Intelligent sensing and automatic adapting electric pole maintenance prediction system

Abstract

The invention relates to an intelligent sensing and automatic adapting electric pole maintenance prediction system, and aims to improve the electric pole maintenance efficiency and prediction accuracy. The system comprises a sensing unit installed on an electric pole and used for collecting physical state data and environmental factor data of the electric pole. And the prediction unit calculates the residual life and the potential fault risk of the electric pole through a pre-trained neural network prediction model by using the data. The decision-making unit generates specific maintenance suggestions and optimized maintenance schemes according to the prediction result, and then the suggestions and schemes are provided for a maintenance team to be executed. The system further comprises a self-adaptive unit which receives feedback information from a maintenance team and adjusts parameters of a neural network prediction model according to the feedback information, so that the accuracy and responsiveness of prediction are continuously optimized to cope with complex environmental changes and the actual state of the electric pole. The system can carry out efficient and accurate electric pole maintenance planning on the real-time basis, so that the timeliness and effectiveness of maintenance operation are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electric pole maintenance, and in particular to an electric pole maintenance prediction system with intelligent perception and automatic adaptation. Background Art

[0002] In the power industry, maintaining and monitoring the health of utility poles is a critical task for ensuring power supply stability and preventing equipment failure. Existing technologies typically employ various sensors and monitoring devices installed on utility poles to collect data on their physical condition, such as tilt and vibration levels, as well as environmental factors like temperature, humidity, and weather conditions. This data is processed by traditional monitoring systems to assess the condition of the poles and predict potential failures. These systems rely on preset thresholds and simple data processing models to determine when maintenance or replacement is necessary.

[0003] However, existing technologies suffer from several major issues. First, these systems often rely on static data processing methods and lack the ability to deeply analyze complex data patterns, which limits their performance in terms of predictive accuracy. Second, traditional systems are generally unable to update and automatically adjust predictive models in real time to reflect new data or environmental changes. This results in predictions that may not be accurate or timely under rapidly changing environmental conditions. Furthermore, existing systems often have simple feedback mechanisms and fail to fully utilize operational data from maintenance teams to optimize system performance.

[0004] To address these issues, it is necessary to develop a new intelligent perception and automatic adaptation pole maintenance prediction system. Summary of the Invention

[0005] The present application provides an intelligent perception and automatic adaptation power pole maintenance prediction system to improve the efficiency and prediction accuracy of power pole maintenance.

[0006] This application provides an intelligent perception and automatic adaptation power pole maintenance prediction system, including:

[0007] A sensing unit, mounted on a pole, configured to collect physical state data and environmental factor data of the pole, wherein the physical state data includes the inclination angle, vibration, temperature, and humidity of the pole, and the environmental factor data includes wind speed, rainfall, and air pressure;

[0008] The prediction unit is used to predict the remaining life and potential failure risk of the pole based on the physical status data and environmental factor data collected by the sensing unit and using a pre-trained neural network prediction model;

[0009] A decision-making unit, configured to generate maintenance recommendations and optimized maintenance plans for the poles based on the prediction results of the prediction unit; and provide the generated maintenance recommendations and optimized maintenance plans to a maintenance team;

[0010] The adaptive unit is used to receive feedback information from the maintenance team; and adjust the parameters of the neural network prediction model according to the feedback information to improve the prediction accuracy.

[0011] Furthermore, the sensing unit includes:

[0012] Tilt angle sensor, used to measure the tilt angle of the pole;

[0013] Accelerometers, used to measure the vibration of the pole;

[0014] Temperature and humidity sensor, used to measure the temperature and humidity on the surface of the pole;

[0015] Weather monitoring module, including wind speed sensor, rain gauge and barometer, used to measure environmental factor data;

[0016] The data acquisition controller is used to coordinate the work of the sensors and perform preliminary processing and local storage on the collected data.

[0017] Furthermore, the sensing unit further includes:

[0018] Data preprocessing module, used to filter, denoise and standardize the collected raw data;

[0019] A data compression encoder is used to compress and encode the pre-processed data to reduce the amount of data transmission;

[0020] a wireless communication module, configured to periodically transmit the compressed data to the prediction unit;

[0021] An energy management system, including solar panels and batteries, is used to provide a continuous power supply to the sensing unit.

[0022] Furthermore, the decision-making unit includes:

[0023] The risk assessment module is used to calculate the comprehensive risk index of the pole based on the remaining life and potential failure risk output by the prediction unit;

[0024] Maintenance strategy library, which stores predefined maintenance strategies and corresponding triggering conditions;

[0025] A strategy matcher is used to select an appropriate maintenance strategy from the maintenance strategy library based on the calculated risk index;

[0026] Maintenance plan generator for converting selected maintenance strategies into specific maintenance recommendations and optimized maintenance plans.

[0027] Furthermore, the adaptive unit includes:

[0028] Feedback receiving interface, used to receive feedback information submitted by the maintenance team;

[0029] A feedback classifier is used to classify the received feedback information into prediction accuracy feedback, maintenance effect feedback and system improvement suggestions;

[0030] A parameter adjustment calculator, used to calculate the adjustment amount of the neural network prediction model parameters based on the prediction accuracy feedback;

[0031] A model update executor is used to apply the calculated parameter adjustments to the neural network prediction model.

[0032] Furthermore, the neural network prediction model includes:

[0033] The time series encoder is implemented using a bidirectional long short-term memory network structure. It is used to receive and process time series data from the sensing unit, including the physical state data of the pole and environmental factors, and obtain high-dimensional feature representations at each time point.

[0034] A multi-scale attention module, comprising a daily self-attention submodule, a weekly self-attention submodule, a monthly self-attention submodule, and a fusion submodule; wherein the daily self-attention submodule is used to receive and process the high-dimensional feature representation provided by the time series encoder to obtain daily-level features reflecting short-term changes and daily fluctuations; the weekly self-attention submodule is used to receive and process the daily-level features, and in combination with the high-dimensional feature representation provided by the time series encoder, obtain weekly-level features reflecting medium-term data changes; the monthly self-attention submodule is used to receive and process the weekly-level features, and in combination with the high-dimensional feature representation provided by the time series encoder, obtain monthly-level features reflecting long-term data changes; the fusion submodule is used to fuse the daily-level features, weekly-level features, and monthly-level features to obtain comprehensive features;

[0035] The conditional autoregressive generator, implemented using a variational autoencoder, receives and processes the comprehensive features provided by the multi-scale attention module to generate a probability distribution of the future state of the pole;

[0036] The risk assessment decoder, implemented using a multi-layer perceptron, is used to receive and process the probability distribution output by the conditional autoregressive generator to obtain the remaining life prediction and potential failure risk of the pole.

[0037] Furthermore, the time series encoder includes a dynamically adjusted gating mechanism; the gating mechanism uses an adaptive parameter adjustment strategy to adjust the weights of the forget gate and the input gate in the bidirectional long short-term memory network in real time according to the statistical characteristics of the received time series data;

[0038] The input gate is calculated by the following formula 1:

[0039] i t =σ(W i ·[h t-1 ,x t ]+U i ·s t +b i ) (1)

[0040] Among them, i t represents the activation value of the input gate; σ is the sigmoid activation function; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; U i is the adaptive adjustment matrix of the input gate; s t is a statistical feature vector calculated based on the frequency and amplitude of current data changes; b i is the bias term;

[0041] The forget gate is calculated by the following formula 2:

[0042] f t =σ(W f ·[h t-1 ,x t ]+U f ·s t +b f ) (2)

[0043] Among them, f t represents the activation value of the forget gate; σ is the sigmoid activation function; W f is the weight matrix of the forget gate; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; U f is the adaptive adjustment matrix of the forget gate; s t is a statistical feature vector calculated based on the frequency and amplitude of current data changes; b f is the bias term.

