[0030] in, This represents the loss function of the conditional variational autoencoder. This represents taking the mathematical expectation. This indicates that the encoder is given normalized features. The posterior approximate distribution of condition c, with parameters as follows: ; This indicates that the decoder is given a pair of latent variables z and conditions c. The generation distribution, with parameters as ; Denotes KL divergence, Let the prior distribution be denoted as standard normal. Represents the synthesized feature samples, Indicates soft label Mapping to discrete class conditions, when Values The mapping is normal, when Values Mapped as suspicious, when Values The mapping is based on faults; the encoder and decoder structure employs a two-layer perceptron.

[0031] Step S33: Training set generation, assigning training weights to each training sample according to label confidence and confidence interval.

[0032] Further, in step S4, training the anomaly detection model specifically includes the following steps:

[0033] Step S41: Design the model framework, and normalize the multi-scale feature vectors. As input to the model, the model consists of three parts: the first part is the encoder. 3-layer The encoder maps the input to a latent representation. The second part is the discriminant head. A two-layer perceptron is used, and the discriminant head provides a prediction of the node anomaly confidence level. The third part is the prediction head, which includes two branches: a reconstruction branch and a sub-branch. and time series prediction branch The reconstruction branch performs input reconstruction. A mirror encoder structure is used for anomaly detection; the timing prediction branch adopts a single layer. Based on history The potential representation of each time step Predicting latent variables in the short term And solve it inversely into the characteristic space. The overall parameters of the model are: ;

[0034] Step S42: Construct the loss function for the training set. By sample weight Weighted training is represented as follows:

[0035] ;

[0036] in, Indicates the determination of loss. The reconstruction loss is represented by the confidence interval width introduced in the denominator. As an adaptive normalization factor Denotes the square of the L2 norm. Indicates removing zero factors. This represents the time series prediction loss. , and This represents the loss fusion coefficient; all terms are weighted according to sample weights. Weighted optimization using a dynamic adaptive optimizer ;

[0037] Step S43: Inference output. For each node n and time t, the following are obtained through inference: discrimination confidence, reconstruction error, prediction error, and latent variable representation.

[0038] Furthermore, in step S5, the alarm decision specifically includes the following steps:

[0039] Step S51: Construct node anomaly scores, and construct a comprehensive anomaly score based on the inference output. First, normalize the sub-scores, including: the discriminant confidence score. Reconstruction score Predicted score and potential representation score ,in This indicates that the Mahalanobis distance is taken; then, the scores are fused according to weights to obtain the final score, and then mapped to... , represented as:

[0040] ;

[0041] in, This represents the final overall abnormality score. , , and Indicates the score-adaptation weight. Represents the original outlier score. This represents the sigmoid activation function. and These represent the weights and biases used to map outlier scores, respectively.

[0042] Step S52: Graded alarm, output alarm level according to comprehensive anomaly score; when an alarm is triggered, generate and send out a diagnostic package, which includes: alarm trigger node list, time, and comprehensive anomaly score.

[0043] The present invention provides a transmission line anomaly detection system, comprising a data acquisition module, a multi-scale time series feature engineering module, a training set construction module, a training anomaly detection model module, and an alarm decision module;

[0044] The data acquisition module sets the transmission line as N nodes, periodically collects transmission data for each node, and sends the data to the multi-scale time series feature engineering module.

[0045] The multi-scale time series feature engineering module receives data sent by the data acquisition module, first performs time alignment on the node data, and generates a unified resampling sequence using an alignment strategy based on timestamp correction and adaptive search of neighborhood cross-correlation peaks. Then, it calculates time-domain and frequency-domain features using a sliding window on the time scale, and connects the features of each scale in parallel to form a multi-scale feature vector. Finally, it performs adaptive quantile normalization on each feature vector based on the historical quantiles of the nodes to obtain a confidence interval measure, and sends the data to the training set construction module, the anomaly detection model training module, and the alarm decision module.

[0046] The training set construction module receives data sent by the multi-scale time series feature engineering module, generates soft label confidence for each sample based on event records, relay protection records, and manual confirmation records, synthesizes samples using a conditional variational autoencoder in the normalized feature space, and finally assigns training weights to each real and synthetic sample according to the label confidence and confidence interval, constructs a training set containing sample features and labels, and sends the data to the training anomaly detection model module.

