An icing monitoring terminal module fault self-diagnosis and early warning method, system, device and medium

Abstract

This invention relates to the field of fault prediction technology, and discloses a method, system, device, and medium for self-diagnosis and early warning of faults in an icing monitoring terminal module. The method includes: synchronously acquiring multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal, performing adaptive feature fusion to obtain an enhanced health feature vector, and inputting this vector into a pre-trained multi-scale deep residual network for fault classification and diagnosis, outputting fine-grained fault categories and confidence levels for each functional module within the terminal; responding to a diagnosis result of normal but the enhanced health feature vector indicating potential degradation risk, performing probabilistic prediction of the future state of key performance indicators based on quantile regression forest, and performing early degradation warning and probabilistic prediction of remaining effective lifespan based on the prediction results; performing interpretability analysis on the diagnosis and prediction results, and generating a structured diagnostic report. This invention provides solid support for the digital transformation of icing monitoring in transmission lines.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, and in particular to a method, system, device and medium for self-diagnosis and early warning of faults in an icing monitoring terminal module. Background Technology

[0002] Icing on transmission lines is one of the major natural disasters threatening the safe and stable operation of the power grid, especially in plateau, mountainous, and areas with complex microclimates, where icing disasters can lead to serious accidents such as line breaks and tower collapses. To effectively monitor the icing process, the power system has widely deployed icing monitoring terminals, which collect operating parameters in real time through sensors installed on the towers, providing data support for icing early warning and de-icing decisions.

[0003] However, the operation and maintenance management of existing icing monitoring terminals mainly relies on regular manual inspections and simple threshold alarm mechanisms, which have significant technical shortcomings. On the one hand, fault detection is severely delayed. The system usually only passively detects anomalies after the terminal is completely offline or data is interrupted for a long time. By this time, the monitoring window for critical icing processes has often been missed, resulting in the inability to record the complete icing formation and growth process, which seriously affects the accuracy of subsequent disaster analysis. On the other hand, fault location capabilities are vague. Traditional methods can only report general status and cannot distinguish whether the fault originates from the sensor module, power module, communication module, or main control module. On-site maintenance personnel often need to carry multiple spare parts to climb the tower to check one by one, consuming a lot of manpower and time costs. Moreover, the risk of operation increases significantly in severe weather conditions such as high altitude and strong winds. In addition, the existing system lacks early warning capabilities for gradual degradation. The performance degradation process of icing monitoring terminals is slow and gradual, while the traditional threshold alarm method only triggers an alarm after parameters exceed the limit, which cannot identify risks in the early stages of degradation. This results in the system operating in a sub-healthy state for a long time and cannot achieve predictive maintenance.

[0004] The persistent existence of these problems has kept the operation and maintenance of icing monitoring terminals in a reactive mode, often resulting in missed maintenance windows and difficulty in effectively controlling the risk of monitoring interruptions under extreme weather conditions. Therefore, there is an urgent need for an intelligent operation and maintenance method to promote the transformation of icing monitoring systems from passive maintenance to proactive predictive maintenance. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, this invention provides a method, system, device, and medium for self-diagnosis and early warning of icing monitoring terminal module faults, solving the problems of ambiguous fault location, lack of early warning capabilities, and passive and delayed maintenance in the operation and maintenance of icing monitoring terminals.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for self-diagnosis and early warning of faults in an icing monitoring terminal module, comprising: Multi-source heterogeneous data are synchronously collected from multiple functional modules within the icing monitoring terminal, and adaptive feature fusion is performed on the collected data to obtain an enhanced health feature vector. The enhanced health feature vector is input into a pre-trained multi-scale deep residual network for fault classification and diagnosis, and the fine-grained fault categories and confidence levels of each functional module inside the terminal are output. In response to a diagnosis result that is normal but the enhanced health feature vector shows a potential risk of degradation, a probability prediction of the future state of key performance indicators is made based on quantile regression forest, and an early degradation warning and a probability prediction of the remaining effective lifespan are made based on the prediction results. An interpretability analysis is performed on the diagnostic results of the multi-scale deep residual network and the prediction results of the quantile regression forest to generate a structured diagnostic report.

[0008] As a preferred embodiment of the self-diagnosis and early warning method for icing monitoring terminal module faults described in this invention, the multi-scale deep residual network includes three parallel feature extraction branches: The first branch uses small-sized one-dimensional convolutional kernels to construct residual blocks to capture sudden anomaly features at the second to minute level; The second branch uses a medium-sized one-dimensional convolution kernel and introduces dilated convolution to construct residual blocks, which are used to capture periodic changes and trend drift features at the hour to day level; The third branch uses a large-size one-dimensional convolutional kernel to construct residual blocks and connects the output of the residual blocks to a gated recurrent unit to identify slow performance degradation characteristics at the weekly to monthly level. The multi-scale deep residual network also includes a channel attention fusion module, which adaptively fuses the feature maps output by the three branches. The fused feature maps are then fed into a fully connected classification layer after global average pooling, and output the probability distribution of each fine-grained fault category.

[0009] As a preferred embodiment of the self-diagnosis and early warning method for icing monitoring terminal module faults described in this invention, the multi-scale deep residual network is trained using a focal loss function. During the training process, the focal loss function reduces the loss weight of easily classified samples and increases the loss weight of difficult-to-classify samples and samples with fewer fault categories.

