A method, device and medium for predicting the risk of torsional failure of a split conductor based on an LSTM neural network

Abstract

The application discloses a kind of based on LSTM neural network's split conductor torsional failure risk prediction method, equipment and medium, belong to electric power system monitoring technical field, including in split conductor key position deployment comprehensive monitoring system, real-time acquisition multidimensional monitoring data;Ice image data collected by high-definition camera is extracted icing feature, and environmental data is collected using environmental monitoring sensor array;The historical torsional angle data processing collected by angle sensor obtains historical torsional data;Feature fusion is carried out, and multi-dimensional input vector is constructed, and data preprocessing is carried out using standardization and normalization technology;A multi-layer long short-term memory neural network model is constructed to predict failure probability and residual safe life;Risk classification is carried out based on the predicted failure probability, a multi-level early warning mechanism is established, and corresponding maintenance recommendations are automatically generated according to the risk level.The application significantly improves the safety warning capability and operation and maintenance decision level of power transmission line.

Description

A method, device, and medium for predicting the risk of torsional failure of split conductors based on LSTM neural networks. Technical Field

[0001] This invention relates to the field of power system monitoring technology, specifically to a method, device, and medium for predicting the risk of torsional failure of split conductors based on LSTM neural networks. Background Technology

[0002] Split conductors are a widely used conductor structure in high-voltage transmission lines, consisting of multiple sub-conductors bundled together by spacers. Under severe weather conditions such as icing in winter and strong winds, split conductors are prone to torsion. Long-term accumulation of torsional stress can lead to conductor fatigue damage, potentially causing torsional failure accidents, including sub-conductors becoming entangled, spacer damage, and even wire breakage, posing a significant threat to the safe operation of the power system.

[0003] Current methods for monitoring split conductors mainly have the following problems:

[0004] Lack of failure risk prediction capability: Existing monitoring systems mainly perform real-time angle monitoring. Although they can measure the current torsion angle, they cannot predict future failure risks and remaining safe life, making it difficult to provide a scientific basis for maintenance decisions. Traditional threshold judgment methods are too simplistic and cannot consider the cumulative effect of torsional damage and complex failure mechanisms.

[0005] Long Short-Term Memory (LSTM) networks, as an improved version of recurrent neural networks, are capable of processing long-sequence time-series data and can effectively solve the gradient vanishing problem of traditional recurrent neural networks (RNNs), demonstrating excellent performance in complex time-series prediction tasks. Applying LSTM technology to the prediction of split conductor failure risks can fully utilize historical data to uncover hidden failure modes, providing strong technical support for the safe operation of power systems. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the technical problem solved by this invention is: how to address the fact that existing monitoring systems mainly perform real-time angle monitoring, which, while able to measure the current torsion angle, cannot predict future failure risks and remaining safe life, making it difficult to provide a scientific basis for maintenance decisions. Traditional threshold judgment methods are too simplistic and cannot consider the cumulative effects of torsional damage and complex failure mechanisms.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting the risk of torsional failure of split conductors based on LSTM neural networks, comprising: deploying a comprehensive monitoring system at key locations of the split conductor to collect multi-dimensional monitoring data in real time; the comprehensive monitoring system including a high-definition camera, an angle sensor, and an environmental monitoring sensor array; performing deep analysis on icing image data collected by the high-definition camera to identify icing areas and extracting icing features affecting torsional failure; using the environmental monitoring sensor array to collect environmental data such as wind speed, wind direction, temperature, humidity, and air pressure; processing historical torsional angle data collected by the angle sensor to obtain historical torsional data; and combining icing features, Environmental data and historical tortuosity data are fused to construct a multi-dimensional input vector, and data preprocessing is performed using standardization and normalization techniques. A multi-layer long short-term memory neural network model is constructed, including an input layer, multiple long short-term memory network hidden layers, an attention mechanism layer, a fully connected layer, and an output layer. The output layer is designed with a multi-output structure to simultaneously predict failure probability and remaining safe lifetime. The trained multi-layer long short-term memory neural network model is used to predict failure risk, and the output includes short-term failure probability, medium-term failure probability, and remaining safe lifetime estimate. Based on the predicted failure probability, risk classification is performed, a multi-level early warning mechanism is established, and corresponding maintenance recommendations are automatically generated according to the risk level.

[0009] As a preferred embodiment of the LSTM neural network-based method for predicting the risk of torsional failure of split conductors described in this invention, the real-time acquisition of multi-dimensional monitoring data includes the use of a unified time reference; the environmental monitoring array includes an anemometer, a temperature and humidity sensor, and a barometric pressure sensor.

[0010] As a preferred embodiment of the LSTM neural network-based method for predicting the risk of torsional failure of split conductors described in this invention, the following steps are included: performing deep analysis on icing image data acquired by a high-definition camera to identify icing areas and extracting icing features affecting torsional failure. This includes histogram equalization and contrast stretching of the icing image data, Gaussian filtering to remove noise, perspective transformation for geometric correction, and region segmentation. The icing features include icing thickness, asymmetry coefficient, and centroid shift. The icing thickness is calculated based on pixel statistics and geometric measurements. The asymmetry coefficient is expressed as:

[0011] ;

[0012] in, The asymmetric coefficient, This represents the amount of ice accumulation on the left side of the conductor. This represents the amount of ice accumulation on the right side of the conductor. The center of gravity offset is expressed as:

[0013] ;

[0014] in, The x-coordinate of the actual centroid of the conductor after icing. The ordinate of the actual centroid of the conductor after icing is shown. Let x be the x-coordinate of the centroid in conductor theory. The ordinate of the centroid in conductor theory is given by [reference]. This represents the distance of the center of gravity offset.