[0047] The training anomaly detection model module receives data sent by the training set construction module, constructs a joint model, and the model includes an encoder for mapping to latent variable representations, a discriminant head for outputting anomaly confidence, and a prediction head and decoder for short-term time series prediction and reconstruction, respectively. During training, the discriminant loss, reconstruction loss, and prediction loss are minimized by weighting the sample weights, and the reconstruction and prediction errors are adaptively normalized according to the confidence interval. The training optimization uses an adaptive optimizer and outputs the discriminant confidence, reconstruction error, prediction error, and latent variable representation during inference, and sends the data to the alarm decision module.

[0048] The alarm decision module receives data from the multi-scale temporal feature engineering module and the training anomaly detection model module. Based on the inference output, it calculates the discrimination confidence score, normalized reconstruction error score, normalized prediction error score, and latent variable representation score. These scores are then fused and mapped into a unified comprehensive anomaly score according to their weights. Based on the comprehensive score and duration threshold, the alarm is classified into severe, moderate, and normal levels. When an alarm is triggered, a diagnostic package containing the trigger node, trigger time, and comprehensive anomaly score is generated and sent out.

[0049] The beneficial effects achieved by the present invention using the above solution are as follows:

[0050] (1) In view of the problems of low sensitivity, inaccurate positioning and high false alarm rate of traditional transmission line anomaly detection methods that rely on single-scale features and time asynchrony, this scheme adopts topology-weighted timestamp correction and adaptive alignment of neighbor cross-correlation peaks at the edge, and combines multi-scale window extraction of lightweight features such as root mean square in time domain, narrowband energy ratio and neighborhood difference, and performs adaptive normalization based on node historical quantiles. This improves cross-node data consistency and transient information retention, enhances the detection sensitivity and positioning accuracy of short-term faults and weak anomalies, and reduces false alarms caused by noise and asynchrony while taking into account edge computing and bandwidth limitations.

[0051] (2) In view of the problems that traditional transmission line anomaly detection methods rely on sparse, noisy, single-source labels and are difficult to identify rare faults and are prone to overfitting, this scheme integrates event records, relay protection triggers and manual confirmations into soft labels according to weight and time decay, uses conditional variational autoencoders to synthesize rare fault samples in a normalized feature space, and assigns training weights to samples according to label confidence and confidence interval, thereby enhancing the representativeness and labeling robustness of the training set, thereby improving the model's ability to identify faults with few samples and reducing the negative impact of label noise on actual operating performance.

[0052] (3) To address the problem that traditional transmission line anomaly detection methods cannot simultaneously achieve accuracy, early warning and location capabilities with a single detection strategy, this solution constructs a joint framework that enables the encoder, discriminator, reconstruction branch and time-series prediction branch to work together. The loss is trained with sample confidence weighted and adaptively normalized according to the confidence interval, so that the discrimination confidence, reconstruction error, prediction error and latent variable distance can be used together for anomaly scoring and graded alarm, thereby improving detection sensitivity and early warning capability, enhancing fault location accuracy and robustness and interpretability to noise, label uncertainty and distribution drift. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of a method for detecting anomalies in power transmission lines provided by the present invention;

[0054] Figure 2 This is a schematic diagram of a power transmission line anomaly detection system provided by the present invention;

[0055] Figure 3 This is a schematic diagram of step S1;

[0056] Figure 4 This is a schematic diagram of step S2;

[0057] Figure 5 This is a schematic diagram of step S3;

[0058] Figure 6 This is a schematic diagram of step S4;

[0059] Figure 7 This is a schematic diagram of step S5.

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0062] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0063] Example 1, see Figure 1 The present invention provides a method for detecting anomalies in power transmission lines, the method comprising the following steps:

[0064] Step S1: Data acquisition. Set the transmission line as N nodes and periodically collect transmission data for each node, including phase voltage, current, phase angle, switch status, ambient temperature, wind speed, vibration amplitude and electromagnetic noise. Record event logs and relay protection trigger information, and save node metadata such as unique identifier, coordinates, line to which it belongs, last maintenance time and adjacent node list.

[0065] Step S2: Multi-scale temporal feature engineering. First, the node data is time-aligned. An alignment strategy based on timestamp correction and adaptive search of neighborhood cross-correlation peaks is used to generate a unified resampling sequence. Then, time-domain and frequency-domain features are calculated using a sliding window at multiple time scales. The features at each scale are then connected in parallel to form a multi-scale feature vector. Finally, adaptive quantile normalization is performed on each feature vector based on the historical quantiles of the nodes to obtain the confidence interval measure.

