Power equipment health evaluation method and system based on time sequence analysis and probability modeling, equipment and medium

Abstract

The invention discloses a power equipment health evaluation method and system based on time sequence analysis and probability modeling, equipment and a medium, and relates to the technical field of power equipment state monitoring and fault prediction, and the method comprises the steps: obtaining and preprocessing multi-source operation data of power equipment, outputting a predicted value and a confidence interval of a future parameter through a time sequence prediction model, calculating a residual sequence of an actual observation value and a predicted value, fitting distribution through a probability distribution model, establishing a statistical characteristic model of a normal operation state of the equipment, performing anomaly judgment, calculating a health degree index of the equipment based on a deviation degree and a dynamic weight of a monitoring parameter and weighted accumulation, and dividing equipment state grades according to the index. Quantitative evaluation of the health state of the equipment is realized. According to the method, accurate quantification and early abnormity identification of the health state of the power equipment are realized, a reliable basis is provided for predictive maintenance, and the intelligent level and the safety guarantee capability of power grid operation and maintenance are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring and fault prediction, and in particular to a power equipment health evaluation method, system, device and medium based on time series analysis and probability modeling. BACKGROUND

[0002] As a core component of the power system, the reliability and stability of the operation state of the power equipment are directly related to the safe operation of the power grid; with the in-depth promotion of the construction of the smart grid, the equipment health evaluation technology based on data driving has gradually become a research hotspot; the traditional method mainly relies on periodic maintenance and threshold alarm, but such method is difficult to capture the dynamic evolution characteristics of the equipment state, and is easily affected by environmental interference and sensor noise.

[0003] In recent years, the time series analysis technology provides a new idea for equipment state prediction by mining the potential law of historical operation data; the model based on the recurrent neural network and the long short-term memory network can effectively capture the long-term dependence relationship in the time series data, but the prediction result is usually in the form of point estimation, and lacks the quantification ability of uncertainty; at the same time, the probability modeling method such as Gaussian mixture model performs well in the field of anomaly detection, but its application in multi-dimensional time series data still faces the balance problem between distribution fitting accuracy and computational complexity. The existing technology generally has two key defects: one is that the uncertainty information of time series prediction is not fully fused, which makes it difficult to quantify the confidence of health evaluation; the other is that the static weight allocation mechanism cannot adapt to the dynamic change of the equipment operation state, which reduces the accuracy of the evaluation result. SUMMARY

[0004] In view of the above existing problems, the present application provides a power equipment health evaluation method, system, device and medium based on time series analysis and probability modeling, to solve the problems that the uncertainty information of time series prediction is not fully fused in the prior art, which makes it difficult to quantify the confidence of health evaluation, and the static weight allocation mechanism cannot adapt to the dynamic change of the equipment operation state, which reduces the accuracy of the evaluation result.

[0005] To solve the above technical problems, a power equipment health evaluation method based on time series analysis and probability modeling is proposed, which comprises,

[0006] The multi-source operation data of the power equipment is acquired, the multi-source operation data is preprocessed, the multi-dimensional time series data after preprocessing is modeled by using a time series prediction model, and a prediction value and a confidence interval of future operation parameters are output; a residual sequence between an actual observation value and the prediction value is calculated, a probability distribution model is used to fit the distribution of the residual sequence, a statistical characteristic model under a normal operation state of the equipment is established, and abnormality is determined according to the probability density of the residual points based on the probability distribution model obtained by fitting; a health index of the equipment is calculated by weighted accumulation based on the deviation degree and the weight of each monitoring parameter, and the state grade of the equipment is divided according to the health index.

[0007] As a preferred scheme of the power equipment health evaluation method based on time series analysis and probability modeling, the preprocessing of the multi-source operation data comprises data standardization to eliminate dimensional differences, interpolation to complete missing values, and abnormal value elimination by using an abnormality detection algorithm.

[0008] As a preferred scheme of the power equipment health evaluation method based on time series analysis and probability modeling, the modeling of the multi-dimensional time series data after preprocessing comprises selective memory and forgetting of historical information by using an internal gating mechanism to capture long-term dependence in the time series data.

[0009] In the model inference stage, an uncertainty quantification mechanism is introduced, a distribution of a prediction result is generated by forward propagation sampling, a prediction mean value is output, a prediction variance is calculated, and a prediction interval with a confidence level is formed.

[0010] As a preferred scheme of the power equipment health evaluation method based on time series analysis and probability modeling, the fitting of the distribution of the residual sequence comprises fitting of the residual distribution by using a multi-component distribution combination to depict potential multi-peak and asymmetric characteristics.

[0011] The optimal number of distribution components is determined by selecting among candidate models by using an information criterion, and a balance between model fitting accuracy and complexity is achieved.

