Voltage monitor fault prediction method and apparatus based on time series forecasting and generative model

Abstract

This invention discloses a method and device for predicting voltage monitor faults based on time-series prediction and generative models. First, a multi-dimensional health feature sequence of the voltage monitor is acquired and input into a pre-trained quantile regression time-series prediction model. This model uses a long short-term memory network with an attention mechanism as its backbone and is trained and optimized using a quantile loss function. It outputs multiple quantile prediction values ​​of the comprehensive health index within a preset future period, constituting a health prediction interval. Simultaneously, the sequence is input into a variational autoencoder model to calculate the reconstruction error and the latent space KL divergence, fusing them to generate a comprehensive anomaly score. Exceeding an anomaly threshold triggers a latent anomaly marker. Finally, based on the lower trend of the health prediction interval, the comprehensive anomaly score, and the latent anomaly markers, a Bayesian fusion mechanism is used for collaborative decision-making to generate an early warning signal. This invention overcomes existing limitations, enabling early prediction of equipment performance degradation, detection of latent anomalies, avoidance of various adverse consequences, and ensuring long-term data reliability.

Description

Technical Field

[0001] This invention belongs to the field of smart grid condition monitoring technology, specifically relating to a fault prediction method and device for voltage monitoring instruments based on time-series prediction and generative models. Background Technology

[0002] Voltage monitoring instruments are core equipment for power quality monitoring in distribution networks. They are used to monitor key parameters such as the effective value of grid voltage and harmonic distortion rate in real time. They serve as the data foundation for source-grid-load-storage coordinated control, distributed energy grid-connected management, and active distribution network voltage optimization control. Their operational reliability directly affects the accuracy of statistical data such as voltage qualification rate. Unexpected failures can lead to the failure of source-grid-load-storage coordinated control, making it impossible to accurately implement active voltage regulation strategies, and even causing power fluctuations and islanding effects, posing a dual threat to the personal safety of maintenance personnel and the safety of equipment. With the deepening of smart grid construction, the deployment density and breadth of voltage monitoring instruments continue to increase, generating massive amounts of continuous, multi-dimensional time-series operational data.

[0003] Currently, the industry primarily employs two methods for managing the health status of voltage monitors: one is periodic on-site verification based on industry standards, typically with a verification cycle of several months or even a year; the other relies on in-house alarms with fixed thresholds, triggering an alarm when the monitored value exceeds a preset range. Periodic verification, based on discrete time-point sampling, cannot effectively monitor the gradual degradation processes of equipment components, such as performance aging and reference drift, during the intervals between verifications. This creates a hidden fault window, during which the voltage monitor's data may become distorted without being detected, significantly impacting the long-term reliability of statistical data. Fixed threshold alarms can only respond to significant events where the amplitude has clearly exceeded the limit, failing to detect subtle trends in equipment performance degradation.

[0004] The common drawback of the two methods mentioned above is that they can only detect significant equipment failures that have already occurred, lacking the ability to deeply analyze and mine the raw time-series data generated during continuous equipment operation, and failing to discern the health evolution trajectory of the equipment from subtle changes in the data. Specifically, existing technologies cannot predict the decline trend of equipment performance in advance, nor can they detect latent anomalies that have not yet manifested as functional failure but have deviated from the normal mode in a timely manner. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a voltage monitor fault prediction method and device based on time-series prediction and generative models. Its purpose is to overcome the limitation of existing voltage monitor health status management methods, which can only detect significant faults. By deeply analyzing and mining the raw time-series data generated during continuous operation, it can predict the performance degradation trend of the equipment in advance, promptly detect hidden anomalies that have deviated from normal modes but have not yet manifested as functional failures, and avoid problems such as the failure of source-grid-load-storage coordinated control, the inability to accurately implement active voltage regulation strategies, power fluctuations and islanding effects, and threats to the personal safety of maintenance personnel and equipment safety caused by unexpected voltage monitor failures. This ensures the long-term reliability of statistical data.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: According to a first aspect of the present invention, a method for predicting faults in a voltage monitoring instrument based on time-series prediction and generative models is provided, comprising: Obtain multi-dimensional health feature sequences from the voltage monitor; The multidimensional health feature sequence is input into a pre-trained quantile regression time series prediction model, which outputs multiple quantile prediction values ​​of the comprehensive health index of the voltage monitor within a preset future period. These multiple quantile prediction values ​​constitute the health prediction interval. The quantile regression time series prediction model uses a long short-term memory network with an attention mechanism as the backbone network and is trained and optimized using a quantile loss function as the objective function. The multi-dimensional health feature sequence is simultaneously input into a pre-trained variational autoencoder model to calculate the reconstruction error and latent space KL divergence of the current health feature segment. A comprehensive anomaly score is generated by fusing the reconstruction error and the latent space KL divergence. When the comprehensive anomaly score exceeds the anomaly threshold dynamically calculated based on historical normal data, a latent anomaly label is triggered. Based on the lower trend of the health prediction interval, the comprehensive anomaly score, and the latent anomaly markers, collaborative decision-making is carried out through a Bayesian fusion mechanism, and a corresponding level of early warning signal is generated according to the decision results.