[0066] Step S3: Training set construction. Soft label confidence scores are generated for each sample based on event records, relay protection records, and manual confirmation records. Samples are synthesized using a conditional variational autoencoder in the normalized feature space. Finally, training weights are assigned to each real and synthetic sample based on the label confidence scores and confidence intervals, and a training set containing sample features and labels is constructed.

[0067] Step S4: Train the anomaly detection model and construct a joint model. The model includes an encoder for mapping to latent representations, a discriminant head for outputting anomaly confidence, and a prediction head and decoder for short-term time series prediction and reconstruction, respectively. During training, the discriminant loss, reconstruction loss, and prediction loss are minimized by weighting the sample weights, and the reconstruction and prediction errors are adaptively normalized according to the confidence interval. The training optimization uses an adaptive optimizer and outputs the discriminant confidence, reconstruction error, prediction error, and latent variable representation during inference.

[0068] Step S5: Alarm decision-making. Based on the inference output, calculate the discrimination confidence score, normalized reconstruction error score, normalized prediction error score, and latent representation score. These scores are then fused and mapped into a unified comprehensive anomaly score according to their weights. Alarms are classified into severe, moderate, and normal levels based on the comprehensive score and duration threshold. When an alarm is triggered, a diagnostic package containing the triggering node, triggering time, and comprehensive anomaly score is generated and sent out.

[0069] Example 2, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S1, the data acquisition specifically includes the following steps:

[0070] Step S11: Define the sampling fields. Set the transmission line as N nodes and define the transmission information fields collected by the nodes. Each field includes a node identifier n and a timestamp t during collection. The transmission information includes: the phase voltage of node n at time t. Current Phase angle Switch status Ambient temperature Wind speed Vibration amplitude Electromagnetic noise Event Log Relay protection records And node metadata; the node metadata includes the node's unique identifier, coordinates, line ID, last maintenance timestamp, and a list of adjacent nodes; wherein, Indicates the time of the event recording. Indicates the event type, including fault confirmation, sensor anomaly, and manual confirmation;

[0071] Step S12: Set the sampling period. For each node N, collect power transmission information according to a fixed sampling period. For transient sensitive channels that include phase voltage, current, phase angle, vibration amplitude and electromagnetic noise, set the sampling period to 0.5 milliseconds; after alignment, the unified resampling period is 5 milliseconds; for environmental slow variables that include ambient temperature and wind speed, set the sampling period to 10 seconds; for switch status and relay protection records, set the sampling period to 100 milliseconds.

[0072] Example 3, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S2, the multi-scale temporal feature engineering specifically includes the following steps:

[0073] Step S21: Time alignment. Time alignment is performed on the node data. An adaptive alignment strategy based on timestamp correction and local cross-correlation maximum value is adopted to achieve low-complexity cross-correlation peak search at the edges and output a unified resampling sequence, as shown below:

[0074] ;

[0075] in, This represents the time series data after alignment and resampling. Indicates that node n at time n The original data, This indicates resampling, which downsamples the time series data to the sampling period. , Represents the time offset, in the candidate time offset set Select the optimal value to align the adjacent node sequences. This represents the cross-correlation confidence value, which measures the sequence of node n shifted by the offset. During alignment, with the set of adjacent nodes The overall similarity of the sequences; a higher value indicates better alignment; m represents the index of adjacent nodes. This represents an entry in the node topological adjacency matrix. If nodes n and m have a direct conductive topological connection, then... ,otherwise ; This represents the vector dot product operation; This represents the set of all nodes adjacent to node n;

[0076] Step S22: Construct feature vectors by calculating a set of lightweight time-domain and frequency-domain features on each uniformly sampled sequence, using a multi-scale window set. These correspond to 1 second, 10 seconds, and 60 seconds, respectively, and the features at each scale are concatenated into a feature vector, as shown below:

[0077] ;

[0078] in, This represents the multi-scale eigenvector of node n at time t; , and These represent the time windows. , and The time series data after alignment and resampling of the nth node. Calculate the root mean square value; Indicates within the time window Upper calculation subband Spectral energy percentage, subband , This represents the fundamental frequency, with a value of 50 Hz. The value is 2 Hz; Indicates a pair of sub-bands Short-time Fourier transform coefficient extraction; Indicates time window A set of time points; Indicates in window The average difference between the features of node n and its neighboring nodes is calculated. This indicates that the mean of the sequence is calculated within the window. This indicates taking the absolute value. Represents a set The number of nodes;