[0012] As a preferred scheme of the power equipment health evaluation method based on time series analysis and probability modeling, the probability distribution model comprises a Gaussian mixture model, an expectation maximization algorithm is used for iterative optimization, a posterior probability of each residual sample belonging to each Gaussian component in the model is calculated based on a current parameter estimate value, and the weight, mean value and variance parameters of each Gaussian component in the model are re-estimated by using the calculated posterior probability until the log-likelihood function increment of adjacent two iterations is less than a preset threshold value, and the Gaussian mixture model parameters are obtained.

[0013] The residual probability distribution is represented as:

[0014]

[0015] wherein, is the residual is the overall probability density under the Gaussian mixture model, is the residual, i.e., the difference between the observed value of the equipment operating parameter and the predicted value of the model, is the total number of Gaussian components in the Gaussian mixture model, is the mixing weight of the kth Gaussian component, is the probability density function of the kth Gaussian component, is the mean of the kth Gaussian component, i.e., the center position of the current component distribution, is the variance of the kth Gaussian component, i.e., the degree of dispersion of the current component distribution;

[0016] The total number of Gaussian components in the Gaussian mixture model is determined using the Bayesian information criterion, which is expressed by the formula:

[0017]

[0018] wherein, is the likelihood function under the given model parameters, is the natural logarithm of the likelihood function, is the number of free parameters of the model, i.e., the mean, variance, and mixing weight parameters of each Gaussian component, is the total number of training samples.

[0019] As a preferred scheme of the power equipment health evaluation method based on time series analysis and probability modeling, the abnormality determination includes presetting a probability density threshold value 95% matching the confidence level, and when the probability density value of a certain residual sample point under the Gaussian mixture model is lower than the threshold value 95%, it is determined that the current equipment operating state is abnormal.

[0020] As a preferred scheme of the power equipment health evaluation method based on time series analysis and probability modeling, the calculation of the health index of the equipment includes defining the deviation measure of each monitoring parameter, which is composed of the ratio of the absolute value of the residual to the standard deviation of the prediction confidence interval, and dynamically determining the weight of each monitoring parameter using the entropy weight method, multiplying the deviation measure of each parameter by the matching weight, and accumulating, deducting from the baseline score, mapping to the interval of 0 to 100, and obtaining the health index.

[0021] The weight of each monitoring parameter is dynamically determined by using the entropy weight method, including range standardization of a residual sequence of each parameter, calculation of a normalized proportion of each parameter in all samples, calculation of information entropy of each parameter based on the normalized proportion, calculation of a difference coefficient of each parameter according to the information entropy, and normalization processing of the difference coefficient to obtain the final weight of each parameter.

[0022] The health index formula is obtained as follows:

[0023]

[0024]

[0025] wherein, is the health index of the equipment, m is the number of key monitoring parameters participating in the health calculation, i is the index of the monitoring parameter, is the deviation measure of the i th monitoring parameter, is the weight of the i th monitoring parameter in the health calculation, is the residual of the i th monitoring parameter at time t, is the standard deviation of the confidence interval of the predicted value of the i th monitoring parameter at time t.

[0026] The division of the equipment state level includes that when the health index is , the health state is determined, when , the sub-health state is determined, when , the early warning state is determined, and when , the high-risk state is determined.

[0027] The preferred technical scheme has the beneficial effects that by constructing a dynamic weight distribution mechanism based on the entropy weight method and a deviation measure formula considering the confidence, the objective and comprehensive evaluation of the equipment health state is realized, the entropy weight method automatically determines the weight according to the variation degree of the monitoring parameter data itself, avoids the deviation of subjective weighting, and the standard deviation of the prediction confidence interval is included in the design of the deviation measure, so that the evaluation result reflects the parameter deviation degree and the model confidence level.

[0028] As a preferred scheme of the power equipment health evaluation system based on time series analysis and probability modeling, the system comprises a multi-source data preprocessing module, a time series dynamic prediction module, a residual probability modeling module, a health quantification evaluation module and an intelligent diagnosis and decision support module.

[0029] The multi-source data preprocessing module is configured to eliminate dimensional differences of different physical parameters through standardization, adopt a bucket standardization strategy combined with time characteristics, process data missing by using an autoregressive completion method, and remove abnormal points introduced by interference by using a Hampel filter, and output a standardized data set.

[0030] The time series dynamic prediction module is configured to use a long short-term memory network to perform deep modeling on the preprocessed multi-dimensional time series data, learn and remember long-term dependencies in historical data through a gating mechanism, predict future periodic operation parameters of the equipment, and output a prediction mean value by performing random forward propagation sampling in an inference stage through an integrated Monte Carlo Dropout technology, calculate a prediction variance, and form a prediction interval.

[0031] The residual probability modeling module is configured to calculate a residual sequence between an actual observation value and a prediction value, fit a probability distribution of the sequence by using a Gaussian mixture model, and determine an optimal number of components according to a Bayesian information criterion.