[0044] Furthermore, the fusion submodule of the multi-scale attention module is implemented using an adaptive weight adjustment algorithm driven by environmental factors, and the adaptive weight adjustment algorithm includes the following steps:

[0045] According to the following formula 3, the fusion weight ω at the daily level is calculated d , the fusion weight of the weekly level features ω w And the fusion weight ω of the monthly level features m :

[0046]

[0047] Among them, d, w, m represent the index of the fusion weight of the daily, weekly and monthly features respectively; i is one of the indexes d, w, m, ω i represents the i-th fusion weight in the fusion weights of daily, weekly and monthly features; γ is the adjustment factor; RMSE i is the root mean square error of the i-th time level on the historical data; RMSE j is the root mean square error of the jth time level on the historical data;

[0048] According to the following formula 4, the fusion weight ω i Make adjustments:

[0049]

[0050] Where α is the learning rate; ΔE i It is a quantitative indicator of the impact of environmental factors on the characteristics of the i-th time level; E total It is the total effect of all environmental factors; is the updated fusion weight ω i ; is the fusion weight ω before update i ; Quantitative indicator ΔE i Calculated by the following formula 5:

[0051]

[0052] in, Indicates the rate of environmental change; Corr(E,PE i ) is the environmental factor E and the prediction error PE of the i-th level i The correlation coefficient between the total impact of all environmental factors E total Calculated by the following formula 6:

[0053]

[0054] Among them, w k is the weight of the kth environmental factor; E env,k It is the sensitivity score of the kth environmental factor, which reflects the impact of the kth environmental factor on the operating status of the pole.

[0055] Furthermore, the variational autoencoder structure adopted by the conditional autoregressive generator includes:

[0056] The encoding unit is used to receive comprehensive features from the multi-scale attention module and enhance the representation ability of the features through nonlinear transformation to adapt to the specific needs of pole maintenance prediction;

[0057] The mapping and sampling unit is used to optimize the coverage and accuracy of the probability distribution through an adaptive sampling mechanism to ensure that the probability distribution can effectively reflect the future state of the pole;

[0058] A decoding unit that converts the representation of the latent space into a detailed probability distribution of the future states of the poles.

[0059] Furthermore, the risk assessment decoder uses a multi-layer perceptron comprising:

[0060] The input layer is used to receive the probability distribution output by the conditional autoregressive generator;

[0061] Multiple hidden layers, where each hidden layer contains several neurons, and the number of neurons between adjacent hidden layers decreases, forming a funnel-shaped structure; the neurons in each hidden layer process the output of the previous layer through a nonlinear activation function;

[0062] The skip connection module is used to pass the information of the input layer directly to the intermediate hidden layer to preserve the key features of the original probability distribution;

[0063] Attention gating unit, set after each hidden layer, is used to dynamically adjust the importance weight of each feature;

[0064] Adaptive dropout layer, which automatically adjusts the dropout ratio based on the uncertainty of the input data to improve the robustness of the model;

[0065] Forked output layer, including:

[0066] The life prediction sublayer is used to output the predicted value of the remaining life of the pole; and

[0067] The risk assessment sublayer is used to output the probability distribution of potential failure risks.

[0068] This application has the following beneficial technical effects:

[0069] (1) The system uses sensing units installed on poles to collect data on the poles’ physical condition and environmental factors, and combines this data with a pre-trained neural network model for data analysis and fault risk prediction. This integrated prediction approach enables a more accurate assessment of the health and remaining life of the poles, thereby optimizing maintenance scheduling and reducing unnecessary maintenance activities and associated costs.

[0070] (2) The system’s decision-making unit can generate specific maintenance recommendations and optimize maintenance plans based on real-time analysis results. This not only helps the maintenance team quickly respond to potential failures, but also enables them to adjust maintenance strategies based on actual conditions, improving resource utilization efficiency and operational safety.

[0071] (3) The system includes an adaptive unit that receives feedback from the maintenance team and adjusts the parameters of the prediction model accordingly. This self-learning mechanism enables the system to continuously improve its performance, optimize prediction accuracy based on accumulated data and experience, and cope with complex and changing environmental conditions.

[0072] (4) By comprehensively analyzing physical status data and environmental factors, the system can proactively identify potential faults and problems and take preventive measures in advance. This preventive maintenance approach helps avoid sudden failures and accidents, ensures the stability and reliability of power supply, and thus improves overall service quality and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a schematic diagram of an intelligent perception and automatic adaptation pole maintenance prediction system provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0074] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0075] The first embodiment of the present application provides an intelligent perception and automatic adaptation power pole maintenance prediction system. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1 The first embodiment of the present application provides a detailed description of an intelligent perception and automatic adaptation pole maintenance prediction system.

[0076] The intelligent perception and automatic adaptation pole maintenance prediction system includes a perception unit 101 , a prediction unit 102 , a decision unit 103 and an adaptation unit 104 .

[0077] The sensing unit 101 is installed on the pole and is configured to collect physical state data and environmental factor data of the pole, wherein the physical state data includes the tilt angle, vibration, temperature and humidity of the pole, and the environmental factor data includes wind speed, rainfall and air pressure.

[0078] In this embodiment, the sensing unit 101 is the core component of the intelligent sensing and adaptive pole maintenance prediction system. Designed to be reliably mounted on a pole, it collects key physical status and environmental data. This unit is comprised of multiple sensors, each dedicated to detecting specific parameters to ensure comprehensive and accurate data.

[0079] Sensing unit 101 includes a tilt angle sensor for measuring the tilt of the pole, a key indicator for assessing its structural stability. A vibration sensor monitors the frequency and intensity of the pole's vibrations. This vibration data helps identify potential structural issues caused by wind, traffic, or other external factors. Temperature and humidity sensors provide temperature and humidity readings of the pole and its surroundings. This information is particularly important for analyzing the impact of climate change on material fatigue in the pole.

[0080] In addition, the sensing unit 101 is equipped with an anemometer for measuring wind speed, a rain gauge for recording rainfall, and a barometer for measuring air pressure. The data collected by these environmental sensors is crucial for assessing how weather conditions affect the stability and safety of the poles.

[0081] The output data from all these sensors undergoes preliminary analysis via an integrated data processing unit, which processes and synchronizes data in real time, ensuring its timeliness and accuracy. The processed data is then sent to the system’s prediction unit 102 for more in-depth analysis and lifespan prediction.

[0082] To ensure high performance and long-term reliability, the Sensing Unit 101 is designed with waterproof and environmentally resistant materials. It also includes a power management system that can draw energy from the pole's own power supply or use solar panels as a backup power option, ensuring continuous operation even without external power.

[0083] Furthermore, the sensing unit includes:

[0084] Tilt angle sensor, used to measure the tilt angle of the pole;

[0085] Accelerometers, used to measure the vibration of the pole;

[0086] Temperature and humidity sensor, used to measure the temperature and humidity on the surface of the pole;

[0087] Weather monitoring module, including wind speed sensor, rain gauge and barometer, used to measure environmental factor data;

[0088] The data acquisition controller is used to coordinate the work of the sensors and perform preliminary processing and local storage on the collected data.

[0089] The various components of the perception unit of the intelligent perception and automatic adaptation pole maintenance prediction system work together to collect key data about the pole and its surrounding environment, which is the basis for the system to make high-precision predictions.

[0090] First, the perception unit includes a tilt angle sensor, a key component for real-time monitoring of the pole's tilt angle. Typically mounted at the top or center of a pole, the tilt angle sensor ensures accurate measurement of the pole's tilt. The sensor's ability to detect subtle changes and provide precise angle data is crucial for early identification of potential structural issues.

[0091] Next, accelerometers are used to monitor the pole's vibration. These sensors are particularly important for capturing dynamic changes in the pole caused by environmental factors such as wind and traffic. Vibration data helps analyze the pole's performance under everyday conditions and is essential for assessing its structural integrity during extreme weather events.