[0015] As a preferred embodiment of the LSTM neural network-based method for predicting the risk of torsion failure of split conductors described in this invention, the processing of historical torsion angle data collected by the angle sensor to obtain historical torsion data includes: performing statistical feature analysis on the historical torsion angle data to obtain the mean, variance, peak value, and 95th percentile; fatigue damage assessment uses the rainflow counting method, and the damage degree is expressed as:

[0016] ;

[0017] in, For cumulative damage, For the first The actual number of stress cycles that occur at the first stress level. For the first Fatigue life under stress level.

[0018] As a preferred embodiment of the LSTM neural network-based method for predicting the risk of torsion failure of split conductors according to the present invention, the step of fusing icing features, environmental data, and historical torsion data to construct a multi-dimensional input vector includes: constructing a standardized feature vector; the icing features include thickness, asymmetry coefficient, and centroid offset; the environmental data includes wind speed, temperature, humidity, and air pressure parameters; and the historical torsion data includes torsion mean, variance, dominant frequency, and damage degree parameters; the standardized representation is as follows:

[0019] ;

[0020] in, For standardized data, The original data, The mean, The standard deviation is denoted as: ; the normalized representation is:

[0021] ;

[0022] in, For the normalized data, To be the minimum value, This is the maximum value.

[0023] As a preferred embodiment of the LSTM neural network-based method for predicting the risk of torsion failure of split conductors according to the present invention, wherein: the plurality of long short-term memory network hidden layers include a first long short-term memory network, a second long short-term memory network, and a third long short-term memory network; the first long short-term memory network serves as the basic feature extraction layer, and is represented as:

[0024] ;

[0025] in, This is for the first layer of the Long Short-Term Memory network operation. for The input vector at time t, For a moment, for The hidden state output of the first layer of the Long Short-Term Memory network at each time step. for The hidden state output of the first layer of the Long Short-Term Memory network at each time step. for The cell states of the first layer of the Long Short-Term Memory (LSTM) network at any given time; the second layer of the LTM network, as the pattern association layer, is represented as follows:

[0026] ;

[0027] in, This is for the second layer of the Long Short-Term Memory network operation. for The hidden state output of the second layer of the Long Short-Term Memory network at time step. for The hidden state output of the second layer of the Long Short-Term Memory network at time step. for The cell state of the second-layer long short-term memory network at any given time; the third-layer long short-term memory network, as the risk reasoning layer, is represented as:

[0028] ;

[0029] in, This is for the third layer of the Long Short-Term Memory network operation. for The hidden state output of the third layer of the Long Short-Term Memory network at time step. for The hidden state output of the third layer of the Long Short-Term Memory network at time step. for The cellular state of the third layer of the long short-term memory network at any given moment.

[0030] This preferred solution adopts a three-layer LSTM neural network structure, with each layer undertaking the tasks of feature extraction, pattern recognition, and risk reasoning, respectively. This makes the model more hierarchical and expressive in capturing time-series information and complex patterns, thereby improving the accuracy of torsional failure risk prediction.

[0031] As a preferred embodiment of the LSTM neural network-based method for predicting the risk of torsional failure of split conductors according to the present invention, the attention mechanism layer includes attention weights, attention scores, and context vectors; the attention weights are expressed as:

[0032] ;

[0033] in, Let be the attention weight at time t. Let t be the attention score. This represents the total length of the time series. for Attention score at any moment For time intervals; attention scores are represented as:

[0034] ;

[0035] in, This is the transpose of the attention weight vector. For transpose, Here is the hidden state weight matrix. The sequence state weight matrix is... for The sequence state at time step 1; the hidden state at all time steps is weighted and fused using context vectors, and represented as:

[0036] ;

[0037] in, Let be the context vector at time t. Let k be the attention weight. Let k be the hidden state output of the Long Short-Term Memory network at time k; the failure probability is expressed as:

[0038] ;

[0039] in, This is the predicted failure probability value. This is the failure probability weight matrix. This is the failure probability bias vector. The sigmoid function is used; the remaining safe lifetime is expressed as:

[0040] ;

[0041] in, This is the predicted value of the remaining safe life. This is the remaining lifetime weight matrix. This is the remaining lifetime bias vector. This is the ReLU function.

[0042] By introducing an attention mechanism, this preferred scheme enables the model to dynamically monitor the hidden states at key time steps and improve the accuracy and interpretability of predictions by weighted integration of contextual information.

[0043] As a preferred embodiment of the LSTM neural network-based method for predicting the torsional failure risk of split conductors according to the present invention, the method involves: classifying risks based on the predicted failure probability, establishing a multi-level early warning mechanism, and automatically generating corresponding maintenance recommendations according to the risk level, including calculating a comprehensive risk index, expressed as:

[0044] ;

[0045] in, This is a comprehensive risk index. This represents the probability of failure. For the designed service life, For remaining lifespan, The current damage status, These are the weighting coefficients for the failure probability. The weighting factor is the service life. The weighting coefficients represent the damage status; a four-level risk classification is established based on the comprehensive risk index, with a first threshold, a second threshold, and a third threshold set. If the risk level is below the first threshold, it corresponds to a low-risk condition and a healthy state; if... If the risk level is greater than or equal to the first threshold but less than the second threshold, it is considered medium risk and monitoring should be strengthened; if... If the value is greater than or equal to the second threshold but less than the third threshold, it is considered high-risk and requires preventative maintenance; if If the value is greater than or equal to the third threshold, it is considered an extremely high risk and requires emergency action.