[0010] As a preferred embodiment of the self-diagnosis and early warning method for icing monitoring terminal module faults described in this invention, the step of predicting the future state of key performance indicators based on quantile regression forest, and predicting the early degradation warning and remaining effective lifespan based on the prediction results, includes: A quantile regression forest model is established for key performance indicators to predict the conditional distribution of the key performance indicators under multiple specified quantiles within a preset time domain in the future. Execute first criterion warning and second criterion warning based on the predicted conditional distribution; The first criterion warning includes: estimating the actual slope of the current key performance indicator within a recent time window, and the normal slope baseline of the indicator during the historical normal operation phase; constructing a t-statistic to perform hypothesis testing on the difference between the actual slope and the normal slope baseline; and triggering a trend deviation warning when the test results determine that the difference is significant and the actual slope is greater than the normal slope baseline. The second criterion warning includes: estimating the probability that a key performance indicator will exceed a preset failure threshold at a future preset time domain endpoint based on the predicted condition distribution; and triggering a threshold exceedance risk warning when the probability exceeds a preset risk level threshold. When the trend deviates from the warning or the threshold exceeds the risk warning, the system issues an early degradation warning and simultaneously calls a pre-trained deep survival analysis model to output the probability density function of the remaining effective lifetime of the module corresponding to the key performance indicators.

[0011] As a preferred embodiment of the self-diagnosis and early warning method for icing monitoring terminal module faults described in this invention, the interpretability analysis includes: The contribution of each input feature in the diagnostic results of the multi-scale deep residual network is calculated using the SHAP method, and the key features that contribute the most to fault classification are identified. A gradient-based time importance mapping method is used to calculate the sensitivity of each time step in the input feature sequence to the fault classification decision. After aggregation along the feature dimension, the importance weight of each time step is obtained, and a time series saliency map is generated to highlight the key time points that have the greatest impact on the decision. The diagnostic results, prediction results, SHAP contribution analysis results, and time series saliency plots are integrated to automatically generate a structured diagnostic report and push it to the operations and maintenance personnel.

[0012] As a preferred embodiment of the self-diagnosis and early warning method for icing monitoring terminal module faults described in this invention, the step of synchronously acquiring multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal includes: The physical measurements and status parameters of the sensing module, power module, communication module and main control module are collected synchronously at fixed intervals. The parameters collected by the sensing module include the laser transmitter drive current, receiver signal-to-noise ratio, and internal reference voltage; the parameters collected by the power supply module include the solar panel output voltage, battery terminal voltage, battery internal resistance, and load current; the parameters collected by the communication module include the transmit power, received signal strength, link bit error rate, and network registration status; and the parameters collected by the main control module include the CPU load rate, memory usage, internal temperature, and watchdog reset count.

[0013] As a preferred embodiment of the self-diagnosis and early warning method for icing monitoring terminal module faults described in this invention, the step of adaptively fusing the collected data to obtain an enhanced health feature vector includes: Timestamp alignment and Z-score normalization are performed on each feature sequence; Within a preset sliding time window, the coefficient of variation of each standardized feature and its mutual information with other features are calculated. Based on the coefficient of variation and the mutual information, the fusion weight of each feature is calculated using the Softmax function. The enhanced health feature vector is obtained by weighting and summing the standardized features according to the fusion weights.

[0014] Secondly, the present invention provides a fault self-diagnosis and early warning system for an icing monitoring terminal module, comprising: The multi-source data feature reconstruction module is used to synchronously collect multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal, and to perform adaptive feature fusion on the collected data to obtain an enhanced health feature vector. The multi-scale residual fault classification module is used to input the enhanced health feature vector into a pre-trained multi-scale deep residual network for fault classification and diagnosis, and output the fine-grained fault categories and confidence levels of each functional module inside the terminal. The degradation risk probability prediction module is used to respond to a diagnosis result that is normal but the enhanced health feature vector shows potential degradation risk, to make a probability prediction of the future state of key performance indicators based on quantile regression forest, and to make an early degradation warning and a probability prediction of the remaining effective lifespan based on the prediction results. The diagnostic report parsing and generation module is used to perform interpretability analysis on the diagnostic results of the multi-scale deep residual network and the prediction results of the quantile regression forest, and generate a structured diagnostic report.