[0007] In one possible implementation of the first aspect, acquiring the multi-dimensional health feature sequence of the voltage monitor includes: The raw time-series operating data of the voltage monitor is acquired, the raw time-series operating data is cleaned, missing values ​​are supplemented by linear interpolation, and all features are normalized. Direct monitoring sequences were extracted from the preprocessed data, including the three-phase voltage RMS sequence and the total harmonic distortion rate sequence. Statistical characteristics are calculated using a sliding time window, and the deviation characteristics between the device's measured values ​​and the standard reference values ​​are calculated to construct a derived performance sequence. The direct monitoring sequence and the derived performance sequence are aligned on the time axis to form the multidimensional health feature sequence.

[0008] In one possible implementation of the first aspect, the comprehensive health index is a scalar obtained by training a regressor by combining expert experience with historical data in an offline phase, used to quantify the overall health level of the voltage monitor, with a value range of [0,1].

[0009] In one possible implementation of the first aspect, the quantile loss function is defined as: For a single quantile The loss function is:

[0010] in, This represents the true health score. quantiles The corresponding predicted value; For multiple quantiles In multi-step prediction scenarios, the total loss function is the average of the sums of losses for all samples, all prediction times, and all quantiles:

[0011] in, For the sample size, To predict the step size, For the first The sample in the future The actual health value at any given moment. For the first The sample in the future quantiles at each time point q The corresponding predicted value.

[0012] In one possible implementation of the first aspect, the variational autoencoder model is trained using only health feature fragments from the historical normal operating state of the voltage monitor during the training phase. The variational autoencoder model parameters are optimized by maximizing the lower bound of evidence, the expression for which is:

[0013] in, For the input health feature fragment, For latent spatial variables, For decoder network parameters, For encoder network parameters, For the potential distribution of the encoder output, The reconstructed distribution of the decoder output, Based on the prior of the standard normal distribution, Let KL divergence be denoted as KL divergence.

[0014] In one possible implementation of the first aspect, the comprehensive anomaly score is calculated using a weighted fusion algorithm, with the specific calculation formula as follows:

[0015] in, This represents a fragment of current health characteristics. The reconstructed data output by the variational autoencoder decoder. λ The weighting coefficients have values ​​in the range [0,1]. This represents the potential distribution of the encoder's output for the current healthy feature fragment.

[0016] In one possible implementation of the first aspect, the anomaly threshold is determined by fitting the tail of the comprehensive anomaly score of historical normal data to a generalized Pareto distribution based on extreme value theory, specifically as follows: Using the generalized Pareto distribution Fit the score portion that exceeds the initial threshold, based on a preset significance level. ,by Calculate the dynamic anomaly threshold for the target cumulative probability. ,in, The amount by which an abnormal score exceeds the initial threshold. For shape parameters, This is the scale parameter.

[0017] In one possible implementation of the first aspect, the collaborative decision-making through a Bayesian fusion mechanism, and the generation of a corresponding level of early warning signal based on the decision result, includes: Set event For voltage monitoring instruments in the future Δ t A failure occurred within a certain time period, and the event... As the lower edge of the health prediction range continues to decline, the event... The overall anomaly score exceeds the anomaly threshold; Estimating prior failure rate based on historical data and conditional probability , ; When events are detected simultaneously and events When it occurs, calculate the posterior probability of failure:

[0018] When the posterior fault probability is greater than a preset high probability threshold, the highest level warning signal is generated; When only the event is detected or event When a single event occurs, calculate the corresponding marginal posterior probability. or Based on a preset low-probability threshold, a corresponding level of prompt or warning signal will be triggered.

[0019] In one possible implementation of the first aspect, the lower trend of the health prediction interval is calculated using the average value of a sliding time window, and the calculation formula is as follows:

[0020] in, As a health trend indicator, The length of the sliding time window. In time step The lower quantile predicted value at that point is the lower bound of the health prediction interval.

[0021] According to a second aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the aforementioned method for predicting voltage monitor faults based on timing prediction and generative models.