[0046] This preferred solution constructs a comprehensive risk index based on the predicted failure probability, and combines lifespan and damage status to achieve graded early warning. This not only improves the intelligence level of power grid equipment management, but also provides quantitative basis and decision support for the scientific formulation of maintenance strategies.

[0047] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the method for predicting the risk of torsion failure of split conductors based on LSTM neural networks.

[0048] The present invention provides a computer-readable storage 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 for predicting the risk of torsion failure of a split conductor based on an LSTM neural network.

[0049] The beneficial effects of this invention are as follows: By deploying a comprehensive monitoring system, this invention collects multi-source data such as icing images, torsion angles, and environmental parameters in real time. Deep learning technology is used to extract the unevenness characteristics of icing distribution and analyze historical torsion patterns and damage accumulation patterns. A specialized multi-layer LSTM neural network is constructed, fusing multi-dimensional features for time-series modeling, and a gating mechanism is used to effectively handle long-term dependencies and complex time-series patterns. The network adopts a multi-output structure, simultaneously predicting failure probability and remaining safe lifetime, and introduces an attention mechanism to enhance the identification ability of critical time periods. A multi-level risk assessment system and dynamic early warning mechanism are established, automatically generating maintenance suggestions based on prediction results. This provides strong technical support for the safe operation and preventive maintenance of power systems, significantly improving the safety early warning capability and operation and maintenance decision-making level of transmission lines. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 is a flowchart of an overall method for predicting the risk of torsional failure of split conductors based on LSTM neural networks according to an embodiment of the present invention.

[0052] Figure 2 is a schematic diagram of the neural network principle of a split conductor torsion failure risk prediction method based on LSTM neural network provided in an embodiment of the present invention. Detailed Implementation

[0053] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0054] Example 1, referring to Figure 1, is an embodiment of the present invention. This embodiment provides a method for predicting the risk of torsional failure of split conductors based on LSTM neural networks, including:

[0055] S1. Deploy a comprehensive monitoring system at key locations of the split conductor to collect multi-dimensional monitoring data in real time. The comprehensive monitoring system includes high-definition cameras, angle sensors, and environmental monitoring sensor arrays.

[0056] S2. Perform in-depth analysis on the icing image data collected by the high-definition camera to identify the icing area, extract the icing characteristics that affect torsional failure, and use an environmental monitoring sensor array to collect environmental data such as wind speed, wind direction, temperature, humidity, and air pressure.

[0057] S3. Process the historical torsion angle data collected by the angle sensor to obtain historical torsion data.

[0058] S4. Feature fusion is performed on icing characteristics, environmental data, and historical torsion data to construct a multi-dimensional input vector, and data preprocessing is carried out using standardization and normalization techniques.

[0059] S5. Construct a multi-layer long short-term memory neural network model including an input layer, multiple long short-term memory network hidden layers, an attention mechanism layer, a fully connected layer, and an output layer. The output layer is designed as a multi-output structure to simultaneously predict the failure probability and remaining safe lifetime.

[0060] S6. Use the trained multi-layer long short-term memory neural network model to predict failure risk, and output short-term failure probability, medium-term failure probability and remaining safe lifetime estimate.

[0061] S7. Based on the predicted failure probability, risk classification is carried out, a multi-level early warning mechanism is established, and corresponding maintenance suggestions are automatically generated according to the risk level.

[0062] It should be noted that split conductors are susceptible to icing, wind loads, and other factors under complex weather conditions, leading to asymmetrical torsion. Prolonged abnormal torsion angles can cause conductor instability, damage, or even breakage. Traditional manual inspection methods are time-consuming and have significant delays, making it impossible to dynamically identify and warn of torsion failure risks. Especially in freezing weather, high-wind areas, or frequent thunderstorms, conductors face even higher operational risks.

[0063] This invention achieves comprehensive monitoring and prediction of conductor torsion risk by deploying integrated sensing devices at key parts of conductors and combining depth image processing and time series modeling technology. The use of LSTM neural networks to fuse multi-source data significantly improves prediction accuracy and timeliness. The accompanying risk classification mechanism and maintenance recommendation strategy make power operation and maintenance more targeted, intelligent and forward-looking, effectively ensuring the safe and stable operation of the conductor system.

[0064] Example 2, referring to Figure 2, is an embodiment of the present invention. Based on the previous embodiment, it provides a method for predicting the risk of torsional failure of split conductors using an LSTM neural network, including:

[0065] Furthermore, in step S1, a comprehensive monitoring system is deployed at key locations of the split conductor to collect multi-dimensional monitoring data in real time. The comprehensive monitoring system includes high-definition cameras to acquire images of the conductor icing, angle sensors to measure the actual torsion angle, and an environmental monitoring sensor array to collect environmental parameters such as wind speed, wind direction, temperature, humidity, and air pressure. The simultaneous acquisition of multi-source data provides a comprehensive information foundation for subsequent failure risk analysis, ensuring that the predictive model can capture various key factors affecting conductor failure.

[0066] The high-definition camera employs binocular stereo vision technology with a resolution of at least 1920*1080, capturing images every 10 minutes. An infrared supplemental lighting system is activated at night to ensure image quality. The high-precision angle sensor utilizes MEMS technology, achieving an accuracy of at least 0.1°, capturing images every minute, and features temperature compensation and vibration resistance.

[0067] The micro-weather station integrates multiple sensors such as temperature, humidity, wind speed, wind direction, and air pressure, collecting data once per minute, and supports real-time data transmission and local storage. The system operates in a temperature range of -40°C to +70°C, has an IP67 protection rating, and is adaptable to harsh environments such as lightning strike protection, corrosion resistance, and icing prevention, ensuring stable operation under various extreme weather conditions.