[0079] Step S23: Multi-scale normalization. Adaptive quantile normalization is performed on the feature vector of each node. Linear normalization is performed based on the historical quantiles of the nodes, as shown below:

[0080] ;

[0081] in, This represents the normalized eigenvector of node n. This represents the vector formed by calculating the p-th percentile of the multi-scale feature vector of node n within the historical window, with the historical window length being the past day. Let represent the vector formed by calculating the qth percentile of the multi-scale feature vector of node n within the historical window, where , ; This indicates a measure of the width of the confidence interval.

[0082] By performing the above operations, this scheme addresses the problems of low sensitivity, inaccurate location, and high false alarm rate in traditional transmission line anomaly detection methods that rely on single-scale features and time asynchrony. It adopts topology-weighted timestamp correction and adaptive alignment of neighborhood cross-correlation peaks at the edge, combines multi-scale window extraction of lightweight features such as root mean square in the time domain, narrowband energy ratio, and neighborhood difference, and performs adaptive normalization based on the node's historical quantile. This improves cross-node data consistency and transient information retention, enhances the detection sensitivity and location accuracy of short-term faults and weak anomalies, reduces false alarms caused by noise and asynchrony, and takes into account edge computing and bandwidth limitations.

[0083] Example 4, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S3, the construction of the training set specifically includes the following steps:

[0084] Step S31: Multi-source label fusion. Initial labels for training samples are generated by weighted fusion of event records and relay protection records, as shown below:

[0085] ;

[0086] in, This represents the overall label confidence score of node n at time t, with a value range of [value range missing]. , as soft labels for training samples; , and These represent three confidence weight constants, corresponding to the confidence level of the event record, the confidence level in the relay protection record, and the confidence level of manual confirmation, respectively. This represents the event indicator function. If all event records except for manually confirmed events occur at time t, the value is 1; otherwise, it is 0. The time-decay confidence function representing manual confirmation is derived from the event log. The event type is manual confirmation. It is composed of and decays exponentially with respect to time difference; This indicates a relay protection record; the value is 1 if the record appears, and 0 otherwise. Indicates the time decay parameter. This indicates an indicator function; it takes a value of 1 if the condition is true, and a value of 0 otherwise. Represents an exponential function with the natural constant as its base;

[0087] Step S32: Rare sample synthesis. For rare fault types, a conditional variational autoencoder is used in the feature space. The above synthesized sample, with the condition variable being the soft label of the fault category, is represented as follows:

[0088] ;

[0089] in, This represents the loss function of the conditional variational autoencoder. This represents taking the mathematical expectation. This indicates that the encoder is given normalized features. The posterior approximate distribution of condition c, with parameters as follows: ; This indicates that the decoder is given a pair of latent variables z and conditions c. The generation distribution, with parameters as ; Denotes KL divergence, Let the prior distribution be denoted as standard normal. Represents the synthesized feature samples, Indicates soft label Mapping to discrete class conditions, when Values The mapping is normal, when Values Mapped as suspicious, when Values The mapping is based on faults; the encoder and decoder structure employs a two-layer perceptron.

[0090] Step S33: Training set generation, for each training sample, according to label confidence and confidence interval Training weights are assigned, with higher weights having a larger proportion in the classification loss, as shown below:

[0091] ;

[0092] in, The training samples are assigned positive real-valued training weights. Indicates the label exponentiation factor. This represents the confidence interval penalty coefficient. This represents the final training set, which includes real samples and synthetic samples. N represents the total number of samples, and T represents the maximum time step.

[0093] By performing the above operations, this scheme addresses the problems of traditional transmission line anomaly detection methods relying on sparse, noisy, single-source labels and struggling to identify rare faults, as well as being prone to overfitting. It integrates event records, relay protection triggers, and manual confirmations into soft labels based on weights and time decay. A conditional variational autoencoder is then used to synthesize rare fault samples in a normalized feature space. Training weights are assigned to the samples based on label confidence and confidence intervals, enhancing the representativeness and robustness of the training set. This improves the model's ability to identify faults with few samples and reduces the negative impact of label noise on actual operational performance.