[0022] According to a third aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for predicting faults in a voltage monitor based on a timing prediction and generative model.

[0023] According to a fourth aspect of the present invention, a computer program product is provided, which, when executed by a processor, implements the aforementioned method for predicting voltage monitor faults based on timing prediction and generative models.

[0024] Compared with the prior art, the present invention has at least the following beneficial effects: This invention provides a fault prediction method for voltage monitors based on time-series prediction and generative models, employing an architecture that coordinates a quantile regression time-series prediction model and a variational autoencoder anomaly detection model. The quantile regression time-series prediction model, based on historical operating data of the equipment, outputs multiple quantile prediction values ​​of the comprehensive health index of the voltage monitor within a preset future period, forming a health prediction interval. The width of this interval adaptively reflects the model's prediction confidence: when the equipment is in a stable operating phase, historical data patterns are clear, and the prediction interval is narrow; when the equipment approaches a state transition critical point, data pattern complexity is high, and the prediction interval naturally widens, providing a quantitative basis for decision-making regarding uncertainty. The variational autoencoder anomaly detection model is trained only on historical normal data of the equipment, learning the data distribution patterns of normal operating modes. During online detection, it calculates the reconstruction error and latent spatial KL divergence of the current health feature segment and fuses them to generate a comprehensive anomaly score. When this score exceeds a dynamic threshold, a latent anomaly marker is triggered, thereby enabling the perception of subtle anomalies that have deviated from the normal mode but have not yet manifested as functional failure. Training is conducted only on operating data segments that have been verified as qualified through on-site inspection throughout the equipment's entire lifecycle, ensuring the model's accurate characterization of the normal mode boundaries. Unlike traditional supervised anomaly detection, which requires labeling positive and negative samples, this method avoids model bias problems caused by scarce fault samples and incomplete fault modes. Two models analyze equipment health status from two dimensions: trend prediction and anomaly detection. This allows for the prediction of equipment health degradation trajectories days to weeks in advance, and the accurate capture of unknown anomaly patterns that traditional methods cannot detect. This enables control over equipment risks at the source, providing early warnings before significant performance degradation or failure occurs.

[0025] This invention employs a quantile loss function to train a Long Short-Term Memory (LSTM) network with an attention mechanism. This enables the prediction model to output not only point predictions of the health index but also predictions of multiple quantiles, forming a prediction interval. The width of the prediction interval directly reflects the degree of uncertainty in the prediction result. Simultaneously, the reconstruction error generated by the variational autoencoder and the KL divergence are weighted and fused to form a continuously quantified comprehensive anomaly score. The magnitude of this score reflects the degree to which the current equipment operating status deviates from the normal mode. The uncertainty information of the prediction interval and the continuously quantified information of the comprehensive anomaly score provide decision-making basis for maintenance personnel, achieving a refined matching between maintenance decisions and risk levels, avoiding the limitation of traditional methods that can only provide a binary judgment of normal or abnormal. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a voltage monitoring instrument fault prediction method based on time-series prediction and generative models, as described in an embodiment of the present invention.

[0028] Figure 2 This is an overall architecture diagram of the voltage monitoring instrument fault prediction device based on time-series prediction and generative model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the quantile regression time series prediction model structure used in this invention; Figure 4 This is a schematic diagram of the training and anomaly detection structure of the variational autoencoder model used in this invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] like Figure 1 As shown, this invention provides a method for predicting voltage monitoring instrument faults based on time-series prediction and generative models, specifically including the following steps: S1. Obtain the multi-dimensional health feature sequence of the voltage monitor.

[0031] It should be noted that the multi-dimensional health feature sequence is obtained by preprocessing and feature extraction of the original time-series operating data of the voltage monitor. It includes various quantitative indicators that can reflect the health status of the voltage monitor itself, such as the effective value of three-phase voltage, total harmonic distortion rate, sliding time window statistics, and measurement error.

[0032] S2. Input the multi-dimensional health feature sequence into the pre-trained quantile regression time series prediction model, and output multiple quantile prediction values ​​of the comprehensive health index of the voltage monitor within a preset period in the future. The multiple quantile prediction values ​​constitute the health prediction interval. The quantile regression time series prediction model uses a long short-term memory network with an attention mechanism as the backbone network and uses the quantile loss function as the objective function for training and optimization.

[0033] In other words, after receiving input, the quantile regression time-series prediction model outputs multiple quantile predictions of the comprehensive health index from the voltage monitor within a preset future period. These multiple quantile predictions together constitute the health prediction interval. The width of the health prediction interval directly reflects the degree of uncertainty in the prediction results of the quantile regression time-series prediction model.