[0068] Furthermore, in step S2, deep analysis is performed on the acquired icing images to identify icing regions and extract icing feature parameters that affect torsional failure. First, image preprocessing is performed, including noise reduction filtering, contrast enhancement, and edge detection, to improve image quality and highlight icing features. Then, an improved deep learning object detection algorithm is used to identify icing regions and accurately segment the icing portions. The non-uniformity of icing distribution is analyzed in detail, calculating parameters such as icing thickness differences, asymmetry coefficient, and centroid shift. These parameters directly affect the torsional stress distribution and failure risk of the conductor.

[0069] Icing features include icing thickness, asymmetry coefficient, and centroid offset; icing thickness is calculated based on pixel statistics and geometric measurements.

[0070] The asymmetry coefficient is expressed as:

[0071] ;

[0072] in, This is the asymmetry coefficient, ranging from 0 to 1. A larger value indicates a more uneven icing. This represents the amount of ice accumulation on the left side of the conductor; This represents the amount of ice accumulation on the right side of the conductor. .

[0073] The center of gravity offset is expressed as:

[0074] ;

[0075] in, The x-coordinate of the actual centroid of the conductor after icing. The ordinate of the actual centroid of the conductor after icing is shown. Let x be the x-coordinate of the centroid in conductor theory. The ordinate of the centroid in conductor theory is given by [reference]. This represents the distance of the center of gravity offset.

[0076] Furthermore, in step S3, the historical torsion angle data collected by the angle sensor is processed to obtain historical torsion data, including:

[0077] Based on historical torsional angle data collected by angle sensors, this study analyzes the torsional behavior patterns and damage accumulation laws of conductors. Statistical characteristics of the torsional angles, including mean, variance, and peak frequency, are calculated, and the amplitude distribution and frequency characteristics of torsional cycles are analyzed. Rainflow counting is used to statistically analyze stress cycles and assess the degree of fatigue damage accumulation. Through in-depth analysis of historical data, regularities and abnormal patterns in conductor torsional behavior are identified, providing crucial prior knowledge for failure risk prediction.

[0078] Statistical characteristic analysis was performed on historical torsional angle data to obtain the mean, variance, peak value, and 95th percentile. Fatigue damage assessment was conducted using the rainflow counting method, and the damage degree was expressed as follows:

[0079] ;

[0080] in, This represents the cumulative damage level, ranging from 0 to 1, with values ​​close to 1 indicating near failure. For the first The actual number of stress cycles that occur at each stage; For the first Fatigue life under stress level.

[0081] Furthermore, in step S4, icing features, environmental data, and historical torsion data are fused to construct a multi-dimensional input vector. Data preprocessing is then performed using standardization and normalization techniques, including:

[0082] Feature fusion was performed on icing characteristics, environmental data, and historical torsion patterns to construct a multidimensional input vector. Considering the differences in the dimensions and numerical ranges of different data types, standardization and normalization techniques were employed for data preprocessing.

[0083] A standardized feature vector is constructed. The icing features include parameters such as thickness, asymmetry coefficient, and center of gravity shift. The environmental data includes parameters such as wind speed, temperature, humidity, and air pressure. The historical torsion data includes parameters such as torsion mean, variance, dominant frequency, and damage degree.

[0084] The standardized representation is:

[0085] ;

[0086] in, For standardized data, The original data, The mean, Standard deviation;

[0087] Normalized representation is:

[0088] ;

[0089] in, For the normalized data, To be the minimum value, This is the maximum value.

[0090] Standardized data is beneficial for the training and convergence of neural networks, ensuring that the contributions of each dimension of features to model training are relatively balanced.

[0091] Furthermore, in step S5, a multi-layered long short-term memory neural network model is constructed, including an input layer, multiple long short-term memory network hidden layers, an attention mechanism layer, a fully connected layer, and an output layer. The output layer is designed as a multi-output structure, simultaneously predicting the failure probability and remaining safe lifetime, as shown in Figure 2, including:

[0092] A multi-layer LSTM neural network model specifically designed for failure risk prediction was constructed. The network structure includes an input layer, multiple LSTM hidden layers, an attention mechanism layer, a fully connected layer, and an output layer. The input layer receives multi-dimensional temporal feature vectors; the LSTM layer extracts temporal dependencies and long-term evolution patterns; the attention mechanism enhances focus on key time periods; and the fully connected layer performs feature fusion and non-linear mapping. The output layer is designed as a multi-output structure to simultaneously predict failure probability and remaining safe lifetime.

[0093] LSTM addresses the long-term dependency problem of traditional RNNs by introducing a gating mechanism, a characteristic particularly suitable for failure risk prediction tasks. Torsional failure of a conductor is a long-term cumulative process involving complex load history effects and damage evolution patterns. Traditional recurrent neural networks suffer from vanishing or exploding gradients when processing long sequences, preventing the network from effectively learning long-term dependencies. An LSTM unit contains three key gating structures: a forget gate, an input gate, and an output gate, as well as a cell state to store long-term information.

[0094] The forget gate determines what historical information is discarded from the cell state, corresponding to the fading of certain outdated load histories in failure prediction. The input gate determines what new information is stored in the cell state, corresponding to the memorization of new damage information. The output gate controls what information is output based on the cell state, corresponding to the failure risk assessment in the current state. This ingenious gating mechanism allows LSTM to selectively remember or forget historical information, thus effectively handling long-term cumulative effects and complex time dependencies in failure prediction.

[0095] Furthermore, the computation process of LSTM is a sophisticated mathematical model, where each gated unit has a specific function and physical meaning in failure prediction. For time... The complete LSTM computation process includes the following steps:

[0096] The formula for calculating the forgetting gate is:

[0097] ;

[0098] in, for The output of the forget gate; Here is the weight matrix for the forget gate; This is the bias vector for the forget gate; It is the sigmoid activation function; for The hidden state at any given time carries historical risk information; for The input vector at any given time contains the current load and environmental conditions.