[0015] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor, when executing the computer-executable instructions, implements the steps of a method for self-diagnosis and early warning of faults in an icing monitoring terminal module.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of a method for self-diagnosis and early warning of faults in an icing monitoring terminal module.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention improves the accuracy of fault diagnosis from the general terminal level to the refined module level, enabling precise identification and location of faults in specific modules such as sensors, power supplies, communication, or main control systems. It further subdivides these faults into specific fault modes such as battery aging and laser attenuation, providing clear elimination targets for on-site maintenance and significantly improving operational efficiency. This invention pioneers a new paradigm of early warning based on probabilistic prediction, overcoming the insensitivity of traditional threshold alarms to slow performance degradation. It can issue early warnings and estimate remaining effective lifespan in a probabilistic form days or even weeks before a module's function completely fails, truly realizing a fundamental shift from passive post-event repair to proactive predictive maintenance. The core model of this invention possesses strong adaptive and generalization capabilities. Through multi-scale temporal feature extraction and adaptive feature fusion mechanisms, it can automatically learn and adapt to different environments, different terminal models, and complex fault modes ranging from transient interference to long-term degradation, and can continuously iterate and optimize under the drive of new operational data. This invention forms a closed-loop intelligent system that not only outputs diagnostic and early warning results, but also provides decision-making basis through interpretability analysis and automatically generates structured maintenance reports. Furthermore, the maintenance results can be fed back to the system to form a learning loop, thereby constructing a complete technology chain from state perception and intelligent analysis to precise execution and knowledge accumulation, providing solid support for the digital transformation of transmission line icing monitoring. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0019] Figure 1 This is a schematic diagram of the overall process logic of the self-diagnosis and early warning method for icing monitoring terminal module faults provided in an embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for self-diagnosis and early warning of faults in an icing monitoring terminal module is provided, such as... Figure 1 The specific steps shown are as follows: S100: Simultaneously collects multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal, and performs adaptive feature fusion on the collected data to obtain an enhanced health feature vector. In this embodiment of the invention, the synchronous acquisition of multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal includes: With the help of the monitoring circuit and self-diagnostic interface embedded in the terminal, the physical measurement values ​​and status parameters of the sensing module, power module, communication module and main control module are collected synchronously at fixed intervals. The parameters collected by the sensing module include the laser emitter drive current. (mA), receiver signal-to-noise ratio (dB), internal reference voltage (V); The parameters collected by the power module include the output voltage of the solar panel. (V) Battery terminal voltage (V) and internal resistance (mΩ), load current (mA); The parameters collected by the communication module include transmit power. (dBm) Received signal strength (dBm), link bit error rate Network registration status The parameters collected by the main control module include CPU load rate. (%) Memory utilization (%), internal temperature (°C), number of times the watchdog resets .

[0022] In this embodiment of the invention, adaptive feature fusion is performed on the collected data to obtain an enhanced health feature vector, including: Timestamp alignment and Z-score normalization are performed on each feature sequence; Within a preset sliding time window, calculate the coefficient of variation for each standardized feature. and mutual information with other features Based on the coefficient of variation and mutual information, the fusion weight of each feature is calculated using the Softmax function; The enhanced health feature vector is obtained by weighting and summing the standardized features according to the fusion weights.

[0023] Specifically, the fusion weights for each feature are calculated using the Softmax function, expressed by the following formula: in, as well as These are adjustable hyperparameters used to balance the importance of variability and correlation. Specifically, the weighted fusion results in a unified enhanced health feature vector that highlights abnormal changes. ,in These are the standardized features.

[0024] It should be noted that step S100 above, by synchronously collecting heterogeneous parameters from multiple modules and adopting an adaptive weighted fusion strategy based on the coefficient of variation and mutual information, can effectively eliminate the dimensional differences and asynchronous problems between multi-source data, dynamically highlight key features with significant fluctuations and strong correlations, and suppress redundant or stable secondary information, thereby generating an enhanced health feature vector with a higher signal-to-noise ratio. This provides a pure and highly representative input for the subsequent diagnostic model, avoids misjudgments caused by too many interference terms in the original data, and significantly improves the distinguishability of fault features and the robustness of the system.

[0025] S200: The enhanced health feature vector is input into a pre-trained multi-scale deep residual network for fault classification and diagnosis, and the fine-grained fault categories and confidence levels of each functional module inside the terminal are output. In this embodiment of the invention, the multi-scale deep residual network includes three parallel feature extraction branches: The first branch uses small-sized one-dimensional convolutional kernels to construct residual blocks to capture sudden anomaly features at the second to minute level; The second branch uses a medium-sized one-dimensional convolution kernel and introduces dilated convolution to construct residual blocks, which are used to capture periodic changes and trend drift features at the hour to day level; The third branch uses a large-size one-dimensional convolutional kernel to construct residual blocks and connects the output of the residual blocks to a gated recurrent unit to identify slow performance degradation characteristics at the weekly to monthly level. The multi-scale deep residual network also includes a channel attention fusion module, which adaptively fuses the feature maps output by the three branches. The fused feature maps are then fed into a fully connected classification layer after global average pooling, and output the probability distribution of each fine-grained fault category.

[0026] Specifically, the multi-scale deep residual network adopts an architecture of multi-scale parallel residual network + channel attention fusion. The overall structure includes 5 steps: first, the input feature sequence is connected to the first branch (local fine-grained branch), the second branch (trend analysis branch), and the third branch (long-period degradation branch). The first, second, and third branches are connected to the channel attention fusion module, then to the fully connected classification layer, and finally output the probability distribution P.

[0027] Specifically, the time step T takes 60 past sampling points (1 minute per sampling point, totaling 1 hour of historical data). This can be dynamically adjusted based on actual needs, but a fixed dimension facilitates batch processing. The feature dimension D is the total number of features collected from the four main modules (16 dimensions) plus the enhanced health features generated after adaptive fusion (an additional dimension) → approximately 17 dimensions in total. In practical engineering, this can be extended to 28 dimensions (including statistics such as variance and slope). This paper uses D=28 as an example, with input tensor dimensions [batch, T=60, D=28].