[0094] Example 5, see Figure 1 and Figure 6 This embodiment is based on the above embodiment. In step S4, training the anomaly detection model specifically includes the following steps:

[0095] Step S41: Design the model framework, and normalize the multi-scale feature vectors. As input to the model, the model consists of three parts: the first part is the encoder. 3-layer The encoder maps the input to a latent representation. The second part is the discriminant head. A two-layer perceptron is used, and the discriminant head provides a prediction of the node anomaly confidence level. The third part is the prediction head, which includes two branches: a reconstruction branch and a sub-branch. and time series prediction branch The reconstruction branch performs input reconstruction. A mirror encoder structure is used for anomaly detection; the timing prediction branch adopts a single layer. Based on history The potential representation of each time step Predicting latent variables in the short term And solve it inversely into the characteristic space. The overall parameters of the model are ;

[0096] Step S42: Construct the loss function for the training set. By sample weight Weighted training is represented as follows:

[0097] ;

[0098] in, Indicates the determination of loss. The reconstruction loss is represented by the confidence interval width introduced in the denominator. As an adaptive normalization factor Denotes the square of the L2 norm. Indicates removing zero factors. This represents the time series prediction loss. , and This represents the loss fusion coefficient; all terms are weighted according to sample weights. Weighted optimization using a dynamic adaptive optimizer ;

[0099] Step S43: Inference output, for each node n and time t, infer the following: discriminant confidence score. Reconstruction error Prediction error Latent variables .

[0100] By performing the above operations, this solution addresses the problem that traditional transmission line anomaly detection methods, relying on a single detection strategy, struggle to simultaneously achieve accuracy, early warning, and location capabilities. Instead, it constructs a joint framework that enables the encoder, discriminator, reconstruction branch, and temporal prediction branch to work collaboratively. The framework is trained using a loss weighted by sample confidence and adaptively normalized according to confidence intervals. This allows the discriminator confidence, reconstruction error, prediction error, and latent variable distance to be used synergistically for anomaly scoring and graded alarms. This improves detection sensitivity and early warning capabilities, enhances fault location accuracy, and strengthens robustness and interpretability against noise, label uncertainty, and distribution drift.

[0101] Example 6, see Figure 1 and Figure 7 This embodiment is based on the above embodiment. In step S5, the alarm decision specifically includes the following steps:

[0102] Step S51: Construct node anomaly scores, and construct a comprehensive anomaly score based on the inference output. First, normalize the sub-scores, including: the discriminant confidence score. Reconstruction score Predicted score and potential representation score ,in This indicates that the Mahalanobis distance is taken; then, the scores are fused according to weights to obtain the final score, and then mapped to... , represented as:

[0103] ;

[0104] in, This represents the final overall abnormality score. , , and Indicates the score-adaptation weight. Represents the original outlier score. This represents the sigmoid activation function. and These represent the weights and biases used to map outlier scores, respectively.

[0105] Step S52: Tiered alarm, output alarm level based on comprehensive anomaly score. , means as follows:

[0106] ;

[0107] in, Indicates the alarm level. Indicates a severe alert. Indicates a medium-level alarm. No alarms were detected. , ; This represents logical AND. The duration of abnormal scores within a score range is indicated by the persistence threshold. When an alarm is triggered, a diagnostic package is generated and sent out. The diagnostic package includes: a list of alarm triggering nodes, the time, and a comprehensive anomaly score.

[0108] Example 7, see Figure 1 and Figure 2 Based on the above embodiments, the present invention provides a transmission line anomaly detection system, including a data acquisition module, a multi-scale time series feature engineering module, a training set construction module, a training anomaly detection model module, and an alarm decision module.

[0109] The data acquisition module sets the transmission line as N nodes, periodically collects transmission data for each node, including phase voltage, current, phase angle, switch status, ambient temperature, wind speed, vibration amplitude and electromagnetic noise, and records event logs and relay protection trigger information. At the same time, it saves node metadata such as unique identifier, coordinates, line to which it belongs, last maintenance time and adjacent node list, and sends the data to the multi-scale time series feature engineering module.

[0110] The multi-scale time series feature engineering module receives data sent by the data acquisition module, first performs time alignment on the node data, and generates a unified resampling sequence using an alignment strategy based on timestamp correction and adaptive search of neighborhood cross-correlation peaks. Then, it calculates time-domain and frequency-domain features using a sliding window at multiple time scales, and connects the features at each scale in parallel to form a multi-scale feature vector. Finally, it performs adaptive quantile normalization on each feature vector based on the historical quantiles of the nodes to obtain a confidence interval measure, and sends the data to the training set construction module, the anomaly detection model training module, and the alarm decision module.