[0032] The health quantification evaluation module is configured to fuse deviation information of monitoring parameters, calculate a health index, calculate a dynamic weight based on a data dispersion degree of each parameter residual sequence by using an entropy weight method, obtain a final health score by weighted accumulation, and divide a device state into four levels of health, sub-health, early warning, and high risk.

[0033] The intelligent diagnosis and decision support module is configured to integrate abnormal detection results and health scores, generate visual output, provide a global perception of a full life cycle operation state of the equipment for an operation and maintenance personnel, discover potential risks in a timely manner, and develop an operation and maintenance strategy according to a state level.

[0034] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements steps of a method for power equipment health evaluation based on time series analysis and probability modeling when executing the computer program.

[0035] A computer readable storage medium stores a computer program, and the computer program implements steps of a method for power equipment health evaluation based on time series analysis and probability modeling when executed by a processor.

[0036] The application has the beneficial effects that: the application eliminates dimension difference, fills in data loss and eliminates abnormal values through data preprocessing, provides high-quality data basis for subsequent models, and significantly improves the robustness of overall evaluation; in time series prediction modeling, an uncertainty quantification mechanism is introduced to output prediction values and provide a prediction interval with a confidence level, which provides a scientific and adaptive dynamic benchmark for abnormality determination; a Gaussian mixture model is used to fit the residual distribution, the model complexity is optimized by information criterion, the multi-peak and asymmetric statistical characteristics of the device in normal operation state are accurately described, and the false positive and false negative rates of abnormality determination are reduced; and by combining the deviation measurement of the prediction confidence interval and the dynamic weight distribution based on the entropy weight method, the health index calculated can objectively and sensitively reflect the overall health status of the device, and through state grade division, clear decision basis from "abnormality detection" to "health management" is provided for operation and maintenance personnel, and the purpose of evaluating the health status of the device in a forward-looking and high-confidence manner is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0038] Figure 1 The overall flowchart of the power equipment health evaluation method based on time series analysis and probability modeling provided by an embodiment of the application is shown.

[0039] Figure 2 The system scheme flowchart of the power equipment health evaluation system based on time series analysis and probability modeling provided by an embodiment of the application is shown. DETAILED DESCRIPTION

[0040] In order to make the above-mentioned purposes, features and advantages of the application more apparent and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are only part of the embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor should be within the protection scope of the application.

[0041] Embodiment 1, refer to Figure 1 For an embodiment of the application, a power equipment health evaluation method based on time series analysis and probability modeling is provided, which comprises:

[0042] S100: Obtain multi-source operation data of the power equipment, pre-process the multi-source operation data, model the pre-processed multi-dimensional time series data by using a time series prediction model, and output a predicted value and a confidence interval of a future operation parameter.

[0043] S200: Calculate a residual sequence between an actual observation value and a predicted value, fit a distribution of the residual sequence by using a probability distribution model, establish a statistical characteristic model under a normal operation state of the equipment, and perform abnormality determination according to a probability density of a residual point based on the probability distribution model obtained by fitting.

[0044] S300: Calculate a health index of the equipment by weighted accumulation based on a deviation degree and a weight of each monitoring parameter, and divide a state level of the equipment according to the health index.

[0045] It should be noted that, in the data pre-processing stage, the robust standardization and intelligent interpolation method are used to effectively eliminate the dimensional difference and abnormal interference of the multi-source data, ensure the high quality and consistency of the input data, and lay a reliable foundation for subsequent analysis; in the trend prediction stage, the time series model with a gating mechanism is used to capture long-term dependence, and the uncertainty quantification technology is used to output a prediction interval, which not only improves the prediction accuracy, but also provides a dynamic reference benchmark for abnormality detection; in the abnormality identification stage, the Gaussian mixture model is used to fit the complex distribution characteristics of the residual, and the abnormality is determined based on statistical significance, which significantly improves the detection sensitivity and reliability of early weak fault signals; in the health degree evaluation stage, the entropy weight method is used to dynamically assign weights, and a deviation measurement formula integrating the prediction confidence is constructed to generate an intuitive health index and state classification, which provides a scientific basis for operation and maintenance decision-making.

[0046] Embodiment 2, with reference to Figure 1 As a second embodiment of the present application, the embodiment provides a power equipment health evaluation method based on time series analysis and probability modeling, comprising:

[0047] In step S100, the pre-processing of the multi-source operation data comprises steps S101-S103.