[0034] S3. Simultaneously input the multi-dimensional health feature sequence into the pre-trained variational autoencoder model, calculate the reconstruction error and latent space KL divergence of the current health feature segment, and generate a comprehensive anomaly score based on the reconstruction error and the latent space KL divergence; when the comprehensive anomaly score exceeds the anomaly threshold dynamically calculated based on historical normal data, trigger the latent anomaly labeling.

[0035] In other words, the variational autoencoder model analyzes the current health feature segments, calculates their reconstruction error in the feature space and KL divergence in the latent space, and generates a comprehensive anomaly score based on the fusion of these two indicators. When the calculated comprehensive anomaly score exceeds an anomaly threshold dynamically calculated based on historical normal data, the system triggers a latent anomaly flag, indicating that the current operating state of the device has deviated from its historical normal mode.

[0036] S4. Based on the lower trend of the health prediction interval, the comprehensive anomaly score and the latent anomaly marker, a collaborative decision is made through a Bayesian fusion mechanism, and a warning signal of the corresponding level is generated according to the decision result.

[0037] In other words, the Bayesian fusion mechanism takes into account the outputs of both trend prediction and anomaly detection, and generates corresponding warning signals based on the decision results, such as prompts, warnings, or the highest level alarm.

[0038] This implementation combines trend prediction and anomaly detection models to predict the future evolution of equipment health in advance and to sense in real time whether the current operating status deviates from the normal mode. This enables early warning and accurate identification of voltage monitor faults and solves the shortcomings of existing technologies that cannot deeply mine equipment health information from time series data.

[0039] In one possible implementation, the acquisition of the multi-dimensional health feature sequence of the voltage monitor is specifically implemented as follows: S11. Obtain the raw timing data of the voltage monitor, and perform data cleaning on the raw timing data to remove obviously abnormal jump points. If there are missing values ​​in the data, use linear interpolation to supplement them. The interpolation formula is as follows:

[0040] in, for t Missing values ​​at time points and These are the known values ​​at two different times.

[0041] After data supplementation, all features are normalized to scale their numerical range to the [0,1] interval. The normalization formula is as follows:

[0042] in, These are the original eigenvalues. and These are the minimum and maximum values ​​of the feature, respectively. These are the normalized eigenvalues.

[0043] S12. Extract the direct monitoring sequence from the preprocessed data, mainly including the three-phase voltage RMS sequence and the total harmonic distortion (THD) sequence. The calculation formula for the three-phase voltage RMS sequence is:

[0044] in, for t time i Phase voltage RMS value for i Phase voltage in one cycle k One sampling point, N The number of sampling points within one period. This represents the voltage of phase A, phase B, and phase C.

[0045] The formula for calculating the total harmonic distortion rate sequence is:

[0046] in, for t Total harmonic distortion at time t, This is the effective value of the fundamental voltage. For the first h RMS value of subharmonic voltage H The highest harmonic order under consideration.

[0047] S13. Calculate statistical characteristics through a sliding time window, and calculate the deviation characteristics between the equipment measurement value and the standard reference value to construct a derived performance sequence.

[0048] Specifically, statistical characteristics are calculated using a sliding time window to construct a derived performance series. The formulas for calculating the mean and standard deviation of the sliding time window are as follows:

[0049] in, The mean of the data within the window. The standard deviation of the data within the window. The length of the sliding time window. For the first in the window i Data points.

[0050] Specifically, the deviation characteristic between the measured value of the calculation device and the standard reference value is calculated using the following formula:

[0051] in, for t Measurement error of time, For equipment measurements, This is the standard reference value.

[0052] S14. Align the direct monitoring sequence and the derived performance sequence on the time axis to form the multidimensional health feature sequence.

[0053] This implementation method transforms raw monitoring data into a multi-dimensional feature sequence that can effectively characterize the health status of equipment through systematic data preprocessing and feature extraction, ensuring the reliability of input data for prediction and detection models and making the model analysis results more accurate and reliable.

[0054] In one possible implementation, the comprehensive health index is a scalar obtained by training a regressor through an offline phase that combines expert experience with historical data. Specifically, the construction process of the comprehensive health index is as follows: First, historical operational data throughout the voltage monitor's entire lifecycle is collected, including various samples such as normal operation phases, performance degradation processes, and pre-fault states. Simultaneously, based on expert experience, corresponding weights are assigned to each verification error indicator. The regressor learns the mapping relationship between multi-dimensional health features and the actual health status of the equipment, compressing the multi-dimensional feature space into a one-dimensional scalar.