[0092] Temperature and humidity sensors are deployed on the poles to monitor their surface temperature and humidity. This information is crucial for assessing the impact of weather conditions on the pole materials. For example, temperature changes can affect the expansion or contraction of the pole material, while humidity levels can affect the corrosion rate of wood or metal.

[0093] The weather monitoring module provides the system with a comprehensive environmental data measurement solution, including a wind speed sensor, rain gauge, and barometer. The wind speed sensor measures the potential impact of wind on the power poles; the rain gauge helps assess precipitation, which is crucial for determining humidity conditions and moisture impact; and the barometer provides air pressure change data, which helps analyze weather trends.

[0094] Data from all these sensors is managed by a data acquisition controller. This controller not only coordinates the operation of the sensors to ensure real-time data collection, but also performs preliminary data processing and local storage. This design enables the system to quickly and accurately assess the condition of the poles without relying on a constant external data connection.

[0095] Furthermore, the sensing unit further includes:

[0096] Data preprocessing module, used to filter, denoise and standardize the collected raw data;

[0097] A data compression encoder is used to compress and encode the pre-processed data to reduce the amount of data transmission;

[0098] a wireless communication module, configured to periodically transmit the compressed data to the prediction unit;

[0099] An energy management system, including solar panels and batteries, is used to provide a continuous power supply to the sensing unit.

[0100] In the intelligent perception and adaptive pole maintenance prediction system, the perception unit not only includes multiple sensors but also expands its functionality to include data preprocessing, compression, wireless communication, and an energy management system. These components work together to ensure efficient data processing and transmission while maintaining the system's energy independence and long-term operational capabilities.

[0101] First, the data preprocessing module is a crucial component of the system, responsible for processing the raw data collected by each sensor. This module uses advanced algorithms for filtering, denoising, and normalization, steps that are crucial for ensuring data quality. Filtering and denoising help remove environmental interference and potential errors introduced by the device itself, while normalization ensures that data from different sensors is compared and analyzed using the same metrics. This preprocessing not only improves data usability but also optimizes the accuracy of subsequent analysis.

[0102] Next, a data compression encoder compresses the preprocessed data. This step optimizes data storage and transmission efficiency, particularly when transmitting data to a remote server or prediction unit over a wireless network. Data compression reduces data redundancy, reducing required storage space and transmission bandwidth, thereby alleviating network burden and accelerating data transmission.

[0103] The wireless communication module is responsible for regularly transmitting compressed data to the prediction unit. This module includes communication technologies that adapt to various network conditions, such as cellular networks, Wi-Fi, and other wireless technologies, ensuring reliable and secure data transmission. The design of the wireless communication module takes into account the system's actual application environment, selecting the communication technology most suitable for the area where the poles are located to ensure continuous and efficient data transmission.

[0104] Finally, the sensing unit includes an energy management system consisting of solar panels and batteries to provide a continuous power supply. The solar panels collect energy based on ambient sunlight conditions, while the batteries store this energy for use when power is insufficient. This energy management solution enables the sensing unit to operate independently without an external power source, making it particularly suitable for remote locations or those with limited access to the power grid.

[0105] The prediction unit 102 is used to predict the remaining life and potential failure risk of the pole based on the physical state data and environmental factor data collected by the sensing unit using a pre-trained neural network prediction model.

[0106] Prediction Unit 102 plays a crucial role in the intelligent perception and adaptive pole maintenance prediction system. It is responsible for conducting in-depth analysis and predictions based on the data collected by Perception Unit 101. This unit includes a pre-trained neural network prediction model, which is the core of the system and is used to assess the remaining life of the pole and identify possible failure risks.

[0107] Prediction unit 102 first receives various physical state data and environmental factor data collected from the poles by perception unit 101. This data includes the pole's tilt angle, vibration, temperature, humidity, wind speed, rainfall, and air pressure. This data is transmitted to the prediction unit via a specific interface to ensure data integrity and accuracy.

[0108] To effectively process and analyze this data, the neural network model within the prediction unit is pre-trained on a large amount of historical data. This training enables the network to learn the behavior patterns of power poles under different physical and environmental conditions, enabling it to make accurate predictions when new data is input. This neural network typically includes multiple hidden layers, each containing dozens to hundreds of neurons. Each neuron processes information from the previous layer and passes it on to the next layer.

[0109] Within a neural network, data first undergoes preprocessing steps, such as normalization, to ensure that different types of data have a balanced impact on the model. This data is then fed into the first layer of the network, passing through each layer. Each layer uses a specific activation function to increase the model's nonlinear processing capabilities, allowing the model to better capture complex data relationships.

[0110] Once the data is processed in the neural network, the final output layer generates a prediction of the pole’s remaining life and an assessment of its potential failure risk. These predictions are based on the model’s learning of the pole’s historical performance under similar conditions, taking into account real-time data on its current environmental and physical state.

[0111] Finally, the prediction results are passed to the decision-making unit 103, which generates maintenance recommendations and optimization plans that guide the maintenance team in actual operations. Furthermore, the prediction unit receives feedback from the adaptive unit 104, adjusting the neural network parameters based on actual maintenance results and team feedback to continuously optimize the accuracy of the predictions.

[0112] Through such a configuration, the prediction unit 102 ensures the efficient operation of the pole maintenance prediction system, enabling the system to make fast and accurate predictions based on real-time data, greatly improving the maintenance efficiency and safety of the poles.

[0113] Furthermore, the neural network prediction model includes:

[0114] The time series encoder is implemented using a bidirectional long short-term memory network structure. It is used to receive and process time series data from the sensing unit, including the physical state data of the pole and environmental factors, and obtain high-dimensional feature representations at each time point.

[0115] A multi-scale attention module, comprising a daily self-attention submodule, a weekly self-attention submodule, a monthly self-attention submodule, and a fusion submodule; wherein the daily self-attention submodule is used to receive and process the high-dimensional feature representation provided by the time series encoder to obtain daily-level features reflecting short-term changes and daily fluctuations; the weekly self-attention submodule is used to receive and process the daily-level features, and in combination with the high-dimensional feature representation provided by the time series encoder, obtain weekly-level features reflecting medium-term data changes; the monthly self-attention submodule is used to receive and process the weekly-level features, and in combination with the high-dimensional feature representation provided by the time series encoder, obtain monthly-level features reflecting long-term data changes; the fusion submodule is used to fuse the daily-level features, weekly-level features, and monthly-level features to obtain comprehensive features;

[0116] The conditional autoregressive generator, implemented using a variational autoencoder, receives and processes the comprehensive features provided by the multi-scale attention module to generate a probability distribution of the future state of the pole;

[0117] The risk assessment decoder, implemented using a multi-layer perceptron, is used to receive and process the probability distribution output by the conditional autoregressive generator to obtain the remaining life prediction and potential failure risk of the pole.

[0118] The neural network prediction model is a complex and sophisticated component of the intelligent perception and adaptive pole maintenance prediction system, specifically designed to process and analyze large amounts of data from the perception unit, ultimately generating detailed predictions about the pole's condition. This model incorporates multiple advanced machine learning techniques to ensure accurate and reliable predictions.

[0119] First, the time series encoder uses a bidirectional long short-term memory (Bi-LSTM) architecture, which is well-suited for processing time series data because it can capture patterns in the data that change over time. The time series encoder's primary task is to receive data on the pole's physical state (such as tilt angle and vibration) and environmental factors (such as temperature, humidity, and wind speed) from the sensing unit. Through this encoder, the data is converted into a high-dimensional feature representation, providing a rich layer of information for further analysis.