[0048] S101: Standardize the original data to eliminate dimensional differences;

[0049] S102: Adopting interpolation method to complete missing values, when the observation value of a certain time is missing, the missing value is expressed as a linear combination of the previous historical observation values to predict by constructing an autoregressive model, and the autoregressive coefficients of the model are estimated by least squares method using historical sample data; In order to determine the optimal order of the model, that is, the number of historical observation values used for prediction, the Bayesian information criterion is used to weigh, the model under each order is evaluated, the goodness of fit and complexity of the model are considered comprehensively, and the order that makes the Bayesian information criterion value minimum is selected as the optimal model order;

[0050] The formula is:

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] Wherein, is the original sample data value at time t, that is, the original data directly read from the monitoring device without processing, is the first, second, and qth original sample data before time t, is the autoregressive coefficient, that is, the contribution weight of the original sample data to the prediction of the current value, is the random error term at time t, is the design matrix, the dimension of the matrix is , is the response vector, each element in the vector corresponds to the target value to be predicted by each row of the design matrix X, is the estimated value of the autoregressive coefficient vector, S is the total sample amount for estimating the Bayesian information criterion, and T is the transpose operation, is the maximum likelihood value of the model when the model order is q, is the Bayesian information criterion score of the model with order q.

[0057] S103: Using anomaly detection algorithm to eliminate outliers, calculating the median of all data points in the window as the local normal level of the current time point, calculating the absolute value of the difference between each data point and the median, and obtaining the median absolute deviation by calculating the median of the absolute value, measuring the fluctuation amplitude of the local data, and defining the abnormal degree of the current data point, reflecting the degree of deviation of the current point from the local normal level, comparing the calculated abnormal degree with the preset abnormal degree threshold, and the data points exceeding the threshold are determined as outliers and are eliminated, The formula is:

[0058]

[0059] wherein, is the raw sample data value at time t, i.e. the raw, unprocessed data read directly from the monitoring device, is the anomaly score at time t, is the local normal level at time t represented by the median within a sliding window, is the median absolute deviation within the window, and is a correction factor with a constant value of 1.4826.

[0060] Further, in the embodiments of the present application, in step S101, the data standardization includes adopting "robust-bucket" standardization for data standardization, specifically including steps A1-A2:

[0061] A1: For the original monitoring data sequence, the median and interquartile range are calculated, and robust scaling is performed by subtracting the data value from the median and dividing by the interquartile range to resist the influence of extreme outliers, which is expressed by the formula:

[0062]

[0063]

[0064] wherein, is the raw sample data value at time t, i.e. the raw, unprocessed data read directly from the monitoring device, is the median of the entire original sample sequence, is the quartile, is the upper quartile, i.e. 75% of the data in the sequence is less than the upper quartile, is the lower quartile, i.e. 25% of the data in the sequence is less than the lower quartile, is the data value after robust standardization.

[0065] A2: According to the obvious day-night and seasonal cycles of power data, the data is divided into multiple buckets according to time characteristics such as hours, weekdays / weekends or seasons, the mean and variance of the data after robust scaling are calculated independently in each bucket, and the final Z-score standardization is completed, which is expressed by the formula:

[0066]

[0067]

[0068]

[0069] wherein, is a data subset, is the number of samples contained in the subset b, is the mean of all robustly normalized data within the data subset b, is the variance of all robustly normalized data within the data subset b, is the final normalization result, is the standard deviation of data within the data subset b, and t is the time point.

[0070] In an optional embodiment, in step S101, the performing data normalization further comprises using sliding window adaptive normalization to establish a time sliding window centered on the current data point for each data point, calculating the data mean and standard deviation within the current window, and normalizing the current center data point, and the window size is dynamically adjusted according to the data sampling frequency.

[0071] In another optional embodiment, in step S101, the performing data normalization can further comprise using decimal scaling normalization to scale the absolute value maximum of the original data to below 1 by finding a suitable power of 10, and map the entire data sequence into the interval [-1, 1] or [0, 1], and the power is the logarithm of the maximum absolute value of the data.

[0072] Further, in the embodiments of the present application, in S100, the modeling the preprocessed multi-dimensional time series data comprises using a long short-term memory network (LSTM) as the core prediction model, which specifically comprises steps S111-S113:

[0073] S111: The model receives the preprocessed multi-dimensional time series data, and selectively remembers long-term historical information and forgets useless information through internal input gate, forget gate and output gate mechanisms.

[0074] S112: The network structure contains two layers of LSTM hidden layers to extract deep time series features, and obtains the next time running parameter point prediction value through a fully connected output layer.

[0075] S113: To quantify uncertainty, the Monte Carlo Dropout is enabled during inference, multiple forward propagation samples are performed, the mean of the prediction results is used as the final prediction, and the variance is used to construct the confidence interval.

[0076] In an optional embodiment, in S100, the modeling the preprocessed multi-dimensional time series data further comprises using a gated recurrent unit (GRU) network, which is a variant of LSTM, combines the input gate and the forget gate of LSTM into an update gate, and mixes the cell state and the hidden state, the structure is more simplified, the training and prediction processes are similar to LSTM, the input historical sequence is input, and the future prediction value is output.

[0077] In another optional implementation, in S100, the modeling on the preprocessed multi-dimensional time series data can further include using an autoregressive integrated moving average (ARIMA) model, the model is for a single variable time series, the sequence is made stationary by difference operation, and a linear combination model of autoregressive term and moving average term is established to predict future values, and model parameters need to be determined according to autocorrelation and partial autocorrelation diagrams.