[0055] The comprehensive health index is used to quantify the overall health level of the voltage monitor, and its value range is set to the interval [0,1]. A value of 1 indicates that the device is in a fully healthy state, with all performance indicators at their optimal level; a value of 0 indicates that the device is completely faulty and cannot function properly. Within the [0,1] interval, a higher value indicates a better health status, while a lower value indicates a more severe performance degradation.

[0056] This implementation method simplifies complex multidimensional health characteristics into an intuitive quantitative indicator by defining a comprehensive health index, making it easier for maintenance personnel to quickly grasp the overall health status of the equipment.

[0057] In one feasible approach, such as Figure 3 The diagram shows a Long Short-Term Memory network with an attention mechanism as the backbone network of the prediction model.

[0058] The core of the attention mechanism lies in calculating the contribution weights of different time steps or different feature dimensions to the current prediction task. Let's assume the LSTM network at time step... i The hidden state is Current time step t The hidden state is Then attention weight The calculation formula is:

[0059] in, The attention scoring function can take the form of dot product, additive attention, etc. By using attention weights to sum the hidden states at each time step, the model can focus on the historical information most critical to predicting future health.

[0060] The quantile loss function is used for training, aiming to establish a mapping from historical multivariate time-series features of devices to future health index quantiles, thereby quantifying the uncertainty of prediction. Let the input historical multivariate time-series features be... ,in L Given the historical time window length, the model outputs the future. T Health quantile predicted value at each time point For a given quantile The goal of model learning is to make:

[0061] in, For the future The actual health value at any given moment. For parameters The model represents the mapping function, and this probability equation shows that the quantiles predicted by the model are consistent with the quantiles of the true distribution.

[0062] For a single quantile quantile loss function Defined as:

[0063] in, This represents the true health score. quantiles The corresponding predicted value. This loss function assigns different weights to positive and negative errors, enabling the model to learn the predictive ability of specific quantiles.

[0064] In practical applications, it is necessary to predict multiple quantiles to form a prediction interval. Let the set of quantiles be... Meanwhile, considering multi-step prediction scenarios (predicting the future) T (at each time step), the total loss function is defined as the average of the sum of the losses for all training samples, all prediction times, and all quantiles:

[0065] in, For the sample size, To predict the step size, For the first The sample in the future The actual health value at any given moment. For the first The sample in the future quantiles at each time point q The corresponding predicted value.

[0066] Minimize the total loss using the backpropagation algorithm. The model parameters are optimized, and it can ultimately output health prediction values ​​corresponding to multiple quantiles for each future time point, thus providing a prediction interval for each future time point. The width of the prediction interval directly reflects the degree of uncertainty in the model's prediction at that point; the wider the interval, the greater the uncertainty of the prediction result.

[0067] This implementation introduces an attention mechanism and a quantile loss function, enabling the prediction model to not only target key historical information but also output confidence intervals for health prediction values, thus quantifying the uncertainty of the prediction process.

[0068] In one feasible approach, such as Figure 4 As shown on the left, the training phase of the variational autoencoder model only uses health feature fragments from the voltage monitor's historical normal operating conditions.

[0069] A variational autoencoder contains an encoder and a decoder ,in, For the input health feature fragment, For latent spatial variables, For encoder network parameters, Here are the decoder network parameters. The training objective is to optimize the model parameters by maximizing the lower bound of evidence, which is expressed as:

[0070] Among them, the first item The reconstruction loss measures the decoder's understanding of latent space variables. Reconstruct the original input The accuracy; the second item The KL divergence constrains the latent distribution of the encoder output. Approximating the prior distribution Prior distribution The model is set to a standard normal distribution. By maximizing the lower bound of evidence, the model learns optimal parameters that can reconstruct normal data with high fidelity.

[0071] This implementation method trains the variational autoencoder model using only normal data, enabling it to learn the data distribution patterns under normal operating conditions of the voltage monitor, thus establishing an accurate digital benchmark for anomaly detection.

[0072] In one feasible approach, such as Figure 4 As shown on the right, during online detection, for new health feature fragments... Then, input it into the trained variational autoencoder model.

[0073] First, the latent distribution corresponding to the current health feature fragment is obtained through the encoder. Then, samples are taken from this distribution and reconstructed by the decoder to obtain... The comprehensive anomaly score is calculated using a weighted fusion algorithm. The specific calculation formula is as follows:

[0074] in, This represents a fragment of current health characteristics. The reconstructed data output by the variational autoencoder decoder. λ This is a weighting coefficient with a value range of [0,1], used to balance the contributions of the two indicators. The first term... The deterministic reconstruction error measures the difference between the input data and the reconstructed data in the original feature space; the second term... It measures the degree of deviation between the latent distribution corresponding to the current health characteristic segment and the prior distribution (standard normal distribution).