[0028] Specifically, the input of the first branch is [T,D] → reshaped into [T,D] (as a single-channel feature map, the number of channels = the feature dimension); Convolutional layer 1 is Conv1D(kernel=3, stride=1, padding='same', out_channels=32) → BN → ReLU; Residual block 1 is Conv1D(3,32)→BN→ReLU→Conv1D(3,32)→add input (shortcut)→ReLU; Residual block 2 is the same as above, and the output channel remains 32; Convolutional layer 2 is Conv1D(3,64)→BN→ReLU→output feature map F1:[T,64].

[0029] Specifically, the input to the second branch is [T,D]; Convolutional layer 1 is ; Residual blocks (dilated convolution) are ; Convolutional layer 2 is .

[0030] Specifically, the input to the third branch is [T,D]; Convolutional layer 1 is ; The residual block is ; The GRU layer takes input [T, 32], has 64 hidden units, and outputs [T, 64]. Furthermore, to align the outputs of the three branches, global average pooling is performed along the time dimension on the feature map of each branch, resulting in a

[64] -dimensional vector. The three

[64] vectors are stacked into [3,64], and global pooling is performed along the feature dimension → or more commonly: the weights of each branch are computed independently through a small fully connected network. Finally, Global average pooling along the time dimension yields a

[64] -dimensional feature vector, which is then fed into a fully connected classification layer.

[0031] Furthermore, the fully connected layer 1 from Fully connected layer 2 is The output layer is 16→C (C=9, the number of fine-grained fault categories). Output probability distribution P.

[0032] Specifically, the steps of global average pooling include: outputting feature maps for each branch. Average along the time dimension ; Specifically, the steps for weight calculation include: dividing the three... spliced ​​as Send it into a two-layer fully connected network. ,in Output 16 dimensions. ,in Output three weights ,satisfy ; Specifically, the weighted fusion steps include: applying weights to each time step and channel of the original feature map. Output .

[0033] Specifically, in the module-level fault diagnosis of icing monitoring terminals, based on engineering experience, fine-grained fault categories can be divided into the following 9 types:

[0034] In practical applications, adjustments can be made based on the specific hardware composition of the terminal.

[0035] In this embodiment of the invention, the multi-scale deep residual network is trained using a focal loss function. During the training process, the focal loss function reduces the loss weight of easily classified samples and increases the loss weight of difficult-to-classify samples and few-class fault samples.

[0036] Specifically, to address the extreme imbalance problem where faulty samples are far fewer than normal samples in reality, an improved focus loss function is used to train the model, expressed by the formula: The total number of fault categories is The true label of the sample (one-hot encoded) is The model predicts that the sample belongs to the category. The probability is The focusing parameter is ( This operation can reduce the number of easily classified samples. The loss weights, which are close to 1), allow the model to focus more on processing difficult-to-distinguish and rare fault samples.

[0037] Specifically, the training set consists of historical labeled data, with various fault samples oversampled or synthesized using SMOTE. The optimizer is... The batch size is 64. Early stopping occurs when the validation set loss does not decrease for 10 consecutive rounds. Inference uses a real-time sliding window input, the model outputs probabilities, and the class with the highest probability is taken as the diagnosis result, while the probability value is also output as the confidence level.

[0038] It should be noted that the multi-scale deep residual network constructed in step S200 above captures second-level burst pulses, hour-level trend drifts, and month-level slow degradation features respectively. Therefore, it can simultaneously perceive transient anomalies and long-term performance degradation, overcoming the limitation of single-scale models that cannot take into account multiple time-scale fault modes. By combining the channel attention mechanism to adaptively fuse the features of the three branches, the network can dynamically adjust the branch weights according to different fault types, achieving accurate module-level fault location, accurately distinguishing fine-grained faults, and significantly reducing on-site troubleshooting time and blind replacement costs.

[0039] S300: In response to a diagnosis result that is normal but the enhanced health feature vector shows potential degradation risk, it uses quantile regression forest to predict the future state of key performance indicators and makes early degradation warnings and probabilistic predictions of remaining effective lifespan based on the prediction results. In this embodiment of the invention, step S300 includes: A quantile regression forest model is established for key performance indicators to predict the conditional distribution of key performance indicators at multiple specified quantiles within a preset time domain in the future. Execute first criterion warning and second criterion warning based on the predicted conditional distribution; The first criterion for early warning includes: estimating the actual slope of the current key performance indicator within a recent time window, and the normal slope baseline of the indicator during its historical normal operation phase. The difference between the actual slope and the normal slope baseline is tested by constructing a t-statistic. When the test results indicate that the difference is significant and the actual slope is greater than the normal slope baseline, a trend deviation warning is triggered. The second criterion warning includes: estimating the probability that key performance indicators will exceed a preset failure threshold at the future preset time domain endpoint based on the predicted condition distribution; when the probability exceeds a preset risk level threshold, a threshold exceedance risk warning is triggered. When a trend deviation warning or a threshold exceeding a risk warning is triggered, the system issues an early degradation warning and simultaneously calls a pre-trained deep survival analysis model to output the probability density function of the remaining effective lifetime of the module corresponding to the key performance indicators.