[0078] In step S100, training the LSTM model includes steps B1-B4:

[0079] B1: After standardization, missing data completion and abnormal data elimination, multi-source operation data from the power grid management platform, the online monitoring system and the periodic test are constructed into a time series sample set, and an input-output pair is generated in a sliding time window manner, the input is a continuous multi-dimensional monitoring parameter sequence, and the output is a predicted value at the next moment, and the input sequence is represented as: .

[0080] B2: The LSTM network is composed of an input layer, two layers of hidden LSTM unit layers and one layer of fully connected output layer, wherein the hidden layer adopts a tanh activation function, the gate unit adopts a sigmoid function, the selective memory and forgetting of historical information are realized, the output layer adopts linear mapping, and the predicted result of the operation parameter is obtained.

[0081] B3: Mean square error (MSE) is used as a loss function, and the parameter is updated through an Adam optimization algorithm to speed up the convergence speed, and the loss function formula is represented as:

[0082]

[0083] Wherein, is the value of the loss function, M is the number of samples in the training batch, and j is the index of the sample in the batch, is the true observation value of the jth sample, is the predicted value of the LSTM model for the sample in the jth batch, and j is the index of the sample in the batch.

[0084] B4: In the model inference stage, the preprocessed monitoring sequence is input into the LSTM network to obtain the predicted value of the next cycle operation parameter.

[0085] In step S113, Monte Carlo Dropout includes steps C1-C6:

[0086] C1: Dropout activation is maintained in the training and inference stages of the model, each forward propagation is equivalent to randomly sampling a model from the model subset for prediction, and the sampling estimation of the Bayesian neural network is realized.

[0087] C2: By averaging the multiple sampling results in the prediction set, the prediction result of the next cycle operation parameter point is obtained, and it is used as the estimation of the future operation trend of the device. The prediction mean value formula of the next cycle operation parameter is:

[0088]

[0089] Where F is the number of Monte Carlo sampling, is the predicted value of the model at the next time t+1 in the fth Monte Carlo sampling, is the sample mean of the predicted value at the next time t+1.

[0090] C3: By calculating the variance of the sampling result, the fluctuation degree of the model prediction is quantified. The variance directly reflects the stability and credibility of the prediction result. The calculation formula of the prediction variance is:

[0091]

[0092] Where, is the sample variance of the predicted value at the next time t+1.

[0093] C4: Combined with the prediction mean, the standard deviation and the quantile of the normal distribution under a certain confidence level, the prediction confidence interval of the next cycle operation parameter is constructed. The interval directly describes the uncertainty boundary of the model prediction, which provides an important reference threshold for subsequent anomaly detection. The confidence interval is expressed as:

[0094]

[0095] Where, is the prediction mean of the ith monitoring parameter at the t+1 time point, is the quantile of the standard normal distribution, is the prediction standard deviation of the ith monitoring parameter at the t+1 time, is the marginal error.

[0096] C5: By calculating the residual between the actual observation value and the predicted value, the deviation degree of the device operation state relative to the normal trend is described. When the device is working normally, the residual sequence will fluctuate within a relatively stable range, and the statistical distribution characteristics are consistent.

[0097] C6: When the device has potential anomalies or performance degradation, the distribution of the residual shows obvious deviation or thickening of the tail. Continuous monitoring and analysis of the residual sequence, the residual formula is:

[0098]

[0099] Where, is the actual observation at time t, is the actual observation at time t, is the predicted value of the model at time t.

[0100] In the embodiments of the present application, in step S200, fitting the distribution of the residual sequence comprises fitting the residual distribution using a Gaussian Mixture Model (GMM), specifically comprising steps S201-S203:

[0101] S201: Selecting among different Gaussian component numbers using the Bayesian Information Criterion (BIC) to determine the optimal model complexity.

[0102] The residual probability distribution is expressed as:

[0103]

[0104] wherein, is the residual is the overall probability density under the entire Gaussian mixture model, is the difference between the observed value of the equipment operating parameter and the predicted value of the model, is the total number of Gaussian components in the Gaussian mixture model, is the mixing weight of the kth Gaussian component, is the probability density function of the kth Gaussian component, is the mean of the kth Gaussian component, i.e. the center position of the current component distribution, is the variance of the kth Gaussian component, i.e. the dispersion degree of the current component distribution;

[0105] The Bayesian Information Criterion is used to determine the total number of Gaussian components in the Gaussian mixture model, and the formula is expressed as:

[0106]

[0107] wherein, is the likelihood function under the given model parameters, is the natural logarithm of the likelihood function, is the number of free parameters of the model, i.e. the mean, variance and mixing weight parameters of each Gaussian component, is the total number of training samples.

[0108] S202: The Expectation Maximization (EM) algorithm is used for parameter estimation, the posterior probability of each residual point belonging to each Gaussian component is calculated, and the weight, mean and variance of each component are updated according to the posterior probability.