[0075] This implementation method integrates two metrics, reconstruction error and KL divergence, to comprehensively measure the degree of anomaly in the feature space and latent space of the data. It is more robust than using reconstruction error alone and can more sensitively capture various types of anomaly patterns.

[0076] In one feasible approach, the anomaly threshold is determined by fitting the tail of the comprehensive anomaly score from historical normal data using a generalized Pareto distribution based on extreme value theory. The specific process is as follows: First, comprehensive anomaly score data from the voltage monitor's historical normal operating conditions is collected using a sliding time window. An initial threshold is set, and only scores exceeding this initial threshold are retained as excess values. Then, the generalized Pareto distribution is used to fit these excess values. The expression for the generalized Pareto distribution is:

[0077] in, The amount by which an abnormal score exceeds the initial threshold. The shape parameter determines the thickness of the distribution tail. As a scale parameter, it controls the degree of diffusion in the distribution, affecting fluctuations beyond the range of the quantity, and >0.

[0078] Based on the preset significance level ,by Calculate the dynamic anomaly threshold for the target cumulative probability:

[0079] in, This is the final determined dynamic anomaly threshold. When the overall anomaly score obtained from online detection exceeds this threshold, the system triggers a latent anomaly marker.

[0080] This implementation collects abnormal rating data through a sliding time window and uses extreme value theory to fit the tail of the rating distribution, making the setting of the abnormal threshold statistically rigorous and adaptive. It can be dynamically adjusted according to the distribution characteristics of historical data, which avoids the problem of too many false alarms caused by fixed thresholds and ensures sensitivity to rare abnormal events.

[0081] In one feasible approach, two key features need to be extracted before decision fusion: a health trend indicator and an anomaly score sequence. The anomaly score sequence is defined as the comprehensive anomaly score at each time step within a recent time window.

[0082] in, It is an anomaly scoring sequence, containing the anomaly scores for each time step within a recent period; This is the width of the sliding time window, the size of the window used for statistical anomaly scoring; For time step i The comprehensive anomaly score is calculated by the variational autoencoder model.

[0083] Define related events: Set the event For voltage monitoring instruments in the future Δ t A failure occurred within a certain time period, and the event... As the lower edge of the health prediction range continues to decline, the event... The overall anomaly score exceeds the anomaly threshold; Based on historical operational data, the prior failure rate can be estimated. and conditional probability , .

[0084] When the system detects events simultaneously and events When it occurs, calculate the posterior probability of failure:

[0085] When the calculated posterior failure probability is greater than the preset high probability threshold, the system generates the highest level warning signal, prompting maintenance personnel to take immediate intervention measures.

[0086] When the system only detects the event or event When a single event occurs, calculate the corresponding marginal posterior probability. or It triggers corresponding levels of prompts or warning signals based on preset low probability thresholds. For example, if only the lower edge of the prediction interval continues to decline but the abnormal score does not exceed the standard, a "pay attention" level prompt can be generated; if only the abnormal score briefly exceeds the standard but the prediction trend is stable, a "check" level prompt can be generated.

[0087] This implementation method integrates the output results of two models through a Bayesian fusion mechanism, achieving more stable and refined decision-making, avoiding inappropriate alarms caused by misjudgment from a single model, and providing differentiated response strategies for operation and maintenance personnel through a hierarchical early warning mechanism.

[0088] In one possible implementation, the lower trend of the health prediction interval is calculated using the average value of a sliding time window, as follows:

[0089] in, This is a health trend indicator, reflecting the recent trend of changes in equipment health. The sliding time window length is the number of time steps used to calculate the trend. The specific value can be set according to the actual application scenario, for example, taking... =5 or =10; In time step The lower quantile predicted value, i.e. the lower bound of the health prediction interval, is usually the 10th percentile or the 5th percentile.

[0090] By calculating the average of the lower bounds of the prediction interval across multiple historical moments, short-term fluctuations can be smoothed out, providing a clearer reflection of the overall trend in health status. When the score shows a continuous downward trend, it indicates that the health status of the equipment is deteriorating rapidly. Even if the current abnormal score has not exceeded the standard, it should still be a cause for concern.

[0091] This implementation calculates the health trend index by averaging over a sliding time window, providing a quantitative basis for event judgment for the Bayesian fusion mechanism. This transforms the continuously decreasing lower edge into a calculable numerical index, making the decision-making process more objective.