[0099] In failure prediction, the forget gate determines the degree to which historical damage information is retained. The output value is between 0 and 1, where 1 indicates complete retention of historical damage memory and 0 indicates complete forgetting of outdated load history. This mechanism allows the model to adaptively adjust the importance of historical information.

[0100] The formula for calculating the input gate is:

[0101] ;

[0102] in, for The output of the input gate at all times, Here is the weight matrix of the input gate. is the bias vector of the input gate.

[0103] The input gate determines how new damage information at the current moment is memorized and stored. It works with candidate values ​​to update the damage accumulation state and control the impact of new loads and environmental conditions on the failure risk.

[0104] The formula for calculating the candidate value is:

[0105] ;

[0106] in, for Candidate values ​​at time 10:00 The weight matrix for the candidate values, is the bias vector for candidate values.

[0107] Candidate values ​​represent new information that may be added to the damage state at the current moment, including current load conditions, environmental stress, and icing status. The tanh activation function is used to ensure that the values ​​are between -1 and 1, providing a bidirectional damage effect (increasing or decreasing damage).

[0108] The cell state update formula is:

[0109] ;

[0110] in, for Cellular state at any given moment for Cellular state at any given moment This indicates element-wise multiplication, enabling selective filtering through a gating mechanism.

[0111] Cell state updates are central to failure prediction, representing the cumulative state of damage. This formula combines historical damage states (filtered through a forget gate) with newly added damage (filtered through an input gate), enabling the dynamic accumulation and evolution of damage.

[0112] The formula for calculating the output gate is:

[0113] ;

[0114] in, for The output of the output gate is always being output; The weight matrix of the output gate is used to learn the characteristics of various failure modes through training. The bias vector for the output gate provides flexibility and adaptability to the model.

[0115] The formula for outputting in the hidden state is:

[0116] ;

[0117] in, for The status output is always hidden.

[0118] The output gate determines which parts of the current damage state will be used for failure risk assessment. It controls how the internal damage state is transformed into externally observable risk indicators, achieving a mapping from latent damage to explicit risk.

[0119] In the above formula, For the sigmoid activation function:

[0120]

[0121] in, This is the input variable for the sigmoid activation function.

[0122] The output range is (0, 1), which is suitable for representing probabilities and weights.

[0123] Furthermore, this invention employs a three-layer LSTM structure, specifically optimized for failure prediction tasks. The multi-layer structure can extract feature representations at different levels, from basic temporal patterns at the lower layers to complex failure patterns at the higher layers. The three layers are as follows:

[0124] The first layer of the Long Short-Term Memory (LSTM) network serves as the basic feature extraction layer, and the calculation formula is as follows:

[0125] ;

[0126] in, This is for the first layer of the Long Short-Term Memory network operation. for The input vector at time t, For a moment, for The hidden state output of the first layer of the Long Short-Term Memory network at each time step. for The hidden state output of the first layer of the Long Short-Term Memory network at each time step. for The cellular state of the first layer of the Long Short-Term Memory network at any given moment.

[0127] The main focus is on learning basic time-series patterns, such as daily cycles and seasonal variations.

[0128] The second layer of the Long Short-Term Memory (LSTM) network serves as the pattern association layer, and its calculation formula is as follows:

[0129] ;

[0130] in, This is for the second layer of the Long Short-Term Memory network operation. for The hidden state output of the second layer of the Long Short-Term Memory network at time step. for The hidden state output of the second layer of the Long Short-Term Memory network at time step. for The cellular state of the second layer of the long short-term memory network at any given moment.

[0131] Learn more complex association patterns based on the first layer of features.

[0132] The third layer of the Long Short-Term Memory (LSTM) network serves as the risk inference layer, and its calculation formula is as follows:

[0133] ;

[0134] in, This is for the third layer of the Long Short-Term Memory network operation. for The hidden state output of the third layer of the Long Short-Term Memory network at time step. for The hidden state output of the third layer of the Long Short-Term Memory network at time step. for The cellular state of the third layer of the long short-term memory network at any given moment.

[0135] It is specifically responsible for advanced inference of failure risk, integrating multi-layer feature information. The decreasing dimensionality design of each LSTM layer realizes a hierarchical representation from concrete features to abstract concepts, improving the model's ability to understand complex failure modes.

[0136] The attention mechanism layer dynamically focuses on the time period most critical to failure risk.

[0137] The formula for calculating attention weights is:

[0138] ;

[0139] in, for The attention weight at any given moment indicates the importance of that moment in predicting the current failure risk. for The attention score at any given moment is calculated using the following formula; This represents the total time series length. The softmax function ensures that the sum of the weights is 1.

[0140] The formula for calculating attention score is:

[0141] ;

[0142] in, This is the transpose of the attention weight vector. For transpose, This is the hidden state weight matrix, used to process the LSTM output; for The hidden state of the LSTM at time t, derived from the output of the LSTM in the third layer; This is the sequence state weight matrix, used to process historical sequence information; It represents the sequence state from the previous moment and is used to carry historical information.

[0143] The formula for calculating the hidden state across all time steps using context vector weighted fusion is:

[0144] ;

[0145] in, For a moment The context vector is a weighted sum of all historical information; For a moment The attention weight determines the degree of contribution of information at that moment; For a moment The hidden state of the LSTM.