[0120] The output of the time series encoder is then further processed by the multi-scale attention module. This module comprises several self-attention sub-modules at different levels: daily, weekly, and monthly, each sensitive to data changes at a different timescale. The daily sub-module captures short-term data fluctuations, the weekly sub-module processes data changes over longer timeframes, and the monthly sub-module analyzes long-term trends. These features are then processed by the fusion sub-module to form a comprehensive feature set that encompasses a comprehensive view of the data, from immediate changes to long-term trends.

[0121] Furthermore, a conditional autoregressive generator uses variational autoencoder (VAE) technology to process the output of the multi-scale attention module. The core function of this generator is to generate a probability distribution of the future state of the pole based on the processed comprehensive features. This probability distribution model can provide predictions about the remaining lifespan of the pole and the potential risk of failure.

[0122] Finally, the risk assessment decoder uses a multi-layer perceptron (MLP) structure to process the probability distribution output by the conditional autoregressive generator. This decoder converts the probability distribution into specific prediction results, such as the specific remaining life of the pole and the failure risk level.

[0123] The design of the entire neural network prediction model ensures that every step from data collection to risk assessment is accurate and can provide scientific decision-making support for the maintenance team.

[0124] The implementation code of the neural network prediction model is provided below:

[0125]

[0126]

[0127]

[0128]

[0129]

[0130] Training a neural network predictive model requires following a detailed series of steps. This process involves data preparation, model building, training process, and performance evaluation.

[0131] First, the data required to train the model needs to be collected and prepared. This includes time series data from the sensing unit, such as the physical state of the pole (tilt angle, vibration, etc.) and environmental factors (temperature, humidity, wind speed, etc.). The data should be formatted in a form suitable for input to the neural network, which usually means that each data point needs to have a timestamp and all sensor readings should be synchronized.

[0132] Model construction:

[0133] 1. Time Series Encoder**:

[0134] Implement a bidirectional long short-term memory (Bi-LSTM) network, which will allow the model to capture the previous and next dependencies in time series data. The output of the Bi-LSTM layer will be a high-dimensional feature representation of each time point.

[0135] 2. Multi-scale Attention Module:

[0136] Three self-attention submodules are designed, corresponding to the daily, weekly, and monthly levels. Each module processes the output from the Bi-LSTM to extract and enhance the important features of its corresponding time scale.

[0137] These features are then merged through a fusion submodule to form a comprehensive feature set that contains information from all time scales.

[0138] 3. Conditional Autoregressive Generator:

[0139] The fused features are processed using a variational autoencoder (VAE) to generate a probability distribution over the pole’s future states. The VAE should be designed to receive the combined features and output distribution parameters that represent the likelihood of the pole’s state.

[0140] 4. Risk Assessment Decoder:

[0141] A multi-layer perceptron (MLP) is used to parse the output of the conditional autoregressive generator to generate the final prediction results, such as the remaining life of the pole and the potential failure risk.

[0142] Training process:

[0143] Use an appropriate loss function to train the model, such as mean squared error (MSE) or cross entropy, depending on the specific nature of the prediction task. During training, split the dataset into a training set and a validation set. Use the training set to train the model, and the validation set to adjust the model parameters and avoid overfitting.

[0144] Set the number of iterations and use an optimization algorithm (such as Adam or SGD) to optimize the parameters. At the end of each training cycle (epoch), evaluate the performance of the model on the validation set and monitor indicators such as loss and accuracy.

[0145] Performance evaluation and adjustment:

[0146] After model training is complete, the test set is used to evaluate the model's performance, focusing on the model's accuracy and reliability in predicting the remaining life of the pole and the risk of failure.

[0147] Adjust the model structure or training parameters based on the test results to optimize performance. This may require returning to the model building phase to make adjustments, such as adding or removing network layers, adjusting layer sizes, or modifying the implementation of the attention mechanism.

[0148] By implementing the above steps, we can build, train, and deploy an efficient neural network model that can accurately predict maintenance needs for utility poles and effectively support maintenance decision-making. Such a system can significantly improve the efficiency and safety of utility pole maintenance, bringing far-reaching impacts to related fields.

[0149] Furthermore, the time series encoder includes a dynamically adjusted gating mechanism; the gating mechanism uses an adaptive parameter adjustment strategy to adjust the weights of the forget gate and the input gate in the bidirectional long short-term memory network in real time according to the statistical characteristics of the received time series data;

[0150] The input gate is calculated by the following formula 1:

[0151] i t =σ(W i ·[h t-1 ,x t ]+U i ·s t +b i ) (1)

[0152] Among them, i t represents the activation value of the input gate; σ is the sigmoid activation function; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; U i is the adaptive adjustment matrix of the input gate; s t is a statistical feature vector calculated based on the frequency and amplitude of current data changes; b i is the bias term;

[0153] The forget gate is calculated by the following formula 2:

[0154] f t =σ(W f ·[h t-1 ,x t ]+U f ·s t +b f ) (2)

[0155] Among them, f t represents the activation value of the forget gate; σ is the sigmoid activation function; W f is the weight matrix of the forget gate; ht-1 is the hidden state of the previous time step; x t is the input of the current time step; U f is the adaptive adjustment matrix of the forget gate; s t is a statistical feature vector calculated based on the frequency and amplitude of current data changes; b f is the bias term.

[0156] In the intelligent perception and adaptive pole maintenance prediction system, the time series encoder employs a dynamically adjusted gating mechanism, which increases the model's adaptability to changes in time series data. Specifically, this mechanism uses an adaptive parameter adjustment strategy to adjust the weights of the forget gate and input gate in the bidirectional long short-term memory (Bi-LSTM) network in real time, enabling the network to more accurately respond to dynamic changes in the data.

[0157] Gating mechanism of time series encoder:

[0158] Calculation of input gate (Formula 1):

[0159] The function of the input gate is to control the current input x t The degree of influence on the current cell state. The input gate is calculated as follows:

[0160] i t =σ(W i ·[h t-1 ,x t ]+U i ·s t +b i ) (1)

[0161] Among them, i t Represents the activation value of the input gate.

[0162] σ is the sigmoid activation function, which is used to compress any input between 0 and 1, which makes it ideal for controlling the flow of information.

[0163] W i is the weight matrix of the input gate, which directly acts on the hidden state h of the previous time step. t-1 and the input x at the current time step t .

[0164] [h t-1 ,x t ] is the connection between the hidden state of the previous time step and the input of the current time step.

[0165] U i It is the adaptive adjustment matrix of the input gate, which is used to adjust the behavior of the input gate according to the statistical characteristics of the data.

[0166] s t It is a statistical feature vector calculated from the frequency and amplitude of changes in the current data, which provides a quantitative representation of the real-time data characteristics.

[0167] b i It is a bias term used in neural networks to increase the flexibility and nonlinearity of the model.

[0168] The function of the forget gate is to control how much past information the cell should discard from its internal state. The forget gate is calculated as follows:

[0169] f t =σ(W f ·[h t-1 ,x t ]+U f ·s t +b f ) (2)

[0170] in:

[0171] f t Represents the activation value of the forget gate.

[0172] σ is the sigmoid activation function.

[0173] W f is the weight matrix of the forget gate, which acts on the hidden state of the previous time step and the input of the current time step.

[0174] [h t-1 ,x t ] is the connection between the hidden state of the previous time step and the input of the current time step.

[0175] U f It is the adaptive adjustment matrix of the forget gate, which is used to adjust the behavior of the forget gate based on the statistical characteristics of the current data.

[0176] s t It is a statistical feature vector calculated from the frequency and amplitude of changes in the current data.

[0177] b f is the bias term.