[0111] The training set construction module receives data sent by the multi-scale time series feature engineering module, generates soft label confidence for each sample based on event records, relay protection records, and manual confirmation records, synthesizes samples using a conditional variational autoencoder in the normalized feature space, and finally assigns training weights to each real and synthetic sample according to the label confidence and confidence interval, constructs a training set containing sample features and labels, and sends the data to the training anomaly detection model module.

[0112] The training anomaly detection model module receives data sent by the training set construction module, constructs a joint model, and the model includes an encoder for mapping to latent representations, a discriminant head for outputting anomaly confidence, and a prediction head and decoder for short-term time series prediction and reconstruction, respectively. During training, the discriminant loss, reconstruction loss, and prediction loss are minimized by weighting the sample weights, and the reconstruction and prediction errors are adaptively normalized according to the confidence interval. The training optimization uses an adaptive optimizer and outputs the discriminant confidence, reconstruction error, prediction error, and latent variable representation during inference, and sends the data to the alarm decision module.

[0113] The alarm decision module receives data sent by the multi-scale temporal feature engineering module and the training anomaly detection model module. Based on the inference output, it calculates the discrimination confidence score, normalized reconstruction error score, normalized prediction error score, and latent representation score. These scores are then fused and mapped into a unified comprehensive anomaly score according to their weights. Based on the comprehensive score and duration threshold, the alarm is classified into severe, moderate, and normal levels. When an alarm is triggered, a diagnostic package containing the trigger node, trigger time, and comprehensive anomaly score is generated and sent out.

[0114] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0116] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for detecting anomalies in transmission lines, characterized in that, The method includes the following steps: Step S1: Data acquisition. Set the transmission line as N nodes and periodically collect transmission data for each node. Step S2: Multi-scale temporal feature engineering. First, the node data is time-aligned. An alignment strategy based on timestamp correction and adaptive search of neighborhood cross-correlation peaks is used to generate a unified resampling sequence. Then, the time domain and frequency domain features are calculated, and the features at each scale are connected in parallel to form a multi-scale feature vector. Finally, quantile normalization is performed on each feature vector to obtain the confidence interval measure. Step S21: Time alignment. Time alignment is performed on the node data. An adaptive alignment strategy based on timestamp correction and local cross-correlation maximum value is adopted to achieve low-complexity cross-correlation peak search at the edges and output a unified resampling sequence, as shown below: ； in, This represents the time series data after alignment and resampling. Indicates that node n at time n The original data, This indicates resampling, which downsamples the time series data to the sampling period. , Represents the time offset, in the candidate time offset set Select the optimal value to align the adjacent node sequences. This represents the cross-correlation confidence value, which measures the sequence of node n shifted by the offset. During alignment, with the set of adjacent nodes The overall similarity of the sequences; a higher value indicates better alignment; m represents the index of adjacent nodes. This represents an entry in the node topological adjacency matrix. If nodes n and m have a direct conductive topological connection, then... ,otherwise ; This represents the vector dot product operation; This represents the set of all nodes adjacent to node n; Step S3: Training set construction, generating soft label confidence for each sample, and using a conditional variational autoencoder to synthesize samples, finally constructing a training set containing sample features and labels; Step S4: Train the anomaly detection model and build a joint model. The model includes an encoder, a discriminator, a predictor, and a decoder. During training, the discriminator loss, reconstruction loss, and prediction loss are minimized by weighting the sample weights, and the reconstruction and prediction errors are adaptively normalized according to the confidence interval. An adaptive optimizer is used for training optimization. Step S41: Design the model framework, and normalize the multi-scale feature vectors. As input to the model, the model consists of three parts: the first part is the encoder. 3-layer The encoder maps the input to a latent representation. The second part is the discriminant head. A two-layer perceptron is used, and the discriminant head provides a prediction of the node anomaly confidence level. The third part is the prediction head, which includes two branches: a reconstruction branch and a sub-branch. and time series prediction branch The reconstruction branch performs input reconstruction. A mirror encoder structure is used for anomaly detection; the timing prediction branch adopts a single layer. Based on history The potential representation of each time step Predicting latent variables in the short term And solve it inversely into the characteristic space. The overall parameters of the model are: ; Step S42: Construct the loss function for the training set. By sample weight Weighted training is represented as follows: ； in, Indicates the determination of loss. The reconstruction loss is represented by the confidence interval width introduced in the denominator. As an adaptive normalization factor Denotes the square of the L2 norm. Indicates removing zero factors. This represents the time series prediction loss. , and This represents the loss fusion coefficient; all terms are weighted according to sample weights. Weighted optimization using a dynamic adaptive optimizer ; Step S43: Inference output, for each node n and time t, the following are obtained through inference: discrimination confidence, reconstruction error, prediction error and latent variable representation; Step S5: Alarm decision-making, calculate the comprehensive anomaly score, classify alarms, and generate and send a diagnostic package containing the triggering node, triggering time and comprehensive anomaly score when an alarm is triggered.