[0146] This attention mechanism can automatically identify the most important historical moments for current failure risk, improving the accuracy and interpretability of predictions. By visualizing the attention weight distribution, operations personnel can intuitively understand the model's decision-making basis, enhancing the credibility of prediction results.

[0147] Furthermore, the multi-output layer design adopts a multi-task learning architecture to predict multiple failure risk indicators simultaneously.

[0148] The formula for calculating the failure probability output is:

[0149] ;

[0150] in, This is the predicted failure probability value. This is the failure probability weight matrix. This is the failure probability bias vector. This is the sigmoid function.

[0151] The formula for calculating the remaining lifetime output is as follows:

[0152] ;

[0153] in, This is the predicted value of the remaining safe life. This is the remaining lifetime weight matrix. This is the remaining lifetime bias vector. The ReLU function is used to ensure that lifetime predictions are non-negative.

[0154] Output These represent low, medium, high, and extremely high risks.

[0155] The multi-task learning architecture improves the model's generalization ability and prediction accuracy by sharing underlying feature representations. It also outputs multiple relevant metrics, providing more comprehensive information support for operational decisions and avoiding the limitations of relying on a single metric.

[0156] Furthermore, the model training employs a multi-task loss function optimization strategy.

[0157] The failure probability loss uses binary cross-entropy, which is suitable for binary classification problems. The calculation formula is as follows:

[0158] ;

[0159] in, For failure probability loss, The total number of samples, As a sample, For the first The true failure label of each sample For the first The probability of prediction failure for each sample.

[0160] The remaining lifetime loss is calculated using Huber loss, and the formula is as follows:

[0161] ;

[0162] in, For remaining life loss, For the first The true remaining lifetime of each sample For the first Predicted remaining lifespan for each sample The threshold for Huber loss is determined by analyzing the historical prediction error distribution and is typically set to 1-2 times the standard deviation of the prediction error.

[0163] The total loss function is:

[0164] ;

[0165] in, The weighting coefficients for the failure probability loss are... This is the weighting factor for remaining lifetime loss. These are the weight coefficients for the L2 regularization term. This is the L2 regularization term. Huber loss is robust to outliers and suitable for data noise that may exist in lifetime prediction tasks. Multi-task learning improves the overall performance and stability of the model by jointly optimizing multiple related tasks.

[0166] Furthermore, in step S7, risk classification is performed based on the predicted failure probability, a multi-level early warning mechanism is established, and corresponding maintenance suggestions are automatically generated according to the risk level, including:

[0167] The comprehensive risk index is calculated as follows:

[0168] ;

[0169] in, This is a comprehensive risk index. This represents the probability of failure. For the designed service life, For remaining lifespan, The current damage status, These are the weighting coefficients for the failure probability. The weighting factor is the service life. The weighting coefficient for the damage state is used to comprehensively consider risk factors in three dimensions: probability, time, and state.

[0170] The risk level is divided into four levels based on the comprehensive risk index, and a first threshold, a second threshold, and a third threshold are set. In this invention, the first threshold is 0.3, the second threshold is 0.6, and the third threshold is 0.8.

[0171] like A risk index of <0.3, meaning a comprehensive risk index less than the first threshold, indicates low risk and corresponds to a healthy state.

[0172] If 0.3≤ If the comprehensive risk index is less than 0.6, meaning it is greater than or equal to the first threshold but less than the second threshold, it is considered a medium risk, corresponding to a mild abnormal state, and requires enhanced monitoring.

[0173] If 0.6≤ A risk index of <0.8, meaning a comprehensive risk index greater than or equal to the second threshold but less than the third threshold, indicates high risk, corresponding to a severe emergency situation, requiring preventative maintenance.

[0174] like A risk index of ≥0.8, or greater than or equal to the third threshold, indicates extremely high risk and a critical state requiring immediate action.

[0175] Example 3 is an embodiment of the present invention, which provides a method for predicting the risk of torsional failure of split conductors based on LSTM neural networks. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0176] Integrated monitoring devices are deployed at transmission line monitoring points. The visual monitoring module uses industrial-grade high-definition cameras with a resolution of 1920×1080, IP67 protection rating, and an operating temperature range of -40℃ to +70℃. The angle sensor module uses a 9-axis inertial measurement unit with a measurement accuracy of no less than 0.1° and a sampling frequency of once every 10 minutes. The environmental monitoring array includes anemometers, temperature and humidity sensors, and barometric pressure sensors.

[0177] Data acquisition uses a unified time base: visual data is collected every ten minutes, angle data once per minute, and environmental data once per minute. The communication solution employs a redundant configuration of a 4G / 5G primary channel and a satellite backup channel to ensure reliable data transmission.

[0178] Furthermore, the image processing employs a multi-step standardized process, including histogram equalization and contrast stretching to improve image quality, Gaussian filtering to remove noise, perspective transformation for geometric correction, and region segmentation and recognition based on deep learning algorithms.

[0179] Ice accretion feature extraction enables multi-dimensional quantitative analysis. Ice accretion thickness is calculated based on pixel statistics and geometric measurements, and the asymmetry coefficient is obtained using the formula:

[0180] ;

[0181] in, This represents the amount of ice accumulation on the left side of the conductor. This represents the amount of ice accumulation on the right side of the conductor. , This is the asymmetry coefficient, ranging from 0 to 1. A larger value indicates a more uneven icing.

[0182] The center of gravity offset is expressed by the formula:

[0183] ;

[0184] in, The x-coordinate of the actual centroid of the conductor after icing. The ordinate of the actual centroid of the conductor after icing is shown. Let x be the x-coordinate of the centroid in conductor theory. The ordinate of the centroid in conductor theory is given by [reference]. This represents the distance of the center of gravity offset.