[0178] Furthermore, the fusion submodule of the multi-scale attention module is implemented using an adaptive weight adjustment algorithm driven by environmental factors, and the adaptive weight adjustment algorithm includes the following steps:

[0179] According to the following formula 3, the fusion weight ω at the daily level is calculated d , the fusion weight of the weekly level features ω w And the fusion weight ω of the monthly level features m :

[0180]

[0181] Among them, d, w, m represent the index of the fusion weight of the daily, weekly and monthly features respectively; i is one of the indexes d, w, m, ω i represents the i-th fusion weight in the fusion weights of daily, weekly and monthly features; γ is the adjustment factor; RMSE i is the root mean square error of the i-th time level on the historical data; RMSE j is the root mean square error of the jth time level on the historical data;

[0182] According to the following formula 4, the fusion weight ω i Make adjustments:

[0183]

[0184] Where α is the learning rate; ΔE i It is a quantitative indicator of the impact of environmental factors on the characteristics of the i-th time level; E total It is the total effect of all environmental factors; is the updated fusion weight ω i ; is the fusion weight ω before update i ; Quantitative indicator ΔE i Calculated by the following formula 5:

[0185]

[0186] in, Indicates the rate of change of environmental factor E; Corr(E,PE i ) is the environmental factor E and the prediction error PE of the i-th level i The correlation coefficient between the two variables is shown in Figure 2. Variable E represents the total environmental factor, which combines all the environmental variables considered to represent the overall state of the environment in which the pole is located. This total environmental factor E can include multiple specific environmental parameters such as temperature, humidity, wind speed, and rainfall. These parameters together describe the overall environmental conditions of the pole.

[0187] The total impact of all environmental factors E total Calculated by the following formula 6:

[0188]

[0189] Among them, w k is the weight of the kth environmental factor; E env,kIt is the sensitivity score of the kth environmental factor, which reflects the impact of the kth environmental factor on the operating status of the pole.

[0190] In the intelligent perception and adaptive pole maintenance prediction system, the fusion submodule of the multi-scale attention module uses an adaptive weight adjustment algorithm driven by environmental factors. This algorithm dynamically adjusts weights based on the performance of features at different time scales and changes in environmental factors. This design aims to optimize the model's predictive capabilities, making it more sensitive and accurate in reflecting the impact of real-time environmental changes.

[0191] First, the algorithm calculates the fusion weight (ω) of each time level feature according to Formula 3 d ,ω w ,ω m ), these weights are based on the performance of each time level and are obtained by the following formula:

[0192]

[0193] in:

[0194] ω i Represents the fusion weight of the i-th time level.

[0195] γ is a regulation factor used to control the impact of the error on the strength of weight adjustment.

[0196] RMSE i is the root mean square error of the i-th time level on the historical data, which is used to evaluate the accuracy of the forecast at that level. The smaller the error, the greater its corresponding weight.

[0197] RMSE j is the root mean square error at the jth time level.

[0198] d, w, and m represent the daily, weekly, and monthly levels, respectively.

[0199] The weight adjustment is implemented through Formula 4, taking into account the real-time changes in environmental factors:

[0200]

[0201] in:

[0202] α is the learning rate, which controls the speed of adjustment to the effects of environmental changes.

[0203] ΔE i It is a quantitative indicator of the impact of changes in environmental factors on the characteristics of the i-th time level.

[0204] E total It is the total impact of all environmental factors.

[0205] Quantitative indicator ΔEi It is calculated by formula 5:

[0206]

[0207] Indicates the rate at which the environment changes.

[0208] Corr(E,PE i ) is the environmental factor E and the prediction error PE of the i-th level i The correlation coefficient between them measures the association between environmental changes and prediction errors.

[0209] Finally, Equation 6 is used to calculate the total impact of all environmental factors:

[0210]

[0211] w k is the weight of the kth environmental factor.

[0212] E env,k is the sensitivity score of the kth environmental factor, which reflects the impact of this factor on the operating status of the pole. The sensitivity score can be provided by expert knowledge.

[0213] This adaptive weight adjustment algorithm enables the prediction model to automatically adjust its behavior based on real-time environmental changes and historical prediction performance, thereby improving prediction accuracy and system responsiveness.

[0214] Furthermore, the variational autoencoder structure adopted by the conditional autoregressive generator includes:

[0215] The encoding unit is used to receive comprehensive features from the multi-scale attention module and enhance the representation ability of the features through nonlinear transformation to adapt to the specific needs of pole maintenance prediction;

[0216] The mapping and sampling unit is used to optimize the coverage and accuracy of the probability distribution through an adaptive sampling mechanism to ensure that the probability distribution can effectively reflect the future state of the pole;

[0217] A decoding unit that converts the representation of the latent space into a detailed probability distribution of the future states of the poles.

[0218] In the intelligent perception and adaptive pole maintenance prediction system, the conditional autoregressive generator uses a variational autoencoder (VAE) structure to more accurately predict the future state of the pole and provide output in the form of a probability distribution. Variational autoencoders are particularly well-suited for this type of problem because they not only generate data but also optimize this generation process by learning the latent representation of the input data.

[0219] The encoding unit's primary task is to receive comprehensive features from the multi-scale attention module. These features encompass information about the pole's physical state and environmental factors at different time scales. The encoding unit enhances the representational power of these features through a series of nonlinear transformations, typically comprising several layers of neural networks. This processing not only extracts the most important information from the data but also adapts it to the specific needs of pole maintenance prediction. The nonlinear layers can be either fully connected or convolutional, depending on the nature of the input features and the desired output.

[0220] The mapping and sampling unit is the core of the variational autoencoder architecture, responsible for defining and optimizing the representation of the probability distribution. In this unit, features from the encoding unit are first mapped to the parameters of a latent space, typically consisting of two statistics: the mean (μ) and the variance (σ). These parameters describe the underlying distribution of the pole states and provide the basis for generating the probability distribution.

[0221] Furthermore, an adaptive sampling mechanism, known as the reparameterization technique, is employed to draw samples from the distribution defined by these parameters. This mechanism not only optimizes the coverage of the probability distribution but also improves its precision, ensuring that the generated probability distribution effectively reflects the future states of the poles. This allows the model to backpropagate gradients during training, resulting in a more efficient learning of the distribution parameters.

[0222] The decoding unit converts the latent space representation output by the mapping and sampling unit back into the original data space. This means converting the latent variables into a probability distribution over the specific future states of the pole. The decoding process typically involves a similar but inverse neural network structure to the encoding unit. By gradually increasing the data dimensionality and complexity, the decoding unit ultimately outputs detailed information such as the remaining life of the pole and potential failure risk.

[0223] To implement such a variational autoencoder structure, it is necessary to design and adjust the network architecture according to the specific requirements of the prediction task, select an appropriate nonlinear activation function, optimize the depth and width of the network, and adjust hyperparameters during training, such as the learning rate and batch size. In addition, it is necessary to ensure that the data is properly preprocessed and normalized to suit the network input requirements.

[0224] In this way, variational autoencoders not only provide a powerful framework to predict and analyze the future states of poles, but also optimize the process of learning these predictions from real data.

[0225] Furthermore, the risk assessment decoder uses a multi-layer perceptron comprising:

[0226] The input layer is used to receive the probability distribution output by the conditional autoregressive generator;

[0227] Multiple hidden layers, where each hidden layer contains several neurons, and the number of neurons between adjacent hidden layers decreases, forming a funnel-shaped structure; the neurons in each hidden layer process the output of the previous layer through a nonlinear activation function;

[0228] The skip connection module is used to pass the information of the input layer directly to the intermediate hidden layer to preserve the key features of the original probability distribution;

[0229] Attention gating unit, set after each hidden layer, is used to dynamically adjust the importance weight of each feature;

[0230] Adaptive dropout layer, which automatically adjusts the dropout ratio based on the uncertainty of the input data to improve the robustness of the model;

[0231] Forked output layer, including:

[0232] The life prediction sublayer is used to output the predicted value of the remaining life of the pole; and

[0233] The risk assessment sublayer is used to output the probability distribution of potential failure risks.