[0092] Based on the above implementation methods, this invention, by integrating the predictive maintenance capabilities of a quantile regression time-series prediction model with the latent fault identification capabilities of a variational autoencoder generative model, can both predict equipment health degradation trajectories days to weeks in advance, avoiding unplanned downtime, and accurately capture subtle and unknown anomalies that traditional methods cannot detect, thus controlling equipment risks at the source. Compared to traditional methods, this invention possesses latent fault detection capabilities, enabling timely warnings before faults occur, thereby achieving earlier intervention and reducing system damage.

[0093] This invention employs a quantile loss function as its technical support, enabling prediction results to inherently include confidence intervals. Simultaneously, through continuously quantified anomaly scores, it clearly defines the degree to which equipment deviates from its normal state, achieving refined matching of maintenance decisions with risk levels. Furthermore, this invention adopts a fully data-driven approach, eliminating the need for manually defining complex fault rules or thresholds. The model can autonomously learn the normal operating patterns and health evolution trajectories of the equipment from the historical operating data of the voltage monitor. Faced with newly emerging fault modes or maintenance needs of different types of equipment, it can be quickly adapted through retraining or model fine-tuning, without the need to redevelop the entire system. The self-learning and adaptive capabilities make this method universally applicable and scalable in the maintenance of different scenarios and types of equipment, reducing system deployment and long-term maintenance costs.

[0094] This invention uses a software algorithm as its core architecture and can be directly deployed on existing cloud platforms or edge computing nodes. Its input data can directly reuse the raw time-series operational data and historical calibration records uploaded daily by the voltage monitor, without requiring additional sensors or data acquisition equipment, nor any modification to existing data transmission channels and hardware. This lightweight deployment approach allows this method to be quickly integrated into existing operation and maintenance systems at a low cost, achieving intelligent management of the voltage monitor's health status without increasing hardware investment.

[0095] like Figure 2As shown, this invention provides a voltage monitoring instrument fault prediction device based on time-series prediction and generative models, including a data acquisition and processing module, a server analysis module, and a human-computer interaction module.

[0096] The data acquisition and processing module is configured to acquire multi-dimensional health feature sequences of the voltage monitor.

[0097] The server analysis module is configured to: input the multi-dimensional health feature sequence into a pre-trained quantile regression time-series prediction model, and output multiple quantile prediction values ​​of the comprehensive health index of the voltage monitor within a preset future period, wherein the multiple quantile prediction values ​​constitute a health prediction interval; wherein the quantile regression time-series prediction model uses a long short-term memory network with an attention mechanism as the backbone network and uses a quantile loss function as the objective function for training and optimization; and simultaneously input the multi-dimensional health feature sequence into a pre-trained variational autoencoder model, calculate the reconstruction error and latent space KL divergence of the current health feature segment, and generate a comprehensive anomaly score based on the fusion of the reconstruction error and the latent space KL divergence; when the comprehensive anomaly score exceeds an anomaly threshold dynamically calculated based on historical normal data, a latent anomaly label is triggered; The human-computer interaction module is configured to: make collaborative decisions based on the lower trend of the health prediction interval, the comprehensive abnormal score and the latent abnormal marker, and generate a warning signal of the corresponding level according to the decision results.

[0098] All relevant content regarding the steps involved in the aforementioned embodiment of a voltage monitor fault prediction method based on time-series prediction and generative models can be referenced from the functional description of the corresponding functional module of a voltage monitor fault prediction device based on time-series prediction and generative models in this invention embodiment, and will not be repeated here. The module division in this embodiment is illustrative and only represents a logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this invention can be integrated into a processor, exist as separate physical entities, or have two or more modules integrated into one module. The integrated modules can be implemented in hardware or as software functional modules.

[0099] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a voltage monitoring instrument fault prediction method based on timing prediction and generative models.

[0100] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the voltage monitoring instrument fault prediction method based on timing prediction and generative models in the above embodiments.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] This invention also provides a computer program product, which is used to execute any of the above-described voltage monitoring instrument fault prediction methods based on time-series prediction and generative models. Since the computer program product provided by this invention and the voltage monitoring instrument fault prediction method based on time-series prediction and generative models described above belong to the same inventive concept, the computer program product provided by this invention possesses all the advantages of the voltage monitoring instrument fault prediction method based on time-series prediction and generative models described above. Therefore, the beneficial effects of the computer program product provided by this invention will not be elaborated upon here.