[0040] Specifically, in response to a diagnostic result that is normal but whose enhanced health feature vector indicates potential degradation risk, key performance indicators (KPIs) such as battery internal resistance are targeted. or sensor signal-to-noise ratio A quantile regression forest model was established. This model predicts future time-domain events. Within, KPIs at a specific quantile (like Conditional distribution under ) .

[0041] Specifically, the first criterion for early warning includes: For each key performance indicator, during the early stages of normal operation of the terminal (such as the first month after commissioning), its numerical series are collected, and linear regression is used to estimate the slope of normal variation. The formula is expressed as: Simultaneously calculate the residual standard deviation. In real-time monitoring, the past... Perform linear regression on each point to obtain the current slope. and its standard error Test whether the current slope deviates significantly from the normal baseline. Null hypothesis And construct the t-statistic: in This represents the standard error of the normal slope. If... If the null hypothesis is rejected, it is considered that there is a significant deviation, where the degrees of freedom are... .

[0042] like Furthermore, a significant deviation indicates an accelerated degradation trend. A significant negative deviation may indicate measurement anomalies or recovery, and is usually not flagged as a warning.

[0043] When the statistical significance (p<0.05) and the deviation ratio are met... When the deviation is detected, a "Trend Deviation Warning" will be triggered. Two levels can be set: 30%~60% is medium level, and >60% is high level.

[0044] Furthermore, the second criterion warning includes: Calculate future time intervals When the KPI value exceeds the preset failure threshold probability This probability can be approximated using quantile prediction results, expressed by the formula: when Once the risk level exceeds the preset value (e.g., >70% is a medium warning, >90% is a high warning), the system will immediately issue a warning.

[0045] Furthermore, using full lifecycle data from similar modules, a deep survival analysis model is trained to output the remaining effective lifetime of key modules. probability density function This system provides a quantitative basis for spare parts management and planned maintenance. The early warning level classification uses a dual-indicator fusion: a probability-based risk level and a trend deviation-based level. The higher level is ultimately selected, with the specific criteria as follows: Based on threshold exceedance probability Level:

[0046] Levels based on trend deviation:

[0047] The highest level of the two assessments mentioned above will be selected. For example, if the probability risk is level two (70%~90%) and the trend deviation is level one (>60%), then the final warning level will be level one. The structured warning message pushed by the system should include the following five fields: Warning Title: [Warning] Terminal U107 Power Module - Battery Aging Risk Warning levels: Level 1 (Red) / Level 2 (Orange) / Level 3 (Yellow) Current status: internal resistance, degradation rate, number of days running Forecast information: Estimated remaining useful life (RUL 90% confidence interval) Recommended measures: Level 3: Strengthen monitoring and generate a degradation report weekly; Level 2: Include in the monthly maintenance plan and prepare spare parts; Level 1: Immediately arrange for personnel replacement to avoid affecting icing monitoring.

[0048] It should be noted that step S300 above can output the distribution range of key performance indicators in the future, rather than a single predicted value, and issue early warnings through a dual criterion of trend deviation hypothesis testing and threshold exceedance probability. This allows for risk warnings to be issued in the early stages of degradation, achieving true early warning. At the same time, combined with the remaining effective life probability density function output by the deep survival analysis model, the system can quantify the time window of failure risk, providing a scientific basis for spare parts management and planned maintenance. This transforms the traditional passive emergency repair mode into proactive predictive maintenance, effectively avoiding the risk of monitoring interruption under extreme weather conditions.

[0049] S400: Perform interpretability analysis on the diagnostic results of multi-scale deep residual networks and the prediction results of quantile regression forests, and generate structured diagnostic reports. In this embodiment of the invention, step S400 includes: The SHAP method is used to calculate the contribution of each input feature in the diagnostic results of multi-scale deep residual networks, and to identify the key features that contribute the most to fault classification. A gradient-based time importance mapping method is used to calculate the sensitivity of each time step in the input feature sequence to the fault classification decision. After aggregation along the feature dimension, the importance weight of each time step is obtained, and a time series saliency map is generated to highlight the key time points that have the greatest impact on the decision. The diagnostic results, prediction results, SHAP contribution analysis results, and time series saliency plots are integrated to automatically generate a structured diagnostic report and push it to operations and maintenance personnel.

[0050] It should be noted that, to increase the credibility and operability of the diagnostic results, SHAP and time feature mapping technology are used to make the model's decisions easier to understand. This allows for a clear view of the key features and corresponding times that are most helpful in fault classification. The cloud platform automatically merges diagnostic results, warnings, lifetime predictions, and interpretive analysis results to generate a structured diagnostic report. The report includes: "Confirmed faulty module and root cause analysis," "Performance degradation curve and predicted trajectory of the problematic module," "Estimated remaining effective lifetime (mean and 90% confidence interval)," "Recommended maintenance measures (such as immediate replacement, focused attention, continued observation)," and "Similar historical examples for reference," and is pushed to relevant operations and maintenance personnel, forming a complete process from intelligent perception to decision assistance.