[0234] In the intelligent perception and automatic adaptation pole maintenance prediction system, the risk assessment decoder adopts a multi-layer perceptron (MLP) structure to accurately parse and transform the probability distribution output by the conditional autoregressive generator, ultimately generating detailed information about the remaining life prediction and potential failure risks of the pole.

[0235] The input layer of the risk assessment decoder receives the output from the conditional autoregressive generator. These outputs are typically representations of the probability distribution of the pole's future states, such as encoded vectors in latent space. The input layer ensures that these probability distributions can be correctly processed by subsequent network layers.

[0236] The decoder consists of several hidden layers, where the number of neurons decreases from the input layer to the output layer, forming a funnel-shaped structure. This design helps to gradually refine and strengthen the features extracted from the input data. Each hidden layer uses a nonlinear activation function (such as ReLU or tanh) to process the output of the previous layer. These nonlinear transformations are key to improving the network's ability to handle complex problems.

[0237] In order to preserve the key features of the original probability distribution in the input layer and avoid the feature attenuation problem that occurs in multi-layer networks, a skip connection module is provided. This module allows part of the information in the input layer to be directly passed to the intermediate hidden layer, which helps the network maintain the key information of the input data at a deep level and helps the gradient flow more efficiently during training.

[0238] Attention gating units are placed after each hidden layer. The purpose of these units is to dynamically adjust the importance weights of each feature passed to the next layer. In this way, the model can focus more on features that have a greater impact on the predicted output, thereby improving prediction accuracy and efficiency.

[0239] To enhance the model's adaptability to input data uncertainty and improve its robustness, adaptive dropout layers are introduced. These layers automatically adjust the dropout ratio based on the characteristics of the input data and the current state of the model, thereby reducing overfitting while maintaining model performance.

[0240] The final output layer is bifurcated and consists of two sub-layers:

[0241] Life prediction sublayer: It is specifically used to output the remaining life prediction value of the pole. This is achieved by analyzing the latent features and converting them into specific life values.

[0242] Risk Assessment Sublayer: This sublayer outputs a probability distribution of potential failure risks. This sublayer processes the features extracted by the network and converts them into probabilistic information describing the possible failure of the pole.

[0243] Implementing this risk assessment decoder requires careful design of the dimensions and parameters of each layer, ensuring that the choice of nonlinear activation function matches the expected data processing requirements. Furthermore, during training, network parameters such as the learning rate, weight initialization, and the number of training cycles must be carefully tuned to optimize model performance and ensure accurate and reliable output.

[0244] The decision unit 103 is configured to generate maintenance suggestions and an optimized maintenance plan for the poles according to the prediction results of the prediction unit; and provide the generated maintenance suggestions and optimized maintenance plan to the maintenance team.

[0245] Decision-making unit 103 is a key component of the intelligent perception and adaptive pole maintenance prediction system. It is specifically designed to process the pole condition prediction results received from prediction unit 102 and generate practical maintenance recommendations and optimized maintenance plans based on them. This unit's design ensures that it can translate complex data analysis into specific, actionable steps that directly support effective pole maintenance and management.

[0246] After receiving the remaining life estimate and potential failure risk information from the prediction unit, the decision-making unit begins its core operations. This unit incorporates an advanced decision support system, comprised of algorithms and logic models, to interpret the prediction data and determine the most appropriate maintenance strategy. These algorithms formulate recommendations based on a variety of factors, including the pole's urgent maintenance needs, scheduled maintenance intervals, cost-benefit analysis, and possible risk mitigation strategies.

[0247] The decision-making unit's functionality relies on a set of predefined rules and dynamic learning algorithms that optimize maintenance recommendations based on historical maintenance data, seasonal variations, and the specific environmental conditions of the pole's location. For example, if a pole exhibits high-risk tilt and vibration patterns, the decision-making unit will recommend immediate inspection and possible reinforcement. It also considers the area's current climate conditions, such as strong winds or rainfall, which could exacerbate the pole's risk.

[0248] In addition to generating maintenance recommendations, the decision-making unit also includes an interface that directly transmits these suggestions and solutions to the ground maintenance team. This real-time information transmission ensures that the maintenance team can quickly respond to system-generated recommendations and perform necessary maintenance operations, thereby minimizing potential service interruptions and maintenance costs.

[0249] The decision-making unit also receives feedback from the maintenance team. This feedback is used to further adjust and optimize the algorithms within the decision-support system, ensuring continuous improvement and adaptability in the decision-making process. Through this self-optimization mechanism, the decision-making unit continuously improves the accuracy of its decisions and more effectively supports pole maintenance.

[0250] The following uses a specific example to illustrate how the decision unit 103 generates maintenance recommendations and optimized maintenance plans for electric poles based on the results of the prediction unit 102 and ensures that these recommendations can be implemented.

[0251] Assume that in a specific monitoring area, prediction unit 102 analyzes the data collected by sensing unit 101, including the pole's tilt angle, vibration, temperature, and humidity, and concludes that a pole has a high risk of tilting. Furthermore, environmental factor data indicates that the area has recently experienced continuous heavy rain, making the soil around the pole slippery and increasing the risk of toppling.

[0252] 1. Analysis and decision-making process:

[0253] Prediction unit 102 transmits the data analysis results to decision unit 103. The algorithm within the decision unit immediately assesses the need for urgent maintenance on the pole. Considering the pole's tilt risk and the harsh environmental conditions, the decision unit determines that immediate action is necessary to prevent the pole from collapsing, which could lead to greater safety hazards and power outages.

[0254] 2. Maintenance suggestion generation:

[0255] The maintenance recommendations generated by the decision-making unit include an on-site inspection of the pole to assess whether the tilt exceeds a safe threshold. The recommendations also include reinforcing the pole base and, if necessary, installing support equipment to prevent further tilt.

[0256] 3. Optimize maintenance plan:

[0257] The optimized plan, taking into account the upcoming rainfall, recommends deploying additional drainage measures to reduce the impact of groundwater on the pole's stability after the inspection and reinforcement work is completed. This plan also recommends conducting similar risk assessments on other nearby poles to identify other structures that may be at similar risk.

[0258] 4. Plan communication and implementation:

[0259] These maintenance recommendations and optimization plans are directly transmitted to the ground maintenance team through an interface designed by the decision-making unit. After receiving this information, the maintenance team can quickly dispatch and implement on-site operations according to the recommendations to ensure the safety of the poles and the stability of the power supply.

[0260] Furthermore, the decision-making unit includes:

[0261] The risk assessment module is used to calculate the comprehensive risk index of the pole based on the remaining life and potential failure risk output by the prediction unit;

[0262] Maintenance strategy library, which stores predefined maintenance strategies and corresponding triggering conditions;

[0263] A strategy matcher is used to select an appropriate maintenance strategy from the maintenance strategy library based on the calculated risk index;

[0264] Maintenance plan generator for converting selected maintenance strategies into specific maintenance recommendations and optimized maintenance plans.

[0265] In the intelligent perception and automatic adaptation pole maintenance prediction system, the decision-making unit plays a core role, responsible for converting the data provided by the prediction unit into a practical and feasible maintenance action plan.

[0266] First, the Risk Assessment module within the Decision-Making Unit is the core of the analysis. It uses data from the Prediction Unit—specifically, information about the remaining lifespan of the poles and potential failure risks—to calculate a comprehensive risk index for each pole. This module assesses and quantifies risk using a series of algorithms that consider different risk factors, such as physical damage, environmental impact, and age. These factors are combined to produce a risk index that reflects the level of urgent maintenance needed for each pole.

[0267] Next, the maintenance strategy library is a database of predefined maintenance strategies and their corresponding triggering conditions. The strategies in this library cover a wide range of situations, from routine inspections to emergency repairs. Each strategy has clear conditions that indicate when it should be triggered. These strategies are developed and optimized based on historical data, industry standards, and previous maintenance experience.