[0109] The responsibility formula is calculated as:

[0110]

[0111] wherein, is the responsibility of the lth residual sample is the posterior probability that the sample belongs to the kth Gaussian component, is the observation value of the lth residual sample, is the total number of Gaussian components in the Gaussian mixture model, is the probability density of the kth Gaussian component at the sample is the observation value of the lth residual sample, is a normalization term;

[0112] Using the current responsibility, the model parameters are re-estimated, which is formulated as:

[0113]

[0114]

[0115]

[0116] wherein, is the new weight of the kth Gaussian component, E is the total number of residual samples, is the new mean of the kth Gaussian component, l is the residual sample index, is the observation value of the lth residual sample, is the new variance of the kth Gaussian component

[0117] S203: When the log-likelihood function increment of the adjacent two iterations is less than , it is considered that the model has converged and the iteration is stopped; when the iteration number reaches 500 times and still does not converge, the calculation is terminated and the current parameter result is output, obtaining a probability density function composed of a plurality of Gaussian distributions weighted and combined, describing the complex distribution form of normal residuals.

[0118] In an optional embodiment, in step S200, the fitting of the distribution of the residual sequence further comprises using kernel density estimation (KDE) to obtain a continuous probability density function of the entire residual sequence by placing a kernel function (such as a Gaussian kernel) at each residual data point position, and then superimposing and normalizing all kernel functions.

[0119] In another optional embodiment, in step S200, the fitting of the distribution of the residual sequence can further comprise using a single Gaussian distribution model, assuming that the residual when the device is normally running obeys a single normal distribution, and directly calculating the sample mean and sample variance of the historical normal residual sequence to obtain the parameters of the current distribution.

[0120] It should be noted that the probability distribution model of the residual in the normal operation state is obtained by the Gaussian mixture model, which can reflect the first-order and second-order statistical characteristics such as the mean and variance of the residual, and can also capture the potential multi-peak structure and complex distribution characteristics, and is more accurate and flexible than the traditional single Gaussian distribution fitting method.

[0121] Further, in step S200, the performing of the abnormality determination includes, after the residual distribution modeling, performing the abnormality determination based on the probability density value of the residual, and when the probability density of a certain residual point is less than 95% of the threshold value, it is determined that the occurrence probability of the current point is extremely low, that is, there is an abnormal situation at the current time, and the reliability of the abnormality detection in the statistical sense is ensured.

[0122] In the embodiment of the application, in step S300, the health index of the computing device includes steps S301-S304:

[0123] S301: The range standardization is performed on the residual sequence of each parameter, and is mapped to the [0, 1] interval, and the formula is represented as:

[0124]

[0125] wherein, is the original residual value of the i th monitoring parameter at the o th sample (time point), is the minimum residual value of the i th monitoring parameter in all samples, is the maximum residual value of the i th monitoring parameter in all samples, is the standardized residual value, and o is the sample index.

[0126] S302: The normalized proportion of each parameter in all samples is calculated, and the formula is represented as:

[0127]

[0128] wherein, is the normalized proportion of the i th monitoring parameter at the o th sample, and O is the total number of samples, that is, the total number of time points.

[0129] S303: The information entropy of each parameter is calculated according to the information entropy formula; the smaller the information entropy, the greater the amount of information provided by the current parameter, and the higher the variation degree, and the formula is represented as:

[0130]

[0131]

[0132] wherein, is the information entropy of the i th monitoring parameter, is a normalization constant, is the natural logarithm.

[0133] S304: Calculate the difference coefficient, and normalize the difference coefficient to obtain the objective weight of each parameter. The health index is calculated by the weighted accumulation of the deviation of each parameter after standardization and the weight, and is mapped to 0-100 points to obtain the health index formula:

[0134]

[0135]

[0136]

[0137]

[0138] wherein, is the health index of the device, m is the number of key monitoring parameters participating in the health calculation, and i is the index of the monitoring parameter, is the difference coefficient of the i-th monitoring parameter, is the deviation measure of the i-th monitoring parameter, is the weight of the i-th monitoring parameter in the health calculation, is the residual of the i-th monitoring parameter at time t, is the standard deviation of the confidence interval of the predicted value of the i-th monitoring parameter at time t.

[0139] In an optional embodiment, in step S300, the health index of the computing device further comprises: using the analytic hierarchy process (AHP) combined with expert scoring to establish a hierarchical structure model of health evaluation, and inviting field experts to compare each parameter pairwise to construct a judgment matrix, calculating the characteristic vector of the matrix as the initial weight, and performing consistency check, and obtaining the comprehensive health degree by weighted synthesis.

[0140] In another optional embodiment, in step S300, the health index of the computing device can further comprise: using the equal weight weighting method, all monitoring parameters participating in the health calculation are equally important, and each parameter is assigned the same weight, and the health index is determined by the average value of the standardized deviations of all parameters.