2. The method for detecting anomalies in transmission lines according to claim 1, characterized in that: In step S2, the multi-scale temporal feature engineering specifically includes the following steps: Step S21: Time alignment; Step S22: Construct a feature vector. Calculate a set of time-domain and frequency-domain features for each uniformly sampled sequence, using a multi-scale window set, and concatenate the features at each scale into a feature vector, as shown below: ； in, This represents the multi-scale eigenvector of node n at time t; , and These represent the time windows. , and The time series data after alignment and resampling of the nth node. Calculate the root mean square value; Indicates within the time window Upper calculation subband Spectral energy percentage, subband , This represents the fundamental frequency, with a value of 50 Hz. The value is 2 Hz; Indicates a pair of sub-bands Short-time Fourier transform coefficient extraction; Indicates time window A set of time points; Indicates in window The average difference between the features of node n and its neighboring nodes is calculated. This indicates that the mean of the sequence is calculated within the window. This indicates taking the absolute value. Represents a set The number of nodes; Step S23: Multi-scale normalization. Adaptive quantile normalization is performed on the feature vector of each node. Linear normalization is performed based on the historical quantiles of the nodes, as shown below: ； in, This represents the normalized eigenvector of node n. This represents the vector formed by calculating the p-th percentile of the multi-scale feature vector of node n within the historical window, with the historical window length being the past day. Let represent the vector formed by calculating the qth percentile of the multi-scale feature vector of node n within the historical window, where , ; This indicates a measure of the width of the confidence interval.

3. The method for detecting anomalies in transmission lines according to claim 1, characterized in that: In step S3, the construction of the training set specifically includes the following steps: Step S31: Multi-source label fusion. Initial labels for training samples are generated by weighted fusion of event records and relay protection records, as shown below: ； in, This represents the overall label confidence score of node n at time t, with a value range of [value range missing]. , as soft labels for training samples; , and These represent three confidence weight constants, corresponding to the confidence level of the event record, the confidence level in the relay protection record, and the confidence level of manual confirmation, respectively. This represents the event indicator function. If all event records except for manually confirmed events occur at time t, the value is 1; otherwise, it is 0. The time-decay confidence function representing manual confirmation is derived from the event log. The event type is manual confirmation. It is composed of and decays exponentially with respect to time difference; This indicates a relay protection record; the value is 1 if the record appears, and 0 otherwise. Indicates the time decay parameter. This indicates an indicator function; it takes a value of 1 if the condition is true, and a value of 0 otherwise. Represents an exponential function with the natural constant as its base; Step S32: Rare sample synthesis. For rare fault types, a conditional variational autoencoder is used to synthesize samples in the feature space. The conditional variable is the soft label of the fault category, as shown below: ； in, This represents the loss function of the conditional variational autoencoder. This represents taking the mathematical expectation. This indicates that the encoder is given normalized features. The posterior approximate distribution of condition c, with parameters as follows: ; This indicates that the decoder is given a pair of latent variables z and conditions c. The generation distribution, with parameters as ; Denotes KL divergence, Let the prior distribution be denoted as standard normal. Represents the synthesized feature samples, Indicates soft label Mapping to discrete class conditions, when Values The mapping is normal, when Values Mapped as suspicious, when Values The mapping is based on faults; the encoder and decoder structure employs a two-layer perceptron. Step S33: Training set generation, assigning training weights to each training sample according to label confidence and confidence interval.