[0051] Specifically, gradient-based temporal importance mapping includes: For a pre-trained multi-scale deep residual network, select the score of the last classification layer. As target category Decision score; calculate For the input feature sequence gradient: By taking the absolute value, summing, or calculating the L2 norm along the feature dimension, the importance weight of each time step can be obtained. Normalization Plot a time-importance curve (i.e., a time significance plot) in the interval [0,1]. The time point corresponding to the peak value represents the moment with the greatest impact on the decision.

[0052] When the model diagnoses "battery aging", the system calculates the importance of each time step in the input sequence, highlighting the time periods when the internal resistance suddenly increases or fluctuates drastically. Maintenance personnel can use this information to trace the cause of the anomaly and present it in the report in the form of a heat map or curve.

[0053] It should be noted that step S400 above introduces SHAP contribution analysis and gradient-based time importance mapping technology to visualize and explain the decision-making basis of the deep learning model, significantly improving the credibility and operability of the diagnostic results. On this basis, the system automatically integrates diagnostic conclusions, degradation trajectories, life predictions and historical similar cases to generate a structured report containing graded maintenance recommendations, realizing a complete closed loop from intelligent perception to decision support. This facilitates maintenance personnel to quickly understand the equipment status and take targeted measures, effectively improving the intelligent operation and maintenance level and management efficiency of the icing monitoring terminal.

[0054] Example 2, based on the previous example, provides an application example of a method for self-diagnosis and early warning of faults in an icing monitoring terminal module, to verify and illustrate the technical effect of the method.

[0055] This embodiment uses a 110kV transmission line in a high-altitude mountainous area as an example. The line L101 is about 25 kilometers long and is equipped with 15 icing monitoring terminals (numbered U101-U115). Icing disasters often occur in winter. This time, we focus on observing terminal U107 and show the entire process of its internal module fault being detected, diagnosed and issued as a warning.

[0056] Terminal U107 contains four monitoring units and sensing modules that use laser ranging principles. The internal optical power monitoring circuit collects the laser drive current in real time. The power module includes a 100Ah lithium iron phosphate battery, and the management chip collects the terminal voltage. and internal resistance The communication module is a 4G DTU, and the module provides the transmit power. The signal quality is monitored by RSRP, the main control module is an ARM processor, and the operating system reports the CPU load rate. All data is packaged in 1-minute increments and uploaded to the cloud-based intelligent diagnostic platform via a 4G network.

[0057] The platform receives the time-series data stream uploaded by U107. Before entering the key model, the data first passes through an adaptive feature fusion layer, for example... The platform calculates the coefficient of variation (CV) and mutual information (MI) with other features for each feature within a one-hour time window, based on a set of data collected in real time. The fluctuation is very small (CV=0.02), but it is similar to... The correlation is very strong (MI=0.85), and The self-fluctuation is significant (CV=0.15). Let the hyperparameter be... =1.0, =0.8, The fusion weight is calculated as follows: It should be noted that this weighting is characterized by low volatility and weak correlation, which becomes more prominent after weighted fusion. Overall health feature vector of changing information This allows subsequent analysis to utilize inputs with a high signal-to-noise ratio.

[0058] Furthermore, the fused feature sequences The data is fed into a pre-trained multi-scale deep residual network, and local branches quickly capture the data. Several instantaneous spikes occurred within a few minutes. Trend branch analysis shows... The average value for the most recent day has maintained a slow but steady upward trend. Comparing long-term branches with historical data reveals... The baseline value has increased by approximately 25% over the past three months. The network, utilizing channel attention, automatically assigns higher weights to trending and long-period branches. This is because the current form better reflects slow degradation rather than instantaneous disturbances. After passing through fully connected layers, the model outputs a probability distribution. Since power failure samples constitute a small percentage of the training set, a focus loss function is used during model training to enhance learning of the minority class. The formula is as follows: The model calculations showed that the probability of power module-battery aging was 0.76, the normal probability was 0.21, and the other probability was 0.03. The system selected the result with the highest probability, and the diagnosis was "power module failure - battery aging", with a confidence level of 76%. Similarly, the network analysis results for the sensing, communication, and main control modules were all normal. This achieved accurate module-level fault location, rather than a general terminal abnormality.

[0059] Furthermore, after the diagnosis is completed, for batteries confirmed to be aging, the system initiates an early warning and lifespan estimation process, and the platform calls upon the appropriate technologies... The trained quantile regression forest model estimates the next 7 days' (t) based on current and historical H(t). =10080 minutes) The trajectory range. The estimation results indicate that on the 7th day... The estimated value for the 0.5 quantile (median) is 45 mΩ, and the estimated value for the 0.9 quantile is 50 mΩ.

[0060] The failure threshold of this type of battery is known. The resistance is 55mΩ, and the system calculates for the next 7 days. The probability of exceeding this threshold is: At this point, although the absolute probability of exceeding the threshold (7%) has not reached the advanced warning line, the system simultaneously performed a trend deviation analysis, and the calculation showed that currently... The battery resistance was increasing by 0.5mΩ per day, much faster than the normal 0.1mΩ. Both criteria indicated a significant risk of degradation, so the system issued a medium-level alarm: "Alarm target: Terminal U107 power module; Problem type: Accelerated decline in battery performance; Reminder: Check frequently, once a month."