[0141] Further, in step S300, the division of the device state level comprises: when the health index is determined that the device is in a healthy state, and runs stably without additional intervention; when is determined to be in a sub-healthy state, and the device runs with slight fluctuations, and it is recommended to increase the monitoring frequency; when When this occurs, it is determined to be a warning state, indicating that multiple equipment indicators have deviated significantly, and a maintenance plan needs to be formulated; when If the equipment is deemed to be in a high-risk state, it indicates that the equipment has been in an abnormal state for a long time. Immediate shutdown or replacement measures should be taken to prevent serious accidents.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0143] Example 3, referring to Figure 2 This is the third embodiment of the present invention, which provides a power equipment health evaluation system based on time series analysis and probabilistic modeling, including a multi-source data preprocessing module, a time series dynamic prediction module, a residual probability modeling module, a health quantification assessment module, and an intelligent diagnosis and decision support module.

[0144] The multi-source data preprocessing module is used to eliminate the dimensional differences of different physical parameters through standardization, and adopts a bucketing standardization strategy that combines time features. It uses an autoregressive completion method to handle missing data and uses Hampel filtering to remove outliers introduced by interference, and outputs a standardized dataset.

[0145] The time-series dynamic prediction module is used to perform deep modeling of preprocessed multidimensional time-series data using a long short-term memory network. It learns and memorizes long-term dependencies in historical data through a gating mechanism to predict the future cycle operating parameters of the equipment. By integrating Monte Carlo Dropout technology, it performs random forward propagation sampling during the inference stage, outputs the prediction mean, calculates the prediction variance, and forms the prediction interval.

[0146] The residual probability modeling module is used to calculate the residual sequence between the actual observed value and the predicted value, and to fit the probability distribution of the sequence using a Gaussian mixture model. The module also uses the expectation-maximization algorithm to iteratively optimize the weight, mean, and variance of each Gaussian component, and determines the optimal number of components based on the Bayesian information criterion.

[0147] The health measurement and assessment module is used to integrate the deviation information of the monitoring parameters, calculate the health index, and use the entropy weight method to calculate the dynamic weight based on the data dispersion of the residual sequence of each parameter. The final health score is obtained through weighted accumulation, and the equipment status is divided into four levels: healthy, sub-healthy, warning, and high risk.

[0148] The intelligent diagnosis and decision support module is used to integrate anomaly detection results and health scores to generate visual output, providing maintenance personnel with a global awareness of the equipment's operational status throughout its entire lifecycle, enabling them to promptly identify potential risks and formulate maintenance strategies based on the status level.

[0149] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0150] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that:

[0151] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0152] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0153] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0154] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A method for health assessment of power equipment based on time series analysis and probabilistic modeling, characterized in that: include, Acquire multi-source operating data of power equipment, preprocess the multi-source operating data, and use a time series prediction model to model the preprocessed multi-dimensional time series data, outputting the predicted values ​​and confidence intervals of future operating parameters; Calculate the residual sequence between the actual observed values ​​and the predicted values, fit the distribution of the residual sequence using a probability distribution model, establish a statistical characteristic model under normal operating conditions of the equipment, and make anomaly judgments based on the probability density of the residual points according to the fitted probability distribution model. Based on the degree of deviation and weight of each monitoring parameter, the health index of the equipment is calculated by weighted accumulation, and the equipment status level is classified according to the health index.

2. The power equipment health assessment method based on time series analysis and probabilistic modeling as described in claim 1, characterized in that: The preprocessing of multi-source operational data includes standardizing the raw data to eliminate dimensional differences, using interpolation methods to fill in missing values, and using anomaly detection algorithms to remove outliers.

3. The power equipment health assessment method based on time series analysis and probabilistic modeling as described in claim 2, characterized in that: The modeling of the preprocessed multidimensional time series data includes selectively remembering and forgetting historical information through an internal gating mechanism to capture long-term dependencies in the time series data. During the model inference stage, an uncertainty quantification mechanism is introduced. The distribution of the prediction results is generated through forward propagation sampling, the prediction mean is output and the prediction variance is calculated to form a prediction interval with a confidence level.

4. The power equipment health assessment method based on time series analysis and probabilistic modeling as described in claim 3, characterized in that: The fitting of the residual sequence distribution includes fitting the residual distribution by combining multiple component distributions to characterize potential multimodal and asymmetric characteristics. By selecting from candidate models using information criteria, the optimal number of distribution components is determined, achieving a balance between model fitting accuracy and complexity.