[0106] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0107] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A time-series prediction and generative model based voltage monitor failure prediction method, characterized in that, include: Obtain multi-dimensional health feature sequences from the voltage monitor; The multidimensional health feature sequence is input into a pre-trained quantile regression time series prediction model, which outputs multiple quantile prediction values ​​of the comprehensive health index of the voltage monitor within a preset future period. These multiple quantile prediction values ​​constitute the health prediction interval. The quantile regression time series prediction model uses a long short-term memory network with an attention mechanism as the backbone network and is trained and optimized using a quantile loss function as the objective function. The multi-dimensional health feature sequence is simultaneously input into a pre-trained variational autoencoder model to calculate the reconstruction error and latent space KL divergence of the current health feature segment. A comprehensive anomaly score is generated by fusing the reconstruction error and the latent space KL divergence. When the comprehensive anomaly score exceeds the anomaly threshold dynamically calculated based on historical normal data, a latent anomaly label is triggered. Based on the lower trend of the health prediction interval, the comprehensive anomaly score, and the latent anomaly markers, collaborative decision-making is carried out through a Bayesian fusion mechanism, and a corresponding level of early warning signal is generated according to the decision results.

2. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The multi-dimensional health feature sequence of the voltage monitor includes: The raw time-series operating data of the voltage monitor is acquired, the raw time-series operating data is cleaned, missing values ​​are supplemented by linear interpolation, and all features are normalized. The direct monitoring sequences, including the three-phase voltage RMS sequence and the total harmonic distortion rate sequence, are extracted from the preprocessed data. Statistical characteristics are calculated using a sliding time window, and the deviation characteristics between the device's measured values ​​and the standard reference values ​​are calculated to construct a derived performance sequence. The direct monitoring sequence and the derived performance sequence are aligned on the time axis to form the multidimensional health feature sequence.

3. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The comprehensive health index is a scalar obtained by training a regressor by combining expert experience with historical data in an offline phase. It is used to quantify the overall health level of the voltage monitor and its value ranges from [0,1].

4. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The quantile loss function is defined as follows: For a single quantile The loss function is wherein, is the healthiness true value, is the quantile is the corresponding predicted value; For multiple quantiles And multi-step prediction scenarios, the total loss function is the average of the sum of the losses for all samples, all prediction times, and all quantiles: in, For the sample size, To predict the step size, For the first The sample in the future The actual health value at any given moment. For the first The sample in the future quantiles at each time point q The corresponding predicted value.

5. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The variational autoencoder model is trained using only health feature segments from the historical normal operating state of the voltage monitor during the training phase. The model parameters are optimized by maximizing the lower bound of evidence, which is expressed as follows: in, For the input health feature fragment, For latent spatial variables, For decoder network parameters, For encoder network parameters, For the potential distribution of the encoder output, The reconstructed distribution of the decoder output, Given a standard normal distribution prior, Let KL divergence be denoted as KL divergence.

6. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 5, characterized in that, The comprehensive anomaly score is calculated using a weighted fusion algorithm, and the specific calculation formula is as follows: in, This represents a fragment of current health characteristics. The reconstructed data output by the variational autoencoder decoder. λ The weighting coefficients have values ​​in the range [0,1]. This represents the potential distribution of the encoder's output for the current healthy feature fragment.

7. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The anomaly threshold is determined by fitting the tail of the comprehensive anomaly score of historical normal data to the generalized Pareto distribution in extreme value theory. Specifically: Using the generalized Pareto distribution Fit the score portion that exceeds the initial threshold, based on a preset significance level. ,by Calculate the dynamic anomaly threshold for the target cumulative probability. ,in, The amount by which an abnormal score exceeds the initial threshold. For shape parameters, This is the scale parameter.

8. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The collaborative decision-making process using a Bayesian fusion mechanism, and the generation of corresponding level early warning signals based on the decision results, includes: Set event For voltage monitoring instruments in the future Δ t A failure occurred within a certain time period, and the event... As the lower edge of the health prediction range continues to decline, the event... The overall anomaly score exceeds the anomaly threshold; Estimating prior failure rate based on historical data and conditional probability , ; When events are detected simultaneously and events When it occurs, calculate the posterior probability of failure: When the posterior fault probability is greater than a preset high probability threshold, the highest level warning signal is generated; When only the event is detected or event When a single event occurs, calculate the corresponding marginal posterior probability. or Based on a preset low-probability threshold, a corresponding level of prompt or warning signal will be triggered.

9. The voltage monitoring instrument fault prediction method based on time-series prediction and generative model according to claim 1, characterized in that, The lower trend of the health prediction interval is calculated using the average value of a sliding time window, and the calculation formula is as follows: in, As a health trend indicator, The length of the sliding time window. In time step The lower quantile predicted value at that point is the lower bound of the health prediction interval.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a voltage monitoring instrument fault prediction method based on timing prediction and generative models as described in any one of claims 1 to 9.

Images