[0268] The policy matcher bridges risk assessment and actual operations. Based on the risk index calculated by the risk assessment module, it automatically selects the most appropriate policy from the maintenance policy library. This selection process, based on the risk index and the trigger conditions defined in the policy library, ensures that each selected policy is the optimal response for a specific situation.

[0269] Finally, the Maintenance Plan Generator is responsible for translating the selected maintenance strategy into specific action steps and plans. It refines the strategy into detailed maintenance recommendations and optimized maintenance plans, including required resources, estimated timelines, and expected outcomes. These recommendations and plans are formatted as implementation guidelines that can be directly communicated to the maintenance team, ensuring that maintenance activities are carried out efficiently and in an orderly manner.

[0270] Through this coherent process, the decision-making unit can not only ensure that the risks of the poles are effectively managed, but also optimize the use of maintenance resources, reduce unnecessary expenditures, and improve the stability and security of the entire power system.

[0271] The adaptive unit 104 is configured to receive feedback from the maintenance team and adjust the parameters of the neural network prediction model according to the feedback to improve prediction accuracy.

[0272] Adaptive unit 104 plays a crucial role in the intelligent perception and adaptive pole maintenance prediction system. It is primarily responsible for implementing the system's self-learning and optimization capabilities. By receiving real-time feedback from the maintenance team, the adaptive unit can dynamically adjust the parameters of the neural network prediction model in prediction unit 102, thereby continuously improving prediction accuracy and the system's overall responsiveness.

[0273] In practice, adaptive unit 104 includes a specially designed receiving mechanism for collecting and organizing feedback data provided by the maintenance team. This data may include the results of actual maintenance activities, additional issues observed during maintenance, and evaluations of the effectiveness of current maintenance recommendations. The received information first undergoes preprocessing, including data cleaning and formatting, to ensure the quality and consistency of the input data.

[0274] The processed feedback data is then used to adjust the predictive model. This process is accomplished through a built-in learning algorithm, which evaluates the accuracy of the predictions based on the feedback and optimizes the weights and parameters in the neural network accordingly. For example, if the feedback indicates that certain predicted fault types do not match the actual situation, the learning algorithm will adjust the relevant parameters to reduce the occurrence of similar errors in the future.

[0275] To achieve this function, the adaptive unit 104 is equipped with an advanced machine learning framework that supports various optimization techniques such as gradient descent, back propagation, etc. These technologies help the unit to update and adjust the model in real time based on the ever-changing environment and accumulated maintenance experience.

[0276] The adaptive unit also includes an evaluation module that regularly assesses the improved model to ensure that each adjustment improves the system's predictive performance. This includes monitoring and testing key performance indicators such as model stability, generalization ability, and response speed.

[0277] Through this integrated feedback processing and model optimization mechanism, adaptive unit 104 not only enhances the system's adaptability and flexibility but also ensures continuous improvement over time, enabling more accurate predictions of pole maintenance needs and potential risks. This makes pole maintenance more efficient while also reducing the potential for wasted resources or unexpected failures caused by inaccurate predictions.

[0278] For example, after a particular maintenance inspection, the maintenance team discovered that although the predictive model had predicted a high risk of tilting a pole, the inspection actually showed the pole was much more stable than predicted. This type of feedback is crucial because it indicates that the predictive model may be overly sensitive in assessing certain types of risks or that certain parameters may be improperly configured.

[0279] In this case, the maintenance team will send this finding as feedback to the adaptive unit 104 through an interface designed for this purpose. First, the received feedback data includes the specific physical inspection data of the pole and the observation results of the maintenance team. These data will be formatted and cleaned through a data preprocessing process to ensure that their quality meets the analysis requirements.

[0280] Next, the adaptive unit 104 uses this cleaned data to adjust the neural network model in the prediction unit 102. Specifically, this may involve adjusting the weights and bias parameters of the neural network that handle skewed data inputs to reduce the model's oversensitivity to skewed data. For example, if the original model predicted high risk even with relatively small skews, the learning algorithm might adjust the activation function parameters of the relevant neurons to make them less responsive to small skews.

[0281] This parameter adjustment is achieved through built-in learning algorithms, such as backpropagation, which automatically optimizes model parameters based on the deviation between the predicted results and the actual detection results. In this process, gradient descent may be used to gradually reduce the error and fine-tune the model to better reflect the actual situation.

[0282] Furthermore, the adaptive unit regularly evaluates the performance of the adjusted model to ensure that the improvements have been effective and that the model's predictive accuracy has actually improved. This includes assessing the model's stability, accuracy, and performance on new data to monitor its overall performance and adaptability.

[0283] In this way, adaptive unit 104 not only responds to direct feedback from the maintenance team but also continuously self-optimizes to ensure that the prediction model evolves over time, more accurately matching actual maintenance needs. This dynamic adjustment mechanism ensures continuous improvement and adaptability of the entire system, making pole maintenance more accurate and efficient, significantly reducing the waste of resources and potential risks caused by prediction errors.

[0284] Furthermore, the adaptive unit includes:

[0285] Feedback receiving interface, used to receive feedback information submitted by the maintenance team;

[0286] A feedback classifier is used to classify the received feedback information into prediction accuracy feedback, maintenance effect feedback and system improvement suggestions;

[0287] A parameter adjustment calculator, used to calculate the adjustment amount of the neural network prediction model parameters based on the prediction accuracy feedback;

[0288] A model update executor is used to apply the calculated parameter adjustments to the neural network prediction model.

[0289] In the intelligent perception and adaptive pole maintenance prediction system, the adaptive unit plays a crucial role, ensuring the system continuously optimizes and adapts to the actual operating environment. This unit comprises four main components, which work together to receive and analyze feedback from the maintenance team and subsequently adjust and update the neural network prediction model. The following details the functions and implementation of these components.

[0290] First, the adaptive unit features a feedback interface, serving as a bridge between the system and the maintenance team. This interface is specifically designed to receive a variety of feedback from the maintenance team, including evaluations of the accuracy of predictions, feedback on the effectiveness of completed maintenance tasks, and suggestions for improving system operation and performance. This feedback is typically submitted digitally, such as through a dedicated online platform or mobile application, ensuring that information is delivered directly and quickly to the adaptive unit.

[0291] Next, the received feedback is processed by the feedback classifier. This component sorts and files the complex feedback information into categories, primarily categorizing it into prediction accuracy feedback, maintenance effectiveness feedback, and system improvement suggestions. This classification process relies on pre-set keywords and data parsing technology to ensure that each type of feedback is correctly identified and processed, providing accurate input for subsequent data analysis and model adjustments.

[0292] The parameter adjustment calculator, a core technical component of the adaptive unit, calculates the required model parameter adjustments based on the classification's prediction accuracy feedback. This process involves complex mathematical calculations, including but not limited to error analysis and gradient descent algorithms, to determine which parameters require adjustment and by how much. The calculator incorporates advanced algorithms that identify the specific factors contributing to prediction errors and, accordingly, optimize the neural network model's weights and biases to improve future prediction accuracy.

[0293] Finally, the model update executor applies the adjustments derived by the parameter adjustment calculator to the neural network prediction model. This step enables real-time model updates, ensuring they remain optimal. The executor includes software and hardware resources that ensure the model update process is both secure and efficient, enabling model upgrades to be completed without disrupting normal system operations.