[0185] Furthermore, historical torsional data analysis employs a multi-level time-series analysis method. Torsional angle data processing includes statistical characteristic analysis, calculating the mean, variance, peak-to-peak value, and 95th percentile. Frequency analysis uses Fast Fourier Transform to identify the main frequency components. Fatigue damage assessment uses the rainflow counting method; the damage degree calculation formula is:

[0186] ;

[0187] in, This represents the cumulative damage level, ranging from 0 to 1, with values ​​close to 1 indicating near failure. For the first The actual number of stress cycles that occur at the first stress level. For the first Fatigue life (the total number of theoretically withstandable cycles) under stress level.

[0188] Furthermore, standardized feature vectors are constructed. The icing feature vector includes parameters such as thickness, asymmetry coefficient, and center of gravity shift. The environmental feature vector includes parameters such as wind speed, temperature, humidity, and air pressure. The historical pattern feature vector includes parameters such as torsional mean, variance, dominant frequency, and damage degree.

[0189] Data preprocessing employs a systematic quality control process, using linear interpolation to fill in missing data, the 3σ principle to detect outliers, and Z-score standardization for features of different dimensions.

[0190] Furthermore, the LSTM network employs a multi-layer architecture, with the input dimension being the feature vector length × a 168-hour time window. The LSTM layers are configured with decreasing neurons from 128 to 64 to 32, with a dropout rate of 0.2-0.3. The fully connected layers consist of two dense layers with 16 and 8 neurons respectively. The output layer uses a multi-output design, including outputs of failure probability, remaining lifetime, and risk level.

[0191] The model was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The dataset was divided in a 7:2:1 ratio. The loss function adopted was a weighted combination of multi-task learning loss functions.

[0192] Furthermore, the system runs hourly, automatically inputting the latest 168 hours of monitoring data and outputting the failure probability and remaining life prediction results for the next 1-7 days. The prediction results are smoothed using a moving average, and abnormal predictions trigger a manual review mechanism.

[0193] The model is updated regularly through an online learning mechanism, and an ensemble learning strategy is used to improve prediction stability. The model performance is continuously monitored and retraining is automatically triggered.

[0194] Furthermore, a multi-indicator risk assessment system is established, and the comprehensive risk index is calculated using the following formula:

[0195] ;

[0196] in, This represents the failure probability, ranging from 0 to 1. Design service life (years). Remaining lifespan (years). This represents the percentage of lifespan consumed. The current damage status, This is a comprehensive risk index, ranging from 0 to 1. The failure probability accounts for 40% of the weight, the normalized remaining lifetime accounts for 40%, and the current damage state accounts for 20%.

[0197] The early warning decision-making process employs a tiered response mechanism: low-risk areas maintain routine inspections, medium-risk areas increase monitoring frequency to every 30 minutes, high-risk areas arrange special inspections and prepare emergency materials, and extremely high-risk areas activate emergency response. Early warning information is disseminated through multiple channels, including SMS, email, and mobile apps.

[0198] Furthermore, a layered software architecture is adopted, with the data processing layer responsible for multi-source data acquisition and storage, the model service layer deploying LSTM models to provide prediction service APIs, and the application presentation layer developing web and mobile applications to achieve visualization.

[0199] Performance optimizations include deploying GPU servers to accelerate model inference, using data compression and incremental transmission to optimize data transmission, establishing off-site backups to ensure high availability, and building a comprehensive performance monitoring system to track system operation status.

[0200] Example 4 is an embodiment of the present invention, which provides a device and medium for predicting the risk of torsional failure of split conductors based on LSTM neural networks, including:

[0201] This embodiment also provides an electronic device applicable to a method for predicting the risk of torsion failure of split conductors based on an LSTM neural network, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for predicting the risk of torsion failure of split conductors based on an LSTM neural network as proposed in the above embodiment.

[0202] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for predicting the risk of torsion failure of split conductors based on an LSTM neural network as proposed in the above embodiment.

[0203] The storage medium proposed in this embodiment and the method for predicting the risk of torsion failure of split conductors based on LSTM neural networks proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0204] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0205] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the risk of torsional failure of split conductors based on LSTM neural networks, characterized in that: This includes deploying a comprehensive monitoring system at key locations on the split conductor to collect multi-dimensional monitoring data in real time. The system comprises high-definition cameras, angle sensors, and an environmental monitoring sensor array. Deep analysis is performed on icing images collected by the high-definition cameras to identify icing areas and extract icing features affecting torsional failure. The environmental monitoring sensor array collects environmental data such as wind speed, wind direction, temperature, humidity, and air pressure. Historical torsional angle data collected by the angle sensors is processed to obtain historical torsional data. Icing features, environmental data, and historical torsional data are fused to construct a multi-dimensional input vector, and data preprocessing is performed using standardization and normalization techniques. A multi-layered long short-term memory neural network model is constructed, including an input layer, multiple long short-term memory hidden layers, an attention mechanism layer, a fully connected layer, and an output layer. The layered design employs a multi-output structure to simultaneously predict failure probability and remaining safe lifetime. A trained multi-layer long short-term memory neural network model is used for failure risk prediction, outputting short-term failure probability, medium-term failure probability, and remaining safe lifetime estimates. Risk grading is performed based on the predicted failure probabilities, establishing a multi-level early warning mechanism and automatically generating corresponding maintenance recommendations according to the risk level. The deep analysis of icing image data acquired by a high-definition camera identifies icing areas and extracts icing features affecting torsional failure. This includes histogram equalization and contrast stretching of the icing image data, Gaussian filtering to remove noise, perspective transformation for geometric correction, and region segmentation. Icing features include icing thickness, asymmetry coefficient, and centroid shift. Icing thickness is calculated based on pixel statistics and geometric measurements. The asymmetry coefficient is expressed as... in, The asymmetric coefficient, This represents the amount of ice accumulation on the left side of the conductor. This represents the amount of ice accumulation on the right side of the conductor. The centroid offset is expressed as, in, The x-coordinate of the actual centroid of the conductor after icing. The ordinate of the actual centroid of the conductor after icing is shown. Let x be the x-coordinate of the centroid in conductor theory. The ordinate of the centroid in conductor theory is given by [reference]. This represents the distance of the center of gravity offset.