5. The power equipment health assessment method based on time series analysis and probabilistic modeling as described in claim 4, characterized in that: The probability distribution model includes using a Gaussian mixture model and employing the expectation-maximization algorithm for iterative optimization. Based on the current parameter estimates, the posterior probability of each residual sample belonging to each Gaussian component in the model is calculated. Using the calculated posterior probability, the weight, mean, and variance parameters of each Gaussian component in the model are re-estimated until the log-likelihood function increment between two adjacent iterations is less than a preset threshold, thereby obtaining the parameters of the Gaussian mixture model. The residual probability distribution is expressed as: in, For residuals The overall probability density under the entire Gaussian mixture model The residual is the difference between the observed values ​​and the model predictions of the equipment's operating parameters. This represents the total number of Gaussian components in the Gaussian mixture model. The mixing weights for the k-th Gaussian component. Let be the probability density function of the k-th Gaussian component. The mean of the k-th Gaussian component represents the center position of the current component distribution. The variance of the k-th Gaussian component represents the degree of dispersion of the current component distribution. The total number of Gaussian components in a Gaussian mixture model is determined using the Bayesian information criterion, expressed by the following formula: in, Let be the likelihood function given the model parameters. Let be the natural logarithm of the likelihood function. This represents the number of free parameters of the model, i.e., the mean, variance, and mixture weights of each Gaussian component. This represents the total number of training samples.

6. The power equipment health assessment method based on time series analysis and probabilistic modeling as described in claim 5, characterized in that: The anomaly determination includes setting a probability density threshold of 95% that matches the confidence level. When the probability density value of a certain residual sample point under the Gaussian mixture model is lower than the threshold of 95%, the current device operating status is determined to be abnormal.

7. The power equipment health assessment method based on time series analysis and probabilistic modeling as described in claim 6, characterized in that: The health index of the computing device includes defining a deviation measure for each monitoring parameter, which is composed of the ratio of the absolute value of the residual to the standard deviation of the prediction confidence interval. The weight of each monitoring parameter is dynamically determined using the entropy weight method. The deviation measure of each parameter is multiplied by the matching weight and accumulated. The result is deducted from the benchmark score and mapped to the range of 0 to 100 to obtain the health index. The entropy weight method is used to dynamically determine the weight of each monitoring parameter, which includes standardizing the range of the residual sequence of each parameter, calculating the standardization ratio of each parameter in all samples, calculating the information entropy of each parameter based on the standardization ratio, calculating the difference coefficient of each parameter based on the information entropy, and normalizing the difference coefficient to obtain the final weight of each parameter. The formula for obtaining the health index is expressed as: in, Here, m represents the health index of the equipment, i is the number of key monitoring parameters involved in the health calculation, and i is the index of the monitoring parameter. This is a measure of the deviation of the i-th monitoring parameter. Let i be the weight of the i-th monitoring parameter in the health score calculation. Let be the residual of the i-th monitoring parameter at time t. Let be the standard deviation of the confidence interval of the predicted value of the i-th monitoring parameter at time t; The classification of device status levels includes, when the health index... When it is determined to be a healthy state, when When it is determined to be a sub-healthy state, when When it is determined to be a warning state, when It was determined to be a high-risk situation.

8. A power equipment health assessment system based on time series analysis and probabilistic modeling, employing the power equipment health assessment method based on time series analysis and probabilistic modeling as described in any one of claims 1 to 7, characterized in that, It includes a multi-source data preprocessing module, a time-series dynamic prediction module, a residual probability modeling module, a health measurement and assessment module, and an intelligent diagnosis and decision support module. The multi-source data preprocessing module is used to eliminate the dimensional differences of different physical parameters through standardization, and adopts a bucket standardization strategy combined with time characteristics, uses an autoregressive completion method to handle missing data, and uses Hampel filtering to remove outliers introduced by interference, and outputs a standardized dataset. The time-series dynamic prediction module is used to perform deep modeling of preprocessed multidimensional time-series data using a long short-term memory network, learn and memorize long-term dependencies in historical data through a gating mechanism, predict the future cycle operating parameters of the equipment, and perform random forward propagation sampling during the inference stage by integrating Monte Carlo Dropout technology, output the prediction mean, calculate the prediction variance, and form the prediction interval. The residual probability modeling module is used to calculate the residual sequence between the actual observed value and the predicted value, and to fit the probability distribution of the sequence using a Gaussian mixture model. The module also uses the expectation-maximization algorithm to iteratively optimize the weight, mean, and variance of each Gaussian component, and determines the optimal number of components based on the Bayesian information criterion. The health measurement and assessment module is used to integrate the deviation information of the monitoring parameters, calculate the health index, and use the entropy weight method to calculate the dynamic weight based on the data dispersion of the residual sequence of each parameter. The final health score is obtained through weighted accumulation, and the equipment status is divided into four levels: healthy, sub-healthy, warning and high risk. The intelligent diagnosis and decision support module is used to integrate anomaly detection results and health scores to generate visual output, providing maintenance personnel with a global awareness of the equipment's operational status throughout its entire lifecycle, enabling them to promptly identify potential risks and formulate maintenance strategies based on the status level.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power equipment health assessment method based on time series analysis and probabilistic modeling as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power equipment health evaluation method based on time series analysis and probabilistic modeling as described in any one of claims 1 to 7.

Images