4. The method for detecting anomalies in transmission lines according to claim 1, characterized in that: In step S5, the alarm decision specifically includes the following steps: Step S51: Construct node anomaly scores, and construct a comprehensive anomaly score based on the inference output. First, normalize the sub-scores, including: the discriminant confidence score. Reconstruction score Predicted score and potential representation score ,in This indicates that the Mahalanobis distance is taken; then, the scores are fused according to weights to obtain the final score, and then mapped to... , represented as: ； in, This represents the final overall abnormality score. , , and Indicates the score-adaptation weight. Represents the original outlier score. This represents the sigmoid activation function. and These represent the weights and biases used to map outlier scores, respectively. Step S52: Graded alarm, output alarm level according to comprehensive anomaly score; when an alarm is triggered, generate and send out a diagnostic package, which includes: alarm trigger node list, time, and comprehensive anomaly score.

5. The method for detecting anomalies in transmission lines according to claim 1, characterized in that: In step S1, the data acquisition specifically includes the following steps: Step S11: Define the sampling fields. Set the transmission line as N nodes and define the transmission information fields collected by the nodes. Each field is accompanied by a node identifier n and a timestamp t during collection. The transmission information includes: phase voltage, current, phase angle, switch status, ambient temperature, wind speed, vibration amplitude, electromagnetic noise, event records, relay protection records, and node metadata of node n at time t. The node metadata includes the node's unique identifier, coordinates, line ID, last maintenance timestamp, and a list of adjacent nodes. Step S12: Set the sampling period. For each node N, collect power transmission information according to a fixed sampling period. For transient sensitive channels that include phase voltage, current, phase angle, vibration amplitude and electromagnetic noise, set the sampling period to 0.5 milliseconds; after alignment, the unified resampling period is 5 milliseconds; for environmental slow variables that include ambient temperature and wind speed, set the sampling period to 10 seconds; for switch status and relay protection records, set the sampling period to 100 milliseconds.

6. A transmission line anomaly detection system, used to implement the transmission line anomaly detection method as described in any one of claims 1-5, characterized in that: It includes a data acquisition module, a multi-scale time series feature engineering module, a training set construction module, an anomaly detection model training module, and an alarm decision module.

7. The power transmission line anomaly detection system according to claim 6, characterized in that: The data acquisition module sets the transmission line as N nodes, periodically collects transmission data for each node, and sends the data to the multi-scale time series feature engineering module. The multi-scale time series feature engineering module receives data sent by the data acquisition module, first performs time alignment on the node data, and generates a unified resampling sequence using an alignment strategy based on timestamp correction and adaptive search of neighborhood cross-correlation peaks. Then, it calculates time-domain and frequency-domain features using a sliding window on the time scale, and connects the features of each scale in parallel to form a multi-scale feature vector. Finally, it performs adaptive quantile normalization on each feature vector based on the historical quantiles of the nodes to obtain a confidence interval measure, and sends the data to the training set construction module, the anomaly detection model training module, and the alarm decision module. The training set construction module receives data sent by the multi-scale time series feature engineering module, generates soft label confidence for each sample based on event records, relay protection records, and manual confirmation records, synthesizes samples using a conditional variational autoencoder in the normalized feature space, and finally assigns training weights to each real and synthetic sample according to the label confidence and confidence interval, constructs a training set containing sample features and labels, and sends the data to the training anomaly detection model module. The training anomaly detection model module receives data sent by the training set construction module, constructs a joint model, and the model includes an encoder for mapping to latent variable representations, a discriminant head for outputting anomaly confidence, and a prediction head and decoder for short-term time series prediction and reconstruction, respectively. During training, the discriminant loss, reconstruction loss, and prediction loss are minimized by weighting the sample weights, and the reconstruction and prediction errors are adaptively normalized according to the confidence interval. The training optimization uses an adaptive optimizer and outputs the discriminant confidence, reconstruction error, prediction error, and latent variable representation during inference, and sends the data to the alarm decision module. The alarm decision module receives data from the multi-scale temporal feature engineering module and the training anomaly detection model module. Based on the inference output, it calculates the discrimination confidence score, normalized reconstruction error score, normalized prediction error score, and latent variable representation score. These scores are then fused and mapped into a unified comprehensive anomaly score according to their weights. Based on the comprehensive score and duration threshold, the alarm is classified into severe, moderate, and normal levels. When an alarm is triggered, a diagnostic package containing the trigger node, trigger time, and comprehensive anomaly score is generated and sent out.

Images