2. The method for predicting the risk of torsional failure of split conductors based on LSTM neural networks as described in claim 1, characterized in that: The real-time acquisition of multi-dimensional monitoring data includes the use of a unified time reference; the environmental monitoring array includes an anemometer, temperature and humidity sensor, and air pressure sensor.

3. The method for predicting the risk of torsional failure of split conductors based on LSTM neural networks as described in claim 2, characterized in that: The historical torsional angle data collected by the angle sensor is processed to obtain historical torsional data, including statistical characteristic analysis of the historical torsional angle data to obtain the mean, variance, peak value, and 95th percentile. Fatigue damage assessment uses the rainflow counting method, and the damage degree is expressed as... in, For cumulative damage, For the first The actual number of stress cycles that occur at the first stress level. For the first Fatigue life under stress level.

4. The method for predicting the risk of torsional failure of split conductors based on LSTM neural networks as described in claim 3, characterized in that: The process of fusing icing features, environmental data, and historical torsion data to construct a multi-dimensional input vector includes: constructing a standardized feature vector; icing features including thickness, asymmetry coefficient, and center of gravity shift; environmental data including wind speed, temperature, humidity, and air pressure parameters; and historical torsion data including torsion mean, variance, dominant frequency, and damage degree parameters; the standardized representation is as follows: in, For standardized data, The original data, The mean, The standard deviation is denoted as ; the normalized expression is . in, For the normalized data, To be the minimum value, This is the maximum value.

5. The method for predicting the risk of torsional failure of split conductors based on LSTM neural networks as described in claim 4, characterized in that: The multiple long short-term memory (LSTM) network hidden layers include a first LTM network, a second LTM network, and a third LTM network; the first LTM network serves as the basic feature extraction layer, and is represented as follows: in, This is for the first layer of the Long Short-Term Memory network operation. for The input vector at time t, For a moment, for The hidden state output of the first layer of the Long Short-Term Memory network at each time step. for The hidden state output of the first layer of the Long Short-Term Memory network at each time step. for The cell states of the first layer of the Long Short-Term Memory (LSTM) network at any given time; the second layer of the LTM network, as the pattern association layer, is represented as follows: in, This is for the second layer of the Long Short-Term Memory network operation. for The hidden state output of the second layer of the Long Short-Term Memory network at time step. for The hidden state output of the second layer of the Long Short-Term Memory network at time step. for The cell state of the second-layer long short-term memory network at any given time; the third-layer long short-term memory network, serving as the risk reasoning layer, is represented as follows: in, This is for the third layer of the Long Short-Term Memory network operation. for The hidden state output of the third layer of the Long Short-Term Memory network at time step. for The hidden state output of the third layer of the Long Short-Term Memory network at time step. for The cellular state of the third layer of the long short-term memory network at any given moment.

6. The method for predicting the risk of torsional failure of split conductors based on LSTM neural networks as described in claim 5, characterized in that: The attention mechanism layer includes attention weights, attention scores, and context vectors; the attention weights are represented as follows: in, Let be the attention weight at time t. Let t be the attention score. This represents the total length of the time series. for Attention score at any moment For time; attention score is represented as, in, This is the transpose of the attention weight vector. For transpose, Here is the hidden state weight matrix. The sequence state weight matrix is... for The sequence state at time step 1; the hidden state at all time steps is weighted and fused using context vectors, and represented as follows: in, Let be the context vector at time t. Let k be the attention weight. Let be the hidden state output of the Long Short-Term Memory network at time k; the failure probability is expressed as . in, This is the predicted failure probability value. This is the failure probability weight matrix. This is the failure probability bias vector. The sigmoid function is used; the remaining safe lifetime is expressed as... in, This is the predicted value of the remaining safe life. This is the remaining lifetime weight matrix. This is the remaining lifetime bias vector. This is the ReLU function.

7. The method for predicting the risk of torsional failure of split conductors based on LSTM neural networks as described in claim 6, characterized in that: The system employs risk grading based on predicted failure probabilities, establishes a multi-level early warning mechanism, and automatically generates corresponding maintenance recommendations according to the risk level. This includes calculating the comprehensive risk index, expressed as follows: in, This is a comprehensive risk index. This represents the probability of failure. For the designed service life, For remaining lifespan, The current damage status, These are the weighting coefficients for the failure probability. The weighting factor is the service life. The weighting coefficients represent the damage status; a four-level risk classification is established based on the comprehensive risk index, with a first threshold, a second threshold, and a third threshold set. If the risk level is below the first threshold, it corresponds to a low-risk condition and a healthy state; if... If the risk level is greater than or equal to the first threshold but less than the second threshold, it is considered medium risk and monitoring should be strengthened; if... If the value is greater than or equal to the second threshold but less than the third threshold, it is considered high-risk and requires preventative maintenance; if If the value is greater than or equal to the third threshold, it is considered an extremely high risk and requires emergency action.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for predicting the risk of torsion failure of a split conductor based on an LSTM neural network, as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for predicting the risk of torsion failure of a split conductor based on an LSTM neural network, as described in any one of claims 1 to 7.

Images