[0061] More importantly, the deep survival network within the platform uses a large amount of full lifecycle data from similar batteries to predict the battery's remaining effective lifespan. The model outputs the probability density function of the battery's RUL (Right-Use Life). This indicates a 90% probability of failure within 15 to 40 days, providing a precise time window for planned maintenance.

[0062] Furthermore, after receiving the alert, maintenance personnel can use the platform to view the interpretability analysis results. The SHAP analysis chart clearly indicates that the top three features contributing most to this "battery aging" diagnosis are, in order: 72-hour trend items and covariance, and Its own instantaneous fluctuations. This uses data to illustrate the diagnostic basis, making the results more reliable.

[0063] The platform automatically integrated all the information and generated the following structured diagnostic report: Fault diagnosis conclusion: The power module battery is aging (76% confidence level).

[0064] Key evidence: internal resistance The rate of increase is accelerating, and the current value is 25% higher than the baseline three months ago.

[0065] Risk prediction: The probability of performance degradation within the next week is 7%, but the degradation rate will accelerate. The remaining effective lifespan is expected to have a 90% confidence interval of [15, 40] days.

[0066] Maintenance suggestion: It is best to replace it during the planned inspection within two weeks. If a cold wave warning is issued, it should be dealt with in advance.

[0067] Historical reference: There were 12 similar cases, and the fault was eliminated after replacing the battery.

[0068] Furthermore, as required by the report, the maintenance team replaced the U107 battery during the scheduled maintenance window 20 days later. The data uploaded by the terminal after the replacement... The data immediately returned to normal, and the relevant warnings were automatically lifted. The entire process of capturing this latent degradation, confirming its accurate diagnosis, and executing predictive maintenance was fully recorded and fed back into the model training set as a new positive sample. The system continuously learns from such closed-loop instances, constantly improving diagnostic accuracy and early warning lead time, thus achieving true intelligent evolution.

[0069] This embodiment fully illustrates how the present invention deeply integrates deep learning models with domain knowledge, transforming the traditional "fault response" approach into a new intelligent operation and maintenance paradigm of "fault early warning," significantly enhancing the reliability and economy of the icing monitoring system.

[0070] Example 3: This example provides a self-diagnosis and early warning system for icing monitoring terminal module faults, including: The multi-source data feature reconstruction module is used to synchronously collect multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal, and to perform adaptive feature fusion on the collected data to obtain an enhanced health feature vector. The multi-scale residual fault classification module is used to input the enhanced health feature vector into a pre-trained multi-scale deep residual network for fault classification and diagnosis, and output the fine-grained fault categories and confidence scores of each functional module inside the terminal. The degradation risk probability prediction module is used to respond to a diagnosis result that is normal but the enhanced health feature vector shows potential degradation risk. It uses quantile regression forest to predict the future state of key performance indicators and to make early degradation warnings and probability predictions of remaining effective lifespan based on the prediction results. The diagnostic report parsing and generation module is used to perform interpretability analysis on the diagnostic results of multi-scale deep residual networks and the prediction results of quantile regression forests, and generate structured diagnostic reports.

[0071] It should be noted that the technical solution of the self-diagnosis and early warning system for the icing monitoring terminal module fault is based on the same concept as the technical solution of the aforementioned self-diagnosis and early warning method for the icing monitoring terminal module fault. For details not described in detail in the technical solution of the self-diagnosis and early warning system for the icing monitoring terminal module fault in this embodiment, please refer to the description of the technical solution of the aforementioned self-diagnosis and early warning method for the icing monitoring terminal module fault.

[0072] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0073] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for self-diagnosis and early warning of icing monitoring terminal module faults. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0074] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0075] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. 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.

[0076] 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, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended 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 present invention.

Claims

1. A method for self-diagnosis and early warning of faults in an icing monitoring terminal module, characterized in that, include: Multi-source heterogeneous data are synchronously collected from multiple functional modules within the icing monitoring terminal, and adaptive feature fusion is performed on the collected data to obtain an enhanced health feature vector. The enhanced health feature vector is input into a pre-trained multi-scale deep residual network for fault classification and diagnosis, and the fine-grained fault categories and confidence levels of each functional module inside the terminal are output. In response to a diagnosis result that is normal but the enhanced health feature vector shows a potential risk of degradation, a probability prediction of the future state of key performance indicators is made based on quantile regression forest, and an early degradation warning and a probability prediction of the remaining effective lifespan are made based on the prediction results. An interpretability analysis is performed on the diagnostic results of the multi-scale deep residual network and the prediction results of the quantile regression forest to generate a structured diagnostic report.

2. The method for self-diagnosis and early warning of icing monitoring terminal module faults as described in claim 1, characterized in that, The multi-scale deep residual network includes three parallel feature extraction branches: The first branch uses small-sized one-dimensional convolutional kernels to construct residual blocks to capture sudden anomaly features at the second to minute level; The second branch uses a medium-sized one-dimensional convolution kernel and introduces dilated convolution to construct residual blocks, which are used to capture periodic changes and trend drift features at the hour to day level; The third branch uses a large-size one-dimensional convolutional kernel to construct residual blocks and connects the output of the residual blocks to a gated recurrent unit to identify slow performance degradation characteristics at the weekly to monthly level. The multi-scale deep residual network also includes a channel attention fusion module, which adaptively fuses the feature maps output by the three branches. The fused feature maps are then fed into a fully connected classification layer after global average pooling, and output the probability distribution of each fine-grained fault category.