[0294] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. An intelligent perception and automatic adaptation power pole maintenance prediction system, characterized by: include: A sensing unit, mounted on a pole, configured to collect physical state data and environmental factor data of the pole, wherein the physical state data includes the inclination angle, vibration, temperature, and humidity of the pole, and the environmental factor data includes wind speed, rainfall, and air pressure; The prediction unit is used to predict the remaining life and potential failure risk of the pole based on the physical status data and environmental factor data collected by the sensing unit and using a pre-trained neural network prediction model; A decision-making unit, configured to generate maintenance recommendations and optimized maintenance plans for the poles based on the prediction results of the prediction unit; and provide the generated maintenance recommendations and optimized maintenance plans to a maintenance team; The adaptive unit is used to receive feedback information from the maintenance team; and adjust the parameters of the neural network prediction model according to the feedback information to improve the prediction accuracy.

2. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 1 is characterized in that: The sensing unit includes: Tilt angle sensor, used to measure the tilt angle of the pole; Accelerometers, used to measure the vibration of the pole; Temperature and humidity sensor, used to measure the temperature and humidity on the surface of the pole; Weather monitoring module, including wind speed sensor, rain gauge and barometer, used to measure environmental factor data; The data acquisition controller is used to coordinate the work of the sensors and perform preliminary processing and local storage on the collected data.

3. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 2 is characterized in that: The sensing unit further includes: Data preprocessing module, used to filter, denoise and standardize the collected raw data; A data compression encoder is used to compress and encode the pre-processed data to reduce the amount of data transmission; a wireless communication module, configured to periodically transmit the compressed data to the prediction unit; An energy management system, including solar panels and batteries, is used to provide a continuous power supply to the sensing unit.

4. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 1 is characterized in that: The decision-making unit includes: The risk assessment module is used to calculate the comprehensive risk index of the pole based on the remaining life and potential failure risk output by the prediction unit; Maintenance strategy library, which stores predefined maintenance strategies and corresponding triggering conditions; A strategy matcher is used to select an appropriate maintenance strategy from the maintenance strategy library based on the calculated risk index; Maintenance plan generator for converting selected maintenance strategies into specific maintenance recommendations and optimized maintenance plans.

5. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 1 is characterized in that: The adaptive unit comprises: Feedback receiving interface, used to receive feedback information submitted by the maintenance team; A feedback classifier is used to classify the received feedback information into prediction accuracy feedback, maintenance effect feedback and system improvement suggestions; A parameter adjustment calculator, used to calculate the adjustment amount of the neural network prediction model parameters based on the prediction accuracy feedback; A model update executor is used to apply the calculated parameter adjustments to the neural network prediction model.

6. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 1 is characterized in that: The neural network prediction model includes: The time series encoder is implemented using a bidirectional long short-term memory network structure. It is used to receive and process time series data from the sensing unit, including the physical state data of the pole and environmental factors, and obtain high-dimensional feature representations at each time point. A multi-scale attention module, comprising a daily self-attention submodule, a weekly self-attention submodule, a monthly self-attention submodule, and a fusion submodule; wherein the daily self-attention submodule is used to receive and process the high-dimensional feature representation provided by the time series encoder to obtain daily-level features reflecting short-term changes and daily fluctuations; the weekly self-attention submodule is used to receive and process the daily-level features, and in combination with the high-dimensional feature representation provided by the time series encoder, obtain weekly-level features reflecting medium-term data changes; the monthly self-attention submodule is used to receive and process the weekly-level features, and in combination with the high-dimensional feature representation provided by the time series encoder, obtain monthly-level features reflecting long-term data changes; the fusion submodule is used to fuse the daily-level features, weekly-level features, and monthly-level features to obtain comprehensive features; The conditional autoregressive generator, implemented using a variational autoencoder, receives and processes the comprehensive features provided by the multi-scale attention module to generate a probability distribution of the future state of the pole; The risk assessment decoder, implemented using a multi-layer perceptron, is used to receive and process the probability distribution output by the conditional autoregressive generator to obtain the remaining life prediction and potential failure risk of the pole.

7. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 6 is characterized in that: The time series encoder includes a dynamically adjusted gating mechanism; the gating mechanism uses an adaptive parameter adjustment strategy to adjust the weights of the forget gate and the input gate in the bidirectional long short-term memory network in real time according to the statistical characteristics of the received time series data; The input gate is calculated by the following formula 1: i t =σ(W i ·[h t-1 ,x t ]+U i ·s t +b i ) (1) Among them, i t represents the activation value of the input gate; σ is the sigmoid activation function; W i is the weight matrix of the input gate; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; U i is the adaptive adjustment matrix of the input gate; s t is a statistical feature vector calculated based on the frequency and amplitude of current data changes; b i is the bias term; The forget gate is calculated by the following formula 2: f t =σ(W f ·[h t-1 ,x t ]+U f ·s t +b f ) (2) Among them, f t represents the activation value of the forget gate; σ is the sigmoid activation function; W f is the weight matrix of the forget gate; h t-1 is the hidden state of the previous time step; x t is the input of the current time step; U f is the adaptive adjustment matrix of the forget gate; s t is a statistical feature vector calculated based on the frequency and amplitude of current data changes; b f is the bias term.

8. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 6 is characterized in that: The fusion submodule of the multi-scale attention module is implemented using an adaptive weight adjustment algorithm driven by environmental factors. The adaptive weight adjustment algorithm includes the following steps: According to the following formula 3, the fusion weight ω at the daily level is calculated d , the fusion weight of the weekly level features ω w And the fusion weight ω of the monthly level features m : Among them, d, w, m represent the index of the fusion weight of the daily, weekly and monthly features respectively; i is one of the indexes d, w, m, ω i represents the i-th fusion weight in the fusion weights of daily, weekly and monthly features; γ is the adjustment factor; RMSE i is the root mean square error of the i-th time level on the historical data; RMSE j is the root mean square error of the jth time level on the historical data; According to the following formula 4, the fusion weight ω i Make adjustments: Where α is the learning rate; ΔE i It is a quantitative indicator of the impact of environmental factors on the characteristics of the i-th time level; E total It is the total effect of all environmental factors; is the updated fusion weight ω i ; is the fusion weight ω before update i ; Quantitative indicator ΔE i Calculated by the following formula 5: in, Indicates the rate of environmental change; Corr(E,PE i ) is the environmental factor E and the prediction error PE of the i-th level i The correlation coefficient between the total impact of all environmental factors E total Calculated by the following formula 6: Among them, w k is the weight of the kth environmental factor; E env,k It is the sensitivity score of the kth environmental factor, which reflects the impact of the kth environmental factor on the operating status of the pole.

9. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 6 is characterized in that: The variational autoencoder structure adopted by the conditional autoregressive generator includes: The encoding unit is used to receive comprehensive features from the multi-scale attention module and enhance the representation ability of the features through nonlinear transformation to adapt to the specific needs of pole maintenance prediction; The mapping and sampling unit is used to optimize the coverage and accuracy of the probability distribution through an adaptive sampling mechanism to ensure that the probability distribution can effectively reflect the future state of the pole; A decoding unit that converts the representation of the latent space into a detailed probability distribution of the future states of the poles.

10. The intelligent perception and automatic adaptation pole maintenance prediction system according to claim 6, characterized in that: The multi-layer perceptron used in the risk assessment decoder includes: The input layer is used to receive the probability distribution output by the conditional autoregressive generator; Multiple hidden layers, where each hidden layer contains several neurons, and the number of neurons between adjacent hidden layers decreases, forming a funnel-shaped structure; the neurons in each hidden layer process the output of the previous layer through a nonlinear activation function; The skip connection module is used to pass the information of the input layer directly to the intermediate hidden layer to preserve the key features of the original probability distribution; Attention gating unit, set after each hidden layer, is used to dynamically adjust the importance weight of each feature; Adaptive dropout layer, which automatically adjusts the dropout ratio based on the uncertainty of the input data to improve the robustness of the model; Forked output layer, including: The life prediction sublayer is used to output the predicted value of the remaining life of the pole; and The risk assessment sublayer is used to output the probability distribution of potential failure risks.

Images