3. The method for self-diagnosis and early warning of icing monitoring terminal module faults as described in claim 2, characterized in that, The multi-scale deep residual network is trained using a focal loss function, which reduces the loss weight of easily classified samples and increases the loss weight of difficult-to-classify samples and few-class fault samples during training.

4. The method for self-diagnosis and early warning of icing monitoring terminal module faults as described in claim 3, characterized in that, The method of using quantile regression forests to predict the future state of key performance indicators and, based on the prediction results, to predict the probability of early degradation warning and remaining effective lifespan includes: A quantile regression forest model is established for key performance indicators to predict the conditional distribution of the key performance indicators under multiple specified quantiles within a preset time domain in the future. Execute first criterion warning and second criterion warning based on the predicted conditional distribution; The first criterion warning includes: estimating the actual slope of the current key performance indicator within a recent time window, and the normal slope baseline of the indicator during the historical normal operation phase; constructing a t-statistic to perform hypothesis testing on the difference between the actual slope and the normal slope baseline; and triggering a trend deviation warning when the test results determine that the difference is significant and the actual slope is greater than the normal slope baseline. The second criterion warning includes: estimating the probability that a key performance indicator will exceed a preset failure threshold at a future preset time domain endpoint based on the predicted condition distribution; and triggering a threshold exceedance risk warning when the probability exceeds a preset risk level threshold. When the trend deviates from the warning or the threshold exceeds the risk warning, the system issues an early degradation warning and simultaneously calls a pre-trained deep survival analysis model to output the probability density function of the remaining effective lifetime of the module corresponding to the key performance indicators.

5. The method for self-diagnosis and early warning of icing monitoring terminal module faults as described in claim 4, characterized in that, The interpretability analysis includes: The contribution of each input feature in the diagnostic results of the multi-scale deep residual network is calculated using the SHAP method, and the key features that contribute the most to fault classification are identified. A gradient-based time importance mapping method is used to calculate the sensitivity of each time step in the input feature sequence to the fault classification decision. After aggregation along the feature dimension, the importance weight of each time step is obtained, and a time series saliency map is generated to highlight the key time points that have the greatest impact on the decision. The diagnostic results, prediction results, SHAP contribution analysis results, and time series saliency plots are integrated to automatically generate a structured diagnostic report and push it to the operations and maintenance personnel.

6. The method for self-diagnosis and early warning of icing monitoring terminal module faults as described in claim 1, characterized in that, The simultaneous acquisition of multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal includes: The physical measurements and status parameters of the sensing module, power module, communication module and main control module are collected synchronously at fixed intervals. The parameters collected by the sensing module include the laser transmitter drive current, receiver signal-to-noise ratio, and internal reference voltage; the parameters collected by the power supply module include the solar panel output voltage, battery terminal voltage, battery internal resistance, and load current; the parameters collected by the communication module include the transmit power, received signal strength, link bit error rate, and network registration status; and the parameters collected by the main control module include the CPU load rate, memory usage, internal temperature, and watchdog reset count.

7. The method for self-diagnosis and early warning of icing monitoring terminal module faults as described in claim 6, characterized in that, The adaptive feature fusion of the collected data to obtain the enhanced health feature vector includes: Timestamp alignment and Z-score normalization are performed on each feature sequence; Within a preset sliding time window, the coefficient of variation of each standardized feature and its mutual information with other features are calculated. Based on the coefficient of variation and the mutual information, the fusion weight of each feature is calculated using the Softmax function. The enhanced health feature vector is obtained by weighting and summing the standardized features according to the fusion weights.

8. A fault self-diagnosis and early warning system for an icing monitoring terminal module, employing the fault self-diagnosis and early warning method for an icing monitoring terminal module as described in any one of claims 1 to 7, characterized in that, include: The multi-source data feature reconstruction module is used to synchronously collect multi-source heterogeneous data from multiple functional modules within the icing monitoring terminal, and to perform adaptive feature fusion on the collected data to obtain an enhanced health feature vector. The multi-scale residual fault classification module is used to input the enhanced health feature vector into a pre-trained multi-scale deep residual network for fault classification and diagnosis, and output the fine-grained fault categories and confidence levels of each functional module inside the terminal. The degradation risk probability prediction module is used to respond to a diagnosis result that is normal but the enhanced health feature vector shows potential degradation risk, to make a probability prediction of the future state of key performance indicators based on quantile regression forest, and to make an early degradation warning and a probability prediction of the remaining effective lifespan based on the prediction results. The diagnostic report parsing and generation module is used to perform interpretability analysis on the diagnostic results of the multi-scale deep residual network and the prediction results of the quantile regression forest, and generate a structured diagnostic report.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the self-diagnosis and early warning method for the icing monitoring terminal module fault as described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the self-diagnosis and early warning method for icing monitoring terminal module faults as described in any one of claims 1 to 7.

Images